From enhancement to exploitation The dual role of LLMs in recommender systems

Recommender systems (RSs) are essential for delivering personalized suggestions across digital platforms, yet conventional neural models often lack commonsense knowledge, limiting their ability to provide contextually relevant recommendations. Large language models (LLMs), with their advanced commonsense reasoning and text understanding capabilities, offer significant potential to enhance RSs. However, the misalignment between language-oriented tasks (e.g., text generation) and recommendation tasks (e.g., user preference modeling) introduces challenges such as hallucinations (irrelevant or incorrect outputs) and under-representation (discrepancies between user behavior and item semantics). These issues are particularly pronounced in text-rich domains like news recommendation, where reliance on textual content also exposes vulnerabilities to adversarial textual attacks that manipulate recommendation rankings.
This thesis investigates the dual role of LLMs in RSs, focusing on: (1) enhancing recommendation effectiveness by mitigating hallucinations and under-representation, and (2) exposing vulnerabilities through adversarial textual attacks, with an emphasis on preserving media bias orientation. On the enhancement side, we propose frameworks like ToolRec, which leverages LLMs as surrogate users to simulate human-like decision-making, and fine-tuning strategies to align LLMs with user preferences in news recommendation. On the exploitation side, we introduce LANCE and BALANCE, two frameworks that exploit LLMs to craft textual attacks, with BALANCE specifically designed to maintain media bias orientation for stealthier manipulations. Through these contributions, the thesis advances the understanding of LLMs’ potential to improve RSs while highlighting the need for robust defenses to ensure their responsible use.