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.
Introduction
In the current landscape of recommender systems (RS), algorithms typically operate in a "feedback loop." By observing and reinforcing existing user tastes, these systems often trap users in "filter bubbles" or "echo chambers." While effective for short-term engagement, this passive approach limits discovery and can lead to opinion polarization. Researchers from the University of Science and Technology of China and the National University of Singapore have proposed a shift from passive catering to proactive recommendation through a framework called Iterative Preference Guidance (IPG).
The Problem: Passive vs. Proactive Recommendation
Most modern recommenders are passive: they observe a user's interest in categories like games or movies and provide more of the same. This reinforces the user's current state, narrowing their interests over time. Proactive recommendation, by contrast, aims to strategically nudge the user toward a new target interest (e.g., "Art") that they might not have explored otherwise.
Existing proactive methods often fail because they assume users will follow any suggested path, ignoring the reality that users only click on items that align with their current (or slightly evolved) preferences. Furthermore, these methods often require complex, specialized architectures that are difficult to integrate into existing industrial pipelines.
The IPG Framework
IPG is a model-agnostic, post-processing framework designed to steer user interests toward a specific target. It identifies a sequence of "bridge" items that are both likely to be clicked and effective at shifting the user’s long-term preferences.
The Dual-Objective Score
Instead of ranking items solely by click probability, IPG uses a combined score:
$$\text{IPG Score} = \text{Interaction Probability} \times \text{Guiding Score}$$
- Interaction Probability: Ensures the item is relevant enough for the user to engage with it immediately.
- Guiding Score: Quantifies how much the user's preference for the target item will increase following an interaction with the candidate item.
Efficiency through Linear Assumption
To make real-time ranking feasible, the authors utilize a linear assumption. Rather than performing a full neural network pass to predict the user's next state for every potential item, they approximate the next user state as a weighted combination of the current state and the new item's embedding. This allows the system to evaluate thousands of candidates efficiently.
Evaluation and Results
Because standard offline datasets are static, the researchers developed an Environment Simulator to model preference evolution and boredom effects. They evaluated IPG using two primary metrics: Hit Ratio (HR) (immediate accuracy) and Increase of Interest (IoI) (success in reaching the target).
Key findings include:
- Superior Guidance: IPG significantly outperformed baselines in IoI. In a 20-round simulation, IPG achieved an IoI of 0.2649, compared to 0.0647 for the previous state-of-the-art method (IRN).
- Balanced Accuracy: While steering users away from their current interests naturally incurs a slight penalty in Hit Ratio, IPG maintained a much higher accuracy than simple heuristic-based steering methods.
- Visual Trajectories: Embedding heatmaps confirmed that under IPG, user representations systematically migrated toward the target features over multiple interaction rounds, whereas passive models remained stagnant.
Conclusion
The IPG framework offers a flexible, effective way to break filter bubbles by proactively guiding user interests. By acting as a post-processing layer, it can be easily integrated into existing sequential recommendation models like SASRec. Future work will focus on bridging the "sim-to-real" gap through real-world A/B testing and exploring the use of Large Language Models for more sophisticated path planning.
Summary updated