We were inspired to pursue this challenge because it tasked us with the impossible; making sense out of randomness.
What it does
In our project, we developed a deep learning model that can predict when a loot box will provide a rare or legendary item. We then performed an interpretability analysis of this model to shed light on its blackbox and determines what factors it considered and the patterns it learned.
How we built it
We developed this neural network using Python Tensorflow, along with several other processing libraries such as NumPy, Sklearn, Pandas, and SciPy. Plots were visualized using Matplotlib pyplot. Our team collaborated across a number of platforms, including Jupyter Notebook, Google Colab, and VS Code.
Challenges we ran into
The main challenge we initially faced was inaccuracies in the dataset. Many of our analyses were predicated on the raw dataset provided, but we were later told to splice the majority of the entries as they did not match the attached description. Thus, we were forced to restart from square one about half way through the hackathon. Nevertheless, we bounced back and found novel insights to explore.
Accomplishments that we're proud of
We are proud that we were able to train a deep learning model on such limited data (only 5 inputs) in a constrained time frame.
What we learned
We learned to collaborate more efficiently, maximizing the speed at which we progress in our analyses. Moreover, we found alternative solutions to computationally expensive algorithms, such as replacing saliency analysis with a leave-one-out proxy.
What's next for Deep Learning Models Stochastic Processes
The next step for this project is to explore additional architectures, such as Attention based mechanisms in transformers to further model the data. Moreover, a more comprehensive dataset that includes additional information about the users can reveal deeper insights into the loot algorithm.
Log in or sign up for Devpost to join the conversation.