Method, non-transitory computer-readable medium, and system for inferring user brand sensitivity using a machine learning model

1 . A computer-implemented method, comprising, at a computer system comprising a processor and a computer-readable medium:

receiving, from a customer, a selection of an item that is associated with a first brand;

extracting a first set of features associated with the customer and a second set of features associated with the item;

inputting the first set of features and the second set of features to a brand-association machine learning model that is trained to predict a degree of association between the customer and the first brand associated with the item, the degree of association being specific to the customer, wherein the brand-association machine learning model is trained using a pairwise learning-to-rank training that comprises;

obtaining past replacement records each comprising a customer choice when offered a replacement item;

generating training samples by comparing a first brand associated with an original item and a second brand associated with a replacement item;

assigning higher weights to training samples where the customer selected a replacement item associated with the same brand as the original item or requested a refund, and assigning lower weights to training samples where the customer selected a replacement item associated with a different brand; and

minimizing inversions between predicted brand preferences and actual customer choices by adjusting model weights through backpropagation based on the weighted training samples;

identifying a plurality of candidate alternatives for replacing the item, the plurality of candidate alternatives comprising a first alternative that is associated with the first brand and a second alternative that is associated with a second brand different from the first brand;

predicting, using an availability machine-learning model that is trained to output an availability score for an item at a physical store location, a plurality of availability scores that respectively indicates a likelihood that a candidate alternative in the plurality of candidate alternatives is in stock at the physical store location;

computing composite replacement scores that are functions of the availability scores predicted by the availability machine learning model and the degree of association predicted by the brand-association machine learning model; and

selecting, based on the degree of association between the customer and the first brand, one or more candidate alternatives to be presented to the customer to replace the item.

2 . The computer-implemented method of claim 1 , wherein training of the brand-association machine learning model comprises:

classifying the pairwise learning-to-rank training as at least two categories, a first category associated past replacement records indicating that a brand switch occurred and a second category associated past replacement records indicating a brand switch did not occur; and

predicting the degree of association using the two categories.

3 . The computer-implemented method of claim 1 , wherein training of the brand-association machine learning model comprises:

receiving the pairwise learning-to-rank training that comprises a plurality of past replacement records, each past replacement records assigned with labels indicating actions taken by the customer when offered replacements;

using, in a forward propagation, the brand-association machine learning model to generate predicted actions taken by the customer in the past replacement records;

comparing the predicted actions with the labels of the past replacement records; and

adjusting, in a backpropagation, weights of the brand-association machine learning model based on the comparing.

4 . The computer-implemented method of claim 1 , wherein the pairwise learning-to-rank training comprises a first past replacement record where the customer picked another item from the first brand when offered a replacement, a second past replacement record where the customer asked for a refund when offered a replacement, and a third past replacement record where the customer replaced an item associated with the first brand with an item associated with another brand.

5 . The computer-implemented method of claim 1 , wherein the first set of features is selected from one or more of: brand search history of the customer, chat history of the customer that mentioned any brands, profile of the customer, or past transaction records of the customer.

6 . The computer-implemented method of claim 1 , wherein the second set of features is selected from one or more of: type of product, customers who purchased the item, features of the first brand, or features of other brands that offer competitive products.

7 . The computer-implemented method of claim 1 , further comprising:

conducting a comparison between the degree of association and a threshold; and

responsive to the comparison indicating that the degree of association is higher than the threshold, selecting the first alternative that is associated with the first brand to be presented to the customer.

8 . The computer-implemented method of claim 1 , wherein identifying the plurality of candidate alternatives comprises:

retrieving a taxonomy of products, wherein the item is defined as a type of product under the taxonomy; and

identifying the candidate alternatives under the type of product defined under the taxonomy.

9 . The computer-implemented method of claim 1 , wherein one of the candidate alternatives selected to be presented to the customer is a sponsored item.

10 . The computer-implemented method of claim 1 , wherein selecting one or more candidate alternatives to be presented to the customer to replace the item is in response to determining that the item is unavailable.

