Recurrent Neural Networks with Top-k Gains for Session-based Recommendations (1706.03847v3)

Abstract: RNNs have been shown to be excellent models for sequential data and in particular for data that is generated by users in an session-based manner. The use of RNNs provides impressive performance benefits over classical methods in session-based recommendations. In this work we introduce novel ranking loss functions tailored to RNNs in the recommendation setting. The improved performance of these losses over alternatives, along with further tricks and refinements described in this work, allow for an overall improvement of up to 35% in terms of MRR and Recall@20 over previous session-based RNN solutions and up to 53% over classical collaborative filtering approaches. Unlike data augmentation-based improvements, our method does not increase training times significantly. We further demonstrate the performance gain of the RNN over baselines in an online A/B test.

• The paper introduces novel ranking loss functions that improve top-k recommendation quality by targeting gradient vanishing in traditional losses.
• The study details an efficient mini-batch sampling strategy combined with GPU optimization to manage large output spaces without added computational cost.
• Experimental results demonstrate up to 53% improvement in Recall@20 and MRR, outperforming classical collaborative filtering and previous RNN methods.

Recurrent Neural Networks with Top-k Gains for Session-based Recommendations

The paper "Recurrent Neural Networks with Top-k Gains for Session-based Recommendations" addresses the significant challenge of providing accurate recommendations without access to a user's historical data, relying solely on current session interactions. This common problem is encountered in domains including e-commerce, video, and music recommendations. Previous approaches centered on item-based collaborative filtering and content-based methods have generally underperformed compared to sequential models like Recurrent Neural Networks (RNNs).

Hidasi et al. propose several improvements to the RNN-based session-based recommendation models, focusing primarily on novel ranking loss functions tailored for the recommendation tasks. The researchers introduce loss functions that are explicitly designed to handle large output spaces efficiently and ensure that top-k ranking performance improves without significantly increasing training times.

Key Contributions

Ranking Loss Functions

The authors introduce a set of ranking loss functions designed for RNNs, which provide substantial performance improvements:

1. Cross-entropy Loss Adjustment: By stabilizing the traditional cross-entropy loss through the addition of a small constant $\epsilon$ or an alternative direct computation method, they circumvent numerical instability issues.
2. TOP1 and BPR-based Losses: Extending beyond the traditional BPR (Bayesian Personalized Ranking) and TOP1 losses, the paper introduces TOP1-max and BPR-max. These new losses focus on leveraging the maximum of the sampled negative examples rather than averaging all the pairwise losses. This ensures that the most relevant negative samples influence the gradients, addressing the gradient vanishing problem inherent in the original losses.

Efficient Sampling Strategies

The authors propose an advanced sampling strategy to handle RNNs' large output spaces, crucial for scalability:

1. Mini-batch based Sampling: They extend the negative sampling to include additional samples drawn from a distribution defined by $\mathrm{supp}_i^\alpha$, where $\alpha$ balances between uniform and popularity-based sampling. This mixed strategy proves effective, as demonstrated by experiments leveraging thousands of additional samples without introducing significant computational overhead.
2. GPU Optimization: Efficient pre-sampling and caching techniques are employed to maintain computational efficiency, sidestepping frequent GPU-CPU data transfer issues.

Experimental Validation

The improvements are validated across various datasets, including RSC15 (RecSys Challenge 2015 dataset), proprietary datasets from online video services (VIDEO and VIDXL), and a classified ads platform (CLASS). The experiments show significant gains in Recall@20 and Mean Reciprocal Rank (MRR@20), with improvements up to:

• 53% in MRR and Recall@20 over classical collaborative filtering approaches.
• 35% over previous session-based RNN solutions.

Practical Implications

The paper extends RNN applicability in session-based recommendation scenarios by addressing both scalability and top-k ranking effectiveness:

1. Scalability and Efficiency: By introducing advanced sampling and efficient computation strategies, the authors ensure that the proposed method remains scalable, even for large datasets.
2. Improved Recommendation Quality: Substantially better recommendation quality is achieved, which is validated through comprehensive offline experiments and confirmed through an online A/B test on a large-scale video platform. In live settings, improvements in key performance indicators such as watch time, video plays, and user clicks were noted, thus demonstrating strong business implications.

Theoretical and Future Directions

The research sets a foundation for further exploration:

1. Broader Application: The proposed loss functions and sampling techniques are not limited to RNNs and can potentially apply to other machine learning models like matrix factorization and autoencoders.
2. NLP Applications: Given the methodological parallels between recommendation systems and NLP, similar techniques could be employed to enhance NLP tasks such as machine translation and text generation.
3. Combining Approaches: Future work may combine these methodological enhancements with data augmentation strategies to push performance further, as indicated by compared state-of-the-art approaches.

In summary, this paper enhances RNN-based recommendations' effectiveness and scalability by introducing innovative ranking losses and sampling strategies, setting new benchmarks in both offline metrics and real-world applications. These advancements open new avenues for broader deployment and cross-domain applicability, solidifying the role of RNNs in dynamic, session-based recommendation environments.