The document describes a multi-criteria recommender system that exploits aspect-based sentiment analysis of user reviews. It involves a two-step methodology: 1) performing aspect extraction and sentiment analysis on user reviews using an algorithm based on SABRE to identify aspects, sub-aspects, and sentiment, and 2) creating and populating a multi-criteria data model with the extracted information and using it to generate recommendations. The system aims to develop a multi-criteria data model for recommendations without overwhelming users by automatically extracting product aspects and sentiments from reviews rather than requiring users to manually evaluate each aspect.
80 ĐỀ THI THỬ TUYỂN SINH TIẾNG ANH VÀO 10 SỞ GD – ĐT THÀNH PHỐ HỒ CHÍ MINH NĂ...
A Multi-Criteria Recommender System Exploiting Aspect-based Sentiment Analysis of Users' Reviews
1. @cataldomusto @pasqualelops
@semeraro_g @SWAP_research
A Multi-criteria Recommender System
Exploiting Aspect-based Sentiment
Analysis of Users’ Reviews
CATALDO MUSTO, MARCO DE GEMMIS, GIOVANNI SEMERARO, PASQUALE LOPS
UNIVERSITÀ DEGLI STUDI DI BARI ‘ALDO MORO’ - ITALY
RecSys 2017 - 11th ACM Conference on
Recommender Systems
Como, Italy
August 30, 2017
cataldo.musto@uniba.it
2. Multi-Criteria RecSys
Cataldo Musto, Marco de Gemmis, Giovanni Semeraro. Pasquale Lops
A Multi-criteria Recommender System Exploiting Aspect-based Sentiment Analysis of Users’ Reviews. RECSYS 2017. Como, Italy. August 30, 2017
Not a new concept (*)
Each user evaluates
each aspect of
the item
(*) Adomavicius, Gediminas, and YoungOk Kwon. "Multi-criteria recommender
systems." Recommender Systems Handbook. Springer US, 2015. 847-880.
3. Multi-Criteria RecSys
Cataldo Musto, Marco de Gemmis, Giovanni Semeraro. Pasquale Lops
A Multi-criteria Recommender System Exploiting Aspect-based Sentiment Analysis of Users’ Reviews. RECSYS 2017. Como, Italy. August 30, 2017
Not a new concept (*)
Each user evaluates
each aspect of
the item
(*) Adomavicius, Gediminas, and YoungOk Kwon. "Multi-criteria recommender
systems." Recommender Systems Handbook. Springer US, 2015. 847-880.
Problem:
Overwhelming!
4. Multi-Criteria RecSys
Cataldo Musto, Marco de Gemmis, Giovanni Semeraro. Pasquale Lops
A Multi-criteria Recommender System Exploiting Aspect-based Sentiment Analysis of Users’ Reviews. RECSYS 2017. Como, Italy. August 30, 2017
Not a new concept (*)
Each user evaluates
each aspect of
the item
(*) Adomavicius, Gediminas, and YoungOk Kwon. "Multi-criteria recommender
systems." Recommender Systems Handbook. Springer US, 2015. 847-880.
Problem: Aspects
are not fixed!
5. Multi-Criteria RecSys
Cataldo Musto, Marco de Gemmis, Giovanni Semeraro. Pasquale Lops
A Multi-criteria Recommender System Exploiting Aspect-based Sentiment Analysis of Users’ Reviews. RECSYS 2017. Como, Italy. August 30, 2017
Not a new concept (*)
Each user evaluates
each aspect of
the item
(*) Adomavicius, Gediminas, and YoungOk Kwon. "Multi-criteria recommender
systems." Recommender Systems Handbook. Springer US, 2015. 847-880.
Problem: Aspects can
be further modeled as a
hierarchy
6. Research Question
Cataldo Musto, Marco de Gemmis, Giovanni Semeraro. Pasquale Lops
A Multi-criteria Recommender System Exploiting Aspect-based Sentiment Analysis of Users’ Reviews. RECSYS 2017. Como, Italy. August 30, 2017
How to develop a
multi-criteria data model
without overwhelming
the user ?
