With the increasing demand for personalization in modern digital ecosystems, Recommender Systems (RS) have become essential for delivering user-centric services. However, traditional RS approaches, which are primarily driven by algorithmic optimization over historical interaction data, still face well-known challenges such as the cold-start problem, limited interpretability, and insufficient personalization quality. Meanwhile, the rapid evolution of Large Language Models (LLMs) has showcased remarkable capabilities in semantic understanding, reasoning, and generalization, presenting a promising opportunity to rethink the design of recommendation paradigms. A central research question is how to systematically leverage LLMs to address fundamental RS challenges while avoiding unnecessary complexity, inefficiency, or unintended consequences. In this paper, we provide the first problem-driven synthesis of LLM-empowered RS. We propose a comprehensive taxonomy that organizes existing integration efforts into four major problem domains: Cold-start problem, poor interpretability, suboptimal user experience, and novel challenges introduced by LLMs themselves (e.g., hallucination, bias, and inefficiency). We begin by formally characterizing each problem domain and reviewing its significance in recommendation research. We then categorize recent advances by methodological perspective. Finally, we offer insights and future directions for building adaptive, efficient, and value-aligned RS enhanced by LLM capabilities.

In the Internet of Things (IoT) environment, user-service interaction data are often stored in multiple distributed platforms. In this situation, recommender systems need to integrate the distributed user-service interaction data across different platforms for making a comprehensive recommendation decision, during which user privacy is probably disclosed. Moreover, as user-service interaction records accumulate over time, they significantly reduce the efficiency of recommendations. To tackle these issues, we propose a lightweight and privacy-preserving service recommendation approach named SerRec_L2H. In SerRec_L2H, we employ Learning to Hash (L2H) to encapsulate sensitive user-service interaction data into less-sensitive user indices, which facilitates identifying users with similar preferences efficiently for accurate recommendations. We then validate the feasibility of our proposed SerRec_L2H approach through massive experiments conducted on the popular WS-DREAM dataset. The comparative analysis with other competitive approaches demonstrates that our proposal surpasses other approaches in terms ofrecommendation accuracy and efficiency while protecting user privacy.