[Paper Review] Tensor-Based Link Prediction in Intermittently Connected Wireless Networks

Mohamed-Haykel Zayani, Vincent Gauthier+2 more|arXiv (Cornell University)|Aug 12, 2011

Opportunistic and Delay-Tolerant Networks28 references3 citations

TL;DR

This paper proposes a tensor-based link prediction method for intermittently connected wireless networks (DTNs) that models spatio-temporal node behavior as a third-order tensor to compute similarity via the Katz measure. The approach achieves high prediction accuracy—especially on real mobility traces—and maintains strong performance even when computed in a distributed manner using only local two-hop information.

ABSTRACT

Through several studies, it has been highlighted that mobility patterns in mobile networks are driven by human behaviors. This effect has been particularly observed in intermittently connected networks like DTN (Delay Tolerant Networks). Given that common social intentions generate similar human behavior, it is relevant to exploit this knowledge in the network protocols design, e.g. to identify the closeness degree between two nodes. In this paper, we propose a temporal link prediction technique for DTN which quantifies the behavior similarity between each pair of nodes and makes use of it to predict future links. Our prediction method keeps track of the spatio-temporal aspects of nodes behaviors organized as a third-order tensor that aims to records the evolution of the network topology. After collapsing the tensor information, we compute the degree of similarity for each pair of nodes using the Katz measure. This metric gives us an indication on the link occurrence between two nodes relying on their closeness. We show the efficiency of this method by applying it on three mobility traces: two real traces and one synthetic trace. Through several simulations, we demonstrate the effectiveness of the technique regarding another approach based on a similarity metric used in DTN. The validity of this method is proven when the computation of score is made in a distributed way (i.e. with local information). We attest that the tensor-based technique is effective for temporal link prediction applied to the intermittently connected networks. Furthermore, we think that this technique can go beyond the realm of DTN and we believe this can be further applied on every case of figure in which there is a need to derive the underlying social structure of a network of mobile users.

Motivation & Objective

• To improve link prediction in delay-tolerant networks (DTNs) by leveraging human mobility patterns driven by social intentions.
• To model the dynamic topology of intermittently connected networks using a time-evolving tensor structure capturing spatio-temporal interactions.
• To quantify node behavior similarity through the Katz measure to predict future link occurrences based on historical contact patterns.
• To evaluate the method’s performance in both centralized and distributed deployment scenarios, particularly under limited information exchange.
• To demonstrate that social behavior similarity, derived from correlated mobility patterns, is a strong predictor of future network contacts.

Proposed method

• Construct a third-order tensor from time-ordered adjacency matrices representing node contacts across T time periods, encoding spatio-temporal network topology.
• Collapse the tensor to extract structural information for each node pair, enabling comparison of their behavioral patterns over time.
• Apply the Katz measure to compute a similarity score between each node pair, where higher scores indicate stronger behavioral correlation and higher likelihood of future contact.
• Use the Katz measure’s property of capturing both direct and indirect connections through paths of varying lengths to reflect social closeness.
• Implement the prediction algorithm in a distributed fashion, where each node uses only local knowledge (up to two-hop neighbors) to compute its own prediction scores.
• Compare the proposed tensor-based method with a baseline similarity metric based on preferred locations only, without temporal correlation.

Experimental results

Research questions

• RQ1 Can modeling node behavior as a tensor of spatio-temporal contacts improve link prediction accuracy in DTNs compared to traditional similarity metrics?
• RQ2 How does the performance of the tensor-based prediction method vary across real-world mobility traces, synthetic traces, and different time window sizes?
• RQ3 To what extent can the link prediction be effectively performed in a distributed manner using only local two-hop information without global network knowledge?
• RQ4 How does the inclusion of temporal correlation in contact patterns (via the tensor) compare to methods relying only on spatial preferences (e.g., preferred locations)?
• RQ5 What is the impact of network structure—such as multi-hop vs. one-hop knowledge—on the prediction performance in terms of AUC and top scores ratio?

Key findings

• The tensor-based method achieves high prediction accuracy on real mobility traces (Dartmouth and MIT campus), with AUC values consistently above 0.85 and top scores ratios at T+1 exceeding 0.75 in favorable conditions.
• The method performs better on real traces than on synthetic traces (TVC model), where the lack of social correlation results in lower AUC (below 0.6) and poor top scores ratios.
• Prediction performance remains strong in a distributed setting: the method achieves comparable AUC and top scores ratios to the centralized version, validating its scalability and practicality.
• The inclusion of temporal correlation via the tensor model significantly outperforms a baseline method based solely on location similarity, as shown by higher AUC and better top scores ratios.
• The prediction of entirely new, previously unseen links (i.e., links not observed in training data) is inaccurate, with AUC values below 0.5 and top scores ratios near zero, indicating the method relies on historical behavioral patterns.
• Shorter tensor time windows (fewer recorded periods) yield better performance on real traces, suggesting that recent behavior is more predictive than long-term historical data.