7. Research Question
Cataldo Musto, Marco de Gemmis, Giovanni Semeraro. Pasquale Lops
A Multi-criteria Recommender System Exploiting Aspect-based Sentiment Analysis of Users’ Reviews. RECSYS 2017. Como, Italy. August 30, 2017
What is the performance of
such a data model in a
collaborative
recommendation scenario?
8. Our contribution
Cataldo Musto, Marco de Gemmis, Giovanni Semeraro. Pasquale Lops
A Multi-criteria Recommender System Exploiting Aspect-based Sentiment Analysis of Users’ Reviews. RECSYS 2017. Como, Italy. August 30, 2017
A multi-criteria collaborative
recommendation methodology exploiting
aspect-based sentiment analysis of users’ reviews
9. Methodology
Cataldo Musto, Marco de Gemmis, Giovanni Semeraro. Pasquale Lops
A Multi-criteria Recommender System Exploiting Aspect-based Sentiment Analysis of Users’ Reviews. RECSYS 2017. Como, Italy. August 30, 2017
Input: textual reviews
Step 1: aspect extraction and sentiment analysis
Step 2: creating multi-criteria data model
Output: recommendations
10. Methodology
Cataldo Musto, Marco de Gemmis, Giovanni Semeraro. Pasquale Lops
A Multi-criteria Recommender System Exploiting Aspect-based Sentiment Analysis of Users’ Reviews. RECSYS 2017. Como, Italy. August 30, 2017
Input: textual reviews
Step 1: aspect extraction and sentiment analysis
Step 2: creating multi-criteria data model
Output: recommendations
11. Methodology
Cataldo Musto, Marco de Gemmis, Giovanni Semeraro. Pasquale Lops
A Multi-criteria Recommender System Exploiting Aspect-based Sentiment Analysis of Users’ Reviews. RECSYS 2017. Como, Italy. August 30, 2017
Input: textual reviews
Step 1: aspect extraction and sentiment analysis
Step 2: creating and filling our multi-criteria data model
Output: recommendations
12. Methodology
Cataldo Musto, Marco de Gemmis, Giovanni Semeraro. Pasquale Lops
A Multi-criteria Recommender System Exploiting Aspect-based Sentiment Analysis of Users’ Reviews. RECSYS 2017. Como, Italy. August 30, 2017
Input: textual reviews
Step 1: aspect extraction and sentiment analysis
Step 2: creating and filling our multi-criteria data model
Output: recommendations
13. Aspect Extraction and Sentiment Analysis
Cataldo Musto, Marco de Gemmis, Giovanni Semeraro. Pasquale Lops
A Multi-criteria Recommender System Exploiting Aspect-based Sentiment Analysis of Users’ Reviews. RECSYS 2017. Como, Italy. August 30, 2017
Algorithm based on SABRE(*)
(Sentiment Aspect-based Retrieval Engine)
(*) Caputo, A., Basile, P., de Gemmis, M., Lops, P., Semeraro, G., & Rossiello, G. (2017). SABRE: A
Sentiment Aspect-Based Retrieval Engine. In Information Filtering and Retrieval (pp. 63-78).
𝑅 = {𝑟1, 𝑟2 … 𝑟 𝑛}Input: set of reviews
Output: quintuples < 𝑟𝑖, 𝑎𝑖𝑗, 𝑎𝑖𝑗𝑘, 𝑟𝑒𝑙 𝑎𝑖𝑗𝑘, 𝑟𝑖 , 𝑠𝑒𝑛𝑡(𝑎𝑖𝑗𝑘, 𝑟𝑖) >
14. Aspect Extraction and Sentiment Analysis
Cataldo Musto, Marco de Gemmis, Giovanni Semeraro. Pasquale Lops
A Multi-criteria Recommender System Exploiting Aspect-based Sentiment Analysis of Users’ Reviews. RECSYS 2017. Como, Italy. August 30, 2017
Algorithm based on SABRE(*)
(Sentiment Aspect-based Retrieval Engine)
(*) Caputo, A., Basile, P., de Gemmis, M., Lops, P., Semeraro, G., & Rossiello, G. (2017). SABRE: A
Sentiment Aspect-Based Retrieval Engine. In Information Filtering and Retrieval (pp. 63-78).
