A Multi Classification Problem to Recognize Gestures from Accelerometer Time Series Data
By Amitav Chris Mostafa ,with care and support from good friends and family ,library textbooks and scripts at TU Darmstadt , and python library documentations :D
Celonis Code Challenge Task 1
Abstract___
Due to lack of skill and experience, I'll try to keep the complexity as low as possible and just to convey the working paradigms and decision designs behind the project, as would have been normally done in the interview. The challenge is meant to be kept to 3-4 hour so I hope the following is sufficient in its breadth.
For the project, I used the numpy, pyplot and os libraries only.
For better project management I used self made functions as modules
- prototype.ipynb is the Jupyter notebook I used to test and try the different approached to working with the regression. Since I've never implemented a regression from scratch before, I decided to get my hands dirty by trying out a linear regression. As my experience later would reinforce, this was a lot of learning and relearning as I realized the skills in pythong and maths needed is something I possessed to start with, but I was inexperienced in how to combine them. Much of that has been deleted, but I did use the scripts I used to access the txt files and make the X Matrix features and y target vector, all of which I later put into functions for ease of use. Later, I would go on to implement a logisitic regression model, a multi classification variant, through help from university scripts, friend and some online resources. I tried to follow the math and in some cases read the documentation in sklearn and numpy libraries to try to keep the task as grounded as possible. I try to hint to my inspirations as well as I can in the comments.
Since the task is due to be submitted today on friday within working hours, I'm avoiding burning more midnight oil by attempting the optional exercises, things which I am already familiar with working through online resources. Rather I focus on the implementation with numpy aspect, something which I haven't done before and also presents a novel challenge.
I tried to use as type hinting as much as possible and used scales and Black to keep my code clean and readable. I understand however that it might still be convovulated and quite far from best practices. This is an are I'm continually working on.
Although I tried to use type hinting as much as possible on advice from my fnf, this was not always possible with some functions as typing couldn't be imported in VS Code. I have to read more into declaring my own types as well. Additionally I tried to ensure as much docstring documentation as possible.
There were some solutions I considered such as loading the features in a different way (multi dimensional arrays), interpolating data instead of padding with 0, using other coordinate axes (polar coordinate) etc. But due to time crunch I decided to keep it as ideas for now and for later implementation for my own use, as they would raise complexity of concept.
For reasons unknown to me the normalization is raising errors. This isn't working with any of the fixes I know of. Maybe I'll leave it out of the final solution and just present it as an area I need further work on.
The One Hot Column in numpy was handled quite nicely. The Numpy documentation and sklearn documentation helped a lot for this. There was the idea of using the column numbers to implement unit testing with pytest on github. I don't think I'll come around to doing that due to time constraints. I wanted to add some Github Actions to automate testing, touching on Task 2 a bit practically. I also wanted to CLI but right now my aim is to send it via 1400 CEST. I'm currently 5 or 6 hours in already!
Also for the logistical regression, I chose to forego forward gradient ascent too much, instead relying on the sigmoid fucntion. As this was the approach taken by the Springer Book on Numerical Analysis (2012), and the mathematics for this topic was huge. It took a lot of cross referencing with sklearn documentation to constraint myself to the bare minimum.
I was made aware that my memory management with numpy could be much more efficient and smart. Indeed, out of jupyter, but even in it, the code runs notoriously slow. I tried to initiate zero arrays and flow values into it iteratively or else as much as possible. But sometimes I decided the convenience was worth it as a first proof of concept.
As I intended to work with low epochs to begin with i did not necessarily choose hyper optimized methods, but also I was limited by my capacity in knowledge and experience . As such, the speeds are quite slow when compared to its load.
I noticed more efficiency in my code, such as not calling functions in other functions for efficiency. That is something that is needful to be addressed as well
I wanted to put in a accuracy plot over epoch but the way i set up my code made it hard to incorporate. Note to self, plan out your important ideas in the future.
Endtime: 14:06 CEST (took about 6-7h excluding reading literature)
I hope you are happy with the solution! Couldn't make it as elegant as I'd hoped.
I think the plots illustrate well the high enough accuracies with this minimal solution. The math works! There are some false positives on the confusion matrix but the density along the diagonal looks great. Loss is also that logarithmly deacreased quickly over time. Ideally I'd make a list of hyperparameters and compare them to find the best find. For now, I'm sticking to test values.
Thank you for your considerations! I think I improved myself a little too with the challenge :D