Abstract
Predicting content diffusion in Device-to-Device (D2D) social networks is a significant challenge due to the opportunistic and offline nature of these networks. This paper introduces D2D-LSTM, a deep learning framework designed to predict the entire propagation tree of large files—such as mobile applications and videos—within D2D networks. By clustering users into representative "prototypes" and transforming complex tree structures into sequences for Long Short-Term Memory (LSTM) networks, the model achieves a 95% precision in identifying terminal nodes in sharing chains. Our findings highlight that geographical proximity is a more critical driver of content diffusion in D2D networks than content preference itself.
Introduction
In modern mobile networks, a small fraction of popular files often accounts for the majority of data traffic. To alleviate network congestion, Device-to-Device (D2D) communication—sharing files directly via Bluetooth or WiFi-Direct—has emerged as a promising solution. However, understanding and predicting how content diffuses in these networks is difficult. Unlike Online Social Networks (OSNs) where the social graph is explicit, D2D networks are opportunistic; the topology is invisible and changes constantly as users move.
Existing research has largely focused on link prediction (predicting if one user will share with another), but this fails to capture the entire diffusion path. Understanding the full tree structure is essential for network operators to design effective content caching and offloading strategies.
Methodology
The D2D-LSTM framework addresses the complexity of offline sharing through a three-stage process:
1. User Prototyping
Predicting the behavior of tens of millions of individual users is computationally prohibitive and prone to noise. To solve this, the authors employ K-means clustering to group users into 80 distinct "prototype users." These prototypes are defined by features such as content preferences, geographical patterns, and temporal activity. This abstraction allows the model to generalize behavior across the network.
2. Tree-to-Sequence Transformation
While LSTMs are optimized for sequential data, diffusion paths are inherently tree-like (one sender to multiple receivers). The authors utilize Breadth-First Search (BFS) traversal to flatten these trees into sequences. To signal the end of a branch, "Virtual Nodes" (V-NODEs) are introduced, allowing the LSTM to learn the specific conditions under which a diffusion path terminates.
3. D2D-LSTM Architecture
The core model utilizes an LSTM cell that processes multivariate features, including content category, time, geography, and transmission preferences. At each step $t$, the model concatenates the feature vectors of the parent node and the previous sibling node to predict the next recipient in the diffusion tree.
Dataset and Experimental Setup
The study is grounded in a massive, real-world dataset from a global D2D file-sharing application:
- Scale: 90GB of trace data containing 4.4 billion records.
- Scope: 30.4 million users and 16.7 million unique files over three months (August to October 2016).
- Enrichment: The authors scraped Apkpure to categorize shared applications into 48 distinct types, providing a granular view of user content preferences.
Key Insights
- Geography Dominates Content: A primary finding is that in D2D networks, location is a more significant predictor than content type. Because D2D requires physical proximity, the homophily of shared location drives sharing behavior even among users with divergent interests.
- Generalization via Prototypes: By focusing on user "types" rather than individuals, the model can predict diffusion paths for new users it has not previously encountered, provided they align with an established prototype.
- Terminal Node Precision: The model excels at identifying "leaves"—the points where a file stops spreading. This "terminal precision" is vital for optimizing cache storage, as it identifies when a file no longer needs to be prioritized for local sharing.
Results and Performance
Compared against standard RNNs and Fully Connected (FC) baselines, D2D-LSTM demonstrated superior performance:
- Accuracy: A 39.2% increase in mean Average Precision (mAP) and an 88% improvement in tree hit rate over state-of-the-art methods.
- Precision: The model achieved 95% average precision in identifying terminal nodes.
- Ablation Study: Removing the "Location" feature resulted in a significantly steeper performance drop than removing "Content Category," confirming the dominance of geography in offline social dynamics.
Conclusion and Future Work
D2D-LSTM represents a significant advancement in modeling offline content diffusion. While highly effective, the authors note that early errors in sequential prediction can propagate through the tree. Future research may explore dynamic clustering techniques to allow user prototypes to evolve over time and investigate the implementation of these models in live, decentralized networks where global information is not readily available.
Summary updated