𝑅 = {𝑟1, 𝑟2 … 𝑟 𝑛}Input: set of reviews
Output: quintuples < 𝑟𝑖, 𝑎𝑖𝑗, 𝑎𝑖𝑗𝑘, 𝑟𝑒𝑙 𝑎𝑖𝑗𝑘, 𝑟𝑖 , 𝑠𝑒𝑛𝑡(𝑎𝑖𝑗𝑘, 𝑟𝑖) >
𝑟𝑖 =
𝑎𝑖𝑗, 𝑎𝑖𝑗𝑘=
𝑟𝑒𝑙 𝑎𝑖𝑗𝑘, 𝑟𝑖 , 𝑠𝑒𝑛𝑡(𝑎𝑖𝑗𝑘, 𝑟𝑖) =
i-th review
j-th aspect and k-th sub-aspect in the i-th review
relevance and sentiment
15. Aspect Extraction and Sentiment Analysis
Cataldo Musto, Marco de Gemmis, Giovanni Semeraro. Pasquale Lops
A Multi-criteria Recommender System Exploiting Aspect-based Sentiment Analysis of Users’ Reviews. RECSYS 2017. Como, Italy. August 30, 2017
Algorithm based on SABRE(*)
(Sentiment Aspect-based Retrieval Engine)
(*) Caputo, A., Basile, P., de Gemmis, M., Lops, P., Semeraro, G., & Rossiello, G. (2017). SABRE: A
Sentiment Aspect-Based Retrieval Engine. In Information Filtering and Retrieval (pp. 63-78).
𝑅 = {𝑟1, 𝑟2 … 𝑟 𝑛}Input: set of reviews
Output: quintuples < 𝑟𝑖, 𝑎𝑖𝑗, 𝑎𝑖𝑗𝑘, 𝑟𝑒𝑙 𝑎𝑖𝑗𝑘, 𝑟𝑖 , 𝑠𝑒𝑛𝑡(𝑎𝑖𝑗𝑘, 𝑟𝑖) >
𝑟𝑖 =
𝑎𝑖𝑗, 𝑎𝑖𝑗𝑘=
𝑟𝑒𝑙 𝑎𝑖𝑗𝑘, 𝑟𝑖 , 𝑠𝑒𝑛𝑡(𝑎𝑖𝑗𝑘, 𝑟𝑖) =
i-th review
j-th aspect and k-th sub-aspect in the i-th review
relevance and sentiment
How do we extract aspects, relevance and sentiment?
16. Aspect Extraction and Sentiment Analysis
Cataldo Musto, Marco de Gemmis, Giovanni Semeraro. Pasquale Lops
A Multi-criteria Recommender System Exploiting Aspect-based Sentiment Analysis of Users’ Reviews. RECSYS 2017. Como, Italy. August 30, 2017
Aspect Extraction
(*) Caputo, A., Basile, P., de Gemmis, M., Lops, P., Semeraro, G., & Rossiello, G. (2017). SABRE: A
Sentiment Aspect-Based Retrieval Engine. In Information Filtering and Retrieval (pp. 63-78).
Statistical approach based on the Kullback-Leibler (KL) Divergence
17. Aspect Extraction and Sentiment Analysis
Cataldo Musto, Marco de Gemmis, Giovanni Semeraro. Pasquale Lops
A Multi-criteria Recommender System Exploiting Aspect-based Sentiment Analysis of Users’ Reviews. RECSYS 2017. Como, Italy. August 30, 2017
Aspect Extraction
(*) Caputo, A., Basile, P., de Gemmis, M., Lops, P., Semeraro, G., & Rossiello, G. (2017). SABRE: A
Sentiment Aspect-Based Retrieval Engine. In Information Filtering and Retrieval (pp. 63-78).
