Xue Li

Analytical Queries on Big Data

Description:

Traditional database queries are used to search for facts from structured database such as RDB (Relational Databases) to satisfy user search conditions. With big data currently available in many ways such as structured and unstructured multi-modalities, user queries should be constructed not only for searching facts, but also for searching patterns, emerging events, and outliers from available big data. This PhD research is to propose a new type of query language that can query on analytic results , to satisfy user requirements for informed decision support. In order to make such a language to be implementable on general big dataset, this PhD research will also define and design a framework that can answer declarative analytic queries by a data-driven approach to apply transparent machine learning algorithms in order to discover unexpected patterns, emerging trends, various correlations from big data. The challenges of this research will be on how to use an end-to-end black-box mechanism to provide big data analytic services to make big data available for general queries beyond classical data warehousing technologies.

Background:

In classical DSS systems based on data warehouses and OLAP operations, the queries such as Canned and Continuous Queries would not involve procedural operations that can reflect the dynamic parameters of queries. The operators such as Role-Up, Drill-Down, Slice/Dice, Cube, Pivoting etc, cannot reflect the context of the query objects in their business context. This PhD research will try to introduce more flexible analytical data manipulation operations based on machine learning algorithms that can provide end-to-end queries for strategic DSS with baselines.

Privacy Preservation for Sharing Distributed Big Data

Description:

Predictive data analytics usually involves Big Data that is distributed in different locations and owned by different organizations, such as the Taxation Office Data, Boarder-Control Customs Data, Crime-Stop Police Data, and Social Security Data. The organizations are legally responsible for the privacy preservation of their data which is of highly risk and sensitive. However, this should not prevent the sharing of those de-identified, privacy preserved data sets for the predictions of pending social-economic events, emerging trends, patterns of relationships, or correlations among entities. Currently, there are many algorithms that can preserve privacy for computing data from multiple owners, such as SMC (secure multi-party computation), Differential Privacy algorithms. However, the predictive tasks often require to use all original raw data for the learning. This would involve the individual organizations to conduct local learning tasks and contribute to global learning with their local models, instead of their sensitive data. Federated learning therefore coming to being as a promising and useful approach to learn from individual datasets and producing a general model for the required predicting tasks. This project is to research on the Federated Learning algorithms that can deal with large distributed, sensitive datasets and derive a computational model to predict some pre-defined tasks. The challenges of this project would be the following three issues in one solution, i.e., data shareability, data privacy, and computational utility.

Key Terms: Federated Learning, Deep Learning, Distributed Database Technology, Privacy Preservation, Mathematical Modelling, Data Shareability, Computational Utility

The First Principle AI (FAI) Research

Description:

Artificial Intelligence (AI) applications are mostly based on the first-order thinking that is reasoning based on deduction, abduction, induction, or eduction. In this way, AI is limited and unable to discover the First Principles such as those in sciences and complex Math Equations, and laws in Physics and Chemistry. However, this should not prevent AI to be used together with the First Principles in those discovery projects. This research is to design an architecture of AI Application platform that can use First Principle in AI to speed up the human trial-and-error process of experiments, to use First Principle in a more intelligent way to converge an optimization process which has a large number of iterations faster and scalable for human's research problems.