Machine Learning for Recommender Systems MLSS 2015 Sydney

1. MLSS Sydney, February 18, 2015 – Machine Learning for Recommender Systems MLSS 2015 Sydney Machine Learning for Recommender Systems Alexandros Karatzoglou Senior Research Scientist @ Telefonica Research, Barcelona alexk@tid.es @alexk_z

2. MLSS Sydney, February 18, 2015 – Machine Learning for Recommender Systems Telefonica Research Machine Learning Recommender Systems HCI Mobile Computing Systems Networks http://www.tid.es

3. MLSS Sydney, February 18, 2015 – Machine Learning for Recommender Systems Tuenti

4. MLSS Sydney, February 18, 2015 – Machine Learning for Recommender Systems Mozilla FirefoxOS Marketplace

5. MLSS Sydney, February 18, 2015 – Machine Learning for Recommender Systems Recent Publications SIGIR 2014: Gaussian Process Factorization Machines for Context-aware Recommendation RecSys 2014: Coverage, Redundancy and Size-Awareness in Genre Diversity for Recommender Systems RecSys 2014: Question Recommendation for Collaborative Question Answering Systems with RankSLDA CIKM 2013: GAPfm: Optimal Top-N Recommendations for Graded Relevance Domains RecSys 2013: xCLiMF: Optimizing Expected Reciprocal Rank for Data with Multiple Levels of Relevance ECML/PKDD 2013: Socially Enabled Preference Learning from Implicit Feedback Data CIKM 2012: Climbing the App Wall: Enabling Mobile App Discovery through Context-Aware Recommendations RecSys 2012: CLiMF: Learning to Maximize Reciprocal Rank with Collaborative Less-is-More Filtering * Best Paper Award SIGIR 2012: TFMAP: Optimizing MAP for Top-N Context-aware Recommendation RecSys 2011: Collaborative Temporal Order Modeling RecSys 2011: Implicit Feedback Recommendation via Implicit-to-Explicit Ordinal Logistic Regression Mapping RecSys 2010: Multiverse Recommendation: N-dimensional Tensor Factorization for Context-Aware Collaborative Filtering EC-Web 2010: Quantile Matrix Factorization for Collaborative Filtering AISTATS 2010: Collaborative Filtering on a Budget RecSys 2009: Maximum Margin Code Recommendation RecSys 2008: Adaptive Collaborative Filtering Machine Learning Journal, 2008: Improving Maximum Margin Matrix Factorization * Best Machine Learning Paper Award at ECML PKDD 2008 NIPS 2007: CoFiRank - Maximum Margin Matrix Factorization for Collaborative Ranking

6. MLSS Sydney, February 18, 2015 – Machine Learning for Recommender Systems Index 1.Introduction: What is a Recommender System? 2. Approaches 1. Collaborative Filtering 2. Content-based Recommendations 3. Context-aware Recommendations 4. Ranking 5. Diversity 3. Research Directions 4. Conclusions 5. References

7. MLSS Sydney, February 18, 2015 – Machine Learning for Recommender Systems From Search to Recommendation “The Web is leaving the era of search and entering one of discovery. What's the difference? SearchSearch is what you do when you're looking for something. DiscoveryDiscovery is when something wonderful that you didn't know existed, or didn't know how to ask for, finds you.” – CNN Money, “The race to create a 'smart' Google

8. MLSS Sydney, February 18, 2015 – Machine Learning for Recommender Systems The value of recommendations Netflix: 2/3 of the movies watched are recommended Google News: recommendations generate 38% more click-throughs Amazon: 35% sales from recommendations Choicestream: 28% of the people would buy more music if they found what they liked.

9. MLSS Sydney, February 18, 2015 – Machine Learning for Recommender Systems The “Recommender problem” Estimate a utility functionutility function to predictpredict how a user will likelike an item.

10. MLSS Sydney, February 18, 2015 – Machine Learning for Recommender Systems The “Recommender problem” C:= {users} S:= {recommendable items} u:= utility function, measures the usefulness of item s to user c, u : C X S→ R where R:= {recommended items}. For each user c, we want to choose the items s that maximize u. ∀c∈C , s'c=argmax (u(c , s)) s∈S

11. MLSS Sydney, February 18, 2015 – Machine Learning for Recommender Systems A good recommendation is relevant to the user: personalized

12. MLSS Sydney, February 18, 2015 – Machine Learning for Recommender Systems A good recommendation is diverse: it represents all the possible interests of one user

13. MLSS Sydney, February 18, 2015 – Machine Learning for Recommender Systems A good recommendation Does not recommend items the user already knows or would have found anyway. Expands the user's taste into neighboring areas. SerendipitySerendipity = Unsought finding= Unsought finding

14. MLSS Sydney, February 18, 2015 – Machine Learning for Recommender Systems Top k recommendations Usually users take into account only few suggestions. There is a need to do better on the top scoring recommended items

15. MLSS Sydney, February 18, 2015 – Machine Learning for Recommender Systems Index 1. Introduction: What is a Recommender System? 2. Approaches 1. Collaborative Filtering 2. Content-based Recommendations 3. Context-aware Recommendations 4. Ranking 5. Diversity 3. Research Directions 4. Conclusions 5. References

16. Alexandros Karatzoglou – September 06, 2013 – Recommender Systems Approaches to Recommendation Recommendations are based on: Users' past behaviour (Collaborative Filtering – CF): Find similar users to me and recommend what they liked (Memory-based) Model taste patterns in large amounts of user-item interactions (Model-based). Item features (Content-based) Users' features (Demographic) Social recommendations (Trust-based)

17. MLSS Sydney, February 18, 2015 – Machine Learning for Recommender Systems What works? Depends on the domain and particular problem Currently, the best approach is Collaborative Filtering. Other approaches can be combined to improve results What matters? Data preprocessing: outlier removal, denoising, removal of global effects Combining methods

18. MLSS Sydney, February 18, 2015 – Machine Learning for Recommender Systems 2.1. Collaborative Filtering

19. MLSS Sydney, February 18, 2015 – Machine Learning for Recommender Systems Index 1. Introduction: What is a Recommender System? 2. Approaches 1. Collaborative Filtering 2. Content-based Recommendations 3. Context-aware Recommendations 4. Other Approaches 5. Hybrid Recommender Systems 3. Research Directions 4. Conclusions 5. References

20. MLSS Sydney, February 18, 2015 – Machine Learning for Recommender Systems Collaborative Filtering The task of predictingpredicting (filtering) user preferences on new items by collectingcollecting taste information from many users (collaborative). Challenges: many items to choose from very few recommendations to propose few data per user no data for new user very large datasets

21. MLSS Sydney, February 18, 2015 – Machine Learning for Recommender Systems 2 5 4 5 4 4 1 5 5 4 1 2 5 4 1 2 Example: User-based CF Target (or Active) user for whom the CF recommendation task is performed

22. MLSS Sydney, February 18, 2015 – Machine Learning for Recommender Systems 2 5 4 5 4 4 1 5 5 4 1 2 5 2 4 1 Example Each user has expressed an opinion for some items: Explicit opinion: rating score Implicit: purchase records or listen to tracks

23. MLSS Sydney, February 18, 2015 – Machine Learning for Recommender Systems 2 5 4 5 4 4 1 5 5 4 1 2 5 4 1 2 Example: User-based CF 1. Identify set of items rated by the target user

24. MLSS Sydney, February 18, 2015 – Machine Learning for Recommender Systems 2 5 4 5 4 4 1 5 5 4 1 2 5 4 1 2 Example: User-based CF 1. Identify set of items rated by the target user 2. Identify which other users rated 1+ items in this set (neighbourhood formation)

25. MLSS Sydney, February 18, 2015 – Machine Learning for Recommender Systems 3. Compute how similar each neighbour is to the target user (similarity function) 4. In case, select k most similar neighbours User-based Similarity

26. MLSS Sydney, February 18, 2015 – Machine Learning for Recommender Systems 5. Predict ratings for the target user's unrated items (prediction function) 6. Recommend to the target user the top N products based on the predicted ratings User-based CF

27. MLSS Sydney, February 18, 2015 – Machine Learning for Recommender Systems User-based CF Target user u, ratingsratings matrix Y yv,i → rating by user v for item i Similarity Pearson r correlation sim(u,v) between users u & v Predicted rating

28. MLSS Sydney, February 18, 2015 – Machine Learning for Recommender Systems 2 5 4 5 4 4 1 5 5 4 1 2 5 4 1 2 sim(u,v) NA NA Example: User-based CF

29. MLSS Sydney, February 18, 2015 – Machine Learning for Recommender Systems 2 5 4 5 4 4 1 5 5 4 1 2 5 4 1 2 Example: User-based CF 0.87 sim(u,v) NA NA

30. MLSS Sydney, February 18, 2015 – Machine Learning for Recommender Systems 2 5 4 5 4 4 1 5 5 4 1 2 5 4 1 2 1 sim(u,v) 0.87 NA NA Example: User-based CF

31. MLSS Sydney, February 18, 2015 – Machine Learning for Recommender Systems 2 5 4 5 4 4 1 5 5 4 1 2 5 4 1 2 -1 0.87 1 sim(u,v) NA NA Example: User-based CF

32. MLSS Sydney, February 18, 2015 – Machine Learning for Recommender Systems 2 5 3.51* 4 5 4 4 1 3.81* 5 5 4 2.42* 1 2 5 2.48* 4 1 2 -1 0.87 1 sim(u,v) NA NA Example: User-based CF

33. MLSS Sydney, February 18, 2015 – Machine Learning for Recommender Systems 2 5 4 5 4 4 1 5 5 4 1 2 5 4 1 2 Example: Item-based CF Target item: item for which the CF prediction task is performed.

