The Machine Learning Group of the Computer Science Department, Turin University, dates back to the 1984. The group belongs to the Artificial Intelligence area and it is composed of two associate professors, three assistant professors and five short term collaborators.
The Machine Learning group (ML) began its research by studying and developing: i) systems to learn symbolic concepts (ML-SMART, SMART+, ENIGMA, and RIGEL), ii) inductive database systems (MINERULE), and iii) neural networks (local receptive models, radial basis functions, and sigmoidal activation functions). These models have been actively used for classification, diagnostic and prediction.
The ML group also studied in depth how symbolic knowledge learnt using classical learning approaches could be used to initialize a neural network that would then be optimized by means of reinforcement learning techniques. Finally, the ML group maintained a focus on Computational Learning Theory (COLT) developments. In particular it studied systems supporting the recursive definitions of concepts and methods for semi-automatic knwoledge extraction.
Today the ML group continues to develop the MINERULE system and to study other important and facinating areas of Machine Learning. Actual interests include:
- extendig the MINERULE system to support the analysis of geo-localized information (GIS) to allow land monitoring;
- privacy preserving techniques for social networks;
- fraud detection in bank transactions;
- creation of tematic maps based on annotations derived from social sensors and crowdsourcing;
- clustering and co-cloustering algorithms, (hierarchical and multidimentional), allowing building models for knowledge representation and inference;
- using neural models for web mining and efficient representation of complex and heterogeneous data by means of projections into semantical spaces or by matrix decomposition;
- efficient and exact algorithms for MAP inference in graphical models;
- learning in context where examples have poor quality labels or are only partially labelled, such as in constrained clustering problems, semi-supervised learning, anomaly detection, or "one-class" learning.