DSFSI Monthly Deep Dive Talks

Jessica Nemasisi: Dynamic event detection on social media data streams. [1 March 2019]

Abstract:
Event detection has attracted a lot of significant amount of research interest in social media analysis. However, most existing approaches from related literature focus on event detection applied on batch data.In this study we present an algorithm that efficiently detects events on text streams. The algorithm addresses the challenges of prior algorithms in that it does not need to constantly reuse all prior text documents, it further adaptively resolves the temporal scale needed to connect prior documents and tries to reduce the computational overhead by using a hashing approach. We apply our algorithm to the challenge of dynamic event detection on social media data.

Bio: Jessica Nemasisi is a Data Science masters student at the University of Witwatersrand and based at the CSIR. Jessica holds a computer science honours degree from the University of Venda. She was a data science intern at the CSIR (Modelling and Digital Science) from September 2015 until September 2016. Jessica hailds from Venda, Limpopo, in a rural area called Tshidimbini. One of her life goals is to introduce more rural girls like herself to all the cool things one can do as a Data Scientist.

Rendani Mbuvha: Bayesian Inference in Neural Networks via Hamiltonian Monte Carlo [29 March 2019]

Bio: Rendani is a PhD Candidate in Artificial Intelligence at the University of Johannesburg and a lecturer in the School of Actuarial Science and Statistics at the University of Witwatersrand. He is a fellow of the Actuarial Society of South Africa and a Chartered Enterprise Risk Actuary. He holds a Bsc Honours in Actuarial Science from the University of Cape Town and an Msc in Machine Learning from KTH, Royal Institute of Technology in Sweden.

Abiodun Modupe: Authorship Identification in Social Media: Deep Learning Approach Presenter [29 March 2019]

Abstract: To identify and track cybercriminals in social media platforms, investigators employ carefully engineered textual features along with traditional text mining techniques to capture criminal signatures. This process is usually manual, time-consuming and highly dependant on the engineered textual features which may result in retrieved pieces of evidences not trustworthy for pre-trial investigations. In this work, we aim to enhance this process. We employ a multi-channel convolutional neural network over textual characters to extract linguistic features.

Bio: Abiodun is a PhD Candidate at the University of the Witwatersrand and a visiting PhD student at the CSIR Modelling and Digital Science.