Previous Research

I have made significant contributions to the field of machine learning and nature-inspired computational techniques through my publications. My research journey began by introducing a tabu local search heuristic to particle swarm optimization, which effectively addressed a Quadratic Assignment Problem (QAP). I then presented further enhancements to this algorithm in subsequent papers.

Expanding my work into data mining approaches, I enhanced the Ant Miner approach by introducing a new pheromone model that greatly improved runtime efficiency. Moreover, I demonstrated the advantages of a hybrid approach that combined archive and graph pheromone models, resulting in better accuracy and runtime for the algorithm. These algorithms are deeply rooted in comprehensive machine learning models.

I extended the Ant Miner framework to tackle stream mining problems as well, enabling the use of explainable models in dynamic environments. This contribution marked a significant stride in incorporating explainable models into stream mining.

In collaboration with the Plead team, I started working explainability and model acceptance, working alongside Experian. Together, we utilized provenance as an explanatory tool for the machine learning models of companies. The central idea was to trace the provenance of the data used to train the machine learning models, enabling companies to provide transparent and justifiable explanations for their methodologies. Our interdisciplinary research led to publications in legal conferences, highlighting the practical applications of our work. Notably, our provenance-based approach, adopted by PLEAD, was cited by the Information Commissioner’s Office (ICO) in their official recommendations for all organizations involved in data collection and processing.

Recent projects that was supervised with masters students.

  • Using explainable machine learning models to improve the shelf life of COVID diagnostic test strips.

  • Using natural language processing models and machine learning models to detect the risk factor for people undergoing substance abuse treatment.

  • Using hybrid deep learning to predict the mortality of patients with heart failure.

Future work grants in the pipeline.

  • Credit card fraud detection using stream mining approaches to help with continuously changing environment.

  • Detect the risk factor for substance abuse treatment using temporal machine learning models with a sequential learning approach that refines predictions as treatment progresses.

  • Federated learning for health care data using explainable machine learning models.