Statistical approach based on the Kullback-Leibler (KL) Divergence
Measures the difference between the distribution of a term
in a generic corpus (e.g. BNC) and its distribution in a domain corpus
(e.g. hotel reviews)
Insight: the higher the divergence, the higher the
importance of the term in the domain
18. Aspect Extraction and Sentiment Analysis
Cataldo Musto, Marco de Gemmis, Giovanni Semeraro. Pasquale Lops
A Multi-criteria Recommender System Exploiting Aspect-based Sentiment Analysis of Users’ Reviews. RECSYS 2017. Como, Italy. August 30, 2017
Aspect Extraction
(*) Caputo, A., Basile, P., de Gemmis, M., Lops, P., Semeraro, G., & Rossiello, G. (2017). SABRE: A
Sentiment Aspect-Based Retrieval Engine. In Information Filtering and Retrieval (pp. 63-78).
Statistical approach based on the Kullback-Leibler (KL) Divergence
Measures the difference between the distribution of a term
in a generic corpus (e.g. BNC) and its distribution in a domain corpus
(e.g. hotel reviews)
Insight: the higher the divergence, the higher the
importance of the term in the domain
KL(room, BNC, hotel-reviews) >> 0
KL(food, BNC, hotel-reviews) > 0
KL(place, BNC, hotel-reviews) ~ 0
KL(politics, BNC, hotel-reviews) ~ 0
We label as ‘aspects’ the
nouns whose
KL-divergence is higher
than zero
19. Aspect Extraction and Sentiment Analysis
Cataldo Musto, Marco de Gemmis, Giovanni Semeraro. Pasquale Lops
A Multi-criteria Recommender System Exploiting Aspect-based Sentiment Analysis of Users’ Reviews. RECSYS 2017. Como, Italy. August 30, 2017
Aspect Extraction
(*) Caputo, A., Basile, P., de Gemmis, M., Lops, P., Semeraro, G., & Rossiello, G. (2017). SABRE: A
Sentiment Aspect-Based Retrieval Engine. In Information Filtering and Retrieval (pp. 63-78).
Statistical approach based on the Kullback-Leibler (KL) Divergence
Measures the difference between the distribution of a term
in a generic corpus (e.g. BNC) and its distribution in a domain corpus
(e.g. hotel reviews)
Insight: the higher the divergence, the higher the
importance of the term in the domain
KL(room, BNC, hotel-reviews) >> 0 YES
KL(food, BNC, hotel-reviews) > 0 YES
KL(place, BNC, hotel-reviews) ~ 0 NO
KL(politics, BNC, hotel-reviews) ~ 0 NO
20. Aspect Extraction and Sentiment Analysis
Cataldo Musto, Marco de Gemmis, Giovanni Semeraro. Pasquale Lops
A Multi-criteria Recommender System Exploiting Aspect-based Sentiment Analysis of Users’ Reviews. RECSYS 2017. Como, Italy. August 30, 2017
Aspect Extraction
(*) Caputo, A., Basile, P., de Gemmis, M., Lops, P., Semeraro, G., & Rossiello, G. (2017). SABRE: A
Sentiment Aspect-Based Retrieval Engine. In Information Filtering and Retrieval (pp. 63-78).
Statistical approach based on the Kullback-Leibler (KL) Divergence
Measures the difference between the distribution of a term
in a generic corpus (e.g. BNC) and its distribution in a domain corpus
(e.g. hotel reviews)
Insight: the higher the divergence, the higher the
importance of the term in the domain
KL(room, BNC, hotel-reviews) >> 0 YES
KL(food, BNC, hotel-reviews) > 0 YES
KL(place, BNC, hotel-reviews) ~ 0 NO
KL(politics, BNC, hotel-reviews) ~ 0 NO
Distinguishing aspect:
the set is not fixed!
21. Aspect Extraction and Sentiment Analysis
Cataldo Musto, Marco de Gemmis, Giovanni Semeraro. Pasquale Lops
A Multi-criteria Recommender System Exploiting Aspect-based Sentiment Analysis of Users’ Reviews. RECSYS 2017. Como, Italy. August 30, 2017
Sub-aspect Extraction
(*) Caputo, A., Basile, P., de Gemmis, M., Lops, P., Semeraro, G., & Rossiello, G. (2017). SABRE: A
Sentiment Aspect-Based Retrieval Engine. In Information Filtering and Retrieval (pp. 63-78).
