Automated knowledge graph populator for data selection

1 . A method of training a graph neural network to classify documents stored in a document datastore based on a document query, the method comprising:

generating, for each document, feature embeddings by at least applying an embeddings machine learning model to one or more elements of the document;

recording, for each document, the feature embeddings of the document as a corresponding document node of a knowledge graph, the knowledge graph including a set of document nodes corresponding to the documents, each document node corresponding to a respective document of the documents;

connecting each document node of the knowledge graph with one or more engagement edges based on engagement telemetry data indicating a measure of user access with the documents stored in the document datastore, wherein the engagement telemetry data includes a number of times two documents corresponding to document nodes connected by a corresponding engagement edge have been accessed by a same entity;

training the graph neural network using the knowledge graph populated with each document node and the one or more engagement edges;

generating, by a computing device, a feature embedding for the document query; and

classifying, by the computing device, one or more documents from the document datastore as relevant to the document query by executing the graph neural network on the feature embedding of the document query into the graph neural network, after training the graph neural network using the knowledge graph.

2 . The method of claim 1 , wherein recording the feature embeddings for each document comprises:

extracting elements of the document, each element being associated with a format of different modalities;

generating a feature embedding for each element of each document; and

fusing the feature embeddings for each element of the document into a weighted average feature embedding for the document, wherein feature embeddings corresponding to the different modalities in the document are weighted with fusion weights.

3 . The method of claim 2 , wherein the elements extracted from the document include elements in different modalities.

4 . The method of claim 3 , wherein recording the feature embeddings further comprises:

connecting the feature embeddings for the documents into a machine learning model with a fully connected layer.

5 . The method of claim 4 , further comprising:

training weights of the fully connected layer and the fusion weights for each document based on cross entropy loss to yield a trained document embedding machine learning model that generates the feature embeddings of the document query.

6 . The method of claim 1 , further comprising:

classifying one or more documents from the document datastore as relevant to the document query using the graph neural network; and

presenting the one or more documents as relevant to the document query.

7 . A computing system for training a graph neural network to classify documents stored in a document datastore based on a document query, the computing system comprising:

one or more hardware processors;

a document element extractor executable by the one or more hardware processors and configured to generate, for each document, feature embeddings by at least applying an embeddings machine learning model to one or more elements of the document;

a knowledge graph populator executable by the one or more hardware processors and configured to record, for each document the feature embeddings of the document as a document node of a knowledge graph and to connect each document node of the knowledge graph with one or more engagement edges based on engagement telemetry data indicating a measure of engagement with the documents stored in the document datastore, wherein the engagement telemetry data includes a number of times two documents corresponding to document nodes connected by a corresponding engagement edge have been accessed by a same entity, the knowledge graph including a set of document nodes corresponding to the documents, each document node corresponding to a respective document of the documents;

a graph-based model trainer executable by the one or more hardware processors and configured to train the graph neural network using the knowledge graph populated with each document node and the one or more engagement edges;

an embeddings machine learning model executable by the one or more hardware processors and configured to generate a feature embedding for the document query; and

a document selector executable by the one or more hardware processors and configured to classify one or more documents from the document datastore as relevant to the document query by executing the graph neural network on the feature embedding of the document query into the graph neural network, after training the graph neural network using the knowledge graph.

8 . The computing system of claim 7 , wherein the knowledge graph populator is further configured to extract elements of each document, each element being associated with a format of different modalities, to generate a feature embedding for each element of each document, and to fuse the feature embeddings for each element of a corresponding document into a weighted average feature embedding for the corresponding document, wherein feature embeddings corresponding to the different modalities in the corresponding document are weighted with fusion weights.

9 . The computing system of claim 8 , wherein the elements extracted from at least one of the documents include elements in different modalities.

10 . The computing system of claim 9 , wherein the knowledge graph populator is further configured to connect the feature embeddings for documents into a machine learning model with a fully connected layer.

11 . The computing system of claim 10 , further comprising:

a model trainer executable by the one or more hardware processors and further configured to train weights of the fully connected layer and the fusion weights for each document based on cross entropy loss to yield a trained document embedding machine learning model that generates the feature embeddings of the document query.

12 . The computing system of claim 7 , wherein the engagement telemetry data includes a number of times two documents connected by a corresponding edge have been accessed by a same entity.

13 . One or more tangible processor-readable storage media embodied with instructions for executing on one or more processors and circuits of a computing device a process of training a graph neural network to classify documents stored in a document datastore based on a document query, the process comprising:

generate, for each document, feature embeddings by at least applying an embeddings machine learning model to one or more elements of the document;

recording, for each document, the feature embeddings of the document as a document node of a knowledge graph, the knowledge graph including a set of document nodes corresponding to the documents, each document node corresponding to a respective document of the documents;

connecting each document node of the knowledge graph with one or more engagement edges based on engagement telemetry data indicating a measure of engagement with the documents stored in the document datastore, wherein the engagement telemetry data includes a number of times two documents corresponding to document nodes connected by a corresponding engagement edge have been accessed by a same entity;

training the graph neural network using the knowledge graph populated with each document node and the one or more engagement edges,

generating, by a computing device, a feature embedding for the document query; and

classifying, by the computing device, one or more documents from the document datastore as relevant to the document query by executing the graph neural network on the feature embedding of the document query into the graph neural network, after training the graph neural network using the knowledge graph.

14 . The one or more tangible processor-readable storage media of claim 13 , wherein recording the feature embeddings for each document comprises:

extracting elements of the document, each element being associated with a format of different modalities;

generating a feature embedding for each element of the document; and

fusing the feature embeddings for each element of the document into a weighted average feature embedding for the document, wherein feature embeddings corresponding to the different modalities in the document are weighted with fusion weights.

15 . The one or more tangible processor-readable storage media of claim 14 , wherein the elements extracted from the document include elements in different modalities.

16 . The one or more tangible processor-readable storage media of claim 15 , wherein recording the feature embeddings for each document further comprises:

connecting the feature embeddings for the document into a machine learning model with a fully connected layer.

17 . The one or more tangible processor-readable storage media of claim 16 , wherein the process further comprises:

training weights of the fully connected layer and the fusion weights for each document based on cross entropy loss to yield a trained document embedding machine learning model that generates the feature embeddings of the document query.

18 . The method of claim 1 , wherein the engagement telemetry data are recorded in the one or more engagement edges.

19 . The method of claim 1 , wherein at least one of the one or more engagement edges connects between two nodes, wherein the measure of user access is a measure of user access of the two documents by a same user.

20 . The method of claim 1 , wherein at least one of the one or more engagement edges connects a single node, wherein the measure of access is a measure of how many times a document corresponding to the node has been accessed.