34. MLSS Sydney, February 18, 2015 – Machine Learning for Recommender Systems Item-based CF The basic steps: Identify set of users who rated the target item i Identify which other items (neighbours) were rated by the users set Compute similarity between each neighbour & target item (similarity function) In case, select k most similar neighbours Predict ratings for the target item (prediction function)

35. MLSS Sydney, February 18, 2015 – Machine Learning for Recommender Systems Item Based Similarity

36. MLSS Sydney, February 18, 2015 – Machine Learning for Recommender Systems Item Based Similarity Target item I yu,j → rating of user u for item j, average rating for j. Similarity sim(i,j) between items i and j (Pearson- correlation) Predicted rating

37. MLSS Sydney, February 18, 2015 – Machine Learning for Recommender Systems 2 5 4 5 4 4 1 5 5 4 1 2 5 4 1 2 Example: Item-based CF -1sim(i,j)

38. MLSS Sydney, February 18, 2015 – Machine Learning for Recommender Systems 2 5 4 5 4 4 1 5 5 4 1 2 5 4 1 2 Example: Item-based CF -1-1sim(i,j)

39. MLSS Sydney, February 18, 2015 – Machine Learning for Recommender Systems 2 5 4 5 4 4 1 5 5 4 1 2 5 4 1 2 Example: Item-based CF 0.86-1-1sim(i,j)

40. MLSS Sydney, February 18, 2015 – Machine Learning for Recommender Systems 2 5 4 5 4 4 1 5 5 4 1 2 5 4 1 2 Example: Item-based CF 1-1-1 0.86sim(i,j)

41. MLSS Sydney, February 18, 2015 – Machine Learning for Recommender Systems 2 5 4 5 4 4 1 5 5 4 1 2 5 4 1 2 Example: Item-based CF sim(6,5) cannot be calculated 1-1-1 0.86sim(i,j) NA

42. MLSS Sydney, February 18, 2015 – Machine Learning for Recommender Systems 2 5 4 5 4 4 1 5 5 4 1 2 5 4 2.48* 2.94* 1 2 1.12* Example: Item-based CF 1-1-1 0.86sim(i,j) NA

43. MLSS Sydney, February 18, 2015 – Machine Learning for Recommender Systems PearsonPearson r correlation-based Similarityr correlation-based Similarity does not account for user rating biases Cosine-basedCosine-based SimilaritySimilarity does not account for user rating biases AdjustedAdjusted Cosine SimilarityCosine Similarity takes care of user rating biases as each pair in the co-rated set corresponds to a different user. Item Similarity Computation

44. MLSS Sydney, February 18, 2015 – Machine Learning for Recommender Systems Performance Implications Bottleneck: Similarity computation. Time complexity, highly time consuming with millions of users & items in the database. Two-step process: “off-line component” / “model”: similarity computation, precomputed & stored. “on-line component”: prediction process.

45. MLSS Sydney, February 18, 2015 – Machine Learning for Recommender Systems Two-step process OfflineOffline OnlineOnline Learning Process Model Or Clusters Decision Process Context Recommended Items

46. MLSS Sydney, February 18, 2015 – Machine Learning for Recommender Systems Performance Implications User-based similarity is more dynamic. Precomputing user neighbourhood can lead to poor predictions. Item-based similarity is static. We can precompute item neighbourhood. Online computation of the predicted ratings.

47. MLSS Sydney, February 18, 2015 – Machine Learning for Recommender Systems Memory based CF + Requires minimal knowledge engineering efforts + Users and products are symbols without any internal structure or characteristics + Produces good-enough results in most cases - Requires a large number of explicit and reliable “ratings” - Requires standardized products: users should have bought exactly the same product - Assumes that prior behaviour determines current behaviour without taking into account “contextual” knowledge

48. MLSS Sydney, February 18, 2015 – Machine Learning for Recommender Systems The Sparsity Problem Typically large product sets & few user ratings e.g. Amazon: in a catalogue of 1 million books, the probability that two users who bought 100 books each, have a book in common is 0.01 in a catalogue of 10 million books, the probability that two users who bought 50 books each, have a book in common is 0.0002 CF must have a number of users ~ 10% of the product catalogue size

49. MLSS Sydney, February 18, 2015 – Machine Learning for Recommender Systems The Sparsity Problem Methods for dimensionality reduction Matrix Factorization Clustering Projection (Principal Component Analysis ...)

50. MLSS Sydney, February 18, 2015 – Machine Learning for Recommender Systems Personalised vs Non-Personalised CF CF recommendations are personalized: the prediction is based on the ratings expressed by similar users; neighbours are different for each target user A non-personalized collaborative-based recommendation can be generated by averaging the recommendations of ALL users How would the two approaches compare?

51. MLSS Sydney, February 18, 2015 – Machine Learning for Recommender Systems Personalised vs Non-Personalised CF 0,1510,2230,0222811718164974424EachMovie 0,1790,2330,041100020939526040MovieLens 0,1520,2200,725351944910048483Jester MAE Pers MAE Non Pers density total ratings itemsusersData Set Not much difference indeed! vij is the rating of user i for product j and vj is the average rating for product j MAENP = ∑i, j ∣vij−v j∣ num.ratings Mean Average Error Non Personalized:

52. MLSS Sydney, February 18, 2015 – Machine Learning for Recommender Systems Model-Based Collaborative Filtering

53. MLSS Sydney, February 18, 2015 – Machine Learning for Recommender Systems Model Based CF Algorithms Models are learned from the underlying data rather than heuristics. Develop a model of user ratings: Bayesian network (probabilistic) Clustering (classification) Association rules Other models: linear regression, probabilistic latent semantic analysis ...

54. MLSS Sydney, February 18, 2015 – Machine Learning for Recommender Systems Matrix Factorization 2 5 4 5 4 4 1 5 5 4 1 2 5 2 4 1 d d

55. MLSS Sydney, February 18, 2015 – Machine Learning for Recommender Systems Learning: Stochastic Gradient Descent

56. MLSS Sydney, February 18, 2015 – Machine Learning for Recommender Systems Learning: Stochastic Gradient Descent For each rating of item j by user i:

57. MLSS Sydney, February 18, 2015 – Machine Learning for Recommender Systems Matrix Factorization 2 5 4 5 4 4 1 5 5 4 1 2 5 2 4 1 d d

58. MLSS Sydney, February 18, 2015 – Machine Learning for Recommender Systems Objective Function

59. MLSS Sydney, February 18, 2015 – Machine Learning for Recommender Systems Loss Functions for MF Squared error loss: Mean Average Error: Binary Hinge loss:

60. MLSS Sydney, February 18, 2015 – Machine Learning for Recommender Systems Restricted Boltzmann Machines A (generative stochastic) Neural Network Learns a probability distribution over its inputs Used in dimensionality reduction, CF, topic modeling, feature learning Essential components of Deep Learning methods (DBN's, DBM's)

61. MLSS Sydney, February 18, 2015 – Machine Learning for Recommender Systems Restricted Boltzmann Machines

62. MLSS Sydney, February 18, 2015 – Machine Learning for Recommender Systems Restricted Boltzmann Machines Each unit is in a state which can be active or not active. Each input of a unit is associated to a weight The transfer function Σ calculates for each unit a score based on the weighted sum of the inputs This score is passed to the activation function φ which calculated the probability that the unit state is active.

