This repository provides the source code of our paper: Item-side Ranking Regularized Distillation for Recommender System, accepted in Information Sciences'21.
In the paper, we propose Item-side ranking Regularization (IR) method for ranking distillation in Recommender System. The proposed IR method utilizes item-side ranking information, effectively preventing the student model from being overfitted and enabling the student model to more accurately learn the teacher’s prediction results.
We evaluate the effectiveness of IR with the state-of-the-art ranking distillation method, RRD (CIKM'20).
We provide the leave-one-out evaluation protocol used in the paper. The protocol is as follows:
- For each test user
- randomly sample two positive (observed) items
- each of them is used for test/validation purpose.
- randomly sample 499 negative (unobserved) items
- evaluate how well each method can rank the test item higher than these sampled negative items.
- randomly sample two positive (observed) items
We provide three ranking metrics broadly adopted in the recent papers: HR@N, NDCG@N, MRR@N. The hit ratio simply measures whether the test item is present in the top-N list, which is defined as follows:
where δ is the indicator function, Utest is the set of the test users, pu is the hit ranking position of the test item for the user u. On the other hand, the normalized discounted cumulative gain and the mean reciprocal rank are ranking position-aware metrics that put higher scores to the hits at upper ranks. N@N and M@N are defined as follows:
We also provide the training log and the learning curves of the proposed method. You can find them in /logs folder and the attached jupyter notebook.
Please note that the proposed IR method is for transferring knowledge from model's predictions. Topology Distillation (KDD'21), which is a follow-up study of DE and transfers the latent knowledge, is available in https://github.com/SeongKu-Kang/Topology_Distillation_KDD21