[Paper Review] On Temporal Regularity in Social Interactions: Predicting Mobile Phone Calls

Mehwish Nasim, Aimal Rextin+2 more|arXiv (Cornell University)|Dec 26, 2015

Human Mobility and Location-Based Analysis14 references3 citations

TL;DR

This paper proposes a machine learning approach to predict outgoing mobile phone calls using temporal patterns and recent calling behavior. By modeling calling behavior as a multiclass classification problem with day-of-week and time-of-day features, the method achieves 44% accuracy for top-1 predictions and over 78% for top-5 predictions, outperforming baseline methods like last-k calls and most-frequently-called numbers.

ABSTRACT

In this paper we predict outgoing mobile phone calls using a machine learning approach. We analyze to which extent the activity of mobile phone users is predictable. The premise is that mobile phone users exhibit temporal regularity in their interactions with majority of their contacts. In the sociological context, most social interactions have fairly reliable temporal regularity. If we quantify the extension of this behavior to interactions on mobile phones we expect that caller-callee interaction is not merely a result of randomness, rather it exhibits a temporal pattern. To this end, we tested our approach on an anonymized mobile phone usage dataset collected specifically for analyzing temporal patterns in mobile phone communication. The data consists of 783 users and more than 12,000 caller-callee pairs. The results show that users' historic calling patterns can predict future calls with reasonable accuracy.

Motivation & Objective

• To investigate whether mobile phone calling behavior exhibits predictable temporal regularity based on sociological principles of routine social interaction.
• To develop a machine learning model that predicts future outgoing calls using historical calling patterns and temporal features.
• To evaluate the model’s performance against baseline approaches such as last-k calls and most-frequently-called numbers.
• To assess the generalizability of call prediction across diverse user populations using datasets from Pakistan and the US.
• To explore implications for improved mobile phone user interfaces and personalized service delivery.

Proposed method

• Models call prediction as a multiclass classification problem, where the target is the next outgoing call.
• Uses temporal features including day of the week and time of day as key predictors.
• Incorporates recent calling behavior by including the last-k called numbers as features.
• Applies standard machine learning classifiers trained on anonymized call logs from 783 users and 12,000 caller-callee pairs.
• Evaluates performance using top-k prediction accuracy, comparing against last-k calls and most-frequently-called lists.
• Tests the model on two datasets: one from Pakistan and one from the Reality Mining project in the US.

Experimental results

Research questions

• RQ1 To what extent can temporal regularity in social interactions be leveraged to predict future mobile phone calls?
• RQ2 Can a machine learning model using time-based and recent calling features outperform simple baselines like last-k calls?
• RQ3 How does the prediction accuracy vary across different user groups and datasets?
• RQ4 Does the assumption of periodic calling behavior hold across diverse populations?
• RQ5 Can such a model inform the design of more intelligent and efficient mobile calling interfaces?

Key findings

• The proposed method achieves 44% accuracy in predicting the top-1 outgoing call, significantly outperforming the baseline last-k calls approach.
• With a top-5 prediction list, the model achieves over 78% accuracy on the Reality Mining dataset, demonstrating strong performance at scale.
• The model outperforms both the last-k calls and most-frequently-called numbers baselines across both the Pakistan and Reality Mining datasets.
• Temporal features such as day of the week and time of day are critical predictors, supporting the hypothesis of circadian rhythm in calling behavior.
• The results suggest that users' calling patterns are highly predictable based on historical temporal patterns, with implications for interface design and service personalization.
• The study identifies a gap in prior work by including a direct comparison with last-k calls, a more realistic baseline than previously used in similar studies.