Recreating the State of the Player
summary
Rebuilds complete player profiles by replaying daily event streams, joining game events, store transactions and account snapshots into ordered timelines and snapshotting state so models see the true pre-game features. Claims: 500 billion daily points, 12–20 PB warehouse; tools: Honu ingest to S3, Databricks/Spark RDDs with mapPartitions, recurrent nets for time-to-event; uses Kaplan-Meier and Cox proportional hazards to predict churn and feature store patterns to serve low-latency features. Email test shows non-ranked players respond; ranked players show no lift.
AI-written summary · about
chapters (15)
- 0:00Start
- 0:20Riot Games
- 6:25Player Dynamics
- 8:44Survival Analysis
- 9:35Estimate a Survival Curve
- 10:26Cox Proportional Hazards
- 10:54Formula for Estimating the Hazard Rate
- 20:20Cohort Analysis
- 22:23Data Gathering Phase
- 35:05Takeaways
- 36:52Central Data Store
- 38:52Summary
- 40:21How Do You Deal with the Toxic Community Which Can Be a Reason for some Players To Leave the Game for Good
- 48:25How Do You Determine Which Elements of a Patch Use Causally Impact User Retention When There Are So Many Interaction Events
- 54:10What Are the Data Challenges You'Ve Found in Your Etl Solution
related talks
23mCustom-made games with machine learning and big data
11mBig Data Analytics: Immutability & Idempotence
35mReal Time Marketing with Kafka, Storm, Cassandra & Spark #bbuzz
17mCreating & operating ML models from event-based data using feature stor...
43mUsing Queryable State to update ML Algorithms per datapoint for...
27mBig Data in Mobile Games
56mSpark-driven Audience counting
22mTurn Data Into Insights