A model of contiguous sequential patterns of songs, in order to acquire the knowledge about which songs sound well together in a playlist.
A research paper that I have presented in September 2006 at the European Conference on Case Based Reasoning (ECCBR '06).
By Claudio Baccigalupo and Enric Plaza
A Case-Based Song Scheduler for Group Customised RadioUri Levanon
Presented in Aug 2007 at the International Conference on Case Based Reasoning (ICCBR '07), where it was granted the Best Application Paper Award
The idea discussed is called Poolcasting - a Case-Based Reasoning technique to schedule the songs to play on each channel, combining several different techniques (CBR, Pattern Mining, Group Satisfaction).
By Claudio Baccigalupo and Enric Plaza
Pay-per-click (PPC) is an advertising model in which advertisers pay for click-throughs to their website. Ads are served based on keywords or themes. An effective PPC campaign can help you reduce the cost per click and increase profit. There are four key components of a winning PPC advertising program: Explore, Evaluate, Expand and Enhance.
A Case-Based Song Scheduler for Group Customised RadioUri Levanon
Presented in Aug 2007 at the International Conference on Case Based Reasoning (ICCBR '07), where it was granted the Best Application Paper Award
The idea discussed is called Poolcasting - a Case-Based Reasoning technique to schedule the songs to play on each channel, combining several different techniques (CBR, Pattern Mining, Group Satisfaction).
By Claudio Baccigalupo and Enric Plaza
Pay-per-click (PPC) is an advertising model in which advertisers pay for click-throughs to their website. Ads are served based on keywords or themes. An effective PPC campaign can help you reduce the cost per click and increase profit. There are four key components of a winning PPC advertising program: Explore, Evaluate, Expand and Enhance.
Applications of AI and NLP to advance Music Recommendations on Voice AssistantsAI Publications
This paper builds on the popular use case of music requests on voice assistants like Siri, Google Assistant, Alexa, and others and explores the different AI and NLP techniques. The paper particularly focuses on how each of these techniques can be applied in the context of musical recommendations and experiences on voice assistants. It enumerates specific problems in the space of music recommendations and illustrates how specific techniques like multi-armed bandits can be applied.
DETECTING AND LOCATING PLAGIARISM OF MUSIC MELODIES BY PATH EXPLORATION OVER ...csandit
To the best of our knowledge, the issues of automatic detection of music plagiarism have never
been addressed before. This paper presents the design of an Automatic Music Melody
Plagiarism Detection (AMMPD) method to detect and locate the possible plagiarism in music
melodies. The key contribution of the work is an algorithm proposed to address the challenging
issues encountered in the AMMPD problem, including (1) the inexact matching of noisy and
inaccurate pitches of music audio and (2) the fast detection and positioning of similar
subsegments between suspicious music audio. The major novelty of the proposed method is that
we address the above two issues in temporal domain by means of a novel path finding approach
on a binarized 2-D bit mask in spatial domain. In fact, the proposed AMMPD method can not
only identify the similar pieces inside two suspicious music melodies, but also retrieve music
audio of similar melodies from a music database given a humming or singing query.
Experiments have been conducted to assess the overall performance and examine the effects of
various parameters introduced in the proposed method.
The kusc classical music dataset for audio key findingijma
In this paper, we present a benchmark dataset based on the KUSC classical music collection and provide
baseline key-finding comparison results. Audio key finding is a basic music information retrieval task; it
forms an essential component of systems for music segmentation, similarity assessment, and mood
detection. Due to copyright restrictions and a labor-intensive annotation process, audio key finding
algorithms have only been evaluated using small proprietary datasets to date. To create a common base for
systematic comparisons, we have constructed a dataset comprising of more than 3,000 excerpts of classical
music. The excerpts are made publicly accessible via commonly used acoustic features such as pitch-based
spectrograms and chromagrams. We introduce a hybrid annotation scheme that combines the use of title
keys with expert validation and correction of only the challenging cases. The expert musicians also provide
ratings of key recognition difficulty. Other meta-data include instrumentation. As demonstration of use of
the dataset, and to provide initial benchmark comparisons for evaluating new algorithms, we conduct a
series of experiments reporting key determination accuracy of four state-of-the-art algorithms. We further
show the importance of considering factors such as estimated tuning frequency, key strength or confidence
value, and key recognition difficulty in key finding. In the future, we plan to expand the dataset to include
meta-data for other music information retrieval tasks.
Mining Melodic Patterns in Large Audio Collections of Indian Art MusicSankalp Gulati
More info: http://mtg.upf.edu/node/3108
Abstract: Discovery of repeating structures in music is fundamental to its analysis, understanding and interpretation. We present a data-driven approach for the discovery of short-time melodic patterns in large collections of Indian art music. The approach first discovers melodic patterns within an audio recording and subsequently searches for their repetitions in the entire music collection. We compute similarity between melodic patterns using dynamic time warping (DTW). Furthermore, we investigate four different variants of the DTW cost function for rank refinement of the obtained results. The music collection used in this study comprises 1,764 audio recordings with a total duration of 365 hours. Over 13 trillion DTW distance computations are done for the entire dataset. Due to the computational complexity of the task, different lower bounding and early abandoning techniques are applied during DTW distance computation. An evaluation based on expert feedback on a subset of the dataset shows that the discovered melodic patterns are musically relevant. Several musically interesting relationships are discovered, yielding further scope for establishing novel similarity measures based on melodic patterns. The discovered melodic patterns can further be used in challenging computational tasks such as automatic raga recognition, composition identification and music recommendation.
Performance Comparison of Musical Instrument Family Classification Using Soft...Waqas Tariq
Nowadays, it appears essential to design automatic and efficacious classification algorithm for the musical instruments. Automatic classification of musical instruments is made by extracting relevant features from the audio samples, afterward classification algorithm is used (using these extracted features) to identify into which of a set of classes, the sound sample is possible to fit. The aim of this paper is to demonstrate the viability of soft set for audio signal classification. A dataset of 104 (single monophonic notes) pieces of Traditional Pakistani musical instruments were designed. Feature extraction is done using two feature sets namely perception based and mel-frequency cepstral coefficients (MFCCs). In a while, two different classification techniques are applied for classification task, which are soft set (comparison table) and fuzzy soft set (similarity measurement). Experimental results show that both classifiers can perform well on numerical data. However, soft set achieved accuracy up to 94.26% with best generated dataset. Consequently, these promising results provide new possibilities for soft set in classifying musical instrument sounds. Based on the analysis of the results, this study offers a new view on automatic instrument classification
Applications of AI and NLP to advance Music Recommendations on Voice AssistantsAI Publications
This paper builds on the popular use case of music requests on voice assistants like Siri, Google Assistant, Alexa, and others and explores the different AI and NLP techniques. The paper particularly focuses on how each of these techniques can be applied in the context of musical recommendations and experiences on voice assistants. It enumerates specific problems in the space of music recommendations and illustrates how specific techniques like multi-armed bandits can be applied.
DETECTING AND LOCATING PLAGIARISM OF MUSIC MELODIES BY PATH EXPLORATION OVER ...csandit
To the best of our knowledge, the issues of automatic detection of music plagiarism have never
been addressed before. This paper presents the design of an Automatic Music Melody
Plagiarism Detection (AMMPD) method to detect and locate the possible plagiarism in music
melodies. The key contribution of the work is an algorithm proposed to address the challenging
issues encountered in the AMMPD problem, including (1) the inexact matching of noisy and
inaccurate pitches of music audio and (2) the fast detection and positioning of similar
subsegments between suspicious music audio. The major novelty of the proposed method is that
we address the above two issues in temporal domain by means of a novel path finding approach
on a binarized 2-D bit mask in spatial domain. In fact, the proposed AMMPD method can not
only identify the similar pieces inside two suspicious music melodies, but also retrieve music
audio of similar melodies from a music database given a humming or singing query.
Experiments have been conducted to assess the overall performance and examine the effects of
various parameters introduced in the proposed method.
