lxd789 / rumordetectionrnn

RumorDetectionRNN

Libraries used

1. gensim
2. keras
3. pickle

Algorithm

Dataset pre-process (preprocessData.py)

1. for each train file f in twitter_json,
   1. put value of 'text' key in a list X_train, do this for all lines
   2. Tokenize X_train
   3. Convert each text in X_train to sequences
   4. pad X_train
2. dump X_train and Y_train using pickle
3. do step 1 and 2 for test every test file

Training neural network

1. load trainX, trainY, testX, testY using loadTensorInput() Each item is now a list of list
2. categorize trainY and testY to two classes
3. build the neural net model using tflearn (LSTM RNN)
   1. activation='softmax'
   2. optimizer='adam'
   3. learning_rate=0.001
   4. loss='categorical_crossentropy'
4. fit the model
   1. n_epoch=20
5. save the model

before running any file:

1. make sure that the dataset folder is inside the project and has name 'rumor'

To create folder resources:

1. run preprocessData.py

To train neural network:

1. run RumorRNN.py

Result

Train on 875 samples, validate on 118 samples

20 epochs

875/875 [==============================] - 15s - loss: 0.5089 - acc: 0.7646 - val_loss: 0.4852 - val_acc: 0.7542 Epoch 13/20 875/875 [==============================] - 15s - loss: 0.3863 - acc: 0.8274 - val_loss: 0.7699 - val_acc: 0.7203 Epoch 14/20 875/875 [==============================] - 15s - loss: 0.2909 - acc: 0.8720 - val_loss: 0.8753 - val_acc: 0.7373 Epoch 15/20 875/875 [==============================] - 15s - loss: 0.1825 - acc: 0.9314 - val_loss: 1.3211 - val_acc: 0.7119 Epoch 16/20 875/875 [==============================] - 15s - loss: 0.1228 - acc: 0.9543 - val_loss: 1.5710 - val_acc: 0.6695 Epoch 17/20 875/875 [==============================] - 15s - loss: 0.0728 - acc: 0.9794 - val_loss: 2.1107 - val_acc: 0.6525 Epoch 18/20 875/875 [==============================] - 15s - loss: 0.0792 - acc: 0.9749 - val_loss: 2.3427 - val_acc: 0.6695 Epoch 19/20 875/875 [==============================] - 15s - loss: 0.0710 - acc: 0.9783 - val_loss: 2.8942 - val_acc: 0.6356 Epoch 20/20 875/875 [==============================] - 15s - loss: 0.0675 - acc: 0.9794 - val_loss: 1.9913 - val_acc: 0.6695 118/118 [==============================] - 0s Accuracy: 66.95% Process finished with exit code 0