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

Модель Похожих товаров и её приложения

Модель Похожих товаров и её приложения

Айгуль Газимова (Joom.ru, Junior Machine Learning Developer) @ Moscow Python №78

"Похожие товары – это важная часть для любой e-commerce платформы, а особенно для маркетплейсов. Они помогают покупателям найти лучший товар за меньшие деньги. Joom – международная группа e-commerce и финтех компаний, которая работает как с B2B, так и с B2C сегментами с одной основной командой разработки поиска. Мы поговорим про:
- построение Модели Похожих Товаров для Joom Marketplace (B2C-продукт);
- разработку Модели Похожих Товаров для JoomPro (B2B-продукт), где у нас мало пользовательской истории;
- как с помощью такой модели мы улучшили алгоритмы рекомендаций".

Видео: https://moscowpython.ru/meetup/78/similar-products-search/

Moscow Python Meetup

July 14, 2022
Tweet

More Decks by Moscow Python Meetup

Other Decks in Programming

Transcript

  1. SIMILAR PRODUCTS MODEL
    AND ITS’ APPLICATIONS
    Aigul Gazimova

    View full-size slide

  2. Aliexpress
    SHEIN
    Similar Products Model
    Amazon

    View full-size slide

  3. Joom Search Team
    ML engineers,
    Software engineers,
    Analytics

    View full-size slide

  4. About me
    AIGUL GAZIMOVA
    Bachelor’s degree in applied mathematics and
    computer science
    Cloud storage → eCommerce
    Search@Joom

    View full-size slide

  5. JoomPro
    platform for crossborder wholesale trade
    International group
    of e-commerce and fintech companies
    Joom Marketplace
    platform for shopping
    from all over the world
    Onfy
    pharmaceutical marketplace
    in Germany
    Joompay
    fintech service for daily financial
    transactions in Europe
    Joom Logistics
    business that provides logistics,
    technology and infrastructure services
    for crossborder eCommerce

    View full-size slide

  6. Similar Products Model

    View full-size slide

  7. Similar products: idea of using Word2Vec
    device_id product_id1
    … product_idn

    View full-size slide

  8. .
    .
    .
    device_id1
    product_id1
    … product_idn1
    device_id2
    product_id1
    … product_idn2
    device_idm
    product_id1
    … product_idnm
    Similar products: idea of using Word2Vec

    View full-size slide

  9. Word2Vec: recap

    View full-size slide

  10. Center
    Center embedding and context embedding

    View full-size slide

  11. V(KINGS) - V(KING) ≈ V(QUEENS) - V(QUEEN)
    Interpretation of embeddings

    View full-size slide

  12. Issue: new products

    View full-size slide

  13. Issue: new products

    View full-size slide

  14. Image embeddings

    View full-size slide

  15. Customer experience

    View full-size slide

  16. Customer experience

    View full-size slide

  17. Customer experience

    View full-size slide

  18. Customer experience

    View full-size slide

  19. How to build the Similar
    Products Model if we
    lack of user activity
    history?

    View full-size slide

  20. How to build the Similar Products Model
    if we lack of user activity history?
    - word2vec approach

    View full-size slide

  21. How to build the Similar Products Model
    if we lack of user activity history?
    - word2vec approach

    View full-size slide

  22. How to build the Similar Products Model
    if we lack of user activity history?
    - word2vec approach
    - image-based embeddings

    View full-size slide

  23. How to build the Similar Products Model
    if we lack of user activity history?
    - word2vec approach
    - image-based embeddings

    View full-size slide

  24. How to build the Similar Products Model
    if we lack of user activity history?
    - word2vec approach
    - image-based embeddings
    to use content information for word2vec approach
    and to train the model on joom users’ history

    View full-size slide

  25. category id product name
    from EUR 12
    Content Similar Products Model: product features

    View full-size slide

  26. Optimus Prime!!!

    View full-size slide

  27. category_id product name
    BOW BPE
    category emb name emb
    Transformer encoder
    center embedding context embedding
    product_id category_id
    BOW BOW
    emb categories emb
    +
    Content Similar Products Model: architecture

    View full-size slide

  28. category_id product name
    BOW BPE
    category emb name emb
    Transformer encoder
    center embedding context embedding
    product_id category_id
    BOW BOW
    emb categories emb
    +
    Content Similar Products Model: applications

    View full-size slide

  29. Category embeddings: 2D representation with TSNE

    View full-size slide

  30. Clusterization

    View full-size slide

  31. ● Similar Products Model for B2C
    ● Content Similar Products Model for B2B
    ● Why it’s useful in e-commerce platforms
    Conclusion

    View full-size slide

  32. BTW, we are hiring!

    View full-size slide