Consumer Location
Agenda
1 Product delivery lifecycle
2 Time to delivery vs Customer experience
3 Challenges and data driven decisions
Event based data logging
4
5 Analytics and machine learning model
6 Final recommendations
Customer - Busy bee
Product delivery lifecycle
Order
1st Mile
Cooking time
Last Mile
Customer Jobs to be done
• Order Status
• Estimated time of delivery (ETOD)
• Track the runner
Runner Restaurant Order ready
Order fulfilled
ETOD
1
Runner
arrives
Runner picks
up order
Time to delivery vs Customer experience
Time to delivery
Customer
Experience
Fast delivery of quickly cooked food
Hot but not with love
Slow delivery of bad food
Why am I paying so much and waiting?
Fast delivery of well cooked food
Hot food delivered with love
Slow delivery of well cooked food
My food is here finally! It tastes great too
2
Challenges and data driven decisions
Availability of runners in
near the restaurant and
customer
Inaccurate estimation
of restaurant food
preparation time
Challenges Outcome Potential solutions
Route inspection problem is
simulated for different
order/ restaurant / runner
scenarios for best practices
Poor runner availability
leads to longer ETOD and
customer dissatisfaction
Poor customer experience
due to poor restaurant
performance and waste of
runner time
Insights from previous orders to
predict location based order
volume and rider utilization
Data driven decisions
Delays due to
externalities like traffic,
how to best
communicate to
customers
Communicate with
restaurants on order load
and incentivize for quicker
delivery to rider
Reward top performers with better
listing on platform and bad
performers with opportunities for
improvement and penalties
External externalities still
create a bad experience
for the customer, reflects
poorly on the platform
Automate notification flow
depending on previous
consumer interactions to
avoid bad customer
feedback
Collect information constantly on
local traffic to flag reduced
restaurant availability in app, divert
more runners to reducing delivery
load in challenging locations
3
Event based data logging for each order
4
Location Dimension
Address ID
City
State
Country
Sales region
Sales Event
Customer ID
Product ID
Address ID
Day ID
Restaurant ID
Quantity
Amount
Order time stamp
Fulfilment time stamp
External Dimension
Day ID
Month
Year
Traffic status
Sales region
Restaurant
Dimension
Day ID
Restaurant ID
Quantity
Order load
Prep time
Sales Volume
Cooking time
estimate
Time Dimension
Day ID
Month
Year
Time stamp ID
Runner Dimension
Product ID
Restaurant ID
Restaurant Arrival
time
Order load
Restaurant Exit time
Runner idle/ waiting
time
Fulfilment time
stamp
Analytics and machine learning models
Features and dimensions
2. Context based learning
(Runner time stamp analysis, Extenal metadata – Delay and route optimization)
1. Collaborative learning
(User / Order similarity, higher data density, rating normalization to de-bias)
Order
Dimension
Restaurant
Dimension
Runner
Dimension
External
Dimension
User
persona
Order
Volume
Time stamp
at restaurant
Berlin based APIs
Transport
mode
X Idle / Waiting
time
X
Order time
stamp
X Prevailing
traffic
conditions
X Historical delivery
estimate based
weights
Historical cooking
efficiency
X Cooking
time estimate
X
Sorted by time and
frequency
X Itemized
Order history
X
Current
Order
Prenzlauerberg
ETOD and live
tracking
Objective
Predict ETOD for
current order
based on previous
customer order,
fulfillement history,
traffic data and
restaurant load
and runner
availability
How do we make sequential
decisions under uncertainty?
Try a Multi- task reinforcement
learning model using quadratic
convergence, from order heat maps,
rider availability and order load
Wedding
Neukölln
Hybrid models
5
Source :chrome-extension://ohfgljdgelakfkefopgklcohadegdpjf/http://papers.neurips.cc/paper/8106-
distributed-multitask-reinforcement-learning-with-quadratic-convergence.pdf
Final recommendations
6
The customer is always unsatisfied and hungry for faster and shorter delivery windows
Providing accurate time estimates in a set if sequential decisions plagued by uncertainty is a doable task
Providing clear communication about order status and time of arrival enforces good service and feedback
Having insight into region specific challenges, runner and restaurant behaviour helps make better predictions
Incentivizing runners to follow a suggested path and having step by step plan is key to last mile fulfilment
Tracking of customer behaviours and ordering patterns is a good starting point to model for order loads

DeliveryHero _ Customer Location_old 2.pdf

  • 1.
