This project aims to design a deep learning model with certain schemes to analyze the sentiment of each user review which is then evaluated with predefined objectives.
- Training and test accuracy of at least 92%
- The model was able to classify 3 different classes: negative, neutral, and positive
To support the defined objectives, the dataset was obtained from independent scrapping with a total of 139500 reviews in Indonesian.
| Scheme | Model | Feature Extraction | Data Splitting |
|---|---|---|---|
| 1 | Sequential | TF-IDF | 80/20 |
| 2 | CNN | TF-IDF | 70/30 |
| 3 | CNN | TF-IDF | 80/20 |
| Scheme | Train Accuracy | Test Accuracy |
|---|---|---|
| 1 | 97 | 92 |
| 2 | 97 | 94 |
| 3 | 97 | 91 |
Seen from the summary and prediction results of the model that has fulfilled all the points that are the objective of this project.
During the model development process, there were several challenges encountered. Below are the main challenges and the solutions implemented to overcome them:
Challenge: The dataset had an imbalance in the number of reviews for each sentiment class.
Solution: Applied oversampling techniques such as SMOTE (Synthetic Minority Over-sampling Technique) to balance the dataset. This helped in improving the model's ability to accurately classify the minority class.
Challenge: Achieving the desired accuracy on the test set while avoiding overfitting.
Solution: Implemented various model architectures and regularization techniques. For instance, added Dropout and Batch Normalization layers in the Convolutional Neural Network (CNN) model to prevent overfitting and improve generalization.
Challenge: Training deep learning models with a large dataset required significant computational power and time.
Solution: Utilized cloud-based GPU instances to accelerate the training process and ensure timely completion of the experiments.
Challenge: Finding the optimal hyperparameters for the model to achieve the best performance.
Solution: I experimenting with different values for learning rate, batch size, and the number of layers/neurons.