11 . A non-transitory computer-readable medium configured to store code comprising instructions, the instructions, when executed by a processor, cause the processor to:

receive, from a customer, a selection of an item that is associated with a first brand;

extract a first set of features associated with the customer and a second set of features associated with the item;

input the first set of features and the second set of features to a brand-association machine learning model that is trained to predict a degree of association between the customer and the first brand associated with the item, the degree of association being specific to the customer, wherein the brand-association machine learning model is trained using pairwise learning-to-rank training that comprises:

obtaining past replacement records each comprising a customer choice when offered a replacement item;

generating training samples by comparing a first brand associated with an original item and a second brand associated with a replacement item;

assigning higher weights to training samples where the customer selected a replacement item associated with the same brand as the original item or requested a refund, and assigning lower weights to training samples where the customer selected a replacement item associated with a different brand; and

minimizing inversions between predicted brand preferences and actual customer choices by adjusting model weights through backpropagation based on the weighted training samples;

identify a plurality of candidate alternatives for replacing the item, the plurality of candidate alternatives comprising a first alternative that is associated with the first brand and a second alternative that is associated with a second brand different from the first brand;

predict, using an availability machine-learning model that is trained to output an availability score for an item at a physical store location, a plurality of availability scores that respectively indicates a likelihood that a candidate alternative in the plurality of candidate alternatives is in stock at the physical store location;

compute composite replacement scores that are functions of the availability scores predicted by the availability machine learning model and the degree of association predicted by the brand-association machine learning model; and

select, based on the degree of association between the customer and the first brand, one or more candidate alternatives to be presented to the customer to replace the item.

12 . The non-transitory computer-readable medium of claim 11 , wherein training of the brand-association machine learning model comprises:

classifying the pairwise learning-to-rank training as at least two categories, a first category associated past replacement records indicating that a brand switch occurred and a second category associated past replacement records indicating a brand switch did not occur; and

predicting the degree of association using the two categories.

13 . The non-transitory computer-readable medium of claim 11 , wherein training of the brand-association machine learning model comprises:

receiving the pairwise learning-to-rank training that comprises a plurality of past replacement records, each past replacement records assigned with labels indicating actions taken by the customer when offered replacements;

using, in a forward propagation, the brand-association machine learning model to generate predicted actions taken by the customer in the past replacement records;

comparing the predicted actions with the labels of the past replacement records; and

adjusting, in a backpropagation, weights of the brand-association machine learning model based on the comparing.

14 . The non-transitory computer-readable medium of claim 11 , wherein the pairwise learning-to-rank training comprises a first past replacement record where the customer picked another item from the first brand when offered a replacement, a second past replacement record where the customer asked for a refund when offered a replacement, and a third past replacement record where the customer replaced an item associated with the first brand with an item associated with another brand.

15 . The non-transitory computer-readable medium of claim 11 , wherein the first set of features is selected from one or more of: brand search history of the customer, chat history of the customer that mentioned any brands, profile of the customer, or past transaction records of the customer.

16 . The non-transitory computer-readable medium of claim 11 , wherein the second set of features is selected from one or more of: type of product, customers who purchased the item, features of the first brand, or features of other brands that offer competitive products.

17 . The non-transitory computer-readable medium of claim 11 , wherein identifying the plurality of candidate alternatives comprises:

retrieving a taxonomy of products, wherein the item is defined as a type of product under the taxonomy; and

identifying the candidate alternatives under the type of product defined under the taxonomy.

18 . The non-transitory computer-readable medium of claim 11 , wherein one of the candidate alternatives selected to be presented to the customer is a sponsored item.

19 . The non-transitory computer-readable medium of claim 11 , wherein selecting one or more candidate alternatives to be presented to the customer to replace the item is in response to determining that the item is unavailable.

20 . A system comprising:

a processor; and

memory configured to store code comprising instructions, wherein the instructions, when executed by the processor, cause the processor to:

receive, from a customer, a selection of an item that is associated with a first brand;

extract a first set of features associated with the customer and a second set of features associated with the item;

input the first set of features and the second set of features to a brand-association machine learning model that is trained to predict a degree of association between the customer and the first brand associated with the item, the degree of association being specific to the customer, wherein the brand-association machine learning model is trained using pairwise learning-to-rank training that comprises:

obtaining past replacement records each comprising a customer choice when offered a replacement item;

generating training samples by comparing a first brand associated with an original item and a second brand associated with a replacement item;

assigning higher weights to training samples where the customer selected a replacement item associated with the same brand as the original item or requested a refund, and assigning lower weights to training samples where the customer selected a replacement item associated with a different brand; and

minimizing inversions between predicted brand preferences and actual customer choices by adjusting model weights through backpropagation based on the weighted training samples;

identify a plurality of candidate alternatives for replacing the item, the plurality of candidate alternatives comprising a first alternative that is associated with the first brand and a second alternative that is associated with a second brand different from the first brand;

predict, using an availability machine-learning model that is trained to output an availability score for an item at a physical store location, a plurality of availability scores that respectively indicates a likelihood that a candidate alternative in the plurality of candidate alternatives is in stock at the physical store location;

compute composite replacement scores that are functions of the availability scores predicted by the availability machine learning model and the degree of association predicted by the brand-association machine learning model; and

select, based on the degree of association between the customer and the first brand, one or more candidate alternatives to be presented to the customer to replace the item.