TL;DR
This paper introduces a novel approach to detect geolocated events in social media by combining location and time burstiness with user feedback. Their proposed algorithms, SIGLER-Cov and SIGLER-Samp, significantly outperform existing methods in terms of speed, robustness to noise, and the quality of detected events tailored to user interests.
What Problem Does This Paper Address?
Social media platforms are a rich source of real-time information about events happening worldwide. While detecting events based on keyword bursts is a well-studied problem, identifying geolocated events in an unsupervised manner, and crucially, incorporating user interests into this detection process, has remained a significant challenge. Existing methods often fail to pinpoint local events or require extensive labeled data. This paper aims to fill this gap by developing methods that can automatically discover geolocated events and allow users to guide the discovery towards events they find most relevant.
Methods
The core of the proposed solution lies in a unified framework for both data-driven and user-driven geolocated event detection. The authors define a geolocated event as a set of terms that exhibit burstiness within a specific geographical area and time interval.
- Data-Driven Detection: They first identify "bursting" terms by measuring their significance relative to their average occurrences in a given space-time zone, weighted by inverse document frequency (IDF). This forms the basis of a data-driven quality measure, M(P).
- User-Driven Integration: To incorporate user interests, they introduce a user-driven quality measure, Mu(P). This measure amplifies the quality score of events that contain terms or are located in areas that the user has previously indicated interest in (through liking events). This user interest is modeled dynamically based on feedback.
- Algorithms: Two primary algorithms are proposed:
- SIGLER-Cov: This algorithm uses a generate-and-test approach with pruning techniques to efficiently enumerate and filter geolocated events, ensuring a "coverage guaranteed event summary."
- SIGLER-Samp: This algorithm employs a pattern sampling approach, using a random walk on the space of potential events, biased towards events with higher Mu scores. This is designed for faster, interactive event discovery.
- Post-processing: A Louvain clustering algorithm is used to group similar terms within detected events, ensuring semantic coherence, and further pattern similarity is used to reduce redundancy in the final output.
Datasets
The authors used three real-world Twitter datasets collected from:
- New York (NYC)
- Los Angeles (LA)
- London
These datasets were filtered to include only geolocated tweets containing hashtags or user mentions, and bot-generated tweets were removed. For evaluation, they also generated artificial datasets with known ground-truth events to test robustness to noise.
Key Insights and Ideas
- Location Matters: Geolocated events cannot be effectively detected by methods that ignore spatial information.
- User Feedback is Crucial: Incorporating user preferences significantly improves the relevance and quality of detected events, making them more interesting to the end-user.
- Efficiency and Robustness: The proposed SIGLER algorithms are not only effective but also substantially faster and more robust to noisy data compared to existing state-of-the-art methods.
- Interactive Discovery: SIGLER-Samp offers a promising approach for real-time, interactive event discovery, balancing speed with the ability to find high-quality events.
Key Results
- Performance: SIGLER-Cov is orders of magnitude faster than MED and three times faster than GeoBurst.
- Robustness: SIGLER algorithms are more robust to noise than MED and GeoBurst.
- Effectiveness: On artificial datasets, SIGLER achieved higher F-scores, especially as noise levels increased.
- User Studies: Experiments with real users (crowdsourcing) and simulated users showed that the user-driven approach consistently leads to a higher number of liked events and better average ranks for liked events compared to data-driven methods.
- Event Quality: The user-driven approach produces events that are more relevant and of higher quality for the user.
Limitations and Future Work
- Optimal Summary is NP-hard: The paper acknowledges that finding a minimal-sized summary of events is NP-hard, and their approach aims for a "coverage guaranteed" summary, which may still contain some redundancy.
- Semantic Coherence: While post-processing helps, merging distinct events that occur in the same space-time is a potential issue.
- Future Directions: The authors suggest enhancing event detection by incorporating prior knowledge to find unexpected events, propagating user feedback more effectively to semantic neighborhoods, and exploring active learning-based heuristics for interactive processes. They also mention the potential for further research into learning explicit user interest models.
Summary updated