Scenario-Adaptive Fine-Grained Personalization Network: Tailoring User Behavior Representation to the Scenario Context

TL;DR

The paper tackles the problem of effectively modeling user interests across multiple scenarios by moving beyond coarse-grained, sequence-wide adaptation. It introduces SFPNet, a deep architecture built from Scenario-Tailoring Blocks that combine a Scenario-Adaptive Module (SAM) with a Distribution-Aware Pooling mechanism and a Residual-Tailoring Module (RTM) to produce fine-grained, per-behavior context-aware representations. Through extensive offline experiments on industrial Lazada data and Ali-CCP, along with ablation studies, hyperparameter analyses, and online A/B testing, SFPNet consistently outperforms state-of-the-art scenario-aware methods, achieving higher AUC and S-GAUC as well as practical gains (e.g., 6.4% revenue and 9.2% CTR). The work demonstrates the importance of fine-grained, context-rich sequence modeling for tracking user-interest migrations across scenarios and provides a scalable framework for real-world multi-scenario recommendations.

Abstract

Existing methods often adjust representations adaptively only after aggregating user behavior sequences. This coarse-grained approach to re-weighting the entire user sequence hampers the model's ability to accurately model the user interest migration across different scenarios. To enhance the model's capacity to capture user interests from historical behavior sequences in each scenario, we develop a ranking framework named the Scenario-Adaptive Fine-Grained Personalization Network (SFPNet), which designs a kind of fine-grained method for multi-scenario personalized recommendations. Specifically, SFPNet comprises a series of blocks named as Scenario-Tailoring Block, stacked sequentially. Each block initially deploys a parameter personalization unit to integrate scenario information at a coarse-grained level by redefining fundamental features. Subsequently, we consolidate scenario-adaptively adjusted feature representations to serve as context information. By employing residual connection, we incorporate this context into the representation of each historical behavior, allowing for context-aware fine-grained customization of the behavior representations at the scenario-level, which in turn supports scenario-aware user interest modeling.

Figures (5)

• Figure 1: An illustrative example of multi-scenarios for the online e-commerce platform Lazada.
• Figure 2: Examples of previous methods and ours. In (a), $\boldsymbol{V}_1=w_1\boldsymbol{V}$ and $\boldsymbol{V}_2=w_2 \boldsymbol{V}$, where $\boldsymbol{V} = f(\boldsymbol{b}_1, ..., \boldsymbol{b}_L)$. While in (b), $\boldsymbol{V}_1=f(w_1^{s_1}\boldsymbol{b}_1, ..., w_L^{s_1}\boldsymbol{b}_L)$ and $\boldsymbol{V}_2= f(w_1^{s_2}\boldsymbol{b}_1, ..., w_L^{s_2}\boldsymbol{b}_L)$. In our method, $\boldsymbol{V}_1=f(\boldsymbol{b}_1^{s_1}, ..., \boldsymbol{b}_L^{s_1})$, where $\boldsymbol{b}_1^{s_1}$ is obtained by $(\boldsymbol{V}, \boldsymbol{C}, \boldsymbol{S}_1, \boldsymbol{b}_1)$.
• Figure 3: The network architecture of our proposed Scenario-Adaptive Fine-Grained Personalization Network (SFPNet).
• Figure 4: The impact of multi-scenario joint training.
• Figure 5: The performance of SFPNet under different predefined hyper-parameters on the both datasets.