Supercharging Business Decision with AI: Insights, Optimize and Personalize
1.Supercharging Business Decisions with AI: Ins ights , Optimize and Pers onalize
3.How many Drivers do we need in December 2019 in San Francisco?
4.How far do we plan? Tactic Strategic 4-6 12-18 weeks months
5.How do we know how many people are going to take Uber? Trip Forecasting Historical Trips Historical Trips data for a city Long-Term Forecasts Acquisition Spend Upto 52 Weeks Marketing (acquiring new Time Series Model riders/drivers) Used for year long budget planning Events Big events in the city
6. Time-Series Forecas ting Algorithms Reference https://www.datascience.com/blog/time-series-forecasting-machine-learning-differences
7.Forecasting - Predictive Model Cohort Rider Retention Trips/Active Driver First Trips Trips Rate User Eater Month of joining
8. Forecasting - Bayes ian Model It’s a probabilis tic graphical model that repres ents a s et of variables and their conditional dependencies via a graph (DAG). For example, a Bayes ian network could repres ent the probabilis tic relations hips between dis eas es and s ymptoms . Given s ymptoms , the network can be us ed to compute the probabilities of the pres ence of various dis eas es . Reference https://en.wikipedia.org/wiki/Bayesian_network#/media/File:SimpleBayesNetNodes.svg
9.Forecasting - Black Box Model Cohort Ensemble Bayesian Black Box Output Classical Backtesting
10.Forecasting Models - Neural Net Trip transactions Cost Trip Product Curve Mix Model Model Model Marketing spending Cost Trips Trip Fare Curve Net-Inflow Cost Model Model Model Ensemble Model Model Model GB Driver/Rider signup Service Net-Inflow Cost Retention Trip Fee Model Model Model Model Revenue Model User behavior Incentive Trip UFP Model Model Up/Down Holiday/events Model Planning-as-a-Service Finance Modeling and Computation Platform
11.Number of trips Number of Drivers needed But how do you balance the Market with Drivers & Riders
13. Optimization Process Financial Planning Tool: Previous slides Bi-annual Output: monthly/weekly spend budget, by lever/PU Continuous On-demand Cross Lever Budgeting Budget Setting Regional Growth Leads Tool: Cross Lever Optimizer (CLOe) Finance Teams: Strategic + Regional Finance, Perf Marketing, Central Ops Ops Output: updated weekly spend budget, by lever/PU Incentive Budget Paid Budget Referrals Budget Incentive Spend Paid Spend Referral Spend Tools: Finplan Tools: Mixed media model Tools: Web referrals Output: Weekly EDI/ERI/UFP Output: Weekly marketing Output: City referral structures Spenders spend by city channel spend by city Ops Marketplace Marketing
14.Trip Forecasting & Optimization Historical Trips Long-Term Forecasts Historical Trips data for a city Upto 52 Weeks Time Series Model Acquisition Spend Used for year long budget Marketing (acquiring new planning riders/drivers) Driver Incentives Short-Term Rolling (~ $1Bn) Forecasts 1-12 Weeks Rider Promotions Personalized Model (~ $500m) Adjust budgets in spend levers to achieve trip targets - Represents a spending lever
15.Scenario Planning Scenario Generation Deviation from Forecast Weekly/monthly Incentive planning Which subset of users should we focus integrated with trip forecasting on to meet goal? Combining insights like these can help Uber adjust its budget in the short term.
16.Optimization Model Overview User Level Model Historical FTs Incentive spend Paid + Organic Driver Pe r trip me trics Re fe rral First time Driver RR & TPA fore cast (Fare ) Drive r (FTD) by channe l Trips = FT x RR x TPA Drive r FT model (base d on channe l cost curve s) Trips production GB function Trips = FT x RR x TPA Paid + Organic First time Ride r (FTR) by channe l Legend Re fe rral Rider RR & TPA Model signal/input Cost curve Lever Promo Historical FTs YYY Incentive spend Rider FT model (base d on channe l cost curve s) Rider 16 CLOe helps to optimize growth spend across marketing, referrals and incentives.
17.Trip Model Detail (LSTM) y1 y2 yn y1 y2 yn λ π λ π λ π λ π λ π λ π FC FC FC FC FC FC LSTM 2 LSTM 2 LSTM 2 LSTM 2 LSTM 2 LSTM 2 I_t+1 I_t+2 I_t+n I_t’+1 I_t’+2 I_t’+n LSTM 1 LSTM 1 F_t F_t’ training n = 12 predict n = 12
19.LTV is an estimate of LifeTime contribution of each user in order to drive efficiency in marketing, incentive spend and as a KPI to inform product improvements
20.Model Overview Model Overview: ● We use Gradient Boosting Trees Mode l which consolidate s pre dictions of hundre ds of inde pe nde ntly traine d tre e s. Its an ite rative Mode l and Pre diction syste m. Figure shows the use r le ve l GB Mode l pre diction. ● We are using the Gamma-Gamma BG/NBD mode l to pre dict the ne xt 2-ye ars of rolling gross bookings for e ach use r. Eng Platform: Apache Spark Ecosystem ● PySpark platform to proce ss Pe tabyte s of data ● Combine s Que ry, data frame s and machine le arning Machine Graph ● Ability to acce ss data across Ube r’s data store s: Hive , Spark Streaming Learning Analytics HDFS, Cassandra, and S3. SQL (MLlib) (GraphX) Spark Core API R Python Scala Java
22.Finance Intelligence at Uber Planning Forecasting Budgeting Optimization Lifetime Value Scenario Management Service Analytics Model Orchestrator S Metrics Computation Management APIs Security Model Computation Service Forecasting Models Optimization Models LTV Models Finance Data Warehouse Data Pipelines Metrics Store Dashboards
23.Data Platform Overview
24.Financial Data Store (FDS)
26.Looking forward to.. ● Uber Freight ● Uber Health ● Drones (Food Delivery) ● Uber Elevate (Air Transportation) ● Autonomous Vehicles ● Facilitate better Transportation
27.Proprietary and confidential © 2018 Uber Technologies, Inc. All rights reserved. No part of this docume nt may be re produce d or utilize d in any form or by any me ans, e le ctronic or me chanical, including photocopying, re cording, or by any information storage or re trie val syste ms, without pe rmission in writing from Ube r. This docume nt is inte nde d only for the use of the individual or e ntity to whom it is addre sse d and contains information that is privile ge d, confide ntial or othe rwise e xe mpt from disclosure unde r applicable law. All re cipie nts of this docume nt are notifie d that the information containe d he re in include s proprie tary and confide ntial information of Ube r, and re cipie nt may not make use of, disse minate , or in any way disclose this docume nt or any of the e nclose d information to any pe rson othe r than e mploye e s of addre sse e to the e xte nt ne ce ssary for consultations with authorize d pe rsonne l of Ube r.
28.Business Facts ● X Ride sharing cities, YUberEats cities ● $zz bn gross bookings (excluding Uber Eats) ● ??M+ active riders, ?M+ active drivers ● ??M+ trips/day Goal: Enable intelligent models to make data-driven financial decisions faster and more accurately.
WiredTiger B-Tree and WiredTiger In-Memory
18/12 - SASI, Cassandra on the full text search ride - Apache Bi
19_08 - CassandraTokenManagement
16_07 - cassandrasummit2016-runningcassandraonapachemesosacrossm
19_10 - strapdata_Elassandra_cassandra_es