Excerpt of code analysing data to develop automated classifier This code will not run as certain code has been removed A lot of the data loading / cleaning / preprocessing is removed

Background:

Had ~3 weeks to read material, data, then make models / code Whole 3 Month project (I am not joining) intends to create classifier, and explain it They had:

1. engineered features from company,
2. Also had a deep neural network for classification from previous project. I was interested in determining whether a deep neural network was needed, or if it simply inhibits the follow-up task of explaining the classifier. I did not pursue attempts to optimise a given model - this was more a scoping project

There were 2 datasets:

1. Dataset 1 had only their engineered features and classifications.
2. Dataset 2 was split across multiple excel / csv files and was partially labelled.

They had transformed their engineered features via scaling but not documented how too clearly, difficult to replicate. The results between Dataset 1 and Dataset 2 using same features were very different. This indicates need to follow up for further clarity on how their transformation was done.

Conclusion:

1. If they can replicate their scaling and get similar performance as they did for Dataset 1: They should simply keep their features, as they are meaningful and performant.
2. If they cannot be replicated, proposed model: mlp_I3B4Y1, is a recommended candidate: It uses their features (with my attempt at their scaling) combined with quantile-based data. The features are therefore interpretable. The model is a simple MLP that should be further optimised. It had an accuracy of ~93% versus the project's DNN's existing 85.8%.* *Care with these metrics: I created my own data pipeline, so tests are not like-for-like.
3. If manually selected features are not desired, proposed model: mlp_I2AY1, is good starting point It had an accuracy of ~92%. I think this can be easily increased with additional tuning. I also think errors in the autoencoder should be fed in as additional inputs, but out of time!