Sentiment-Analysis-with-Playstore-reviews

Performed sentiment analysis for XYZ company on playstore reviews to categorize customer reviews as 'POSITIVE' or 'NEGATIVE'

INTRODUCTION:

AIM: To perform sentiment analysis on Google Play Store app reviews, classifying them as either "Positive" or "Negative"
DATASET USED: Sample dataset was sourced from kaggle.
TOOLS AND LIBRARIES: This project is made with Python and uses:

• NLTK for text preprocessing
  
• sci-kit learn for machine learning for ML models (Logistic regression, naive bayes)
  
• Pandas for data manipulation
  
• Seaborn and Matplotlib for data visualization (making confusion matrices)
  

DATA UNDERSTANDING:

Dataset had 2 useful columns with user reviews and another one with score for those reviews on a scale of 1 to 5, where:

• 1 = Very Negative
  
• 2 = Negative
  
• 3 = Neutral
  
• 4 = Positive
  
• 5 = Very Positive
  

DATA PREPROCESSING:

Text Cleaning: Used NLTK for:

• Tokenization
  
• Stop word removal
  
• Lemmatization
  

Label Assignment:
Scores of 1 and 2 are labeled as negative
Scores of 4 and 5 are labeled as positive
Neutral Scores (3) are removed from the dataset

FEATURE EXTRACTION:

TF-IDF (Text-frequency inverse document frequency):
Used TfidfVectorizer to convert the cleaned text into numerical features suitable for machine learning models.
Limited the feature size to 6000 terms for efficient computation while preventing overfit.

MODEL IMPLEMENTATION:

Logistic Regression:

• Initially implemented logistic regression
  
• Accuracy achieved: 87%
  
• Pros: Simple and easy to interpret, excellent for binary classification
  
• Cons: Assumes linear relation between features, and best useful when datasets are small- medium sized.
  

Naive Bayes:

• Decided to implement naive bayes to compare accuracy
  
• Achieved accuracy of 85%
  
• Pros: Simple and effective for text processing
  
• Cons: Assumes no interrelation between words, hence ‘naive’
  

CONCLUSION:

Logistic Regression was the best-performing model with an accuracy of 87%. Naive Bayes came close but was slightly lower in performance.

Future Work:

Use advanced models and explore word embeddings.