63. MLSS Sydney, February 18, 2015 – Machine Learning for Recommender Systems Learning to Rank - Usually we care to make accurate ranking and not rating prediction - Square loss optimizes to accurately predict 1s and 5s. - RS should get the top items right -> Ranking problem - Why not to learn how to rank directly - Learning to Rank methods provide up to 30% performance improvements in off-line evaluation

64. MLSS Sydney, February 18, 2015 – Machine Learning for Recommender Systems Restricted Boltzmann Machines Each unit in the visible layer vi corresponds to one item The number of the hidden units hj is a parameter. Each vi is connected to each hj through a weight wij In the trainingtraining phase, for each user: if the user purchased the item the corresponding vi is activated. The activation states of all vi are the input of each hj Based on this input the activation state of each hj is calculated The activation state of all hj become now the input of each vi The activation state of each vi is recalculated For each vi the difference between the present activation state and the previous is used to update the weights wij and thresholds θj

65. MLSS Sydney, February 18, 2015 – Machine Learning for Recommender Systems Restricted Boltzmann Machines

66. MLSS Sydney, February 18, 2015 – Machine Learning for Recommender Systems Restricted Boltzmann Machines In the predictionprediction phase, using a trained RBM, when recommending to a user: For the items of the user the corresponding vi is activated. The activation states of all vi are the input of each hj Based on this input the activation state of each hj is calculated The activation state of all hj become now the input of each vi The activation state of each vi is recalculated The activation probabilities are used to recommend items

67. MLSS Sydney, February 18, 2015 – Machine Learning for Recommender Systems Recurrent Neural Networks for CF RNN's are Neural Networks designed to model sequences RNN for CF tries to predict the next item given all previous ones e.g. predict the next song, video, given the history of streams. RNN's model the current output as a function of the previous output and a hidden (latent) state RNN models the items

68. MLSS Sydney, February 18, 2015 – Machine Learning for Recommender Systems Recurrent Neural Networks for CF h0 y0 h1 y1 h2 y2 ...

69. MLSS Sydney, February 18, 2015 – Machine Learning for Recommender Systems Recurrent Neural Networks for CF 2 vectors of dimension k: time steps: 0... t-1. “hidden” internal states: h0 ... ht-1 . Output vectors y At each time step: Predict the output yi given the hidden state hi (we need to model P(yi |hi )) Feed yi back into the next hidden state hi+1 : hi+1 = f(a(hi ) + b(yi )). where f is generally some nonlinear function (e.g. sigmoid or tanh), a, b are usually a simple linear transform.

70. MLSS Sydney, February 18, 2015 – Machine Learning for Recommender Systems RNN for CF: example Predict a sequence (vector yi ) of items: Assume P(yi |hi ) is a normal distribution the log-likelihood of the loss is the (negative) L2 loss: -(yt – ht )2 hi+1 = tanh(Uyi + Vhi ) & h0 = 0. yi = W hi Learn the k2 unknown parameters (U, V, and W) that maximize the log-likelihood over all:

71. MLSS Sydney, February 18, 2015 – Machine Learning for Recommender Systems RNN for CF: example Maximize the log-likelihood through back-propagation. we need to do two steps through each sequence. First propagation: For i=0... t-1: Calculate all h_i. the total log-likelihood is the sum of all log-probabilities of observing each output:

72. MLSS Sydney, February 18, 2015 – Machine Learning for Recommender Systems RNN for CF: example Define the partial derivatives of the log-L with respect to the hidden state: Since each hi only influences later hidden states, is just a function of all future j = i... t-1. We can rewrite as a function of : We can evaluate this backwards from t-1...0.

73. MLSS Sydney, February 18, 2015 – Machine Learning for Recommender Systems RNN for CF: example P(yi | hi ) and are specified by our model Derive the gradients for the unknown parameter U: Derive similar gradients for V, W. Optimize using SGD over lots of samples.

74. MLSS Sydney, February 18, 2015 – Machine Learning for Recommender Systems RNN for CF Given a sequence of watched movies, predict the next one. Now yi is not a scalar, but one out of many items. We need a probability distribution P(yi |hi ). The Softmax distribution over all possible outputs. This means we have to learn a vector aj for each item. P(yi |hi ) is now proportional to :

75. MLSS Sydney, February 18, 2015 – Machine Learning for Recommender Systems RNN for CF We also need something linking back the output to the next hidden state. we will learn another set of vectors bj – one for each item. we don’t need the matrices U and V. Our model now becomes: Again, we maximize the total log-likelihood Note that we have the unknowns aj ‘s and bj ‘s for each item. [Erik Bernhardsson

76. MLSS Sydney, February 18, 2015 – Machine Learning for Recommender Systems Clustering ClusterCluster customers into categories based on preferences & past purchases ComputeCompute recommendations at the cluster level: all customers within a cluster receive the same recommendations

77. MLSS Sydney, February 18, 2015 – Machine Learning for Recommender Systems Clustering B, C & D form 1 CLUSTER vs. A & E form another cluster. « Typical » preferences for CLUSTER are: Book 2, very high Book 3, high Books 5 & 6, may be recommended (Books 1 & 4, not recommended)

78. MLSS Sydney, February 18, 2015 – Machine Learning for Recommender Systems Clustering Customer F is classified as a new member of CLUSTER will receive recommendations based on the CLUSTER's preferences : Book 2 will be highly recommended to Customer F Book 6 will also be recommended to some extent

79. MLSS Sydney, February 18, 2015 – Machine Learning for Recommender Systems Clustering + It can also be applied for selecting the k most relevant neighbours in a CF algorithm + Faster: recommendations are per cluster - less personalized: recommendations are per cluster vs. in CF they are per user

80. MLSS Sydney, February 18, 2015 – Machine Learning for Recommender Systems Association rules Past purchases used to find relationships of common purchases

81. MLSS Sydney, February 18, 2015 – Machine Learning for Recommender Systems Association rules + Fast to implement + Fast to execute + Not much storage space required + Not « individual » specific + Very successful in broad applications for large populations, such as shelf layout in retail stores - Not suitable if preferences change rapidly - Rules can be used only when enough data validates them. False associations can arise

82. MLSS Sydney, February 18, 2015 – Machine Learning for Recommender Systems Limitations of CF ColdCold Start: It needs to have enough users in the system. New items need to get enough ratings. Sparsity: it is hard to find users who rated the same items. Popularity Bias: Cannot recommend items to users with unique tastes. Tends to recommend popular items.

83. MLSS Sydney, February 18, 2015 – Machine Learning for Recommender Systems Cold-start New User ProblemNew User Problem: the system must first learn the user’s preferences from the ratings. Hybrid RS, which combines content-based and collaborative techniques, can help. New Item ProblemNew Item Problem: Until the new item is rated by a substantial number of users, the RS is not able to recommend it.

85. MLSS Sydney, February 18, 2015 – Machine Learning for Recommender Systems Content-Based Recommendations Recommendations are based on the information on the contentcontent of itemsof items rather than on other users’ opinions. Use a machine learning algorithm to model the users' preferences from examples based on a description of the content.

86. MLSS Sydney, February 18, 2015 – Machine Learning for Recommender Systems What is content of an item? Explicit attributes or characteristics e.g. for a movie: Genre: Action / adventure Feature: Bruce Willis Year: 1995 Textual content e.g. for a book: title, description, table of content

87. MLSS Sydney, February 18, 2015 – Machine Learning for Recommender Systems In Content-Based Recommendations... The recommended items for a user are based on the profile built up by analysing the content of the items the user has liked in the past

88. MLSS Sydney, February 18, 2015 – Machine Learning for Recommender Systems Content-Based Recommendation Suitable for text-based products (web pages, books) Items are “described” by their features (e.g. keywords) Users are described by the keywords in the items they bought Recommendations based on the match between the content (item keywords) and user keywords The user model can also be a classifier (Neural Networks, SVM, Naïve Bayes...)

89. MLSS Sydney, February 18, 2015 – Machine Learning for Recommender Systems Advantages of CB Approach + No need for data on other users. + No cold-start or sparsity problems. + Can recommend to users with unique tastes. + Can recommend new and unpopular items + No first-rater problem. + Can provide explanations of recommended items by listing content-features that caused an item to be recommended.

