Publications

ACM-BCB 2022 - The 13th ACM Conference on Bioinformatics, Computational Biology and Health Informatics, Chicago, IL, 2022

Abstract: Supervised machine learning models are, by definition, data-sighted, requiring to view all or most parts of the training dataset which are labeled. This paradigm presents two bottlenecks which are inter-twined: risk of exposing sensitive data samples to the third-party site with machine learning engineers, and time-consuming, laborious, bias-prone nature of data annotations by the personnel at the data source site. In this paper we studied learning impact of data adequacy as bias source in a data-blinded semi-supervised learning model for covid chest X-ray classification. Data-blindedness was put in action on a semi-supervised generative adversarial network to generate synthetic data based only on a few labeled data samples and concurrently learn to classify targets. We designed and developed a data-blind COVID–19 patient classifier that classifies whether an individual is suffering from COVID–19 or other type of illness with the ultimate goal of producing a system to assist in labeling large datasets. However, the availability of the labels in the training data had an impact in the model performance, and when a new disease spreads, as it was COVID9-19 in 2019, access to labeled data may be limited. Here, we studied how bias in the labeled sample distribution per class impacted in classification performance for three models: a Convolution Neural Network based classifier (CNN), a semi-supervised GAN using the source data (SGAN), and finally our proposed data-blinded semi-supervised GAN (BSGAN). Data-blind prevents machine learning engineers from directly accessing the source data during training, thereby ensuring data confidentiality. This was achieved by using synthetic data samples, generated by a separate generative model which were then used to train the proposed model. Our model achieved com- parable performance, with the trade–off between a privacy–aware model and a traditionally–learnt model of 0.05 AUC–score, and it maintained stable, following the same learning performance as the data distribution was changed.

Authors: Javier Pastorino, Ashis Kumer Biswas

Paper Link - Presentation - Source Code (Github)

ACMSE 2022 - ACM Southeast Conference, Virtual Event, 2022

Abstract: Wildfires play a critical role in determining ecosystem structure and function and pose serious risks to human life, property and ecosystem services. Burn probability (BP) models the likelihood that a location could burn. Simulation models are typically used to predict BP but are computationally intensive. Machine learning (ML) pipelines can predict BP and reduce computational intensity. In this work, we tested approaches to reduce the set of input features used in an ML model to estimate BP for the state of California, USA, without loss of predictive performance. We used Principal Component Analysis (PCA) to determine the optimal set of features to use in our ML pipeline. Then, we mapped BP and compared model performance when using the reduced set and when using the whole set of features. Models using optimized input achieved similar prediction performance while using less than 50% of the input features.

Authors: Javier Pastorino, Joseph W. Director, Ashis Kumer Biswas, Todd J. Hawbaker

Paper Link - Presentation - Source Code (Github)

IEEE AIKE 2021 - IEEE International Conference on Artificial Intelligence and Knowledge Engineering (AIKE), Virtual Event, 2021 - Virtual

Abstract: Traditional Machine Learning (ML) pipeline development requires the ML practitioner to directly access the data to analyze, clean and preprocess it, in order to develop an ML model, train it and evaluate its performance. When the data owner has no infrastructure for in-house development, such pipelines are outsourced. It is common that data has some level of privacy constraints that will impose a laborious and maybe expensive infrastructure, including among others contracts drafting and infrastructure improvement. Traditional approaches rely either on anonymization which does not entirely protect from identity disclosure, or on synthetic data generation which requires expertise not necessarily available to the organization. In this paper, we present Data-Blind ML, an automated framework, fueled by synthetic generative learning and distributed computing paradigms, which enables an organization to outsource the development and training of ML models without sharing any sample from the real dataset. In addition, the framework allows the ML practitioner to get feedback of the model’s performance against the actual real data without accessing it directly.

Authors: Javier Pastorino, Ashis Kumer Biswas.

Paper Link - Presentation - Source Code (Github)

IEEE AIKE 2020 - IEEE International Conference on Artificial Intelligence and Knowledge Engineering (AIKE), Virtual Event, 2020 - Virtual

Abstract: Machine learning (ML) algorithms are data-driven and given a goal task and a prior experience dataset relevant to the task, one can attempt to solve the task using ML seeking to achieve high accuracy. There is usually a big gap in the understanding between an ML experts and the dataset providers due to limited expertise in cross disciplines. Narrowing down a suitable set of problems to solve using ML is possibly the most ambiguous yet important agenda for data providers to consider before initiating collaborations with ML experts. We proposed an ML-fueled pipeline to identify potential problems (i.e., the tasks) so data providers can, with ease, explore potential problem areas to investigate with ML. The autonomous pipeline integrates information theory and graph-based unsupervised learning paradigms in order to generate a ranked retrieval of top-$k$ problems for the given dataset for a successful ML based collaboration. We conducted experiments on diverse real-world and well-known datasets, and from a supervised learning standpoint, the proposed pipeline achieved $72\%$ top-$5$ task retrieval accuracy on an average, which surpasses the retrieval performance for the same paradigm using the popular exploratory data analysis tools. Detailed experiment results with our source code are available at out Github.

Authors: Javier Pastorino, Ashis Kumer Biswas.

Paper Link - Presentation - Source Code (Github)