A complete workflow through the Cleveland heart disease dataset. From data analysis, through model creation to experimenting on improving the model.
Historically - this was my first actual notebook that I created and wanted to share. Since it's inception I worked on some more projects, but after a nearly 5 months long hiatus I returned back to this repository. I wanted to use a simple, clean dataset to experiment with building my very own neuraln networks.
In this notebook we will take a look at the slightly altered Cleveland UCI heart disease dataset, available on Kaggle under this link. In here I am going through an entire workflow as depicted below (author is Daniel Bourke, you can read about these steps right here on his website)
The notebook in its current state should be available on Kaggle, and you can see it right here.
Please, as I stated in the notebook - leave your comments and opinions either here or on Kaggle since I want to improve my skills and knowledge in the Machine Learning topic.
Thanks to I.J. for helping with the initial notebook!
-- 16.09.2023 --
Well that took some time, but August was really busy for me. This is the last update for another while, since I feel I am done with this notebook and the dataset. I completed a run of extensive (in my opinion) experimentation and exported the best network I had. I went through multiple iterations of different structures, tweaking the units and dropout rates as I went along. As I say at the bottom of the new notebook - if you've got anything you'd like to say or teach me, please don't hesitate to contact me via e-mail or the contact form on my website.
-- 25.07.2023 --
After a long break from this dataset I wanted to try setting up my very own binary classification neural network. So far I put up a simple one, with an input layer, one hidden Dense layer, one hidden Dropout layer and then the output layer. This simple network managed to score 83.333% which sounds satisfactory, but I'm willing to experiment with the structure of the network. That's gonna happen in the next update though, for now please check out the headrt_disease_project_2.ipynb.
-- 05.03.2023 --
Another week another update. I believe it's the last one for some time now, since I feel I'm done with this notebook. This time I managed to finish the third round - I created and evaluated the default, RandomizedSearchCV() and GridSearchCV() versions of the RandomForestClassifier(). I think I'm finally understanding the meaning behind cross-validated metrics and what they mean in terms of overfitting the model. I did not reach the goal I set out to achieve but it doesn't really matter. What matters is I went through the whole workflow three times, I started experimenting with the datasets and what can be removed/improved in them and in general I had good time, improved my skills probably a bit. So that's a wrap for now, see you round!
-- 26.02.2023 --
Tough week so not much of an update - I reworked the second round so that it functions as it should, and we actually are quite a bit happier with the exported model. I checked to see if the features in Round 2 were used to their fullest potential and I initiated the third round by creating a new dataframe. Oh and I also Kaggle-proofed the code so that it won't break each time I upload it there and forget to add an url to the dataset. See you next time!
-- 20.02.2023 --
Another week another update. Round 2 is finished (for now at least) - I went over fine tuning the LogisticRegression() model with both RandomizedSearchCV() and GridSearchCV(). Then I recreated the best models that these functions found and went through all the metrics as in Round 1. Finally, I attempted to manually tune the parameters that were different between two best models to see, if changing them in some way makes any difference in performance. I also decided to dump two models, even though they both underperform.
-- 13.02.2023 --
Some more changes added to almost all files. I changed the environment name in the enivronment.yml from a generic and bland "env" to "ML-env". The notebook has been updated with inspection of feature importances for both LogisticRegression() and RandomForestClassifier(). Round 1 of the modelling and tuning is done, now we're off to Round 2 and Round 3. Each of them will be dedicated to a single model that I'll attempt to tune to get as close to the accuracy goal as possible. I am also adding the notebook in the current state to Kaggle - you can visit it right here.
-- 07.02.2023 --
I deleted the .gitignore file. We can almost say that the notebook is "done", although it still needs some changes, and more thorough tuning. I tuned the models with both RandomizedSearchCV and GridSearchCV, then I evaluated the models with both the classification report and cross-validated metrics. It's almost completed, after I'll finish the whole workflow I'll go back to the start and go into more small details.
-- 05.02.2023 --
The README.md was rewritten, a correct environment.yml was pushed. The current steps that I (more or less) cover in this notebook are:
- Problem definition
- Data analysis
- Evaluation
- Features (kind of but not really)
Currently working on step 5. Modelling. So far I have created three models, LogisticRegression(), KNearestNeighbors() and RandomForestClassifier(). After experimenting with different values for n_neigbors I decided to abandon KNN as it seems to not be good enough of a model for this particular dataset.
-- 04.02.2023 --
The repository was created, unedited files were pushed.