Another distinguishing aspect: we can extract a hierarchy of terms
Based on Phraseness and Informativeness: They measure the gain
in information if two terms are modeled together
Insight: if phraseness and informativeness are high,
the terms have an high cohesion
22. Aspect Extraction and Sentiment Analysis
Cataldo Musto, Marco de Gemmis, Giovanni Semeraro. Pasquale Lops
A Multi-criteria Recommender System Exploiting Aspect-based Sentiment Analysis of Users’ Reviews. RECSYS 2017. Como, Italy. August 30, 2017
Sub-aspect Extraction
(*) Caputo, A., Basile, P., de Gemmis, M., Lops, P., Semeraro, G., & Rossiello, G. (2017). SABRE: A
Sentiment Aspect-Based Retrieval Engine. In Information Filtering and Retrieval (pp. 63-78).
Another distinguishing aspect: we can extract a hierarchy of terms
Based on Phraseness and Informativeness: They measure the gain
in information if two terms are modeled together
Insight: if phraseness and informativeness are high,
the terms have an high cohesion
SUB(room, food, hotel-reviews) ~ 0 NO
SUB(room, shower, hotel-reviews) > 0 YES
SUB(food, wine, hotel-reviews) > 0 YES
SUB(food, service, hotel-reviews) ~ 0 NO
23. Aspect Extraction and Sentiment Analysis
Cataldo Musto, Marco de Gemmis, Giovanni Semeraro. Pasquale Lops
A Multi-criteria Recommender System Exploiting Aspect-based Sentiment Analysis of Users’ Reviews. RECSYS 2017. Como, Italy. August 30, 2017
Sabre@Work
(*) Real review of the hotel
we actually stay in Como :)
(*)
24. Aspect Extraction and Sentiment Analysis
Cataldo Musto, Marco de Gemmis, Giovanni Semeraro. Pasquale Lops
A Multi-criteria Recommender System Exploiting Aspect-based Sentiment Analysis of Users’ Reviews. RECSYS 2017. Como, Italy. August 30, 2017
< 𝑟, 𝑏𝑟𝑒𝑎𝑘𝑓𝑎𝑠𝑡,∗, 1.5, 0.5 >
Sabre@Work
(*) Real review of the hotel
we actually stay in Como :)
(*)
25. Aspect Extraction and Sentiment Analysis
Cataldo Musto, Marco de Gemmis, Giovanni Semeraro. Pasquale Lops
A Multi-criteria Recommender System Exploiting Aspect-based Sentiment Analysis of Users’ Reviews. RECSYS 2017. Como, Italy. August 30, 2017
< 𝑟, 𝑏𝑟𝑒𝑎𝑘𝑓𝑎𝑠𝑡,∗, 1.5, 0.5 >
No sub-aspects Relevance=KL-divergence score
Sentiment = lexicon-based approach based on AFINN
wordlist (*) or machine-learning based approach based on
CoreNLP (^)
(*) http://neuro.imm.dtu.dk/wiki/AFINN
(^) https://nlp.stanford.edu/sentiment/code.html
Sabre@Work
(*) Real review of the hotel
we actually stay in Como :)
(*)
26. Aspect Extraction and Sentiment Analysis
Cataldo Musto, Marco de Gemmis, Giovanni Semeraro. Pasquale Lops
A Multi-criteria Recommender System Exploiting Aspect-based Sentiment Analysis of Users’ Reviews. RECSYS 2017. Como, Italy. August 30, 2017
< 𝑟, 𝑏𝑟𝑒𝑎𝑘𝑓𝑎𝑠𝑡,∗, 1.5, 0.5 >
< 𝑟, 𝑟𝑜𝑜𝑚, 𝑠ℎ𝑜𝑤𝑒𝑟, 1.2, −0.5 >
Sabre@Work
(*) Real review of the hotel
we actually stay in Como :)
(*)
27. Aspect Extraction and Sentiment Analysis
Cataldo Musto, Marco de Gemmis, Giovanni Semeraro. Pasquale Lops
A Multi-criteria Recommender System Exploiting Aspect-based Sentiment Analysis of Users’ Reviews. RECSYS 2017. Como, Italy. August 30, 2017
< 𝑟, 𝑏𝑟𝑒𝑎𝑘𝑓𝑎𝑠𝑡,∗, 1.5, 0.5 >
< 𝑟, 𝑟𝑜𝑜𝑚, 𝑠ℎ𝑜𝑤𝑒𝑟, 1.2, −0.5 >
< 𝑟, 𝑟𝑜𝑜𝑚,∗, 1.3, 0.2 >
Sabre@Work
(*) Real review of the hotel
we actually stay in Como :)
(*)
28. Aspect Extraction and Sentiment Analysis
Cataldo Musto, Marco de Gemmis, Giovanni Semeraro. Pasquale Lops
A Multi-criteria Recommender System Exploiting Aspect-based Sentiment Analysis of Users’ Reviews. RECSYS 2017. Como, Italy. August 30, 2017
< 𝑟, 𝑏𝑟𝑒𝑎𝑘𝑓𝑎𝑠𝑡,∗, 1.5, 0.5 >
< 𝑟, 𝑟𝑜𝑜𝑚, 𝑠ℎ𝑜𝑤𝑒𝑟, 1.2, −0.5 >
< 𝑟, 𝑟𝑜𝑜𝑚,∗, 1.3, 0.2 >
Sabre@Work
… … . 𝑒𝑡𝑐.
