Interpretable Vehicle Insurance Claim Fraud Detection Using Categorical Preprocessing, SMOTE-NC and a Decision Tree
This paper demonstrates that a carefully preprocessed Decision Tree classifier, utilizing SMOTE-NC for oversampling and categorical encoding, can effectively prioritize vehicle insurance fraud with high recall and full interpretability, despite achieving modest precision on an imbalanced dataset.
Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
Every year, insurance companies process millions of vehicle claims, a massive stream of paperwork that includes everything from minor fender benders to total losses. While most of these requests are legitimate, a small but costly fraction are fraudulent, involving staged accidents or inflated repair bills. This deception drives up costs for everyone, forcing insurers to raise premiums for honest drivers and requiring human investigators to spend countless hours sifting through files to find the fakes. The challenge for data scientists is to build a system that can spot these rare bad actors without getting lost in the noise of millions of normal claims. The difficulty lies in two main areas: the sheer rarity of fraud, which makes it easy for a computer to simply guess "no fraud" and still be right most of the time, and the messy nature of the data itself, which is filled with text-based categories like car brands, policy types, and accident locations rather than neat numbers.
A researcher at Sanjivani University in India set out to see if a computer could learn to spot these frauds using a very simple, transparent method, provided the data was prepared with extreme care. Instead of building a complex, black-box artificial intelligence that is difficult to understand, the study focused on a single decision tree, a type of model that works like a flowchart of yes-or-no questions. The goal was to determine whether the key to success lay in the sophistication of the model or in the discipline of cleaning and organizing the data before the computer ever saw it. The researcher took a public dataset of over 15,000 vehicle insurance claims, where only about six percent were actually fraudulent, and applied a strict series of steps to clean the information. This included removing unique identifiers that could trick the computer into memorizing specific cases, converting text categories into a format the machine could understand, and carefully balancing the training data so the computer would pay attention to the rare fraud cases rather than ignoring them.
The results showed that this disciplined approach worked remarkably well for a simple tool. When tested on a fresh set of claims it had never seen before, the model correctly identified more than half of the fraudulent cases, catching 150 out of 277 hidden fakes. This was a significant improvement over a standard approach that would simply label every claim as honest, which would have missed every single fraud. However, the study also highlighted a necessary trade-off: to catch that many fakes, the system also flagged about 700 honest claims as suspicious, meaning that for every six alerts the system generated, only one was actually a fraud. This is a familiar reality in fraud detection, where the cost of missing a scam is high, but the cost of annoying an honest customer is also real. The model did not try to automatically deny claims; instead, it acted as a highly efficient filter, sorting the massive pile of paperwork and handing the most suspicious ones to human investigators for a closer look.
What made this work stand out was not the complexity of the algorithm, but the clarity of the rules it produced. Because the model is a decision tree, every prediction it makes can be traced back to a specific set of conditions, such as the type of vehicle, the policy details, or the nature of the accident. The study found that the most important factors in spotting fraud were the fault assigned in the accident, the base policy type, and the specific kind of insurance policy held. These findings align with what human investigators already suspect, but the computer arrived at these conclusions by learning from the data patterns alone. The research suggests that for industries where transparency is required, such as insurance, a well-prepared simple model can be more valuable than a complex one that cannot explain its reasoning.
The study concludes that while the system is not perfect and requires human oversight to handle the false alarms, it offers a powerful, low-cost way to prioritize investigations. By focusing on how the data is prepared rather than on building more complicated machines, the researcher demonstrated that a single, interpretable decision tree can effectively separate the signal from the noise in a sea of insurance claims. This approach allows regulators and companies to see exactly why a claim was flagged, building trust in the system while ensuring that the few bad actors are not hiding in the crowd of millions of honest drivers. The work serves as a reminder that in the world of data science, sometimes the most effective tool is not the most powerful one, but the one that is most carefully tuned to the reality of the problem.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.