[Paper Review] A Machine Learning Approach to Predicting Continuous Tie Strengths

James Flamino, Ross DeVito+2 more|arXiv (Cornell University)|Jan 23, 2021

Mechanical stress and fatigue analysis24 references4 citations

TL;DR

This paper proposes a machine learning framework to continuously predict evolving tie strengths in dynamic social networks using communication data from the NetSense study. By applying random forest and LSTM models to time-series communication features—frequency, recency, duration, and volume—it generates dynamic, continuous tie strength estimates that outperform static surveys and reveal longitudinal relationship patterns.

ABSTRACT

Relationships between people constantly evolve, altering interpersonal behavior and defining social groups. Relationships between nodes in social networks can be represented by a tie strength, often empirically assessed using surveys. While this is effective for taking static snapshots of relationships, such methods are difficult to scale to dynamic networks. In this paper, we propose a system that allows for the continuous approximation of relationships as they evolve over time. We evaluate this system using the NetSense study, which provides comprehensive communication records of students at the University of Notre Dame over the course of four years. These records are complemented by semesterly ego network surveys, which provide discrete samples over time of each participant's true social tie strength with others. We develop a pair of powerful machine learning models (complemented by a suite of baselines extracted from past works) that learn from these surveys to interpret the communications records as signals. These signals represent dynamic tie strengths, accurately recording the evolution of relationships between the individuals in our social networks. With these evolving tie values, we are able to make several empirically derived observations which we compare to past works.

Motivation & Objective

• To develop a scalable, continuous method for predicting evolving social tie strengths in dynamic networks, overcoming limitations of static survey-based approaches.
• To model tie strength as a time-varying signal derived from communication patterns rather than discrete snapshots.
• To create a generalizable framework applicable beyond single platforms by using communication metadata instead of platform-specific features.
• To validate the model against ground-truth survey data from the NetSense study, ensuring empirical relevance.
• To enable longitudinal analysis of relationship evolution by generating time-series edge weights from raw communication logs.

Proposed method

• Uses a Borda count-based ranking system to aggregate pairwise comparisons from machine learning models into a global ranking of social tie strength for each ego.
• Employs an Ensemble model with a random forest classifier trained on difference vectors of communication features (frequency, recency, duration, volume) between two individuals.
• Applies an LSTM model that processes two-channel time series of call and text activity, binned into 21-day intervals, to capture temporal dynamics.
• Computes pairwise comparison probabilities $ M_{ij} $ representing the likelihood that individual $ i $ has a stronger tie to the ego than individual $ j $, with $ M_{ii} = 0.5 $.
• Transforms comparison matrices into win/loss/tie indicators via Eq. 2 and computes Borda scores using Eq. 3 to generate final rankings.
• Converts Borda counts into winning percentages (Eq. 4) to produce continuous, time-varying edge weights representing dynamic tie strengths.

Experimental results

Research questions

• RQ1 Can machine learning models accurately predict continuous tie strengths from communication metadata in a dynamic social network?
• RQ2 How do different communication features—frequency, recency, duration, volume—contribute to predicting evolving relationship strength?
• RQ3 To what extent do time-series patterns in communication (e.g., call and text histories) improve tie strength prediction compared to static features?
• RQ4 How do the predicted dynamic tie strengths compare to ground-truth survey data collected over multiple semesters?
• RQ5 Can the proposed framework generalize beyond platform-specific features to enable scalable, real-time relationship tracking in large networks?

Key findings

• The proposed machine learning models, particularly the LSTM and Ensemble models, successfully generate continuous, time-varying tie strength estimates that closely track ground-truth survey data.
• The inclusion of Bow Tie Overlap as a feature significantly degraded model performance and was therefore excluded from the final Ensemble model.
• The Borda count-based ranking system effectively aggregates noisy pairwise comparisons into stable, interpretable rankings of social tie strength.
• The winning percentage transformation (Eq. 4) produces a smooth, continuous time series of dynamic edge weights suitable for network analysis.
• The models demonstrate that communication frequency, recency, duration, and volume are strong predictors of evolving tie strength, even without explicit sentiment or content analysis.
• The framework enables longitudinal observation of relationship evolution, revealing patterns such as strengthening, weakening, or stagnation of ties over time, which static surveys cannot capture.