DIET: Customized Slimming for Incompatible Networks in Sequential Recommendation

TL;DR

The paper addresses edge-device limitations in sequential recommendation by proposing DIET, a slimming framework that learns edge-specific diets (binary masks) via a sequence extractor and hypernetworks to personalize subnets for each edge. It introduces a connections-correction module to exploit inter-layer dependencies, improving accuracy and reducing inference cost, and further proposes DIETING to represent the entire model with a single layer, lowering storage and transmission needs. The approach yields superior accuracy and efficiency compared to multiple baselines across four real-world datasets and two model families, while demonstrating robustness to distribution shifts and offering clear practical benefits for edge-enabled recommendations. Overall, DIET enables fast, bandwidth-efficient, and storage-conscious edge personalization for sequential recommenders, facilitating scalable edge-cloud collaboration.

Abstract

Due to the continuously improving capabilities of mobile edges, recommender systems start to deploy models on edges to alleviate network congestion caused by frequent mobile requests. Several studies have leveraged the proximity of edge-side to real-time data, fine-tuning them to create edge-specific models. Despite their significant progress, these methods require substantial on-edge computational resources and frequent network transfers to keep the model up to date. The former may disrupt other processes on the edge to acquire computational resources, while the latter consumes network bandwidth, leading to a decrease in user satisfaction. In response to these challenges, we propose a customizeD slImming framework for incompatiblE neTworks(DIET). DIET deploys the same generic backbone (potentially incompatible for a specific edge) to all devices. To minimize frequent bandwidth usage and storage consumption in personalization, DIET tailors specific subnets for each edge based on its past interactions, learning to generate slimming subnets(diets) within incompatible networks for efficient transfer. It also takes the inter-layer relationships into account, empirically reducing inference time while obtaining more suitable diets. We further explore the repeated modules within networks and propose a more storage-efficient framework, DIETING, which utilizes a single layer of parameters to represent the entire network, achieving comparably excellent performance. The experiments across four state-of-the-art datasets and two widely used models demonstrate the superior accuracy in recommendation and efficiency in transmission and storage of our framework.

Figures (6)

• Figure 1: (a): There always exists data distribution shifts across cloud and edges, where cloud holds massive user historical interaction data and each edge only accesses their own data. (b): In each edge, user's interest would change frequently due to some other factors. This would cause cloud to send the latest models to each edge for adaptation, causing massive transmission delays. (c): In recommender systems, different scenarios need different models to provide services, which makes that user devices are flooded with models. (d): Due to the difference in computing resources between cloud and edges, although the model can quickly complete gradient updates and inference on cloud, it still takes a long time on edges.
• Figure 2: Overview of DIET. (a): Cloud will generate edge-specific diets(subnets) condition on real-time interactions for various data distribution on edges. (b): Another module on cloud to discover inter-layer filter relationships in networks, generated filter-level importance will post-process diets from (a) to correct less important connections. (c): A brief schematic diagram of DIET and DIETING(the portion marked with orange dashed lines). The edges send their real-time samples and cloud customizes diets for them. The transmission over the network consists of a series of binary masks, with all edges sharing the same network. (d): With specific diets from cloud, edges construct their own models, enabling fast adaptation.
• Figure 3: The variation on the test set during training.
• Figure 4: Sensitivity of each model towards $\alpha$.
• Figure 5: Emperical study of the influence caused by connection correlation. We calculate the percent of non-zero rows/filters in each layer before and after correction.