4LLab : UniSA Data Analytics Group

Scholarships

Assuring competency of machine learning systems through causal explanation

For International or Australian domestic applicants

Supervisors: Prof. Lin Liu

Further information can be found at Project Website

The goal of machine learning (ML) systems is prediction accuracy. ML is not concerned with explanations for outcomes and thus is incapable of articulating its competency. With ambiguous cases and cases where there are conflicting opinions between humans and ML models, it is essential for a ML system to explain not only how reached its decision, but also why the other decision options were not taken.
There are methods of explanation which are based on statistical associations, but they may not imply causation. We need to focus on the causal explanation for ML which is just starting to emerge in research. The mainstream methods look at the causal impact of structural components of a model – for example, neuros or a layer structure in a neural network on predictions, which is not the explanation we want, since they explain a model instead of predictions.
Our project will develop methods to provide causal explanation for why a prediction is made for a case. We will also develop methods to provide contrasting explanations as to why one prediction and not another is made on a case.

If you are keen on working on cutting-edge research within a dynamic research group, please feel free to contact me with your CV, transcripts etc. via Lin.Liu@unisa.edu.au.

Advancing machine learning for human-machine partnership in decision making

For International or Australian domestic applicants

Supervisors: A/Prof Lin Liu and Prof Jiuyong Li.

Further information can be found at Project Website

Recent development in machine learning has seen machine learning models with extremely high accuracy in areas such as image classification and speech recognition. However, for decision support, it has become clear that we do not expect machine learning systems to replace human, but we do not want to only use them as tools either. Rather, we want them to contribute to our decision making as partners. In critical areas, decision makers constantly face complex environments with both unintentional and intentional unknown situations. To support human-machine partnership in such situations, beyond accuracy, it is essential for machine learning systems to be competency aware and explanatory, i.e. when they see a case, first they need tell their human partners whether they know the case or not, and why, rather making a prediction for the case bluntly. Meanwhile, they need to be adaptable to unknown situations. This project aims to develop novel machine learning techniques to transform machine learning systems from tools to partners of human decision makers.

If you are keen on working on cutting-edge research within a dynamic research group, please feel free to contact me with your CV, transcripts etc. via Lin.Liu@unisa.edu.au.

Developing causal-based methods for recommending the repurposed drugs for a disease and applications in breast cancer and SARS-CoV-2

For International or Australian domestic applicants

Supervisors: A/Prof Thuc Le, A/Prof Lin Liu and Prof Jiuyong Li

Further information can be found at Project Website

It costs about 2.6 billion US dollars to develop a new drug and can take up to 17 years for FDA approval. Finding new uses for already approved drugs (drug repurposing) avoids the expensive and lengthy process of drug development. For example, nearly 70 existing FDA approved drugs are currently being investigated to see if they can be repurposed to treat COVID-19. Existing approaches, however, have the following drawbacks:

  • Existing approaches find the association between drugs and proteins. However, a drug represents an intervention in the system and only a causal framework allows for predicting the effect of an intervention. It is, therefore, critical to develop causal-based approaches for recommending the candidate repurpose drugs.
  • Existing approaches aim to find the repurposed drugs that treat disease in general, i.e. assuming everyone who has the disease will be treated the same way. Working toward personalised medicine, we will need a new approach to finding repurposed drugs for individuals.
In this project, we will develop causal recommending systems for suggesting repurposed drugs for breast cancer and SARS-CoV-2. Given the severity of the COVID-19 pandemic, our work contributes to the international challenge of rapidly repurposing the existing approved drugs for clinical interventions.

If you are keen on working on cutting-edge research within a dynamic research group, please feel free to contact me with your CV, transcripts etc. via Thuc.Le@unisa.edu.au.

Using raw satellite imagery to detect early fire smokes on small satellite

For Australian domestic applicants

Supervisors: A/Prof Jixue Liu and Prof Jiuyong Li

Further information can be found at Project Website

A large number of devastating wildfires caused life threats and property damage across the globe. Early detection of wildfires is important to mitigate the threats and damage. Detection of fires via the detection of smokes is seen as an effective approach because early fires are small in size, low in temperature and hard to detect directly. Satellite image-based fire smoke detection provides a good approach to early fire detection. Existing studies depend on pre--processed data which is transformed from satellite raw data. Preprocessed means that the readings (raw data) from satellite sensors have been modified to rectify distortion, remove certain physical and atmospheric effects, and improve the accuracy of data. For example, raw data from satellites is modified to remove the effect of mist, time, and cloud on the reflectance. Preprocessing helps data visualization and makes the machine learning model development easy. However, such data processing consumes a lot of time and power and is not suitable for early fire detection on small satellites which have limited computation, communication, and electrical power. Pre-processing of imagery data is time consuming and should be simplified, and it prolongs the fire smoke detection time. When fire smoke detection is done on the satellite (onboard processing), preprocessing causes even more problems. Satellite senses hundreds of squared kilometers per second. As sensor technologies develop, the spatial and spectral resolutions of sensing become higher, which leads to larger amount of data to be produced per second. At the same time, onboard processing is limited by power consumption, limited computation power (like the number of GPUs) in addition to a large amount of data. All these raise the need to simplify the existing smoke detection workflow.

If you are keen on working on cutting-edge research within a dynamic research group, please feel free to contact me with your CV, transcripts etc. via Jixue.Liu@unisa.edu.au.

Econometric model selection through testing

International student or Australian domestic student

Supervisors: Dr Zen Lu , Prof Chandra Krishnamurti and Prof Jiuyong Li

Econometrics plays a vital role in informing and shaping government policy in relation to gathering quantitative evidence. Econometric study has become an important part of contemporary economic study. This is evidenced by the many recent Nobel Economics prizes that have been awarded for the important econometric contributions by the economists/econometricians James Heckman (2000), Daniel McFadden (2000), Robert Engle (2003), Clive Granger (2003), Thomas Sargent (2011), Christopher Sims (2011) and Lars Hansen (2013). The objectives of an econometric study are to discover and to verify the relationship between economic variables based on information extracted from observational or experimental studies. Therefore, it is crucial that more accurate and appropriate models are identified, not only to advance our knowledge, but also to assist governments to make decisions in an increasingly uncertain global economy and financially challenging times.

The aims of this project are:

  • to develop a new class of tests for model selection and evaluation.
  • to develop a test-based approach for model selection without a need to estimate models
  • to apply our new technology for model selection and evaluation in a range of finance models.

If you are keen on working on cutting-edge research within a dynamic research group, please feel free to contact me with your CV, transcripts etc. via Zen.Lu@unisa.edu.au.

Positions