Trjtdtk

I am trying to apply text classification with keras. Previously, I used Random Forest and lightgbm with accurancy score arround 65%. Then, I made some first attempts with neural networks, to improve my score, but the results are really bad (score less than 20%).

After searching, I have found approaches for multi-label classification when trying to predict multiple labeles for each input texte, such as the following

multi-label text classification

But this not at all my case!

My dataset is not extremely unbalanced but I have hundreds of classes to predict. I want to predict only one class per text.

After many attempts to find the right values to tune the neural network, I wonder if NN actually do not performe well in this case? Why the Random Forest out performs with the minimum hyper-parametre tuning? Maybe I should transform my problem into a multi-label binary classification?

This is a code example, taken from kaggle:

data = pd.read_csv('multiClassExample.csv', delimiter=';', encoding="latin-1")


enc=LabelEncoder()
enc.fit(data['typeChord'])
data['typeChord']=enc.transform(data['typeChord']) 
labels = to_categorical(data['typeChord'], num_classes=len(data.typeChord.unique()))

n_most_common_words = 8000
max_len = 130
tokenizer = Tokenizer(num_words=n_most_common_words, lower=True)
tokenizer.fit_on_texts(data['text'].values)
sequences = tokenizer.texts_to_sequences(data['text'].values)
word_index = tokenizer.word_index
print('Found %s unique tokens.' % len(word_index))

X = pad_sequences(sequences, maxlen=max_len) 
X_train, X_test, y_train, y_test = train_test_split(X , labels, test_size=0.25, random_state=42)

epochs = 20
emb_dim = 128
batch_size = 256
print(labels[:2]) 

print((X_train.shape, y_train.shape, X_test.shape, y_test.shape))

model = Sequential()
model.add(Embedding(n_most_common_words, emb_dim, input_length=X.shape[1]))
model.add(SpatialDropout1D(0.7))
model.add(LSTM(64, dropout=0.7, recurrent_dropout=0.7))
model.add(Dense(len(data.typeChord.unique()), activation='softmax'))
model.compile(optimizer='adam', loss='categorical_crossentropy', metrics=['acc'])
print(model.summary())
history = model.fit(X_train, y_train, epochs=epochs, batch_size=batch_size,validation_split=0.2,callbacks=[EarlyStopping(monitor='val_loss',patience=7, min_delta=0.0001)]) 

accr = model.evaluate(X_test,y_test)
print('Test setn Loss: :0.3fn Accuracy: :0.3f'.format(accr[0],accr[1]))