  • 2.
    Agenda 1 Product deliverylifecycle 2 Time to delivery vs Customer experience 3 Challenges and data driven decisions Event based data logging 4 5 Analytics and machine learning model 6 Final recommendations
  • 3.
    Customer - Busybee Product delivery lifecycle Order 1st Mile Cooking time Last Mile Customer Jobs to be done • Order Status • Estimated time of delivery (ETOD) • Track the runner Runner Restaurant Order ready Order fulfilled ETOD 1 Runner arrives Runner picks up order
  • 4.
    Time to deliveryvs Customer experience Time to delivery Customer Experience Fast delivery of quickly cooked food Hot but not with love Slow delivery of bad food Why am I paying so much and waiting? Fast delivery of well cooked food Hot food delivered with love Slow delivery of well cooked food My food is here finally! It tastes great too 2
  • 5.
    Challenges and datadriven decisions Availability of runners in near the restaurant and customer Inaccurate estimation of restaurant food preparation time Challenges Outcome Potential solutions Route inspection problem is simulated for different order/ restaurant / runner scenarios for best practices Poor runner availability leads to longer ETOD and customer dissatisfaction Poor customer experience due to poor restaurant performance and waste of runner time Insights from previous orders to predict location based order volume and rider utilization Data driven decisions Delays due to externalities like traffic, how to best communicate to customers Communicate with restaurants on order load and incentivize for quicker delivery to rider Reward top performers with better listing on platform and bad performers with opportunities for improvement and penalties External externalities still create a bad experience for the customer, reflects poorly on the platform Automate notification flow depending on previous consumer interactions to avoid bad customer feedback Collect information constantly on local traffic to flag reduced restaurant availability in app, divert more runners to reducing delivery load in challenging locations 3
  • 6.
    Event based datalogging for each order 4 Location Dimension Address ID City State Country Sales region Sales Event Customer ID Product ID Address ID Day ID Restaurant ID Quantity Amount Order time stamp Fulfilment time stamp External Dimension Day ID Month Year Traffic status Sales region Restaurant Dimension Day ID Restaurant ID Quantity Order load Prep time Sales Volume Cooking time estimate Time Dimension Day ID Month Year Time stamp ID Runner Dimension Product ID Restaurant ID Restaurant Arrival time Order load Restaurant Exit time Runner idle/ waiting time Fulfilment time stamp
  • 7.
    Analytics and machinelearning models Features and dimensions 2. Context based learning (Runner time stamp analysis, Extenal metadata – Delay and route optimization) 1. Collaborative learning (User / Order similarity, higher data density, rating normalization to de-bias) Order Dimension Restaurant Dimension Runner Dimension External Dimension User persona Order Volume Time stamp at restaurant Berlin based APIs Transport mode X Idle / Waiting time X Order time stamp X Prevailing traffic conditions X Historical delivery estimate based weights Historical cooking efficiency X Cooking time estimate X Sorted by time and frequency X Itemized Order history X Current Order Prenzlauerberg ETOD and live tracking Objective Predict ETOD for current order based on previous customer order, fulfillement history, traffic data and restaurant load and runner availability How do we make sequential decisions under uncertainty? Try a Multi- task reinforcement learning model using quadratic convergence, from order heat maps, rider availability and order load Wedding Neukölln Hybrid models 5 Source :chrome-extension://ohfgljdgelakfkefopgklcohadegdpjf/http://papers.neurips.cc/paper/8106- distributed-multitask-reinforcement-learning-with-quadratic-convergence.pdf
  • 8.
    Final recommendations 6 The customeris always unsatisfied and hungry for faster and shorter delivery windows Providing accurate time estimates in a set if sequential decisions plagued by uncertainty is a doable task Providing clear communication about order status and time of arrival enforces good service and feedback Having insight into region specific challenges, runner and restaurant behaviour helps make better predictions Incentivizing runners to follow a suggested path and having step by step plan is key to last mile fulfilment Tracking of customer behaviours and ordering patterns is a good starting point to model for order loads