[Paper Review] RepeatNet: A Repeat Aware Neural Recommendation Machine for Session-based Recommendation

Pengjie Ren, Zhumin Chen+4 more|arXiv (Cornell University)|Dec 6, 2018

Recommender Systems and Techniques16 references17 citations

TL;DR

RepeatNet is a novel neural recommendation model that explicitly models repeat consumption in session-based recommendation by introducing a repeat-explore mechanism with dual decoders for reusing past items and discovering new ones. It outperforms state-of-the-art methods on three benchmark datasets, with significant gains in MRR and Recall—especially as repeat ratios and dataset sizes increase.

ABSTRACT

Recurrent neural networks for session-based recommendation have attracted a lot of attention recently because of their promising performance. repeat consumption is a common phenomenon in many recommendation scenarios (e.g., e-commerce, music, and TV program recommendations), where the same item is re-consumed repeatedly over time. However, no previous studies have emphasized repeat consumption with neural networks. An effective neural approach is needed to decide when to perform repeat recommendation. In this paper, we incorporate a repeat-explore mechanism into neural networks and propose a new model, called RepeatNet, with an encoder-decoder structure. RepeatNet integrates a regular neural recommendation approach in the decoder with a new repeat recommendation mechanism that can choose items from a user's history and recommends them at the right time. We report on extensive experiments on three benchmark datasets. RepeatNet outperforms state-of-the-art baselines on all three datasets in terms of MRR and Recall. Furthermore, as the dataset size and the repeat ratio increase, the improvements of RepeatNet over the baselines also increase, which demonstrates its advantage in handling repeat recommendation scenarios.

Motivation & Objective

• Address the lack of explicit modeling of repeat consumption in neural session-based recommendation systems.
• Improve recommendation performance in scenarios where users frequently re-consume the same items (e.g., e-commerce, music, TV).
• Develop a mechanism that dynamically decides when to recommend previously interacted items versus new ones.
• Enable end-to-end learning of mode switching (repeat vs. explore) and item scoring within a unified framework.

Proposed method

• Use an encoder-decoder architecture with a GRU-based encoder to represent session history.
• Introduce a repeat-explore mechanism that learns switch probabilities between two recommendation modes: repeat (recommending past items) and explore (recommending new items).
• Employ two decoders: one for computing recommendation probabilities in the repeat mode (over historical items), and one for the explore mode (over unseen items).
• Combine mode switch probabilities with item scores from both decoders using a probabilistic fusion to produce final recommendation scores.
• Train the model end-to-end using joint optimization of recommendation loss and mode prediction loss.
• Incorporate attention mechanisms to enhance representation learning, with ablation studies to assess their complementarity with the repeat mechanism.

Experimental results

Research questions

• RQ1 Can a neural model effectively learn when to recommend previously consumed items versus new items in session-based recommendation?
• RQ2 How does explicit modeling of repeat consumption impact performance on standard benchmark datasets?
• RQ3 What is the relative contribution of the repeat mechanism versus attention mechanisms in improving recommendation quality?
• RQ4 Does the performance gain of RepeatNet increase with higher repeat ratios and larger dataset sizes?
• RQ5 Is joint learning of mode prediction and item recommendation beneficial, or does it lead to overfitting toward repeat recommendations?

Key findings

• RepeatNet outperforms state-of-the-art baselines on all three benchmark datasets (YOOCHOOSE, DIGINETICA, LASTFM) in both MRR and Recall.
• On YOOCHOOSE, RepeatNet achieves a Recall@20 of 70.71% and MRR@20 of 31.03%, significantly outperforming Improved-GRU4REC.
• The repeat mechanism contributes more to MRR improvement, while attention helps boost Recall, and their combination yields the best performance.
• Performance gains of RepeatNet increase with higher repeat ratios and larger datasets, confirming its effectiveness in repeat-heavy scenarios.
• Joint training of recommendation and mode prediction losses slightly degrades performance, suggesting that the recommendation loss alone is a strong supervisor for mode learning.
• Ablation studies show that RepeatNet with the repeat mechanism achieves 97.41% Recall@20 on repeated sessions in YOOCHOOSE, compared to 93.70% without it, demonstrating the mechanism's impact on repeat-aware recommendations.