|
|
---
|
|
|
library_name: sklearn
|
|
|
tags:
|
|
|
- text-classification
|
|
|
- cyberbullying
|
|
|
- nlp
|
|
|
- social-impact
|
|
|
- ensemble-learning
|
|
|
dataset_info:
|
|
|
name: Cyberbullying Classification
|
|
|
source: Kaggle (andrewmvd/cyberbullying-classification)
|
|
|
metrics:
|
|
|
- accuracy
|
|
|
model_file: voting_classifier_model.pkl
|
|
|
---
|
|
|
|
|
|
# Cyberbullying Classification Model (Scikit-Learn) |
|
|
|
|
|
This is a traditional Machine Learning model that classifies tweets into different categories of cyberbullying. It is an ensemble **Voting Classifier** combining **Logistic Regression** and **Random Forest**, achieving approximately **91% accuracy**. |
|
|
|
|
|
## Model Details |
|
|
|
|
|
- **Developed by:** Kohil Sharma |
|
|
- **Model Type:** Voting Classifier (Logistic Regression + Random Forest) |
|
|
- **Feature Extraction:** TF-IDF (Term Frequency-Inverse Document Frequency) |
|
|
- **Library:** Scikit-Learn |
|
|
- **Language:** English |
|
|
|
|
|
## Intended Use |
|
|
This model is designed to detect specific types of cyberbullying in text. It is lightweight and faster than transformer models, making it suitable for low-resource environments. |
|
|
|
|
|
### Classification Labels |
|
|
The model classifies text into **5 categories** (mapped as follows): |
|
|
- `0`: **Not Cyberbullying** |
|
|
- `1`: **Gender** (Sexist) |
|
|
- `2`: **Religion** |
|
|
- `3`: **Age** |
|
|
- `4`: **Ethnicity** (Racist) |
|
|
|
|
|
*(Note: The 'Other' category was removed during preprocessing to improve accuracy.)* |
|
|
|
|
|
## Training Data |
|
|
- **Dataset:** [Cyberbullying Classification Tweets](https://www.kaggle.com/datasets/andrewmvd/cyberbullying-classification) |
|
|
- **Original Size:** ~47,000 tweets |
|
|
- **Processed Size:** ~38,000 tweets (after cleaning duplicates and removing the 'Other' class) |
|
|
|
|
|
## Training Procedure |
|
|
|
|
|
### 1. Preprocessing |
|
|
The text underwent rigorous cleaning using the `tweet-preprocessor` library and custom functions: |
|
|
- Removal of Usernames (@), Hashtags (#), and Links (http). |
|
|
- Removal of punctuation and special characters. |
|
|
- Conversion to lowercase. |
|
|
- **Lemmatization** using NLTK's `WordNetLemmatizer`. |
|
|
- Stopword removal (including Twitter-specific stopwords like "rt", "mkr"). |
|
|
|
|
|
### 2. Feature Engineering |
|
|
- **TF-IDF Vectorizer** was used to convert text into numerical vectors. |
|
|
|
|
|
### 3. Model Architecture |
|
|
- **Base Models:** |
|
|
1. `LogisticRegression` (C=100, penalty='l2') |
|
|
2. `RandomForestClassifier` (n_estimators=100) |
|
|
- **Ensemble:** `VotingClassifier` (Hard Voting) combining the above two. |
|
|
|
|
|
## Evaluation Results |
|
|
- **Accuracy:** ~91% on the test set. |
|
|
- **Strengths:** High precision in distinguishing Ethnicity, Religion, and Age-based bullying. |
|
|
|
|
|
## How to Use |
|
|
|
|
|
To use this model in Python, you need to load both the vectorizer and the model using `joblib`. |
|
|
|
|
|
```python |
|
|
import joblib |
|
|
import preprocessor as p # pip install tweet-preprocessor |
|
|
import string |
|
|
|
|
|
# 1. Load the saved files |
|
|
model = joblib.load('model.pickle') |
|
|
vectorizer = joblib.load('tfidf.pickle') |
|
|
|
|
|
# 2. Define the cleaning function (Must match training!) |
|
|
def clean_text(text): |
|
|
text = p.clean(text) |
|
|
text = text.lower() |
|
|
text = "".join([char for char in text if char not in string.punctuation]) |
|
|
return text |
|
|
|
|
|
# 3. Make a prediction |
|
|
text = "You are dumb and you should go back to school." |
|
|
clean_input = clean_text(text) |
|
|
|
|
|
# Vectorize the text |
|
|
vectorized_input = vectorizer.transform([clean_input]) |
|
|
|
|
|
# Predict |
|
|
prediction = model.predict(vectorized_input) |
|
|
classes = {0: 'Not Cyberbullying', 1: 'Gender', 2: 'Religion', 3: 'Age', 4: 'Ethnicity'} |
|
|
|
|
|
print(f"Prediction: {classes[prediction[0]]}") |
