🎬 IMDB Movie Review Sentiment Analysis using Deep Learning (Optimized ANN)

---

📘 Project Overview

This project applies Deep Learning techniques to perform Sentiment Analysis on the IMDB Movie Review Dataset using an Artificial Neural Network (ANN).
The model determines whether a movie review expresses a positive or negative sentiment.

It uses TF-IDF for text feature extraction and a carefully optimized neural network with:

• Batch Normalization
• Leaky ReLU Activation
• Dropout Regularization
• L2 Weight Regularization
• Early Stopping

The architecture is fine-tuned to achieve robust accuracy (~87%) with controlled overfitting.

---

🎯 Objective

To develop a robust text classification model that can accurately predict the sentiment polarity of IMDB movie reviews using optimized neural network architecture and advanced NLP preprocessing.

---

📂 Dataset Information

Dataset Name: IMDB Movie Review Dataset
Size: 50,000 reviews (Balanced)

• 25,000 for training
• 25,000 for testing

Label Distribution: | Sentiment | Label | Count | |————|——–|——–| | Positive | 1 | 25,000 | | Negative | 0 | 25,000 |

Each review is a textual paragraph expressing a user’s opinion about a movie.

---

⚙️ Project Workflow

1️⃣ Data Preprocessing

• Convert all text to lowercase
• Remove HTML tags and punctuation
• Keep only alphabetic characters
• Encode sentiment labels (positive → 1, negative → 0)

2️⃣ Data Splitting

• 80% Training Data
• 20% Testing Data
• Random seed fixed at 42 for reproducibility

3️⃣ Feature Extraction

• TF-IDF Vectorization
  • max_features = 15000
  • ngram_range = (1, 2)
  • stop_words = 'english'

This converts textual reviews into a 15,000-dimensional numeric vector, representing word importance.

4️⃣ Model Architecture

Layer Type	Units	Activation	Regularization	Dropout	Notes
Dense	1024	LeakyReLU(0.1)	L2(0.001)	0.5	Input Layer
Dense	512	LeakyReLU(0.1)	L2(0.001)	0.4	Hidden Layer
Dense	256	LeakyReLU(0.1)	L2(0.001)	0.3	Hidden Layer
Dense	128	LeakyReLU(0.1)	L2(0.001)	0.3	Hidden Layer
Dense	64	LeakyReLU(0.1)	L2(0.001)	0.2	Hidden Layer
Dense	1	Sigmoid	-	-	Output Layer

Additional Enhancements:

• BatchNormalization after every layer for faster convergence
• LeakyReLU prevents neuron death
• Dropout ensures better generalization
• L2 Regularization reduces weight explosion

---

🧠 Model Compilation and Training

Parameter	Value
Optimizer	Adam
Learning Rate	0.0005
Loss Function	Binary Crossentropy
Metric	Accuracy
Epochs	15 (Early Stopping)
Batch Size	64
Validation Split	20%
Early Stopping	Patience = 2

Early Stopping: Stops training when val_loss no longer improves, ensuring the best weights are restored.

---

📊 Results and Analysis

✅ Final Evaluation Metrics

Metric	Score
Accuracy	0.8712
Precision (Positive)	0.86
Recall (Positive)	0.88
F1-Score (Positive)	0.87
Precision (Negative)	0.88
Recall (Negative)	0.86
F1-Score (Negative)	0.87

---

🔍 Classification Report

       precision    recall  f1-score   support

   0       0.88      0.86      0.87      4961
   1       0.86      0.88      0.87      5039

accuracy 0.87 10000

🧩 Confusion Matrix

	Predicted Negative	Predicted Positive
Actual Negative (0)	4255	706
Actual Positive (1)	582	4457

Interpretation:

• 4255 reviews were correctly identified as negative
• 4457 reviews were correctly identified as positive
• The model misclassified only around 12.9% of the total reviews

---

📈 Training Behavior

• Training Accuracy: ↑ 93% → 94%
• Validation Accuracy: Stabilized around 87%
• Loss Curve: Validation loss converges early due to L2 regularization and early stopping
• No overfitting observed — training and validation performance remain closely aligned

---

🧰 Technologies Used

🔹 Programming Language

• Python 3.8+

🔹 Data Handling & Preprocessing

• pandas → Dataset manipulation
• numpy → Array operations
• re (Regex) → Text cleaning and pattern matching

🔹 Natural Language Processing

• scikit-learn
  • TfidfVectorizer → Text vectorization
  • train_test_split → Data partitioning
  • accuracy_score, classification_report, confusion_matrix → Performance metrics

🔹 Deep Learning Framework

• TensorFlow / Keras
  • Sequential, Dense, Dropout, BatchNormalization, LeakyReLU → Neural network architecture
  • Adam Optimizer → Adaptive gradient optimization
  • EarlyStopping → Regularization and convergence control
  • l2 → Weight regularization

🔹 Optional Visualization (Recommended)

• Matplotlib → For plotting accuracy/loss curves
• Seaborn → For visualizing confusion matrix

---

🚀 Future Improvements

Enhancement	Description
🔤 Word Embeddings	Replace TF-IDF with Word2Vec, GloVe, or FastText
🧩 Deep Architectures	Use LSTM, GRU, or BiLSTM for sequential learning
🌐 Transfer Learning	Integrate BERT or DistilBERT for contextual embeddings
📈 Visualization Dashboard	Add training analytics via TensorBoard or Streamlit
🧮 Hyperparameter Optimization	Tune learning rate, dropout, and regularization strength using Optuna

---

👨‍💻 Author

Ali Khan AI Engineer 📧 alikhan132311@gmail.com

💡 Passionate about Deep Learning, NLP, and Model Optimization.

⭐ If you find this project helpful, please give it a star on GitHub! # Sentiment-Analysis-using-MLP # Sentiment-Analysis-using-MLP