Interactive Gender Inference with Integer Linear Programming

Shoushan Li, Jingjing Wang, Guodong Zhou, Hanxiao ShiIJCAI 2015

Abstract

Traditional gender inference models often treat social media users as isolated entities, failing to capture the rich contextual information present in interpersonal interactions. This paper proposes an "interactive gender inference" framework that jointly predicts the genders of conversation participants. By employing Integer Linear Programming (ILP) to enforce consistency across multiple interactions and individual profiles, the method achieves significant accuracy gains over standard classification techniques, demonstrating that gender is a relational property within social networks.

Introduction

Inferring a user's gender from their digital footprint is essential for personalized advertising, market research, and human-computer interaction. However, most existing models analyze a single person's text in isolation. On platforms like Sina Weibo or Twitter, individual messages are often brief and ambiguous. This paper argues that "individual gender inference" misses vital information hidden in social interactions. For instance, relational cues like "my wife" or honorifics like "sistah" provide clear signals about the genders of both parties involved. The challenge lies in mathematically modeling these interconnected relationships to improve prediction accuracy across a social network.

Methodology

The researchers frame the problem as a four-category classification task: identifying the gender pair of an interaction (male-to-male, male-to-female, female-to-female, and female-to-male).

Feature Extraction and Base Classifiers

The system utilizes standard NLP features, including Bag-of-Words (BOW), Part-of-Speech (POS) patterns, and an "F-measure" to calculate text contextuality. Initial probability distributions are generated using three base classifiers: Naive Bayes (NB), Support Vector Machines (SVM), and Maximum Entropy (ME).

Table 1: Baseline performance of various classifiers across different feature sets for the interactive task.

Integer Linear Programming (ILP) Optimization

The core innovation is the use of ILP for global optimization. Instead of relying solely on local classifier outputs, the model introduces constraints to ensure consistency:

  • Intra-task Constraints: These force the model to assign the same gender to a user across all their different interactions (e.g., if User A interacts with both User B and User C).
  • Extra-task Constraints: These link the interactive model with an individual model (analyzing the user's own posts), allowing high-confidence individual predictions to correct ambiguous interactive ones.

Experimental Results

The study utilized data from Sina Weibo, focusing on users with at least 10 comments to ensure sufficient interaction data. The results demonstrate that relational constraints lead to dramatic performance improvements over standard models.

Table 2: Accuracy improvements using various ILP constraints compared to the baseline.

As shown in the results, a standard Maximum Entropy classifier using all features achieved a baseline accuracy of 0.622. By introducing intra-task consistency constraints (ILP C1+C2+C3), the accuracy jumped to 0.771. When the full model incorporating both intra-task and extra-task (individual) constraints was applied, the accuracy reached 0.809—representing an 18.7% absolute improvement over the baseline. Furthermore, the interactive approach provided a mutual benefit, raising individual gender inference accuracy from 0.873 to 0.890.

Conclusion

This research highlights that gender in a social network is a relational property. By treating the genders of interacting users as joint variables and applying global optimization via ILP, models can resolve ambiguities that baffle standard classifiers. While the model currently faces challenges with data sparsity among "lurkers" and relies on surface-level linguistic features, it provides a robust framework for leveraging social graphs to improve demographic inference.

Summary updated

BibTeX
@inproceedings{li2015interactive,
  author = {Shoushan Li and Jingjing Wang and Guodong Zhou and Hanxiao Shi},
  title = {{Interactive Gender Inference with Integer Linear Programming}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2015},
}
Figure