Fake News Classification using DL

PROJECT TITLE

Fake News Detection using Deep Learning

GOAL

To identify whether the given news is fake or not.

DATASET

The link for the dataset used in this project: https://www.kaggle.com/competitions/fake-news/data?select=train.csv

DESCRIPTION

This project aims to identify whether the given news is fake or not by extracting meaning and semantics of the given news.

WHAT I HAD DONE

1. Data collection: From the link of the dataset given above.
2. Data preprocessing: Preprocessed the news by combining title and text to create a new feature and did some augementation like tokeinizing and vectorising before passing them to model training
3. Model selection: Self Designed model having a Embedding Layer followed by Global Pooling Layer and then 2 Dense layers and then output layer.Second model had a Embedding layer followed by a RNN layer and a Dense output layer.
4. Comparative analysis: Compared the accuracy score of all the models.

MODELS SUMMARY

Model-1: "sequential"

---

Layer (type) Output Shape Param #

embedding (Embedding) (None, 12140, 182) 30222010

global_average_pooling1d ( (None, 182) 0
GlobalAveragePooling1D)

dense (Dense) (None, 96) 17568

dense_1 (Dense) (None, 24) 2328

dense_2 (Dense) (None, 1) 25

================================================================= Total params: 30241931 (115.36 MB) Trainable params: 30241931 (115.36 MB) Non-trainable params: 0 (0.00 Byte)

Model-2: "sequential_3"

---

Layer (type) Output Shape Param #

embedding_3 (Embedding) (None, 12140, 100) 16605500

simple_rnn (SimpleRNN) (None, 10) 1110

dense_5 (Dense) (None, 1) 11

================================================================= Total params: 16606621 (63.35 MB) Trainable params: 16606621 (63.35 MB) Non-trainable params: 0 (0.00 Byte)

LIBRARIES NEEDED

The following libraries are required to run this project:

• nltk
• pandas
• matplotlib
• tensorflow
• keras
• sklearn

EVALUATION METRICS

The evaluation metrics I used to assess the models:

• Accuracy
• Loss

It is shown using Confusion Matrix in the Images folder

RESULTS

Results on Val dataset: For Model-1: Accuracy:96.11% loss: 0.1350

For Model-2: Accuracy:85.03% loss: 0.1439

CONCLUSION

Based on results we can draw following conclusions:

1.The model-1 showed high validation accuracy of 96.11% and loss of 0.1350.Thus the model-1 worked fairly well identifying 2874 fake articles from a total of 3044.The first model performed better.The second model had good training accuracy but less test accuracy hinting towards overfitting.Maybe the key reason being in fake news it is important to capture overall sentiment better than individual word sentiment.