Supervisors info:
Ντούλας Αλέξανδρος, Επίκουρος Καθηγητής, Πληροφορικής και Τηλεπικοινωνιών, Εθνικό και Καποδιστριακό Πανεπιστήμιο Αθηνών
Summary:
In recent years, with the constant evolution of technology and due to the great computing
power, that computer systems have acquired, the field of machine learning has made
rapid progress. As a result, deep machine learning applications now play a significant role
in modern life. The science of data mining and processing is one of most well-known and
at the same time interesting uses which aims to improve our daily lives.
This thesis with the help of data, is an application that focuses on the field of psychology
of people. Its goal is to better understand their personality and consequently, to improve
their lives through it. It is an application of deep machine learning, where with the help of
neural networks, data corresponding to publications of internet users are studied.
Through them, the Myers-Briggs personality type indices corresponding to these people
are predicted. The development of this application was carried out with two different
implementations and in the end the predictions they exported were compared. In the first
implementation, the data are downloaded and then processed so that they are given to
the neural network to be trained. Because the data are quite long sequences, the neural
network model chosen is that of the recurrent neural network. Specifically, the type
chosen to be developed is the LSTM neural network, which is suitable for such data.
However, given the nature of the prediction he is called upon to make, which can be
broken down into four sub-predictions, it was considered more suitable to train four
different models. The four different predictions together make up the Myers-Briggs
personality type index. After completing its training, the models are able to extract their
predictions. The results are considered quite satisfactory, as the models to a large extent
managed to determine the personality type of the users who had made the posts.
The second implementation took place using the BERT method which is recently
developed (2018) by Google. Initially, the data was also downloaded here, but it was
processed much less. The reason is that the operation of this model is based on
bidirectional neural networks, which manage non processed sequences better. Then the
training of four models took place, as in the previous implementation and followed by the
export of forecasts. The evaluation of the results of this implementation is quite close to
the previous one, as the models showed similar behavior during their training.
Finally, with the implementation of this thesis, it is perfectly understood that artificial
intelligence offers unlimited possibilities in improving people's daily lives. Its applications
can deal with problems with minimal use of resources with quick and efficient solutions.
Keywords:
Myers-Briggs personality type, Deep machine learning, Neural networks, LSTM, BERT