90. MLSS Sydney, February 18, 2015 – Machine Learning for Recommender Systems Disadvantages of CB Approach - Only for content that can be encoded as meaningful features. - Some types of items (e.g. movies, music)are not amenable to easy feature extraction methods - Even for texts, IR techniques cannot consider multimedia information, aesthetic qualities, download time: a positive rating could be not related to the presence of certain keywords - Users’ tastes must be represented as a learnable function of these content features. - Hard to exploit quality judgements of other users. - Difficult to implement serendipity

91. MLSS Sydney, February 18, 2015 – Machine Learning for Recommender Systems Content-based Methods Content(s):= item profile, i.e. a set of attributes/keywords characterizing item s. weight wij measures the 'Importance” (or “informativeness”) of word kj in document dj term frequency/inverse document frequency(TF-IDF) is a popular weighting technique in IR

92. MLSS Sydney, February 18, 2015 – Machine Learning for Recommender Systems Content-based User Profile ContentBasedProfile(c):= profile of user c profiles are obtained by: analysing the content of the previous items using keyword analysis techniques e.g., ContentBasedProfile(c):=(wc1, . . . , wck) a vector of weights, where wci denotes the importance of keyword ki to user c

93. MLSS Sydney, February 18, 2015 – Machine Learning for Recommender Systems Similarity Measurements In content-based systems, the utility function u(c,s) is defined as: where ContentBasedProfile(c) of user c and Content(s) of document s are both represented as TF-IDF vectors of keyword weights. uc ,s=scoreContentBasedProfile c ,Content s

94. MLSS Sydney, February 18, 2015 – Machine Learning for Recommender Systems Similarity Measurements Utility function u(c,s) usually represented by some scoring heuristic defined in terms of vectors, such as the cosine similarity measure. uc ,s=cos wc , ws= wc⋅w s ∥ wc∥2×∥ ws∥2 = ∑i=1 K wi , c wi , s ∑i=1 K wi , c 2 ∑i=1 K wi , s 2

95. MLSS Sydney, February 18, 2015 – Machine Learning for Recommender Systems Content-based Recommendation. An (unrealistic) example How to compute recommendations of books based onlyon their title? A customer buys the book: Building data mining applications for CRM 7 Books are possible candidates for a recommendation: Accelerating Customer Relationships: Using CRM and Relationship Technologies Mastering Data Mining: The Art and Science of Customer Relationship Management Data Mining Your Website Introduction to marketing Consumer behaviour Marketing research, a handbook Customer knowledge management

96. MLSS Sydney, February 18, 2015 – Machine Learning for Recommender Systems building data mining applications for crm Accelerating Customer Relationships: Using CRM and Relationship Technologies Mastering Data Mining: The Art and Science of Customer Relationship Management Data Mining Your Website Introduction to marketing consumer behavior marketing research, a handbook customer knowledge management a 1 accelerating 1 and 1 1 application 1 art 1 behavior 1 building 1 consumer 1 crm 1 1 customer 1 1 1 data 1 1 1 for 1 handbook 1 introduction 1 knowledge 1 management 1 1 marketing 1 1 mastering 1 mining 1 1 1 of 1 relationship 2 1 research 1 science 1 technology 1 the 1 to 1 using 1 website 1 your 1 COUNT

97. MLSS Sydney, February 18, 2015 – Machine Learning for Recommender Systems Content-based Recommendation Computes distances between this book & all others Recommends the « closest » books: #1: Data Mining Your Website #2: Accelerating Customer Relationships: Using CRM and Relationship Technologies #3: Mastering Data Mining: The Art and Science of Customer Relationship Management

98. MLSS Sydney, February 18, 2015 – Machine Learning for Recommender Systems building data mining applications for crm Accelerating Customer Relationships: Using CRM and Relationship Technologies Mastering Data Mining: The Art and Science of Customer Relationship Management Data Mining Your Website Introduction to marketing consumer behavior marketing research, a handbook customer knowledge management a 0.000 0.000 0.000 0.000 0.000 0.000 0.537 0.000 accelerating 0.000 0.432 0.000 0.000 0.000 0.000 0.000 0.000 and 0.000 0.296 0.256 0.000 0.000 0.000 0.000 0.000 application 0.502 0.000 0.000 0.000 0.000 0.000 0.000 0.000 art 0.000 0.000 0.374 0.000 0.000 0.000 0.000 0.000 behavior 0.000 0.000 0.000 0.000 0.000 0.707 0.000 0.000 building 0.502 0.000 0.000 0.000 0.000 0.000 0.000 0.000 consumer 0.000 0.000 0.000 0.000 0.000 0.707 0.000 0.000 crm 0.344 0.296 0.000 0.000 0.000 0.000 0.000 0.000 customer 0.000 0.216 0.187 0.000 0.000 0.000 0.000 0.381 data 0.251 0.000 0.187 0.316 0.000 0.000 0.000 0.000 for 0.502 0.000 0.000 0.000 0.000 0.000 0.000 0.000 handbook 0.000 0.000 0.000 0.000 0.000 0.000 0.537 0.000 introduction 0.000 0.000 0.000 0.000 0.636 0.000 0.000 0.000 knowledge 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.763 management 0.000 0.000 0.256 0.000 0.000 0.000 0.000 0.522 marketing 0.000 0.000 0.000 0.000 0.436 0.000 0.368 0.000 mastering 0.000 0.000 0.374 0.000 0.000 0.000 0.000 0.000 mining 0.251 0.000 0.187 0.316 0.000 0.000 0.000 0.000 of 0.000 0.000 0.374 0.000 0.000 0.000 0.000 0.000 relationship 0.000 0.468 0.256 0.000 0.000 0.000 0.000 0.000 research 0.000 0.000 0.000 0.000 0.000 0.000 0.537 0.000 science 0.000 0.000 0.374 0.000 0.000 0.000 0.000 0.000 technology 0.000 0.432 0.000 0.000 0.000 0.000 0.000 0.000 the 0.000 0.000 0.374 0.000 0.000 0.000 0.000 0.000 to 0.000 0.000 0.000 0.000 0.636 0.000 0.000 0.000 using 0.000 0.432 0.000 0.000 0.000 0.000 0.000 0.000 website 0.000 0.000 0.000 0.632 0.000 0.000 0.000 0.000 your 0.000 0.000 0.000 0.632 0.000 0.000 0.000 0.000 TFIDF Normed Vectors Data 0.187 0.316 Data mining your website Mastering Data Mining: The Art and Science of Customer Relationship Management

100. MLSS Sydney, February 18, 2015 – Machine Learning for Recommender Systems Context Context is a dynamic set of factors describing the state of the user at the moment of the user's experience Context factors can rapidly change and affect how the user perceives an item

101. MLSS Sydney, February 18, 2015 – Machine Learning for Recommender Systems Context in Recommendations Temporal: Time of the day, weekday/end Spatial: Location, Home, Work etc. Social: with Friends, Family Recommendations should be tailored to the user & to the current Context of the user

102. MLSS Sydney, February 18, 2015 – Machine Learning for Recommender Systems Level of Adaptation

103. MLSS Sydney, February 18, 2015 – Machine Learning for Recommender Systems Context-Aware RS: Pre-filtering

104. MLSS Sydney, February 18, 2015 – Machine Learning for Recommender Systems Context-Aware RS: Post-filtering

105. MLSS Sydney, February 18, 2015 – Machine Learning for Recommender Systems Context-Aware RS: Tensor Factorization

106. MLSS Sydney, February 18, 2015 – Machine Learning for Recommender Systems Factorization Machines

107. MLSS Sydney, February 18, 2015 – Machine Learning for Recommender Systems Context-Aware RS: Pre-filtering + Simple + Works with large amounts of data - Increases sparseness - Does not scale well with many Context variables Post-filtering + Single model + Takes into account context interactions - Computationally expensive - Increases data sparseness - Does not model the Context directly Tensor Factorization & Factorization Machines + Performance + Linear scalability + Models context directly

109. MLSS Sydney, February 18, 2015 – Machine Learning for Recommender Systems Ranking Most recommendations are presented in a sorted list Recommendation is a ranking problem Popularity is the obvious baseline Users pay attention to few items at the top of the list

110. MLSS Sydney, February 18, 2015 – Machine Learning for Recommender Systems Ranking: Approaches (I) Re-rank: based on features e.g. predicted rating, popularity, etc (II) Learning to Rank: Build Ranking CF models

111. MLSS Sydney, February 18, 2015 – Machine Learning for Recommender Systems Learning to rank Machine learning task: Rank the most relevant items as high as possible in the recommendation list Does not try to predict a rating, but the order of preference Training data have partial order or binary judgments (relevant/not relevant) Can be treated as a standard supervised classification problem

112. MLSS Sydney, February 18, 2015 – Machine Learning for Recommender Systems Matrix Factorization for Ranking 2 5 4 5 4 4 1 5 5 4 1 2 5 2 4 1 d d

113. MLSS Sydney, February 18, 2015 – Machine Learning for Recommender Systems Learning to rank - Approaches 1) Pointwise Ranking function minimizes loss function defined on individual relevance judgment e.g. Ranking score based on regression or classification Ordinal regression, Logistic regression, SVM

114. MLSS Sydney, February 18, 2015 – Machine Learning for Recommender Systems Learning to rank - Approaches 2) Pairwise Loss function is defined on pair-wise preferences Goal: minimize number of inversions in ranking BPR, RankBoost, RankNet, FRank…

