SURGE: A Graph Neural Network Based Sequential Recommendation Framework

Sequential recommendation aims to predict a user's next interaction by exploiting their historical behavior sequence, but implicit and noisy signals in long histories hinder accurate interest modeling. To address these challenges, researchers from Tsinghua University and Kuaishou propose SURGE, a graph neural network (GNN) based framework that dynamically fuses and extracts core user interests.

SURGE consists of four components: (A) Interest‑graph construction, where metric learning transforms a user's interaction sequence into an item‑item graph, clustering similar preferences; (B) Interest‑fusion graph convolution layer, which applies cluster‑aware and query‑aware attention to aggregate weak implicit feedback into strong interest signals; (C) Interest‑extraction graph pooling layer, a dynamic pooling mechanism that compresses the large graph into a compact subgraph while preserving temporal order via regularizations on the assignment matrix; and (D) Prediction layer, which flattens the pooled graph into an enhanced interest sequence and models its evolution with an attention‑augmented GRU (AUGRU) to forecast the next item.

Experiments on two large‑scale datasets (Taobao e‑commerce with sequence length 50 and Kuaishou short‑video with length 250) evaluate SURGE using AUC, GAUC, MRR, and NDCG. Compared with strong baselines such as DIN, DIEN, PLASTIC, and SLi‑Rec, SURGE consistently achieves higher accuracy and ranking scores, especially on the longer Kuaishou sequences, demonstrating its superiority in handling noisy, long‑range dependencies.

Further analyses show that SURGE converges faster and more stably than baselines, reducing training time by over 20% on the Kuaishou dataset. Ablation studies on different sequence lengths and prediction layers confirm that the graph‑based pooling and attention mechanisms are key to its robustness and generalization.

Overall, SURGE provides a new perspective for sequential recommendation by compressing user histories into interest graphs, mitigating noise, and simplifying interest modeling, with reported real‑world deployment at Kuaishou yielding approximately a 1% increase in watch time.