Key Highlights#

• Dataset: 39,797 articles with 61 features
• Target: Social media shares (discretized into 5 classes)
• Challenge: Handling severely imbalanced classes
• Best Model: Support Vector Machine with strategic class balancing

Target Variable Distribution#

The original dataset showed extreme class imbalance, with the majority of articles having low share counts (class 0), making this a challenging classification problem.

1. Data Preprocessing & Cleaning#

Noise Detection and Removal:

• Identified and removed 135 rows with formatting issues
• Eliminated samples with values outside expected column domains
• Handled missing values and ‘n.a.’ entries
• Applied Z-score outlier detection (threshold: 5) removing 1,875 additional samples

Feature Engineering:

• Used Random Forest for feature importance analysis
• Retained only features with importance > 0.02
• Reduced from 61 to 18 most informative features
• Applied MinMax scaling for model compatibility

2. Class Imbalance Strategy#

The most critical challenge was the severe class imbalance. Instead of uniform balancing, I implemented a strategic approach:

Class Distribution Strategy:
- Class 0 (low shares): 1/3 of samples
- Class 1: 1/4 of samples  
- Class 2: 1/5 of samples
- Class 3: 1/6 of samples
- Class 4 (viral): 1/7 of samples

Rationale: This distribution acknowledges that viral content is naturally rare while ensuring all classes have sufficient representation for learning.

Implementation:

1. RandomUnderSampler: Reduced overrepresented classes
2. SMOTE: Generated synthetic samples for underrepresented classes
3. Maintained realistic class proportions reflecting real-world news sharing patterns

3. Model Selection & Evaluation#

Models Tested:

• Decision Trees
• Support Vector Machines
• AdaBoost Ensemble
• Random Forest
• Multi-Layer Perceptron Neural Networks

Evaluation Strategy:

• 66% training, 33% testing split
• Cross-validation with both accuracy and F1-macro scoring
• Emphasis on F1-macro to ensure sensitivity across all classes

Before Class Balancing#

After Class Balancing#

Hyperparameter Tuning Results#

Best SVM Configuration:

{
    'C': 1,
    'kernel': 'rbf',
    'gamma': 'scale',
    'decision_function_shape': 'ovo',
    'random_state': 42
}

1. The Class Imbalance Challenge#

The original dataset’s extreme imbalance (majority class dominance) led to models that achieved high accuracy by simply predicting the majority class. This highlighted the importance of F1-macro as an evaluation metric for imbalanced problems.

2. Strategic Balancing vs. Uniform Balancing#

Unlike uniform class balancing, the strategic approach maintained the natural hierarchy of news sharing patterns while ensuring model sensitivity to all classes. This proved crucial for real-world applicability.

3. Feature Importance Impact#

Reducing features from 61 to 18 using Random Forest importance analysis improved model efficiency without sacrificing performance, demonstrating the value of feature selection in high-dimensional problems.

4. Model Performance Trade-offs#

The class balancing strategy improved F1-macro scores significantly (from 0.18 to 0.25 for SVM) while maintaining reasonable accuracy, showing the successful navigation of the precision-recall trade-off.

Dependencies#

import pandas as pd
import numpy as np
import matplotlib.pyplot as plt
from sklearn.ensemble import RandomForestRegressor, RandomForestClassifier, AdaBoostClassifier
from sklearn.svm import SVC
from sklearn.tree import DecisionTreeClassifier
from sklearn.neural_network import MLPClassifier
from sklearn.preprocessing import MinMaxScaler, KBinsDiscretizer
from sklearn.model_selection import train_test_split, cross_val_score, GridSearchCV
from sklearn.metrics import classification_report, confusion_matrix
from imblearn.under_sampling import RandomUnderSampler
from imblearn.over_sampling import SMOTE
from scipy.stats import zscore

Key Functions#

• clean_and_load_dataset(): Comprehensive data cleaning pipeline
• get_hist(): Visualization utility for distribution analysis
• fit_models() & evaluate_models(): Model training and evaluation framework
• tune_model(): Hyperparameter optimization wrapper