FeClustRE: Feature Clustering and Semantic Tagging of Mobile App Review Feature Taxonomies

Overview

FeClustRE is a system for extracting app features from reviews, grouping them into hierarchical taxonomies, and assigning semantic labels to each cluster using LLMs. The pipeline stores taxonomies in Neo4j and provides utilities to assess the quality and coherence of these structures.

This repository serves as the full replication package for our paper, including code, metrics, and automated evaluation framework.

---

Quick Start - Replication Package

1. Install Dependencies

git clone https://github.com/your-org/feclustre.git
cd feclustre
pip install -r requirements.txt

2. Start Neo4j

docker run -d --name neo4j \
  -p7474:7474 -p7687:7687 \
  -e NEO4J_AUTH=neo4j/12345678 \
  neo4j:5.15

3. Start Ollama with Qwen

ollama run qwen:1.8b

Ensure your config.py includes:

OLLAMA_BASE_URL = "http://localhost:11434"
OLLAMA_MODEL = "qwen:1.8b"

4. Start the Backend API Server

python app.py

The API server will be available at http://localhost:3000

5. Run Systematic Experiments

All experiments from the paper can be replicated using the systematic testing framework:

# Mock test (quick validation with 2 apps, small samples)
python test_config.py mock

# Full comprehensive test (all apps, multiple configurations)
python test_config.py full

# Semantic experiment (hybrid model, maximum data, balanced strategy)
python test_config.py semantic

Experiment Configurations

Mock Test:

• Apps: ChatGPT, Claude by Anthropic
• Models: transfeatex, t-frex, hybrid
• Embeddings: allmini, sentence-t5
• Sample sizes: 10, 20, 50
• Strategies: balanced, silhouette, conservative

Full Test:

• Apps: All 7 AI assistants (ChatGPT, Claude, DeepSeek, Gemini, Le Chat, Copilot, Perplexity)
• Models: t-frex, transfeatex, hybrid
• Embeddings: allmini, sentence-t5
• Sample sizes: 1000, 2000, 5000, 50000
• Strategies: balanced, silhouette, conservative

Semantic Experiment:

• Apps: All 7 AI assistants
• Models: hybrid
• Embeddings: allmini
• Sample sizes: No limit (all available reviews)
• Strategies: balanced

---

Dataset

The dataset is available in /data/input/endpoint_1_process_reviews/ directory with two categories:

AI Assistants (/data/input/endpoint_1_process_reviews/ai_assistants/)

• ChatGPT.csv
• Claude_by_Anthropic.csv
• DeepSeek_-_AI_Assistant.csv
• Google_Gemini.csv
• Le_Chat_by_Mistral_AI.csv
• Microsoft_Copilot.csv
• Perplexity_-_Ask_Anything.csv

Query All Taxonomies

MATCH (app:App)-[:HAS_MINI_TAXONOMY]->(root:MiniTaxonomyNode)
WHERE NOT ()-[:HAS_CHILD]->(root)
OPTIONAL MATCH (root)-[:HAS_CHILD*]->(leaf)
WHERE NOT (leaf)-[:HAS_CHILD]->()
RETURN app.name as app_name, 
       root.llm_tag as taxonomy_label,
       count(DISTINCT leaf) as leaf_count,
       collect(DISTINCT leaf.feature) as features