(*) Real review of the hotel
we actually stay in Como :)
(*)
29. Multi-Criteria Data Model
Cataldo Musto, Marco de Gemmis, Giovanni Semeraro. Pasquale Lops
A Multi-criteria Recommender System Exploiting Aspect-based Sentiment Analysis of Users’ Reviews. RECSYS 2017. Como, Italy. August 30, 2017
A multi-criteria
data model is
automatically
filled by
exploiting the
aspects
extracted from
the review and
their sentiment
Finer-Grained
Representation!
30. Providing Recommendations
Cataldo Musto, Marco de Gemmis, Giovanni Semeraro. Pasquale Lops
A Multi-criteria Recommender System Exploiting Aspect-based Sentiment Analysis of Users’ Reviews. RECSYS 2017. Como, Italy. August 30, 2017
Similarity is
calculated through
multi-criteria
Euclidean distance
Recommendations
are provided by
exploiting both
User-to-User and
Item-to-Item
Collaborative
Filtering
Recommendation
31. Framework Recap
Cataldo Musto, Marco de Gemmis, Giovanni Semeraro. Pasquale Lops
A Multi-criteria Recommender System Exploiting Aspect-based Sentiment Analysis of Users’ Reviews. RECSYS 2017. Como, Italy. August 30, 2017
32. Experiments
Which combination of the
parameters led to the best
predictive accuracy?
How does our framework perform
when compared to single-criteria
recommendations and matrix
factorization tecniques?
Cataldo Musto, Marco de Gemmis, Giovanni Semeraro. Pasquale Lops
A Multi-criteria Recommender System Exploiting Aspect-based Sentiment Analysis of Users’ Reviews. RECSYS 2017. Como, Italy. August 30, 2017
33. Datasets
Yelp
45,981 users
11,537 items
229,606 ratings(*)
99.95% sparsity
TripAdvisor
536,952 users
3,945 items
796,958 ratings(*)
99.96% sparsity
Amazon
826,773 users
50,210 items
1,324,759 ratings(*)
99.99% sparsity
(*) Ratings = ratings + reviews
Cataldo Musto, Marco de Gemmis, Giovanni Semeraro. Pasquale Lops
A Multi-criteria Recommender System Exploiting Aspect-based Sentiment Analysis of Users’ Reviews. RECSYS 2017. Como, Italy. August 30, 2017
34. Datasets
Yelp
45,981 users
11,537 items
229,606 ratings(*)
99.95% sparsity
TripAdvisor
536,952 users
3,945 items
796,958 ratings(*)
99.96% sparsity
Amazon
826,773 users
50,210 items
1,324,759 ratings(*)
99.99% sparsity
(*) Ratings = ratings + reviews
Cataldo Musto, Marco de Gemmis, Giovanni Semeraro. Pasquale Lops
A Multi-criteria Recommender System Exploiting Aspect-based Sentiment Analysis of Users’ Reviews. RECSYS 2017. Como, Italy. August 30, 2017
35. Experimental Settings
Review Processing
◦ Stop-Word removed
◦ Entity and Collocations recognized
SABRE parameters
◦ With/without subaspects
◦ #aspects/#subaspects = 10, 50
◦ KL-divergence threshold = 0.1
◦ Only nouns!
