Scalable Dynamic Embedding Size Search for Streaming Recommendation

TL;DR

SCALL tackles the memory bottleneck in streaming recommender systems by dynamically allocating embedding sizes under a fixed budget. It combines a base recommender with adjustable embeddings and a reinforcement-learning based embedding-size predictor trained via Soft Actor-Critic, using a fixed-length state built from mean-pooled frequency features. Embedding sizes are sampled from distributions (e.g., Power Law) to meet the memory budget $B$ or maintain a mean embedding size $c$, enabling adaptation to evolving user/item frequencies and unseen entities. Empirical results on Yelp and Amazon-Book show SCALL outperforms static and other dynamic baselines across budgets, while providing explicit budget enforcement and robust performance under streaming conditions.

Abstract

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.

Figures (5)

• Figure 1: An overview of SCALL.
• Figure 2: The total parameters are allocated to to 3 users and 3 items in this example. The size of each circle corresponds to the number of parameters, with a larger circle indicating a higher parameter count.
• Figure 3: Mean pooling is used to achieve fixed-length vectors incorporating frequency information. The users or items in this example are divided into 5 groups.
• Figure 4: Sensitivity analysis of $\gamma$ w.r.t. Recall@20 and NDCG@20 on the Yelp and Amazon-Book dataset.
• Figure 5: Case study on the interaction counts (left) and embedding sizes (right) of four users from the Yelp dataset.