Proactive Recommendation with Iterative Preference Guidance

Shuxian Bi, Wenjie Wang, Hang Pan, Fuli Feng, Xiangnan He

In the world of recommender systems (RS), we are often trapped in a "feedback loop." Algorithms observe what we like and give us more of it, which reinforces our existing tastes and eventually leads to the dreaded "filter bubble." But what if a recommender system didn't just follow your lead? What if it could strategically—and successfully—nudge you toward new interests?

A new paper from researchers at the University of Science and Technology of China and the National University of Singapore, titled "Proactive Recommendation with Iterative Preference Guidance," introduces a framework called IPG designed to do exactly that.

TL;DR

IPG is a model-agnostic, post-processing framework that proactively steers user interests toward a specific target item or topic. By combining interaction probability with an explicit "guiding score," it identifies a sequence of items that are both likely to be clicked and effective at shifting the user’s long-term preferences.

What Problem Does This Paper Address?

Most modern recommenders are passive. They cater to inferred interests, which can lead to opinion polarization and narrowed horizons. Proactive recommendation aims to solve this by actively guiding users to "jump out" of their historical interests toward a predefined target.

Existing attempts at proactive recommendation (like the Influential Recommender System, or IRS) suffer from three main flaws:

  1. Unrealistic Paths: They often assume users will blindly accept every recommendation, ignoring the fact that users won't click on things that don't interest them right now.
  2. Implicit Objectives: They lack a clear, mathematical objective for "guidance," often relying on simple attention mechanisms.
  3. Inflexibility: They usually require specialized architectures, making them hard to plug into existing industrial systems.

Methods

The authors propose Iterative Preference Guidance (IPG), which operates as a flexible post-processing layer on top of standard sequential models (like SASRec).

The IPG Score

Instead of ranking items solely by click probability, IPG uses a dual-objective score: $$\text{IPG Score} = \text{Interaction Probability} \times \text{Guiding Score}$$

  • Interaction Probability: Ensures the item is relevant enough for the user to actually engage with it.
  • Guiding Score: Quantifies how much the user's preference for the target item will increase if they interact with the current candidate item.

The Efficiency Hack: Linear Assumption

Calculating how a user's embedding changes after every potential interaction is computationally expensive (requiring a full pass through a neural network for every item in the catalog). To solve this, the authors use a linear assumption: 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 rank thousands of items in real-time without a massive performance hit.

Datasets & Evaluation

Evaluating proactive recommendation is notoriously difficult because standard offline datasets (like MovieLens) are static—they don't show how a user would have changed if you had shown them something different.

To solve this, the authors built an Environment Simulator.

  • User/Item Embeddings: 20-dimensional vectors representing different categories.
  • Preference Evolution: A mechanism where a user's embedding moves toward the items they interact with.
  • Boredom Effects: The simulator accounts for "item-level" and "category-level" boredom, where users lose interest if over-exposed to the same things.
  • Scale: The simulated dataset included 6,034 users and 3,533 items.

Key Insights and Ideas

  • The "Bridge" Strategy: IPG doesn't just recommend the target item immediately (which the user might ignore). It finds "bridge" items that share features with both the user's current taste and the target goal.
  • Model Agnosticism: Because IPG is a post-processing step, you don't need to retrain your core recommendation model. You can take a well-trained SASRec model and simply apply the IPG scoring logic during inference.
  • Explicit Guidance: By mathematically defining the "Guiding Score" as the change in inner product between the user and the target, the system has a clear "north star" for its trajectory.

Key Results

The researchers compared IPG against standard SASRec and the state-of-the-art IRS. They measured success using Hit Ratio (HR) (did they click?) and Increase of Interest (IoI) (did they get closer to the target?).

  • Superior Guidance: SASRec-IPG significantly outperformed all baselines in IoI. For example, at 20 rounds of recommendation, IPG achieved an IoI of 0.2649, compared to just 0.0647 for the previous state-of-the-art (IRN).
  • Balanced Accuracy: While proactive guidance naturally leads to a slight drop in immediate Hit Ratio (since you aren't just "giving them what they want"), IPG maintained a much higher HR than heuristic-based steering methods.
  • Visual Proof: Case studies (visualized via embedding heatmaps) showed that under IPG, user embeddings systematically migrated toward the target item's features over time, whereas other methods stayed stuck in the user's initial preference zone.

Limitations and Future Work

  • Simulation vs. Reality: The biggest hurdle remains the "sim-to-real" gap. While the simulator is sophisticated, human behavior is far more unpredictable. Real-world A/B testing is the necessary next step.
  • Fixed Targets: Currently, the target item is manually selected. Future iterations could explore how to automatically choose "healthy" or "diverse" targets to automatically break filter bubbles.
  • LLM Integration: The authors suggest that Large Language Models (LLMs) could be used for better high-level planning of the "guiding path," potentially making the transitions feel even more natural to the user.

Summary updated

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.
BibTeX
@inproceedings{bi2024proactive,
  author = {Shuxian Bi and Wenjie Wang and Hang Pan and Fuli Feng and Xiangnan He},
  title = {{Proactive Recommendation with Iterative Preference Guidance}},
  booktitle = {The Web Conference},
  year = {2024},
  doi = {10.1145/3589335.3651548},
}

Loading related content...

Figure