Dallas Mavericks SMM

Interface

Inspiration

This project was inspired by the Dallas Mavericks' very own social media accounts. Seeing their video tweets with highlights of the best plays in every Mavs game as they happen live is one of the best parts of being a fan, and we wanted to use data science and AI to help optimize the process of tweet creation, making it easier to make engaging posts at opportune moments.

What it does

Our project uses a machine learning model to tell users (Dallas Mavericks social media managers) the optimal time to send a highlight tweet about a live game. Given live play-by-play data from the game, the model will generate a "social metric score" which indicates how much engagement might be expected for a tweet sent at that time. Given this live play-by-play data our model will also automatically generate custom potential tweets for that specific highlight, complete with sponsor messages.

How we built it

Our machine learning model had several components. We used multi-factor analysis (probabilistic principal component analysis) and principal component analysis to reduce the predictor and target variables in the PBP and Twitter data set respectively. Further transformations, including inverse Gaussian ranks, were used on the reduced factors. We then used time series regression to derive factors over the time they gave us to make predictions on social engagement score (dependent upon play by play data). Since the target variable has been normalized scores are very interpretable, i.e if the model returned a score of 3 then the predicted social media engagement would be in the 99th percentile. Using Flask, we displayed the model’s output score on our dashboard in both a live ticket graph and a score-box format.

We then used Respell.ai to generate tweets using GPT-4, passing in play-by-play data and a specialized prompt via Flask, and then returning the output tweet to our dashboard.

Challenges we ran into

The Respell.ai API had some lag since we refresh all plays at the same time. The provided Twitter data also required a significant amount of cleaning and processing.

Accomplishments that we're proud of

This is the first hackathon for a couple of the team members, and we have managed to come up with a solid idea and work together effectively to implement it within a limited time constraint. We’re proud of the application we built and its practicality.

What we learned

We learned a lot about working as a team and creating a full stack project leveraging the power of machine learning. We learned about how to use commercial Twitter data in conjunction with live sports data for training a machine learning model. We also learned about Respell and how it can be used to easily create and streamline AI workflows.

What's next for SMM

Hook up to live game API data Multimedia (images, video clips from live broadcast) Dashboard performance improvement Twitter publishing and live chat integration Multiple social media platforms Bayesian Regressions with arbitrary distributions SMM risk scores via coefficient covariance