Machine Learning in non-standard problems
In recent years, new problems and paradigms have appeared in Machine Learning, which, in some way, do not fit into the classical representation (input data vector, output value) or do not follow the conventional dynamics of classification and/or regression paradigms.
Thus, as far as non-standard representations are concerned, in multi-instance learning (MIL), each pattern is represented by a bag containing a variable number of instances, each instance having the same number of attributes. On the other hand, relational learning (RL) aims at working directly with data stored in a relational system, which contrasts with most of the techniques developed so far, which deal with a single table containing all the data. Finally, the paradigms of multi-output learning (MTL) present the output space more flexibly, namely with a vector representing several simultaneous outputs, which correspond to behaviors or functionalities learned concurrently.
As for paradigms that do not fit the learning process’s standard dynamics, we can cite semi-supervised learning or active learning, where both labeled and unlabelled data sets are used to carry out more effective learning. On the other hand, in monotonic classification, the target variable is not of a numerical type but has an implicit ordering, which can enhance the learning process rather than considering this problem as one of conventional classification. Other fields of interest are Multi-Source Learning (MSL) and Multi-View Learning (MVL), which seek to combine heterogeneous and complementary information obtained from various data sets to learn models that bring better results than independent models generated from independent information.
Contact: Sebastián Ventura Soto