The goal of this project is to apply machine learning techniques on a set of data samples from CERN in order to be able to discernate between Higgs Boson particles and background particles.
In order to run our code, all you need to do is to type in the console at the location of run.py:
python run.py
The data csv files (test.csv and train.csv) must be located in the same folder as run.py. The output submission.csv file will be written at the same location as run.py.
The only external libraries that we use are numpy and csv. In order to be executed, the run.py file needs the following files in the same folder:
csv_helper.py
data_cleaning.py
enhanced_implementations.py
poly_expansion.py
All those files should be already included in the code.zip file at the right location.
Our program will work well with the parameters that we have set in it, but these can be modified to test our code. Some extreme choices for those parameters might make the program run into errors, or extreme running time, so make sure to keep normal parameters.
In order for our code to reproduce exactly our best kaggle submission, we set a shuffle parameter to False at some point in run.py. This makes our program totally deterministic (except maybe for some differences in floating point precision, depending on the computer it is running on). Please note that in this situation 'Stochastic gradient descents' are not stochastic unless you set the shuffle parameter to True.
The file implementations.py contains all 6 functions we were asked to write. Additionally, it also contains some helper functions at the end of the file, that help to build the 6 required functions. We suppose that the data that these functions take as parameters are already cleaned and standardized, so it should be tested with clean datasets. Also, for the logistic_regression reg_logistic_regression functions, we suppose that the provided y will only contain values that are either 0 or 1 (as opposed to the load_csv function that yields y with values -1 or 1). Using our function such y can result in errors. Some functions from implementations.py are needed for the final program (run.py) so they are exactly the same as in enhanced_implementations.py, but we decided to copy them in order to make run.py not dependent on implementations.py.
We used some provided functions for csv_helpers.py, and we used a slightly modified version of batch_iter from a lab's solutions. Some of the rest of our functions can be reusing some of the functions in the lab's solutions, but apart from those cited before, we always reworked them quite a bit to do it our way or to make them fit our situation and needs.