The kusc classical music dataset for audio key findingijma
In this paper, we present a benchmark dataset based on the KUSC classical music collection and provide
baseline key-finding comparison results. Audio key finding is a basic music information retrieval task; it
forms an essential component of systems for music segmentation, similarity assessment, and mood
detection. Due to copyright restrictions and a labor-intensive annotation process, audio key finding
algorithms have only been evaluated using small proprietary datasets to date. To create a common base for
systematic comparisons, we have constructed a dataset comprising of more than 3,000 excerpts of classical
music. The excerpts are made publicly accessible via commonly used acoustic features such as pitch-based
spectrograms and chromagrams. We introduce a hybrid annotation scheme that combines the use of title
keys with expert validation and correction of only the challenging cases. The expert musicians also provide
ratings of key recognition difficulty. Other meta-data include instrumentation. As demonstration of use of
the dataset, and to provide initial benchmark comparisons for evaluating new algorithms, we conduct a
series of experiments reporting key determination accuracy of four state-of-the-art algorithms. We further
show the importance of considering factors such as estimated tuning frequency, key strength or confidence
value, and key recognition difficulty in key finding. In the future, we plan to expand the dataset to include
meta-data for other music information retrieval tasks.
Mining Melodic Patterns in Large Audio Collections of Indian Art MusicSankalp Gulati
More info: http://mtg.upf.edu/node/3108
Abstract: Discovery of repeating structures in music is fundamental to its analysis, understanding and interpretation. We present a data-driven approach for the discovery of short-time melodic patterns in large collections of Indian art music. The approach first discovers melodic patterns within an audio recording and subsequently searches for their repetitions in the entire music collection. We compute similarity between melodic patterns using dynamic time warping (DTW). Furthermore, we investigate four different variants of the DTW cost function for rank refinement of the obtained results. The music collection used in this study comprises 1,764 audio recordings with a total duration of 365 hours. Over 13 trillion DTW distance computations are done for the entire dataset. Due to the computational complexity of the task, different lower bounding and early abandoning techniques are applied during DTW distance computation. An evaluation based on expert feedback on a subset of the dataset shows that the discovered melodic patterns are musically relevant. Several musically interesting relationships are discovered, yielding further scope for establishing novel similarity measures based on melodic patterns. The discovered melodic patterns can further be used in challenging computational tasks such as automatic raga recognition, composition identification and music recommendation.
Performance Comparison of Musical Instrument Family Classification Using Soft...Waqas Tariq
Nowadays, it appears essential to design automatic and efficacious classification algorithm for the musical instruments. Automatic classification of musical instruments is made by extracting relevant features from the audio samples, afterward classification algorithm is used (using these extracted features) to identify into which of a set of classes, the sound sample is possible to fit. The aim of this paper is to demonstrate the viability of soft set for audio signal classification. A dataset of 104 (single monophonic notes) pieces of Traditional Pakistani musical instruments were designed. Feature extraction is done using two feature sets namely perception based and mel-frequency cepstral coefficients (MFCCs). In a while, two different classification techniques are applied for classification task, which are soft set (comparison table) and fuzzy soft set (similarity measurement). Experimental results show that both classifiers can perform well on numerical data. However, soft set achieved accuracy up to 94.26% with best generated dataset. Consequently, these promising results provide new possibilities for soft set in classifying musical instrument sounds. Based on the analysis of the results, this study offers a new view on automatic instrument classification
Similar to Case-based Sequential Ordering of Songs for Playlist Recommendation (20)
Discover more inspiring stuff at HTTP://WWW.FACEBOOK.COM/INDIECULTUREBOX
Does being part of the commodity crowd erodes your value?
By Rajesh Setty
(a ChangeThis manifesto)
Brief: Traditional marketing practices no longer work. They're being usurped by customer-driven referrals as the valuable new currency in an organization's growth.
(A ChangeThis manifesto)
By Ben McConnel & Jackie Huba
While creating a growth business can be exhilarating,
many entrepreneurs—especially those starting a
company for the first time—donʼt pay enough
attention to some core issues surrounding the
financial management of their businesses.
By Brad Feld
Marketing guru and agent of change Seth Godin writes that for your company to do more, sometimes you need to do less! Stop trying to be all things to all people (or customers), and focus (in your life and in your work) on your core strengths. Your business (and probably your sanity) are likely to improve.
By Seth Godin
a ChangeThis manifest
Grown-ups never understand anything for themselves,
and it is tiresome for children to be always and forever
explaining things to them.
By Jason Kotecki
a ChangeThis manifest
Product benefit doesnʼt excite us. Belief in humanity and human potential excites us.
Think less about what your product does, and think more about human potential.
What statement about humanity does your product make?
By Hugh MacLeod
a ChangeThis manifest
A Physics Of Ideas - Measuring the Physical Properties of MemesUri Levanon
One of the main obstacles to productivity today is the
growing problem of information overload. Information
overload results because we lack effective tools for
automatically organizing information collections into
meaningful and relevant chunks.
By Nova Spivack
a ChangeThis Manifesto
Great new things often come from the margins,
and yet the people who discover them are looked
down on by everyone, including themselves.
By Paul Graham (Y Combinator)
A ChangeThis manifest
Modern Radio Frequency Access Control Systems: The Key to Efficiency and SafetyAITIX LLC
Today's fast-paced environment worries companies of all sizes about efficiency and security. Businesses are constantly looking for new and better solutions to solve their problems, whether it's data security or facility access. RFID for access control technologies have revolutionized this.
240529_Teleprotection Global Market Report 2024.pdfMadhura TBRC
The teleprotection market size has grown
exponentially in recent years. It will grow from
$21.92 billion in 2023 to $28.11 billion in 2024 at a
compound annual growth rate (CAGR) of 28.2%. The
teleprotection market size is expected to see
exponential growth in the next few years. It will grow
to $70.77 billion in 2028 at a compound annual
growth rate (CAGR) of 26.0%.
Meet Dinah Mattingly – Larry Bird’s Partner in Life and Loveget joys
Get an intimate look at Dinah Mattingly’s life alongside NBA icon Larry Bird. From their humble beginnings to their life today, discover the love and partnership that have defined their relationship.
As a movie director, I am often asked about the process of creating an indie film. It's a journey of passion, perseverance, and planning, and today, I'm going to take you through it.
Maximizing Your Streaming Experience with XCIPTV- Tips for 2024.pdfXtreame HDTV
In today’s digital age, streaming services have become an integral part of our entertainment lives. Among the myriad of options available, XCIPTV stands out as a premier choice for those seeking seamless, high-quality streaming. This comprehensive guide will delve into the features, benefits, and user experience of XCIPTV, illustrating why it is a top contender in the IPTV industry.
Young Tom Selleck: A Journey Through His Early Years and Rise to Stardomgreendigital
Introduction
When one thinks of Hollywood legends, Tom Selleck is a name that comes to mind. Known for his charming smile, rugged good looks. and the iconic mustache that has become synonymous with his persona. Tom Selleck has had a prolific career spanning decades. But, the journey of young Tom Selleck, from his early years to becoming a household name. is a story filled with determination, talent, and a touch of luck. This article delves into young Tom Selleck's life, background, early struggles. and pivotal moments that led to his rise in Hollywood.
Follow us on: Pinterest
Early Life and Background
Family Roots and Childhood
Thomas William Selleck was born in Detroit, Michigan, on January 29, 1945. He was the second of four children in a close-knit family. His father, Robert Dean Selleck, was a real estate investor and executive. while his mother, Martha Selleck, was a homemaker. The Selleck family relocated to Sherman Oaks, California. when Tom was a child, setting the stage for his future in the entertainment industry.
Education and Early Interests
Growing up, young Tom Selleck was an active and athletic child. He attended Grant High School in Van Nuys, California. where he excelled in sports, particularly basketball. His tall and athletic build made him a standout player, and he earned a basketball scholarship to the University of Southern California (U.S.C.). While at U.S.C., Selleck studied business administration. but his interests shifted toward acting.
Discovery of Acting Passion
Tom Selleck's journey into acting was serendipitous. During his time at U.S.C., a drama coach encouraged him to try acting. This nudge led him to join the Hills Playhouse, where he began honing his craft. Transitioning from an aspiring athlete to an actor took time. but young Tom Selleck became drawn to the performance world.
Early Career Struggles
Breaking Into the Industry
The path to stardom was a challenging one for young Tom Selleck. Like many aspiring actors, he faced many rejections and struggled to find steady work. A series of minor roles and guest appearances on television shows marked his early career. In 1965, he debuted on the syndicated show "The Dating Game." which gave him some exposure but did not lead to immediate success.
The Commercial Breakthrough
During the late 1960s and early 1970s, Selleck began appearing in television commercials. His rugged good looks and charismatic presence made him a popular brand choice. He starred in advertisements for Pepsi-Cola, Revlon, and Close-Up toothpaste. These commercials provided financial stability and helped him gain visibility in the industry.
