Learning algorithms have been proposed for a wide variety of tasks: autonomous cars, diagnosis of human diseases, face recognition, translation, etc... However, developing such learning algorithms require expertise in machine learning, mathematics, and computer science. The breakthrough that machine learning has made across many disciplines shows that there are fields where non-experts would benefit from applying machine learning to their domain-specific problems. However, the lack of expertise is a major obstacle preventing the application of machine learning to those problems.
We propose a new method for making machine learning more accessible to non-experts. This method is based on deterministic annealing with mean-field approximations. This method facilitates the design of miners of latent structures for a wide variety of learning tasks, including learning access control policies, business processes, king-and-rook endgame configurations, and elementary mathematical functions.
[PDF]
Organizations define access control policies that restrict who can do what in the organization. Organizations often change in many ways. New users join the organization. Other users leave. New resources may be acquired and others are removed. All these changes induce changes in the policy. When these changes accumulate through time, policies become convoluted and difficult to maintain. Moreover, policy changes are done manually. As a result, unauthorized users may obtain access to sensitive resources and potentially harm the organization.
To assist with the maintenance of access control policies, policy miners have been proposed. These are algorithms that use machine learning and/or combinatorial algorithms to analyze the current assignment of permissions to users and then “mine” a simple policy that is as consistent as possible with the current permission assignment.
The thesis’s goal is to develop a universal policy miner that is independent from the policy language used to specify policies. This universal miner would receive as input a policy language L and an assignment of permission to users. The miner would then output a policy that can be specified in L and that is as consistent as possible with the permission assignment. To develop this universal miner, we follow Unicorn, a method proposed to build policy miners that applies for a wide variety of policy languages, including RBAC, ABAC, and even XACML.
[PDF]
Process mining offers organizations the possibility to better understand how internal processes work. Process mining consists of algorithms that analyze logs produced by the organization’s systems to discover process models that illustrate how products and services are created. Those models are aimed at giving insights for optimizing those processes and, therefore, reducing costs and increasing productivity.
Process mining is gaining traction in industry. Gartner published a market guide for process mining in 2018 that included several common use cases for process mining and an overview of current vendors. However, to our knowledge, process mining uses only combinatorial algorithms and we see potential for algorithms based on machine learning and artificial intelligence.
In this thesis, we approach the field of process mining from a machine-learning perspective. We shall develop a probabilistic process miner that, in contrast to the current state of the art, produces probability distributions over process models rather than producing actual models. Moreover, we shall develop model validation techniques for process mining.
[PDF]