Custom-Made Games with Machine Learning and Big Data
1 like159 views
Yokozuna Data, founded by Dr. África Periáñez, specializes in using machine learning and big data to enhance personalized gaming experiences by predicting player behaviors and preferences. The company focuses on developing actionable insights from game data to improve player retention, optimize game design, and boost revenue through targeted strategies. Their advanced techniques include the use of ensemble learning and convolutional neural networks for behavioral prediction and item recommendation within the video game industry.
![CONVOLUTIONAL NEURAL NETWORK (CNN, CONVNET)
Convolutional layers Max pooling layers
Training:
Optimize weights and bias
to minimize error.
Downsampling
With filters that cover more than one input, CNNs can learn local
connectivity between inputs.
Proposed by LeCun et al. in 1998 [1], CNNs have been widely applied
to image processing, signal processing, and time series prediction.
[1] LeCun, Yann, et al. "Object recognition with gradient-based learning." Shape, contour and grouping in computer vision. Springer, Berlin, Heidelberg, 1999. 319-345.
Reference:
https://adeshpande3.github.io/A-Beginner%27s-Guide-To-Understanding-Convolutional-Neural-Networks/
https://medium.freecodecamp.org/an-intuitive-guide-to-convolutional-neural-networks-260c2de0a050](https://image.slidesharecdn.com/ydperianezgdc2019-190402094346/85/Custom-Made-Games-with-Machine-Learning-and-Big-Data-34-320.jpg)
Custom-Made Games with Machine Learning and Big Data
- 1. Custom-Made Games with Machine Learning and Big Data África Periáñez, PhD CEO, Yokozuna Data GDC2019 San Francisco
- 2. @aperianez Dr. África Periáñez Founder and CEO of Yokozuna Data PhD Mathematics (Ensemble Learning) - 2015 University of Reading MSc String Theory - 2006 CERN MSc Theoretical Physics - 2003 UAM BSc Physics - 2001 UAM
- 4. Founded in 2015, joined Keywords Studios in 2018 to push back the frontiers of General Behavioral Machine Learning and to revamp the video-game industry: Personalized games What is Yokozuna Data?
- 5. Mission To unlock the knowledge of big game databases To convert unstructured data into actionable information in order to understand and predict individual player behavior
- 6. THE TEAMTHE TEAM Vitor Santos, MA DESIGN & BUSINESS DIRECTOR Álvaro de Benito, MA PR & COMMUNICATION LEAD Nitin Kumar, MSc FULL-STACK ENGINEER África Periáñez, PhD Founder & CEO Pooja Revanna, MSc BACKEND ENGINEER Omid Aladini, MSc DATA INFRASTRUCTURE ADVISOR Dexian Tang, MSc BIG DATA ENGINEER Javier Grande, PhD SCIENTIFIC EDITOR Yu-Kai Hung, MSc COMMUNITY MANAGER FOR ASIA Ana Fernández, MSc SENIOR RESEARCH DATA SCIENTIST Pei Pei Chen, MSc MACHINE LEARNING ENGINEER LEAD Anna Guitart, MSc DATA SCIENTIST Jing Li, PhD MACHINE LEARNING ENGINEER Cristian Conteduca, MSc BACKEND ENGINEER LEAD Peng Xiao, BSc BIG DATA ENGINEER Shi Hui Tan, MSc DATA SCIENTIST
- 7. + +AI CLOUD DEVOPS
- 8. Yokozuna Data AWS GameSparks AI Cloud DevOps
- 9. PUSHING BACK THE FRONTIERS OF GAME DATA SCIENCE: A state-of-the-art machine learning engine that predicts individual player behavior
- 11. Highly sophisticated games allow players to express nuanced emotions through their in-game actions
- 14. WHEN WILL PLAYERS LEAVE THE GAME? DATE LEVEL PLAYTIME MONEY
- 15. WHICH ITEM WILL THEY PURCHASE NEXT? Billing history Distribution of item probabilities Time to next purchase 08/06 08/15 08/27 09/06 Playtime GAME APPLICATIONS Players who may stop purchasing Upcoming churners
- 16. RECOMMENDS THE BEST SEQUENCE OF EVENTS TO MAXIMIZE PLAYER ENGAGEMENT CONSIDERING EXTERNAL AND ENVIRONMENTAL FACTORS
- 17. Personalized matching Who should you compete against in Mario Kart? Which clan is your best opponent in Clash of Clans? PERSONALIZATION
- 19. Item recommendation system Action recommendation Rewards and discounts Engagement and retention-motivated actionable recommendations PERSONALIZATION
- 20. VIDEO GAME CHALLENGES Predicting player behavioral outcomes is key to success of game developers
- 21. Man is a deterministic device thrown into a probalistic Universe Amos Tversky Daniel Kahneman Nobel prize 2002
