Compressing Recommender and Ranking Models for Edge Deployment: A Survey

Abstract

Edge deployment of recommender and ranking models enables low-latency personalization, privacy-preserving inference, reduced network dependence, and context-aware decisions on mobile, embedded, and resource-constrained platforms. Yet contemporary recommenders combine massive embedding tables, interaction networks, sequential encoders, multi-stage retrieval and ranking components, and dynamically accessed features whose computational behavior differs fundamentally from conventional classifiers and language models. This survey reviews compression methods specifically for recommender and ranking architectures, focusing on quantization, pruning, knowledge distillation, embedding compression, low-rank factorization, hashing, and hardware-aware optimization. It develops a ranking-centric taxonomy that distinguishes techniques by the component compressed, compression granularity, training requirement, candidate-generation or ranking stage, and effect on ranking quality. The review evaluates how compression influences Recall, Hit Rate, Mean Reciprocal Rank, Normalized Discounted Cumulative Gain, area under the curve, calibration, diversity, fairness, and long-tail coverage, alongside model size, memory bandwidth, latency, throughput, energy consumption, and accelerator utilization. Particular attention is given to the irregular memory access, sparse features, dynamic candidate sets, and user-specific state that complicate deployment on CPUs, GPUs, neural accelerators, and mobile systems. The survey compares post-training and trainingaware methods, identifies interactions among compression techniques, and examines the limitations of accuracy-only evaluation. It also analyzes deployment case studies involving collaborative filtering, deep interaction models, sequential recommenders, graph recommenders, and learning-to-rank pipelines. The gap analysis reveals limited evidence on end-to-end multistage compression, user-level robustness, online adaptation, accelerator-aware embedding design, privacy, fairness preservation, and reproducible device benchmarking. It emphasizes workload-aware measurement under realistic catalog sizes, request patterns, batch sizes, and thermal conditions. The survey concludes with evaluation guidelines and research priorities for producing compact recommenders that retain ranking utility while satisfying practical edge constraints.