Code Implementation of V-A. Tran, et al. Hierarchical Latent Relation Modeling for Collaborative Metric Learning. In: Proceedings of the 15th ACM Conference on Recommender Systems (RecSys 2021), September 2021.

Paper

Paper Summary(Korean)

• Previous recommendation approaches focused heavily on Collaborative Filtering(CF) methods
  • Matrix Factorization (MF)
    • Factorization of the user-item matrix into dense, lower dimensional latent vectors
    • Prediction based on User-Item Similarity (usually dot-product)
• Metric Learning introduced as an alternative to previous approaches.
  • Interested in learning the metrics among data points
• Limitations
  • Over-simplification of the user-item relations
  • Each user&item represented with a single mapped vector
  • Does not incorporate any item-item relations

• CML DL Based approach to learn hierarchical relations of user-item over item-item relations based on the following assumption...

  "there exists a hierarchical structure in different relation types, and that user-item relations are built on top of item-item relations"

• The author proposes an enhanced consideration over the user-item relations with the User Attention & Item Attention Module
• For detailed descriptions of the model implementation & model experiments, refer to the Paper Summary(Korean)

• Train a specific HLRM model from model.py

$ python train.py --train_batch_size --eval_batch_size --num_inter \
    --hlrm_type --epochs --lr --lr_scheduler_gamma --clip --patience \
    --emb_size --num_relations --is_pretrained_embs --freeze_embs \
    --save_model -- save_dir --loss_margin

• Evaluate a specific model checkpoint on the metrics HitRate, MAP, MRR, NDCG, PREC, REC

$ python eval.py --eval_batch_size --num_inter --hlrm_type --save_dir \
    --eval_size --emb_size --topk --num_relations --is_pretrained_embs

• Paper
• Official Source Code