I have been trying to write a naive bayes classifier from scratch that is supposed to predict the class label of the nominal car.arff dataset. However the classifier always predicts the most common one. I have tried log probabilities and laplace correction, both to no avail. Also I have noticed that the conditional probabilities for any attribute is always the greatest for the most common label. Is this because of my dataset? What can be done about it?

Here is my code:

import numpy as np
import pandas as pd
from scipy.io import arff

def parser(path):
    function which parses the data from an arff file
    @param path: string containig the path to file
    @return array containing the data
    @raise FileNotFoundError exception in case if the path does not point to a valid file

    start = 0  # check if data really occured
    # Declaratives as constant to avoid misspelling in code
    RELATION = 'relation'
    ATTRIBUTE = 'attribute'
    DATA = 'data'

    # Create dictionary holding the arff information
    data = {RELATION: [],
            ATTRIBUTE: [],
            DATA: []}

    # Read the file and analyse the data
    with open(path) as file:
        for line in file.readlines():
            # Check if line is empty
            if line.strip() == '':

            # Check if line contains the relation
            elif '@' + RELATION in line:
                data[RELATION].append(line.replace('@' + RELATION, '').strip())

            # Check if line contains an attribute
            elif line.startswith('@attribute'):
                tmp = line.replace("{", "").replace("}", "").replace("\n", "").replace("'", "")
                # checks if whitespaces between commas in attributes occur
                if (len(tmp.split(" ")) > 3):
                    values = tmp.replace(",", "").split(" ")[2:]
                    values = tmp.split(" ")[2].split(",")

                data[ATTRIBUTE].append({'name': tmp.split(" ")[1], 'values': values})

            # check if @data exists
            elif '@' + DATA in line:
                start = 1

            # If the line is not one of the others, it has to be data
            elif '@' + DATA not in line and start:
                line = line.split(',')
                # strip each element of the line
                for i in range(len(line)):
                    line[i] = line[i].strip()
                # Add data to dictionary

    attributes = np.array(data['attribute'])
    out = []
    for i in range(len(data['data'])):
        data_dict = {}
        for j in range(len(attributes)):
            data_dict.update({attributes[j]['name']: data['data'][i][j]})
    out = np.array(out)
    return out, data[ATTRIBUTE]

class NaiveBayes():

    def __init__(self, data, atts, class_label):
        self.data = data
        self.atts = atts
        self.class_label = class_label

    def prior(self):

        prior_probabilities = [0,0,0,0]
        for i in range(len(self.data)):
            if self.data[i]['class'] == 'unacc': prior_probabilities[0] += 1
            if self.data[i]['class'] == 'acc': prior_probabilities[1] += 1
            if self.data[i]['class'] == 'good': prior_probabilities[2] += 1
            if self.data[i]['class'] == 'vgood': prior_probabilities[3] += 1
        prior_probabilities = [x/len(self.data) for x in prior_probabilities]

        return prior_probabilities

    def conditionalProbability(self,key,value,length):
        #returns (in our case) 4 vector for one attribute with probabilities for each outcome
        conditional_probabilities = [0]*length
        #definetly not the most efficient way
        for i in range(len(self.data)):
            if self.data[i][key] == value:
                if self.data[i]['class'] == 'unacc': conditional_probabilities[0] += 1
                if self.data[i]['class'] == 'acc': conditional_probabilities[1] += 1
                if self.data[i]['class'] == 'good': conditional_probabilities[2] += 1
                if self.data[i]['class'] == 'vgood': conditional_probabilities[3] += 1

        s = np.sum(conditional_probabilities)
        conditional_probabilities = [x/s for x in conditional_probabilities]

        return conditional_probabilities

    def classification(self, instance):

        cprobs = []
        probs = self.prior()
        for key in instance.keys():
        #get probabilities
        predicted_class = "unacc"

        for i in range(len(cprobs)-1):
            for j in range(4):


        return probs.index(max(probs))

raw,atts = parser('car.arff')
class_attribute = 'class'

classifier = NaiveBayes(raw,atts,class_attribute)

results = [0,0,0,0]
for i in range(len(classifier.data)):

This is the class distribution and some more information:

% 5. Number of Instances: 1728
%    (instances completely cover the attribute space)
% 6. Number of Attributes: 6
% 7. Attribute Values:
%    buying       v-high, high, med, low
%    maint        v-high, high, med, low
%    doors        2, 3, 4, 5-more
%    persons      2, 4, more
%    lug_boot     small, med, big
%    safety       low, med, high
% 8. Missing Attribute Values: none
% 9. Class Distribution (number of instances per class)
%    class      N          N[%]
%    -----------------------------
%    unacc     1210     (70.023 %) 
%    acc        384     (22.222 %) 
%    good        69     ( 3.993 %) 
%    v-good      65     ( 3.762 %) 

and here is some sample data:


The complete dataset can be found here


Looking at your distribution over classes, it is heavily unbalanced and this can skew the model to predicting the majority class, which in this case is 'unacc'. So, one recommendation would be to balance out the classes, typically by adding more instances of the minority classes to be equal to the majority class.

Also, looking at your sample data, there seems to be little, if not no variation between the buying, maint, doors and persons and here it looks like these features would not impact the classification decision.

In this case, I would go back to exploring the data and seeing which features could affect the classification decision. This can be done with bar plots and histograms. When doing this divide the data into the classes and plot the distribution of the features by class so you can see if there is any noticeable variation in distribution of these features by their classes.

| improve this answer | |

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.