Skip to content

My project for Insight DS 2020A. Predicts engagement for instagram posts to make better advertising decisions

Notifications You must be signed in to change notification settings

wdlcameron/DeepEngage

Repository files navigation

DeepEngage

DeepEngage is a engagement prediction tool for posts within specific instagram niches. An implementation for food-related posts can be found here

The final product accepts a group of images (simulated through preset groups in the demo) and the sponsored content creator's username, then selects the post with the greatest engagement potential. The number overlayed on each image represents how much engagement it is forecasted to receive relative to the average (e.g. a score of 1.2 is predicted to get 1200 likes if the average for that user is 1000). The "Recommended Optimizations" pane applies a series of filters to the chosen image and recommends changes if they are expected to perform better than the original.

Core Requirements and Installation Instructions

The core packages required are: FastAI V1 (conda install -c fastai fastai)
ImageIO (conda install -c conda-forge imageio)
Selenium (conda install -c conda-forge selenium)

Note: there is currently an incompatibility with Pillow and PyTorch, which requires you to downgrade Pillow (conda install pillow=6.1)

Selenium also requires that the path have access to chromedriver.exe, which can be downloaded from here

Modules

Web Scraping

Models are trained using images and tabular data scraped from Instagram. This process is outlined in the "Web Scraping Pipeline.ipynb" notebook in the main directory

Model Training

Model training can be performed using the "Model Training and Prediction.ipynb" notebook.

Performace Validation

Performance of models can be visualized and explored in the "Post_Processing_Exploration.ipynb" notebook

About

My project for Insight DS 2020A. Predicts engagement for instagram posts to make better advertising decisions

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published