On the quality of inferring interests from social neighbors

Zhen Wen, Ching-Yung LinKDD 2010 · DOI

TL;DR

This paper investigates how accurately we can infer a person's interests by looking at their social network connections. They developed a method to predict the quality of these inferences, allowing applications to intelligently decide when to use social network data, significantly improving the accuracy of personalized recommendations.

What Problem Does This Paper Address?

In today's digital world, online platforms like social networks are awash with user data. A key challenge for these platforms is understanding individual user interests to provide personalized content, recommendations, and search results. While explicit user profiles offer some insight, they are often incomplete or outdated. Inferring interests from social connections (friends, friends of friends) offers a promising alternative, especially for "inactive" users who don't contribute much content. However, the effectiveness of this approach varies wildly, and it's unclear when and how well it works. This paper tackles this by quantifying the quality of interest inference from social neighbors and, crucially, developing a way to predict this quality beforehand.

Methods

The authors employ a multi-pronged approach:

  • Data Collection: They built a privacy-preserving distributed social sensing system within a large IT company to collect data from over 30,000 employees. This data includes communication logs (email, instant messaging) and Web 2.0 activities (social bookmarking, file sharing, blogging). They also gathered formal network information like organizational hierarchy.
  • Interest Modeling:
    • Implicit Interests: They used Latent Dirichlet Allocation (LDA) to extract latent topics from user-generated content (emails, bookmarks, shared files). These topics, weighted by their prevalence, form a user's implicit interest profile.
  • Explicit Interests: They used user-specified interests from their profiles, which were drawn from a predefined taxonomy.
  • Social Network Analysis: They constructed social networks based on communication frequency and analyzed both "informal" (communication-based) and "formal" (organizational hierarchy) networks.
  • Interest Inference: They used a network autocorrelation model, a standard technique in social sciences, to infer a user's interests based on the interests of their neighbors within a three-degree ego network.
  • Quality Prediction: This is the core novelty. They developed a regression model (specifically, Support Vector Regression - SVR) that uses network features (user activeness, in-degree, out-degree, betweenness centrality, management role) to predict the quality (precision and recall) of the inferred interests.

Datasets

The primary dataset is derived from the activities of over 30,000 employees within a large global IT company. This includes:

  • Electronic Communication Data: Emails and instant messages from 8,952 volunteer employees, anonymized and represented as term-count vectors. This data was used to infer social networks for over 400,000 employees.
  • Web 2.0 Social Content:
    • Social bookmarks from ~16,000 users.
  • Shared files from ~14,000 users.
  • Blogs from 5,000 contributors.
  • Audio/video media library from 4,200 users. The study specifically focused on social bookmarking and file sharing due to their wider coverage.

  • Formal Network Data: Organizational hierarchy and job roles.

  • Explicit User Interests: Optional information preferences specified by a subset of employees.

The dataset used for evaluation comprised 25,315 users who contributed social content.

Key Insights and Ideas

  • "Birds of a Feather" Holds, But Varies: The study confirms that social neighbors' interests are indeed correlated with an individual's interests, but the strength of this correlation varies significantly.
  • Combining Sources is Key: Inferring interests from multiple social media sources (email, social bookmarking, file sharing) dramatically improves accuracy compared to using any single source. This is attributed to the broader coverage and complementary nature of different media.
  • Predicting Inference Quality is Crucial: The most significant contribution is the ability to predict how good an inferred interest will be. This allows applications to intelligently decide when to rely on inferred interests, avoiding noisy or inaccurate results.
  • Network Structure Matters: Features like user activeness, network in-degree, and the collective influence of neighbors (sum of activeness/centrality) are strong indicators of inference quality.
  • Inactive Users are a Challenge, But Solvable: The methods are particularly valuable for inferring interests of users with limited personal activity, a common scenario in enterprise environments.

Key Results

  • Implicit Interest Inference:
    • Combining all three data sources (email, bookmarking, file sharing) yielded a mean inference quality of 45.1%, a significant improvement over single sources (e.g., electronic communication alone achieved 29.6%).
  • The quality prediction model, when used to filter for "high quality" inferences (e.g., >50% quality), improved the mean inference quality by 42.8%.
  • Explicit Interest Inference:
    • The quality prediction model improved the mean recall of explicit interests by 39.3% and the mean precision by an impressive 101%.
  • Feature Importance:
    • Electronic communication data was the most impactful source due to its sheer volume.
  • Social bookmarking was more informative than file sharing, likely due to its more concise and direct nature.
  • Neighbors within three degrees of separation are relevant, with two-degree neighbors sometimes showing greater impact due to their sheer number.
  • The collective influence of neighbors (sum of features) was more important than the influence of a single most prominent neighbor.

Limitations and Future Work

  • Data Source Bias: The study relies heavily on enterprise data, which might have specific characteristics (e.g., work-related interests). Generalizing to public social networks might require adjustments.
  • Privacy Concerns: While addressed through anonymization and legal review, the data collection itself is a significant undertaking.
  • Dynamic Interests: The current work focuses on static interests. Future work could explore how to infer the evolution of user interests over time.
  • Broader Network Properties: Investigating other network properties, like network backbone structure, could further enhance inference.
  • Integration into Applications: The authors plan to integrate these inferred interests into enterprise expert finding and content recommendation systems.

Summary updated

BibTeX
@inproceedings{wen2010quality,
  author = {Zhen Wen and Ching-Yung Lin},
  title = {{On the quality of inferring interests from social neighbors}},
  booktitle = {ACM SIGKDD Conference on Knowledge Discovery and Data Mining},
  year = {2010},
  doi = {10.1145/1835804.1835853},
}
Figure