Abstract
Traditional recommendation systems often struggle with the "cold start" problem, where limited user history prevents accurate content delivery. This paper investigates methods to predict user interest similarity by leveraging demographic, social, and interest-based data. By utilizing tag-based user profiling and a novel hybrid tree-encoded linear model, we demonstrate that it is possible to accurately predict how "like-minded" users are, even in the absence of extensive viewing history. Our approach, validated on a large-scale dataset from Tencent, outperforms traditional linear and tree-based models, offering a robust solution for video recommendation systems.
1. Introduction
Collaborative Filtering (CF) is the cornerstone of modern recommendation engines, operating on the principle that users who agreed in the past will agree in the future. However, CF faces two significant hurdles: data sparsity and the "cold start" problem. When a user is new or inactive, the system lacks the historical data necessary to find neighbors. Furthermore, item-level comparisons are often too granular; two users may share a love for a genre but never watch the same specific titles, leading the system to overlook their similarity.
This research proposes a framework to bridge these gaps by mining auxiliary information—specifically demographics and social interactions—to predict interest similarity at a more conceptual level.
2. Methodology
The authors propose a two-pronged approach involving advanced user profiling and a sophisticated machine learning architecture.
2.1 Tag-Based User Profiling
To move beyond the limitations of specific video IDs, the study employs a folksonomy-based tagging system. Videos are labeled with descriptive tags, allowing the system to map user behavior to interest categories. Two weighting methods were developed:
- Popular Tag-based Profiling (PTP): This method weights tags based on the frequency of user interaction, emphasizing a user's primary interests.
- Representative Tag-based Profiling (RTP): Inspired by TF-IDF, RTP penalizes globally common tags (e.g., "movie") to highlight a user's unique or niche interests.
2.2 Hybrid Tree-Encoded Linear Model
Predicting similarity from heterogeneous social and demographic data requires a model capable of capturing complex interactions. The researchers developed a hybrid model consisting of two stages:
- Feature Encoding: A Gradient Boosting Decision Tree (GBDT) processes raw features (e.g., age, location, message frequency). The leaves of the trees serve as binary features that represent specific non-linear combinations of the input data.
- Prediction: These encoded features are combined with original features and fed into a regularized linear model. This architecture leverages the GBDT's ability to discover high-dimensional feature interactions while maintaining the stability and efficiency of a linear predictor.
3. Experimental Dataset
The study utilized a massive real-world dataset provided by Tencent Inc. The data encompasses:
- User Profiles: Demographic data including gender, age, and geographic location.
- Social Graphs: Rich interaction data including friendship links, group memberships, and the frequency of instant messaging via QQ and WeChat.
- Video Logs: Ground truth data from Tencent Video used to establish and validate similarity scores.
The experiments were conducted on a sample of one million user pairs, providing a statistically significant basis for the findings.
4. Key Insights and Results
The evaluation focused on the model's ability to predict similarity and its utility in recommendation tasks.
4.1 Model Performance
The hybrid tree-encoded linear model consistently achieved the highest Area Under the Curve (AUC) for classification and the lowest Mean Absolute Error (MAE) for regression. It outperformed standalone models, including Random Forests and standard GBDTs, proving that the combination of tree-based encoding and linear prediction is highly effective for social network data.
4.2 The Value of Social Signals
The research highlights that not all social links are equal. While simple friendship is a useful indicator, "interaction intensity" (messaging frequency) and shared community memberships (groups) are much stronger predictors of shared interests. These signals provide a nuanced view of user relationships that demographic data alone cannot capture.
4.3 Accuracy vs. Discovery
A critical finding was the distinction between PTP and RTP. PTP proved superior for raw recommendation accuracy. Conversely, RTP was more effective for "discovery," surfacing long-tail content and diverse interests that users might not have encountered through traditional popularity-biased methods.
5. Conclusion
By shifting the focus from item-level overlap to tag-based interest similarity and leveraging a hybrid machine learning model, this research provides a viable path forward for mitigating the cold start problem. The ability to predict user similarity using social and demographic data ensures that even new or inactive users can receive relevant, high-quality recommendations. Future work will explore the application of these methods to other domains, such as music and news, and examine the temporal evolution of user interests.
Summary updated