Introduction
The rapid identification of missing person reports on social media platforms like Twitter or Sina Weibo is critical, particularly during the "golden hour" following a disappearance. However, extracting these reports from the massive volume of daily social media traffic is a significant challenge. This "needle in a haystack" problem is compounded by data sparsity (short tweets), information heterogeneity (text, images, and social links), and an extremely low signal-to-noise ratio, where relevant reports constitute a tiny fraction of total data.
Methodology: The $r$-instance (RI) Learning Framework
To address these challenges, we propose the $r$-instance (RI) learning model. This framework treats each tweet as a "bag" of word instances and leverages both content and social network structure.
Content Modeling
Instead of representing a tweet as a single static feature vector, the RI model utilizes high-dimensional word embeddings (word2vec). It applies an $\ell_0$-norm constraint to automatically select the top-$r$ most "positive-related" words within a tweet. A tweet is classified as a missing person report if it contains a significant proportion of these representative keywords (e.g., "lost," "missing," or specific physical descriptions).
Network Modeling and Homophily
A key innovation of the RI model is the incorporation of social "homophily"—the principle that related information tends to cluster. As illustrated in Figure 2, the model transforms a traditional user-user interaction network into a tweet-tweet network based on retweet and reply behaviors.
By applying a Laplacian matrix to this network, the model "smooths" the classification labels. This ensures that tweets linked by social behavior are more likely to be assigned the same label, effectively using the community structure to reduce false positives. The resulting non-convex optimization problem is solved efficiently using Block Coordinate Descent (BCD) and Nesterov’s acceleration.
Experimental Setup
The RI model was evaluated using a real-world dataset crawled from Sina Weibo between September 2014 and February 2015. The dataset comprised 40,373 tweets, of which only 1,404 (approximately 3.48%) were manually verified as genuine missing person reports. This low positive ratio reflects the actual scarcity and difficulty of the identification task in real-world scenarios.
Results and Discussion
The RI model demonstrated superior performance compared to standard baselines, including Support Vector Machines (SVM), Logistic Regression, and Random Forests.
- Performance Metrics: The RI model achieved an F1-measure of 0.863, outperforming the next best method (SVM at 0.849). A t-test confirmed that these improvements over standard Logistic Regression were statistically significant ($p < 0.05$).
- Precision and Recall: The model provided a balanced trade-off between precision and recall. While some baseline models achieved high recall, they suffered from excessive false alarms. The RI model's ability to filter noise is crucial for authorities who require high-confidence leads.
- Feature Efficacy: High-dimensional word embeddings consistently outperformed manually engineered features, such as punctuation counts or specific metadata tags.
The inclusion of the network Laplacian acted as a powerful regularizer. By leveraging the fact that missing person reports tend to circulate within specific helpful communities, the model significantly boosted precision.
Conclusion and Future Work
While the $r$-instance learning framework represents a significant step forward, challenges remain. Extremely short tweets lacking context and semantic overlaps (e.g., users searching for "charming strangers") can still lead to misclassification. Future research will focus on integrating multi-modal data, such as image recognition for attached photos, to further enhance identification accuracy and provide more actionable information for search efforts.
Summary updated