Detecting Car Insurance Fraud Using Improved Clustering with Genetic Algorithm

Detecting Car Insurance Fraud Using Improved Clustering with Genetic Algorithm
By

Behnam Yousefimehr, Mehdi Ghatee, Sina Moradi, Yasamin Tafakor, Sajed Tavakoli
has been accepted in “Iranian Journal of Insurance Research (IJIR)”

Abstract:

BACKGROUND AND OBJECTIVES: Clustering is one of the fundamental techniques in data mining and machine learning used to divide a dataset into homogeneous subsets. Various methods exist for clustering, each with its own strengths and weaknesses. One of the main challenges in clustering is finding the optimal number of clusters and the optimal assignment of data to these clusters. The genetic algorithm, as an optimization method based on natural evolution, has high capabilities in solving complex problems and searching large solution spaces and can be effectively used as a tool in clustering. This paper aims to evaluate the performance and accuracy of the genetic algorithm in clustering data and compare it with traditional clustering methods. To assess the performance of this algorithm, several insurance datasets were used, and the results were analyzed using various metrics such as accuracy, speed, and convergence. Additionally, different parameters of the genetic algorithm were examined to determine their impact on the final performance of the algorithm and to find the optimal settings for clustering data.

METHODS: In this study, to form chromosomes, the number of clusters was first specified. Given that each cluster center has as many features as the dataset, the length of each chromosome was determined by the product of the number of clusters and the number of features. Novel and diverse methods were employed for the processes of Crossover, Mutation, and Survival. Additionally, the fitness score was chosen similar to the k-means algorithm to optimize clustering performance. This innovative approach led to improved accuracy and efficiency in the clustering process.

FINDINGS: By applying the method described in this article to three insurance datasets for fraud detection, interesting results were obtained: a 27% improvement in F1 and a 6% increase in accuracy in the first dataset, a 1% improvement in F1 and a 2% improvement in accuracy in the second dataset, and finally a 2% improvement in F1 and a 9% improvement in accuracy in the third dataset compared to the K-means method and other approaches. Given the binary nature of the data in these datasets, the problem was solved using the algorithm for two clusters, and the best label for each cluster was chosen based on the actual labels of the data. The results were presented in the form of classification outcomes.CONCLUSION: The genetic algorithm has the capability to solve complex problems without definitive solutions and can outperform traditional methods like k-means in clustering data. Using various evaluation metrics such as Silhouette Score, Davies-Bouldin Score, and Adjusted Rand Index (ARI) in this algorithm allows for the improvement of cluster centers. This approach, by incorporating probabilities and randomness, enables the examination of more points as cluster centers and improves clustering performance. The results indicate that this method can, in some cases, perform better than well-known methods and provide a suitable structure for clustering data using the genetic algorithm.

DOI: 10.22056/ijir.2024.369483.3207