Recommendations
◦ Multi-Criteria U2U and I2I
Metric
◦ MAE (calculated with Rival framework)
Cataldo Musto, Marco de Gemmis, Giovanni Semeraro. Pasquale Lops
A Multi-criteria Recommender System Exploiting Aspect-based Sentiment Analysis of Users’ Reviews. RECSYS 2017. Como, Italy. August 30, 2017
36. Baselines
Single-Criteria Recommendations
techniques
◦ User-to-User Collaborative Filtering
◦ Item-to-Item Collaborative Filtering
Static Multi-Criteria Recommendations
◦ Only on TripAdvisor data
Matrix Factorization techniques
◦ SGD (Stochastic Gradient Descent)
◦ ParallelSGD
◦ ALSWR
◦ Implementations available in Mahout
◦ Tuning of parameters
Alessandro Suglia, Claudio Greco, Cataldo Musto, Marco de Gemmis, Pasquale Lops, Giovanni Semeraro.
A Deep Architecture for Content-based Recommendations Exploiting Recurrent Neural Networks. UMAP 2017. Bratislava, Slovakia. July 12, 2017
38. 0,7111
0,7564
0,7269
0,8007
0,65
0,67
0,69
0,71
0,73
0,75
0,77
0,79
0,81
0,83
TripAdvisor
10 neigh. / 10 aspects / sub-aspects 10 neigh. / 10 aspects / no sub-aspects
10 neigh. / 50 aspects / sub-aspects 10 neigh. / 50 aspects / no sub-aspects
Outcomes
Best-results obtained
with 10 aspects
Best-results obtained
by also introducing
sub-aspects
(Amazon had a
different behavior)
Lower MAE!
Results – Multi-Criteria User-to-User CF
Cataldo Musto, Marco de Gemmis, Giovanni Semeraro. Pasquale Lops
A Multi-criteria Recommender System Exploiting Aspect-based Sentiment Analysis of Users’ Reviews. RECSYS 2017. Como, Italy. August 30, 2017
39. 0,7111
0,798
0,8245
0,8429
0,6
0,65
0,7
0,75
0,8
0,85
TripAdvisor
Multi-U2U Static-Multi-U2U Multi-I2I Static-Multi-I2I
Outcomes
TripAdvisors data
included ratings
about six static
aspects (cleanliness,
location, value,
service, sleep quality,
overall)
Our approach based
on unsupervised
aspect extraction also
improved these
results
Results – vs. Static Multi-Criteria RecSys
Cataldo Musto, Marco de Gemmis, Giovanni Semeraro. Pasquale Lops
A Multi-criteria Recommender System Exploiting Aspect-based Sentiment Analysis of Users’ Reviews. RECSYS 2017. Como, Italy. August 30, 2017
40. 0,7111
0,8337
0,745 0,7449
0,9053
0,65
0,7
0,75
0,8
0,85
0,9
0,95
1
TripAdvisor
Multi-U2U Single-U2U Ratings-SGD Parallel-SGD ALSWR
Outcomes
Our approach
overcomes all the
baselines.
Our framework wins
the comparisons to
Single-U2U and
Single-I2I
Also matrix
factorization
techniques got an
higher MAE
Results – Baselines
Cataldo Musto, Marco de Gemmis, Giovanni Semeraro. Pasquale Lops
A Multi-criteria Recommender System Exploiting Aspect-based Sentiment Analysis of Users’ Reviews. RECSYS 2017. Como, Italy. August 30, 2017
41. Recap
Results
☺ Our framework significantly improves all the baselines
☺ Unsupervised Aspect Extraction also overcomes static aspects
Future Work: evaluate data model with more sophisticated algorithms
Cataldo Musto, Marco de Gemmis, Giovanni Semeraro. Pasquale Lops
A Multi-criteria Recommender System Exploiting Aspect-based Sentiment Analysis of Users’ Reviews. RECSYS 2017. Como, Italy. August 30, 2017