Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Towards Explainable Temporal User Profiling with LLMs

Milad Sabouri, Masoud Mansoury, Kun Lin, Bamshad Mobasher

TL;DR

The paper tackles the challenge of making content-based recommender systems both accurate and transparent by modeling temporal user preferences with LLM-generated textual profiles. It creates two narrative profiles per user—short-term and long-term—processes them with BERT to produce embeddings, and fuses them via an attention mechanism before scoring with an MLP. This design yields improved ranking over strong baselines and provides intrinsic explainability through human-readable summaries and attention weights, enabling insights into whether recommendations are driven by recent activity or enduring tastes. The approach is validated on real-world datasets across domains with contrasting profile sizes, demonstrating both performance gains and practical potential for user-facing explanations in transparent recommender systems.

Abstract

Accurately modeling user preferences is vital not only for improving recommendation performance but also for enhancing transparency in recommender systems. Conventional user profiling methods, such as averaging item embeddings, often overlook the evolving, nuanced nature of user interests, particularly the interplay between short-term and long-term preferences. In this work, we leverage large language models (LLMs) to generate natural language summaries of users' interaction histories, distinguishing recent behaviors from more persistent tendencies. Our framework not only models temporal user preferences but also produces natural language profiles that can be used to explain recommendations in an interpretable manner. These textual profiles are encoded via a pre-trained model, and an attention mechanism dynamically fuses the short-term and long-term embeddings into a comprehensive user representation. Beyond boosting recommendation accuracy over multiple baselines, our approach naturally supports explainability: the interpretable text summaries and attention weights can be exposed to end users, offering insights into why specific items are suggested. Experiments on real-world datasets underscore both the performance gains and the promise of generating clearer, more transparent justifications for content-based recommendations.

Towards Explainable Temporal User Profiling with LLMs

TL;DR

The paper tackles the challenge of making content-based recommender systems both accurate and transparent by modeling temporal user preferences with LLM-generated textual profiles. It creates two narrative profiles per user—short-term and long-term—processes them with BERT to produce embeddings, and fuses them via an attention mechanism before scoring with an MLP. This design yields improved ranking over strong baselines and provides intrinsic explainability through human-readable summaries and attention weights, enabling insights into whether recommendations are driven by recent activity or enduring tastes. The approach is validated on real-world datasets across domains with contrasting profile sizes, demonstrating both performance gains and practical potential for user-facing explanations in transparent recommender systems.

Abstract

Accurately modeling user preferences is vital not only for improving recommendation performance but also for enhancing transparency in recommender systems. Conventional user profiling methods, such as averaging item embeddings, often overlook the evolving, nuanced nature of user interests, particularly the interplay between short-term and long-term preferences. In this work, we leverage large language models (LLMs) to generate natural language summaries of users' interaction histories, distinguishing recent behaviors from more persistent tendencies. Our framework not only models temporal user preferences but also produces natural language profiles that can be used to explain recommendations in an interpretable manner. These textual profiles are encoded via a pre-trained model, and an attention mechanism dynamically fuses the short-term and long-term embeddings into a comprehensive user representation. Beyond boosting recommendation accuracy over multiple baselines, our approach naturally supports explainability: the interpretable text summaries and attention weights can be exposed to end users, offering insights into why specific items are suggested. Experiments on real-world datasets underscore both the performance gains and the promise of generating clearer, more transparent justifications for content-based recommendations.
Paper Structure (26 sections, 10 equations, 3 figures, 5 tables)

This paper contains 26 sections, 10 equations, 3 figures, 5 tables.

Table of Contents

  1. Introduction
  2. Related Works
  3. Explainable LLM-Driven Temporal User Profiling
  4. User Profile Creation
  5. Pass 1: Short-Term Profile Generation
  6. Pass 2: Long-Term Profile Generation
  7. Embedding Representation
  8. Recommendation Generation
  9. Model Highlights and Strengths
  10. Experiments & Explainability Demonstration
  11. Datasets
  12. Baselines
  13. Centric Approach (Centric)
  14. Temporal Fusion without LLM-Based Representations (Temp-Fusion)
  15. Popularity-Based Recommendation (Popularity)
  16. ...and 11 more sections

Figures (3)

  • Figure 1: Proposed Architecture for LLM-Driven Temporal User Profiling
  • Figure 2: A conceptual illustration of our framework's potential explainability. (A) and (B) depict the short-term vs. long-term textual profiles our current approach generates and encodes. (C) shows a hypothetical extension wherein the final recommendation is explicitly justified by both sets of preferences—an aspect we plan to explore in future work.
  • Figure 3: Relative improvement of our complete approach over four ablation variants on the Movies&TV dataset. Bars show percentage gains in Recall@K and NDCG@K.