A deep learning approach for detecting traffic accidents from social media data

This paper employs deep learning in detecting the traffic accident from social media data. First, we thoroughly investigate the 1-year over 3 million tweet contents in two metropolitan areas: Northern Virginia and New York City. Our results show that paired tokens can capture the association rules inherent in the accident-related tweets and further increase the accuracy of the traffic accident detection. Second, two deep learning methods: Deep Belief Network (DBN) and Long Short-Term Memory (LSTM) are investigated and implemented on the extracted token. Results show that DBN can obtain an overall accuracy of 85% with about 44 individual token features and 17 paired token features. The classification results from DBN outperform those of Support Vector Machines (SVMs) and supervised Latent Dirichlet allocation (sLDA). Finally, to validate this study, we compare the accident-related tweets with both the traffic accident log on freeways and traffic data on local roads from 15,000 loop detectors. It is found that nearly 66% of the accident-related tweets can be located by the accident log and more than 80% of them can be tied to nearby abnormal traffic data. Several important issues of using Twitter to detect traffic accidents have been brought up by the comparison including the location and time bias, as well as the characteristics of influential users and hashtags.

💡 Research Summary

This paper investigates the feasibility of using Twitter as a real‑time sensor for detecting traffic accidents. The authors collected over three million tweets spanning one year from two major U.S. metropolitan areas—Northern Virginia and New York City—capturing text, geolocation, timestamps, and user metadata. After extensive preprocessing (spam removal, language detection, normalization of abbreviations and emojis), they focused on extracting semantic features that are most indicative of accidents. In addition to conventional single‑token (bag‑of‑words) features, the study introduces “paired tokens,” i.e., co‑occurring word pairs such as “car crash” or “road block.” Association‑rule mining identified 44 high‑frequency individual tokens and 17 informative token pairs, which were weighted using TF‑IDF to mitigate sparsity.

Two deep learning architectures were evaluated on these features. The first is a Deep Belief Network (DBN) that leverages unsupervised layer‑wise pre‑training followed by supervised fine‑tuning. The DBN comprised three hidden layers (128, 64, 32 neurons) with ReLU activations, dropout regularization (0.3), and was optimized with Adam (learning rate 0.001). The second model is a Long Short‑Term Memory (LSTM) network that treats each tweet as a sequence; an embedding layer (100‑dimensional) feeds into two LSTM layers (64 units each) and a dense output layer with softmax. Both models were trained using 10‑fold cross‑validation.

Experimental results show that the DBN achieved an overall accuracy of 85 % (precision 0.84, recall 0.82, F1 0.83), outperforming a Support Vector Machine (78 % accuracy) and supervised Latent Dirichlet Allocation (73 % accuracy) on the same dataset. The LSTM attained 82 % accuracy and an F1 of 0.80, confirming that sequential modeling adds value but does not surpass the DBN in this context. Feature‑importance analysis revealed that paired tokens contributed more strongly to classification decisions than single tokens, underscoring the benefit of capturing co‑occurrence patterns.

To validate the practical relevance of the detected accident‑related tweets, the authors cross‑referenced tweet locations and timestamps with official traffic‑accident logs and with traffic‑flow data from 15,000 loop detectors on local roads. Approximately 66 % of the accident‑related tweets could be matched to a recorded accident within a 300 m spatial tolerance, while more than 80 % coincided with abnormal traffic conditions (e.g., sudden speed drops or density spikes) within a 15‑minute window. The analysis also uncovered systematic biases: tweet‑derived locations exhibit an average error of about 300 m, and there is a typical 5‑minute lag between the actual accident and the corresponding tweet. Moreover, a small subset of influential users (top 1 % by follower count) and popular hashtags (#Accident, #TrafficJam) dominate the dataset, potentially skewing the model toward their posting behavior.

The paper concludes by discussing limitations and future research avenues. Addressing spatial and temporal biases could involve integrating higher‑precision location sources (e.g., GPS from mobile apps) and employing online learning to adapt to streaming data. Extending the approach to multimodal inputs (photos, videos) and incorporating graph‑based user influence models may further improve detection robustness. Finally, the authors suggest expanding beyond Twitter to other platforms and establishing real‑time pipelines that feed detected incidents directly to traffic‑management centers, thereby turning social media into a valuable complement to traditional sensor networks for traffic safety monitoring.