Source-linked AI summary
CORE: Simple and Effective Session-based Recommendation within Consistent Representation Space
Yupeng Hou, Binbin Hu, Zhiqiang Zhang, Wayne Xin Zhao
TL;DR
Session-based recommendation must predict next items from anonymous-session behavior despite inconsistent spaces created by nonlinear session encoders. CORE uses weighted item-embedding combinations for session encoding and robust distance measurement for decoding, and experiments on five public datasets demonstrate effectiveness and efficiency.
Problem
Nonlinear session encoders usually place session embeddings in a different representation space from item embeddings, creating an inconsistent prediction issue in session-based recommendation.
Method
CORE encodes sessions as learned weighted sums of their item embeddings and uses robust distance-measuring techniques while keeping encoding and decoding in a consistent representation space.
Results
CORE outperforms all baselines significantly on 8 of 10 metrics across five public datasets and demonstrates effectiveness and efficiency.
Takeaways & Limitations
Consistent-space encoding and robust distance measurement can improve session-based recommendation and enhance existing models when combined with them.
Takeaways & Limitations
Future work includes studying the expressive ability of the representation-consistent encoder and incorporating side features and useful inductive biases.
Abstract
from arXiv · showhide
Session-based Recommendation (SBR) refers to the task of predicting the next item based on short-term user behaviors within an anonymous session. However, session embedding learned by a non-linear encoder is usually not in the same representation space as item embeddings, resulting in the inconsistent prediction issue while recommending items. To address this issue, we propose a simple and effective framework named CORE, which can unify the representation space for both the encoding and decoding processes. Firstly, we design a representation-consistent encoder that takes the linear combination of input item embeddings as session embedding, guaranteeing that sessions and items are in the same representation space. Besides, we propose a robust distance measuring method to prevent overfitting of embeddings in the consistent representation space. Extensive experiments conducted on five public real-world datasets demonstrate the effectiveness and efficiency of the proposed method. The code is available at: https://github.com/RUCAIBox/CORE.
1 INTRODUCTION
Session-based recommendation predicts next items from short-term behaviors, but nonlinear session encoders can place session and item embeddings in inconsistent spaces. CORE addresses this by learning weighted combinations of item embeddings for session representations while improving distance measurement to reduce overfitting.
- Background: Session-based recommendation targets next-item prediction from short-term behaviors such as clicks within anonymous sessions, commonly using encoder-decoder models and dot-product decoding.Existing encoders include RNNs, Transformers, and GNNs.
- Problem: Nonlinear encoders can produce session embeddings outside the space spanned by item embeddings, making preference representation and distance-based recommendation inconsistent.The issue arises because commonly used neural encoders transform item embeddings before decoding with item-session distances.
- Challenges: Directly summing item embeddings would preserve the shared space but fails to model the sequential nature of session behavior.This motivates a consistent-space encoder that retains neural networks for learning item weights rather than transforming the resulting session embedding.
- Challenges: A unified representation space creates two design challenges: exploiting nonlinear networks while preserving consistency and measuring distances robustly enough to avoid item-embedding overfitting.These challenges arise because item embeddings participate directly in score calculation and optimization.
- Proposed solution: CORE represents sessions as learned weighted combinations of item embeddings, keeping session and item embeddings in a consistent representation space.The weights are learned with a deep nonlinear network, allowing the encoder to incorporate item order and importance.
2 METHODOLOGY
CORE unifies session and item embeddings in one representation space by combining item embeddings with learned weights, then uses robust distance measurement for decoding.
- 2 METHODOLOGY: The framework is presented as the overall CORE architecture for session-based recommendation within a consistent representation space.Figure 2 depicts the complete framework.
- 2 METHODOLOGY: The typical SBR pipeline embeds items, applies a neural session encoder, and decodes the resulting session embedding into next-item probabilities.The decoder outputs a distribution over all items.
- 2.1 Representation-Consistent Encoding: CORE encodes sessions as weighted sums of item embeddings, ensuring session and item representations remain in the same space.The Representation-Consistent Encoder uses deep neural networks to learn item weights while preserving the embedding-space constraint.
- 2.1.1 Learning Weights via Mean Pooling: Mean pooling provides a simple weight-learning variant but ignores item order and importance, whereas Transformer self-attention captures sequential structure through positional encoding.The Transformer variant learns normalized item weights from self-attention representations.
- 2.2 Robust Distance Measuring for Decoding: Because item embeddings directly enter distance calculations after representation unification, CORE replaces the fixed-margin dot-product behavior with more robust distance measurement.The motivating lemma links cross-entropy optimization under dot products to an (N−1)-tuplet loss with fixed margin 2.
3 EXPERIMENTS
Experiments on five real-world datasets evaluate CORE’s accuracy, efficiency, component contributions, and learned session representations. CORE generally outperforms baselines, while its variants and analyses show benefits from consistent representation, robust distance measurement, and efficient pooling.
- Overall comparison: CORE significantly outperforms all baselines on 8 of 10 metrics across five datasets.The evaluation uses top-20 Recall and top-20 Mean Reciprocal Rank on Diginetica, Nowplaying, RetailRocket, Tmall, and Yoochoose.
- Efficiency: CORE-ave is the fastest evaluated model while maintaining competitive performance, whereas CORE-trm matches SASRec’s training time and performs best among the compared baselines.The efficiency comparison measures training time per epoch on an NVIDIA TITAN V GPU using Diginetica.
- Ablation study: The full CORE model ranks above variants without robust distance measurement, without representation-consistent encoding, and SASRec, indicating that both components improve performance.The reported order is CORE > CORE w/o RDM > CORE w/o RCE ≃ SASRec.
- Improving existing methods with RCE & RDM: RCE and RDM consistently improve NARM and SR-GNN in combination, while RDM alone improves performance consistently and RCE helps in most cases.These comparisons are reported in Table 4 for the original methods and their improved variants.
- Visualization of session embeddings: CORE separates session embeddings for different next-item classes more clearly than GRU4Rec and SASRec in the Diginetica t-SNE visualization.The visualization samples sessions associated with 15 ground-truth next items from the test set.
- Parameter tuning: CORE performs best at τ=0.07 on Diginetica and τ=0.08 on RetailRocket, with stable performance around 0.04 ≤ τ ≤ 0.08.Item dropout improves robustness when ρ≤0.3, but performance decreases sharply when ρ>0.3.
4 CONCLUSION
CORE unifies session and item representations across encoding and decoding by combining weighted item embeddings with robust distance measurement. Experiments on five public datasets support its effectiveness and efficiency, while future work targets encoder expressiveness and side features.
- CORE unifies representation spaces across encoding and decoding, addressing inconsistent prediction between session and item embeddings.
- Weighted item sums encode sessions consistently, while robust distance measurements help prevent overfitting of item embeddings.
- Experiments on five public datasets demonstrate CORE’s effectiveness and efficiency and show its techniques can improve existing methods.
- Future work will study the encoder’s expressive ability and incorporate side features and additional inductive biases.