115. MLSS Sydney, February 18, 2015 – Machine Learning for Recommender Systems Learning to rank - Metrics Metrics evaluate the quality of a recommendation list - Normalized Discounted Cumulative Gain NDCG Computed for the first k items The NDCG@k of a list of items ratings Y, permuted by π is: where πs is the permutation which sorts Y decreasingly

116. MLSS Sydney, February 18, 2015 – Machine Learning for Recommender Systems Learning to rank - Metrics 1/log(7)3/log(6)7/log(5)15/log(4)31/log(3)

117. MLSS Sydney, February 18, 2015 – Machine Learning for Recommender Systems Learning to rank - Metrics Mean Reciprocal Rank (MRR)

118. MLSS Sydney, February 18, 2015 – Machine Learning for Recommender Systems Learning to rank - Metrics Mean Average Precision (MAP) S the set of relevant items, N #users

119. MLSS Sydney, February 18, 2015 – Machine Learning for Recommender Systems Non-smoothness of Metrics

122. MLSS Sydney, February 18, 2015 – Machine Learning for Recommender Systems Learning to rank - Approaches 3) List-wise Direct optimization of ranking metrics, List-wise loss minimization for CF a.k.a Collaborative Ranking CoFiRank: optimizes an upper bound of NDCG (Smooth version) CLiMF : optimizes a smooth version of MRR TFMAP: optimizes a smooth version of MAP AdaRank: uses boosting to optimize NDCG

123. MLSS Sydney, February 18, 2015 – Machine Learning for Recommender Systems Learning to rank - Overview Point-wise Pair-wise List-wise Matrix factorization [Koren 2009] BPR [Rendle 2009] CofiRank [Weimer 2007] SVD++ [Koren 2008] EigenRank [Liu 2008] ListRank [Shi 2010] OrdRec [Koren 2011] pLPA [Liu 2009] WLT [Volkovs 2012] Factorization machines [Rendle 2012] CR [Balakrishnan 2012] TFMAP [Shi 2012a] (All rating prediction methods) CLiMF [Shi 2012b] GAPfm [Shi 2013a] xCLiMF [Shi 2013b] Classification by What to Model

124. MLSS Sydney, February 18, 2015 – Machine Learning for Recommender Systems Learning to rank - Approaches 3) List-wise Direct optimization of ranking metrics, List-wise loss minimization for CF a.k.a Collaborative Ranking CoFiRank: optimizes an upper bound of NDCG (Smooth version) CLiMF : optimizes a smooth version of MRR TFMAP: optimizes a smooth version of MAP AdaRank: uses boosting to optimize NDCG

125. MLSS Sydney, February 18, 2015 – Machine Learning for Recommender Systems Learning to rank - Overview Proxy of Rankings Structured Estimation Non-smooth optimization Smoothing ranking measures ( Rating predicting methods CofiRank [Weimer 2007] WLT [Volkovs 2012] BPR [Rendle 2009] EigenRank [Liu 2008] TFMAP [Shi 2012a] pLPA [Liu 2009] CLiMF [Shi 2012b] ListRank [Shi 2010] GAPfm [Shi 2013a] CR [Balakrishnan 2012] XCLiMF [Shi 2013b] Classification by How to Model

126. MLSS Sydney, February 18, 2015 – Machine Learning for Recommender Systems Ordrec [Koren 2011] Pointwise, proxy 11 22 33 44 55 t1 t2 t3 t4

127. MLSS Sydney, February 18, 2015 – Machine Learning for Recommender Systems Ordrec [Koren 2011] Pointwise, proxy 11 22 33 44 55 t1 t2 t3 t4 Parameterize the threshold into: User factor, item factor and previous threshold Full distribution estimation

128. MLSS Sydney, February 18, 2015 – Machine Learning for Recommender Systems BPR [Rendel 2009] Pair-wise, smoothing Main idea: learn by comparing items that the user likes with items that he does not. When dealing with implicit data sample the unseen data as negative Optimize the AUC, use a smooth version for Optimization purposes

129. MLSS Sydney, February 18, 2015 – Machine Learning for Recommender Systems BPR [Rendel 2009] Pair-wise, smoothing Representation of pairwise preference data What Alice watched What Alice’s pairwise preferences (Row to Column) Transformers (T) ? Inception (I) ? Scent of a woman (S) 1 Forrest Gump (F) 1 Zombieland (Z) ? T I S F Z T ? - - ? I ? - - ? S + + ? + F + + ? + Z ? ? - -

130. MLSS Sydney, February 18, 2015 – Machine Learning for Recommender Systems Ordrec [Koren 2011] Ratings are not numeric, but ordinal. Ordinal modeling by learning the thresholds Items rated similarly are similar with each other. An item rated 4 should be more similar to an item rated 5 rather than to an item rated 1 Rating 1 2 3 4 5

131. MLSS Sydney, February 18, 2015 – Machine Learning for Recommender Systems i1 i2 i3 u1 2 3 4 u2 4 2 5 EigenRank [Liu 2008] Pair-wise, proxy Ranking-based similarity instead of rating based similarity User-user similarity based on Kendall Rank Correlation Neighbour selection for a user & item preference estimation Random walk with restart over pair-wise item preference graph

132. MLSS Sydney, February 18, 2015 – Machine Learning for Recommender Systems List-wise Learn to model individual ratings CofiRank ListRank CLiMF TFMAP GAPfm xCLiMF Learn to model ranked list, or to model evaluation metrics that capture the quality of ranked list List-wise perspective

133. MLSS Sydney, February 18, 2015 – Machine Learning for Recommender Systems CoFiRank [Weimer 2007] list-wise, structured est. Based on Structured estimation i.e. Machine Learning methods for complex output domains (graphs, sequences, etc.) Cast the Ranking problem as a Structured estimation problem i.e. Learn a function that is maximized for the best possible ranking with respect to a ranking measure e.g. nDCG

134. MLSS Sydney, February 18, 2015 – Machine Learning for Recommender Systems CoFiRank [Weimer 2007] list-wise, structured est. Step 1: Instead of maximizing Minimize

135. MLSS Sydney, February 18, 2015 – Machine Learning for Recommender Systems CoFiRank [Weimer 2007] list-wise, structured est. Step 2:Step 2: Use Create a linear mapping That is maximized for Denote by a decreasing non-negative sequence

136. MLSS Sydney, February 18, 2015 – Machine Learning for Recommender Systems CoFiRank [Weimer 2007] list-wise, structured est. Step 3:Step 3: Find convex upper on non-convex optimization problems: Let Maximization by solving a Linear Assignment Problem to get an estimate for the worst case loss (maximum margin)

137. MLSS Sydney, February 18, 2015 – Machine Learning for Recommender Systems CoFiRank [Weimer 2007] list-wise, structured est. [3, 2 ,4 ,1] [4, 2 ,1 ,3] [1, 0.7 ,0.57 ,0.5] [-0.5, 0.0 ,0.42 ,0.1] 2.9 2.6 1.9 1.6 1.6 2.6 2.9 1.9

138. MLSS Sydney, February 18, 2015 – Machine Learning for Recommender Systems CoFiRank [Weimer 2007] list-wise, structured est.

139. MLSS Sydney, February 18, 2015 – Machine Learning for Recommender Systems CLiMF [Shi 2012b] List-wise, Smoothing Ranking Metrics Why optimize mean reciprocal rank (MRR)? Focus at the very top of the list Try to get at least one interesting item at the top of the list How to optimize MRR? Find a smooth version of MRR Find a lower bound of the smoother MRR

140. MLSS Sydney, February 18, 2015 – Machine Learning for Recommender Systems Graphical Notation RelevantRelevant Irrelevant Irrelevant Irrelevant Irrelevant Irrelevant RelevantRelevant RelevantRelevant

141. MLSS Sydney, February 18, 2015 – Machine Learning for Recommender Systems Ranking using latent representation 55 44 11 44 22 11 22 11 [0.180, 0.019][0.180, 0.019] [0.013, 0.487][0.013, 0.487] [0.165, 0.632] [0.730, 0.109][0.730, 0.109] [0.420, 0.485] [0.463, 0.725] [0.894, 0.857] [0.942, 0.851] If user = [-100, -100] • 2d latent factor • We get the corresponding ranking

142. MLSS Sydney, February 18, 2015 – Machine Learning for Recommender Systems 2 1 1 44 44 1 55 2 [0.894, 0.857] [0.463, 0.725] [0.942, 0.851] [0.730, 0.109][0.730, 0.109] [0.013, 0.487][0.013, 0.487] [0.165, 0.632] [0.180, 0.019][0.180, 0.019] [0.420, 0.485] Matrix Factorization (for ranking) • Randomly initialize item vectors • Randomly initialize user vectors • While not converged • Compute rating prediction error • Update user factors • Update item factors • Let's say user is [-100, -100] • Compute the square error (5-<[-100, -100], [0.180, 0.19]>)2=1764 • Update user & item factors to the direction where the error is reduced (according to the gradient of the loss) 8 items with ratings & random factors

