Techniques

  • Reasoning under uncertainty is an important topic in Machine Learning. Bayesian Networks (BNs), a marriage between graph and probability theory have been introduced as powerful tools for this task. Within the Computer Modeling Group, BNs have been studied as tools to model and reason about causal knowledge. Several structure learning algorithms and advancements on causal inference learning algorithms, under simplified assumptions, have been implemented. BNs have also been used as classification tools for several applications.

  • Among data mining tools, the applicability of cluster analysis is perhaps the broadest. It can be used in the design of any strategy for analysising data from all kind of applications. We use our expertise in cluster analysis to solve problems from computational biology (disease discovery) and also remote sensing imaging (endmembers discovery) by combining it with application specific analytical tools (for spectral unmixing in this case).

  • Evolutionary Game Theory (EGT) studies the behavioral dynamics of a population of players, engaged in repeated interactions. EGT uses biological principles like fitness and natural selection to determine the outcome of a repeatedly played game. Successful behaviors spread in the population by genetic reproduction, while strategies with lower fitness become extinct. In this way EGT can determine which behavior is evolutionary stable in a population and which can be invaded by a small group of individuals, playing an alternative strategy.

  • In data mining and machine learning, feature selection and feature extraction play a crucial role. These analytical tools are very often preceeding the algorithms designed for any type of learning from a set of recorded data. In many applications requiring data mining and machine learning tools, the joint use of feature selection and extraction tools could increase the learning accuracy.

  • Linguistic utterances are full of errors and novel expressions, yet linguistic communication is remarkably robust.  Fluid Construction grammar uses a double-layered architecture for open-ended language processing, in which ‘diagnostics’ and ‘repairs’ operate on a meta-level for detecting and solving problems that may occur during habitual processing on a routine layer.

  • One of the best known methods for regression and classification are Neural Networks (NNs). Within the research of the Computer Modeling Group, NNs play an important part as tools and baseline comparisons for newly developed classification/regression algorithms.

  • Reinforcement Learning is the process of learning through trial-and-error interactions with your environment. A reinforcement learner repeatedly performs actions under different circumstances, in order to discover their consequences and determine a sequence of actions that will lead to its goal.

  • Our group has expertise in supervised learning methods for classification problems (Neural Networks, Support Vector Machines, Bayesian Learning). We deal with data recorded from a wide range of real-life applications: genomics, geography, chemistry and we work closely with researchers from other VUB Departments, offering them our expertise.  

  • A Support Vector Machine (SVM) is a very potent technique for solving high-dimensional non-linear classification problems. One of the main issues with SVMs is the identification of the correct Kernel, which can be highly problem-specific. Within CoMo a new class of Kernels has been developed which are based on a metric taking into account context. These are specifically well suited for sequential data classification such as secondary protein structure prediction and text classification.