Struggling Actor in Hollywood
Despite his success in commercials. breaking into large acting roles remained a challenge for young Tom Selleck. He auditioned and took on small parts in T.V. shows and movies. Some of his early television appearances included roles in popular series like Lancer, The F.B.I., and Bracken's World. But, it would take a
As a film director, I have always been awestruck by the magic of animation. Animation, a medium once considered solely for the amusement of children, has undergone a significant transformation over the years. Its evolution from a rudimentary form of entertainment to a sophisticated form of storytelling has stirred my creativity and expanded my vision, offering limitless possibilities in the realm of cinematic storytelling.
Tom Selleck Net Worth: A Comprehensive Analysisgreendigital
Over several decades, Tom Selleck, a name synonymous with charisma. From his iconic role as Thomas Magnum in the television series "Magnum, P.I." to his enduring presence in "Blue Bloods," Selleck has captivated audiences with his versatility and charm. As a result, "Tom Selleck net worth" has become a topic of great interest among fans. and financial enthusiasts alike. This article delves deep into Tom Selleck's wealth, exploring his career, assets, endorsements. and business ventures that contribute to his impressive economic standing.
Follow us on: Pinterest
Early Life and Career Beginnings
The Foundation of Tom Selleck's Wealth
Born on January 29, 1945, in Detroit, Michigan, Tom Selleck grew up in Sherman Oaks, California. His journey towards building a large net worth began with humble origins. , Selleck pursued a business administration degree at the University of Southern California (USC) on a basketball scholarship. But, his interest shifted towards acting. leading him to study at the Hills Playhouse under Milton Katselas.
Minor roles in television and films marked Selleck's early career. He appeared in commercials and took on small parts in T.V. series such as "The Dating Game" and "Lancer." These initial steps, although modest. laid the groundwork for his future success and the growth of Tom Selleck net worth. Breakthrough with "Magnum, P.I."
The Role that Defined Tom Selleck's Career
Tom Selleck's breakthrough came with the role of Thomas Magnum in the CBS television series "Magnum, P.I." (1980-1988). This role made him a household name and boosted his net worth. The series' popularity resulted in Selleck earning large salaries. leading to financial stability and increased recognition in Hollywood.
"Magnum P.I." garnered high ratings and critical acclaim during its run. Selleck's portrayal of the charming and resourceful private investigator resonated with audiences. making him one of the most beloved television actors of the 1980s. The success of "Magnum P.I." played a pivotal role in shaping Tom Selleck net worth, establishing him as a major star.
Film Career and Diversification
Expanding Tom Selleck's Financial Portfolio
While "Magnum, P.I." was a cornerstone of Selleck's career, he did not limit himself to television. He ventured into films, further enhancing Tom Selleck net worth. His filmography includes notable movies such as "Three Men and a Baby" (1987). which became the highest-grossing film of the year, and its sequel, "Three Men and a Little Lady" (1990). These box office successes contributed to his wealth.
Selleck's versatility allowed him to transition between genres. from comedies like "Mr. Baseball" (1992) to westerns such as "Quigley Down Under" (1990). This diversification showcased his acting range. and provided many income streams, reinforcing Tom Selleck net worth.
Television Resurgence with "Blue Bloods"
Sustaining Wealth through Consistent Success
In 2010, Tom Selleck began starring as Frank Reagan i
Skeem Saam in June 2024 available on ForumIsaac More
Monday, June 3, 2024 - Episode 241: Sergeant Rathebe nabs a top scammer in Turfloop. Meikie is furious at her uncle's reaction to the truth about Ntswaki.
Tuesday, June 4, 2024 - Episode 242: Babeile uncovers the truth behind Rathebe’s latest actions. Leeto's announcement shocks his employees, and Ntswaki’s ordeal haunts her family.
Wednesday, June 5, 2024 - Episode 243: Rathebe blocks Babeile from investigating further. Melita warns Eunice to stay clear of Mr. Kgomo.
Thursday, June 6, 2024 - Episode 244: Tbose surrenders to the police while an intruder meddles in his affairs. Rathebe's secret mission faces a setback.
Friday, June 7, 2024 - Episode 245: Rathebe’s antics reach Kganyago. Tbose dodges a bullet, but a nightmare looms. Mr. Kgomo accuses Melita of witchcraft.
Monday, June 10, 2024 - Episode 246: Ntswaki struggles on her first day back at school. Babeile is stunned by Rathebe’s romance with Bullet Mabuza.
Tuesday, June 11, 2024 - Episode 247: An unexpected turn halts Rathebe’s investigation. The press discovers Mr. Kgomo’s affair with a young employee.
Wednesday, June 12, 2024 - Episode 248: Rathebe chases a criminal, resorting to gunfire. Turf High is rife with tension and transfer threats.
Thursday, June 13, 2024 - Episode 249: Rathebe traps Kganyago. John warns Toby to stop harassing Ntswaki.
Friday, June 14, 2024 - Episode 250: Babeile is cleared to investigate Rathebe. Melita gains Mr. Kgomo’s trust, and Jacobeth devises a financial solution.
Monday, June 17, 2024 - Episode 251: Rathebe feels the pressure as Babeile closes in. Mr. Kgomo and Eunice clash. Jacobeth risks her safety in pursuit of Kganyago.
Tuesday, June 18, 2024 - Episode 252: Bullet Mabuza retaliates against Jacobeth. Pitsi inadvertently reveals his parents’ plans. Nkosi is shocked by Khwezi’s decision on LJ’s future.
Wednesday, June 19, 2024 - Episode 253: Jacobeth is ensnared in deceit. Evelyn is stressed over Toby’s case, and Letetswe reveals shocking academic results.
Thursday, June 20, 2024 - Episode 254: Elizabeth learns Jacobeth is in Mpumalanga. Kganyago's past is exposed, and Lehasa discovers his son is in KZN.
Friday, June 21, 2024 - Episode 255: Elizabeth confirms Jacobeth’s dubious activities in Mpumalanga. Rathebe lies about her relationship with Bullet, and Jacobeth faces theft accusations.
Monday, June 24, 2024 - Episode 256: Rathebe spies on Kganyago. Lehasa plans to retrieve his son from KZN, fearing what awaits.
Tuesday, June 25, 2024 - Episode 257: MaNtuli fears for Kwaito’s safety in Mpumalanga. Mr. Kgomo and Melita reconcile.
Wednesday, June 26, 2024 - Episode 258: Kganyago makes a bold escape. Elizabeth receives a shocking message from Kwaito. Mrs. Khoza defends her husband against scam accusations.
Thursday, June 27, 2024 - Episode 259: Babeile's skillful arrest changes the game. Tbose and Kwaito face a hostage crisis.
Friday, June 28, 2024 - Episode 260: Two women face the reality of being scammed. Turf is rocked by breaking
Hollywood Actress - The 250 hottest galleryZsolt Nemeth
Hollywood Actress amazon album eminent worldwide media, female-singer, actresses, alhletina-woman, 250 collection.
Highest and photoreal-print exclusive testament PC collage.
Focused television virtuality crime, novel.
The sheer afterlife of the work is activism-like hollywood-actresses point com.
173 Illustrate, 250 gallery, 154 blog, 120 TV serie logo, 17 TV president logo, 183 active hyperlink.
HD AI face enhancement 384 page plus Bowker ISBN, Congress LLCL or US Copyright.
In the vast landscape of cinema, stories have been told, retold, and reimagined in countless ways. At the heart of this narrative evolution lies the concept of a "remake". A successful remake allows us to revisit cherished tales through a fresh lens, often reflecting a different era's perspective or harnessing the power of advanced technology. Yet, the question remains, what makes a remake successful? Today, we will delve deeper into this subject, identifying the key ingredients that contribute to the success of a remake.
From the Editor's Desk: 115th Father's day Celebration - When we see Father's day in Hindu context, Nanda Baba is the most vivid figure which comes to the mind. Nanda Baba who was the foster father of Lord Krishna is known to provide love, care and affection to Lord Krishna and Balarama along with his wife Yashoda; Letter’s to the Editor: Mother's Day - Mother is a precious life for their children. Mother is life breath for her children. Mother's lap is the world happiness whose debt can never be paid.
Scandal! Teasers June 2024 on etv Forum.co.zaIsaac More
Monday, 3 June 2024
Episode 47
A friend is compelled to expose a manipulative scheme to prevent another from making a grave mistake. In a frantic bid to save Jojo, Phakamile agrees to a meeting that unbeknownst to her, will seal her fate.
