Self-Monitoring Large Language Models for Click-Through Rate Prediction
Keywords: Large Language Model, Click-through Rate Prediction, Feature-Click learning
Abstract: Click-through rate (CTR) prediction tasks traditionally aim to model extensive user-
item feature interactions. Recent approaches fine-tune Large Language Models
(LLMs) using user-item features as input and click labels as output. However,
due to the sparsity of click labels, the attention mechanism may focus on a subset
of features rather than all features. This can hinder LLMs’ ability to accurately
match features to click labels, resulting in performance that does not consistently
exceed traditional state-of-the-art CTR approaches. To address this, we introduce
a SLLM4CTR framework which uses adaptive temperature and label matching loss
to improve fine-tuning and inference process of LLMs. The adaptive temperature
serves as a confidence score to calibrate CTR predictions by quantifying the LLMs’
attention to user-item features. The label matching loss clearly distinguish between
click-inducing and non-click-inducing features by constraining the representation
space of click labels. By combining these two designs, SLLM4CTR improves feature
utilization in LLMs and enhances the matching of user-item features to click
labels. Experimental results demonstrate
that SLLM4CTR significantly outperforms state-of-the-art baselines, including both
traditional and LLM-based CTR approaches. The code will be open-sourced.
Primary Area: applications to computer vision, audio, language, and other modalities
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.
Reciprocal Reviewing: I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 5458
Loading