*Note: There is a pdf version for your convenience.

Repo structure:

• main.ipynb contains data processing and model training.
• model_explanation.ipynb contains some beautiful visualizations with SHAP library

The original dataset is given here

This dataset contains 18K job descriptions out of which about 800 are fake. The data consists of both textual information and meta-information about the jobs. The dataset is highly imbalanced with only 4.84% positive values.

Due to imbalanced classes possible metrics were AUC-ROC or AUC-PR. I prefer AUC-PR because finding positive class is more important for the task. AUC-ROC treats each class equally, however AUC-PR is more sensitive for the positive label.

Initially dataframe consists of 16 features. Majority of them are strings. There are some categorical features.

For text data I simply add symbol-length feature.

Furthermore, I wanted to know the importance of each feature for the prediction. For this reason I simply trained several gradient boosting models on each single feature.

len_company_profile has the biggest importance for the target value. Due to this fact I supposed to add some features with company_profile description. It is a text feature so that tf-idf representation could help to increase performance.

There were trained various model.

Comments:

1. LogReg was used as a simple baseline. It has really poor quality. It means data has a lot of nonlinear dependencies.

2. There were trained three gradient boosting models.

The best model gets 0.87 AUC-PR on test dataset which is quite good final score.

After training the model I used SHAP library for interpretation. More examples can be found in model_explanation.ipynb file.

The main interest is feature importance for model prediction.

Obviously len_company_profile and has_company_logo has the biggest impact on predictions.

So empirically we can predict that post with poor company description and without logo is a fake.

In this plot it is clearly visible that shorter company description leads to higher impact. Furthermore there is a correlation between the len_company_profile and has_company_logo. Companies with short description usually do not have logo.