🎮 A Video Game Review Experiment 🔫
👾 Review Sentiment Classification 😊😐🤨
Game Review Text
Prediction Results
stellar gameplay and capt ##ivating story make celestial odyssey a must - play ! engaging mechanics and stunning visuals keep you hooked from start to finish . a true gem in gaming !
Sentiment Analysis: Implementation
GOAL: Classify reviews into positive, neutral, and negative
sentiments.
CHALLENGE: The training dataset had ratings from 1 to 5, which didn’t directly map to these
three
sentiment labels. Additionally, the distribution of ratings was highly imbalanced (skewed
distribution).
Training Process and Results
- Approach: Instead of classifying reviews, DistilBERT predicts a continuous sentiment score (range: 0 to 1).
- Labeling Strategy: I converted this score into three sentiment categories.
- Binning: Positive (0.67-1.0) | Neutral (0.34-0.66) | Negative (0 - 0.33)
- Training Accuracy (5 Bins): Achieved 83% accuracy during testing.
- Validation Accuracy (3 Bins): Improved to 91% on generated reviews.
- Confidence Level: Monte Carlo Sampling (10 iterations) to estimate the model’s prediction confidence.
👾 Game Genres: Chaos to Clarity ✨
Cluster 1 💢 Combat-Focused Gameplay
Games featuring FPS, tactical shooters, and MOBAs, focused on teamwork and combat.
Unleash Your Inner Warrior
Cluster 2 💢 Engaging Simulated Worlds
Games featuring sports, racing, team challenges, and life or vehicle-based simulations.
Immerse Yourself in Simulated Realities
Cluster 3 💢 Action and Tactical Strategy
Games featuring action combat, exploration, abilities, and strategic planning and tactics.
Thrilling Action Meets Tactical Brilliance
Cluster 4 💢 Open Worlds and Discovery
Games featuring story-driven adventures, open worlds, survival, or sandbox gameplay.
Embark on Epic Journeys of Exploration
👾 Clustering & Summarization Tech 🧩📜
Topic Modeling: Implementation 🌐 Topic Modeling
GOAL: Group 23 game genres into a few main clusters for better analysis and
categorization.
CHALLENGE: Finding a balance between reducing genres and maintaining clear,
meaningful
clusters that
are easy to interpret.
Latent Dirichlet Allocation (LDA) 🌐 Latent Dirichlet Allocation
- Stopwords List: Carefully refined to remove irrelevant words and improve cluster clarity.
- Lemmatization: Standardized words to their base forms for consistent representation.
- Bag of Words: Transformed text data into numerical vectors for model input.
- Key Insights & Fine-Tuning:
- Optimizing the stopword list was crucial for generating meaningful clusters.
- Visualizations helped interpret and refine the clusters effectively.
Text Summarization: Implementation 🌐 LLM Text Summarization
GOAL: Summarize the top 3 games per cluster by extracting key pros and cons.
CHALLENGE: Ensuring the summarization processes every single game consistently.
OpenAI API Integration 🌐 OpenAI Api
- Model Selection: Prioritized faster inference for efficiency.
- Token Optimization: Minimized costs while maintaining prompt effectiveness.
- Deployment Optimization:
- HTML Generation: Ensured smooth integration with well-structured code.
- Async Handling & Lazy Loading: Improved response times for a better user experience.
