[Paper Review] Who are Like-minded: Mining User Interest Similarity in Online Social Networks

Chunfeng Yang, Yipeng Zhou+1 more|arXiv (Cornell University)|Mar 7, 2016

Recommender Systems and Techniques4 citations

TL;DR

This paper proposes a tag-based user profiling approach—Popular Tag Profiling (PTP) and Representative Tag Profiling (RTP)—to predict user interest similarity in online social networks using demographic and social features, outperforming traditional video-based and collaborative filtering methods. A novel hybrid tree-encoded linear model achieves superior performance in predicting interest similarity, especially for inactive or new users, and improves video recommendation accuracy and diversity.

ABSTRACT

In this paper, we mine and learn to predict how similar a pair of users' interests towards videos are, based on demographic (age, gender and location) and social (friendship, interaction and group membership) information of these users. We use the video access patterns of active users as ground truth (a form of benchmark). We adopt tag-based user profiling to establish this ground truth, and justify why it is used instead of video-based methods, or many latent topic models such as LDA and Collaborative Filtering approaches. We then show the effectiveness of the different demographic and social features, and their combinations and derivatives, in predicting user interest similarity, based on different machine-learning methods for combining multiple features. We propose a hybrid tree-encoded linear model for combining the features, and show that it out-performs other linear and treebased models. Our methods can be used to predict user interest similarity when the ground-truth is not available, e.g. for new users, or inactive users whose interests may have changed from old access data, and is useful for video recommendation. Our study is based on a rich dataset from Tencent, a popular service provider of social networks, video services, and various other services in China.

Motivation & Objective

• To address the challenge of predicting user interest similarity when video consumption data is unavailable, such as for new or inactive users.
• To investigate the effectiveness of demographic (age, gender, location) and social (friendship, group membership, interaction) features in predicting interest similarity.
• To develop and evaluate a tag-based user profiling method that overcomes limitations of video-based and latent topic modeling approaches like LDA and collaborative filtering.
• To design a hybrid machine learning model that combines tree-based and linear models for improved prediction of user interest similarity.
• To validate the practical utility of the predicted similarity in real-world video recommendation systems.

Proposed method

• Proposes two tag-based user profiling methods: Popular Tag Profiling (PTP), which counts common tags between users based on frequency, and Representative Tag Profiling (RTP), which weights tags by their representativeness.
• Uses video access patterns from Tencent’s platform as ground truth for active users, with tag data from editors as a folksonomy-based categorization system.
• Applies multiple machine learning models—including linear models, tree-based models, and a novel hybrid tree-encoded linear model—to predict interest similarity using demographic and social features.
• The hybrid model integrates decision trees with linear regression, enabling adaptive feature weighting and improved generalization across diverse user profiles.
• Employs a systematic evaluation framework comparing prediction performance across different feature combinations and models.
• Validates the approach in a video recommendation pipeline, comparing results against baseline methods like video-based profiling (VBP).

Experimental results

Research questions

• RQ1 How effective are demographic and social features in predicting user interest similarity when video consumption data is missing?
• RQ2 How do PTP and RTP compare in accuracy and diversity of interest similarity prediction compared to video-based or latent topic modeling methods?
• RQ3 Can a hybrid tree-encoded linear model outperform standard linear and tree-based models in predicting user interest similarity?
• RQ4 To what extent does predicted interest similarity improve video recommendation performance in terms of accuracy and diversity?
• RQ5 How does the similarity between a user’s current interest and her past or average user interest correlate with demographic and social features?

Key findings

• The hybrid tree-encoded linear model significantly outperforms other linear and tree-based models in predicting user interest similarity, achieving the highest AUC and F1 scores.
• PTP-based profiling yields higher prediction accuracy (AUC up to 0.950) compared to RTP, which produces more diverse recommendations by better capturing long-tail interests.
• The model achieves an AUC of 0.930–0.942 across different user groups, with F1 scores ranging from 0.883 to 0.894, indicating strong generalization.
• Social features such as friendship, group membership, and interaction history show strong correlation with interest similarity, especially when combined with demographic attributes.
• Interest similarity with past behavior and average user interest are useful predictors, particularly for users with sparse or outdated consumption data.
• When applied to video recommendation, the tag-based models (especially RTP) produce more diverse recommendations than video-based profiling (VBP), enhancing discovery of niche content.