Tuesday 10 June 2025
Recommending a new approach to e-commerce, researchers at Taobao & Tmall Group of Alibaba have developed a novel method for generating list-level multi-objective optimization in online recommendation systems. This innovative technique, dubbed SORT-Gen, aims to overcome the limitations of traditional item-level multi-objective methods by incorporating sequential ordered regression transformer-generator architecture.
In typical e-commerce recommendation systems, users are presented with lists of items based on their preferences and browsing history. However, these systems often neglect the dynamic user intent and contextual interactions between items. By introducing SORT-Gen, researchers aim to address this shortcoming by optimizing multiple business objectives simultaneously at the list-level.
The core idea behind SORT-Gen is to divide the recommendation process into two stages: sequential ordered regression transformer and generator. In the first stage, the sequential ordered regression transformer accurately estimates multi-objective values for variable-length sub-lists. This is achieved by leveraging a novel mask-driven fast generation algorithm that simulates multi-round candidate selection within one forward pass.
The second stage involves generating diversified item lists based on the estimated multi-objective values. This is done using a generator network that integrates maximal marginal relevance (MMR) to balance accuracy and diversity. By incorporating MMR, SORT-Gen ensures that the generated list not only maximizes click-through rates but also maintains user engagement and satisfaction.
The researchers conducted extensive online experiments on Taobao’s live traffic, comparing SORT-Gen with traditional re-ranking models and other state-of-the-art methods. The results showed a significant improvement in multiple objectives, including clicks, conversions, and gross merchandise volume (GMV). In particular, SORT-Gen outperformed the baseline model by 4.13% in click-through rates and 8.10% in GMV.
One of the key advantages of SORT-Gen is its ability to adapt to different scenarios and user behaviors. By incorporating sequential ordered regression transformer-generator architecture, the system can effectively handle complex user interactions and contextual factors that influence recommendation outcomes.
The implications of this research are far-reaching, with potential applications in various domains beyond e-commerce. For instance, SORT-Gen could be used in content recommendation systems, such as music or video streaming platforms, to improve user engagement and satisfaction.
In summary, the development of SORT-Gen represents a significant step forward in online recommendation systems. By optimizing multiple business objectives simultaneously at the list-level, this innovative technique has the potential to revolutionize e-commerce and beyond.
Cite this article: “SORT-Gen: A Novel Method for List-Level Multi-Objective Optimization in Online Recommendation Systems”, The Science Archive, 2025.
E-Commerce, Recommendation Systems, Online Optimization, Multi-Objective Optimization, List-Level Optimization, Sort-Gen, Sequential Ordered Regression Transformer-Generator Architecture, Mmr, Maximal Marginal Relevance, Click-Through Rates, Gross Merchandise Volume.
