The seminar run from 11am to 12pm. Please register online
Date: 13th June 2019
Abstract: This talk will cover two topics: imbalanced data and interpretability of convolutional models.
First, we consider how to deal with imbalanced classification where one (positive, anomaly) class is rare, and where measures such as the F1 score are often of interest.
Most methods that consider classification accuracy fail in such situations and re-balancing the dataset or using weighted-accuracy are imperfect solutions.
We first show how the k-NN classifier can be adapted to correct for class imbalance.
We also explain how we can derive guarantees both in terms of learning and generalization for models learned to optimize a weighted-accuracy.
In a second part, we use adversarial approaches in a novel way to have more interpretable convolutional models.
By minimizing the Wasserstein distance between the (learned) convolutional filters and teh sub-series from the dataset, filters become more interpretable as they will look like sub-series.
We show that this approach is promising, showing good behavior in terms of scalability, interpretability and accuracy on 85 time series classification datasets.