Upgrade to Pro — share decks privately, control downloads, hide ads and more …

【論文紹介】Latent Factor Models and Aggregation Operators for Collaborative Filtering in Reciprocal Recommender Systems / Latent-factor-models-and-aggregation-operators-for-collaborative-filtering-in-reciprocal-recommender-systems

Yuya Matsumura
October 05, 2019

【論文紹介】Latent Factor Models and Aggregation Operators for Collaborative Filtering in Reciprocal Recommender Systems / Latent-factor-models-and-aggregation-operators-for-collaborative-filtering-in-reciprocal-recommender-systems

2019年7月19日 RecSys2019 論文読み会 (https://connpass.com/event/148001/) での発表資料です。

Latent Factor Models and Aggregation Operators for Collaborative Filtering in Reciprocal Recommender Systems(https://dl.acm.org/citation.cfm?id=3347026) という論文について紹介致しました。

Yuya Matsumura

October 05, 2019
Tweet

More Decks by Yuya Matsumura

Other Decks in Science

Transcript

  1. Yuya Matsumura @yu-ya4
    ʲ࿦จ঺հʳLatent Factor Models and Aggregation
    Operators for Collaborative Filtering in Reciprocal
    Recommender Systems
    RecSys2019 ࿦จಡΈձ
    5 Oct. 2019
    J Neve, I Palomares.2019. Proceedings of the 13th ACM Conference on Recommender Systems, 219-227.

    View full-size slide

  2. ✓ Yuya Matsumura
    ✓ Wantedly, Inc. Recommendation Team
    ✓ Data Scientist, Team Lead
    ✓ Interested in Information Retrieval, Machine Learning
    Self-Introduction
    @yu-ya4
    @yu__ya4

    View full-size slide

  3. ✓ User-to-User recommendation (not User-to-Item recommendation)
    ✓ Aim to create good match between two users
    ✓ Relatively unexplored area
    ✓ Little progression past kNN implementations
    What is Reciprocal Recommender System (RRSs) ?

    View full-size slide

  4. R LuizPizzato,TomekRej,ThomasChung,IrenaKoprinska,andJudyKay.2010. RECON: a reciprocal recommender for online dating.
    Proceedings of the fourth ACM conference on Recommender systems P. 207-214.
    Definition of RRSs

    View full-size slide

  5. Overview of RRSs
    User A -> User B
    Preference Score
    User B -> User A
    Preference Score
    Preference Score
    Aggregation
    Bidirectional
    Preference Relation

    View full-size slide

  6. ✓ RECON, a content-based filtering algorithm for reciprocal recommendation
    ‣ The first work to distinguish RSSs and define its properties.
    ‣ Extract implicit preferences by looking at attributes in common amongst those whom a given user has
    sent messages to. (e.g. Body Shape, Personality, Education…)
    ‣ However, physical appearances is often the main factor taken into consideration…
    ✓ Reciprocal Collaborative Filtering (RCF), a collaborative filtering based algorithm
    ‣ Use a nearest-neighbor collaborative filtering
    ‣ Significantly improve on RECON’s results
    ‣ SoTA!!
    ‣ However, limitation in computational complexity of calculating similarities between all pairs of users
    Related Works
    Peng Xia, Benyuan Liu, Yizhou Sun, and Cindy Chen. 2015. Reciprocal Rec- ommendation System for Online Dating.
    Proceedings of the 2015 IEEE/ACM International Conference on Advances in Social Networks Analysis and Mining P. 234-241. 


    View full-size slide

  7. 1. Reciprocal recommendation based on latent factor models
    (often used in conventional RSs)
    2. Efficient recommendation for large datasets (realistic used in
    real time systems)
    3. Exploring different aggregation functions
    Contributions

    View full-size slide

  8. Latent Factor Model for Reciprocal Recommendation (LFRR)
    ✓ Dataset has over 10 million users, and most users will only see a small
    fraction of other users, and give an opinion on an even smaller fraction.
    ✓ Imputing unknown data was likely to introduce a great deal of inaccuracy
    into the model
    ✓ Using a learning algorithm to generate latent factors using only the known
    data was likely to produce much more accurate results.
    ➡ Use Stochastic Gradient Descent (SGD) , which tends to give better results
    with sparse, explicit data
    ➡ Not other common methods( such as Alternating Least Squares (ALS),
    which tends to give better results with dense, implicit data.)

    View full-size slide

  9. System Architecture
    Latent Factor Model
    for Reciprocal Recommendation (LFRR)

    View full-size slide

  10. Preference Aggregation
    1. Arithmetic Mean
    2. Geometric Mean
    3. Harmonic Mean
    4. Cross-Ratio Uninorm
    (1)
    (2)
    (3)
    (4)

    View full-size slide

  11. Preference Aggregation
    1. Arithmetic Mean
    2. Geometric Mean
    3. Harmonic Mean
    4. Cross-Ratio Uninorm
    (1)
    (2)
    (3)
    (4)
    Previous works used only (3)

    View full-size slide

  12. Preference Aggregation
    1. Arithmetic Mean
    2. Geometric Mean
    3. Harmonic Mean
    4. Cross-Ratio Uninorm
    (1)
    (2)
    (3)
    (4)
    (i) two low values are aggregated to produce a lower value (conjunctive behaviour)
    (ii) two high values are aggregated to produce a higher value (disjunctive behaviour)
    (iii) a high and a low value are aggregated to a value that lies in between both (averaging behaviour)

    View full-size slide

  13. Evaluation
    ✓ Dataset: Pairs (10M users, over 8 years, 9M Likes per week)
    ✓ Preference: Like
    • a binary value
    • a little effort to send
    ✓ Limitations
    1. Users who live in Tokyo and the surrounding areas. These users represent a significant majority
    of Pairs users.
    2. Users between 18 and 30 years of age, for the same reason as above. Users outside this age
    range are outliers in the user base.
    3. Users who have sent at least 10 Likes.
    ✓ Metrics
    • Effectiveness Evaluation: ROC curve, F1-value
    • Efficiency Evaluation

    View full-size slide

  14. Results (ROC Curves)

    View full-size slide

  15. Results (F1 Scores)

    View full-size slide

  16. Results (Efficiency)

    View full-size slide

  17. ✓ Latent factor models produce similar precision and recall to kNN ɹ
    while reducing execution time (So a more realistic algorithm)
    ✓ Aggregation functions significantly impact the results of RRSs.
    Conclusions

    View full-size slide