Hoax Detection

Code Implementation from Big Data Challenge on Satria Data 2020

Technologies

• Machine Learning with NLP
• Neural Network (NN)
• BERT

Dataset

Download dataset from This link

Results

Label Predictions from the experiment test data are on result directory. Validation accuracy are shown below:

Traditional Machine Learning with NLP Technique

NLP	KNN	Naive Bayes	SVM
CountVectorizer	0.77	0.73	0.83
TF-IDF	0.73	0.73	0.83

Neural Network (NN)

Word Embedding and LSTM use text dataset, whereas CNN use image dataset

Neural Network	Accuracy	Loss	Val Accuracy	Val Loss
Word Embedding	0.9501	0.4315	0.8571	0.4939
RNN-LSTM	0.8924	0.2735	0.8300	0.4585
CNN	0.8171	0.6627	0.8388	0.6606

Results from CNN still inaccurate due to imbalance data

BERT

Soon

Conclusion

Actually, validation accuracy is not enough to determine performance from model. We need actual label from this competition organizer to get model's performance in predicting document whether hoax or not. At least, we know the steps binary classification on detecting hoax. I'm waiting for your outstanding project.