Abstract
SimSGR is a streamlined self-supervised learning (SSL) framework designed for edge caching networks to predict content popularity under data scarcity. By replacing destructive random data augmentation with a "Mixing" layer and a "Conversion" layer, it achieves state-of-the-art recommendation accuracy with significantly lower computational overhead than traditional SSL methods.
Introduction
In Mobile Edge Computing (MEC), base stations proactively cache popular content to reduce latency and backhaul traffic. While Deep Learning models are effective at predicting user preferences based on history, they face the cold-start problem: when a new user enters a cell or a new base station is deployed, historical data is insufficient for accurate predictions.
Limitations of Current SSL Methods
Researchers have attempted to use Self-Supervised Learning (SSL) to bridge this gap by generating "pseudo-labels" from the data itself. However, existing graph-based SSL methods suffer from two major flaws:
- Semantic Destruction: Standard SSL often uses random "augmentations" like dropping nodes or edges. Unlike image processing, where a cropped cat is still a cat, removing a single edge in a user-item graph can fundamentally change the data's meaning (e.g., removing a movie a user liked).
- Resource Intensity: These models typically require complex joint training and negative sampling, which are computationally expensive for resource-constrained edge devices.
The SimSGR Framework
The authors propose SimSGR (Simple Self-supervised Graph-based Recommendation), introducing two novel layers to the standard graph convolution pipeline (using LightGCN as a backbone):
The Mixing Layer
Instead of randomly breaking the graph, this layer creates a second "view" of a node by aggregating its most similar neighbors. It uses cosine similarity to filter out "noisy" neighbors (approximately 30% of a node's neighbors), preserving the "semantic soul" of the graph while providing the variety needed for SSL.
The Conversion Layer
Traditional SSL compares abstract vector embeddings. SimSGR’s Conversion layer projects these embeddings into the actual rating space (predicted scores) before calculating loss. This unifies the self-supervised task with the actual recommendation task.
Invariance-Covariance (IC) Loss
Inspired by vision models like Barlow Twins, the IC loss function encourages the two "views" to be identical (invariance) while ensuring the model doesn't collapse into predicting the same output for all users (covariance regularization). This eliminates the need for memory-heavy negative sampling.
Experimental Results
SimSGR was evaluated on three benchmarks: Douban-book, MovieLens-1M, and Yelp2018. It consistently outperformed state-of-the-art baselines like SGL and SimGCL.
- Accuracy Boost: On the Douban-book dataset, SimSGR improved the Recall metric by 36.78% and NDCG by 36.70% compared to the standard LightGCN.
- Cold-Start Resilience: In simulated cold-start environments (5-10 user interactions), SimSGR’s performance remained stable, while supervised models saw drops of nearly 20%.
- Efficiency: By avoiding graph augmentation and negative sampling, SimSGR achieves a lower time complexity (O(Bd + Bd^2)) than competitors like SGL (O(Bd + BMd)).
Conclusion
SimSGR demonstrates that augmentation is not one-size-fits-all. By mixing similar nodes and unifying tasks in the rating space, it provides a robust, efficient solution for edge caching. Future work will focus on performance on actual edge hardware and handling dynamic graphs where user preferences evolve rapidly over time.
Summary updated