Source-linked AI summary
Improved Recurrent Neural Networks for Session-based Recommendations
Yong Kiam Tan, Xinxing Xu, Yong Liu
TL;DR
Session-based recommendation requires accurate, responsive predictions despite changing user behavior and large item catalogs. The paper improves RNN recommenders with data augmentation, temporal-shift handling, distillation, and direct item-embedding prediction, reporting 12.8% and 14.8% relative gains on Recall@20 and MRR@20 over previously reported results.
Problem
Session-based recommendation must remain accurate and responsive while handling changing user behavior and large candidate-item output layers.
Method
The paper studies data augmentation, temporal-shift handling, privileged-information distillation, and a model that directly predicts item embeddings.
Results
12.8% and 14.8% relative improvements were achieved over previously reported results on Recall@20 and Mean Reciprocal Rank@20, respectively, on the RecSys Challenge 2015 dataset.
Takeaways & Limitations
Direct item-embedding prediction reduces prediction time to about 60% of classification-based models and uses fewer parameters, supporting deployment of RNNs in real recommender systems.
Takeaways & Limitations
The embedding-based model performs poorly in predictive accuracy relative to the other models, and its accuracy may depend on better item embeddings or additional item information.
Abstract
from arXiv · showhide
Recurrent neural networks (RNNs) were recently proposed for the session-based recommendation task. The models showed promising improvements over traditional recommendation approaches. In this work, we further study RNN-based models for session-based recommendations. We propose the application of two techniques to improve model performance, namely, data augmentation, and a method to account for shifts in the input data distribution. We also empirically study the use of generalised distillation, and a novel alternative model that directly predicts item embeddings. Experiments on the RecSys Challenge 2015 dataset demonstrate relative improvements of 12.8% and 14.8% over previously reported results on the Recall@20 and Mean Reciprocal Rank@20 metrics respectively.
CCS Concepts
The paper concerns supervised learning, neural networks, and recommender systems.
- The paper is classified under supervised learning and neural networks.
- Recommender systems are a central application area.
- The listed computing and information-systems concepts frame the paper's research domain.
Keywords
The paper focuses on recurrent neural networks for session-based recommendations.
- Recurrent neural networks are a central topic.
- The paper addresses recommender systems.
- The specific task is session-based recommendation.
1. INTRODUCTION
The paper studies improved RNNs for session-based recommendation, targeting cold-start settings, temporal shifts, training improvements, and deployable prediction. On the RecSys Challenge 2015 dataset, data augmentation improved Recall@20 by 12.8% and MRR@20 by 14.8% over previously reported results.
- Motivation: Session-based recommendations use current-session behavior, avoiding the need for an identified user profile and addressing cold-start settings.
- Contributions: The paper extends RNNs with data augmentation, temporal-shift handling, privileged-information distillation, and direct item-embedding prediction.
- Results: 12.8% and 14.8% relative improvements were reported over previous results on Recall@20 and Mean Reciprocal Rank@20, respectively.
- Evaluation: The techniques were evaluated on the RecSys Challenge 2015 dataset.
- Organization: The work was organized around improved RNN models and experiments evaluating those models.
2. RELATED WORK
The related work covers traditional recommender methods, prior RNN-based session recommendation, deep-learning improvements, and privileged-information distillation.
- Matrix factorization and neighborhood-based methods are established recommender-system approaches based on user-item interactions.
- Prior work applied RNNs to session-based recommendation and reported 20% to 30% better performance than existing methods.
- Common deep-learning improvements include data augmentation, dropout, batch normalization, and residual connections.
- LUPI uses additional representations available during training but not testing, while generalized distillation trains a student from teacher-provided soft labels.
3. PROPOSED APPROACHES
The proposed approaches augment session sequences, adapt training to temporal shifts, use privileged future information, and predict item embeddings to reduce output-layer cost.
- 3.1 RNNs for session-based recommendations: RNNs process each click prefix separately to predict the next item, using GRU recurrent layers and trainable input embeddings.
- 3.2 Data augmentation: Sequence preprocessing turns every session prefix into a separate training sequence with its corresponding next-click label.
- 3.2 Data augmentation: Embedding dropout randomly deletes clicks during training, generating shorter sequences and reducing sensitivity to noisy clicks and overfitting.
- 3.3 Adapting to temporal changes: Temporal shifts motivate pre-training on the full dataset, followed by training initialized from that model on a more recent data subset.
- 3.4 Use of privileged information: Privileged future sequences are reversed and used to train a teacher whose soft labels jointly guide the student with real labels through a tradeoff parameter λ.
- 3.5 Output embeddings for faster predictions: Directly predicting item embeddings changes the final-layer parameter count from H ∗N to H ∗D, reducing prediction time and space requirements but requiring good item embeddings.
4. EXPERIMENTS
Experiments evaluate several RNN extensions on the RecSys Challenge 2015 dataset, including preprocessing, temporal adaptation, privileged-information distillation, and direct item-embedding prediction. M1 and M2 improve performance, while M4 reduces prediction cost and M3 is computationally expensive.
- Experimental setup: The experiments use 7,966,257 training sessions, 15,234 test sessions, 37,483 candidate items, and 23,670,981 training sequences.Sessions from the last day form the test set; earlier sessions form the training set.
- Experimental setup: Recall@20 and MRR@20 evaluate next-item rankings, using top-20 softmax predictions for M1–M3 and cosine-nearest item embeddings for M4.Model size and batch prediction time are also reported.
- Compared models: M1 combines softmax outputs, sequence preprocessing, embedding dropout, and a fully connected recurrent-to-output layer; M2 additionally retunes on recent data fractions.M3 uses future sequences as privileged information to provide soft labels for another M1 model.
- Compared models: M4 directly predicts item embeddings through an added hidden layer and retrieves recommendations using cosine distance.The item embeddings used for these models are trained on the full training dataset in M1.
- Results: M1 and M2 yielded strong gains over reported baseline RNN models, while using the entire dataset was slightly worse than training on more recent fractions.The results indicate that recommendation models need to account for changing user behaviour over time.
- Results: M1–M3 did not significantly improve when GRU size increased from 100 to 1000.Figure 4 plots both evaluation metrics against logarithmic dataset fraction; M2 lacks a full-dataset point and larger-GRU M3 results were omitted.
- Results: M3 required extremely long training because computing soft labels and their cross-entropy loss scales poorly with many possible labels.M3 results with GRU size 1000 were omitted, although M3 produced modest gains over M1 on the smallest datasets.
- Results: M4 had poorer predictive accuracy than the other proposed models but improved over the baseline; it used about 60% of their prediction time and fewer parameters.The authors identify richer item information, such as category or brand, as a possible route to improving M4 accuracy.
5. CONCLUSION
The paper empirically evaluates extensions to basic RNNs for session-based recommendation. It reports benefits from data augmentation and temporal-shift handling, while identifying embedding-model tradeoffs and item features as future work.
- The paper presents and empirically evaluates several extensions to a basic RNN model for session-based recommender systems.
- Data augmentation and accounting for temporal shifts in user behaviour enhance recurrent-model performance.
- Future work includes studying tradeoffs of the embedding-based model and incorporating known item features.