[TMP-087] Multivariate Asynchronous Shapelets for Imbalanced Car Crash Predictions

Contact person: Francesco Spinnato (francesco.spinnato@sns.it, francesco.spinnato@isti.cnr.it)

Internal Partners:

1. Università di Pisa
2. ISTI-CNR Pisa
3. Generali Italia

The increasing availability of real-time sequential data, combined with advanced AI decision-making systems, is transforming the mobility industry. Crash Data Recorders (CDRs) are increasingly being used in cars to monitor safety measures, establish human tolerance limits, and quantify vehicle status. These recorders are usually installed on the airbag control module, collecting data before and after a crash. Recently, with the use of powerful Machine Learning (ML) models, these devices have become a valuable source of data for both academic research and businesses, such as insurance companies, to monitor and improve customer service quality.

In this work, we collaborated with Generali Italia, Italy’s biggest insurance company and part of Assicurazioni Generali, one of the largest global insurance and asset management providers. Generali Italia is developing an automatic classification system to provide first aid to its customers. As part of their insurance products, Generali offers to install a CDR in their customers vehicles. This system monitors the vehicle during its use and, among other services, tracks speed and acceleration on the three car axes. This data is used to train a deep learning model that enables the AI system to alert a Generali operator of possible car crashes.

By examining the CDR data and model predictions, the operator can make an informed decision and only contact the customer if assistance is really necessary. Two weaknesses are currently present. First, the high sensitivity of the AI system might cause unnecessary and harassing calls. Second, the AI system is based on a deep learning model that is inherently not interpretable, i.e., it is a black-box. This lack of transparency could hinder the operators understanding of the model’s outcome, potentially leading to a lack of trust, especially if it produces incorrect classifications. Moreover, the opaque nature of a deep learning model makes it difficult to improve the model’s technical performance once a certain plateau is reached (in the specific case under consideration, the reduction of false positives). In such a critical scenario, eXplainable Artificial Intelligence (XAI) is essential for interpreting these black-box predictions to ensure reliability in decision-making. XAI for time series classification is a rapidly emerging field, which presents many challenges due to the nature of time series data, which can be large, multivariate, highly imbalanced, and irregular. These characteristics, as seen in the datasets used by Generali, often cause off-the-shelf XAI approaches to fail due to the limitations of their implementation.

In this work we tackle the challenge of explainability in car crash prediction from different angles, utilizing real-world time series datasets for two distinct tasks:

1) standard time series classification and

2) classification of highly imbalanced time series, which is more akin to anomaly detection.

For the former, we combine existing post-hoc and ante-hoc XAI approaches in a pipeline that provides insights into the logic behind the black-box model used by Generali, enabling the construction of a more transparent predictive model. For the latter, we introduce Multivariate Asynchronous Shapelets, an interpretable-by-design approach based on multivariate shapelets, specifically developed to challenge state-of-the-art classifiers and anomaly detection algorithms, as well as the black-box model currently used by Generali.

Results Summary

A pipeline combining post-hoc and ante-hoc XAI for standard time series classification and the introduction of Multivariate Asynchronous Shapelets, an interpretable method developed to surpass state-of-the-art classifiers and Generali’s black-box model. The results are published.

In addition to the scientific contribution on XAI, it is important to highlight how the application of AI systems to automate the remote detection of potential car accidents by an insurance company has a positive impact on road safety, improving rescue operations and helping to reduce the potential impacts of an accident on the health of the insured.

Tangible Outcomes

1. M. Bianchi, F. Spinnato, R. Guidotti, D. Maccagnola, A. Bencini Farina. “Multivariate Asynchronous Shapelets for Imbalanced Car Crash Predictions”. In: Proceedings of the 27th International Conference on Discovery Science (DS 2024). Accepted for publication. 2024 but proceedings are not published yet