aristocatsrepository

JDA Challenge The Genie of Cervecería Teotihuacan

We began the Hackaton with a small project in mind. The project was simple: make a walker with computer vision that is able to move by itself, to its owner in order to simplify its use.

Arriving at the HM, we decided to accept JDA's challenge. The objective is simple: to build a predictive model. The catch was the format of the dataset and the algorithm that best suited the project.

At first, we agreed to develop a neural network capable of doing time series predictions. However we encountered a more primitive problem: The dataset was in a weird format.

Taking into account none of the team is a Data Scientist, it was difficult at first to edit the dataset in order to make it fit to a certain model. However we found that sorting the data by day was a good approach. We then re-structured the data into 3D tensors with a shape of 50*369*10. However we found that 3 variables were useless for a predictive model (we didn't have the time to make a PCA, but it was a little intuitive), so we changed the data format into 50*369*7.

After that task, we proceeded to develop an LSTM. However we had the doubt if this model would work taking into account it didn't use the 2nd CSV provided by JDA due to the nature of an LSTM. So we also made a Random Forest Tree, in order to compare our results.

We found out that our first algorithm (LSTM) had a RMSE of around 700 (comparing all existing features, which was in a 50*369*7 space) and for the sales quantity, the error measured was of around 12, which is a good result.

In the side of the web development we started first in designing the general outline of the Front end web page as our teammates started creating and defining the Neural Network. After the first results started to pop up, we started designing how we would present the data generated by the Neural Network. We started creating some first base designs over these results, but something happened with the data the Neural Network. The would change because of the weird format of the input. We asked ourselves how to interpret the data that the neural network was producing since the format that it received was weird and we were having a hard time deciphering how to best retrieve and display this data. With the hard work of our teammates we managed to create a good output format that we could use to better display the information. Then as we were working with the graphs to interpret the data we came to the conclusion that we might need a database to improve the performance of the project. After working to connect to the database we were unable to make an good connection with it due to technical limitations. We had to find a workaround to make the project work, which leads us to the current version of the project.

Built With

• javascript
• jupyter-notebook
• quasar