Cross‑Session Aware Temporal Convolutional Network (CA‑TCN) for Session‑Based Recommendation

Session‑based recommendation is a sub‑field of recommender systems that predicts the next user action based solely on the current session’s item sequence, without relying on explicit user profiles. Existing methods either use recurrent neural networks (RNNs) or graph neural networks (GNNs), but they suffer from limited modeling of cross‑session information and loss of positional cues.

Observation 1 : Most current approaches focus only on intra‑session data and ignore valuable cross‑session influences that can improve preference inference.

Observation 2 : Graph‑based methods treat the same item appearing at different time steps as a single node, discarding temporal order and long‑range dependencies.

Related Work : Session‑based recommenders are broadly categorized into collaborative‑filtering methods (e.g., KNN‑RNN, CSRM) and deep‑learning methods (e.g., GRU4Rec, NARM, SR‑GNN, GC‑SAN). While GNN‑based models achieve state‑of‑the‑art performance, they still lack cross‑session modeling and positional awareness.

CA‑TCN Model Overview : The proposed Cross‑Session Aware Temporal Convolutional Network (CA‑TCN) integrates three components:

Cross‑Session Item Graph : Constructs a directed graph G_item where each node is an item and edges represent sequential clicks across all sessions. Edge weights encode both direction (incoming/outgoing adjacency matrices) and co‑occurrence frequency, allowing the GNN to generate global item embeddings that capture cross‑session signals.

Temporal Convolutional Network (TCN) : Applies causal and dilated convolutions to the sequence of item embeddings, producing local and global session representations. An item‑level attention mechanism aggregates item embeddings into a session vector.

Session‑Context Graph : Builds a graph where nodes are sessions and edges reflect similarity (via K‑NN on session vectors). A graph attention network refines session embeddings by incorporating neighboring session information.

The final session representation is obtained by fusing local TCN output, global session vector, and cross‑session context, and is used to predict the probability of each candidate item.

Experimental Evaluation : CA‑TCN was tested on two public benchmarks, Yoochoose and Diginetica. Results show that CA‑TCN outperforms existing RNN‑based and GNN‑based state‑of‑the‑art methods. Ablation studies confirm the contribution of each component (TCN, Cross‑Session Item Graph, Session‑Context Graph).

Future Work : The model has been patented and will be deployed in Meituan’s e‑commerce platforms (e.g., “团好货”, “美团优选”). Further research will explore broader business lines and additional sequence‑based tasks.

References

[1] S. Wang et al., “A survey on session‑based recommender systems,” arXiv:1902.04864, 2019.

[2] B. Hidasi et al., “Session‑based recommendations with recurrent neural networks,” arXiv:1511.06939, 2015.

[3] S. Wu et al., “Session‑based recommendation with graph neural networks,” AAAI, 2019.

[4] C. Xu et al., “Graph contextualized self‑attention network for session‑based recommendation,” IJCAI, 2019.

[5] S. Bai et al., “An empirical evaluation of generic convolutional and recurrent networks for sequence modeling,” arXiv:1803.01271, 2018.

[6] D. Jannach & M. Ludewig, “When recurrent neural networks meet the neighborhood for session‑based recommendation,” RecSys ’17, 2017.

[7] M. Wang et al., “A collaborative session‑based recommendation approach with parallel memory modules,” SIGIR, 2019.

[8] J. Li et al., “Neural attentive session‑based recommendation,” CIKM, 2017.

[9] W. Dong et al., “Efficient k‑nearest neighbor graph construction for generic similarity measures,” WWW, 2011.

[10] P. Veličković et al., “Graph attention networks,” arXiv:1710.10903, 2017.