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

TL;DR

DLCRec tackles the problem of diversity erosion in LLM-based recommender systems by introducing a fine-grained, three-stage task decomposition: Genre Prediction, Genre Filling, and Item Prediction. The framework separates training from control, enabling explicit propagation of user-specified diversity targets through GP→GF→IP, and employs two data augmentation strategies to mitigate data sparsity and distribution skew. Empirical results on MovieLens and Steam show DLCRec achieves precise diversity control with minimal loss in accuracy and outperforms strong baselines, validating the effectiveness of decomposition and augmentation. The approach has practical implications for building more inclusive, discovery-friendly recommender systems that can adapt diversity to individual user needs.

Abstract

The integration of Large Language Models (LLMs) into recommender systems has led to substantial performance improvements. However, this often comes at the cost of diminished recommendation diversity, which can negatively impact user satisfaction. To address this issue, controllable recommendation has emerged as a promising approach, allowing users to specify their preferences and receive recommendations that meet their diverse needs. Despite its potential, existing controllable recommender systems frequently rely on simplistic mechanisms, such as a single prompt, to regulate diversity-an approach that falls short of capturing the full complexity of user preferences. In response to these limitations, we propose DLCRec, a novel framework designed to enable fine-grained control over diversity in LLM-based recommendations. Unlike traditional methods, DLCRec adopts a fine-grained task decomposition strategy, breaking down the recommendation process into three sequential sub-tasks: genre prediction, genre filling, and item prediction. These sub-tasks are trained independently and inferred sequentially according to user-defined control numbers, ensuring more precise control over diversity. Furthermore, the scarcity and uneven distribution of diversity-related user behavior data pose significant challenges for fine-tuning. To overcome these obstacles, we introduce two data augmentation techniques that enhance the model's robustness to noisy and out-of-distribution data. These techniques expose the model to a broader range of patterns, improving its adaptability in generating recommendations with varying levels of diversity. Our extensive empirical evaluation demonstrates that DLCRec not only provides precise control over diversity but also outperforms state-of-the-art baselines across multiple recommendation scenarios.

Figures (6)

• Figure 1: Illustration of the difference between traditional and controllable LLM-based recommender systems.
• Figure 2: Overview of the DLCRec framework. DLCRec decomposes the recommendation task into three sub-tasks: Genre Predicting, Genre Filling, and Item Predicting. The framework consists of two complementary components: the training framework (black lines) and the control framework (read lines). In the training framework, we leverage few-shot data and data augmentation to train each sub-task independently. In contrast, the control framework enables explicit control over the diversity of the final recommendation list by propagating the control number throughout the three sub-tasks.
• Figure 3: Distribution of the number of genres in the Movie and Steam datasets.
• Figure 4: Illustration of the data augmentation method employed in task GF. Two strategies are applied: (a) "GF-N", which introduces noise by replacing the original genre with a noisy genre, and (b) "GF-D", which manipulates the genre distribution by randomly adding or deleting genres up to the sampled controlled threshold, promoting a uniform target distribution.
• Figure 5: Performance comparison of separating and combining tasks GP and GF.