Abstract:
This dissertation investigates various learning paradigms’ behaviour on two pattern mining applications: spoof fingerprint detection and automatic hate speech detection on social media platforms. It argues that learning paradigms must consider properties inherently present in the data while deciding the number of hypotheses to be used for classification. These data properties are vital in applications that require finding a specific pattern in a massive amount of data. In our study, spoof fingerprint detection is regarded as an open-set classification task, and the generalization abilities of hate speech detectors are explored rigorously. Therefore, the emphasis is on the performance under cross-sensor, cross-material and cross-dataset environments