Transition-Aware Graph Network Achieves Faster E-Commerce Recommendation with Multi-Behavior Data

Researchers are tackling the challenge of understanding complex user behaviour on e-commerce platforms, where shoppers click, favourite items, add to baskets and ultimately purchase. Hanqi Jin, Gaoming Yang, and Zhangming Chan, all from the Taobao & Tmall Group of Alibaba, alongside Yapeng Yuan, Longbin Li and Fei Sun from the University of Chinese Academy of Sciences, present a new approach to modelling these sequential interactions, moving beyond computationally expensive transformer models. Their work introduces the Transition-Aware Graph Network (TGA), a linear-complexity method that intelligently captures how users move between different behaviours by constructing a structured graph based on item, category and neighbour transitions. This innovation not only improves recommendation accuracy but crucially offers significant computational savings, demonstrated by its successful deployment and positive impact on business metrics within a large-scale industrial system , a vital step towards more efficient and effective e-commerce recommendations.

Recent approaches relying on transformer-based architectures, while effective, often struggle with polynomial time complexity, hindering their use in large-scale industrial applications with lengthy user sequences. To overcome this, the team achieved a linear-complexity approach for modelling multi-behaviour transitions, offering a substantial performance boost in real-world scenarios.

The study reveals that TGA constructs a structured sparse graph by identifying informative transitions from three distinct perspectives: item-level, category-level, and neighbour-level. Unlike traditional transformers that treat all behaviour pairs equally, this innovative graph construction focuses on the most relevant connections between user actions and items. This structured graph then underpins a transition-aware graph attention mechanism, jointly modelling user-item interactions and behaviour transition types to more accurately capture sequential patterns. Initially, the study employed a structured sparse graph construction process, identifying informative transitions across three key perspectives: item-level, category-level, and neighbour-level interactions. This innovative graph structure explicitly encodes user-item interactions, behaviour types, and the relationships between them, enabling a more nuanced understanding of user preferences.
To build this graph, researchers analysed user behaviour sequences, observing that decision-making involves both specific item interactions and broader category exploration. They identified that high-cost items often trigger repeated interactions, indicating careful evaluation, while users frequently compare similar items within a category before converting. Consequently, the team engineered item-level transition edges to model interactions on a specific item, connecting behaviours exhibited on the same product. Simultaneously, category-level transitions were constructed to model cross-item exploration within the same category, capturing comparative shopping patterns.

Furthermore, the study pioneered neighbour-level transitions, representing local behavioural dependencies based on temporal order. This approach allows the model to understand how actions relate to each other in time, improving the accuracy of sequential pattern recognition. Built upon this structured graph, scientists implemented a Transition-Aware Graph Attention mechanism, jointly modelling user-item interactions and behaviour transition types. This mechanism stacks multiple layers to capture high-order dependencies across multiple behaviour dimensions, enhancing the model’s ability to discern complex user intent. Experiments on the Taobao dataset revealed that TGA achieved an Area Under the Curve (AUC) of 0.7454, surpassing all baseline models tested, including the standard Transformer which achieved 0.7276, and demonstrating a 5.8x speed improvement in both training and inference. Further testing on a large industrial dataset confirmed TGA’s robustness, achieving an AUC of 0.8635, while models like the Transformer failed to run due to computational limitations.

The team meticulously examined the impact of individual components within TGA, revealing that removing item-level transitions resulted in a performance decrease, and category-level transitions also led to a drop in AUC, confirming the importance of these mechanisms in capturing user intent. Investigations into layer depth showed that increasing the number of stacked TGA layers consistently improved performance, validating the effectiveness of deep stacking for capturing complex user behavior dependencies. TGA achieves linear time complexity by constructing a structured sparse graph that captures informative transitions at the item, category, and neighbour levels, allowing for more accurate sequential pattern recognition. Researchers demonstrated that TGA outperforms state-of-the-art models in both performance and computational efficiency.

Deployed within a large-scale industrial production environment, TGA yielded a 1.29% improvement in click-through rate and a 1.79% increase in gross merchandise value, indicating substantial business impact. The authors acknowledge that the model’s performance is contingent on the quality of the transition identification process and the construction of the sparse graph. Future work could explore adaptive graph construction techniques or investigate the integration of additional contextual information to further refine the model’s understanding of user intent.