Item recommendations using convolutions on weighted graphs

Patent metadata

Field	Value
Publication number	US-11995702-B2
Application number	US-202217850701-A
Country	US
Kind code	B2
Filing date	Jun 27, 2022
Priority date	Feb 1, 2019
Publication date	May 28, 2024
Grant date	May 28, 2024

Abstract

Methods and systems for generating item recommendations are disclosed. One method includes sampling from a weighted node-based graph to generate a sampled graph, wherein sampling includes selecting a plurality of nodes and, for each selected node, one or more node pairs. The selection of the node pairs is based at least in part based on a weight assigned to the node pair in the weighted node-based graph. The method further includes aggregating information from the one or more neighboring nodes into each corresponding node of the plurality of nodes in the sampled graph to generate a vector representation of the sampled graph. The method also includes applying a loss function to the vector representation of the sampled graph to generate a modified vector representation. The modified vector representation is used to generate, in response to identification of an item from an item collection, a selection of one or more recommended items from within the item collection.

First claim

The invention claimed is:

1. A method of generating an item recommendation from among a retail item collection at a computing system including one or more processors, the method comprising:
assigning, via the one or more processors, weights between pairs of nodes of a plurality of nodes each representing a different retail item from the retail item collection, where each node includes embeddings that are generated based on an item image and an item description, and the weights are based on a combination of the item images, the item descriptions, and item selection information;
generating, via the one or more processors, a weighted node-based graph that represents likelihoods that a selection of a retail item associated with a first node is associated with selection of another retail item associated with a second node;
sampling, via the one or more processors, from the weighted node-based graph to generate a sampled graph, wherein sampling includes selecting a plurality of nodes and, for each selected node, one or more node pairs, each node pair of the one or more node pairs including the node and one or more neighboring nodes, the selection of the one or more node pairs being based at least in part based on a weight assigned to the node pair in the weighted node-based graph;
aggregating, via the one or more processors, information from the one or more neighboring nodes into each corresponding node of the plurality of nodes in the sampled graph to generate a vector representation of the sampled graph, wherein aggregating information comprises applying one or more convolutional modules of a graph convolutional network;
applying, via the one or more processors, a loss function to the vector representation of the sampled graph to generate a modified vector representation; and
using the modified vector representation to generate, via the one or more processors, in response to identification of an item from an item collection, a selection of one or more recommended items from within the item collection, the selection of the one or more recommended items being provided to a web server for display via a retail website.

2. The method of claim 1 , further comprising presenting the one or more recommended items within the retail website.

3. The method of claim 1 , wherein the weight assigned to each node pair comprises an edge weight assigned to the node pair.

4. The method of claim 1 , wherein the one or more convolutional modules are selected from among the group consisting of: a mean aggregator, a long-short term memory (LSTM) aggregator, a pooling aggregator, a node to vector aggregator, and a graphical convolutional network.

5. The method of claim 1 , wherein the loss function includes the weight assigned to each node pair in the vector representation.

6. The method of claim 1 , wherein the loss function comprises a graph-based loss function utilizing weights assigned to the node pairs.

7. The method of claim 6 , wherein the loss function applies a stochastic gradient descent to the weights assigned to the node pairs and as a parameter provided to the one or more convolutional modules.

8. The method of claim 1 , wherein generating the weighted node-based graph comprises using a Jaccard index.

9. The method of claim 1 , further comprising generating a display of the selection of the one or more recommended items.

10. The method of claim 1 , wherein sampling from the weighted node-based graph to generate a sampled graph comprises generating probability densities of pairwise cosine similarity for image embeddings, text embeddings, graph-based embeddings, and weighted graph-based embeddings, and selection of the retail item comprises selecting the retail item having a probability greater than a predetermined threshold.

11. A method of providing an item recommendation for a customer of a retail enterprise at a computing system including one or more processors, the method comprising:
receiving, from a customer, a selection of an inventory item at a retail website server;
in response to receiving the selection at the computing system from the retail website server, utilizing, via the one or more processors, a weighted graph-based convolution to generate a set of related inventory items from an item collection offered by the retail enterprise via the retail website, wherein generating the set of related inventory items comprises:
assigning, via the one or more processors, weights between a node corresponding to the inventory item and a plurality of nodes each representing a different retail item from a retail item collection, where the weights are based on a combination of item images, item descriptions, and item selection information;
generating, via the one or more processors, a weighted node-based graph that represents the likelihood of selection of each of the plurality of nodes based the assigned weights; and
determining, via the one or more processors, the set of related inventory items based upon a function of the assigned weights and a sampling degree parameter exceeding a predetermined threshold probability; and

generating a display, at the retail website server, of the set of related inventory items as recommended inventory items for the customer.

12. The method of claim 11 , wherein the weighted node-based graph represents an input item collection including image data and textual data associated with each of a plurality of items.

13. The method of claim 11 , wherein the set of related inventory items is generated from inputs including textual item description, item images, and historical customer purchase behavior.

14. The method of claim 13 , wherein the set of related inventory items is further generated from inputs selected from page views, item selections, and purchases.

15. A system configured to generate item recommendations, the system comprising:
a computing system including a programmable circuit operatively connected to a memory, the memory storing computer-executable instructions which, when executed by the programmable circuit, cause the computing system to perform:
associating a plurality of retail items with a corresponding plurality of nodes;
for each node of the plurality of nodes, selecting a set of neighboring nodes each exceeding a predetermined threshold probability, wherein the probability of each node being a neighboring node is proportional to s(u,v) β , where s(u,v) is a weight of the edge and β is a sampling degree parameter;
sampling from a weighted node-based graph to generate a sampled graph, wherein sampling includes selecting a plurality of nodes and, for each selected node, the set of neighboring nodes;
aggregating information from the set of neighboring nodes into each node of the plurality of nodes in the sampled graph to generate a vector representation of the sampled graph, wherein aggregating information comprises applying one or more convolutional modules of a graph convolutional network;
applying a loss function to the vector representation of the sampled graph to generate a modified vector representation;
using the modified vector representation to generate, in response to identification of an item from the plurality of retail items, a selection of one or more recommended items from within the plurality of retails items; and
providing the selection of one or more recommended items to a retail web server communicatively connected to the computing system for display.

16. The system of claim 15 , further comprising the retail web server.

17. The system of claim 16 , wherein the computing system is configured to receive the identif

Assignees

• Target Brands Inc

Inventors

• Pande Amit
• Ni Kai

Classifications

• G06N3/0442

  characterised by memory or gating, e.g. long short-term memory [LSTM] or gated recurrent units [GRU] · CPC title

• G06N3/0464

  Convolutional networks [CNN, ConvNet] · CPC title

• G06N3/09

  Supervised learning · CPC title

• G06Q30/0631 Primary

  Recommending goods or services · CPC title

• G06F16/24578

  using ranking · CPC title