Abstract
Recommendation systems often struggle with the "cold start" problem, where new or inactive users lack sufficient history for effective collaborative filtering. This paper explores predicting user interest similarity by leveraging social and demographic data from Online Social Networks (OSNs). By transitioning from specific video-matching to a tag-based profiling system and employing a hybrid tree-encoded linear model, the authors demonstrate accurate identification of "like-minded" users even in the absence of viewing history.
Introduction
Traditional recommendation engines rely heavily on Collaborative Filtering (CF). However, CF fails when users have sparse histories or when their interests evolve beyond past behaviors. Furthermore, the sheer volume of content means two users with identical tastes might never watch the same specific videos. This research bridges these gaps by using OSN data—such as age, gender, and social interactions—to infer interest similarity.
Methodology
The study introduces a multi-step approach to move beyond simple video-to-video matching:
Tag-based Profiling
Instead of matching raw video IDs, the researchers utilized a vocabulary of approximately 30,000 editor-assigned tags. Two profiling methods were proposed:
- Popular Tag based Profiling (PTP): Weights tags by interaction frequency.
- Representative Tag based Profiling (RTP): Uses a TF-IDF-like approach to penalize common tags and highlight unique "long-tail" interests.
Feature Engineering and Modeling
The team extracted demographic features (age pairs, gender) and social features (friendship, interaction frequency). A key metric, Individuality of Interest, was defined to quantify how unique a user's tastes are relative to the general population.
The core technical contribution is a hybrid tree-encoded linear model. This model uses Gradient Boosting Decision Trees (GBDT) to transform raw features into high-dimensional binary vectors, capturing non-linear interactions, which are then processed by a regularized linear model (Logistic or Lasso regression) for stable global prediction.
Key Insights from OSN Data
Age Pairs vs. Age Differences
A significant finding was that specific age pairs are more predictive than simple age differences. For example, the interest similarity between two 20-year-olds is distinct from that between two 50-year-olds, even though the age difference is zero in both cases.
Interaction Frequency vs. Intensity
The data suggests that the frequency of interaction (number of days communicating) is a stronger predictor of shared interests than the total volume of messages. Consistent, casual interaction often indicates a stronger topical bond.
Gender and Geography
Female users were found to have more concentrated interests and higher mutual similarity compared to male users. Interestingly, physical proximity (living in the same city) did not significantly correlate with shared video interests, highlighting the dominance of digital subcultures.
Results and Conclusion
The hybrid model achieved an AUC of 0.834, significantly outperforming standard models. Tag-based profiling also proved superior to video-based methods. While PTP offered higher accuracy, RTP provided greater recommendation diversity, aiding "long-tail" content discovery. This research underscores the power of social data in overcoming the limitations of traditional recommendation systems.
Summary updated