They deal with the problem to predict users’ based on anonymous sessions. The previous approach has not sufficiently addressed the importance of the items of their relevance to the user's main intent.
key ideas
They estimate item importance in a session to predict the next item by employing a modified self-attention mechanism.
et : item embedding(d dimension) at time step t.
I : a set of items
Q(query), K(key) from self-attention mechanism
C : affinity matrix between Q and K
(cf.
) [4]
quick results
They assert that they achieve the best performance in terms of Recall@10 and MRR@20 on YOOCHOOSE and Recall@20 on DIGINETICA.
However, compared with other papers[1, 2, 3] they do not have better performance in terms of MRR@20 on DIGINETICA and other measurements.
What is the function of the denominator of root d in affinity matrix C?
Why do they just concat long-term preference and current interest without any modification?
What do you like?
I like the way to calculate the user’s preference with Importance extraction module by using concept Self-Attention.
What you don't like?
They use preference fusion to combine long-term preference and current interest. However, it would lower the performance on some datasets (datasets that consist of non-duplicated items)
-Since they only short session length (at most 10), they do not consider each user’s long preference fully.
How to improve?
If we make an IEM for each user, could we improve the performance?
In preference fusion, we can add weight for a user’s long-term preference and his/her current interest.
Any new ideas?
We could use the last layer of encoder of Transformer or BERT instead of the average of affinity matrix.
Reproducing results (if any)
References
[1] Anh, Pham Hoang, Ngo Xuan Bach, and Tu Minh Phuong. 2019. “Session-Based Recommendation with Self-Attention.” In Proceedings of the Tenth International Symposium on Information and Communication Technology, 1–8. SoICT 2019. New York, NY, USA: Association for Computing Machinery.
[2] Qiu, Ruihong, Jingjing Li, Zi Huang, and Hongzhi YIn. 2019. “Rethinking the Item Order in Session-Based Recommendation with Graph Neural Networks.” In Proceedings of the 28th ACM International Conference on Information and Knowledge Management, 579–88. CIKM ’19. New York, NY, USA: Association for Computing Machinery.
[3] Chen, T., and R. C. Wong. 2019. “Session-Based Recommendation with Local Invariance.” In 2019 IEEE International Conference on Data Mining (ICDM), 994–99.
[4] Vaswani, Ashish, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Ł. Ukasz Kaiser, and Illia Polosukhin. 2017. “Attention Is All You Need.” In Advances in Neural Information Processing Systems 30, edited by I. Guyon, U. V. Luxburg, S. Bengio, H. Wallach, R. Fergus, S. Vishwanathan, and R. Garnett, 5998–6008. Curran Associates, Inc.
(3rd Aug)
Paper information
title : Rethinking Item Importance in Session-based Recommendation
authors : Zhiqiang Pan, Fei Cai, Yanxiang Ling, Maarten de Rijke
venue : SIGIR 2020(poster)
pdf link : https://dl.acm.org/doi/pdf/10.1145/3397271.3401274 [video:https://sigir-schedule.baai.ac.cn/poster/sp0093]
github : X
Summary
problems to address
They deal with the problem to predict users’ based on anonymous sessions. The previous approach has not sufficiently addressed the importance of the items of their relevance to the user's main intent.
key ideas
They estimate item importance in a session to predict the next item by employing a modified self-attention mechanism. et : item embedding(d dimension) at time step t. I : a set of items Q(query), K(key) from self-attention mechanism C : affinity matrix between Q and K (cf. ) [4]
quick results
They assert that they achieve the best performance in terms of Recall@10 and MRR@20 on YOOCHOOSE and Recall@20 on DIGINETICA.
However, compared with other papers[1, 2, 3] they do not have better performance in terms of MRR@20 on DIGINETICA and other measurements.
MRR@5 GT = 1
We have items 1, 2, 3 , 4, 5,... 10
[ XXXXXXX ] ?
2, 3, 4, 7, 10, 1 = 0 1, 3, 4, 7, 10 = 1 3, 3, 1, 7, 10 = 1/3 3, 3, 1, 7,2 = 1/5
Questions about the paper?
What do you like?
What you don't like?
How to improve?
Any new ideas?
We could use the last layer of encoder of Transformer or BERT instead of the average of affinity matrix.
Reproducing results (if any)
References
[1] Anh, Pham Hoang, Ngo Xuan Bach, and Tu Minh Phuong. 2019. “Session-Based Recommendation with Self-Attention.” In Proceedings of the Tenth International Symposium on Information and Communication Technology, 1–8. SoICT 2019. New York, NY, USA: Association for Computing Machinery.
[2] Qiu, Ruihong, Jingjing Li, Zi Huang, and Hongzhi YIn. 2019. “Rethinking the Item Order in Session-Based Recommendation with Graph Neural Networks.” In Proceedings of the 28th ACM International Conference on Information and Knowledge Management, 579–88. CIKM ’19. New York, NY, USA: Association for Computing Machinery.
[3] Chen, T., and R. C. Wong. 2019. “Session-Based Recommendation with Local Invariance.” In 2019 IEEE International Conference on Data Mining (ICDM), 994–99.
[4] Vaswani, Ashish, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Ł. Ukasz Kaiser, and Illia Polosukhin. 2017. “Attention Is All You Need.” In Advances in Neural Information Processing Systems 30, edited by I. Guyon, U. V. Luxburg, S. Bengio, H. Wallach, R. Fergus, S. Vishwanathan, and R. Garnett, 5998–6008. Curran Associates, Inc.