Recommender systems are widely applied in numerous online platforms such as shopping and social media platforms. They typically utilize large embedding tables that map users and items to dense vectors of uniform sizes. As the number of users and items continues to grow, this design leads to significant memory consumption and computational inefficiencies. This challenge is particularly pronounced in scenarios such as federated learning, where model parameters are updated locally on edge devices with limited computational resources before being transmitted to a central server for aggregation. Numerous approaches have been proposed to address this issue, among which embedding pruning methods have emerged as a compelling solution. Compared to parameter-sharing and variable-size embedding techniques, embedding pruning methods offer lower training costs and leverage sparse embeddings for improved efficiency. Notably, embedding pruning methods based on the Dynamic Sparse Training (DST) paradigm maintain consistent sparsity throughout training and provide a controllable memory budget, establishing them as state-of-the-art lightweight embedding solutions for resource-constrained environments. However, embedding pruning methods are not without limitations. First, despite the use of sparse embeddings during forward passes, dense gradients are still computed in backward passes, introducing inefficiencies. Second, DST’s weight exploration mechanism tends to prioritize users or items from the most recent batch, reactivating pruned parameters that do not necessarily enhance overall performance. In this work, we introduce SparseRec, a lightweight embedding method designed to overcome these obstacles. SparseRec accumulates gradients to better identify inactive parameters that, when reactivated, contribute more meaningfully to model performance. Additionally, SparseRec avoids dense gradient computation during backpropagation by selectively sampling key vectors. Gradients are calculated only for parameters in this subset, ensuring sparsity throughout both forward and backward passes. Experiments on three benchmark datasets show that SparseRec achieves up to 11.79% performance gains across three base recommenders and multiple density configurations, highlighting its effectiveness in optimizing memory-constrained recommendation systems.

Streaming recommender systems (SRSs) are widely deployed in real-world applications, where user interests shift and new items arrive over time. As a result, effectively capturing users' latest preferences is challenging, as interactions reflecting recent interests are limited and new items often lack sufficient feedback. A common solution is to enrich item representations using multimodal encoders (e.g., BERT or ViT) to extract visual and textual features. However, these encoders are pretrained on general-purpose tasks: they are not tailored to user preference modeling, and they overlook the fact that user tastes toward modality-specific features such as visual styles and textual tones can also drift over time. This presents two key challenges in streaming scenarios: the high cost of fine-tuning large multimodal encoders, and the risk of forgetting long-term user preferences due to continuous model updates. To tackle these challenges, we propose Expandable Side Mixture-of-Experts (XSMoE), a memory-efficient framework for multimodal streaming recommendation. XSMoE attaches lightweight side-tuning modules consisting of expandable expert networks to frozen pretrained encoders and incrementally expands them in response to evolving user feedback. A gating router dynamically combines expert and backbone outputs, while a utilization-based pruning strategy maintains model compactness. By learning new patterns through expandable experts without overwriting previously acquired knowledge, XSMoE effectively captures both cold start and shifting preferences in multimodal features. Experiments on three real-world datasets demonstrate that XSMoE outperforms state-of-the-art baselines in both recommendation quality and computational efficiency.

Recommender systems typically represent users and items by learning their embeddings, which are usually set to uniform dimensions and dominate the model parameters. However, real-world recommender systems often operate in streaming recommendation scenarios, where the number of users and items continues to grow, leading to substantial storage resource consumption for these embeddings. Although a few methods attempt to mitigate this by employing embedding size search strategies to assign different embedding dimensions in streaming recommendations, they assume that the embedding size grows with the frequency of users/items, which eventually still exceeds the predefined memory budget over time. To address this issue, this paper proposes to learn Scalable Lightweight Embeddings for streaming recommendation, called SCALL, which can adaptively adjust the embedding sizes of users/items within a given memory budget over time. Specifically, we propose to sample embedding sizes from a probabilistic distribution, with the guarantee to meet any predefined memory budget. By fixing the memory budget, the proposed embedding size sampling strategy can increase and decrease the embedding sizes in accordance to the frequency of the corresponding users or items. Furthermore, we develop a reinforcement learning-based search paradigm that models each state with mean pooling to keep the length of the state vectors fixed, invariant to the changing number of users and items. As a result, the proposed method can provide embedding sizes to unseen users and items. Comprehensive empirical evaluations on two public datasets affirm the advantageous effectiveness of our proposed method.

At the heart of contemporary recommender systems (RSs) are latent factor models that provide quality recommendation experience to users. These models use embedding vectors, which are typically of a uniform and fixed size, to represent users and items. As the number of users and items continues to grow, this design becomes inefficient and hard to scale. Recent lightweight embedding methods have enabled different users and items to have diverse embedding sizes, but are commonly subject to two major drawbacks. Firstly, they limit the embedding size search to optimizing a heuristic balancing the recommendation quality and the memory complexity, where the trade-off coefficient needs to be manually tuned for every memory budget requested. The implicitly enforced memory complexity term can even fail to cap the parameter usage, making the resultant embedding table fail to meet the memory budget strictly. Secondly, most solutions, especially reinforcement learning based ones derive and optimize the embedding size for each each user/item on an instance-by-instance basis, which impedes the search efficiency. In this paper, we propose Budgeted Embedding Table (BET), a novel method that generates table-level actions (i.e., embedding sizes for all users and items) that is guaranteed to meet pre-specified memory budgets. Furthermore, by leveraging a set-based action formulation and engaging set representation learning, we present an innovative action search strategy powered by an action fitness predictor that efficiently evaluates each table-level action. Experiments have shown state-of-the-art performance on two real-world datasets when BET is paired with three popular recommender models under different memory budgets.

Latent factor models are the most popular backbones for today's recommender systems owing to their prominent performance. Latent factor models represent users and items as real-valued embedding vectors for pairwise similarity computation, and all embeddings are traditionally restricted to a uniform size that is relatively large (e.g., 256-dimensional). With the exponentially expanding user base and item catalog in contemporary e commerce, this design is admittedly becoming memory-inefficient. To facilitate lightweight recommendation, reinforcement learning (RL) has recently opened up opportunities for identifying varying embedding sizes for different users/items. However, challenged by search efficiency and learning an optimal RL policy, existing RL-based methods are restricted to highly discrete, predefined embedding size choices. This leads to a largely overlooked potential of introducing finer granularity into embedding sizes to obtain better recommendation effectiveness under a given memory budget. In this paper, we propose continuous input embedding size search (CIESS), a novel RL-based method that operates on a continuous search space with arbitrary embedding sizes to choose from. In CIESS, we further present an innovative random walk-based exploration strategy to allow the RL policy to efficiently explore more candidate embedding sizes and converge to a better decision. CIESS is also model-agnostic and hence generalizable to a variety of latent factor RSs, whilst experiments on two real-world datasets have shown state-of-the-art performance of CIESS under different memory budgets when paired with three popular recommendation models.