Abstract
Under the explosive growth of information available on the Web, recommender systems have been used as an effective technology to filter useless information and attempt to recommend the most useful items. The proliferation of smart phones, smart wearable devices and other Internet of Thing (IoT) devices has gradually driven many novel emerging services which are latency-sensitive and computation-intensive with a higher quality-of-service. Under such circumstances, the data sources contain four key characteristics (i.e., sparsity, heterogeneity, mobility, volatility). The conventional recommender systems based on cloud computing are incapable of digging the information of user demands. Mobile edge computing is a novel computing paradigm via pushing computation/storage resource from the remote cloud servers to the network edge servers to provide more intelligent and personalized service. This paper comprehensively reviews the state of the art literature on the convergence of recommender systems and edge computing, and identify the future directions along this dimension.
The Problem: Cloud-Based Limitations in an IoT World
Traditional recommender systems (RS), such as those powering Netflix or Amazon, have historically relied on centralized cloud architectures. However, the rapid expansion of the Internet of Things (IoT)—encompassing smartphones, wearables, and connected vehicles—has introduced a data environment that challenges traditional cloud models. The authors identify four "key characteristics" of edge data that necessitate a shift in paradigm:
- Sparsity: Data at a single edge node is often limited to a few local users, making it difficult to build robust global models.
- Heterogeneity: Devices from different manufacturers use varying protocols and data formats.
- Mobility: Users frequently move between base stations, causing their data sources and network connections to be highly volatile.
- Volatility: Fluctuating network conditions and Quality of Service (QoS) metrics impact real-time performance.
Furthermore, cloud-based RS suffers from inherent high latency, which is detrimental to real-time applications like Augmented Reality (AR) or autonomous driving. Privacy also remains a critical concern, as sensitive user data must typically be uploaded to a central server for processing.
A Framework for Convergence
To address these challenges, the survey categorizes the convergence of RS and Mobile Edge Computing (MEC) into a framework of four enabling technologies. This framework highlights the symbiotic relationship where MEC provides the infrastructure and RS provides the intelligence.
1. RS Applications on Edge
This involves deploying recommendation logic for specific edge use cases. Examples include mobile video streaming optimization, navigating Electric Vehicle (EV) charging stations, and providing location-based web services that require immediate response times.
2. RS Inference in Edge
Techniques here focus on optimizing the computational "thinking" of the RS. This includes partial offloading, where specific layers of a neural network model run on the user device while others run on the edge server, and the use of stacked autoencoders to drastically reduce training time.
3. Edge Computing for RS
This dimension uses edge infrastructure to solve traditional RS problems. For instance, the physical proximity of edge servers can be used to alleviate user mobility issues. Additionally, techniques like Locality-Sensitive Hashing (LSH) allow for local data processing, which significantly enhances user privacy by keeping raw data within the local network.
4. RS for Edge Computing
Conversely, recommendation algorithms can be used to optimize the edge network itself. A primary application is "recommending" which content should be cached at specific edge nodes based on predicted user demand, utilizing both human-driven and social-driven strategies.
Architectural Synergy
The integration of these technologies creates a hierarchical architecture that bridges the gap between remote recommendation servers and local city-level infrastructure. This hierarchy allows for a more fluid exchange of data and computation.
Key Insights and Performance Gains
The survey reveals that the relationship between RS and MEC is a "two-way street." MEC is not merely a faster version of the cloud; it is a platform that enables RS to become context-aware. In turn, RS acts as a management tool for MEC, helping to allocate scarce storage and compute resources efficiently.
- Privacy via Decentralization: By utilizing distributed hashing at the edge, systems can provide high personalization without raw data ever leaving the local network.
- Mobility Prediction: Converged systems use RS to predict a user's trajectory, allowing the network to "pre-warm" caches at the next base station the user is likely to encounter.
- The Caching Conflict: There is a natural tension between edge caching (favoring global popularity) and RS (favoring individual relevance). The survey suggests a "soft cache hit" approach to balance these competing needs.
Quantitative results from the literature show significant improvements. Partial offloading using stacked autoencoders can lead to a dramatic reduction in training time compared to traditional matrix factorization. Furthermore, edge-based multimedia recommendations significantly reduce transmission delays, and incorporating local geographic characteristics improves the precision of service recommendations.
Future Directions
Despite the progress, several research gaps remain:
- Deep Learning Integration: There is a need for more research on applying advanced Deep Learning (DL) architectures specifically to the RS-MEC convergence.
- Distributed Scalability: Scaling distributed RS models to thousands of heterogeneous edge devices remains an open challenge.
- New Domains: Beyond mobile phones and EVs, Smart Homes and Smart Factories represent the next frontier for converged systems.
- Specialized Hardware: There is a growing requirement for hardware/software co-design to handle heavy RS inference on resource-constrained edge nodes.
Summary updated