Marky Challenge

Inspiration

Marky's dedication to providing everyone with their own social media manager is an inspiring image of the future. We imagine the world would be filled with more opportunity and more creative content in this vision of the future.

What it does

Our model allows us to predict whether or not a future client would accept or reject the AI generated based upon the restrictions.

How we built it

The first steps we used to build this model was using the tabular dataset to examine the data. We were able to write new features, including the caption_size feature, which appeared to have a significant correlation with whether or not the social media post was accepted. In this stage, we used a XGB Classifier in order to determine whether or not our social media posts were accepted.

The next steps revolved around the images. We attempted to create a CNN, which would allow us to view the image predict, whether or not our client would approve of our post. We attempted to k-cross-validate the CNN. However, this step would not matter much. We will cover the reasons for this in the Challenges section.

Our final step was to use both together to put our two solutions together. For this we used forms of regression, and we recreated our 2 models to give us percent chances of approval, rather than a True or False value.

Challenges we ran into

Most of the major challenges with this problem revolved around the images. Lots of time was spent trying to figure out how to correctly load the images onto the CNN. On top of that, due to many time restrictions and unfamiliarity with CNN architecture, we found it quite difficult to create a useful CNN with predictions.

What we learned

We honestly learned a lot about CNN architecture and many different techniques we can use to optimize our model. Creating features using images, normalizing images, and doing image classification were amongst the many things we learned in relation to image classification in the data science world.

What's next for Marky Challenge

Honestly, the most important point to improve for this challenge was the CNN. We must attempt to figure out how to create the best architecture, which will allow us to actually create a better model.