pydata

Keep Looking, Don't Settle

Recommendation System 02

List some papers to read for better understanding and practical use.

  1. RecSys'16, YouTube | Deep Neural Networks for YouTube Recommendations

fig1

Retrival:

  • Use mean pooling (average) for sequential rfeatures embedding
  • Example age is the the time between user click and training. It's used to describe the bias that user are more likely to click the new video. When applying the model online, the value for this feature is 0.
  • When training offline, the user embedding and view embedding will be saved offline. When running online, it will be called by hash.

fig2

Ranking:

  • The model structure is similar to retrival
  • Features: use more features to describe the users, videos and behaviors.
  • Target: if only use CTR, then some videos with sensational title will be recommended. In the training, it uses weighted loss function and the weight is related to video length.

It is the very early paper to apply deep learning in recommendation system rather than using GBDT or logistic regression.

  1. CIKM'20, jd.com | Deep Multifaceted Transformers for Multi-objective Ranking in Large-Scale E-commerce Recommender Systems
  2. AAAI'20, Alibaba | Deep Match to Rank Model for Personalized Click-Through Rate Prediction
  3. AAAI'21, Tencent | U-BERT: Pre-training User Representations for Improved Recommendation
  4. IJCAI'18, Microsoft | Sequential Recommender System based on Hierarchical Attention Network

To be continued.