Classification of Amazon Fine Food Reviews using LSTM

Team Members

• Aashka Chhabra
• Anchi Kansal
• Anubha Sharma
• Ansh Shekokar
• Anushka Verma

---

Model Architecture

Embedding Layer

The Embedding Layer converts words into dense vectors of fixed size, allowing the model to learn relationships between words.

Example:

• Words like "king" and "queen" might be positioned near each other in vector space, capturing semantic similarity.

---

LSTM Layer (Long Short-Term Memory)

The LSTM Layer is responsible for processing sequences while retaining context over long sentences.

Why LSTM?

• Unlike traditional neural networks, LSTMs handle long-term dependencies.
• It remembers important contextual information, e.g., "not good" has a different sentiment than "good".

---

Dense Layer

The Dense Layer is the output layer responsible for making predictions based on learned patterns.

Fully connected neurons aggregate information from the previous layers.

The number of neurons depends on the output requirement (e.g., 1 neuron for binary classification).

---

Sigmoid Activation Function

The Sigmoid Activation Function is used in the final layer to output probabilities between 0 and 1.

Why Sigmoid?

• Since this is a binary classification problem (positive or negative sentiment), sigmoid helps interpret the model's confidence in a particular class.

Formula:
[ sigma(x) = 1/(1 + e^(-x)) ]
where ( x ) is the input from the previous layer.

---

Dropout Regularization

The Dropout Layer is used to prevent overfitting by randomly disabling neurons during training.

Why Dropout?

• Forces the network to learn redundant patterns, improving generalization.
• Helps in reducing model complexity and preventing over-reliance on specific neurons.

---

Training and Testing the Model

The model will be trained and tested using the Amazon Fine Food Reviews Dataset from Kaggle.

Steps:

1. Data Preprocessing:
   • Remove punctuation, numbers, and stop words.
   • Normalize uppercase and lowercase letters.
2. Tokenization & Padding:
   • Convert text into numerical sequences.
   • Apply padding to ensure uniform input sizes.
3. Model Training:
   • Train using LSTM with different hyperparameters.
4. Evaluation:
   • Compare performance using accuracy and loss metrics.

---

Dataset Information

• Dataset: Amazon Fine Food Reviews
• Total Reviews: 568,454
• Timeframe: Oct 1999 – Oct 2012
• Total Products: 74,258

---

Summary

This project implements sentiment analysis using LSTMs to classify customer reviews as positive or negative. By using embedding layers, LSTMs, dense layers, and dropout regularization, the model learns contextual relationships in text data.

Key Benefits:

• Improved sentiment classification using deep learning.
• Context-aware feature extraction via LSTM.
• Robust performance with dropout regularization.

---