Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Following the TRAIL: Predicting and Explaining Tomorrow's Hits with a Fine-Tuned LLM

Yinan Zhang, Zhixi Chen, Jiazheng Jing, Zhiqi Shen

TL;DR

This work tackles the challenge of applying large language models to recommendation by proposing TRAIL, a fine-tuned LLM that jointly forecasts near-term item popularity and generates faithful explanations for non-personalized ranking. By integrating historical trend, current momentum, and item metadata through contrastive learning, TRAIL aligns explanations with trend signals while achieving strong predictive accuracy. Empirical results on three real-world datasets show TRAIL outperforms strong baselines in both popularity prediction and top-N ranking, and human+LLM evaluations confirm the quality and usefulness of its explanations. The approach enables scalable, interpretable recommendations suitable for large catalogs and cold-start scenarios, highlighting the practical impact of combining trend forecasting with natural-language justifications in recommender systems.

Abstract

Large Language Models (LLMs) have been widely applied across multiple domains for their broad knowledge and strong reasoning capabilities. However, applying them to recommendation systems is challenging since it is hard for LLMs to extract user preferences from large, sparse user-item logs, and real-time per-user ranking over the full catalog is too time-consuming to be practical. Moreover, many existing recommender systems focus solely on ranking items while overlooking explanations, which could help improve predictive accuracy and make recommendations more convincing to users. Inspired by recent works that achieve strong recommendation performance by forecasting near-term item popularity, we propose TRAIL (TRend and explAnation Integrated Learner). TRAIL is a fine-tuned LLM that jointly predicts short-term item popularity and generates faithful natural-language explanations. It employs contrastive learning with positive and negative pairs to align its scores and explanations with structured trend signals, yielding accurate and explainable popularity predictions. Extensive experiments show that TRAIL outperforms strong baselines and produces coherent, well-grounded explanations.

Following the TRAIL: Predicting and Explaining Tomorrow's Hits with a Fine-Tuned LLM

TL;DR

This work tackles the challenge of applying large language models to recommendation by proposing TRAIL, a fine-tuned LLM that jointly forecasts near-term item popularity and generates faithful explanations for non-personalized ranking. By integrating historical trend, current momentum, and item metadata through contrastive learning, TRAIL aligns explanations with trend signals while achieving strong predictive accuracy. Empirical results on three real-world datasets show TRAIL outperforms strong baselines in both popularity prediction and top-N ranking, and human+LLM evaluations confirm the quality and usefulness of its explanations. The approach enables scalable, interpretable recommendations suitable for large catalogs and cold-start scenarios, highlighting the practical impact of combining trend forecasting with natural-language justifications in recommender systems.

Abstract

Large Language Models (LLMs) have been widely applied across multiple domains for their broad knowledge and strong reasoning capabilities. However, applying them to recommendation systems is challenging since it is hard for LLMs to extract user preferences from large, sparse user-item logs, and real-time per-user ranking over the full catalog is too time-consuming to be practical. Moreover, many existing recommender systems focus solely on ranking items while overlooking explanations, which could help improve predictive accuracy and make recommendations more convincing to users. Inspired by recent works that achieve strong recommendation performance by forecasting near-term item popularity, we propose TRAIL (TRend and explAnation Integrated Learner). TRAIL is a fine-tuned LLM that jointly predicts short-term item popularity and generates faithful natural-language explanations. It employs contrastive learning with positive and negative pairs to align its scores and explanations with structured trend signals, yielding accurate and explainable popularity predictions. Extensive experiments show that TRAIL outperforms strong baselines and produces coherent, well-grounded explanations.
Paper Structure (37 sections, 10 equations, 3 figures, 7 tables)

This paper contains 37 sections, 10 equations, 3 figures, 7 tables.

Table of Contents

  1. Introduction
  2. Related Works
  3. Non-Personalized Recommendation
  4. LLM-based Recommendation
  5. Explainable Recommendation
  6. Problem Definition
  7. TRAIL
  8. Similarity Functions
  9. Historical Trend Similarity.
  10. Latest/Current Popularity Levels.
  11. Meta Data Similarity.
  12. Total Similarity Function.
  13. Learning Objectives
  14. Contrastive Loss
  15. Supervised Loss.
  16. ...and 22 more sections

Figures (3)

  • Figure 1: The architecture of TRAIL.
  • Figure 2: (a) Systematic LLM-based evaluation of explanation quality for TRAIL and the the original LLM across four dimensions. (b) Human evaluation of explanation quality for TRAIL and the original LLM baseline without finetuning. (c) Impact of LoRA rank on recommendation accuracy.
  • Figure 3: (a) Performances of TRAIL and SASRec under cold-start and warm-start settings. (b) HR@10 performance across four types of trend categories.