Tuesday, 4 June 2024
Episode 48
A mother, with her son's best interests at heart, finds him unready to heed her advice. Motshabi finds herself in an unmanageable situation, sinking fast like in quicksand.
Wednesday, 5 June 2024
Episode 49
A woman fabricates a diabolical lie to cover up an indiscretion. Overwhelmed by guilt, she makes a spontaneous confession that could be devastating to another heart.
Thursday, 6 June 2024
Episode 50
Linda unwittingly discloses damning information. Nhlamulo and Vuvu try to guide their friend towards the right decision.
Friday, 7 June 2024
Episode 51
Jojo's life continues to spiral out of control. Dintle weaves a web of lies to conceal that she is not as successful as everyone believes.
Monday, 10 June 2024
Episode 52
A heated confrontation between lovers leads to a devastating admission of guilt. Dintle's desperation takes a new turn, leaving her with dwindling options.
Tuesday, 11 June 2024
Episode 53
Unable to resort to violence, Taps issues a verbal threat, leaving Mdala unsettled. A sister must explain her life choices to regain her brother's trust.
Wednesday, 12 June 2024
Episode 54
Winnie makes a very troubling discovery. Taps follows through on his threat, leaving a woman reeling. Layla, oblivious to the truth, offers an incentive.
Thursday, 13 June 2024
Episode 55
A nosy relative arrives just in time to thwart a man's fatal decision. Dintle manipulates Khanyi to tug at Mo's heartstrings and get what she wants.
Friday, 14 June 2024
Episode 56
Tlhogi is shocked by Mdala's reaction following the revelation of their indiscretion. Jojo is in disbelief when the punishment for his crime is revealed.
Monday, 17 June 2024
Episode 57
A woman reprimands another to stay in her lane, leading to a damning revelation. A man decides to leave his broken life behind.
Tuesday, 18 June 2024
Episode 58
Nhlamulo learns that due to his actions, his worst fears have come true. Caiphus' extravagant promises to suppliers get him into trouble with Ndu.
Wednesday, 19 June 2024
Episode 59
A woman manages to kill two birds with one stone. Business doom looms over Chillax. A sobering incident makes a woman realize how far she's fallen.
Thursday, 20 June 2024
Episode 60
Taps' offer to help Nhlamulo comes with hidden motives. Caiphus' new ideas for Chillax have MaHilda excited. A blast from the past recognizes Dintle, not for her newfound fame.
Friday, 21 June 2024
Episode 61
Taps is hungry for revenge and finds a rope to hang Mdala with. Chillax's new job opportunity elicits mixed reactions from the public. Roommates' initial meeting starts off on the wrong foot.
Monday, 24 June 2024
Episode 62
Taps seizes new information and recruits someone on the inside. Mary's new job
Experience the thrill of Progressive Puzzle Adventures, like Scavenger Hunt Games and Escape Room Activities combined Solve Treasure Hunt Puzzles online.
Case-based Sequential Ordering of Songs for Playlist Recommendation
1. Case-based Sequential Ordering of Songs
for Playlist Recommendation
Claudio Baccigalupo and Enric Plaza
IIIA - Artificial Intelligence Research Institute
CSIC - Spanish Council for Scientific Research
Campus UAB, 08193 Bellaterra, Catalonia (Spain)
Vox: +34-93-5809570, Fax: +34-93-5809661
Email: {claudio,enric}@iiia.csic.es
Abstract. We present a CBR approach to musical playlist recommen-
dation. A good playlist is not merely a bunch of songs, but a selected
collection of songs, arranged in a meaningful sequence, e.g. a good DJ
creates good playlists. Our CBR approach focuses on recommending new
and meaningful playlists, i.e. selecting a collection of songs that are ar-
ranged in a meaningful sequence. In the proposed approach, the Case
Base is formed by a large collection of playlists, previously compiled by
human listeners. The CBR system first retrieves from the Case Base the
most relevant playlists, then combines them to generate a new playlist,
both relevant to the input song and meaningfully ordered. Some experi-
ments with different trade-offs between the diversity and the popularity
of songs in playlists are analysed and discussed.
1 Introduction
A typical music recommender considers songs the user likes, has listened to,
or has bought, and proposes similar songs that the user will probably like, be
interested in listening, or buying. Only a few recommenders are addressed to
suggest playlists of similar songs with an inherent sequential structure.
Creating a playlist is a common manual operation in audio user experience.
Imagine a user wants to assemble a playlist, on the basis of a specific song he
likes, has listened to or has bought. He could either choose each song manually,
or follow the advice of a recommender, adding one proposed song after the other.
In both cases, the process would be slow; besides, the user would also have to
order the songs, for a playlist is expected to contain songs in a specific sequence.
In this paper, we describe a case-based approach to playlist recommendation.
Given a song chosen by the user, we don’t focus on finding isolated similar
songs; instead we put the emphasis on recommending good playlists. A playlist
is a sequence—not a set—of songs; thus its quality is given both by the songs it
contains, and by their ordering (their relative positions).
This research is supported in part by a MusicStrands scholarship and by CBR-
ProMusic under the project TIC2003-07776-C02-02.
2. This paper presents a CBR system that takes a song as the input and re-
turns a recommended playlist as the output. The possible applications include:
personalised radio programs, playlist generators, digital music organisers, and in
general any scenario where the user desires a good sequence of songs, related to
a chosen one, without the burden of compiling it manually.
In this paper, after reviewing related works (Sect. 2), we present the problem
of recommending a playlist from a song (Sect. 3). Then we describe the CBR
process: in Sect. 4, we illustrate the structure of the Case Base; in Sect. 5, we
explain the Retrieve process, and introduce the basic concepts of pattern and
relevance; in Sect. 6, we describe how the retrieved playlists are combined in the
Reuse process. In Sect. 7 we report about the tests with an actual Case Base of
playlists. Section 8 ends the paper with a brief conclusion and future works.
2 Related Work
Several approaches have been investigated to develop an Automatic Playlist Gen-
erator [2, 7, 17, 3, 20, 16]. Some of them require a large pre-existent music-related
knowledge, either in form of music metadata [18] or acoustic-based measures
[12]. In this paper we introduce a ‘knowledge-light’ approach to recommenda-
tion, based only on user-related knowledge. In this sense, our approach is com-
parable with SmartRadio and CoCoA Radio, two systems with which we share the
vision that the work involved in compiling a playlist of music can be distributed
to other listeners [10]. SmartRadio [8] is an Automatic Collaborative Filtering
recommender that generates personalised playlists on the basis of playlists of
users whose profiles are similar. Similarly, CoCoA Radio [4] recommends new
playlists, but it also assumes that the knowledge of a playlist is encoded in the
order of its songs. Our system is, in a sense, lighter than SmartRadio and CoCoA
Radio. We also do assume that, in any well chosen collection of music, there is
some great implicit value in the order of its songs, but we use this and only
this knowledge to generate new playlists, without, for instance, collecting users’
profiles or managing explicit ratings.
Some common points can be found between our system and the approach
presented by Ragno et Al. [21]. In their work, similarity is inferred analysing
the occurrences of songs in broadcasted streams of music, and new playlists are
generated using a graph-search process (where the songs are the nodes of the
graph and the similarities between them represent the weights of the arcs).
In our approach, the recommender is implemented as a Case-Based Reasoning
(CBR) process [1, 14]. Every playlist is seen as a case whose relevance is inferred
measuring the co-occurrences of its songs in a large collection of past playlists.
Co-occurrences analysis [11] has been proved effective in many domains where
information is available as large sets of sequential data. Applications range from
semantical similarity recognition [13] to weather forecast [22]. Pachet et al. [15]
have originally proposed co-occurrences analysis as a method to evaluate sim-
ilarity between songs. In this paper, we combine this method with an explicit
notion of sequentiality, in order to infer similarity between songs and playlists.
3. 3 Playlist Recommendation
In our approach, the user chooses an input song s and a desired length λ; the
goal of the recommender is to return a playlist p (with λ songs) such that:
(G1) p contains s; (G2) p is varied; (G3) p is coherently ordered. A playlist is
varied if it does not repeat the same track or artist in the sequence, or at least
if repetitions are not close. We say a playlist is coherently ordered if its songs
make sense in the proposed sequence. The problem we have to solve is then to
determine which sequences of songs are coherently ordered.