143. MLSS Sydney, February 18, 2015 – Machine Learning for Recommender Systems Learning: Stochastic Gradient Descent with Square Loss

144. MLSS Sydney, February 18, 2015 – Machine Learning for Recommender Systems Square Loss User: [3, 1], RMSE=6.7 1 2 44 1 2 1 55 44 [0.942, 0.851] [0.894, 0.857] [0.730, 0.109][0.730, 0.109] [0.463, 0.725] [0.420, 0.485] [0.165, 0.632] [0.180, 0.019][0.180, 0.019] [0.013, 0.487][0.013, 0.487]

145. MLSS Sydney, February 18, 2015 – Machine Learning for Recommender Systems Example: average precision (AP) • AP: we compute the precision at each relevant position and average them

146. MLSS Sydney, February 18, 2015 – Machine Learning for Recommender Systems Why is hard? Non Smoothness Example: AP u:[-20,-20] u:[20,20]

147. MLSS Sydney, February 18, 2015 – Machine Learning for Recommender Systems AP vs. RMSE The minimum in RMS and maximum in MAP do not correspond!

148. MLSS Sydney, February 18, 2015 – Machine Learning for Recommender Systems Non-smoothness of AP ymi =1 if item i is relevant for user m; else ymi = 0 The indicator function = 1 if it is true; else =0 rmi = rank of item i for user m

149. MLSS Sydney, February 18, 2015 – Machine Learning for Recommender Systems AP vs. RMSE The minimum in RMS and maximum in MAP do not correspond!

150. MLSS Sydney, February 18, 2015 – Machine Learning for Recommender Systems Learning to Rank for Top-k RecSys •Usually we care to make accurate ranking (and not rating) prediction •Square loss optimizes to accurately predict 1s and 5s. •RS should get the top k items right -> Ranking problem •Why not to learn how to rank directly? •Learning to Rank methods provide up to 30% performance improvements in off-line evaluations •It is possible, but a more complex task

151. MLSS Sydney, February 18, 2015 – Machine Learning for Recommender Systems How can we get a smooth AP? • We replace non smooth places of MAP with smooth approximation

152. MLSS Sydney, February 18, 2015 – Machine Learning for Recommender Systems How can we get a smooth MAP? • We replace non smooth places of MAP with smooth approximation

153. MLSS Sydney, February 18, 2015 – Machine Learning for Recommender Systems Smooth version of MAP u:[-20,-20] u:[20,20]

154. MLSS Sydney, February 18, 2015 – Machine Learning for Recommender Systems Not always the case: only sometimes approximation is very good…

155. MLSS Sydney, February 18, 2015 – Machine Learning for Recommender Systems Ranking Inconsistencies Achieving a perfect ranking for all users is not possible Two Sources of Inconsistencies: ● Factor Models (all models) have limited expressive power and cannot learn the perfect ranking for all users Ranking functions approximations are inconsistent e.g. A > B & B > C but C > A

156. MLSS Sydney, February 18, 2015 – Machine Learning for Recommender Systems Summary on Ranking 101

157. MLSS Sydney, February 18, 2015 – Machine Learning for Recommender Systems Ranking Metrics: AUC RR AP nDCG

158. MLSS Sydney, February 18, 2015 – Machine Learning for Recommender Systems AUC: Area Under the ROC Curve

159. MLSS Sydney, February 18, 2015 – Machine Learning for Recommender Systems RR: Reciprocal Rank

160. MLSS Sydney, February 18, 2015 – Machine Learning for Recommender Systems AP: Average Precision

161. MLSS Sydney, February 18, 2015 – Machine Learning for Recommender Systems AP vs.RR

162. MLSS Sydney, February 18, 2015 – Machine Learning for Recommender Systems nDCG: Normalized Discounted Cumulative Gain

163. MLSS Sydney, February 18, 2015 – Machine Learning for Recommender Systems AP AUC RR Square Loss

164. MLSS Sydney, February 18, 2015 – Machine Learning for Recommender Systems Top-k Ranking Focus on the very top of the recommendation list since users pay attention only to the first k items Top heavy ranking measures put much more emphasis at the top of the list (e.g. MRR, MAP, NDCG). Drop in measure non-linear with the position in the list AUC gives the same emphasis to the top as to the bottom of the list. Drop in measure linear to the position in the list

165. MLSS Sydney, February 18, 2015 – Machine Learning for Recommender Systems Learning to Rank for RecSys

166. MLSS Sydney, February 18, 2015 – Machine Learning for Recommender Systems Learning to Rank for IR & ML Query-document (Q-D) retrieval: Given a query, rank documents. RankSVM LambdaRank RankNet RankBoostAdaR ank ListNet SVM-MAP SoftRank ... Features of a Q-D pair, relevance label Learn a ranking function Predict relevances of new Q-D pairs Analogy: Query-doc --- User-item

167. MLSS Sydney, February 18, 2015 – Machine Learning for Recommender Systems Learning to Rank for RecSys Focus on the very top of the recommendation list since users pay attention only to the first k items Top heavy ranking measures put much more emphasis at the top of the list (e.g. MRR, MAP, NDCG). Drop in measure non-linear with the position in the list AUC gives the same emphasis to the top as to the bottom of the list. Drop in measure linear to the position in the list

168. MLSS Sydney, February 18, 2015 – Machine Learning for Recommender Systems The non-smoothness of Reciprocal Rank Reciprocal Rank (RR): The inverse of the rank of the first relevant item in a given list. Rank-based component Rank-based component

169. MLSS Sydney, February 18, 2015 – Machine Learning for Recommender Systems Reciprocal Rank

170. MLSS Sydney, February 18, 2015 – Machine Learning for Recommender Systems Lower Bound Jensen’s inequality Concavity of log function

171. MLSS Sydney, February 18, 2015 – Machine Learning for Recommender Systems What's the key? CLiMF reciprocal rank loss essentially pushes relevant items apart In the process at least one items ends up high-up in the list

172. MLSS Sydney, February 18, 2015 – Machine Learning for Recommender Systems Conventional Loss

173. MLSS Sydney, February 18, 2015 – Machine Learning for Recommender Systems CLiMF MRR-Loss

174. MLSS Sydney, February 18, 2015 – Machine Learning for Recommender Systems Optimizing RR vs. AUC Opt. RR Opt. AUC

175. MLSS Sydney, February 18, 2015 – Machine Learning for Recommender Systems List-wise xCLiMF [Shi 2013b] – Generalization to Ratings Expected reciprocal rank (ERR) [Chapelle 2009] A generalization of RR to graded relevance data Cascade nature compared to NDCG Equal contribution independent of the items ranked above Different contributions relative to the items ranked above NDCG ERR 0 1 0 3 0 0 5 0 3 0 0 1 0 3 0 0 5 0 3 0 = <

176. MLSS Sydney, February 18, 2015 – Machine Learning for Recommender Systems xCLiMF vs. CLiMF Results from Tuenti video watching dataset xCLiMF shows its advantage over CLiMF when applied to graded relevance data

177. MLSS Sydney, February 18, 2015 – Machine Learning for Recommender Systems List-wise TFMAP [Shi 2012a] – Generalization to Context Context-aware Recommendation

178. MLSS Sydney, February 18, 2015 – Machine Learning for Recommender Systems Problem Setting Given: Users’ implicit feedback on items under different contexts Target: To recommend a list of items to each user under any given context, as accurate as possible Top-N recommendation Context-aware Optimal in terms of MAP Items Users Contexts

179. MLSS Sydney, February 18, 2015 – Machine Learning for Recommender Systems TFMAP: Main concept CP tensor factorization U, V, C are latent factors (parameters to be learned) U, V, C are learned from optimizing the smoothed MAP

180. MLSS Sydney, February 18, 2015 – Machine Learning for Recommender Systems Social and Trust-based recommenders A social RS recommends items that are “popular” with the friends of the user. Friendship though does not imply trust “trust” in social-based RS can be per-user or topic-specific

181. MLSS Sydney, February 18, 2015 – Machine Learning for Recommender Systems Trust Inference Select 2 users: source (node A) & sink (C) Recommend to the source how much to trust the sink. source sink

182. MLSS Sydney, February 18, 2015 – Machine Learning for Recommender Systems Major Algorithms - Networks Advogato (Levien) Appleseed (Ziegler and Lausen) MoleTrust (Massa and Avesani) TidalTrust (Golbeck)

