Detecting fraudulent activities in financial transactions is a significant challenge, given the dynamic and complicated characteristics of deceptive behaviour. This study presents a category-geospatial fraud detection method that uses category-based risk analysis and geographic distance calculations to improve the identification of fraudulent transactions. The suggested framework incorporates two key features. First, it includes a category risk assessment that evaluates the likelihood of fraud across various transaction categories by examining past fraud rates. This enables the model to pinpoint high-risk categories, such as large purchases or luxury goods. The second feature is calculating the Euclidean distance between the merchant’s location and the user, this feature identifies irregularities, such as transactions occurring in one location that is significantly far away from the user’s usual area of activities. Utilising both of the features and applying them in machine learning models, mainly ensemble techniques like Gradient Boosting significantly enhances fraud detection by spatial analytics and merging contextual insights. The results indicate that mean encoding markedly improves model performance by adeptly integrating previous fraud tendencies into feature representation. The incorporation of distance features slightly enhances fraud detection, indicating a requirement for more advanced spatial analysis. The study’s results emphasise the necessity of integrating sophisticated feature engineering with resilient machine learning methodologies to create scalable and efficient solutions for contemporary financial systems. This study applied comprehensive analyses of the public dataset to demonstrate the effectiveness of the suggested method, despite maintaining interpretative clarity, the study attained significant precision and accuracy in detecting fraudulent activities.