Predicting Occupation with Machine Learning from Turkish Tweets

Abstract

With the spread of social media platforms and the rapid increase in the number of users, the amount of data produced in social media is growing rapidly. One of the goals of scientific studies to extract information from this data is occupation prediction. Social media users' occupation information can be used in many different areas, especially in smart suggestion systems. In this study, it is aimed to make occupation prediction using Turkish tweets. Within the scope of the study, an occupation dataset consisting of 25,000 Turkish tweets was created and shared publicly. Various preprocessing steps were applied on this dataset, and feature sets were extracted using both the words themselves and the word roots. In the tests, tweets were used both singularly and combined in groups of 5 and 10. In the experiments in which Support Vector Machine and Logistic Regression methods were applied, tests were repeated by feature selection. While the best result was obtained as 74.90% accuracy in the experiments with singular tweets, the best performances were reported as 96.20% accuracy in experiments with tweets combined in groups of 5, and 99.00% accuracy in experiments with tweets combined in groups of 10. It has been seen that the using of word roots in the tests has higher success than using the words themselves, and the feature selection generally increases the success. At the end of the study, these results were discussed and suggestions for future studies were presented.