Python Classifiers

Implementations of both N-Nearest Neighbour and Naive Bayes classifiers in Python, also includes some sample data. Some references to Weka are made - it can be downloaded here: http://www.cs.waikato.ac.nz/ml/index.html

Folders & Steps

• 1. Raw Data - Contains the raw pima indians diabetes .data file as well as the .name file as well as the assignment spec.

• 2. Processing - Contains everything needed to generate the processed .csv file:

  • Run the preprocessor.py file to add a header, change the class names and remove invalid results (see Assumptions and Invalid Data).
  • Open the resulting pima_processed.csv file in Weka and go Filter > Choose > Attribute > Normalise > Apply and then save the file as pima.csv to get a normalised CSV file.

• 3. Classifiers - contains python scripts to run the classifiers:

  • classifier.py
    • Run classifier.py -h for more information about argument usage.
    • A log file classifier.log will be created which logs information about the run such as number of correctly and incorrectly identified instances.
    • --folds, -f number of folds you want to split the data into (default 10)
    • --neighbours, -k, number of nearest neighbours (default 3)
    • --algorithm, -a, the algorithm to run (KNN/NB)

• 4. Feature Selection - contains a version of the data that has been run through Weka's CFS feature selection:

  • Open the pima.csv file generated in step 2 in Weka and go Select Attributes > Start > Right click the result > Save Reduced data...
  • Some header information that Weka generates in puts in the file had to be removed manually.

• 5. Results - contains a results spreadsheet and the final report.

  • results_matrix.xlsx - Compares our classifiers to Weka's for both the data with no feature selection and the data with feature selection (this is also included in report.pdf)
  • report.pdf - Contains findings.

Assumptions and Invalid Data

• There are a number of fields in the data where attributes are missing and have been coded as 0. We have decided to remove the rows containing a 0 value in the following fields:
  • Glucose Concentration
  • Blood Pressure
  • Tricep Skin
  • Body Mass Index (BMI)