Inference of demographic attributes based on mobile phone usage patterns and social network topology

Carlos Sarraute, Jorge Brea, Javier Burroni, Pablo Blanc

TL;DR

This paper explores how mobile phone call data and social network structure can be used to infer users' age and gender. They developed a novel diffusion-based algorithm that leverages network topology, significantly outperforming traditional machine learning methods that rely solely on individual user behavior. This work highlights the power of network structure in understanding demographic attributes.

What Problem Does This Paper Address?

Understanding the demographic characteristics of a population is crucial for various applications, from market research to public health initiatives. Traditionally, this information is gathered through surveys or census data, which can be costly, time-consuming, and may not capture the dynamic nature of human behavior. Mobile phone usage generates vast amounts of data that reflect user activity and social interactions. However, directly linking this usage data to demographic attributes like age and gender for large, unlabeled populations is challenging. This paper aims to bridge this gap by inferring these attributes using both individual calling patterns and the structure of the communication network.

Methods

The authors employ a two-pronged approach:

  1. Observational Study: They first analyze mobile phone usage patterns (calls, SMS, duration) for different age and gender groups within their dataset to identify significant differences.
  2. Predictive Modeling:
    • Node Attribute-Based ML: They experiment with standard machine learning algorithms (Naive Bayes, Logistic Regression, SVM) using individual user features (derived from call logs) to predict gender and age.
  3. Network Topology-Based Diffusion Model: Their key contribution is a novel Reaction-Diffusion Algorithm. This method treats age group prediction as a diffusion process on the social network graph. Initial "seed" nodes (users with known age) "diffuse" their age probability to their neighbors. This algorithm leverages the network's structure, assuming that connections between users can reveal information about their attributes.
  4. Population Pyramid Scaling (PPS): To ensure the predicted demographic distributions match the ground truth, they introduce a PPS method. This post-processing step adjusts the probability vectors generated by the prediction algorithms to align with known population demographics.

Datasets

The primary dataset used in this study consists of Call Detail Records (CDRs) and SMS records from a major mobile phone operator in Mexico, covering a period of three months. This dataset includes anonymized transaction logs. A subset of users, referred to as the "ground truth" (NGT), has known demographic information (age and gender). The authors also use census data for comparison with the general population's age pyramid.

Key Insights and Ideas

  • Demographic Differences in Usage: Mobile phone usage patterns (e.g., call duration, frequency) show statistically significant differences between genders and age groups, motivating the prediction task.
  • Gender Homophily: Men tend to call men more, and women tend to call women more, indicating a degree of gender homophily in communication.
  • Age Homophily is Strong: The social network exhibits strong age homophily, meaning users are much more likely to communicate with individuals of their own age. This is a crucial insight for network-based prediction.
  • Network Topology is Powerful: The Reaction-Diffusion algorithm, which explicitly uses network structure, significantly outperforms traditional ML methods that only consider individual user attributes.
  • Topological Features Matter: Specific topological features of a node (like the number of "seed" nodes in its neighborhood) directly influence the accuracy of the prediction algorithm.
  • Robustness of the Diffusion Model: The diffusion model's performance is robust to parameter choices, particularly the balance between "reaction" (memory) and "diffusion" (information spread).

Key Results

  • Gender Prediction Accuracy: Using Linear SVM and Logistic Regression with node attributes, they achieved up to 81.4% accuracy when predicting gender for a subset of users (12.5% of the population).
  • Age Prediction Performance:
    • Node attribute-based Multinomial Logistic Regression achieved up to 52.7% accuracy.
  • The Reaction-Diffusion algorithm alone achieved up to 62.3% accuracy for age prediction when tagging 12.5% of the population.
  • The combined ML + Reaction-Diffusion approach achieved up to 57.2% accuracy.
  • Age Group Performance: The Reaction-Diffusion algorithm performed best for the 35-50 age group (52.3% accuracy), likely due to stronger age homophily and less noise from off-diagonal correlations.
  • Impact of Seed Nodes: Having even one "seed" node in a user's immediate neighborhood significantly increases prediction accuracy.

Limitations and Future Work

  • Dataset Bias: The study acknowledges that the mobile phone user population (NGT) differs from the general population's age pyramid, particularly for younger age groups.
  • Selection Bias: The method used to select the "ground truth" subset of users is not specified, which could introduce bias.
  • Incomplete Network: The constructed graph is based on a single operator's data, meaning it's an incomplete representation of the entire communication network.
  • Future Work:
    • Quasi-experiments: Using the developed methods to investigate whether observed behavioral differences are due to gender/age or ego-network structure.
  • Spending Behavior Prediction: Applying the methodology to infer users' spending habits.
  • Mobility Patterns: Incorporating geolocation data to analyze mobility patterns and their relation to demographics.
  • More Sophisticated Network Metrics: Further exploring how topological features influence prediction accuracy.

Summary updated

Abstract

Mobile phone usage provides a wealth of information, which can be used to better understand the demographic structure of a population. In this paper, we focus on the population of Mexican mobile phone users. We first present an observational study of mobile phone usage according to gender and age groups. We are able to detect significant differences in phone usage among different subgroups of the population. We then study the performance of different machine learning (ML) methods to predict demographic features (namely, age and gender) of unlabeled users by leveraging individual calling patterns, as well as the structure of the communication graph. We show how a specific implementation of a diffusion model, harnessing the graph structure, has significantly better performance over other node-based standard ML methods. We provide details of the methodology together with an analysis of the robustness of our results to changes in the model parameters. Furthermore, by carefully examining the topological relations of the training nodes (seed nodes) to the rest of the nodes in the network, we find topological metrics which have a direct influence on the performance of the algorithm.
BibTeX
@inproceedings{sarraute2015inference,
  author = {Carlos Sarraute and Jorge Brea and Javier Burroni and Pablo Blanc},
  title = {{Inference of demographic attributes based on mobile phone usage patterns and social network topology}},
  booktitle = {Soc. Netw. Anal. Min},
  year = {2015},
  url = {https://openreview.net/forum?id=OQr9TTFn8Wl},
}
Figure