GameLift Analytics
Game Recommendation
Impact Dashboard
Discover how we transformed user engagement through the evolution of our recommendation system, achieving 3x improvement in user retention and 10x boost in game discovery.
Monthly Active Users
Mini-Games
Months Journey
MAU Growth
Gaming Platform Overview
Understanding our ecosystem: a thriving mobile gaming platform with diverse user base and extensive game catalog powering a credits-based economy.
Monthly Active Users
Engaged players across our platform
Mini-Games
Diverse portfolio of gaming experiences
Credits Economy
Monthly credits purchased by users
Daily Active Users
Percentage of MAU active daily
Platform Architecture
Mobile App
iOS & Android gaming platform with seamless user experience
Game Catalog
Extensive library of mini-games across multiple genres
Credits Economy
Virtual currency system driving engagement and monetization
But we faced significant challenges...
Revenue Concentration
Despite our large user base and extensive game catalog, we faced the classic 80/20 problem - a small percentage of users drove most revenue while the majority remained underengaged.
User Segmentation
Segment Analysis
Heavy Spenders
20%Generate 80% of revenue but play limited games
Casual Players
60%Regular users with moderate engagement
Inactive Users
20%Low engagement, at risk of churning
Game Discovery
Users struggled to find games matching their preferences from our 800+ catalog
User Engagement
Average 1.5 games per user indicated limited exploration and repeat engagement
Revenue Growth
Over-reliance on 20% of users created vulnerability and limited monetization potential
Baseline Metrics (Before Recommendations)
Avg Games/User
Limited game exploration
Daily Engagement
Users active daily
Paid Conversion
Users making purchases
Credits/User
Average spend per user
The Core Problem
Revenue Distribution
Classic 80/20 problem visualization
The Challenge
- • Over-reliance on 20% of users for revenue sustainability
- • 80% of users generate minimal revenue despite large user base
- • Risk of revenue loss if heavy spenders churn
- • Untapped potential in casual and inactive user segments
How do we solve this?
User-Based Collaborative Filtering
Our first approach: analyze user similarity patterns to recommend games based on what similar users enjoyed. Using 6 different data types to build comprehensive user profiles.
User Similarity Matrix
User Similarity Calculations
Hover over cells to see similarity scores between users
Recommendation Generation
Recommendation Process
"Users like you will enjoyed also..."
Target User: Alex
Preferences: puzzle, strategy, casual
Similar Users Found
Personalized Recommendations
- Similar users rated it 5 stars
- Matches your puzzle preference
- High engagement in your demographic
- 3 similar users played for 10+ hours
- Aligns with strategy preference
- Popular among casual gamers
- Users with similar tastes loved it
- Combines puzzle and strategy
- Trending in your region
6 Data Types Used
Playtime Data
User session duration, total time spent, and gaming patterns
Sessions Patterns
Login frequency, session timing, and activity streaks
Game Play Frequency
How often users play different games and retry patterns
Achievement Completion
Progress tracking, completion rates, and skill assessment
Credit Purchase Behaviour
In-app purchases, credits spending, and monetization patterns
User Rating Data
Game ratings, reviews, and explicit feedback from users
Live Filtering Demo
Select Sample User
Alex (Casual Gamer)
Maria (Strategy Fan)
Jake (Action Player)
Recommendations for Select a user
How It Works
1. Find Similar Users
Calculate cosine similarity between user behavior vectors
2. Identify Preferences
Analyze games that similar users enjoyed most
3. Generate Recommendations
Rank and serve personalized game suggestions
Let's see how this performed...
Promising Initial Improvements
User-based collaborative filtering delivered measurable improvements across key metrics. Our foundation was solid, but we knew we could do better.
Monthly Active Users
Steady growth from improved recommendations
Game Events
Users engaging more with recommended games
Credits Purchase
Modest revenue increase from better targeting
Feature Adoption
Good user acceptance of new recommendation engine
User Behavior Impact
Technical Performance
Performance vs. Baseline
Monthly Active Users
up from 150K baseline
Game Events
up from 13M baseline
Credits Purchased
Up from 8M baseline
Paid User Conversion
Increased conversion rate
Key Insights from Phase 1
Solid Foundation
User-based filtering provided measurable improvements across all key metrics
Room for Growth
Cold start problem and modest revenue impact highlighted areas for improvement
Next Phase Ready
Strong user adoption set the stage for advanced machine learning approaches
But we knew we could achieve much more...
