Proactive Recommendation with Iterative Preference Guidance

TL;DR

Proactive recommendation addresses filter bubbles by actively shaping user interests toward a target item. The paper proposes Iterative Preference Guidance (IPG), a model-agnostic post-processing framework that ranks candidate items using the IPG score $r_{uij}^{t}= \hat{p}_{ui}^{t} \cdot g_{uij}^{t}$, where $g_{uij}^{t}$ is derived from an explicit, iteratively updated user representation and approximations $\hat{\mathbf{e}}_u^{(t+1)} = \gamma \hat{\mathbf{e}}_u^{(t)} + (1-\gamma) \hat{\mathbf{e}}_i$. By updating user representations with the most recent feedback, IPG provides an explicit guiding objective while maintaining compatibility with existing sequential RS models. An interactive environment simulator is designed to evaluate guiding efficacy via log-collection and guidance phases, using metrics such as HR@K and IoI@K. Experiments show IPG improves guiding performance (IoI@K) with a reasonable trade-off in recommendation accuracy, suggesting practical potential for deployment in industry-scale systems. $r_{uij}^{t}= \hat{p}_{ui}^{t} \cdot g_{uij}^{t}$, $g_{uij}^{t}= \hat{\mathbf{e}}_j^\top \hat{\mathbf{e}}_u^{(t+1)} - \hat{\mathbf{e}}_j^\top \hat{\mathbf{e}}_u^{(t)}$, and $\hat{\mathbf{e}}_u^{(t+1)} = \gamma \hat{\mathbf{e}}_u^{(t)} + (1-\gamma)\hat{\mathbf{e}}_i$ are central to the method.

Abstract

Recommender systems mainly tailor personalized recommendations according to user interests learned from user feedback. However, such recommender systems passively cater to user interests and even reinforce existing interests in the feedback loop, leading to problems like filter bubbles and opinion polarization. To counteract this, proactive recommendation actively steers users towards developing new interests in a target item or topic by strategically modulating recommendation sequences. Existing work for proactive recommendation faces significant hurdles: 1) overlooking the user feedback in the guidance process; 2) lacking explicit modeling of the guiding objective; and 3) insufficient flexibility for integration into existing industrial recommender systems. To address these issues, we introduce an Iterative Preference Guidance (IPG) framework. IPG performs proactive recommendation in a flexible post-processing manner by ranking items according to their IPG scores that consider both interaction probability and guiding value. These scores are explicitly estimated with iteratively updated user representation that considers the most recent user interactions. Extensive experiments validate that IPG can effectively guide user interests toward target interests with a reasonable trade-off in recommender accuracy. The code is available at https://github.com/GabyUSTC/IPG-Rec.

Figures (4)

• Figure 1: Illustration of proactive recommendation. The color of the circle above each item indicates the topics.
• Figure 2: Illustration of the proposed IPG framework.
• Figure 3: The first row shows the embedding evolution of user 61 with target item 3257 and the second row shows the embeddings of recommended items under three methods. The first column of each subfigure is the user's initial embedding and the last column shows the target item's embedding.
• Figure 4: Four guiding cases, the logic of each subfigure are in consist with Figure \ref{['fig:case']}.