Author profiling: A machinelearning approach towards detectinggender, age and native languageof users in social media

The task of analyzing a text of interest in order to find demographicinformation of an unknown author based on how the text is written,called author profiling, has become a problem of growing interest.Author profiling can for instance be used in crime investigations togain more information about a possible suspect, which could forexample be a potential violent lone actor that reveals an intentionof committing targeted violence. Machine learning algorithms offers an efficient approach to performauthor profiling. Previous work in author profiling have mostly usedtraditional machine learning algorithms where features are manuallyextracted, which the algorithms base their prediction on. However,in this thesis we compare two different approaches towards authorprofiling, one where the features are manually extracted using thelearning algorithms Support Vector Machine and Feed-Forward NeuralNetworks and one where the features are learned from raw datausing the learning algorithms Convolutional Neural Networks and LongShort-Term Memory. These learning algorithms were used to predictthe gender, age and native language from a text of an unknown author.Also, to get an understanding of how well author profiling works inthe wild, we tested our models on a different domain than they weretrained on. The results showed that in our setting, manually extracted featuresworked better for author profiling. Also, applying learning algorithms toa different domain than they were trained on significantly decreased theperformance. In addition, the results showed that the characteristics ofthe datasets and the chosen hyperparameters for each learning algorithmwere essential for the performance