Richi Nayak’s research while affiliated with Queensland University of Technology and other places

Understanding the cultural and environmental influences on roadway crash patterns is essential for designing effective prevention strategies. This study applies advanced AI techniques, including Bidirectional Encoder Representations from Transformers (BERT) and Shapley Additive Explanations (SHAP), to examine traffic crash patterns in the United States and Jordan. By analyzing tabular data and crash narratives, the research reveals significant regional differences: in the USA, vehicle overturns and roadway conditions, such as guardrails, are major factors in fatal crashes, whereas in Jordan, technical defects and driver behavior play a more critical role. SHAP analysis identifies “driver” and “damage” as pivotal terms across both regions, while country-specific terms such as “overturn” in the USA and “technical” in Jordan highlight regional disparities. Using BERT/Bi-LSTM models, the study achieves up to 99.5% accuracy in crash severity prediction, demonstrating the robustness of AI in traffic safety analysis. These findings underscore the value of contextualized AI-driven insights in developing targeted, region-specific road safety policies and interventions. By bridging the gap between developed and developing country contexts, the study contributes to the global effort to reduce road traffic injuries and fatalities.

Near-miss traffic incidents, positioned just above "unsafe acts" on the safety triangle theory, offer crucial predictive insights for preventing crashes. However, these incidents are often underrepresented in traffic safety research, which tends to focus primarily on actual crashes. This study introduces a novel AI-based framework designed to detect and analyze near-miss and crash events in crowdsourced dashcam footage. The framework consists of two key components: a deep learning model to segment video streams and identify potential near-miss or crash incidents and a multimodal large language model (MLLM) to further analyze and extract narrative information from the identified events. We evaluated three deep learning models—CNN, Vision Transformers (ViTs), and CNN+LSTM—on a dataset specifically curated for three-class classification (crashes, near-misses, and normal driving events). CNN achieved the highest accuracy (90%) and F1-score (89%) at the frame level. At the event level, ViTs delivered a strong performance with a test accuracy of 77.27% and an F1-score of 67.37%, while CNN+LSTM, although lower in overall performance, demonstrated significant potential with a test accuracy of 78.1% and an F1-score of 68.69%. For a deeper analysis, we applied GPT-4o to process critical safety events (near-misses and crashes), utilizing both zero-shot and few-shot learning for narrative generation and feature extraction. The zero-shot learning method performed better, achieving an accuracy of 81.2% and an F1-score of 81.9%. This study underscores the potential of combining deep learning with MLLMs to enhance traffic safety analysis by integrating near-miss data as a key predictive layer. Our approach highlights the importance of leveraging near-miss incidents to proactively enhance road safety, thereby reducing the likelihood of crashes through early intervention and better event understanding.

Road traffic crashes (RTCs) are a global public health issue, with traditional analysis methods often hindered by delays and incomplete data. Leveraging social media for real-time traffic safety analysis offers a promising alternative, yet effective frameworks for this integration are scarce. This study introduces a novel multitask learning (MTL) framework utilizing large language models (LLMs) to analyze RTC-related tweets from Australia. We collected 26,226 traffic-related tweets from May 2022 to May 2023. Using GPT-3.5, we extracted fifteen distinct features categorized into six classification tasks and nine information retrieval tasks. These features were then used to fine-tune GPT-2 for language modeling, which outperformed baseline models, including GPT-4o mini in zero-shot mode and XGBoost, across most tasks. Unlike traditional single-task classifiers that may miss critical details, our MTL approach simultaneously classifies RTC-related tweets and extracts detailed information in natural language. Our fine-tunedGPT-2 model achieved an average accuracy of 85% across the six classification tasks, surpassing the baseline GPT-4o mini model’s 64% and XGBoost’s 83.5%. In information retrieval tasks, our fine-tuned GPT-2 model achieved a BLEU-4 score of 0.22, a ROUGE-I score of 0.78, and a WER of 0.30, significantly outperforming the baseline GPT-4 mini model’s BLEU-4 score of 0.0674, ROUGE-I score of 0.2992, and WER of 2.0715. These results demonstrate the efficacy of our fine-tuned GPT-2 model in enhancing both classification and information retrieval, offering valuable insights for data-driven decision-making to improve road safety. This study is the first to explicitly apply social media data and LLMs within an MTL framework to enhance traffic safety.