- 22. NEW PLAYERS Acquiring new users is expensive: in Japan, the average cost-per-install for gaming apps reached 6.07 USD in 2018 Video game challenges: Increasing Retention Between 75–90% of new players churn on the first day and inefficient: only 5% remain after one month 1 Liftoff and Adjust, 2018
- 23. VIP PLAYERS Retention of the most valuable players is crucial: the top 10% of paying users contribute 60% of the revenue Thousands of titles are published every year and compete for same playersʼ time and attention Video game challenges: Increasing Retention
- 24. Only a small fraction of users make purchases, identifying these users and predicting their Customer Lifetime Value is crucial A key challenge for game developers is to convert players from non-premium to premium Actionable Goals: 1) tailor marketing efforts: In-game targeting of advertising and price promotions 2) customize game difficulty, e.g. dynamic difficulty adjustment 3) manage user acquisition campaigns Video game challenges: Maximizing the engagement of VIP players
- 25. Targeting the right players by customizing game events and publishing them at the right time Identify the reasons behind behavioral trends Find the best acquisition, marketing and game event strategies Video game challenges: Optimizing Game Events and Marketing Campaigns
- 26. SOLVING VIDEO GAME BUSINESS CHALLENGES
- 27. $1.3M/month sales increase +10% VIP Retention $4M/month sales increase +5% PU Increase $8M/year sales increase Improving development of levels with the highest churn rate NUMBERS FOR AAA MOBILE GAMES
- 28. $200K/month sales increase $70K/month sales increase +5% PU Increase $10M/year sales increase +10% VIP Retention Improving development of levels with the highest churn rate NUMBERS FOR WESTERN CASUAL GAMES
- 29. 1/ OPERATIONAL PLAYER BEHAVIORAL PREDICTION 2/ DEEP LEARNING AND ENSEMBLE LEARNING 3/ BIG DATA
- 30. ENSEMBLE LEARNING Multiple learners are trained to solve the same problem Ordinary machine learning approaches learn one hypothesis from training data Ensemble methods try to construct a set of hypotheses and combine them OUTPUT
- 31. label Reconstruction Error Output nth feature layer 2nd feature layer 1st feature layer Input label Learning Learning&Generalization Deep Learning ModelArtificial Neural Network DEEP LEARNING
- 32. PLAYER LIFETIME VALUE PREDICTION? UPP: UPCOMING PURCHASES PREDICTION Lifetime value (LTV), is an estimate about the amount a player will spend from today until she exits the game
- 33. DATASET Unknown churn day Various lifetime (from 1 day to years) CURRENT DAY
- 34. CONVOLUTIONAL NEURAL NETWORK (CNN, CONVNET) Convolutional layers Max pooling layers Training: Optimize weights and bias to minimize error. Downsampling With filters that cover more than one input, CNNs can learn local connectivity between inputs. Proposed by LeCun et al. in 1998 [1], CNNs have been widely applied to image processing, signal processing, and time series prediction. [1] LeCun, Yann, et al. "Object recognition with gradient-based learning." Shape, contour and grouping in computer vision. Springer, Berlin, Heidelberg, 1999. 319-345. Reference: https://adeshpande3.github.io/A-Beginner%27s-Guide-To-Understanding-Convolutional-Neural-Networks/ https://medium.freecodecamp.org/an-intuitive-guide-to-convolutional-neural-networks-260c2de0a050
- 35. INPUT Feature Time Series Convolutional Layer 32 Filters 300 Nodes 150 60 Output Fully Connected Layers Pooling Layer Subsample ...... - CNNs can learn local connectivity between inputs. - CNNs are able to learn user behavior directly from the raw time series. CONVOLUTIONAL NEURAL NETWORKS (CNNs)
