Recommendation Applications and Systems at Electronic Arts Meng Wu - PowerPoint PPT Presentation

Recommendation Applications and Systems at Electronic Arts Meng Wu 1 | John Kolen 1 | Navid Aghdaie 2 | Kazi A. Zaman 2 1 EADP Intelligent Systems 2 EADP Data Platform

IN INSP SPIRE IRE THE THE WORLD RLD TO PLA LAY EA DIGITAL PLATFORM

• Ap Applica cations • Challenges • Recommendation System • Conclusions

Recommendations everywhere Out Outsi side • Games f for y r you • Info f from th the w web Bet Between een • Maps a and m modes f for y r you • Ma Matchmak akin ing Wi Within Recommendations • Dynamic d difficulty ty As • Ob Objecti tives Next Best Activity

Outside • Games for you

Outside • News and info from the web

Between • Maps and modes for you 8 game modes 20+ maps 4 character classes

Game Mode Recommendation For Optimal Possession Team Death Match Player Journey Old and inactive players Team Death Match > Possession Ne New players Team Death Match < Possession Old and Ol nd rec ecent ent players ers Team Death Match = Possession

Between • Matchmaking Fair is not enough Last 3 3 Ou Outcomes Churn Ra Ch Rate DLW | | L LLW | | L LDW | |DDD 2.6% ~ ~ 2 2.7% … … WWW WWW 3. 3.7% … DLL | | L LWL | | L LDL 4.6% ~ ~ 4 4.7% WWL WW 4. 4.9% LLL LLL 5. 5.1%

� Matchmaking for Maximal Entertainment ℳ ∗ = argmin * Pr(𝑞 . 𝑑ℎ𝑣𝑠𝑜𝑡 | 𝒕 . , 𝒕 9 ) + Pr(𝑞 9 𝑑ℎ𝑣𝑠𝑜𝑡 | 𝒕 9 , 𝒕 . ) ℳ < = ,< > ∈ℳ

Within • Dynamic Difficulty Adjustment 25 Churn Rate DIFFICULTY (#TRIALS) Avg. Trials to 20% Completion 20 CHURN RATE 15 10% 10 5 0 0% 25 50 75 100 LEVEL One EA match three game, 06/2016 – 07/2016

Within • Controlling Player Progression with DDA Level 1 Level 2 Level 3 Level 4 Finished 1 2 3 6 Trial 1 4 7 Trial 2 Churn 5 8 Trial 3 Easy Hard 9 Trial t Difficulty assignment for Level/Trial to maximize engagement

Within • Objectives Generate engaging daily/weekly ”To Do” lists • Make three goals • Play Galactic War mode • Apply a vehicle mod Multiarmed bandit to maximize retention

• Applications • Ch Challen enges es • Recommendation System • Conclusion

Challenges • Games • Players • Data

Game Challenges Genres Franchises Extra Content

Player Challenges Motivations Leader Storywriter Competitor Acquisition Or Organ anic Tr Trial Gu Guest Su Subscrib riber

Data Challenges • • Session days Session days • • Duration Duration • • Last player time Last player time • • Veteran Veteran • • Match count Match count • • Skill Skill • • Wins/loses (indiv) Wins/loses (group) • • Goals Kills/deaths • Real Madrid fan

• Applications • Challenges • Recommendation System • Conclusion

Our Strategy • Systems over models • Process over solutions • One architecture to rule them all

Recommendation Architecture Data Game Warehouse Game Servers Telemetry Recommendation Requests Game Client

Outline • Applications • Challenges • Recommendation System • Co Conclusion

Conclusion Conclusion • Many opportunities for recommendation systems in game industry beyond products. • Recommendations as next best activity. • Recommendation systems have an impact at EA • Improved click-through-rate by 80% and player engagement by 10%. • EA studios share a unified recommendation system platform • Machine learning models as first class citizens • Tightly coupled with experimentation

Questions? Join our Intelligent Systems group jkolen@ea.com