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

Background In the era of big data, recommendation systems are essential for filtering information overload, but many methods rely on explicit user profiles that may be unavailable in certain domains. Session‑based recommendation predicts the next user action from the current session’s item sequence without requiring user profiles.

Traditional RNN‑based models capture only sequential, one‑directional item transitions, while graph‑based approaches (e.g., SR‑GNN, GC‑SAN) model item relationships but ignore positional information and long‑range dependencies. Moreover, existing methods focus solely on intra‑session data, overlooking valuable cross‑session signals.

Observation 1 : Most methods ignore cross‑session information, which can provide complementary cues such as category or brand influences across sessions.

Observation 2 : Graph‑based models treat the same item at different time steps as a single node, losing temporal order and long‑range dependencies; TCNs with causal and dilated convolutions preserve these aspects.

Related Work Session‑based recommendation methods are divided into collaborative‑filtering approaches (KNN‑based, similarity‑based) and deep‑learning approaches (RNN, GRU4Rec, NARM, graph‑based models like SR‑GNN and GC‑SAN). These methods either lack cross‑session modeling or suffer from limited positional awareness.

CA‑TCN Model Overview CA‑TCN addresses these gaps by jointly modeling item‑level and session‑level cross‑session effects. It constructs a Cross‑Session Item‑Graph to link related items across sessions, processes the graph with a GNN to obtain global item embeddings, feeds these embeddings into a Temporal Convolutional Network (TCN) to capture sequential patterns, and finally builds a Session‑Context Graph to model relationships among sessions.

1. Cross‑Session Item‑Graph Each node represents an item; directed edges (v_s_i → v_s_i+1) encode click order within sessions. Unlike prior graphs, this structure aggregates click information from all sessions, allowing the GNN to learn richer item relationships weighted by co‑occurrence frequency.

2. Temporal Convolutional Network (TCN) The TCN receives the item vectors from the graph and applies causal and dilated convolutions to extract both local and global session representations. The last item’s TCN output serves as the session’s local representation, while an attention‑weighted sum of all item representations yields the global session vector.

3. Session‑Context Graph To incorporate inter‑session influences, a Session‑Context Graph is built where each node is a session and edges reflect similarity (computed from session vectors). A K‑NN graph determines neighbors, and a graph‑attention network aggregates contextual information into a session‑aware representation.

4. Click Prediction The final session representation combines local, global, and cross‑session components via a fusion function, and the probability of each candidate item being the next click is computed from the dot product between item and session vectors, followed by ranking.

Experimental Evaluation CA‑TCN was evaluated on the public Yoochoose and Diginetica datasets. Results show that CA‑TCN outperforms existing RNN‑based and graph‑based state‑of‑the‑art methods. Ablation studies confirm that each component (TCN, Cross‑Session Item‑Graph, Session‑Context Graph) contributes to performance gains.

Future Work The model has been patented and will be deployed across Meituan’s business lines for e‑commerce and other recommendation scenarios, such as “团好货” and “美团优选”.

References

S. Wang, L. Cao, and Y. Wang, “A survey on session‑based recommender systems,” arXiv preprint arXiv:1902.04864, 2019.

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

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

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

S. Bai, J. Z. Kolter, and V. Koltun, “An empirical evaluation of generic convolutional and recurrent networks for sequence modeling,” arXiv:1803.01271, 2018.

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

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

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

W. Dong, C. Moses, and K. Li, “Efficient k‑nearest neighbor graph construction for generic similarity measures,” WWW 2011.

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

Author Information The paper was authored by Ye Rui, Zhang Qing, and Heng Liang from Meituan Platform Growth Technology Department.