1
$\begingroup$

I am trying to perform binary classification of search results based on the relevance to the query. I followed this tutorial on how to make an SVM, and I got it to work with a small iris dataset. Now, I am attempting to use the LETOR 4.0 MQ2007 dataset by Microsoft to classify. The dataset has 21 input vectors as well as a score from 0 to 2. I classified 0 as -1 and 1, 2 as 1. My algorithm reaches 57.4% accuracy after 1000 epochs with 500 samples of each classification. My learning rate is 0.0001. here is my code.

from tqdm import tqdm
import numpy as np
from sklearn.metrics import accuracy_score


print("-------------------------------------")
choice = input("Train or Test: ")
print("-------------------------------------")

# HYPERPARAMETERS
feature_num = 21
epochs = 1000
sample_size = 500
learning_rate = 0.0001

if choice == "Train":

    out_file = open('weights.txt', 'w')
    out_file.close()

    print("Serializing Train Data...")

    # SERIALIZE DATA
    file = open('train.txt')
    train_set = file.read().splitlines()
    positive = []
    negative = []

    # GRAB TRAINING SAMPLES
    for i in train_set:
        if (i[0] == '1' or i[0] == '2') and len(positive) < sample_size:
            positive.append(i)
        if (i[0] == '0') and len(negative) < sample_size:
            negative.append(i)

    train_set = positive+negative
    file.close()

    features = []
    query = []

    # CREATE TRAINING VECTORS
    alpha = np.full(feature_num, learning_rate)
    weights = np.zeros((len(train_set), feature_num))
    output = np.zeros((len(train_set), feature_num))
    score = np.zeros((len(train_set), feature_num))

    for i in tqdm(range(len(train_set))):
        elements = train_set[i].split(' ')
        if int(elements[0]) == 0:
            score[i] = [-1] * feature_num
        else:
            score[i] = [1] * feature_num

        query.append(int(elements[1].split(':')[1]))
        tmp = []
        for feature in elements[2:2+feature_num]:
            if feature.split(':')[1] == 'NULL':
                tmp.append(0.0)
            else:
                tmp.append(float(feature.split(':')[1]))
        features.append(tmp)

    features = np.asarray(features)

    print("-------------------------------------")
    print("Training Initialized...")

    # TRAIN MODEL
    for i in tqdm(range(epochs)):

        # FORWARD y = sum(wx)
        for sample in range(len(train_set)):
            output[sample] = weights[sample]*features[sample]
            output[sample] = np.full((feature_num), np.sum(output[sample]))

        # NORMALIZE NEGATIVE SIGNS
        output = output*score
        # UPDATE WEIGHTS
        count = 0
        for val in output:
            if(val[0] >= 1):
                cost = 0
                weights = weights - alpha * (2 * 1/epochs * weights)
            else:
                cost = 1 - val[0]
                # WEIGHTS = WEIGHTS + LEARNING RATE * [X] * [Y]
                weights = weights + alpha * (features[count] * score[count] - 2 * 1/epochs * weights)

            count += 1

    # EXPORT WEIGHTS
    out_file = open('weights.txt', 'a+')
    for i in weights[0]:
        out_file.write(str(i)+'\n')
    out_file.close()

elif choice == "Test":

    print("Serializing Test Data...")

    # SERIALIZE DATA
    file = open('train.txt')
    train_set = file.read().splitlines()
    positive = []
    negative = []
    for i in train_set:
        if (i[0] == '1' or i[0] == '2') and len(positive) < sample_size:
            positive.append(i)
        if (i[0] == '0') and len(negative) < sample_size:
            negative.append(i)

    test_set = positive+negative

    file = open('weights.txt', 'r').read().splitlines()
    weights = np.zeros((len(test_set), feature_num))

    # CREATE TEST SET
    for i in range(len(weights)):
        weights[i] = file
    features = []
    query = []
    output = np.zeros((len(test_set), feature_num))
    score = np.zeros((len(test_set)))

    for i in tqdm(range(len(test_set))):
        elements = test_set[i].split(' ')
        if int(elements[0]) == 0:
            score[i] = -1
        else:
            score[i] = 1

        query.append(int(elements[1].split(':')[1]))
        tmp = []
        for feature in elements[2:2+feature_num]:
            if feature.split(':')[1] == 'NULL':
                tmp.append(0.0)
            else:
                tmp.append(float(feature.split(':')[1]))
        features.append(tmp)

    features = np.asarray(features)

    for sample in range(len(test_set)):
        output[sample] = weights[sample]*features[sample]
        output[sample] = np.full((feature_num), np.sum(output[sample]))

    predictions = []
    for val in output:
        if(val[0] > 1):
            predictions.append(1)
        else:
            predictions.append(-1)

    print("-------------------------------------")
    print("Predicting...")
    print("-------------------------------------")
    print("Prediction finished with "+str(accuracy_score(score, predictions)*100)+"% accuracy.")

My training algorithm

if(val[0] >= 1):
    cost = 0
    weights = weights - alpha * (2 * 1/epochs * weights)
else:
    cost = 1 - val[0]
    # WEIGHTS = WEIGHTS + LEARNING RATE * [X] * [Y]
    weights = weights + alpha * (features[count] * score[count] - 2 * 1/epochs * weights)

What could I do to help the model train? Am I not giving it enough time? Is the algorithm wrong? Are the hyperparameters ok? Thanks for all your help.

$\endgroup$
  • $\begingroup$ Can you clarify this part "I classified 0 as -1 and 1, 2 as 1"? Does it mean that the original datasets has 3 classes and you're joining two classes so that you can perform binary classification? Also, you have only a total of 1k training examples? That's a relatively small dataset. $\endgroup$ – nbro Apr 7 at 1:22
  • $\begingroup$ Yes. The samples are scored 0 - 2, so I classified 0 as -1, and 1 and 2 as 1. Is this not a good way to do it? What would be an adequate sample size? Is my training algorithm ok? $\endgroup$ – iamPres Apr 7 at 3:21

Your Answer

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

Browse other questions tagged or ask your own question.