Group Member

  • Sonali Parbhoo

    PhD candidate

    About me

    I am Sonali Parbhoo. I was born and grew up in Johannesburg, South Africa. My research focuses on developing statistical models for problems in healthcare, medicine and computational biology. I am particularly interested in network inference, non-parametric Bayesian statistics and decision-making in uncertainty. In 2014, I completed an MSc. degree in Computer Science from the University of the Witwatersrand before moving to Basel. My MSc. thesis focused on using reinforcement learning methods for HIV therapy selection. Prior to this, I completed my undergraduate studies specialising in Computer science, Mathematics and Molecular Biology. Apart from research, I enjoy travelling, playing badminton and experimenting with contemporary cuisine.

    My projects

    Systems-X HIV-X: Deciphering Host-Virus Interactions to Cure HIV

    Contact Information

    University of Basel
    Sonali Parbhoo
    Department of Mathematics and Computer Science
    Room 01.001
    Spiegelgasse 1
    CH - 4051 Basel, Switzerland

    E-mail: sonali.parbhoo@unibas.ch
    Phone: +41 61 207 0542



    Publications

  • Bayesian Markov Blanket Estimation. Dinu Kaufmann, Sonali Parbhoo, Aleksander Wieczorek, Sebastian Keller, David Adametz, Volker Roth -- In Proceedings of the 19th International Conference on Artificial Intelligence and Statistics (AISTATS) 2016, Cadiz, Spain. JMLR: W&CP volume 51.

    Graphical models are used for representing dependencies among several random variables. In a typical application, the network of dependencies is unknown and the goal is to identify the dependencies from observations. In a high-dimensional setting, identifying the full network can be diffucult. Also, identifying the full network may also be undesirable or irrelevant because one is not interested in parts of the network. Consider the example in gene analysis where the dependency between only a few clinical factors and hundreds of genetic markers is required. In such situations it is advisable to reduce the focus on estimating a sub-network as opposed to estimating an entire network of all the associations. Here, we are looking at undirected networks and focus on a specific sub-network, namely on the Markov blanket. This is the set of variables that, when conditioned on, render the variables of interest conditionally independent of the rest of the network. The goal is to identify the Markov blanket, i.e. identifying the nodes among a large set of candidates which are the neighbours of the query variables.

    We provide a Bayesian perspective of estimating the Markov blanket in an undirected network. This view enables the computation of a posterior distribution and thus offers a means to assess the (un-)certainty of the network. This contrasts with the maximum likelihood approach of the graphical lasso which only provides a point estimate of the network. We show, how to avoid a limiting inversion when estimating a network in the context of a Markov blanket. Our posterior distribution has an analytic form and can be efficiently sampled from. Overall, we show that the Markov blanket can be estimated efficiently without explicitly inferring the entire network.
  • Combining Kernel and Model Based Learning for HIV Therapy Selection.. Sonali Parbhoo, Jasmina Bogojeska, Maurizio Zazzi, Volker Roth, Finale Doshi-Velez .To appear in Proceedings of the AMIA Summit on Clinical Research Informatics (CRI), 2017.

    We present a mixture-of-experts approach for HIV therapy selection. The heterogeneity in patient data makes it difficult for one particular model to succeed at providing suitable therapy predictions for all patients. An appropriate means for addressing this heterogeneity is through combining kernel and model-based techniques. These methods capture different kinds of information: kernel-based methods are able to identify clusters of similar patients, and work well when modelling the viral response for these groups. In contrast, model-based methods capture the sequential process of decision making, and are able to find simpler, yet accurate patterns in response for patients outside these groups. We take advantage of this information by proposing a mixture-of-experts model that automatically selects between the methods in order to assign the most appropriate therapy choice to an individual. Overall, we verify that therapy combinations proposed using this approach significantly outperform previous methods.