Source-linked AI summary
Session-based Recommendations with Recurrent Neural Networks
Balázs Hidasi, Alexandros Karatzoglou, Linas Baltrunas, Domonkos Tikk
TL;DR
Session-based recommendation must operate with short sessions when reliable long-term user histories are unavailable, limiting the suitability of profile-based methods. The paper applies GRU-based RNNs to model session sequences, adapting training and ranking methods for the recommender setting. The approach significantly outperforms popular baselines, while future work remains needed on item representations built from content.
Problem
Session-based recommendation lacks reliable long-term user histories or profiles, creating a practically important setting where conventional profile-based approaches are limited.
Method
The paper uses a GRU-based RNN whose input represents the session state and adapts training with session-parallel mini-batches, output sampling, and a ranking loss.
Results
The GRU-based approach significantly outperforms popular baselines used for session-based recommendation on both datasets and evaluation metrics.
Takeaways & Limitations
Modeling the whole session with an adapted recurrent network can provide a basis for deep-learning approaches to session-based recommendations.
Takeaways & Limitations
Future work will examine item representations automatically extracted from item content instead of the current input.
Abstract
from arXiv · showhide
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.
1 INTRODUCTION
Session-based recommendation addresses settings where reliable long-term user identities or profiles are unavailable, so systems often rely on simple methods that overlook earlier clicks. The paper motivates modeling the full click sequence with RNNs, adapting them to sparse recommendation data and large item sets.
- Small retailers and media sites often lack reliable long-term user identifiers, while privacy concerns and short user histories limit profile-based recommendation.
- Session-based systems therefore commonly use item similarity, co-occurrence, or transition probabilities instead of user-profile methods.
- These practical methods often consider only the last click, ignoring information from earlier clicks in the session.
- RNNs offer a way to model sequential recommendation, with each output depending on the session's previous clicks.
- The recommendation setting requires handling sparse click-stream data and item sets ranging from tens of thousands to hundreds of thousands.
2 RELATED WORK
Related work includes item-to-item and neighborhood methods, Markov decision processes, session-aware factorization, and neural collaborative filtering. These approaches differ in how they represent sessions, model order, and handle sequential recommendation.
- Item-to-item recommendation addresses missing user profiles by precomputing item similarities from available interactions.
- Markov decision processes model recommendations as sequential stochastic decisions, but representing all selection sequences can make the state space unmanageable.
- The extended General Factorization Framework represents a session by averaging event features but does not consider event order.
- Neural collaborative-filtering work has used restricted Boltzmann machines for user-item interactions and deep networks for content-derived features.
3 RECOMMENDATIONS WITH RNNS
The paper adapts GRU-based recurrent neural networks for session-based recommendation, modifying the architecture and training procedure to handle variable-length sessions, large item sets, and ranking-oriented prediction.
- Recurrent neural networks: RNNs maintain an internal hidden state to model variable-length sequences and output a probability distribution over the next sequence element.The standard update combines the current input with the previous hidden state.
- GRU model: GRUs regulate hidden-state updates with gates, including an update gate, candidate activation, and reset gate, to address vanishing gradients.The GRU activation linearly interpolates between the previous and candidate activations.
- GRU model: The recommendation model uses the current session state as input and predicts the next session item, with 1-of-N encoding outperforming the weighted session-history representation and an added embedding layer.The session state may be the current event or the session so far; earlier events can be discounted and the resulting vector normalized.
- Session-parallel mini-batches: Session-parallel mini-batches preserve session evolution by aligning events from multiple sessions and resetting hidden states when sessions end.New sessions replace completed ones in the batch, while sessions are treated as independent.
- Output sampling: Output sampling avoids scoring every item at each step by evaluating a small subset, using other mini-batch items as popularity-based negative examples.The desired next item is trained to rank above sampled negatives, reducing computation and simplifying implementation.
- Ranking loss: Pairwise ranking losses fit the recommendation objective better than pointwise losses, with BPR and the task-specific TOP1 loss included in the solution.The authors report that pointwise ranking was unstable, whereas pairwise ranking losses performed well.
4 EXPERIMENTS
The experiments evaluate the proposed recurrent network on two datasets against common session-based recommendation baselines. Results show that GRU models, especially with pairwise ranking losses, substantially outperform item-KNN across the reported metrics.
- Experimental setup: The evaluation uses two datasets and ranks the next session event after presenting events one by one.The GRU hidden state is reset when each session finishes.
- Evaluation metrics: Recall@20 measures whether the desired item appears among the top 20, while MRR@20 additionally accounts for its rank.Both metrics are motivated by practical recommendation lists containing only a few visible items.
- Baselines: The baselines include global popularity, session popularity, item-KNN, and BPR-MF, with item-KNN clearly dominating the other baseline methods.The baseline results are reported for Recall@20 and MRR@20.
- Parameter and structure optimization: Pairwise ranking losses performed best, whereas pointwise losses were often numerically unstable; BPR and TOP1 were the strongest tested pairwise losses.Cross-entropy produced only 10 and 6 numerically stable networks out of 100 runs for RSC15 and VIDEO, respectively.
- Parameter and structure optimization: A single GRU layer performed best among the examined architectures, while adding layers worsened training loss, recall, and MRR.The authors associate this result with the generally short lifespan of sessions, but state that the exact reason remains unknown.
- Results: ∼20–30% accuracy gain over the best baseline was achieved by TOP1 on the two datasets, while GRU models improved over item-KNN on both metrics.Increasing hidden units further improved pairwise-loss results but reduced cross-entropy accuracy, with higher training time.
5 CONCLUSION & FUTURE WORK
The paper applies GRU-based recurrent neural networks to session-based recommendation and adapts them with task-specific training procedures. It reports significant improvements over popular baselines and identifies further examination and content-based item representations as future work.
- Conclusion: The method adapts GRUs for session-based recommendation using session-parallel mini-batches, mini-batch output sampling, and a ranking loss function.These modifications target the practical characteristics of session-based recommendation.
- Conclusion: The proposed method significantly outperforms popular baselines used for session-based recommendation.The conclusion presents the work as a basis for deep learning applications in recommender systems and session-based recommendations.
- Future work: Future work will examine the proposed network more thoroughly and train it on item representations automatically extracted from thumbnails, video, or text.The current input is contrasted with these prospective content-based representations.