Predicting DataScienceSalaries from Glassdoor

• Created a tool that estimates data science salaries (MAE ~ $ 11K) to help data scientists negotiate their income when they get a job.
• Scraped over 1000 job descriptions from glassdoor using python and selenium
• Engineered features from the text of each job description to quantify the value companies put on python, excel, aws, and spark.
• Optimized Linear, Lasso, and Random Forest Regressors using GridsearchCV to reach the best model.
• Built a client facing API using flask

Code and Resources Used

Python Version: 3.7
Packages: pandas, numpy, sklearn, matplotlib, seaborn, selenium, flask, json, pickle
For Web Framework Requirements: pip install -r requirements.txt
Scraper Github: https://github.com/arapfaik/scraping-glassdoor-selenium
Scraper Article: https://towardsdatascience.com/selenium-tutorial-scraping-glassdoor-com-in-10-minutes-3d0915c6d905
Flask Productionization: https://towardsdatascience.com/productionize-a-machine-learning-model-with-flask-and-heroku-8201260503d2

Web Scraping

Tweaked the web scraper github repo (above) to scrape 1000 job postings from glassdoor.com. With each job, we got the following:

• Job title
• Salary Estimate
• Job Description
• Rating
• Company
• Location
• Company Headquarters
• Company Size
• Company Founded Date
• Type of Ownership
• Industry
• Sector
• Revenue
• Competitors

Data Cleaning

After scraping the data, I needed to clean it up so that it was usable for our model. I made the following changes and created the following variables:

• Parsed numeric data out of salary
• Made columns for employer provided salary and hourly wages
• Removed rows without salary
• Parsed rating out of company text
• Made a new column for company state
• Added a column for if the job was at the company’s headquarters
• Transformed founded date into age of company
• Made columns for if different skills were listed in the job description:
  • Python
  • R
  • Tableau
  • SQL
  • AWS
  • Spark
• Column for simplified job title and Seniority
• Column for description length

EDA

I looked at the distributions of the data and the value counts for the various categorical variables. Below are a few highlights from the pivot tables.

Model Building

First, I transformed the categorical variables into dummy variables. I also split the data into train and tests sets with a test size of 20%.

I tried three different models and evaluated them using Mean Absolute Error. I chose MAE because it is relatively easy to interpret and outliers aren’t particularly bad in for this type of model.

I tried three different models:

• Multiple Linear Regression – Baseline for the model
• Lasso Regression – Because of the sparse data from the many categorical variables, I thought a normalized regression like lasso would be effective.
• Random Forest – Again, with the sparsity associated with the data, I thought that this would be a good fit.

Model performance

The Random Forest model far outperformed the other approaches on the test and validation sets.

• Random Forest : MAE = 11.22
• Linear Regression: MAE = 18.86
• Ridge Regression: MAE = 19.67

Productionization

In this step, I built a flask API endpoint that was hosted on a local webserver by following along with the TDS tutorial in the reference section above. The API endpoint takes in a request with a list of values from a job listing and returns an estimated salary.