Reddit Post Popularity Classification

Project Overview

This repository presents an advanced end-to-end pipeline for classifying Reddit posts based on their popularity using state-of-the-art Natural Language Processing (NLP) and deep learning techniques. The framework is designed to handle textual content from social media and predict post popularity levels with high accuracy by integrating transformer-based language models, convolutional architectures, and autoencoder-based representations.

By utilizing BERT embeddings as the semantic foundation for the input text, this system demonstrates how modern language modeling can be adapted for classification tasks where subjective popularity (measured via upvote ratios) becomes the target label.

Classification Objective

Posts are labeled based on the upvote_ratio, a Reddit metric indicating the proportion of upvotes compared to total votes:

• Low Popularity: upvote ratio ≤ 0.5
• Average Popularity: 0.5 < upvote ratio ≤ 0.8
• High Popularity: upvote ratio > 0.8

This multi-class classification is valuable for content ranking, moderation prioritization, and community engagement modeling.

---

Technology Stack

Programming Language and Runtime

• Python 3.7+

Core Libraries

• Data Manipulation: pandas, numpy
• Visualization: matplotlib, seaborn, plotly
• Machine Learning: scikit-learn, imbalanced-learn
• Deep Learning Framework: TensorFlow with Keras
• NLP and Transformers: HuggingFace Transformers, tokenizers
• Model Evaluation: scikit-learn.metrics, classification_report, confusion_matrix

These technologies work in harmony to ensure modular, scalable, and reproducible experimentation.

---

System Architecture

1. Data Acquisition and Preprocessing

• Input data is expected in a structured .csv format with relevant metadata and post content.
• Columns used: subreddit, title, selftext, upvote_ratio.
• Preprocessing pipeline includes:
  • Null value removal and filtering of incomplete records.
  • Concatenation of title and selftext fields to capture complete context.
  • Stratified labeling based on upvote ratio thresholds.

2. BERT-based Text Embedding

• Utilizes Bidirectional Encoder Representations from Transformers (BERT) models for feature extraction:
  • Default Model: bert-base-uncased
  • Optional: TinyBERT, RoBERTa, DistilBERT, ALBERT
• Tokenization using WordPiece-based tokenizer.
• Outputs used:
  • CLS Token ([CLS]): Captures summary representation.
  • Pooled Output: Represents contextualized semantic embedding of the entire input.
• Feature vectors are extracted and cached for efficiency.

3. Class Imbalance Handling

Given the real-world skew in popularity data, this project implements robust strategies to mitigate class imbalance:

• Random Oversampling: Synthetic duplication of minority class samples.
• Random Undersampling: Controlled reduction of the majority class.
• SMOTE (Synthetic Minority Over-sampling Technique): Generates synthetic samples in feature space using k-nearest neighbors.

4. Model Architectures

A. Fully Connected Neural Networks

• BERT embeddings are passed through multiple dense layers with non-linear activation functions (ReLU).
• Dropout and batch normalization are employed to prevent overfitting.
• Output layer uses softmax activation for multi-class classification.

B. Convolutional Neural Networks (CNNs)

• Embeddings are reshaped into matrix format and passed through:
  • 1D convolutional filters to capture local n-gram semantics.
  • Max-pooling layers for dimensionality reduction and positional invariance.
  • Final dense layers for classification.

C. Autoencoder-Augmented Networks

• Two-stage process:
  • Stage 1: Train an unsupervised autoencoder to compress high-dimensional embeddings into latent space.
  • Stage 2: Use the encoder's output as input to a classification head (dense or CNN-based).
• Helps in learning compact, noise-reduced representations of semantic content.

---

Evaluation Pipeline

Metrics Tracked:

• Accuracy: Overall prediction correctness.
• Precision, Recall, F1-Score: Per-class and macro-averaged performance.
• Confusion Matrices: For detailed error analysis.
• Support: Number of samples per class to contextualize scores.

Visualization Techniques:

• Comparative plots using seaborn and matplotlib.
• Precision-recall trade-offs for each class.
• Heatmaps for model confusion matrices.
• Overall model benchmarking on a unified axis.

---

Key Python Classes and Methods

Class Name	Purpose
LabelClassification	Transforms continuous upvote ratios into categorical labels.
Preprocessing	Handles cleaning, embedding, and balancing.
NeuralNetwork	Base classifier model using dense architecture.
ConvolutionalDenseNetwork	Hybrid model combining CNN layers with dense outputs.
AutoencoderClassifier	Compresses semantic representations and classifies latent vectors.
Visualization	Provides statistical visualizations and comparisons.

---

Results & Analysis

Visual output is saved to the visualizations/ directory, including:

• confusion_matrices.png: Class-wise misclassifications.
• overall_accuracy.png: Total accuracy across models.
• label_accuracy.png: Accuracy distribution per popularity label.
• f1_score_comparison.png: F1 scores per model architecture.
• precision_recall_per_label.png: Label-specific trade-offs.
• confusion_matrix_difference.png: Delta heatmaps for model comparisons.

---

Project Highlights

• Transformer Integration: Exploits pretrained BERT and its variants for text understanding.
• Multi-architecture Benchmarking: Dense, CNN, and hybrid approaches evaluated in parallel.
• Imbalance Mitigation: Integrates SMOTE and resampling techniques for realistic datasets.
• Flexible Pipeline: Modular components enable adaptation for sentiment analysis, fake news detection, or toxicity classification.

---

Future Enhancements

• Expand to multi-modal analysis by integrating post metadata or image content.
• Incorporate attention-based pooling mechanisms for richer sentence representations.
• Implement cross-validation and hyperparameter optimization using Optuna or Ray Tune.
• Extend to zero-shot classification using sentence-transformers and T5.

---

License

This project is licensed under the MIT License. See the LICENSE file for usage rights and limitations.