Preference Discerning with LLM-Enhanced Generative Retrieval

Sequential recommendation systems aim to provide personalized recommendations for users based on their interaction history. To achieve this, they often incorporate auxiliary information, such as textual descriptions of items and auxiliary tasks, like predicting user preferences and intent. Despite numerous efforts to enhance these models, they still suffer from limited personalization. To address this issue, we propose a new paradigm, which we term preference discerning. In preference discerning, we explicitly condition a generative sequential recommendation system on user preferences within its context. To this end, we generate user preferences using Large Language Models (LLMs) based on user reviews and item-specific data. To evaluate preference discerning capabilities of sequential recommendation systems, we introduce a novel benchmark that provides a holistic evaluation across various scenarios, including preference steering and sentiment following. We assess current state-of-the-art methods using our benchmark and show that they struggle to accurately discern user preferences. Therefore, we propose a new method named Mender (Multimodal Preference Discerner), which improves upon existing methods and achieves state-of-the-art performance on our benchmark. Our results show that Mender can be effectively guided by human preferences, even though they have not been observed during training, paving the way toward more personalized sequential recommendation systems.

As a result, these preferences must be approximated from the user’s interaction history. Recent works have utilized LLMs to extract user preferences from existing datasets and leverage them for auxiliary tasks (Zhang et al., 2023; Cao et al., 2024). However, these approaches do not allow the model to be dynamically steered by user preferences in their context during inference. Therefore, they usually need to be fine-tuned to be used effectively for new users. Furthermore, currently there is no benchmark that effectively evaluates to what extent these models discern preferences.

We propose a novel paradigm, which we term preference discerning. Preference discerning entails training a multimodal generative retrieval model conditioned on user preferences within its context (see figure 1). This requires approximating a user’s preference in textual form from user-specific data, such as reviews via pre-trained LLMs (Kim et al., 2024). By conditioning the sequential recommendation system on user preferences in-context, we unlock steering via generated user preferences, effectively combining the sequential prior from interaction history with the user preferences. Therefore, users can specify in natural language what item properties they wish to avoid or prefer.

we propose a holistic benchmark that comprises five evaluation axes: (1) preference-based recommendation, (2) sentiment following, (3) fine-grained steering, (4) coarse-grained steering, and (5) history consolidation. We evaluate state-of-the-art generative retrieval methods on our benchmark and find they lack several key abilities of preference discerning. Most importantly, they usually do not generalize well to new user sequences. Therefore, we introduce a novel multimodal generative retrieval method named Multimodal preference discerner (Mender) that effectively fuses pre-trained language encoders with the generative retrieval framework for preference discerning. We demonstrate that Mender can generalize to novel synthetic user sequences in our benchmark

Such preferences are often modeled indirectly from user queries and responses to recommended items (Min et al., 2023; Huang et al., 2013; Ma et al., 2018), or represented as edges on graphs (Ying et al., 2018; Li et al., 2019). In query-aware sequential recommendation He et al. (2022) the model is given keywords in its context that represent the user’s intent but do not capture their preferences.

language provides a natural interface for users to express their preferences and allows harnessing the expressive power of LLMs

Language-Based Sequential Recommendation rely on the premise of enhanced transparency and actionable interrogation of recommendation systems (Radlinski et al., 2022). Furthermore, language provides a natural interface for users to express their preferences and allows harnessing the expressive power of LLMs.

Preference approximation refers to the process of inferring a user’s preferences based on user- and itemspecific data. This process has been user-specific data may include user reviews, profiles, posts, demographic information, or any other relevant details. Incorporating item-specific information is crucial, as it provides additional context that can help alleviate the vagueness or incompleteness often encountered in user-specific data. Preference approximation is a necessary prerequisite that enables in-context conditioning on the generated user preferences.