[Paper Review] Session-based Recommendations with Recurrent Neural Networks

Balázs Hidasi, Alexandros Karatzoglou+2 more|arXiv (Cornell University)|Nov 21, 2015

Recommender Systems and Techniques15 references557 citations

TL;DR

This paper proposes a GRU-based recurrent neural network for session-based recommendation, modeling user behavior sequences to improve recommendation accuracy over traditional item-to-item methods. By using a ranking loss function and session-parallel mini-batches, the model achieves up to 31.49% higher MRR@20 than the best baseline on real-world datasets.

ABSTRACT

We apply recurrent neural networks (RNN) on a new domain, namely recommender systems. Real-life recommender systems often face the problem of having to base recommendations only on short session-based data (e.g. a small sportsware website) instead of long user histories (as in the case of Netflix). In this situation the frequently praised matrix factorization approaches are not accurate. This problem is usually overcome in practice by resorting to item-to-item recommendations, i.e. recommending similar items. We argue that by modeling the whole session, more accurate recommendations can be provided. We therefore propose an RNN-based approach for session-based recommendations. Our approach also considers practical aspects of the task and introduces several modifications to classic RNNs such as a ranking loss function that make it more viable for this specific problem. Experimental results on two data-sets show marked improvements over widely used approaches.

Motivation & Objective

• Address the challenge of session-based recommendation in the absence of long-term user profiles, common in e-commerce and media sites.
• Overcome limitations of matrix factorization and item-to-item similarity methods that ignore sequential behavior and past clicks.
• Develop a deep learning approach tailored for sparse, short sessions with high scalability and practical deployment in mind.
• Improve recommendation accuracy by modeling the full sequence of user interactions using RNNs with custom training objectives.

Proposed method

• Adapt gated recurrent units (GRUs) to model sequential user interactions in session-based recommendation tasks.
• Introduce session-parallel mini-batches to efficiently train on long sequences of short user sessions.
• Apply mini-batch based output sampling to reduce computational cost during training with large item sets.
• Use a ranking-based loss function (TOP1, BPR, or cross-entropy) to optimize for top-N recommendation performance.
• Employ 1-of-N encoding for items and tanh activation in the output layer to stabilize training and improve ranking quality.
• Evaluate model performance using recall@20 and mean reciprocal rank (MRR)@20 on two real-world datasets (RSC15 and VIDEO).

Experimental results

Research questions

• RQ1 Can recurrent neural networks effectively model session-based user behavior when user profiles are unavailable?
• RQ2 How does an RNN-based approach compare to traditional item-to-item similarity baselines in session-based recommendation?
• RQ3 Which loss function (ranking-based vs. cross-entropy) yields better performance and stability in session-based recommendation?
• RQ4 How do architectural choices such as network depth, GRU size, and activation functions affect model performance?
• RQ5 What is the impact of input representation (1-of-N vs. embeddings) and sequence modeling strategy on recommendation accuracy?

Key findings

• The GRU-based model with TOP1 ranking loss achieved a 31.49% improvement in MRR@20 on the RSC15 dataset compared to the best baseline (item-KNN).
• On the VIDEO dataset, the TOP1 loss model achieved a 20.27% gain in Recall@20 and a 15.08% gain in MRR@20 over the baseline.
• Increasing the GRU size from 100 to 1000 units improved performance with pairwise losses (TOP1 and BPR), but cross-entropy became numerically unstable at 1000 units.
• Adding deeper layers or additional feed-forward layers after the GRU did not improve performance, suggesting that a single GRU layer is sufficient for session modeling.
• Using 1-of-N encoding outperformed learned item embeddings, and modeling the full sequence of clicks was more effective than using only the last click.
• The model trained efficiently on GPU in a few hours, making frequent retraining feasible for dynamic recommender systems.