Upload
damien-dudognon
View
209
Download
0
Embed Size (px)
DESCRIPTION
Conference Centric 2011, Barcelona, Spain Full paper available at : http://www.thinkmind.org/index.php?view=article&articleid=centric_2011_2_30_30049 Abstract : Recommender systems aim at automatically providing objects related to user’s interests. The angular stone of such systems is a way to identify documents to be recommended. Indeed, the quality of these systems depends on the accuracy of its recommendation selection method. Thus, the selection method should be carefully chosen in order to improve end-user satisfaction. In this paper, we first compare two sets of approaches from the literature to underline that their results are significantly different. We also provide the conclusion of a survey done by thirty four students showing that diversity is considered as important in recommendation lists. Finally, we show that combining existing recommendation selection methods is a good mean to obtain diversity in recommendation lists.
Citation preview
Institut de Recherche en Informatique de Toulouse (IRIT) - UMR 5505
Bridging the gap between users and systems
27/10/11
Laurent CANDILLIER – Max CHEVALIER – Damien DUDOGNON – Josiane MOTHE
Diversity in recommender systems How to recommend documents for a visited one
Maximizing the chances of retrieving at least one relevant document per user [Santos et al., 2010]
Cover a large range of users’ interests
Context
Blog platform
Unknown user => no profile
Diversity of users, diversity of their expectations
27/10/11 2 Candillier L. – Chevalier M. – Dudognon D. – Mothe M.
Diversity in recommender systems How to recommend documents for a visited one
Maximizing the chances of retrieving at least one relevant document per user [Santos et al., 2010]
Cover a large range of users’ interests
Context
Blog platform
Unknown user => no profile
Diversity of users, diversity of their expectations
=> Diversify the recommendations
27/10/11 3 Candillier L. – Chevalier M. – Dudognon D. – Mothe M.
What is diversity? Definitions from the literature
Topicality
Related to a particular topic [Xu and Chen, 2006]
Diversity
Topical diversity
Extrinsic: solve ambiguity [Radlinski et al., 2009]
Intrinsic: avoid redundancy [Clarke et al., 2008]
Serendipity
Attractive and surprising documents [Herlocker et al., 2004]
27/10/11 4 Candillier L. – Chevalier M. – Dudognon D. – Mothe M.
Approaches to diversify IR results Topical diversity
Clustering
Identify aspects
Reorder depending on the aspects covered
Examples
K-Means [Bi et al., 2009]
Hierarchical Clustering [Meij et al., 2010]
27/10/11 5 Candillier L. – Chevalier M. – Dudognon D. – Mothe M.
Approaches to diversify IR results Topical diversity
Sliding Windows
Reorder the retrieved documents
Select documents using metrics
Similarity with the visited document
Similarity with the current recommended document list
Examples
MMR [Carbonell and Goldstein, 1998]
Intra-list similarity [Ziegler et al., 2005]
27/10/11 6 Candillier L. – Chevalier M. – Dudognon D. – Mothe M.
Approaches to diversify IR results Serendipity
Alternative to topical diversity
Similarity not only based on the content
Examples
Organizational similarity [Cabanac et al., 2007]
Temporal diversity [Lathia et al., 2010]
27/10/11 7 Candillier L. – Chevalier M. – Dudognon D. – Mothe M.
Analysis of the TREC Web 2009 results
Hypothesis
Diversity of approaches
No one approach for all users’ needs
Approaches are complementary
Valuable to combine them
Goals
Analyse results obtained with approaches having
Same goal
Similar performances
=> To identify if diversity exists
27/10/11 8 Candillier L. – Chevalier M. – Dudognon D. – Mothe M.
Analysis of the TREC Web 2009 results Experimental framework
Reference IR corpus (TREC Web 2009)
Two IR contexts
Adhoc task
Diversity task
Compare results (runs) of the 4 best approaches of each task
Similar performances according to IR metrics
MAP for adhoc task
NDCG for diversity task
Overlap for each pair of runs underlying diversity
27/10/11 9 Candillier L. – Chevalier M. – Dudognon D. – Mothe M.
Analysis of the TREC Web 2009 results
Adhoc Task
Top 10 documents
Overlap: 22.4%
Precision: 0.384
Overlap max < 30%
27/10/11 10 Candillier L. – Chevalier M. – Dudognon D. – Mothe M.
Analysis of the TREC Web 2009 results
Diversity Task
Top 10 documents
Overlap: 6.3%
Overlap max < 15%
27/10/11 11 Candillier L. – Chevalier M. – Dudognon D. – Mothe M.
