SE Radio 594: Sean Moriarity on Deep Learning with Elixir and Axon

Fraud Models

Designing effective fraud detection models requires careful consideration of feature selection and architecture choice. Sean Moriarity explains that a model trained on a dataset with a high percentage of legitimate transactions might inaccurately label all transactions as legitimate, thus failing to detect fraud. To address this, he emphasizes the importance of precision and tracking true positives and negatives, using gradient descent to minimize loss and update model parameters 1. Gavin Henry summarizes the process, highlighting the role of labeled datasets in training models to differentiate between fraudulent and legitimate transactions 2. Moriarity notes that Axon represents models through Elixir structs, translating them into functions for predictions and training.

   

Data Challenges

Data preparation and feature engineering are crucial in fraud detection, as Sean Moriarity points out. He describes the complexity of maintaining version control over features, which involves updating models with new features identified by data science teams 3. This process can become challenging when dealing with unbalanced datasets, where a small percentage of transactions are fraudulent. Moriarity stresses the importance of differentiating good data from bad, using features like transaction location and amount to identify potential fraud 4.

   

Model Evaluation

Evaluating and training models effectively is essential for successful fraud detection. Sean Moriarity highlights the importance of testing multiple models to find the best fit for a dataset, considering factors like accuracy, precision, and cost 5. He explains that training involves creating a pipeline that applies gradient descent to update model parameters, tracking metrics such as accuracy 6. Moriarity advises against relying solely on complex deep learning models, suggesting that simpler models like decision trees can sometimes outperform them in terms of interpretability and deployment ease.