https://github.com/laavanjan/youtube_comment_analysis_end_to_end

https://github.com/laavanjan/youtube_comment_analysis_end_to_end

Last synced: about 1 month ago
JSON representation

• Host: GitHub
• URL: https://github.com/laavanjan/youtube_comment_analysis_end_to_end
• Owner: laavanjan
• Created: 2025-07-31T06:34:39.000Z (2 months ago)
• Default Branch: main
• Last Pushed: 2025-07-31T11:12:57.000Z (2 months ago)
• Last Synced: 2025-07-31T11:59:59.910Z (2 months ago)
• Language: Jupyter Notebook
• Size: 3.27 MB
• Stars: 0
• Watchers: 0
• Forks: 0
• Open Issues: 0
• Metadata Files:
  • Readme: README.md

Awesome Lists containing this project

README

          


# YouTube Comment Sentiment Analysis: Project Overview

This project analyzes YouTube (and Reddit) comments for sentiment classification using natural language processing (NLP) and machine learning techniques. The workflow includes Jupyter notebooks for data science, a Flask API for model serving, and a Chrome extension frontend.

---

## Demo

Below are screenshots of the Chrome extension and API in action:





  


  

  

  

  

  


  



### Video Demo



  

    Your browser does not support the video tag.

  



---

## API & Frontend Components

### Flask API (`flask_api/main.py`)

- Provides REST endpoints for sentiment prediction.

- Loads the trained model from MLflow Model Registry and a local TF-IDF vectorizer.

- Endpoints:

  - `/predict` (POST): Predicts sentiment for a list of comments.

  - `/predict_with_timestamps` (POST): Predicts sentiment for comments with timestamps.

  - `/generate_chart` (POST): Generates charts (details in code).

- Example request to `/predict`:

  ```json

  {

    "comments": ["This video is awesome"]

  }

  ```

- Returns: List of comments with predicted sentiment.

### Chrome Extension Frontend (`yt-chrome-plugin-frontend/`)

- Contains `manifest.json`, `popup.html`, and `popup.js` for a Chrome extension UI.

- Allows users to interact with the Flask API directly from YouTube.

- UI is now a modern, light theme with color-coded sentiment cards (green for Positive, yellow for Neutral, red for Negative).

- Comments are displayed in clean, minimalist cards with timestamps and sentiment labels.

- Metrics and charts are shown in a visually appealing layout.

- Permissions in `manifest.json` include both `http://localhost/*` and `http://127.0.0.1/*` for local API access.

- See the folder for implementation details and customization options.

---

---

## Data & Model Artifacts

- All processed data is in the `data/` folder (raw and interim CSVs).

- Model artifacts: `lgbm_model.pkl`, `tfidf_vectorizer.pkl`.

- MLflow experiment tracking: `experiment_info.json`, `dvc.yaml`, `dvc.lock`.

---

## Project Structure

- `src/`: Data ingestion, preprocessing, model building, evaluation, and registration scripts.

- `notebooks/`: Jupyter notebooks for EDA, experiments, and model development.

- `flask_api/`: Flask REST API for serving predictions.

- `yt-chrome-plugin-frontend/`: Chrome extension for user interaction.

- `data/`: Raw and processed datasets.

- `requirements.txt`, `setup.py`: Project dependencies and setup.

## 1. 1_Preprocessing_&_EDA.ipynb

**Purpose:**

- Data cleaning and preprocessing

- Exploratory Data Analysis (EDA)

**Key Steps:**

- Loads the Reddit comments dataset from a public URL.

- Cleans the data by removing nulls, duplicates, and empty comments.

- Standardizes text (lowercasing, removing whitespace, newlines, and non-English characters).

- Adds features: word count, character count, stop word count, punctuation count.

- Visualizes data distributions (class balance, word/stopword distributions, boxplots, KDE plots).

- Extracts and visualizes most common words, bigrams, and trigrams.

- Removes stopwords (with some exceptions for sentiment), and applies lemmatization.

- Generates word clouds and stacked bar plots for word frequencies by sentiment category.

**Outcome:**

- Provides a clean, feature-rich dataset ready for modeling.

- Offers insights into comment length, vocabulary, and class imbalance.

## 2. 2_experiment_1_baseline_model.ipynb

**Purpose:**

- Build and evaluate a baseline sentiment classification model.