Transfer learning has gained significant traction in natural language processing due to the emergence of state-of-the-art pre-trained language models (P.L.M.s). Unlike traditional word embedding methods such as TF-IDF and Word2Vec, P.L.M.s are context-dependent and outperform conventional techniques when fine-tuned for specific tasks. This paper proposes an innovative hard voting classifier to enhance crash severity classification by combining machine learning and deep learning models with various word embedding techniques, including BERT, RoBERTa, Word2Vec, and TF-IDF. Our study involves two comprehensive experiments using motorists’ crash data from the Missouri State Highway Patrol. The first experiment evaluates the performance of three machine learning models—XGBoost (X.G.B.), random forest (R.F.), and naive Bayes (N.B.)—paired with TF-IDF, Word2Vec, and BERT feature extraction techniques. Additionally, BERT and RoBERTa are fine-tuned with a Bidirectional Long Short-Term Memory (Bi-LSTM) classification model. All models are initially evaluated on the original dataset. The second experiment repeats the evaluation using an augmented dataset to address the severe data imbalance. The results from the original dataset show strong performance for all models in the “Fatal” and “Personal Injury” classes but a poor classification of the minority “Property Damage” class. In the augmented dataset, while the models continued to excel with the majority classes, only XGB/TFIDF and BERT-LSTM showed improved performance for the minority class. The ensemble model outperformed individual models in both datasets, achieving an F1 score of 99% for “Fatal” and “Personal Injury” and 62% for “Property Damage” on the augmented dataset. These findings suggest that ensemble models, combined with data augmentation, are highly effective for crash severity classification and potentially other textual classification tasks.

Traditionally, various statistical analyses were employed to model crash severity using tabular data. However, textual crash narratives, which contain rich information, were underutilized due to the complexity associated with handling text. Recently, transfer learning and language models have gained substantial popularity in natural language processing. In contrast to traditional word embedding methods like TF-IDF and Word2Vec, Large Language Models (LLMs) are context-dependent and outperform conventional techniques when finetuned on the target domain. A major limitation of LLMs is the lack of explainability, functioning akin to a black box. Furthermore, datasets in certain domains exhibit high imbalances. This study presents a framework to address these challenges. Firstly, we generated a balanced dataset using BERT Language models. Secondly, we compared the performance of traditional embedding TF-IDF with the finetuned BERT/Bi-LSTM in classifying crash severity. The results demonstrated that the finetuned BERT/Bi-LSTM model outperformed XGB with accuracy scores of 98% and 96%, respectively. The finetuned classifier was then input into the SHAP explainable model to identify factors impacting crash severity. The analysis utilized crash data obtained from the Missouri State Highway Patrol. Results indicated that the use of PLM and SHAP techniques revealed factors associated with crash severity levels, thereby enabling deployment for other textual classification problems.

Extracting valuable information from a large corpus of unstructured data poses a formidable challenge in many applications. Transformer-based architectures e.g. BERT, employing transfer learning techniques, have exhibited promising results across diverse NLP tasks. Nonetheless, the practical implementation of these models presents a significant hurdle given the substantial demand for annotated data during the training phase. In this paper, we present GAN-IE, a novel GAN-based model architecture, designed specifically for information extraction from unstructured textual data while accounting for limited annotated resources. In a generative adversarial setting, GAN-IE leverages BERT’s rich semantic and contextual knowledge obtained from unlabelled data while fine-tuning. Experimental results show that GAN-IE achieves a level of accuracy that surpasses the current state-of-the-art models when trained using a fraction of labelled data ($\sim$100–200 annotated samples).

Anomaly detection in time series data, to identify points that deviate from normal behaviour, is a common problem in various domains such as manufacturing, medical imaging, and cybersecurity. Recently, Generative Adversarial Networks (GANs) are shown to be effective in detecting anomalies in time series data. The neural network architecture of GANs (i.e. Generator and Discriminator) can significantly improve anomaly detection accuracy. In this paper, we propose a new GAN model, named Adjusted-LSTM GAN (ALGAN), which adjusts the output of an LSTM network for improved anomaly detection in both univariate and multivariate time series data in an unsupervised setting. We evaluate the performance of ALGAN on 46 real-world univariate time series datasets and a large multivariate dataset that spans multiple domains. Our experiments demonstrate that ALGAN outperforms traditional, neural network-based, and other GAN-based methods for anomaly detection in time series data.

In this paper, we present an efficient novel method for mining discriminative itemsets over data streams using the sliding window model. Discriminative itemsets are the itemsets that are frequent in the target data stream, and their frequency in the target stream is much higher in comparison to their frequency in the rest of the streams. The problem of mining discriminative itemsets has more challenges than mining frequent itemsets, especially in the sliding window model, as during the window frame sliding, the algorithms have to deal with the combinatorial explosion of itemsets in more than one data stream, for the transactions coming in and going out of the sliding window. We propose a single scan algorithm using two novel in-memory data structures for mining discriminative itemsets in a combination of offline and online sliding windows. Offline processing is used for controlling the generation of many unpromising itemsets. Online processing is used for getting more up-to-date and accurate online answers between two offline slidings. The discovered discriminative itemsets are accurately updated in the offline sliding window periodically, and the mining process is continued in the online sliding between two periodic offline slidings. The extensive empirical analysis shows that the proposed algorithm provides efficient time and space complexities with full accuracy. The algorithm can handle large, fast-speed, and complex data streams.