Reddit is an online forum where uses can create groups known as 'subreddits' commonly based on a given topic, however, that is not strictly necessary. As of October 27, 2020, there are over 1.5 million unique subreddits. When users post to Reddit, they must post to a specific subreddit. They are allowed to share posts from one subreddit to another, a practice known as 'cross-posting'. A useful feature might be suggesting relevant subreddits a user might want to share a specific post with.

As with most machine learning tasks, NLP requires quality training and evaluation data. There exist several pre-gathered datasets consisting of posts from various subreddits. However, these datasets can be relatively old, and I would like to use more recent data. Reddit also frequently removes subreddits, and the "tone" of a given subreddit can change as well. As such, I would like to implement a way to easily scrape Reddit and retrain my models at a given time interval.

A model saved on my hard drive doesn't offer much value to many people. To allow people to use our model, we must deploy it and expose some sort of API route. This is often difficult given the requirements of large, modern models. In this particular project, there will not be many, predictable, requests to the model, so having a constant deployment server would be quite ineffective. As such, I will attempt to use Azure Functions, a serverless approach, meaning I will only pay per request. Because of my projected number of users and my financial constraints, I believe the tradeoffs that come from using such an approach will be acceptable.

Once we have data, the next step is trying several different approaches and deciding which one to use for the final deployment. There are so many different models, and parameters, and things to consider when testing out machine learning models. We must also choose a metric with which to judge the models. Once we've made our final decision, we will do one final training with a larger subset of the data and evaluate using a validation subset of the data.

As mentioned earlier, the data on Reddit is rapidly changing. As such, any good model will need to be trained using the most recent data possible. To make the process of gathering data, retraining the model, and updating the deployment easier, I would like to automate the process as much as possible. I will again be using Azure Functions to periodically perform this task.

As with most projects, many lessons were learned. This is a great project to hone your NLP and deployment skills. Moving forward, I may choose to change some aspects of this project, such as migrating to the more popular AWS, or GCP. I will try to keep the model architecture up to date with the current state of the art model types.