Smart Suggestions - Part II

Transcript

1. DRIVER REASSIGNMENT, SMART SUGGEST AND
   NAVIGATION FOR LAST-MILE SHIPPING EFFICIENCY
   Systematic analysis of order demands
   and driver efficiency analysis
   Final Presentation – Sricharan Chiruvolu – BL.EN.U4CSE12505
   

   View Slide

2. Review…
   • About Runnr.
   • Reassignments of orders.
   • Smart Suggestions to drivers and operations.
   • Driver Navigation.
   • In this presentation,
   • Demand prediction and driver efficiency analysis.
   – Mathematical models used.
   – Analytics and machine learning.
   

   View Slide

3. DEMAND AND SUPPLY PREDICTION
   

   View Slide

4. Machine Learning
   • What is machine learning?
   • Supervised Learning.
   • Un-supervised Learning.
   • Reinforcement Learning.
   • Prediction of demand and supply.
   

   View Slide

5. Demand pre-processing
   • Calculation of demand and creating mappings.
   • Generating of a mathematical model.
   • Time-series model with trend detection.
   • Other approaches,
   – Blending: Use the prediction/actual demand gap
   for the previous hour to correct the prediction for
   next hour
   – Predict demand for smaller time
   intervals/areas(cluster rather than localities)
   

   View Slide

6. Demand Knowledge
   • Weekend/holiday demand is much higher than
   weekday demand.
   • Demand profile(demand in a time interval as a
   fraction of daily demand) is roughly same even if
   orders/day can fluctuate.
   • Demand in a locality is independent of demand in
   other localities. In essence we need a `different
   model' for each locality.
   • Depending on the maturity of a locality it can have
   a trend (mostly upwards since we are growing!)
   

   View Slide

7. Generating the demand model
   • Standard Time-series with trend detection.
   • Arima seasonal model with 7 day series.
   

   View Slide

8. Demand model anomalies
   • Arima models need a lot of data to fit and with the
   hyper growth of roadrunnr, using data that is older
   than 30 days old is not advisable.
   • Demand is also affected by supply shocks such as
   driver incentive changes and hence going too far back
   in time leads to too much variation in data that cannot
   be explained by standard arima.
   • Holidays are entirely non seasonal demand bursts
   which also complicates model building and hence we
   defer building that model and instead build a simpler
   model that is entirely different for weekends/holidays
   and weekdays.
   

   View Slide

9. Model generation on weekdays
   • For each locality
   – Pull 30 day historical orders
   – Filter weekend/holiday data
   – For each day calculate demand profile(order
   count in an hour/total orders in a day).
   – Calculate avg demand profile for the locality
   – Calculate total orders for each day
   – Detect a trend in orders/day time series using
   regression data at 70% confidence.
   

   View Slide

10. Model prediction on weekdays
    • Use the regression data per locality to predict
    orders/day.
    • Use the average demand profile and the
    orders/day prediction to get interval demand.
    

    View Slide

11. View Slide

12. View Slide

13. DRIVER EFFICIENCY ANALYSIS
    

    View Slide

14. Measuring efficiency
    • Touchpoints per busy hour (TPBH): TP/ time
    spent on those touchpoints.
    • Touchpoints per available hour (TPAH): TP/
    hours logged in.
    • Trip Time: ​Pickup Time + Wait Time + Drop
    Time.
    • Ideal Pickup Time: 10 mins.
    • Ideal Wait Time​: 1 mins.
    

    View Slide

15. Measuring efficiency
    • Ideal Trip Time: ​Ideal PickupTime + Ideal
    WaitTime + Ideal Drop Off Time
    • Ideal Drop off time: Ideal Travel Time +
    Handshake Time.
    – Ideal Travel Time = (Gmaps Distance)/(12Km/hr)
    – Handshake Time = 5 min.
    

    View Slide

16. TPBH Improvements
    Metric Value % Improvement over
    current
    Current TPBH 2.94 0
    With Ideal pickup time 2.97 1%
    With Ideal wait time 3.32 13%
    With Ideal drop off time 3.82 30%
    With Ideal trip time 5.16 75%
    

    View Slide

17. Pickup leg analytics
    • Pickup leg is defined as the time from which the
    driver receives the order to the time he/she
    reaches the shop.
    • In 67% ​of trips, drivers click the reached shop
    button in their app before reaching the
    merchant’s location. This way, we grossly
    underestimate the number of pickup breaches in
    the system and wrongly calculate the wait time
    for merchants. Drivers need to be trained to
    properly use the system.
    

    View Slide

18. Pickup bleaches
    Factors Contribution to pickup bleaches
    On road distance 60%
    Hour of day 20%
    Driver relevance 11%
    Cross locality pickup 9%
    • Driver Relevance is the number of times the driver has been to the merchant
    before this trip.
    

    View Slide

19. Driver relevance with pickup bleaches
    

    View Slide

20. Pickup leg recommendations
    • Multiple Locality Assignment of Drivers​: The
    current assignment logic gives preference to
    locality over distance when comparing with a
    cross locality driver.
    • This increases the overall distance and as a
    result pickup breaches. If we take three top
    localities of the driver then effective distance
    (based on current assignment logic) will
    decrease.
    

    View Slide

21. Pickup leg recommendations
    • Reducing Assignment Distance: ​Reducing
    assignment distance from 2.5 Kms should
    decrease pickup breaches. However, this need
    to evaluated carefully with stockout events at
    locality level.
    

    View Slide

22. Wait leg analysis
    • Wait Time leg is defined as the time from
    when the driver has reached the shop to
    package shipment.
    • One analysis featured around charging for
    wait times to merchants: Let’s say we charge
    1 Re/min for any minute above a certain
    threshold. Here is the breakdown of
    approximate amount of revenue we will earn
    in a day per city:
    

    View Slide

23. Wait time threshold
    Wait Time Threshold Approximate Total Revenue (in Rs.)
    5 min 8- 10 Lac
    10 mins 3 – 4 Lac
    15 mins 1 – 2 Lac
    

    View Slide

24. Wait time recommendations
    • Merchant Charged, Driver Gains​: We charge for wait
    time and trickle that revenue to drivers as incentive.
    The merchant will ensure that reached shop button is
    clicked only in front of him (because of charge) and
    drivers will have incentive to keep waiting.
    • Delayed Dispatch​: We implement a prediction
    algorithm to predict the food preparation time and
    display the expected preparation time to merchant
    while placing the order. Merchant can override that
    time. According to the time, we incorporate delayed
    dispatch in the system to minimize wait times.
    

    View Slide

25. Drop leg analysis
    • Drop time is the time that the driver took to
    drop the shipment to the customer. It
    approximately contributes to 66% of total trip
    time on an average.
    • Travel Distance and traffic is a major factor
    which contributes to the drop off times. Driver
    Relevance and other things are not so much
    important.
    

    View Slide

26. Drop leg recommendations
    • The Merchant books the order and the delivery driver picks
    the shipment. A notification is sent to the customer to track
    the order.
    • Once the customer clicks on the tracking URL, the
    customer’s current location is found and sent to the server.
    The Customer-to-location relation is established.
    • Meanwhile, the delivery driver is moving towards the
    customer sub-locality specified by the merchant.
    • Once the customer location is found, it is sent to delivery
    driver in the next ping request. A GCM or an MQTT message
    would reach the driver and the trip is updated with the
    customer location.
    • The driver is navigated to the customers location.
    

    View Slide

27. Thank you.
    Name C. SRICHARAN
    REG.NO. BL.EN.U4CSE12505
    

    View Slide