https://github.com/aorosoeon/linkedin_statistical_inference

Linkedin automation through Selenium to conduct an A/B test
https://github.com/aorosoeon/linkedin_statistical_inference

ab-testing gspread linkedin orchestration python selenium shell

Last synced: 3 months ago
JSON representation

Linkedin automation through Selenium to conduct an A/B test

• Host: GitHub
• URL: https://github.com/aorosoeon/linkedin_statistical_inference
• Owner: aorosoeon
• License: mit
• Created: 2024-08-28T18:04:05.000Z (almost 2 years ago)
• Default Branch: main
• Last Pushed: 2024-12-14T17:55:24.000Z (over 1 year ago)
• Last Synced: 2025-04-06T03:44:34.943Z (about 1 year ago)
• Topics: ab-testing, gspread, linkedin, orchestration, python, selenium, shell
• Language: Python
• Homepage:
• Size: 631 KB
• Stars: 5
• Watchers: 1
• Forks: 0
• Open Issues: 0
• Metadata Files:
  • Readme: README.md
  • License: LICENSE

Awesome Lists containing this project

README

          
# linkedin_statistical_inference

Is it better to send LinkedIn connection requests with a personalized message or leave them blank? That’s the question I try to answer in this repo.

To prove that one approach has a higher acceptance rate, we need to conduct an A/B test. The main idea of an A/B test is randomization, so the decision of who gets a blank request and who gets a message should be completely random. If I try to do this manually, I will introduce my bias into the test, and it wouldn’t be within the rules of A/B testing.

The architecture below automates this process to eliminate my bias. Here is how it works:

- A Cron job on a Raspberry Pi activates orchestrator.sh, and any error that occurs after that goes to cron.log.

- An arrow under DAILY PROCESS represents a randomized time delay between actions, so after Cron job activation, there is a delay before main.py.

- main.py opens the Google spreadsheet with LinkedIn links and a Chromium browser, then it opens profiles, pulls needed info, and sends them requests (with random assignment of who gets a message and who gets a blank request).

- To be on the safe side, counter updates happen after some time in a separate file to ensure that each day the bot works with new profiles.

- LinkedIn is tricky, so checking_invites.py verifies that these people are added and logs this in a spreadsheet.

- checking_accepts.py pulls people who accepted my invite so that we can conduct an A/B test afterward.

- LinkedIn doesn’t allow more than 700 sent requests, so removing_old_invite_requests.py keeps it under 600.



Below is what the spreadsheet looks like (private data is hidden, of course):



The A/B test was recently completed with the following stats:



I will be releasing a Jupyter Notebook with statistical tests soon, but a preliminary analysis confirms the hypothesis that sending a connection request with a message works much better (all statistics and p-values support this and make it statistically significant).