| CPC G06Q 30/0605 (2013.01) [G06F 18/214 (2023.01); G06N 3/04 (2013.01)] | 20 Claims |
|
1. A computer-implemented method comprising:
at an online system comprising memory and one or more processors:
receiving orders from a plurality of user devices, each order including one or more items and associated with a different physical delivery location;
generating candidate groups of the received orders, each candidate group including one or more of the received orders that have corresponding physical delivery locations that are within a common geographic region;
extracting a set of characteristics for each candidate group from the one or more orders included in the candidate group, an extracted set of characteristics including a value determined from a number of shoppers available in the common geographic region and a number of orders being within the common geographic region;
training a plurality of selection prediction models, each selection prediction model specific to a geographical region and being a machine learning model configured to generate a predicted time for a shopper to select a candidate group for fulfillment, wherein training a selection prediction model comprises:
retrieving a set of training data comprising previous groups of previous orders, each previous group associated with previously-identified characteristics and a label of a length of time between a time when the previous group was created and a time when a shopper selected the previous group for fulfillment,
generating, using the selection prediction model, predicted times for shoppers to select the previous groups,
comparing the predicted times to the labels associated with the previous groups to generate one or more error terms, and
backpropagating the one or more error terms to modify one or more parameters of the selection prediction model;
retraining the plurality of selection prediction models at a plurality of time intervals to account for changes in fulfillment parameters in the specific geographical region, wherein retraining the selection prediction model in one of the time intervals comprises:
receiving additional training examples captured in the one of the time intervals;
splitting the additional training examples into subsets having different values of the fulfillment parameters; and
recursively generating new nodes of the selection prediction model using the subsets to group one or more orders in the subsets together;
accessing a particular selection prediction model from the plurality of selection prediction models, the particular selection model corresponding to the common geographic region, the particular selection prediction model configured to receive the extracted set of characteristics of a given candidate group and output a predicted time for a shopper to select the given candidate group for fulfillment;
determining the predicted time for the shopper to select each candidate group for fulfillment by applying the machine learning model to the extracted set of characteristics corresponding to each candidate group;
determining an estimated time for fulfillment for each candidate group, the estimated time for fulfillment for the group determined from the predicted time for the shopper to select the candidate group;
identifying a candidate group including a plurality of orders;
selecting the identified candidate group as a selected group for evaluation in response to an estimated time for fulfillment of the selected group being within a threshold amount of time from an estimated time for fulfillment of at least one of the plurality of orders included in the selected group;
causing a shopper mobile device to display the selected group of the plurality of orders for the shopper to select for the fulfillment of the selected group as a whole; and
receiving a fulfillment confirmation of the selected group of multiple orders from the shopper mobile device, wherein a total delivery time for the fulfillment of the selected group as a whole is shorter than a total delivery time to fulfill the plurality of orders individually that are associated with different physical delivery locations.
|