Jobs From Headhunter Portal

This thesis is devoted to the study of such a task, as forecasting wages according to the job description. The main purpose of the work is to explore the possibilities of a number of methods and approaches from the field of data analysis and machine learning in application to solving this problem. The work implemented a system for collecting and preparing data for analysis, parsers were written to collect additional geographical features, to clear job descriptions from HTML tags. An exploratory analysis of the data was carried out, the relationship of the features with the target variable was investigated, the most informative features were highlighted, such as the name of the region, professional role, currency in which wages are indicated, work experience, type of employment, pre-tax or after income were indicated. Quality metrics were selected to evaluate the models - MAE and MAPE were used. A number of models were trained on the processed data, data enrichment and processing were performed using methods such as geocoding, TF-IDF conversion, and embedding extraction from the rubert-tiny model. Models such as random forest and Catboost were studied. For each trained model, the quality of its work was evaluated. The results of experiments with model training were tabulated, and the best approach among the subjects was identified. The best quality was achieved using an approach in which a Catboost model was built on a set of features and a job description, for which suitable hyperparameters were selected using brute force.