Graph-based machine learning model with feature-agnostic training and holistic feature metrics
Various embodiments of the present disclosure describe data evaluation techniques that leverage a graph-based machine learning model to evaluate a knowledge graph. The techniques include using a target graph model to generate a predictive representation for a graph node of a graph training dataset. The techniques include using a feature prediction model to generate predicted feature values for the graph node based on the predictive representation. The techniques include generating a data evaluation score for the graph training dataset based on the predicted feature values. The techniques include using the target graph model to generate a predictive output for the graph node based on the predictive representation and then generating an evaluation output for the target graph model based on the evaluation score and the predictive output.
1 . A computer-implemented method comprising:
generating, by one or more processors and using a target graph model, a predictive representation for a graph node of a graph training dataset;
generating, by the one or more processors and using a feature prediction model, one or more predicted feature values for the graph node based on the predictive representation;
generating, by the one or more processors, a data evaluation score for the graph training dataset based on the one or more predicted feature values, wherein the data evaluation score is based on a graph feature confidence score indicative of a predicted accuracy of the one or more predicted feature values;
generating, by the one or more processors and using the target graph model, a predictive output for the graph node based on the predictive representation, wherein (i) the predictive output comprises a node classification for the graph node, (ii) the one or more predicted feature values correspond to one or more evaluation features of the graph training dataset, and (iii) the target graph model is previously trained to generate an evaluation feature-agnostic predictive representation that at least partially prevents the target graph model from generating the predictive output based on the one or more evaluation features; and
generating, by the one or more processors, an evaluation output for the target graph model based on the data evaluation score and the predictive output.
2 . The computer-implemented method of claim 1 , wherein the target graph model comprises a first graph neural network (GNN) and the feature prediction model comprises a second GNN.
3 . The computer-implemented method of claim 1 , wherein the target graph model and the feature prediction model are jointly trained using a joint objective function.
4 . The computer-implemented method of claim 3 , wherein:
the joint objective function comprises a first objective function and a second objective function,
the first objective function comprises a first optimization function for the target graph model, and
the second objective function comprises a second optimization function for the feature prediction model.
5 . The computer-implemented method of claim 4 , wherein the first objective function comprises a comparison between the predictive output of the target graph model and a ground truth label for the graph node.
6 . The computer-implemented method of claim 1 , wherein the predictive representation comprises a feature embedding that encodes one or more features of the graph node and one or more adjacent features of one or more neighboring nodes of the graph node in the graph training dataset.
7 . The computer-implemented method of claim 6 , wherein the one or more neighboring nodes of the graph node comprise one or more nodes of the graph training dataset that are connected to the graph node by one or more edges.
8 . The computer-implemented method of claim 1 , wherein the graph feature confidence score is based on a predicted feature confidence score indicative of a confidence level for the one or more predicted feature values.
9 . The computer-implemented method of claim 8 , wherein the predicted feature confidence score is generated by the feature prediction model.
10 . A system comprising:
one or more processors; and
at least one memory storing processor-executable instructions that, when executed by the one or more processors, cause the one or more processors to:
generate, using a target graph model, a predictive representation for a graph node of a graph training dataset;
generate, using a feature prediction model, one or more predicted feature values for the graph node based on the predictive representation;
generate data evaluation score for the graph training dataset based on the one or more predicted feature values, wherein the data evaluation score is based on a graph feature confidence score indicative of a predicted accuracy of the one or more predicted feature values;
generate, using the target graph model, a predictive output for the graph node based on the predictive representation, wherein the predictive output comprises a node classification for the graph node, wherein the one or more predicted feature values correspond to one or more evaluation features of the graph training dataset, and wherein the target graph model is previously trained to generate an evaluation feature-agnostic predictive representation that at least partially prevents the target graph model from generating the predictive output based on the one or more evaluation features; and
generate an evaluation output for the target graph model based on the data evaluation score and the predictive output.
11 . The system of claim 10 , wherein the target graph model comprises a first graph neural network (GNN) and the feature prediction model comprises a second GNN.
12 . The system of claim 10 , wherein the target graph model and the feature prediction model are jointly trained using a joint objective function.
13 . The system of claim 12 , wherein:
the joint objective function comprises a first objective function and a second objective function,
the first objective function comprises a first optimization function for the target graph model, and
the second objective function comprises a second optimization function for the feature prediction model.
14 . The system of claim 13 , wherein the first objective function comprises a comparison between the predictive output generated by the target graph model and a ground truth label for the graph node.
15 . The system of claim 10 , wherein the graph feature confidence score is based on a predicted feature confidence score indicative of a confidence level for the one or more predicted feature values.
16 . One or more non-transitory computer-readable storage media including instructions that, when executed by one or more processors, cause the one or more processors to:
generate, using a target graph model, a predictive representation for a graph node of a graph training dataset;
generate, using a feature prediction model, one or more predicted feature values for the graph node based on the predictive representation;
generate data evaluation score for the graph training dataset based on the one or more predicted feature values, wherein the data evaluation score is based on a graph feature confidence score indicative of a predicted accuracy of the one or more predicted feature values;
generate, using the target graph model, a predictive output for the graph node based on the predictive representation, wherein the predictive output comprises a node classification for the graph node, wherein the one or more predicted feature values correspond to one or more evaluation features of the graph training dataset, and wherein the target graph model is previously trained to generate an evaluation feature-agnostic predictive representation that at least partially prevents the target graph model from generating the predictive output based on the one or more evaluation features; and
generate an evaluation output for the target graph model based on the data evaluation score and the predictive output.
17 . The one or more non-transitory computer-readable storage media of claim 16 , wherein the predictive representation comprises a feature embedding that encodes one or more features of the graph node and one or more adjacent features of one or more neighboring nodes of the graph node in the graph training dataset.
18 . The one or more non-transitory computer-readable storage media of claim 16 , wherein the target graph model comprises a first graph neural network (GNN) and the feature prediction model comprises a second GNN.
19 . The one or more non-transitory computer-readable storage media of claim 16 , wherein the target graph model and the feature prediction model are jointly trained using a joint objective function.
20 . The one or more non-transitory computer-readable storage media of claim 19 , wherein:
the joint objective function comprises a first objective function and a second objective function,
the first objective function comprises a first optimization function for the target graph model, and
the second objective function comprises a second optimization function for the feature prediction model.