Abstract:Diversity control is an important task to alleviate bias amplification and filter bubble problems. The desired degree of diversity may fluctuate based on users' daily moods or business strategies. However, existing methods for controlling diversity often lack flexibility, as diversity is decided during training and cannot be easily modified during inference. We propose \textbf{D3Rec} (\underline{D}isentangled \underline{D}iffusion model for \underline{D}iversified \underline{Rec}ommendation), an end-to-end method that controls the accuracy-diversity trade-off at inference. D3Rec meets our three desiderata by (1) generating recommendations based on category preferences, (2) controlling category preferences during the inference phase, and (3) adapting to arbitrary targeted category preferences. In the forward process, D3Rec removes category preferences lurking in user interactions by adding noises. Then, in the reverse process, D3Rec generates recommendations through denoising steps while reflecting desired category preferences. Extensive experiments on real-world and synthetic datasets validate the effectiveness of D3Rec in controlling diversity at inference.

What problem does this paper attempt to address?

This paper attempts to solve the problem of diversity control in recommendation systems, especially how to flexibly adjust the diversity of recommendations during the inference stage. Specifically, existing methods determine the degree of diversity during training and cannot be easily modified during the inference process. This results in a lack of flexibility in the recommendation system, making it difficult to adapt to the ever - changing needs of users or changes in business strategies. ### Background and Problem Description of the Paper Recommendation systems are crucial in the era of information overload. They generate personalized recommendation lists by analyzing users' historical behaviors (such as clicks) and item features (such as categories). However, these systems often focus too much on accuracy and ignore the issue of diversity. Due to the uneven distribution of item categories in user historical data (for example, movie genres), recommendation systems are prone to favor popular items and categories, thus forming a "filter bubble", that is, the recommended content is increasingly limited to the user's known interests, reducing novelty and diversity. ### Limitations of Existing Methods Existing methods can be divided into two categories: post - processing methods and end - to - end methods. Post - processing methods perform diversity control after generating candidate items. However, since these modules are independent of the candidate generation process, diversity signals are not integrated during training, resulting in sub - optimal solutions. End - to - end methods jointly optimize accuracy and diversity during training, but these methods usually lack flexibility because the trade - off between accuracy and diversity is determined by hyper - parameters during training and cannot be easily adjusted during the inference process. ### Proposed Solution To solve the above problems, the authors propose D3Rec (Disentangled Diffusion model for Diversified Recommendation), an end - to - end method that can control the trade - off between accuracy and diversity during the inference stage. The design of D3Rec is based on the following three goals: 1. **Generate Recommendations Based on Category Preferences**: Ensure that the recommendation results reflect the user's category preferences. 2. **Control Category Preferences during the Inference Stage**: Allow for dynamic adjustment of category preferences to meet different needs. 3. **Adapt to Any Target Category Preferences**: Be able to adjust the recommended content according to the user's immediate needs or business strategies. ### Method Overview D3Rec adopts a diffusion model framework to generate recommendations. In the forward process, old category preferences in user interactions are eliminated by adding noise; in the reverse process, recommendations are generated through the denoising step and reflect the target category preferences. Specific implementations include: - **Category Preference Representation**: Use a category embedding matrix to capture category preferences. - **Decoupled Twin - Tower Encoder**: Design a category - preference encoder and a category - independent encoder separately to extract different types of features. - **Category - Preference - Guided Decoder**: Introduce category - preference information during the decoding process to generate recommendations that meet the target preferences. ### Experimental Verification Through experiments on three real - world datasets, the authors verified the effectiveness of D3Rec. The experimental results show that D3Rec outperforms existing methods in both accuracy and diversity, especially in diversity metrics such as coverage and entropy. ### Conclusion D3Rec provides a flexible and effective method for controlling recommendation diversity, which can dynamically adjust the diversity of recommended content according to user needs or business strategies during the inference stage, thereby improving user experience and the practicality of the recommendation system.

Controlling Diversity at Inference: Guiding Diffusion Recommender Models with Targeted Category Preferences

Covering Diversification and Fairness for Better Recommendation (Short Paper).

On Practical Diversified Recommendation with Controllable Category Diversity Framework

FairRec: Fairness Testing for Deep Recommender Systems

Disentangled Representation for Diversified Recommendations

DLCRec: A Novel Approach for Managing Diversity in LLM-Based Recommender Systems

Diversify or Not: Dynamic Diversification for Personalized Recommendation

Recent Advances in Diversified Recommendation

A Joint Optimization Approach for Personalized Recommendation Diversification.

Diversity-Promoting Deep Reinforcement Learning for Interactive Recommendation

Improving End-to-End Sequential Recommendations with Intent-aware Diversification

DifFaiRec: Generative Fair Recommender with Conditional Diffusion Model

Contextual Distillation Model for Diversified Recommendation

Sequential Recommendation with Controllable Diversification: Representation Degeneration and Diversity

Reconciling the accuracy-diversity trade-off in recommendations

A joint optimization approach for personalized recommendation diversification

A Simple Yet Effective Approach for Diversified Session-Based Recommendation

Relieving Popularity Bias in Interactive Recommendation: A Diversity-Novelty-Aware Reinforcement Learning Approach

CDCM: ChatGPT-Aided Diversity-Aware Causal Model for Interactive Recommendation

Divrec: A Framework For Top-N Recommendation With Diversification In E-Commerce

Multi-interest Diversification for End-to-end Sequential Recommendation