- Track experiments using MLflow.

**Key Steps:**

- Installs and configures MLflow for experiment tracking (with remote tracking URI).

- Loads and preprocesses the dataset (mirroring steps from the first notebook).

- Vectorizes comments using Bag of Words (CountVectorizer).

- Splits data into training and test sets.

- Trains a Random Forest classifier as a baseline model.

- Logs parameters, metrics, and artifacts (plots, datasets, models) to MLflow.

- Evaluates model performance (accuracy, classification report, confusion matrix).

- Saves the processed dataset for future use.

**Outcome:**

- Establishes a baseline for sentiment classification.

- Provides experiment tracking and reproducibility via MLflow.

---

## 3. 3_experiment_2_bow_vs_tfidf.ipynb

**Purpose:**

- Compare the performance of Bag of Words (BoW) and TF-IDF vectorization techniques for sentiment classification.

- Evaluate the impact of different n-gram ranges (unigrams, bigrams, trigrams) on model performance.

- Track all experiments using MLflow.

**Key Steps:**

- Loads the preprocessed dataset.

- Configures MLflow for experiment tracking.

- Defines a reusable experiment function to:

  - Vectorize text using either BoW or TF-IDF with specified n-gram ranges and feature limits.

  - Split data into training and test sets.

  - Train a Random Forest classifier.

  - Log parameters, metrics (accuracy, classification report), confusion matrix plots, and models to MLflow.

- Runs experiments for both BoW and TF-IDF with n-gram ranges of (1,1), (1,2), and (1,3).

- Compares results to determine which vectorization method and n-gram configuration performs best.

**Outcome:**

- Provides a systematic comparison of feature engineering strategies for text classification.

- Helps identify the optimal vectorization approach for sentiment analysis on this dataset.

---

## 4. 4_experiment_3_tfidf_(1,3)_max_features.ipynb

**Purpose:**

- Investigate the effect of varying the `max_features` parameter in TF-IDF vectorization (with trigrams) on sentiment classification performance.

- Identify the optimal number of features for the best model accuracy.

- Track all experiments using MLflow.

**Key Steps:**

- Loads the preprocessed dataset.

- Configures MLflow for experiment tracking.

- Defines an experiment function to:

  - Vectorize text using TF-IDF with ngram_range=(1,3) and different max_features values.

  - Split data into training and test sets.

  - Train a Random Forest classifier.

  - Log parameters, metrics (accuracy, classification report), confusion matrix plots, and models to MLflow.

- Iterates over a range of max_features values (e.g., 1000 to 10000) to compare results.

- Analyzes how feature dimensionality impacts model performance.

**Outcome:**

- Provides insight into the trade-off between feature size and classification accuracy.

- Helps select the best max_features value for TF-IDF-based sentiment models.

---

## 5. 5_experiment_4_handling_imbalanced_data.ipynb

**Purpose:**

- Address class imbalance in the sentiment dataset using various resampling and class weighting techniques.

- Evaluate the impact of different imbalance handling methods on model performance.

- Track all experiments using MLflow.

**Key Steps:**

- Loads the preprocessed dataset.

- Configures MLflow for experiment tracking.

- Defines an experiment function to:

  - Vectorize text using TF-IDF with ngram_range=(1,3) and a fixed max_features value.

  - Split data into training and test sets.

  - Apply different imbalance handling strategies to the training set:

    - Class weighting in Random Forest

    - SMOTE (oversampling)

    - ADASYN (oversampling)

    - Random undersampling

    - SMOTEENN (combined over- and under-sampling)

  - Train a Random Forest classifier.

  - Log parameters, metrics (accuracy, classification report), confusion matrix plots, and models to MLflow.

- Compares the results of each imbalance handling method.

**Outcome:**

- Provides insight into the effectiveness of different strategies for handling imbalanced data in sentiment classification.

- Helps select the best approach for improving model performance on minority classes.

---

## 6. 6_experiment_5_xgboost_with_hpt.ipynb

**Purpose:**

- Compare multiple machine learning algorithms for sentiment classification, including XGBoost, Logistic Regression, SVM, LightGBM, KNN, Naive Bayes, and Random Forest.

- Use Optuna for hyperparameter tuning of XGBoost to optimize model performance.

