Intrinsic and Extrinsic Factor Disentanglement for Recommendation in Various Context Scenarios

TL;DR

The paper tackles context-dependent user behavior by formalizing intrinsic (context-invariant) and extrinsic (context-driven) factors and proposing IEDR, a generic framework that differentiates these factors across multiple contexts. It introduces two modules: RP for predicting interactions using a context-graph representation via SIGN, and CIED, which blends context-invariant contrastive learning with bidirectional MI-based disentangling (BiDis) to obtain disentangled factor representations. The authors provide an information-theoretic interpretation of CICL, discuss optimization and complexity, and address potential degeneracies with statistical interaction and a symmetric disentangling approach. Extensive experiments on real-world datasets show IEDR achieving up to 4% gains in NDCG over strong baselines and demonstrate robust factor disentanglement and generalization across diverse contexts.

Abstract

In recommender systems, the patterns of user behaviors (e.g., purchase, click) may vary greatly in different contexts (e.g., time and location). This is because user behavior is jointly determined by two types of factors: intrinsic factors, which reflect consistent user preference, and extrinsic factors, which reflect external incentives that may vary in different contexts. Differentiating between intrinsic and extrinsic factors helps learn user behaviors better. However, existing studies have only considered differentiating them from a single, pre-defined context (e.g., time or location), ignoring the fact that a user's extrinsic factors may be influenced by the interplay of various contexts at the same time. In this paper, we propose the Intrinsic-Extrinsic Disentangled Recommendation (IEDR) model, a generic framework that differentiates intrinsic from extrinsic factors considering various contexts simultaneously, enabling more accurate differentiation of factors and hence the improvement of recommendation accuracy. IEDR contains a context-invariant contrastive learning component to capture intrinsic factors, and a disentanglement component to extract extrinsic factors under the interplay of various contexts. The two components work together to achieve effective factor learning. Extensive experiments on real-world datasets demonstrate IEDR's effectiveness in learning disentangled factors and significantly improving recommendation accuracy by up to 4% in NDCG.

Figures (13)

• Figure 1: An example to compare existing work (consider only the context of social settings) and our approach (consider various contexts) in learning intrinsic and extrinsic factors. The upper part shows the preference fact (upper left) and observed behaviors (upper right) of a user Bob. The bottom part shows the possible factor learning results and corresponding recommendations of existing work (bottom left) and our approach (bottom right).
• Figure 2: An Overview of IEDR. It is a symmetric structure on the user side and the item side. The middle part (the black arrows) represents the recommendation prediction (RP) module (Section \ref{['sec:rpm']}). It generates the intrinsic and extrinsic factor representations ($\bm{o}_{in}$ and $\bm{o}_{ex}$) for producing the recommendation prediction $y^{\prime}$. The side parts are two contrastive intrinsic-extrinsic disentangling (CIED) modules. Each CIED includes a context-invariant contrastive learning component (the red arrows, Section \ref{['sec:method_CICL']}), and a disentangling component (the blue arrows, Section \ref{['sec:method_biDis']}) to ensure the success of the factor learning. The losses generated through these modules ($\mathcal{L}_{RP},\mathcal{L}_{CICL},\mathcal{L}_{bi\text{-}appr},\mathcal{L}_{Dis}$) will be optimized as a two-step multi-task training (Section \ref{['sec:method_multitask']}).
• Figure 3: An illustrative example demonstrating the potential problem of asymmetric learning in vCLUB. The blue circles are intrinsic representations, and the red circles are extrinsic representations. The dotted arrows are the directions that vCLUB will push $\bm{o}_{in}^{u}$ and $\bm{o}_{ex}^{u}$ to move toward their space.
• Figure 4: Ablation studies results with different component(s) removed.
• Figure 5: The performance and variance statistics of vCLUB and BiDis.

Theorems & Definitions (2)

• Definition 1
• Theorem 1: Equivalence of contrastive loss $\mathcal{L}_{\textit{CICL}}^{u}$