- 36. YOKOZUNA DATA PEER-REVIEWED ARTICLES CHURN PREDICTION IN MOBILE SOCIAL GAMES: TOWARDS A COMPLETE ASSESSMENT USING SURVIVAL ENSEMBLES A. Periáñez, A. Saas, A. Guitart and C. Magne IEEE DSAA 2016 Montreal DISCOVERING PLAYING PATTERNS: TIME SERIES CLUSTERING OF FREE-TO-PLAY GAME DATA A. Guitart, A. Periáñez and A. Saas, IEEE CIG 2016 Santorini GAMES AND BIG DATA: A SCALABLE MULTI-DIMENSIONAL CHURN PREDICTION MODEL P. Bertens, A. Guitart and A. Periáñez, A. Saas, IEEE CIG 2017 New York FORECASTING PLAYER BEHAVIORAL DATA AND SIMULATING IN-GAME EVENTS A. Guitart, P. Chen, P. Bertens and A. Periáñez IEEE FICC 2018 Singapore THE WINNING SOLUTION TO THE IEEE CIG 2017 GAME DATA MINING COMPETITION A. Guitart, P. Chen, A. Periáñez A MACHINE-LEARNING ITEM RECOMMENDATION SYSTEM FOR VIDEO GAMES P. Chen, A. Guitart, P. Bertens, A. Periáñez IEEE CIG 2018 Maastricht CUSTOMER LIFETIME VALUE IN VIDEO GAMES USING DEEP LEARNING AND PARAMETRIC MODELS P. Chen, A. Guitart, A. Fernández del Río and A. Periáñez IEEE Big Data 2018 Seattle
- 37. THE WINNING SOLUTION TO THE IEEE CIG 2017 GAME DATA MINING COMPETITION
- 38. Number of registrants in the competition 2 tracks: Which players will leave the game 264 Teams When they will leave the game
- 39. Results Rank YokozunaData (Japan) UTU (Finland) TripleS (Korea) TheCowKing goedleio 0.610098 0.60326 0.57968 0.59370 0.57717 0.63326 0.60370 0.62459 0.60718 0.56205 0.62145 0.60348 0.60130 0.60036 0.58882 Team Test1 Score Test2 Score Total Score Rank 1 2 3 4 5 YokozunaData (Japan) IISLABSKKU UTU (Finland) TripleS (Korea) DTND 0.883248 1.034321 0.927712 0.958308 1.032688 0.616499 0.679214 0.898471 0.891106 0.930417 0.726151 0.819972 0.912857 0.923486 0.978888 Team Test1 Score Test2 Score Total Score Track 1 Which players will leave the game Track 2 When they will leave the game
- 40. CONDITIONAL INFERENCE SURVIVAL ENSEMBLES TWO STEPS ALGORITHM: 1) The optimal split variable is selected: association between covariates and response 2) The optimal split point is determined by comparing two-sample linear statistics for all possible partitions of the split variable RANDOM SURVIVAL FOREST RSF is based on original random forest algorithm RSF favors variables with many possible split points over variables with fewer 4) Ishwaran H. et. al, 2008. Random Survival Forests. 5) Breiman L. et. al, 2001. Random Forests. 4 5
- 41. CENSORED DATA PROBLEM RESULTS Predicted survival curves as a function of playtime (hours), level and days of existing and new players
- 42. ITEM 1 ITEM 2 ITEM 3 SAMPLING TO HANDLE MULTI-LABEL OUTPUTS Players make multiple purchases - Multiple prediction targets Subsample until time t and find next purchase after t - Single label to train on - Enable multiple subsamples to enlarge training set - Reduce overfitting ITEM RECOMMENDATION MODEL
- 43. GAMES WITH LARGE NUMBER OF ITEMS High-dimensional item space Dimensionality reduction to train ML models Operational in-game item recommendation to individual players MACHINE LEARNING MODELS
- 44. AN OPERATIONAL PREDICTION SYSTEM BIG DATA ENGINEERING INFRASTRUCTURE
- 45. The Problem The Solution Supporting data from thousands of games and millions of Monthly Active Users (MAU) A CLOUD DISTRIBUTED-SYSTEM DESIGN FOR: - Data upload - Databases and storage - Parallel computing for data processing and machine learning execution SCALING TO INFINITY
- 46. UPLOAD COMPUTE RESULTS RESULTS DATA UPLOAD Servers/ Storage Compute Instances Cloud Storage Computing Servers Web Servers Database Cluster Cluster KUBERNETES
- 47. CLIENT Web Server + Engine m5.4xlarge DB Instance db.r4.2xlarge X3 nodes WEB APP SERVICE DB CLUSTER YOKOZUNA DATA ARCHITECTURE ON AWS VPC REGION: ASIA PACIFIC TOKYO S3 CASSANDRA SERVICE Cassandra Servers + Engine m5.2xlarge M Storage Storage backup
- 48. The biggest challenge is to be able to retain users with very diverse tastes, desires and motivations - Profile players at the individual level and direct them towards the activities that are more likely to interest them - Tailor game events, pricing and promotions - Identify VIP players sooner to provide a premium service 1 SUMMARY
- 49. SUMMARY Yokozuna Data provide accurate predictions of: - Individual-user actions - Personalized promotions and rewards (recommendations) - Moment and level at which players will leave the game - Money they will spend in the game and their potential 2
- 50. CONTACT aperianez@yokozunadata.com linkedin.com/in/africaperianez @yokozunadata www.yokozunadata.com
- 51. THANK YOU! :)