OPTIMIZING CLUSTERING OF K-MEANS ALGORITHM USING PARTICLE SWARM OPTIMIZATION FOR CREDIT CARD FRAUD DETECTION

Since then, online shopping has grown by leaps and bounds, and most of those digital checkouts now work with credit cards. But plastic used to solely be found in wallets; today it accompanies data packets across multiple servers. The negative is exposure: $24.26 billion in credit card fraud was written off in 2018, leaving actual consumers with a large percentage of the losses. Surveillance companies call a report fraud, and then they have to convince an angry cardholder to wait until the charges are dropped. The consumers who are too elderly to take the jump and start utilizing payment applications are always the first to feel the pain when they scroll by fewer of those pop-up boxes that keep us from seeing the terms and conditions. Supervised learning is the basis of traditional defenses. This means that all hijacked charges must be identified before a detector can be trained. Before any algorithm runs, there is a lot of work to be done in labeling, annotating, and grinding. Researchers still do all of that. The newest chorus attempts to accomplish things on its own, grouping data without pre-markers. This means it has to give up speed for certainty. Here, we look at some of the same no-label techniques to see if they may still be useful as fraud evolves. To achieve that goal, we build a hybrid architecture called PSOKClus, which combines particle-swarm optimization with K-means. The approach lets swarms change the centers of clusters on the fly, which cuts down on distance, tightness, and outlier noise in a single layer. The combination of Particle Swarm Optimization with K-means aims to address the challenge of centroid initialization, a factor often cited as a cause for diminished efficacy in fraud detection. After we have the temporary centroids that have been stabilized and hard assignments are established, the cluster-specific Interquartile Range approach looks for spots that are very different from the regular behavior of each group. A number of criteria, such as the Silhouette Score, Davies–Bouldin Index, Dunn Index, overall accuracy, and Purity, show that the PSOKClus variation beats the traditional K-means by a wide margin. The setup reduces the variance within each cluster and increases the distance between the cluster centroids. It also labels flagged transactions without using labeled samples to help shape the clusters themselves, which creates a unique profile of what confirmed fraud cases look like. This hybrid architecture reliably merges coherence, separation, and functional anomaly detection over several trials. DO NOT BREAK UP PARAGRAPHS.