Project
Unsupervised Representation Learning and Health Insurance Anomaly Detection
One of the most important reasons behind the advances in machine learning is due to their increasingly powerful representation learning capabilities. The most well-known type of representation learning is a neural network, which is a powerful type of model that can automatically extract features out of input data. However, one has little control over the types of features that a neural network learns. This motivates research into learning representations that can capture prior knowledge, offer insight into the model and can improve the learning process.
In this work, a number of unsupervised representation learning methods are discussed to increase performance, tackle biases, improve diversity and explainability of machine learning models. Specifically, the problem of mode collapse and complete mode coverage in generative adversarial networks is tackled.
Additionally, we extend the restricted kernel machines framework to time-series data such that it can learn time-series features and predict unseen future values.
Finally, a novel representation learning technique in combination with other state-of-the-art methods are employed to create a system that can help health insurance experts detect anomalous profiles.