System, method, and non-transitory computer readable medium for generating recommendations

Patent metadata

Field	Value
Publication number	US-12182845-B2
Application number	US-202117483469-A
Country	US
Kind code	B2
Filing date	Sep 23, 2021
Priority date	Sep 23, 2021
Publication date	Dec 31, 2024
Grant date	Dec 31, 2024

Abstract

A system can include a database and a computing device. The computing device is configured to receive an item recommendation request corresponding to an asset from an analyst device and select a set of item identifiers of a plurality of item identifiers. An associated published timeframe of the selected item identifiers is related to a present timeframe. The computing device is further configured to determine a composite similarity value for each item identifier of the set of item identifiers comparing a similarity of the asset to each item identifier of the set of item identifiers. The computing device is also configured to generate an item recommendation list including each item identifier of the set of item identifiers with a corresponding composite similarity value above a threshold value and transmit the item recommendation list to the analyst device for display.

First claim

What is claimed is:

1. A system, comprising:
a database storing a plurality of item identifiers and a plurality of asset identifiers, each item identifier of the plurality of item identifiers being associated with item data and a published timeframe, and each asset identifier of the plurality of asset identifiers being associated with asset data and a previously published indicator; and
a processor;
a non-transitory memory storing instructions, that when executed, cause the processor to:
receive an item recommendation request corresponding to an asset from a computing device;
select, based on the item recommendation request, a set of item identifiers of the plurality of item identifiers, the published timeframe of each of the selected item identifiers being related to a present timeframe;
receive historical interaction data associated with the asset, wherein the historical interaction data is generated by a click tracking module that records user interactions with at least the asset, and wherein the click tracking module associates the asset with user interactions if a threshold number of user interactions is not exceeded;
generate a respective image score representative of a similarity between an item image associated with the item data and an asset image associated with the asset based on a first trained machine learning model, wherein the first trained machine learning model is based on unsupervised linear discriminant analysis to reduce one or more dimensions across extracted feature vectors associated with the item image and the asset image;
determine a composite similarity value for each item identifier of the set of item identifiers comparing a similarity of the asset to each item identifier of the set of item identifiers, wherein the composite similarity value comprises a harmonic mean of:
an order score representative of a likelihood of interaction with the item data of a corresponding item identifier, wherein the order score comprises a first historical order score when the historical interaction data includes user interactions indicating a relevance between an item identifier of the set of item identifiers and the asset, and wherein the order score comprises a predicted order score when the historical interaction data does not include user interactions indicating a relevance between an item identifier of the set of item identifiers and the asset;
a string score representative of a similarity of the item data of the corresponding item identifier and the asset data of the asset;
a keyword score representative of a similarity between keywords associated with the item data and an asset name associated with the asset; and
the respective image score representative of the similarity between the item image associated with the item data and the asset image associated with the asset,
wherein the string score or the keyword score is generated by a second trained machine learning model utilizing a first labeled training dataset comprising historically identified relevant items for the asset and validation labels;

generate an item recommendation list including each item identifier of the set of item identifiers where the respective composite similarity value is above a threshold value;
transmit instructions to the computing device that cause the computing device to display a user interface including interface elements representative of the item recommendation list;
receive feedback data representative of additional user interactions with the displayed user interface including validation or rejection of at least one item identifier in the item recommendation list;
generate an updated second trained machine learning model based on a second labeled training dataset, wherein the second labeled training dataset is generated by removing items from the first labeled training dataset based on the feedback data, and the first trained machine learning model and the second trained machine learning model include a plurality of processing layers, each processing layer of the plurality of processing layers associated with trainable filters, transformations, projections, hashing, pooling, and regularization; and
generate an updated item recommendation list for display including an updated set of item identifiers based on respectively updated composite similarity values, wherein each of the updated composite similarity values includes at least one updated score generated by the updated second trained machine learning model and an updated order score comprising a second historical order score generated at least in part based on the feedback data, wherein the updated set of item identifiers is different than the set of item identifiers.

2. The system of claim 1 , wherein the instructions cause the processor to, in response to determining the asset is a new asset based on the corresponding previously published indicator:
extract asset features from corresponding asset data;
extract item features for each item identifier of the set of item identifiers from respective item data;
reduce a number of the extracted asset features and item features;
compute the order score for each item identifier of the set of item identifiers using a predictive model based on the reduced number of the extracted asset features and item features; and
reduce the set of item identifiers based on the corresponding order score being less than a determined threshold order value.

3. The system of claim 2 , wherein reducing the number of the extracted asset features and item features includes implementing linear discriminant analysis.

4. The system of claim 2 , wherein reducing the set of item identifiers includes:
applying a beta regression to the computed order scores for each item identifier of the set of item identifiers; and
determining the threshold order value as a pre-defined sample quantile.

5. The system of claim 2 , wherein the predictive model comprises a third trained machine learning model generated, at least in part, by a supervised training processes using the feedback received from the computing device.

6. The system of claim 1 , wherein the string score comprises a semantic similarity between a text description of the asset included in the asset data and a text description of the selected item identifier included in the item data.

7. The system of claim 1 , wherein the keyword score comprises a semantic similarity between keywords of the asset included in the asset data and keywords from a category description of at least one category corresponding to the selected item identifier included in the item data.

8. The system of claim 1 , wherein the instructions cause the processor to, in response to determining the asset is an existing asset based on the corresponding previously published indicator, compute the order score for each item identifier of the set of item identifiers based on historical data related to the asset and the respective item.

9. The system of claim 1 , wherein the second trained machine learning model comprises a global vectors (GloVe) model that determines one of the string score or the keyword score, and wherein the first trained machine learning model comprises an EfficientNet model that determines the image score.

10. The system of claim 1 , wherein each of the order score, the string score, the keyword score, and the image score are normalized prior to determining the harmonic mean.

11. A computer-implemented method, comprising:
receiving an item recommendation request corresponding to an asset from a computing device;
selecting, based on the item recommendation request, a set of item identifiers of a plurality of item identifiers, an associated published timeframe of the selected it

Assignees

• Walmart Apollo Llc

Inventors

• Banerjee Debanjana
• Das Amlan Jyoti
• Kaddevarmuth Srujana

Classifications

• G06F16/9538

  Presentation of query results · CPC title

• G06F16/24578

  using ranking · CPC title

• G06Q30/0633 Primary

  Managing shopping lists, e.g. compiling or processing purchase lists (shipping orders G06Q10/083; order filling G06Q10/087) · CPC title

• G06N5/02

  Knowledge representation; Symbolic representation · CPC title

• G06F16/9535

  Search customisation based on user profiles and personalisation · CPC title