183. MLSS Sydney, February 18, 2015 – Machine Learning for Recommender Systems Building Recommender Systems Using Trust Use trust as a way to give more weight to some users Trust for collaborative filtering Use trust in place of (or combined with) similarity Trust for sorting and filtering Prioritize information from trusted sources

184. MLSS Sydney, February 18, 2015 – Machine Learning for Recommender Systems Social Recommendations

185. MLSS Sydney, February 18, 2015 – Machine Learning for Recommender Systems Other ways to use Social Social connections can be used in combination with other approaches In particular, “friendships” can be fed into collaborative filtering methods in different ways e.g. replace or modify user-user “similarity” by using social network information

187. MLSS Sydney, February 18, 2015 – Machine Learning for Recommender Systems Hybridization Methods Hybridization Method Description Weighted Outputs (scores or votes) from several techniques are combined with different degrees of importance to offer final recommendations Switching Depending on situation, the system changes from one technique to another Mixed Recommendations from several techniques are presented at the same time Feature combination Features from different recommendation sources are combined as input to a single technique Cascade The output from one technique is used as input of another that refines the result Feature augmentation The output from one technique is used as input features to another Meta-level The model learned by one recommender is used as input to another

188. MLSS Sydney, February 18, 2015 – Machine Learning for Recommender Systems Weighted Rating for an item is computed as the weighted sum of ratings produced by a pool of different RS. The weights are determined by training and get adjusted as new ratings arrive. Assumption: relative performance of the different techniques is uniform. Not true in general: e.g. CF performs worse for items with few ratings. e.g. P-Tango RS: a CB and a CF recommender equally weighted at first. Weights are adjusted as predictions are confirmed or not. RS with consensus scheme: each recommendation of a specific item counts as a vote for the item.

189. MLSS Sydney, February 18, 2015 – Machine Learning for Recommender Systems Feature Combination CF ratings of users are passed as additional feature to a CB. CB makes recommendations over this augmented data set.

190. MLSS Sydney, February 18, 2015 – Machine Learning for Recommender Systems Switching The system uses a criterion to switch between techniques The main problem is to identify a good switching criterion. e.g. The DailyLearner system uses a CB-CF. When CB cannot predict with sufficient confidence, it switches to CF.

191. MLSS Sydney, February 18, 2015 – Machine Learning for Recommender Systems Mixed Recommendations from more than one technique are presented together e.g. The PTV system recommends a TV viewing schedule for the user by combining recommendations from a CB and a CF system. CB uses the textual descriptions of TV shows; vs CF uses other users’ preferences. When collision occurs, the CB has priority.

192. MLSS Sydney, February 18, 2015 – Machine Learning for Recommender Systems Cascade At each iteration, a first recommendation technique produces a coarse ranking & a second technique refines the recommendation Cascading avoids employing the second, lower-priority, technique on items already well-differentiated by the first Requires a meaningful ordering of the techniques. E.g.: EntreeC is a restaurant RS uses its knowledge of restaurants to make recommendations based on the user’s stated interests. The recommendations are placed in buckets of equal preference, and the collaborative technique breaks ties

193. MLSS Sydney, February 18, 2015 – Machine Learning for Recommender Systems Feature Augmentation Very similar to the feature combination method: Here the output of one RS is incorporated into the processing of a second RS e.g.: Amazon.com generates text data (“related authors” and “related titles”) using its internal collaborative systems Libra system makes content-based recommendations of books based on these text data found in Amazon.com, using a naive Bayes text classifier

195. MLSS Sydney, February 18, 2015 – Machine Learning for Recommender Systems Beyond Explicit Ratings Implicit feedback is more readily available, and less noisy Already many approaches (e.g. SVD++) can make use of implicit feedback Ongoing research in combining explicit and implicit feedback D. H. Stern, R. Herbrich, and T. Graepel. Matchbox: large scale online bayesian recommendations. In Proc.of the 18th WWW, 2009. Koren Y and J. Sill. OrdRec: an ordinal model for predicting personalized item rating distributions. In Rec-Sys ’11, pages 117–124, 2011. Y. Koren. Factorization meets the neighborhood: a multifaceted collaborative filtering model. In Proceedings of the 14th ACM SIGKDD, 2008. Yifan Hu, Y. Koren, and C. Volinsky. Collaborative Filtering for Implicit Feedback Datasets. In Proc. Of the 2008 Eighth ICDM, pages 263–272, 2008.

196. MLSS Sydney, February 18, 2015 – Machine Learning for Recommender Systems Personalized Learning to Rank Better approaches to learning to rank that directly optimize ranking metrics and allow for personalization (e.g. CliMF & TFMAP) Y. Shi, A. Karatzoglou, L. Baltrunas, M. Larson, N. Oliver, and A. Hanjalic. CLiMF: learning to maximize reciprocal rank with collaborative less-is-more filtering. In Proc. of the sixth Recsys, 2012. Y. Shi, A. Karatzoglou, L. Baltrunas, M. Larson,A. Hanjalic, and N. Oliver. TFMAP: optimizing MAP for top-n context-aware recommendation. In Proc. Of the 35th SIGIR, 2012.

197. MLSS Sydney, February 18, 2015 – Machine Learning for Recommender Systems Context-aware Recommendations Beyond the traditional 2D user-item space Recommendations should also respond to user context (e.g. location, time of the day...) Many different approaches such as Tensor Factorization or Factorization Machines A. Karatzoglou, X. Amatriain, L. Baltrunas, and N. Oliver. Multiverse recommendation: n-dimensional tensor factorization for context-aware collaborative filtering. In Proc. of the fourth ACM Recsys, 2010. S. Rendle, Z. Gantner, C. Freudenthaler, and L. Schmidt-Thieme. Fast context- aware recommendations with factorization machines. In Proc. of the 34th ACM SIGIR, 2011.

198. MLSS Sydney, February 18, 2015 – Machine Learning for Recommender Systems User choice and presentation effects We log the recommended items to the users and their choice We can use this information as negative feedback (not chosen) and positive feedback (chosen). S.H. Yang, B. Long, A.J. Smola, H. Zha, and Z. Zheng. Collaborative competitive filtering: learning recommender using context of user choice. In Proc. of the 34th ACM SIGIR, 2011.

199. MLSS Sydney, February 18, 2015 – Machine Learning for Recommender Systems Social Recommendations Beyond trust-based Cold-starting with Social Information Combining Social with CF Finding “experts” J. Delporte, A. Karatzoglou, T. Matuszczyk, S. Canu. Socially Enabled Preference Learning from Implicit Feedback Data. In Proc. of ECML/PKDD 2013 N. N. Liu, X. Meng, C. Liu, and Q. Yang. Wisdom of the better few: cold start recommendation via representative based rating elicitation. In Proc. of RecSys’11, 2011. M. Jamali and M. Ester. Trustwalker: a random walk model for combining trust-based and item-based recommendation. In Proc. of KDD ’09, 2009. J. Noel, S. Sanner, K. Tran, P. Christen, L. Xie, E. V. Bonilla, E. Abbasnejad, and N. Della Penna. New objective functions for social collaborative filtering. In Proc. of WWW ’12, pages 859–868, 2012. X. Yang, H. Steck, Y. Guo, and Y. Liu. On top-k recommendation using social networks. In Proc. of RecSys’12, 2012.

200. MLSS Sydney, February 18, 2015 – Machine Learning for Recommender Systems Diversity in Recommendation (I) Recommendations from a music on-line retailer: No diversity: pop albums from female singers. Some are redundant. Born This Way Pink Friday Dangerously in Love Born This Way – The Remix Femme Fatale Can't be Tamed Teenage Dream Lady Gaga Nicki Minaj Beyoncé Lady Gaga Britney Spears Miley Cyrus Katy Perry

201. MLSS Sydney, February 18, 2015 – Machine Learning for Recommender Systems Diversity in Recommendation (II) Some good music recommendations: Different authors and genres. Not similar between them. These are much better recommendations! Wrecking Ball Not your Kind of People Like a Prayer Choice of Weapon Sweet Heart Sweet Light The Light the Dead See Little Broken Hearts B. Springsteen Garbage Madonna The Cult Spiritualized Soulsavers Norah Jones

202. MLSS Sydney, February 18, 2015 – Machine Learning for Recommender Systems Re-Ranking Recommender Re-ranking top 5 not diverse D top 5 diverse Ziegler et al. 2005 Zhang et al. 2008 Vargas et al. 2011 comedy drama action

203. MLSS Sydney, February 18, 2015 – Machine Learning for Recommender Systems Sub-Profiles comedy action drama

204. MLSS Sydney, February 18, 2015 – Machine Learning for Recommender Systems Novelty and Diversity Metrics

206. MLSS Sydney, February 18, 2015 – Machine Learning for Recommender Systems Conclusions Recommender Systems (RS) are an important application of Machine Learning RS have the potential to become as important as Search is now However, RS are more than Machine Learning HCI Economical models ...