Analysis of the TREC Web 2009 results Conclusions
Two distinct approaches are unlikely to return the same (relevant) documents Low average overlap
Diversity of approaches No approach significantly better than others
A combination can be valuable
TREC tasks focused on topicality and topical diversity Can’t be used to evaluate other types of diversity
Users’ study necessary [Hayes et al., 2002]
27/10/11 12 Candillier L. – Chevalier M. – Dudognon D. – Mothe M.
Users’ Study Our intuitions
Most of the time, users want topicality
Get focused information
Sometime, they want diversity
Topical diversity
Enlarge the subject
Serendipity
Discover new information
27/10/11 13 Candillier L. – Chevalier M. – Dudognon D. – Mothe M.
Users’ Study Goals
Verify our intuitions
Prove that diversified recommendations answer a larger range of users’ needs
Context of experimentation
34 students in M. Sc. (Management field)
Blog post recommendations
27/10/11 14 Candillier L. – Chevalier M. – Dudognon D. – Mothe M.
Users’ Study Experimental Framework
Select a document
27/10/11 15 Candillier L. – Chevalier M. – Dudognon D. – Mothe M.
Users’ Study Experimental Framework
Read the selected document
27/10/11 16 Candillier L. – Chevalier M. – Dudognon D. – Mothe M.
Users’ Study Experimental Framework
Compute the recommendation lists
27/10/11 17 Candillier L. – Chevalier M. – Dudognon D. – Mothe M.
Approach 1
Approach 2
Approach 3
Approach 4
Approach 5
List 1 (random)
List 2 (fused)
Users’ Study Experimental Framework
Compute the recommendation lists
27/10/11 18 Candillier L. – Chevalier M. – Dudognon D. – Mothe M.
List 1 (random)
List 2 (fused)
Approach 1
Approach 2
Approach 3
Approach 4
Approach 5
Users’ Study Experimental Framework
Compute the recommendation lists
27/10/11 19 Candillier L. – Chevalier M. – Dudognon D. – Mothe M.
List 1 (random)
List 2 (fused)
Approach 1
Approach 2
Approach 3
Approach 4
Approach 5
Users’ Study Experimental Framework
Compute the recommendation lists
27/10/11 20 Candillier L. – Chevalier M. – Dudognon D. – Mothe M.
List 1 (random)
List 2 (fused)
Approach 1
Approach 2
Approach 3
Approach 4
Approach 5
Users’ Study Experimental Framework
Compute the recommendation lists
27/10/11 21 Candillier L. – Chevalier M. – Dudognon D. – Mothe M.
List 1 (random)
List 2 (fused)
Approach 1
Approach 2
Approach 3
Approach 4
Approach 5
Users’ Study Experimental Framework
Present recommendation lists for assessment
27/10/11 22 Candillier L. – Chevalier M. – Dudognon D. – Mothe M.
Which list best meets your needs?
Users’ Study Experimental Framework
Present recommendation lists for assessment
27/10/11 23 Candillier L. – Chevalier M. – Dudognon D. – Mothe M.
Which list is the most diversified?
Users’ Study Experimental Framework
Assessment of all documents
27/10/11 24 Candillier L. – Chevalier M. – Dudognon D. – Mothe M.
Topicality
Users’ Study Approaches used
searchsim Vector-space model Document title as query
mlt Apache Solr MoreLikeThis module Document content as query
27/10/11 25 Candillier L. – Chevalier M. – Dudognon D. – Mothe M.
Topical diversity
Users’ Study Approaches used
kmeans K-means classification One element per cluster
27/10/11 26 Candillier L. – Chevalier M. – Dudognon D. – Mothe M.
Serendipity
Users’ Study Approaches used
blogart Random selection from the same blog
topcateg Popular documents in the same category
27/10/11 27 Candillier L. – Chevalier M. – Dudognon D. – Mothe M.
Users’ Study Approaches used
Same analysis than TREC experiments
Same results
Overlap is low (< 10%)
=> High diversity
27/10/11 28 Candillier L. – Chevalier M. – Dudognon D. – Mothe M.
Users’ Study Results
Distribution of relevant documents
35% 0%
65%
blogart fused
52.5% 26.2%
21.3%
kmeans fused
54.7% 12.5%
32.8%
mlt fused
52.4% 8.7%
38.9%
searchsim fused
8.8% 0%
91.2%
topcateg fused
27/10/11 29 Candillier L. – Chevalier M. – Dudognon D. – Mothe M.
