Explainable CTR Prediction via LLM Reasoning

Abstract: Recommendation Systems have become integral to modern user experiences, but lack transparency in their decision-making processes. Existing explainable recommendation methods are hindered by reliance on a post-hoc paradigm, wherein explanation generators are trained independently of the underlying recommender models. This paradigm necessitates substantial human effort in data construction and raises concerns about explanation reliability. In this paper, we present ExpCTR, a novel framework that integrates LLM based explanation generation directly into the CTR prediction process. Inspired by recent advances in reinforcement learning, we employ two carefully designed reward mechanisms, LC alignment, which ensures explanations reflect user intentions, and IC alignment, which maintains consistency with traditional ID-based CTR models. Our approach incorporates an efficient training paradigm with LoRA and a three-stage iterative process. ExpCTR circumvents the need for extensive explanation datasets while fostering synergy between CTR prediction and explanation generation. Experimental results demonstrate that ExpCTR significantly enhances both recommendation accuracy and interpretability across three real-world datasets.

• The paper introduces ExpCTR, embedding LLM reasoning directly into CTR models to generate coherent, user-aligned explanations for enhanced predictions.
• It employs reinforcement learning via Proximal Policy Optimization with dual reward mechanisms (LC and IC alignment) to improve both accuracy and interpretability.
• ExpCTR leverages lightweight fine-tuning with LoRA and dynamic adaptation to outperform state-of-the-art models on multiple datasets.

Explainable CTR Prediction via LLM Reasoning

The paper "Explainable CTR Prediction via LLM Reasoning" presents an ambitious approach, ExpCTR, for enhancing the explanatory power and accuracy of click-through rate (CTR) predictions. This is achieved by integrating LLMs into the recommendation process for concurrent prediction and explanation generation. While previous methods have relied on post-hoc explanations, ExpCTR innovatively embeds the explanatory process directly within the CTR prediction mechanism, thus addressing issues related to data construction and explanation reliability.

In recent years, recommendation systems have predominantly operated as opaque systems ("black boxes"), focusing on accuracy rather than transparency. The ExpCTR framework is motivated by two pivotal enhancements: (1) utilizing LLMs' reasoning capabilities for generating coherent explanations that align with user interactions, and (2) incorporating these explanations as integral components of CTR models to foster accurate predictions. This dual thrust is accomplished through reinforcement learning techniques, specifically Proximal Policy Optimization (PPO), to fine-tune the LLMs for this purpose.

ExpCTR employs two main reward mechanisms—LC alignment (user-centered) and IC alignment (ID-based CTR consistency). The LC alignment ensures that generated explanations align with user intentions, thereby promoting user trust and interpretability. For IC alignment, the explanations are evaluated against traditional ID-based CTR models, ensuring that explanations match underlying model decisions.

One of the major revelations noted in the paper is how ExpCTR circumvents the need for widespread data construction by leveraging LLMs’ capacity for reasoning. This approach eliminates traditionally resource-intensive practices of preparing explanation datasets and permits dynamic adaptation to recommendation contexts. Furthermore, ExpCTR integrates LoRA, a lightweight fine-tuning mechanism, to efficiently train the involved LLMs, significantly reducing the computational overhead typically associated with such large-scale models.

The results on three substantial datasets—BookCrossing, MovieLens-20M, and Amazon Books—demonstrated that ExpCTR outperforms several state-of-the-art models not only in terms of recommendation accuracy, but also in generating qualitatively superior, human-readable explanations. ExpCTR yielded significant improvements in AUC and other CTR performance metrics, reinforcing the notion that explainable AI does not merely serve transparency demands but can enhance accuracy.

From a theoretical perspective, ExpCTR advances the nature of recommendation systems by bridging the gap between prediction tasks and their explanations. It theorizes that, by incorporating explanations in the training model, one can achieve a system that is inherently aligned with both user motivations and recommender logic. The reward-based strategy ensures that generated explanations maintain consistency with model predictions while also being interpretable for end-users.

Looking forward, this research sets forth a foundation for more intuitive, comprehensible, and accurate recommendation systems. By paving the way for direct integration of explanation in model training, it stimulates discussions around improving transparency in AI, indicating potential explorations into explainable AI's role in user satisfaction and decision-making processes. Future work could expand on this foundation by considering real-time user feedback integration or experiment with different architectures of LLMs to further optimize this holistic model-explanation synergy. ExpCTR’s reinforcement learning-based framework is likely to inspire new methodologies in creating intuitive, reliable, and transparent recommender systems.