Source-linked AI summary
Self-Supervised Graph Co-Training for Session-based Recommendation
Xin Xia, Hongzhi Yin, Junliang Yu, Yingxia Shao, Lizhen Cui
TL;DR
Session-based recommendation faces severe sparsity, while dropout-based self-supervision can make sessions even sparser and yield unusable signals. COTREC instead co-trains graph encoders over complementary session views, and experiments show broad benchmark gains, with complete session information preserved during augmentation.
Problem
Dropout-based self-supervised recommendation methods are poorly suited to session-based recommendation because they create sparser sessions and unserviceable self-supervision signals.
Method
COTREC constructs item and session graph views, trains distinct graph encoders, and uses evolving pseudo-labels with contrastive learning to supervise the views mutually.
Results
COTREC almost outperforms all baselines across datasets and shows especially large gains on Tmall, according to the reported experiments.
Takeaways & Limitations
The framework provides more informative and practicable self-supervision while preserving complete session information rather than relying on fragmentary dropout views.
Takeaways & Limitations
Pseudo-labels are not retained in the training set because false information could misguide later training and repeated adjacency reconstruction would reduce efficiency.
Abstract
from arXiv · showhide
Session-based recommendation targets next-item prediction by exploiting user behaviors within a short time period. Compared with other recommendation paradigms, session-based recommendation suffers more from the problem of data sparsity due to the very limited short-term interactions. Self-supervised learning, which can discover ground-truth samples from the raw data, holds vast potentials to tackle this problem. However, existing self-supervised recommendation models mainly rely on item/segment dropout to augment data, which are not fit for session-based recommendation because the dropout leads to sparser data, creating unserviceable self-supervision signals. In this paper, for informative session-based data augmentation, we combine self-supervised learning with co-training, and then develop a framework to enhance session-based recommendation. Technically, we first exploit the session-based graph to augment two views that exhibit the internal and external connectivities of sessions, and then we build two distinct graph encoders over the two views, which recursively leverage the different connectivity information to generate ground-truth samples to supervise each other by contrastive learning. In contrast to the dropout strategy, the proposed self-supervised graph co-training preserves the complete session information and fulfills genuine data augmentation. Extensive experiments on multiple benchmark datasets show that, session-based recommendation can be remarkably enhanced under the regime of self-supervised graph co-training, achieving the state-of-the-art performance.
1 INTRODUCTION
Session-based recommendation addresses next-item prediction when long-term user profiles are unavailable, but limited interactions create severe data sparsity. COTREC combines graph-based views, co-training, and contrastive learning to generate more practicable self-supervision signals.
- Session-based recommendation predicts the next item from short-term behavior when accessible long-term user profiles may be unavailable.Examples include guests who do not log in and users who keep personal information private.
- GNN-based session recommenders improve on RNN-based models, yet session sparsity remains an unresolved challenge.
- Self-supervised learning can discover ground-truth samples from raw data and is considered promising for addressing recommendation data sparsity.
- COTREC constructs item and session views, then trains two asymmetric graph encoders to exchange pseudo-labels and improve the main recommendation encoder.
- The framework is model-agnostic and experiments report statistically significant improvements over state-of-the-art baselines on benchmark datasets.
2 RELATED WORK
Related work progresses from Markov chains and recurrent models to graph neural networks for session recommendation, while self-supervised recommendation commonly uses masking or dropout. These augmentation strategies can produce sparse, unusable signals in session-based settings, motivating COTREC’s graph co-training approach.
- Session-based Recommendation: Early session-based recommenders modeled temporal transitions with Markov chains, followed by recurrent neural networks for sequential data.
- Session-based Recommendation: Graph neural networks model item transitions and aggregate session-graph information, with representative methods capturing item-level and session-level patterns.
- Self-Supervised Learning in Recommendation: Self-supervised learning generates labels from unlabeled data and has been applied to visual representation learning, language modeling, and recommendation.
- Self-Supervised Learning in Recommendation: Random dropout-based self-supervision is unsuitable for session recommendation because it creates sparser sessions and unserviceable signals.
- Self-Supervised Learning in Recommendation: S²-DHCN avoids random dropout but uses fixed ground-truths, leading to merely slight improvements according to the related-work discussion.
3.1 Preliminaries
The paper formalizes sessions as timestamp-ordered item sequences and frames recommendation as ranking next-item probabilities. It also introduces co-training as mutual pseudo-label exchange between models trained on different views.
- Preliminaries: A session is a timestamp-ordered sequence of items interacted with by an anonymous user.Items belong to the item set H, and the session contains m interactions.
- Preliminaries: Session-based recommendation predicts the next item i_s,m+1 for a given session.
- Preliminaries: The recommendation model outputs item probabilities, and the top-K items with the highest probabilities are selected.
- Co-Training: Co-training trains two classifiers on different views and exchanges confident pseudo-labels to construct additional labeled data for each other.