In order to answer this issue, we analyse a large repository of playlists, previ-
ously compiled by human listeners. In this analysis, we seek for those sequences
(of two or more songs) that occur more often, with the same order, in these
playlists. The idea is that the higher the number of playlists where a sequence
occurs, the stronger the evidence that sequence is coherently ordered. This idea
takes inspiration from studies of co-occurrences analysis. Basically, these studies
argue that when two items appear frequently in the same context, this proves
the existence of some similarity between them. For instance, linguists regularly
use co-occurrences analysis to extract clusters of semantically related words from
large collections of texts. We adapt this approach to a musical context: if some
songs occur frequently, in the same order, in a collection of past playlists, then
we assume that the order in which they appear is meaningful. We call such
sequences of songs relevant patterns.
Relevant patterns are the starting point for the recommender we present;
to fulfil the aforementioned goals we employ a Case-Based Reasoning approach
that comprises four subsequent steps:
1. Assemble a Case Base from a repository of playlists (Case Base Setup).
2. Prompt the user for an input song s and the desired length λ of the recom-
mendation (Problem Description).
3. Retrieve from the Case Base a subset of playlists that satisfy most of the
goals. Specifically, the retrieved playlists are varied (G2) and include rele-
vant patterns that contain s (G1) (Retrieve Process).
4. Combine the songs of the retrieved playlists to generate a coherently ordered
playlist (G3) with length λ (Reuse Process).
Notice that the knowledge that guides the CBR process comes from the same
content of the Case Base, in the form of relevant patterns. Hence the quality of
the recommendation depends mostly on the quality of past playlists: the more
accurately they have been compiled by the users with a meaningful order, the
more this order will be reflected in the output. Clearly, different repositories may
lead to different recommendations.
4 The Case Base
Let P be the repository of past playlists. P may include, for example, track-
lists from radio programs, web streams, music compilations, DJ sessions, and
4. INPUT Retrieved
RETRIEVAL
(song) playlists
CASE BASE
(past playlists)
Relevant
Co-occurrence OUTPUT
REUSE
patterns
analysis (playlist)
Desired
length
Fig. 1. General view of the CBR process.
in general lists of songs that have been compiled with a meaningful order. By
the way, the repository could include some noisy playlists, that is, playlists that
have not been compiled with a great accuracy, and that would lower the quality
of the Case Base if used. Examples of noisy playlists are: very long sequences
of songs (that are probably only large sets of songs, with no meaningful order),
very short sequences (from which we cannot extract any relevant pattern), and
sequences in alphabetical order (that is hardly a meaningful order). Before using
the repository, we filter P to remove such playlists; the result C is the Case Base.
Every playlist of C is a sequence of songs: p = (s1 , s2 , .., sn ). Let Λ(p) = n be
the length of p. A song s is contained in a playlist p if the following condition
holds: In(s, p) = ∃i, 1 i n : s = si
Every song is characterised by a set A of attributes (e.g.: an identifier of
the track, an identifier of the album it belongs to, the artist that performs
it, etc.). We assume that the user will disapprove of recommendations where
close songs repeat the same attributes: for example playlists where many songs
belong to the same album or artist. Such playlists lack necessary variety, and
our approach intends to avoid these cases, in order to boost diversity. For every
attribute a ∈ A, we introduce a parameter γa , called safe distance for a, as the
minimum number of songs that must occur before we allow a value of a to be
repeated. For instance, if γa = 10, the same value of a can be repeated without
restriction only after 10 songs. To quantify how many, and how close, are the
potential repetitions of an attribute in a playlist, we introduce a function called
attribute variety:
Definition 1. (Attribute variety) Let p ∈ C be a playlist, p = (s1 , s2 , .., sn ),
and a ∈ A be a song attribute; the variety of a in p is a function Va : C → (0, 1]
5. such that:
j−i
γa if ∃j, i < j min(n, i + γa ) : a(si ) = a(sj )
n
Va (p) = ∧∀k, i < k < j, a(si ) = a(sk )
1 otherwise
i=1
In other words, a playlist has an attribute variety Va (p) = 1 only if no value
of a is repeated within the safe distance γa ; otherwise the more and the closer
the repeated values, the smaller Va (p). Combining the values of Va (p) for all
the song attributes with a t-norm (e.g. product), we obtain a global measure of
variety of a playlist p ∈ C:
V ar(p) = Va (p)
a∈A
The variety of a playlist p equals 1 only if no attribute is repeated in p
within its safe distance. According to (G2), we are interested in recommending
this kind of playlists. We will see in Sect. 5 how V ar(p) will be used, inside the
Retrieve process, to prefer varied playlists over repetitive ones.
5 Retrieve
In a classic CBR approach, the goal of the retrieval task is to return a subset
of cases that contain the solutions for problems similar to the proposed one. In
our approach, we do not intend to characterise the cases in terms of problem
and solution parts: our Retrieve process is not focused on finding which songs
in the Case Base are the most similar to the input song. Instead, we look for
useful cases [6], that is, previous users’ playlists that are useful to achieve the
requested goals. The idea is to retrieve a subset of those playlists that include
s (G1), are varied (G2), and contain relevant patterns, or else are coherently
ordered (G3).
Before presenting the Retrieve process, we introduce the Relevance function
and the Rating function. The Relevance function Rel(q, t), defined in Sect. 5.1,
takes as input a sequence of songs q, and a song t contained in q, and returns
the degree in which q is a relevant pattern for t. The Rating function ρ(p, s),
defined in Sect. 5.2, takes in input a playlist p ∈ C and the chosen song s, and
returns a combined measure of the variety of p and the relevant patterns in p
that contain s. At the end of the section, we present the Retrieve process that,
given all the playlists of C, returns only the k playlists that are more useful to
fulfil our goals.
5.1 Patterns and Relevance
Let q be a sequence of songs. We say q a pattern if q is sub-sequence of at least
one playlist of C, where q = (t1 , t2 , .., tm ) is sub-sequence of p = (s1 , s2 , .., sn )
when the following condition holds: Sub(q, p) = ∃i : ∀j, 1 j m, tj = si+j .
6. The pattern count φ(q) = |{p ∈ C : Sub(q, p)}| is the number of playlists of C
that q is sub-sequence of.
Let t be a song, Q(t) = {q : Λ(q) θ ∧ In(t, q) ∧ ∃p ∈ C : Sub(q, p)}
be the set of patterns, with at most θ songs, that contain t, and q ∈ Q(t)
be one of these patterns. To assess whether q is a relevant pattern for t, we
can consider how many times t occurs in the playlists of C together with the
other songs of q. According to co-occurrences analysis, in fact, the higher this
number, the higher the evidence that q is a relevant pattern for t. Thus, it
makes sense to use the pattern count φ(q) as a measure of the relevance of q
for t. Nevertheless, the actual value of φ(q) can be biased by two properties
of q: its length and the popularity of its songs. In general, short patterns or
patterns made of very popular songs have a higher probability of occurring in
many playlists, independently from their relevance for a specific song.
Example 1. To understand why shorter patterns are likely to have a higher pattern
count, consider the patterns (S2, S3) and (S2, S3, S4) in Table 1. Notice that (S2, S3)
is shorter than (S2, S3, S4), and φ( (S2, S3) ) > φ( (S2, S3, S4) ). This is not surprising,
for (S2, S3) is contained in (S2, S3, S4), so it occurs at least every time that (S2, S3, S4)
occurs in some playlist. To understand why patterns made of popular songs are likely
to have a higher pattern count, consider the patterns (S2, S3) and (S9, S2), both
containing S2. In the first case, S2 is followed by a “popular” song (S3 occurs 6 times
in Table 1); in the second case, S2 is preceded by a less popular song (S9 occurs only
3 times). Notice that φ( (S2, S3) ) > φ( (S9, S2) ); this is not surprising, for S3 has a
higher probability than S9 to co-occur with S2, due to its high popularity.
Table 1. A sample repository and the list of patterns with φ > 1.
