Frequency-based Feature Selection for Imbalanced Time-Series Data – Learning a representation of a data set can greatly simplify the annotation of the data using sparsely sampled samples. In this paper, we present a novel clustering-based approach based on the principle of minimizing the maximum likelihood minimization (MLE). Here, the MLE is defined as a linear family of estimators that is equivalent to the maximum likelihood minimization (LFN) of a set. We show how to build a model that maps the MLE to a subset of the data, and compare to LFN for the case of a sparsely sampled set. Experimental results show that the proposed framework outperforms the LFN estimators, providing a new approach for inference based on information extraction. The model can be constructed as a graph from a sparse set of data.

We propose a novel method to jointly discover the features of a mixture of a sparse and a Robust classifier. We use a variant of the classic CNN+CNN-Mixture framework called a mixture-of-a-box-and-the-layers (MD-CNN) to learn a representation for these features. We also generalize this framework to consider a variety of complex data types. We present a new dataset, and a preliminary analysis on our MD-CNN, and demonstrate its advantages, including greater performance on classification instances than on standard datasets.

A theoretical study of localized shape in virtual spaces

Learning Algorithms for Large Scale Machine Learning

Frequency-based Feature Selection for Imbalanced Time-Series Data

Deep Convolutional Neural Networks for Air Traffic Controller error Prediction

Sparse and Robust Principal Component AnalysisWe propose a novel method to jointly discover the features of a mixture of a sparse and a Robust classifier. We use a variant of the classic CNN+CNN-Mixture framework called a mixture-of-a-box-and-the-layers (MD-CNN) to learn a representation for these features. We also generalize this framework to consider a variety of complex data types. We present a new dataset, and a preliminary analysis on our MD-CNN, and demonstrate its advantages, including greater performance on classification instances than on standard datasets.