207. MLSS Sydney, February 18, 2015 – Machine Learning for Recommender Systems Conclusions RS are fairly new but already grounded on well-proven technology Collaborative Filtering Machine Learning Content Analysis Social Network Analysis … However, there are still many open questions and a lot of interesting research to do!

209. MLSS Sydney, February 18, 2015 – Machine Learning for Recommender Systems "Recommender Systems Handbook." Ricci, Francesco, Lior Rokach, Bracha Shapira, and Paul B. Kantor. (2010). “Recommender systems: an introduction”. Jannach, Dietmar, et al. Cambridge University Press, 2010. “Toward the Next Generation of Recommender Systems: A Survey of the State-of-the-Art and Possible Extensions”. G. Adomavicious and A. Tuzhilin. 2005. IEEE Transactions on Knowledge and Data Engineering, 17 (6) “Item-based Collaborative Filtering Recommendation Algorithms”, B. Sarwar et al. 2001. Proceedings of World Wide Web Conference. “Lessons from the Netflix Prize Challenge.”. R. M. Bell and Y. Koren. SIGKDD Explor. Newsl., 9(2):75–79, December 2007. “Beyond algorithms: An HCI perspective on recommender systems”.K. Swearingen and R. Sinha. In ACM SIGIR 2001 Workshop on Recommender Systems “Recommender Systems in E-Commerce”. J. Ben Schafer et al. ACM Conference on Electronic Commerce. 1999- “Introduction to Data Mining”, P. Tan et al. Addison Wesley. 2005 References

210. MLSS Sydney, February 18, 2015 – Machine Learning for Recommender Systems “Evaluating collaborative filtering recommender systems”. J. L. Herlocker, J. A. Konstan, L. G. Terveen, and J. T. Riedl. ACM Trans. Inf. Syst., 22(1):5–53, 2004. “Trust in recommender systems”. J. O’Donovan and B. Smyth. In Proc. of IUI ’05, 2005. “Content-based recommendation systems”. M. Pazzani and D. Billsus. In The Adaptive Web, volume 4321. 2007. “Fast context-aware recommendations with factorization machines”. S. Rendle, Z. Gantner, C. Freudenthaler, and L. Schmidt- Thieme. In Proc. of the 34th ACM SIGIR, 2011. “Restricted Boltzmann machines for collaborative filtering”.R. Salakhutdinov, A. Mnih, and G. E. Hinton.In Proc of ICML ’07, 2007 “Learning to rank: From pairwise approach to listwise approach”.Z. Cao and T. Liu. In In Proceedings of the 24th ICML, 2007. “Introduction to Data Mining”, P. Tan et al. Addison Wesley. 2005 References

211. MLSS Sydney, February 18, 2015 – Machine Learning for Recommender Systems Balakrishnan, S., & Chopra, S. (2012, February). Collaborative ranking. In Proceedings of the fifth ACM international conference on Web search and data mining (pp. 143-152). ACM. Burges, C., Shaked, T., Renshaw, E., Lazier, A., Deeds, M., Hamilton, N., & Hullender, G. (2005, August). Learning to rank using gradient descent. In Proceedings of the 22nd international conference on Machine learning (pp. 89-96). ACM. Cao, Z., Qin, T., Liu, T. Y., Tsai, M. F., & Li, H. (2007, June). Learning to rank: from pairwise approach to listwise approach. In Proceedings of the 24th international conference on Machine learning (pp. 129-136). ACM. Chapelle, O., Metlzer, D., Zhang, Y., & Grinspan, P. (2009, November). Expected reciprocal rank for graded relevance. In Proceedings of the 18th ACM conference on Information and knowledge management (pp. 621-630). ACM. Joachims, T. (2002, July). Optimizing search engines using clickthrough data. In Proceedings of the eighth ACM SIGKDD international conference on Knowledge discovery and data mining (pp. 133-142). ACM. Koren, Y. (2008, August). Factorization meets the neighborhood: a multifaceted collaborative filtering model. In Proceedings of the 14th ACM SIGKDD international conference on Knowledge discovery and data mining (pp. 426-434). ACM. Koren, Y., Bell, R., & Volinsky, C. (2009). Matrix factorization techniques for recommender systems. Computer, 42(8), 30-37. Koren, Y., & Sill, J. (2011, October). OrdRec: an ordinal model for predicting personalized item rating distributions. In Proceedings of the fifth ACM conference on Recommender systems (pp. 117-124). ACM. Liu, N. N., & Yang, Q. (2008, July). Eigenrank: a ranking-oriented approach to collaborative filtering. In Proceedings of the 31st annual international ACM SIGIR conference on Research and development in information retrieval (pp. 83-90). ACM Liu, N. N., Zhao, M., & Yang, Q. (2009, November). Probabilistic latent preference analysis for collaborative filtering. In Proceedings of the 18th ACM conference on Information and knowledge management (pp. 759-766). ACM. Quoc, C., & Le, V. (2007). Learning to rank with nonsmooth cost functions. Proceedings of the Advances in Neural Information Processing Systems, 19, 193-200. Rendle, S., Freudenthaler, C., Gantner, Z., & Schmidt-Thieme, L. (2009, June). BPR: Bayesian personalized ranking from implicit feedback. In Proceedings of the Twenty-Fifth Conference on Uncertainty in Artificial Intelligence (pp. 452-461). AUAI Press. References

212. MLSS Sydney, February 18, 2015 – Machine Learning for Recommender Systems Salakhutdinov, R., Mnih, A., & Hinton, G. (2007, June). Restricted Boltzmann machines for collaborative filtering. In Proceedings of the 24th international conference on Machine learning (pp. 791-798). ACM. Shi, Y., Larson, M., & Hanjalic, A. (2010, September). List-wise learning to rank with matrix factorization for collaborative filtering. In Proceedings of the fourth ACM conference on Recommender systems (pp. 269-272). ACM. Shi, Y., Karatzoglou, A., Baltrunas, L., Larson, M., Hanjalic, A., & Oliver, N. (2012a, August). TFMAP: Optimizing MAP for top- n context-aware recommendation. In Proceedings of the 35th international ACM SIGIR conference on Research and development in information retrieval (pp. 155-164). ACM. Shi, Y., Karatzoglou, A., Baltrunas, L., Larson, M., Oliver, N., & Hanjalic, A. (2012b, September). CLiMF: learning to maximize reciprocal rank with collaborative less-is-more filtering. In Proceedings of the sixth ACM conference on Recommender systems (pp. 139-146). ACM. Taylor, M., Guiver, J., Robertson, S., & Minka, T. (2008, February). Softrank: optimizing non-smooth rank metrics. In Proceedings of the international conference on Web search and web data mining (pp. 77-86). ACM.. Volkovs, M., & Zemel, R. S. (2012). Collaborative ranking with 17 parameters. In Advances in Neural Information Processing Systems (pp. 2303-2311). Weimer, M., Karatzoglou, A., Le, Q. V., & Smola, A. J. (2007). Cofi rank-maximum margin matrix factorization for collaborative ranking. In Advances in neural information processing systems (pp. 1593-1600). Xu, J., & Li, H. (2007, July). Adarank: a boosting algorithm for information retrieval. In Proceedings of the 30th annual international ACM SIGIR conference on Research and development in information retrieval (pp. 391-398). ACM. Yue, Y., Finley, T., Radlinski, F., & Joachims, T. (2007, July). A support vector method for optimizing average precision. In Proceedings of the 30th annual international ACM SIGIR conference on Research and development in information retrieval (pp. 271-278). ACM. References

213. MLSS Sydney, February 18, 2015 – Machine Learning for Recommender Systems Recsys Wiki: http://recsyswiki.com/ Recsys conference Webpage: http://recsys.acm.org/ Recommender Systems Books Webpage: http://www.recommenderbook.net/ Mahout Project: http://mahout.apache.org/ MyMediaLite Project: http://www.mymedialite.net/ Online resources

214. MLSS Sydney, February 18, 2015 – Machine Learning for Recommender Systems Thank you! Questions? Alexandros Karatzoglou alexk@tid.es @alexk_z

Machine Learning for Recommender Systems MLSS 2015 Sydney

You are reading a preview.

Check these out next

Machine Learning for Recommender Systems MLSS 2015 Sydney

More Related Content

Slideshows for you (20)

Similar to Machine Learning for Recommender Systems MLSS 2015 Sydney (20)

More from Alexandros Karatzoglou (9)

Recently uploaded (20)

Machine Learning for Recommender Systems MLSS 2015 Sydney