Users’ Study Results
Distribution of relevant documents
35% 0%
65% 52.5% 26.2%
21.3%
kmeans fused
54.7% 12.5%
32.8%
mlt fused
52.4% 8.7%
38.9%
searchsim fused
8.8% 0%
91.2%
27/10/11 30 Candillier L. – Chevalier M. – Dudognon D. – Mothe M.
Users’ Study Results
Distribution of relevant documents
35% 0%
65%
blogart fused
52.5% 26.2%
21.3%
54.7% 12.5%
32.8%
52.4% 8.7%
38.9% 8.8% 0%
91.2%
topcateg fused
27/10/11 31 Candillier L. – Chevalier M. – Dudognon D. – Mothe M.
Users’ Study Results
Distribution of relevant documents
Relevant mainly retrieved by topical approaches
But at least 20% are retrieved only by fused
Fused matches with a larger range of needs
27/10/11 32 Candillier L. – Chevalier M. – Dudognon D. – Mothe M.
Conclusions and future work Conclusions
Diversity of users’ expectations
No one approach to rule them all
A diversity of approaches
Complementary
Fused
Diversity helps RS to fit more users’ needs
27/10/11 33 Candillier L. – Chevalier M. – Dudognon D. – Mothe M.
Conclusions and future work Future work
Real scale experiment
OverBlog platform
Renew the user survey
More users (international call for participation)
Avoid revealed biases
e.g. More detailed form
=> Deeper analysis
27/10/11 34 Candillier L. – Chevalier M. – Dudognon D. – Mothe M.
Conclusions and future work Future work
Improve the model
Refining the fusing process
Adding a learning process to weight each approach
For every visited document
Find the proportion of documents coming from each approach (log analysis)
Better match with the real users’ needs
27/10/11 35 Candillier L. – Chevalier M. – Dudognon D. – Mothe M.
27/10/11 Candillier L. – Chevalier M. – Dudognon D. – Mothe M. 36
Questions ?
Thank you for your attention
W. Bi, X. Yu, Y. Liu, F. Guan, Z. Peng, H. Xu, and X. Cheng, “ICTNET at Web Track 2009 diversity task”, Text REtrieval Conf., 2009 G. Cabanac, M. Chevalier, C. Chrisment, and C. Julien, “An Original Usage-based Metrics for Building a Unified View of Corporate Documents”, Inter. Conf. on Database and Expert Systems Applications, 2007, LNCS V. 4653, 2007, pp. 202–212 J. Carbonell and J. Goldstein, “The use of MMR, diversity-based reranking for reordering documents and producing summaries”, ACM Conf. on Research and Development in Information Retrieval, 1998, pp. 335-336 C. L. A. Clarke, M. Kolla, G. V. Cormack, O. Vechtomova, A. Ashkan, S. Buttcher, and I.n MacKinnon, “Novelty and Diversity in Information Retrieval Evaluation”, ACM Conf. on Research and Development in Information Retrieval, 2008, pp. 659-666 C. Hayes, P. Massa, P. Avesani, and P. Cunningham, « An online evaluation framework for recommender systems», Workshop on Personalization and Recommendation in E-Commerce, 2002 J. L. Herlocker, J. A. Konstan, L. G. Terveen, and J. T. Riedl, “Evaluating Collaborative Filtering Recommender Systems”, ACM Trans. Information Systems, 22(1), 2004, pp. 5-53 N. Lathia, S. Hailes, L. Capra, and X. Amatriain, “Temporal diversity in recommender systems”, ACM Conf. on Research and Development in Information Retrieval, 2010, pp. 210-217 E. Meij, J. He, W. Weerkamp, and M. de Rijke, “Topical Diversity and Relevance Feedback”, Text REtrieval Conf., 2010 F. Radlinski, P. N. Bennett, B. Carterette, and T. Joachims. “Redundancy, diversity and interdependent document relevance”, SIGIR Forum, 43(2), 2009, pp. 46–52 R. L. T. Santos, C. Macdonald, and I. Ounis, “Selectively Diversifying Web Search Results”, ACM Inter. Conf. on Information and Knowledge Management, 2010 Y. C. Xu and Z. Chen, “Relevance judgment: What do information users consider beyond topicality”, Journal of the American Society for Information Science and Technology, 57(7), 2006, pp. 961–973 C. Ziegler, S. McNee, J. A. Konstan, and G. Lausen, “Improving recommendation lists through topic diversification”, Inter. Conf. on World Wide Web, 2005, pp. 22–32
References
27/10/11 Candillier L. – Chevalier M. – Dudognon D. – Mothe M. 37