3.2 Self-Supervised Graph Co-Training
COTREC constructs item- and session-level graph views, then uses asymmetric graph encoders and iterative co-training to generate self-supervision signals for session-based recommendation. Contrastive learning and adversarial divergence constraints refine the two views while preserving complementary structural information.
- View Construction: COTREC derives item and session views from intra- and inter-connectivity patterns, capturing item-level and session-level structural information.The item view connects items co-occurring within sessions, while the session view represents session relations.
- Graph Encoding: Two graph-convolution encoders operate over the views as a concrete, adaptable base for the model-agnostic framework.The encoders can be modified for other session-based neural graph recommendation scenarios.
- Graph Encoding: The item encoder learns item connectivity but adds reversed position embeddings because graph convolution alone cannot encode within-session item order.Session representations aggregate item embeddings, with attention-based refinement described for item importance.
- Self-Supervision Mining: Each view predicts high-confidence top-K positive samples and informative hard negatives, which become augmented ground-truth signals for supervising the other encoder.Negatives are sampled from highly ranked items rather than the lowest-scoring items to avoid easy examples and reduce false-negative risk.
- Cross-View Co-Training: The pseudo-labels are iteratively reconstructed across views, allowing high-confidence next-item candidates from one view to refine representations in the other through contrastive learning.The contrastive objective compares the last-clicked item with predicted positive and negative samples in session context.
- Divergence Constraint: An adversarial divergence constraint prevents the two encoders from collapsing to similar representations and generating identical ground-truths.The encoders are trained to resist adversarial examples generated by each other while retaining correct predictions.
4.1 Experimental Settings
The experiments evaluate COTREC on three real-world benchmark datasets using Precision and Mean Reciprocal Rank at K=10 or 20, alongside representative session-based recommendation baselines.
- Datasets: Three benchmark datasets—Tmall, RetailRocket, and Diginetica—are used for evaluation.Tmall contains anonymized shopping logs, RetailRocket contains six months of browsing activity, and Diginetica describes music listening behavior.
- Baseline Methods: COTREC is compared with Markov-chain, recurrent, attention-based, graph-neural, and self-supervised session-based recommendation methods.Baselines include FPMC, GRU4REC, NARM, STAMP, SR-GNN, GCE-GNN, and S2-DHCN.
- Evaluation Metrics: Recommendation quality is measured with P@K and MRR@K, where K is 10 or 20.P@K denotes Precision and MRR@K denotes Mean Reciprocal Rank.
- Implementation Details: The experiments use embedding size 100, batch size 100, L2 regularization 10^-5, Adam optimization with learning rate 0.001, and three graph-convolution layers.Parameters are initialized from N(0, 0.1), and baseline models use their best reported settings.
4.2 Experimental Results
Experiments show that COTREC consistently improves session-based recommendation through self-supervised graph co-training, with gains across datasets, SSL comparisons, session lengths, and component analyses.
- Overall Performance: Graph-based methods outperform RNN-based methods, while modeling inter- and intra-session information improves session-based next-item prediction.Temporal information also benefits recent methods compared with FPMC.
- Overall Performance: COTREC almost outperforms all baselines across datasets, with a particularly large margin on Tmall and competitive performance against GCE-GNN.GCE-GNN encounters an out-of-memory problem on RetailRocket in the reported setup.
- Ablation Study: Self-supervised graph co-training is the strongest ablated component, and removing it causes a remarkable performance drop on both evaluation metrics.The divergence constraint also prevents mode collapse, while soft attention improves session representation over averaging.
- Comparison with Different SSL Methods: Only self-supervised graph co-training improves recommendation on both Tmall and Diginetica and achieves the best performance among compared SSL methods.Item masking is least effective in most cases because limited session behaviors produce sub-optimal self-supervision signals.
- Handling Different Session Lengths: COTREC almost outperforms all baselines for both short and long sessions, demonstrating adaptability to sessions with different lengths.Short sessions perform better than long sessions in the reported results.
- Hyperparameter and Layer Analysis: A three-layer graph convolution setting achieves the best performance on both Tmall and Diginetica, whereas more layers reduce performance through over-smoothing.Diginetica also exhibits an obvious performance fluctuation as the layer count changes.
5 CONCLUSION
The paper proposes self-supervised graph co-training to address session-based recommendation’s weak supervision signals and reports improved performance over recent baselines.
- Self-supervised graph co-training addresses session-based recommendation’s data sparsity by generating informative pseudo-labels from unlabeled data.The framework iteratively selects evolving pseudo-labels as self-supervision examples for each view.
- The framework uses co-training to let different graph views iteratively improve one another.Two views are constructed from session-item relations, and their encoders exchange pseudo-label supervision.
- The approach avoids random dropout-based self-discrimination, which creates sparser data and weak self-supervision signals for session-based recommendation.
- Extensive experiments and empirical studies demonstrate effectiveness and superiority over other recent baselines.