LightFM Implementation
Revolutionary hybrid approach: simultaneously learning user preferences and game characteristics to create powerful, scalable recommendations that solved our initial challenges.
From Collaborative Filtering to Hybrid Intelligence
User-Based CF
Find similar users → recommend what they liked
LightFM Hybrid
Learn user preferences + game features
Technical Deep Dive
💻 LightFM Implementation
from lightfm import LightFM
from lightfm.datasets import fetch_movielens
# User-Based Collaborative Filtering
User_1: [Strategic_Conquest: 5⭐, Adventure_Quest: 5⭐, Speed_Racer_3D: 1⭐]
User_2: [Strategic_Conquest: 5⭐, Adventure_Quest: 5⭐, Action_Heroes: 5⭐]
# LightFM learns this:
User_1: Strategy_preference=HIGH, Racing_preference=LOW
User_2: Strategy_preference=HIGH, Action_preference=HIGH
# Initialize hybrid model
model = LightFM(loss='warp', no_components=100)
model.fit(interactions, user_features, item_features)🎯 Feature Matrix Learning
Feature Learning Matrix
Game Characteristics
Live Preference Learning
New User Interaction
Strategic Conquest
New User
Rating
Word Puzzle Master
New User
Rating
Speed Racer 3D
New User
Rating
Model Learning
Neural Network Ready
Waiting for user interaction...
Hybrid Matrix Factorization
Updated Recommendations
Chess Champions
Strategy + Complex gameplay preference learned
Logic Masters
Puzzle-solving pattern detected
City Builder Deluxe
Strategic thinking + planning preference
Algorithm Comparison
🔄 User-Based Collaborative Filtering
1 Find Similar Users
2 Analyze Their Games
3 Recommend Similar Games
⚠️ Limitations
- • Can't recommend brand new games
- • Slow with large user bases
- • Limited by user similarity only
- • Cold start problem for new users
⚡ LightFM Hybrid Model
1 Learn User Preferences
2 Learn Game Characteristics
3 Hybrid Prediction
✨ Advantages
- • ✅ Handles brand new games perfectly
- • ✅ Fast predictions via dot products
- • ✅ Learns both user and item features
- • ✅ Scales to millions of users/items
- • ✅ Cold start problem solved
Dramatic Performance Improvement
MAU Growth
Revenue
Game Events
Adoption Rate
Unique Games Played
Paid Users
Paid Conversion
Implementation Success
The LightFM implementation delivered exceptional results, transforming our recommendation system from a basic collaborative filter to a sophisticated hybrid model that understands both users and games.
The Journey to Intelligence
From traditional collaborative filtering to advanced machine learning - how we transformed our recommendation system to deliver 25% MAU growth.
The Revenue Concentration Problem
Starting Point - March 2024
80% of revenue from just 20% of users. Poor engagement patterns created revenue concentration issues.
User-Based Collaborative Filtering
Phase 1 - April 2024
Implemented traditional collaborative filtering using cosine similarity across 6 data types.
Solid foundation with room for optimization
The Machine Learning Insight
Research Phase - May 2024
Discovered LightFM's hybrid approach could solve cold start problems while improving accuracy.
- • Hybrid collaborative + content-based filtering
- • Built-in cold start handling
- • Scalable matrix factorization
- • Multiple loss functions for optimization
LightFM Implementation
Phase 2 - June-September 2024
Full migration to LightFM with hybrid features, optimized training pipeline, and A/B testing.
Exceeded all targets, transformed user engagement
Production Excellence
Current State - September 2024
Stable production system serving 150K+ MAU with continuous learning and optimization.
User Engagement
Revenue Impact
AVG Unique Game Play
Paid Users
Transformation Timeline
Let's dive into the technical details...
Production to Data Lake ELT
Real-time data ingestion from production pipelines systems
Google Analytics ETL
Marketing and web analytics data processing
POS/Currency/CRM
Point of sale, currency exchange, and CRM data
Self-Service Dashboards
Automated dashboard data feeds
MASTER Datasets
Consolidated master data management