The number of transactions using electronic financial instruments has been increasing by about 23% year over year. The global COVID-19 pandemic has only accelerated that process. Electronic means have become the primary vehicle of how people purchase their goods. With this sudden increase in transactions, fraud detection systems are stressed. They need to be much more accurate, much faster than they currently are. This can be done by optimized models using AI.
Here are the five key takeaways from a recent webinar I hosted on how AI can detect fraud quicker :
- Compact vs Comprehensive features. A compact feature that shows speed is better than a comprehensive feature that is slow.
- Balance. A balance between the number of features, the type of features, and the complexity of features is important to ensure the model is fast, accurate, and robust.
- Apply Zero-Prior Knowledge Features. Having features that have no or little prior knowledge lightens the load on the model and increases the speed of reaching a decision. Using this type of feature becomes imperative when it provides a value similar to a feature that uses prior information to detect fraud.
- Build a Simple Model. Keep the model simple and fast. Especially, if you handle transactions in volumes. You will reduce risk by volumes, not by value, which might be efficient.
- If GLM works, then use it. The problem tends to increase in complexity when you try to take a complex, comprehensive model to production. A 100+ feature, deep neural network might become complicated to productionize compared to a simple, fast, GLM model that might be equally effective.
Want more details on each key element? Watch the full webinar here