This same example could be extended to credit underwriting, credit card delinquency, or a host of other important class-imbalance binary classification problems.Ī common theme in all financial credit risk modeling is the concern for the expected loss. We also show the simple speedups obtained by an NVTabular data loader for model training. We do this in the context of predicting mortgage delinquencies using the public Fannie Mae mortgage dataset. When GPU-acceleration is utilized many times during model development, the modeler’s time is used much more effectively by amortizing the training time and cost reductions over dozens of model iterations. In this article, we demonstrate how to examine competing models, known as challenger models, and use GPU-acceleration to succeed with easy, cost effective, and understandable application of model explainability. The classic extract-transform-load process (ETL), is a core-starting point in any data science project.įor GPU-accelerated use cases, the NVIDIA MERLIN application framework for deep recommender systems uses NVTabular – an accelerated feature engineering, preprocessing, and data loading library, which can also be used in other domains such as financial services.

Here, we are working with a tabular dataset. RAPIDS makes leveraging GPUs easier for data scientists through interfaces similar to favorites like scikit-learn and pandas. Whichever you pick, GPUs are accelerating data science use cases to the point where any data analysis on large datasets simply requires them for every day convenience, rapid iteration, and results. Deep learning excels in the non-tabular computer vision, language, and speech recognition domains. Machine learning, especially decision trees, and leading up to the more advanced XGBoost models, was maturing earlier than deep learning and has some well-established methods. Others apply the latest deep learning models, while still others pursue an “ensemble” model hoping to get the best of both worlds – explainability and performance. Depending upon the nature of the dataset, some data scientists prefer classical machine-learning approaches.

Data Scientists and Machine Learning Engineers often face the dilemma of “ machine learning compared to deep learning” classifier usage for their business problems.

Edit trees in chief architect library