Skip to content

Latest commit

 

History

History
45 lines (35 loc) · 2.38 KB

README.md

File metadata and controls

45 lines (35 loc) · 2.38 KB
Raw

ITU-ML5G-PS-008-Network-failure-classification-model-using-NDT--MLAB

This code is used for MLP-base feature selection in KDDI challenge. We mainly operate the code in Google colabortory. Therefore, the dataset and model import are all related to the google drive.

Given the randomness of neural network, the code might not produce the exact same result as that in the slides

Also, the code still need further refinement.

Repository structure

data: folder containing the datasets for domain A and domain C

MLP_based_feature_selection: The entire python code to address the NDT problem

Proposed model

We used multi-layer perceptron to train the model on domain A. After that, we applied transfer learning by freezing all the layer but the last layer. We fine-tuned the last layer with training data from domain C

Feature selection

The originality of the research mainly lies on feature selection process. We calculate the mutual information betwwen the label column and feature columns and select different numbers of top features. By changeing the value K, we can evaluate the performance when we use different number of selected features.

t0 = time.time()
k_best_A = SelectKBest(score_func=mutual_info_classif, k=150)
X_new_A = k_best_A.fit_transform(data_input_features_A, encoded_label_A)

Advanced task

We further expand the advanced task by changing the number of train_data_C we use from 10% to 100%. By changing the parameter "frac", we can evaluate the performance of our proposed model when we use different amount of train_data_C

# Split the dataset into normal and abnormal data
normal_data = data_raw_tab_C_train[data_raw_tab_C_train['y_true(fc)'] == "normal"]
abnormal_data = data_raw_tab_C_train[data_raw_tab_C_train['y_true(fc)'] != "normal"]

print("Testing the data separation from noraml and abnormal")
print(normal_data)
print("\n")
print(abnormal_data)

# Change the data from data_train_C we use to train by adjusting the value: frac. When frac is 1.0, we use 100% of train_data_C
sampled_data = normal_data.sample(frac=1.0, random_state=42)
for label in abnormal_data['y_true(fc)'].unique():
    label_data = abnormal_data[abnormal_data['y_true(fc)'] == label]
    sampled_data = pd.concat([sampled_data, label_data.sample(frac=1.0, random_state=42)])

# Reset the index
sampled_data.reset_index(drop=True, inplace=True)
data_raw_tab_C_train = sampled_data.sample(frac=1, random_state=42)