- Address class imbalance using SMOTE.

- Track all experiments using MLflow.

**Key Steps:**

- Loads the preprocessed dataset.

- Remaps class labels for compatibility with XGBoost.

- Splits data into training and test sets.

- Vectorizes text using TF-IDF with trigrams and a fixed max_features value.

- Applies SMOTE to balance the training data.

- Defines a logging function to record model parameters, metrics, and artifacts in MLflow.

- Compares several algorithms by training and evaluating each, logging results to MLflow.

- Uses Optuna to tune XGBoost hyperparameters (n_estimators, learning_rate, max_depth), selects the best model, and logs it.

**Outcome:**

- Provides a comprehensive comparison of popular ML algorithms for sentiment analysis.

- Identifies the best-performing model and hyperparameters for the task.

- Demonstrates the use of automated hyperparameter optimization and experiment tracking.

---

## 7. 7_experiment_6_lightgbm_detailed_hpt.ipynb

**Purpose:**

- Perform detailed hyperparameter tuning for LightGBM using Optuna to maximize sentiment classification performance.

- Address class imbalance using SMOTE.

- Track all experiments and hyperparameter trials using MLflow.

**Key Steps:**

- Loads the preprocessed dataset and remaps class labels for compatibility.

- Vectorizes text using TF-IDF with trigrams and a fixed max_features value.

- Applies SMOTE to balance the dataset.

- Splits the data into training and test sets after resampling.

- Defines a logging function to record model parameters, metrics, and artifacts in MLflow for each Optuna trial.

- Uses Optuna to search a wide hyperparameter space for LightGBM (n_estimators, learning_rate, max_depth, num_leaves, min_child_samples, colsample_bytree, subsample, reg_alpha, reg_lambda).

- Logs each trial and the best model to MLflow, and visualizes parameter importance and optimization history.

**Outcome:**

- Identifies the best hyperparameters for LightGBM on this sentiment analysis task.

- Demonstrates advanced experiment tracking and hyperparameter optimization workflows.

---

## 8. 8_stacking.ipynb

**Purpose:**

- Implement a stacking ensemble model for sentiment classification using LightGBM and Logistic Regression as base learners, and KNN as the meta-learner.

- Compare the performance of the stacking model with individual models.

**Key Steps:**

- Loads the preprocessed dataset and drops rows with missing values.

- Vectorizes text using TF-IDF with trigrams and a fixed max_features value.

- Splits the data into training and test sets.

- Defines LightGBM and Logistic Regression as base models, and KNN as the meta-learner.

- Constructs a StackingClassifier with 5-fold cross-validation.

- Trains the stacking model and evaluates its performance on the test set.

- Compares the stacking model's results to those of the LightGBM model alone.

**Outcome:**

- Demonstrates the use of ensemble learning to potentially improve sentiment classification performance.

- Provides insight into the complexity and benefits of stacking versus single-model approaches.

---

## How to Use

1. Run `1_Preprocessing_&_EDA.ipynb` to explore and preprocess the data.

2. Run `2_experiment_1_baseline_model.ipynb` to train and evaluate the baseline model, and log results to MLflow.

3. Run `3_experiment_2_bow_vs_tfidf.ipynb` to compare BoW and TF-IDF vectorization methods and analyze their impact on model performance.

4. Run `4_experiment_3_tfidf_(1,3)_max_features.ipynb` to experiment with different max_features settings for TF-IDF trigrams and optimize feature selection.

5. Run `5_experiment_4_handling_imbalanced_data.ipynb` to explore and compare different techniques for handling class imbalance in the dataset.

6. Run `6_experiment_5_xgboost_with_hpt.ipynb` to compare ML algorithms and tune XGBoost with Optuna for best performance.

7. Run `7_experiment_6_lightgbm_detailed_hpt.ipynb` to perform detailed hyperparameter tuning for LightGBM with Optuna and MLflow.

8. Run `8_stacking.ipynb` to experiment with stacking ensemble models for sentiment classification.

---

**Note:**

- The project uses Python, pandas, scikit-learn, NLTK, seaborn, matplotlib, MLflow, Flask, and Chrome extension technologies.

- For MLflow tracking, ensure the tracking server URI is accessible.

- The workflow can be extended with more advanced models, APIs, and frontend features.