Abstract
Latent author attribute prediction in social media provides a novel set of conditions for the construction of supervised classification models. With individual authors as training and test instances, their associated content ("features") are made available incrementally over time, as they converse over discussion forums. We propose various approaches to handling this dynamic data, from traditional batch training and testing, to incremental bootstrapping, and then active learning via crowdsourcing. Our underlying model relies on an intuitive application of Bayes rule, which should be easy to adopt by the community, thus allowing for a general shift towards online modeling for social media.
Introduction
Predicting latent user attributes—such as gender, age, or political leanings—is a cornerstone of social media analytics. Traditionally, this task is treated as a static "batch" learning problem: researchers collect a fixed dataset and train a classifier once. However, social media is inherently a dynamic stream. Users generate content at varying rates, and their social context evolves. Static models fail to capture this temporal dimension and often struggle with the "long tail" of users who have not yet produced enough content for a single-shot classification. This paper proposes a shift toward online, incremental modeling that updates predictions as new data arrives.
Methodology
The core of the proposed approach is an Online Bayesian Model. Instead of processing all data at once, each tweet is treated as independent evidence. Using Bayes’ Rule, the model updates the probability of a user’s attribute (e.g., political affiliation) every time a new message is authored.
The authors explored several variations to handle the streaming data:
- Iterative Batch (IB): A baseline model trained on labeled users and applied to a growing test stream.
- Active Without Oracle (AWOO): A bootstrapping method where the model adds highly confident predictions to its own training set.
- Rationale Filtering (IBR/AWR): The model focuses on "rationales"—highly predictive unigrams identified through crowdsourcing or statistical analysis—to reduce noise.
- Social Stream Integration: Beyond individual tweets, the model incorporates the "neighbor" stream, using the communications of a user's social circle as proxy evidence.
The study utilized a Twitter dataset of 300 users (150 Democrats, 150 Republicans). To enrich the context, the authors sampled 20 friends for each user, each with 200 tweets, creating a comprehensive dataset of individual (U) and neighborhood (N/UN) communications.
Results and Discussion
The integration of social network data proved highly effective. Neighbors are often more talkative than the target users, and their content provides strong signals for classification.
Key findings include:
- Social Context: Using neighbor data (N or UN streams) consistently outperformed models relying solely on the user's own tweets. Friends often provide more "talkative" and easily classifiable data.
- Denoising via Rationales: Rationale filtering acted as intelligent dimensionality reduction, leading to accuracy gains of up to 27.7% in early time steps by ignoring irrelevant chatter.
- Confidence Thresholds: The use of a threshold ($\theta$) allows for a trade-off between speed and precision. A high threshold (0.95) ensures accuracy, while a lower one (0.55) enables earlier predictions.
- Performance Bounds: In an "Active With Oracle" (AWO) scenario, the model achieved 85% accuracy after just two iterations when using joint user-neighbor streams.
While effective, the model currently assumes balanced class priors and treats tweets as independent events, ignoring linguistic context between messages. Future work will explore estimating real-world priors and integrating real-time human-in-the-loop labeling for difficult cases.
Conclusion
By treating social media as a continuous stream rather than a static corpus, this research demonstrates that incremental Bayesian models can provide more timely and accurate personal analytics. The inclusion of social context and rationale filtering further enhances performance, paving the way for more responsive social media monitoring systems.
Summary updated