Playlists Patterns
A = (S1, S2, S4) (S1, S2) (φ = 2)
B = (S1, S2, S3, S4, S5, S7, S1, S8) (S2, S3, S4) (φ = 3)
C = (S4, S6, S9, S2, S2, S3) (S2, S3) (φ = 6)
D = (S2, S3, S2, S3, S4) (S3, S4) (φ = 3)
(S4, S6) (φ = 2)
E = (S2, S3, S4, S6, S5)
(S9, S2) (φ = 2)
F = (S9, S2, S3, S5, S9, S1)
In a nutshell, short patterns and patterns made of popular songs are likely
to have a high pattern count, independently from being relevant to a particular
song. Therefore, to estimate the relevance of a pattern q to a song t, we combine
φ(q) with two factors that balance the effect of the length of q, and of the
popularity of its songs. We define the Relevance of a pattern q for a song t as
follows:
αθ−Λ(q)
Rel(q, t) = φ(q) · β
ψ (q, t)
where αθ−Λ(q) and ψ β (q, s) are the two factors intended to reduce the bias of the
length of q, and of the popularity of its songs, on the relevance value. Hereafter,
we explain in detail these two factors.
7. Length factor. Let q1 ∈ Q(t) and q2 ∈ Q(t) be two patterns, with same pattern
count and different length (e.g. Λ(q1 ) > Λ(q2 )). In this case, we consider q1
more relevant for t than q2 , because q1 occurs the same number of times than
the (more probable) shorter pattern q2 . Moreover, since q1 is longer, it contains
more songs co-occurrent with t than q2 . For this reason, we introduce a length
factor αθ−Λ(q) that we multiply φ(q) for, in order to substantially decrease the
relevance of shorter patterns. The shorter q (that is, the more distant is Λ(q) to
the maximum pattern length θ), the smaller this factor, the smaller the relevance
value. The parameter α can be tuned inside (0, 1], to control if the length of q
should affect more (small α) or less (high α) the relevance value.
Popularity factor. Let q1 = (t, u) and q2 = (t, v) be two patterns of two songs
with the same pattern count (e.g. φ(q1 ) = φ(q2 ) = 5), and let u be more popular
than v (e.g. u occurs in 600 playlists of C, while v only in 6). Notice that u
occurs relatively few times in C preceded by t (5 out of 600), while v occurs
relatively many times preceded by t (5 out of 6). In the whole, q2 appears in C
more frequently in relation to the popularity of its songs; in this sense it is more
relevant for t. For this reason, we introduce a popularity factor ψ β (q, t) that we
divide φ(q) by, in order to decrease the relevance of patterns made of popular
songs. This factor acts as a measure of inverse songs’ frequency of q; precisely,
ψ(q, t) is the number of times that the sequence q occurs in C “independently
from t”: ψ(q, t) = |{p ∈ C : ∃i : ∀j, 1 m, qj = pi+j ∨ qj = t}|. The
j
parameter β can be tuned inside [0, 1] to control if the popularity of the songs
should affect more (high β) or less (small β) the relevance value.
Example 2. To understand the effects of α and β on the relevance value, consider the
patterns in Table 1, ranked according to how relevant they are to a specific song, S3.
Table 2 shows the relevance values, using different values of α and β.
Notice that (S2, S3) has the highest pattern count, and indeed is the most relevant
pattern when α = 1 and β = 0.5. However, if we decrease the relevance of short patterns
(α = 0.5), then the most relevant pattern becomes (S2, S3, S4), that occurs less times
but is longer. Similarly, if we decrease the effect of popular songs (β = 1), then the
most relevant pattern is (S2, S3, S4), for it occurs 3 out of the 3 times that S2 and S4
occur in some playlist of Table 1 separated by one song.
Table 2. Relevance to S3 of the patterns from Table 1
Patterns Λ(qi ) φ(qi ) ψ(qi , S3) Rel(qi , S3) Rel(qi , S3) Rel(qi , S3)
qi ∈ Q(S3) (α = 1 β = 0.5) (α = 0.5 β = 0.5) (α = 1, β = 1)
(S2, S3, S4) 3 3 3 1.73 1.73 1
(S2, S3) 2 6 8 2.12 1.06 0.75
(S3, S4) 2 3 4 1.50 0.75 0.75
8. 5.2 Coherence and Retrieval
We are now able to define the Rating function ρ(p, s), that assigns a high value
to playlists that are varied and coherent with the input song s. This function is a
combination of V ar(p) (the variety of the playlist) and Coh(p, s) (the coherence
of the playlist to the input song), where Coh(p, s) is defined as the sum of the
relevance values of the patterns, present in p, that contain s:
Coh(p, s) = Rel(q, s)
q∈Ω(s,p)
where Ω(s, p) = {q : Λ(q) θ ∧ In(s, q) ∧ Sub(q, p)}. The more the relevant
patterns that contain s in p, and the higher their relevance to s, the higher
Coh(p, s). The Rating function ρ(p, s) is defined as a combination of variety and
coherence:
ρ(p, s) = V ar(p) · Coh(p, s), ∀p ∈ C
As a result, a playlist p ∈ C has a high value of ρ(p, s) if it contains s, is varied,
and is coherently ordered. Actually, this is the kind of playlist that we intend to
retrieve from C. Let k be the number of playlists to retrieve, the Retrieve process
returns the first k playlists of C, every playlist p ∈ C ranked according to ρ(p, s).
6 Reuse
The goal of the Reuse process is to transform the k retrieved playlists into a
single, recommended playlist of a specific length λ. Notice that the retrieved
playlists can have any length, so they cannot be directly returned as recommen-
dations. The approach we take for Reuse is that of Constructive Adaptation
[19]. Following this approach, we use the elements of the retrieved cases (i.e. the
songs from the retrieved playlists) to build a new combination of elements (i.e. a
new playlist), that is a solution for the current problem. Building this playlist is
seen as a search process, guided by the information contained in the Case Base
(in the form of relevant patterns).
Basically, Constructive Adaptation [19] is composed of two processes: Hy-
potheses Generation and Hypotheses Ordering. Hypotheses Generation deter-
mines how partial solutions are extended in the search process (i.e. which songs
from the retrieved playlists can be combined and how). Hypotheses Ordering
uses the knowledge from the Case Base to rank the nodes visited during the
search (i.e. the sequences of songs generated so far) in order for the search pro-
cess to follow the best paths to the solution. Thus, the Reuse process is a search
process where:
– the goal is to find a playlist of λ songs, that includes the input song s;
– every node in the search is a sequence made of songs from the retrieved
playlists;
– the singleton sequence (s) is taken as the initial node;
9. – the successors of every node contain the same sequence of songs of that node,
plus another song, either at the beginning or at the end of the sequence.
The fastest method to reach a sequence of length λ starting from (s) (a se-
quence of length 1) is to always explore the search tree “as deep as possible”,
that is, to always visit first the nodes with the longest sequences of songs found
so far. For this reason, the system uses a Depth-First Search process: at first,
it expands the initial node (s) by generating all its successors; at each subse-
quent step, it considers one of the most recently generated nodes, generates its
successors, and ranks them according to a heuristic; if at any moment it finds a
node with λ songs, returns it as the solution; if otherwise it finds a node with
no successors, it backtracks to expand another of the most recently generated
nodes.
6.1 Hypotheses Generation
The nodes of the search are sequences made of elements from the retrieved cases,
that is, of songs from the retrieved playlists. Let T = (t1 , t2 , .., tn ) be a node,
and let T be a successor of T . T contains the same songs of T , and another
song u from the retrieved playlists, either at the beginning or at the end of T .
We distinguish these two cases with the notations u + T and T + u . Let, for
instance, be T = u+T . Notice that in T , u occurs before t1 . In order for T to
be coherently ordered, u should occur before t1 at least in one retrieved playlist;
basically we consider T a valid successor of T if ∃p ∈ Ret(s) : Sub((u, t1 ), p),
where Ret(s) is the set of k retrieved playlists. Similarly, a sequence T + u is
a valid successor of T if ∃p ∈ Ret(s) : Sub((tn , u), p). The set of all the valid
successors of a node T is defined as Succ(T ) = { u + T , ∀u ∈ P re(T ) ∨ T +
u , ∀u ∈ P ost(T )}, where P re(T ) = {u : ∃p ∈ Ret(s) : Sub((u, t1 ), p)} and
P ost(T ) = {u : ∃p ∈ Ret(s) : Sub((tn , u), p)}.
In the next subsection, we will describe how such sets of successors are sorted
in the search process in order to finally reach a sequence of λ songs that is varied,
coherently ordered, and relevant to s.
6.2 Hypotheses Ordering
In ordering the successors of a node, we promote those sequences that are varied
and include relevant patterns. In other words, given any successor T of T (where
u is the added song), we determine if T is a good successor of T using the
heuristic function H(T ) = Rel(T , u) · V ar(T ). Using H, we are able to rank
highly a node T that contains a varied and coherently ordered sequence of songs.
