Abstract
Recommender systems have become an essential component of many onlineplatforms, providing personalized recommendations to users. A crucial aspect isembedding techniques that convert the high-dimensional discrete features, suchas user and item IDs, into low-dimensional continuous vectors, which canenhance the recommendation performance. Embedding techniques haverevolutionized the capture of complex entity relationships, generatingsignificant research interest. This survey presents a comprehensive analysis ofrecent advances in recommender system embedding techniques. We examinecentralized embedding approaches across matrix, sequential, and graphstructures. In matrix-based scenarios, collaborative filtering generatesembeddings that effectively model user-item preferences, particularly in sparsedata environments. For sequential data, we explore various approaches includingrecurrent neural networks and self-supervised methods such as contrastive andgenerative learning. In graph-structured contexts, we analyze techniques likenode2vec that leverage network relationships, along with applicableself-supervised methods. Our survey addresses critical scalability challengesin embedding methods and explores innovative directions in recommender systems.We introduce emerging approaches, including AutoML, hashing techniques, andquantization methods, to enhance performance while reducing computationalcomplexity. Additionally, we examine the promising role of Large LanguageModels (LLMs) in embedding enhancement. Through detailed discussion of variousarchitectures and methodologies, this survey aims to provide a thoroughoverview of state-of-the-art embedding techniques in recommender systems, whilehighlighting key challenges and future research directions.