However we cannot determine whether the successors of T will be varied and
coherently ordered sequences as well. To solve this issue, we introduce in the
heuristic a parameter L, called look-ahead factor, such that:
– If L = 0, we evaluate H on T , as above;
– If L = 1, we evaluate H on all the successors of T ; then we return the
maximum value found as the heuristic value for T ;
10. – If L = 2, we evaluate H on all the successors at depth 2 of T , that is, the
successors of the successors of T ; then we return the maximum value found
as the heuristic value for T ; and so on.
The higher is L, the more reliable is the heuristic; on the other hand, the
number of requested evaluations of H grows steeply with L. The value of L is
usually a compromise between a good heuristic and a reasonable performance.
Finally, we define the heuristic function as follows:
Definition 2. (Look-ahead heuristic) Let T be a node of the search tree; any
successor T of T is ranked using the heuristic function:
H(T ) if L = 0
H (T , L) =
H (T , L − 1) if L > 0
maxT ∈Succ(T )
Summing up, the Reuse process follows a Constructive Adaptation approach
in which a) Hypotheses Generation explores the possible combinations of the
songs of the retrieved playlists and b) Hypothesis Ordering ranks the partial
sequences according to a heuristic based on relevance and variety. At the end,
the result is a playlist of length λ that combines the best sub-sequences in the
retrieved playlists into a varied and coherent recommendation.
Example 3. Let S3 be the input song, and λ = 5 the desired length. Let the repository
of Table 1 be the Case Base, and Ret(S3) = {B, C, E, F } be the subset of k = 4
playlists retrieved. Consider how, in the Reuse process, these playlists are combined
using a search process (for brevity, the computation of H is omitted):
1) Generate the successors of (S3) and sort them using H : (S3, S4), (S2, S3),
(S3, S5). Next node to visit: (S3, S4).
2) Generate the successors of (S3, S4) and sort them using H : (S3, S4, S6),
(S2, S3, S4), (S3, S4, S5). Next node to visit: (S3, S4, S6).
3) Generate the successors of (S3, S4, S6) and sort them using H :
(S2, S3, S4, S6), (S3, S4, S6, S9), (S3, S4, S6, S5). Next node to visit: (S2, S3, S4, S6).
4) Generate the successors of (S2, S3, S4, S6) and sort them using H :
(S9, S2, S3, S4, S6), (S1, S2, S3, S4, S6), (S2, S3, S4, S6, S9), (S2, S3, S4, S6, S5).
Next node to visit: (S9, S2, S3, S4, S6).
5) The node (S9, S2, S3, S4, S6) has λ = 5 songs, hence is returned as the solution.
Notice that the solution is a varied and coherent combination of sub-sequences from
the retrieved playlists: (S2, S3, S4) from B, (S4, S6) and (S2, S3) from C,
(S2, S3, S4, S6) from E and (S9, S2, S3) from F .
7 Experimental Results
We have tested the described CBR recommending system using a collection of
playlists from MusicStrands1 . In the Case Base, we have represented playlists
1
For more information visit: http://www.musicstrands.com.
11. as pairs (track, artist), being track and artist two integers that allow us to
univocally identify every track and every artist in the MusicStrands database.
We have pre-processed the collection of playlists, discarding those with less than
5 songs, with more than 50 songs, and those alphabetically ordered (backwards
and forwards) along the names of the track and/or of the artist. Eventually, we
have assembled the Case Base using a subset (about 300,000) of playlists from
MusicStrands.
Concerning parameters, we were interested in recommending playlists where
tracks were mostly never repeated (γtrack = 200) and artists were hardly repeated
(γartist = 50). We have chosen θ = 5 as the maximum pattern length (longer
patterns very rarely appeared more than once), k = 50 as the number of retrieved
playlists, α = 0.5 and β = 1 as the length factor and popularity factor for the
relevance, and L = 1 as the look-ahead factor.
Table 3 shows six sample recommendations, where the desired length is λ =
10, and the input songs have been chosen from different genres and periods.
Notice that the suggested songs are similar (both acoustically and in terms of
metadata) to the input songs, especially those that appear closely before or after
them in the proposed sequences2 . On the whole, the songs in the playlists sound
well in the presented order, and their genres are never too distant from the one
of the input song.
It is also worth noticing that many of the proposed songs are not very popular,
yet they form a coherently ordered playlist with the input songs. This is due
to effect of the popularity factor in the relevance function (β = 1), by which
we dampen playlists made of popular songs. In this way, we enable the user
to discover new music, because we are suggesting songs the user has probably
never listened to, and yet will appreciate, since they are correlated with the
input song. Should users long for recommendations with more popular songs,
they would just have to lower this value. For instance, in Table 4 we present two
recommendations, obtained with the same input songs of Test 2 and Test 4, but
generated with β = 0, that is, without dampening popular songs. Consider, for
instance, the song American Idiot, recommended in Test 2b (β = 0) before the
input song American Pie. These two songs co-occur in sequence 7 times in the
Case Base, which is, in general, a relevant pattern count. However, this pattern
count becomes irrelevant when compared to the popularity of American Idiot,
which is a very popular song (it appears in 8,479 playlists) and only 7 out of
8,479 times it is followed by American Pie. As a result, the system does not
recommend this sequence in Test 2 (β = 1), where pattern counts are dampened
for sequences made of popular songs.
As a matter of fact, by proposing different recommendations from the same
input, the parameter β allows to improve the diversity of the solutions. This
is indeed an important feature for a recommending system, and a key issue for
CBR systems [5].
2
A 30” excerpt of every song, along with detailed metadata, is available on Music-
Strands web-site.
12. Table 3. Sample of recommended playlists
Test 1. Input song: Test 2. Input song:
Only Shallow (My Bloody Valentine) American Pie (Don McLean)
Kool Thing Sonic Youth We’re An American Band VV.AA.
Stupid Preoccupations Vic Chesnutt Sweet Home Alabama Lynyrd Skynyrd
Seed Toss Superchunk More Than a Feeling Boston
Flyin’ The Flannel Firehose Bad Moon Rising
Only Shallow My Bloody Valentine Creedence Clearwater Revival
Exhibit A The Features American Pie Don McLean
Lover’s Spit Broken Social Scene Mr. Blue Sky Electric Light Orchestra
Gigantic The Pixies Switch Will Smith
Kamera Wilco This Love Maroon 5
From Blown Speakers Walkie Talkie Man Steriogram
The New Pornographers Walkin’ On The Sun Smash Mouth
Test 3. Input song: Test 4. Input song:
Mrs. Robinson (Simon & Garfunkel ) Soldier (Destiny’s Child )
Unchained Melody Let Me Love You Mario
The Righteous Brothers Hush LL Cool J
Kicks Paul Revere & The Raiders Red Carpet (Pause, Flash)
Cherry Hill Park Billy Joe Royal R. Kelly
Windy The Association Hot 2 Nite New Edition
Sunshine Superman Donovan Wonderful Ja Rule
Mrs. Robinson Simon & Garfunkel My Prerogative Britney Spears
Nights In White Satin The Moody Blues Two Step Ciara
Only One Yellowcard Soldier Destiny’s Child
Pain Jimmy Eat World Only U Ashanti
Quiet Ones Tears For Fears Pass Out Ludacris
Test 5. Input song: Test 6. Input song:
Roots Bloody Roots (Sepultura) Strangers In The Night (F. Sinatra)
Satan Spawn, The Caco-Daemon It Had To Be You Steve Tyrell
Deicide Jamaica Farewell Desmond Dekker
Rapture Morbid Angel Just The Way You Are Diana Krall
Serenity In Fire Kataklysm Let’s Fall In Love Diana Krall
Roots Bloody Roots Sepultura Nunca Es Para Siempre
Plague Rages Napalm Death Presuntos Implicados
Blend As Well Coalesce Strangers In The Night F. Sinatra
God Send Death Slayer Candy Man, The Sammy Davis Jr.
Token Damad Unforgettable Nat King Cole
Heaven In Her Arms Converge What A Wonderful World
Fear Of Napalm Louis Armstrong
Terrorizer Falling In Love Again Billie Holiday
13. Table 4. Sample of recommended playlists, without dampening popular songs
Test 2b. Input song: Test 4b. Input song:
American Pie (Don McLean) Soldier (Destiny’s Child )
Behind These Hazel Eyes Kelly Clarkson Disco Inferno 50 Cent
Beverly Hills Weezer Mockingbird Eminem
I Just Wanna Live Good Charlotte Obsession Frankie J
American Idiot Green Day I Just Wanna Live Good Charlotte
American Pie Don McLean Boulevard Of Broken Dreams
Hotel California The Eagles Green Day
Cocaine Eric Clapton Since U Been Gone Kelly Clarkson
Emerald Eyes Fleetwood Mac Two Step Ciara
Carry On Wayward Son Kansas Soldier Destiny’s Child
Sweet Home Alabama Drop It Like It’s Hot Snoop Dogg
Lynyrd Skynyrd Get Back Ludacris
8 Conclusions and Future Work
We have presented a CBR system for recommending playlists made of mean-
ingful sequences of songs. The Case Base is a collection of playlists of different
quality, provided by different people. While most recommendation systems are
concerned with suggesting an unorganised set of songs, “similar” to those se-
lected by the user, our work focuses on recommending meaningful sequences of
songs (playlists). The proposed recommendations are not deeply “personalised”,
since the model we have of the user requesting a playlist is given only by the
selected (input) song. Future works will explore how to further constrain the
playlist to recommend by acquiring a larger input from the user.
This paper contributes to develop Case-based Reasoning in a variety of ways.
First, we address the important issue of reasoning about sequential cases (here,
in the form of coherently ordered playlists). Both the Case Base and the solution
provided by the CBR system are based on the sequential ordering of songs; the
problem-solving process of the CBR system exploits the informational content
of the Case Base to provide a solution with a coherent order.
A second issue of interest is that our Case Base contains only solutions and
not problems. In the usual CBR approach, the Case Base contains pairs (p, s)
(p is a problem, s is a solution); to find the solution for a new problem p ,
the system retrieves from the Case Base a set of similar problems, takes into
account their solutions, and returns one of them, or a combination of them,
as the solution for p . In our approach, instead, the Case Base contains only
solutions (playlists), and no explicit problems: we know that the playlists in the
Case Base are meaningful, but we ignore for which input songs they are good
recommendations. In this context, we have designed the simplest CBR approach
that can exploit the information contained in a large Case Base of meaningful
sequences. We have explicitly avoided complex user models, which are a common
feature of Collaborative Filtering systems. User models are not helpful in relation
14. to our goal: they seek relations between people and preferred songs or playlists,
but disregard any sequential information contained in the playlists.
A third issue of interest is the Reuse process: rather than recommending
an existing playlist, we combine the retrieved playlists using the same criteria
(variety and coherence) that we used in the Retrieve process. In this way we rec-
ommend a playlist that has the desired length λ and higher variety and coherence
values than those of the retrieved playlists.
Further, we have investigated the issue of applying co-occurrences analysis
in a musical context. In this approach, we have opted for a minimal Problem
Description (only one song and the length of the desired recommendation) in
order to generate meaningful playlists by exploiting the co-occurrences present
in the Case Base. We have avoided more constrained inputs (e.g. two songs,
such that the recommendation starts with the first one and ends with the other,
and is coherently ordered). The problem with over-constrained input is that the
system returns low quality solutions when the constraints are too strong (e.g.
a user requests a playlist to start with a rap song and to end with a requiem).
In the future, however, we plan to enlarge the input capabilities of the user,
to partially constrain the recommended playlist. For instance, the user could
provide a short sequence of songs, and the system could recommend a playlist
with most of them; or else the user could indicate a specific genre, or tag, to
consider or to avoid in the recommended playlist.
We have tested the system using a collection of manually-generated playlists,
but in the future we plan to work with other forms of sequential music data. For
instance, MusicStrands provides MyStrands, an iTunes plug-in that recommends
new songs to the user, on the basis of the last songs effectively played in iTunes.
We plan to improve this tool with the introduction of an explicit notion of
sequentiality, in order to recommend on-line, meaningful, personalised playlists,
on the basis of previous sequences of songs effectively played by the user.
The evaluation of the quality of the recommendation is an open issue for
recommender systems. We agree with Hayes et al. [9] that user satisfaction can
only be measured in an on-line context. Although the proposed system has not
yet been evaluated by a significant number of users, we have deployed it on the
Internet3 . Users can select an input song, the desired length, and other param-
eters, and can compare the (meaningfully ordered) recommendation generated
with the presented approach with the playlist proposed by the MusicStrands rec-
ommender. We plan to employ the users’ feedback as an evaluation for our work
and, once recommendations have been positively revised by the users, to retain
in the MusicStrands repository those that have been approved as good playlists,
to use them as new cases for future recommendations.
References
1. A. Aamodt and E. Plaza. Case-based reasoning: Foundational issues, methodological
variations, and system approaches. Artificial Intelligence Communications, 7(1):39–
3
http://labs.musicstrands.com
15. 59, 1994.
2. M. Alghoniemy and A.H. Tewfik. User-defined music sequence retrieval. In Proc.
ACM Multimedia, 356–358, 2000.
3. J.-J. Aucouturier and F. Pachet. Scaling up Music Playlist Generation. In Proc. of
the 3rd IEEE Intl. Conf. on Multimedia and Expo, 2002.
4. P. Avesani, P. Massa, M. Nori, and A. Susi. Collaborative radio community. In
Proc. of Adaptive Hypermedia, 2002.
5. K. Bradley and B. Smyth. Improving recommendation diversity. In Proc. of the
12th Irish Conference on Artificial Intelligence and Cognitive Science, 2001.
6. H.-D. Burkhard. Extending some Concepts of CBR – Foundations of Case Retrieval
Nets. In Case-Based Reasoning Technology – From Foundations to Applications,
9:17–50, 1998.
7. D. B. Hauver and J. C. French. Flycasting: Using Collaborative Filtering to Gen-
erate a Playlist for Online Radio. In Proc. of the Intl. Conf. on Web Delivering of
Music, 2001.
8. C. Hayes and P. Cunningham. Smart radio: Building music radio on the fly. In
Expert Systems, 2000.
9. C. Hayes, P. Massa, P. Avesani and P. Cunningham. An online evaluation framework
for recommender systems. In Proc. of the RPEC Conference, 2002.
10. C. Hayes and P. Cunningham. Context-boosting collaborative recommendations.
Knowledge-Based Systems, 17:131-138, 2004.
11. T. Hofmann and J. Puzicha. Statistical models for co-occurrence data. Memoran-
dum, MIT Artificial Intelligence Laboratory, 1998.
12. B. Logan. Music recommendation from song sets. In Proc. of the 5th ISMIR
Conference, 2004.
13. C. Manning and H. Sch¨tze. Foundations of Natural Language Processing. 1999.
u
14. R. L´pez de M´ntaras, D. McSherry, D. Bridge, D. Leake, B. Smyth, S. Craw,
o a
B. Faltings, M-L. Maher, M.T. Cox, K. Forbus, M. Keane, A. Aamodt, I. Wat-
son. Retrieval, reuse, revision, and retention in case-based reasoning. Knowledge
Engineering Review. In press, 2006.
15. F. Pachet, G. Westerman, and D. Laigre. Musical data mining for electronic music
distribution. In Proc. of the Intl. Conf. on Web Delivering of Music, 2001.
16. E. Pampalk, T. Pohle, and G. Widmer. Dynamic Playlist Generation Based on
Skipping Behaviour. Proc. of the 6th ISMIR Conference, 2005.
17. S. Pauws and B. Eggen. PATS: Realization and User Evaluation of an Automatic
Playlist Generator. In Proc. of the Intl. Conf. on Music Information Retrieval, 2002.
18. J. Platt, C. Burges, S. Swenson, C. Weare, and A. Zheng. Learning a gaussian
process prior for automatically generating music playlists. In Advances in Neural
Information Processing Systems, 14:1425-1432, 2002.
19. E. Plaza and J.-L. Arcos. Constructive adaptation. Lecture Notes in Artificial
Intelligence, 2416:306–320, 2002.
20. T. Pohle, E. Pampalk, and G. Widmer. Generating similarity-based playlists using
traveling salesman algorithms. In Proc. of the Intl. Conf. on Digital Audio Effects,
2005.
21. R. Ragno, C. J. C. Burges, and C. Herley. Inferring Similarity Between Music
Objects with Application to Playlist Generation. ACM Multimedia Information
Retrieval, 2005.
22. K. Wang. Discovering Patterns from Large and Dynamic Sequential Data. Journal
of Intelligent Information System, 8–33, 1995.