Hybrid graph-based prediction machine learning frameworks
Various embodiments of the present invention disclose techniques for determining a graph-based prediction based at least in part on a cross-entity relationship graph data object and using a hybrid graph-based processing machine learning framework. In some embodiments, the hybrid graph-based prediction machine learning framework is configured to generate the graph-based prediction based at least in part on a comprehensive representation of the cross-entity relationship graph data object that is generated based at least in part on output data of a graph convolutional neural machine learning model and an image-based graph convolutional neural network machine learning model.
1 . A computer-implemented method comprising:
processing feature data from a cross-entity relationship graph data object, by one or more processors and using a hybrid graph-based processing machine learning framework that comprises (a) a graph convolutional neural network, (b) an image-based convolutional neural network, and (c) a feedforward neural network, to identify a portion of the cross-entity relationship graph data object for processing by a post-prediction system by:
identifying the cross-entity relationship graph data object associated with a plurality of predictive entities relative to a primary predictive entity, wherein the cross-entity relationship graph data object comprises:
(i) a plurality of entity nodes associated with the plurality of predictive entities,
(ii) an entity node attribute, associated with an entity node of the plurality of entity nodes, that is determined based at least in part on an entity size measure associated with a predictive entity, of the plurality of predictive entities, that is associated with the entity node, and
(iii) a group of directed entity relationship edges, wherein a directed entity relationship edge of the group of directed entity relationship edges is associated with a source entity node and a destination entity node of the plurality of entity nodes,
generating, using the graph convolutional neural network and based at least in part on the cross-entity relationship graph data object, a graph convolutional representation of the cross-entity relationship graph data object,
converting, using a node visualization schema, the cross-entity relationship graph data object to a graph image data object by:
(i) converting the entity node of the cross-entity relationship graph data object to a node icon comprising a defined geographic shape centered in a defined region of the graph image data object, and
(ii) mapping a directed relationship edge of the cross-entity relationship graph data object to an edge line in the graph image data object that stretches from (a) a first node icon corresponding to the source entity node of the directed relationship edge to (b) a second node icon corresponding to the destination entity node of the directed relationship edge,
generating, using the image-based convolutional neural network and based at least in part on the graph image data object of the cross-entity relationship graph data object, an image-based convolutional representation of the cross-entity relationship graph data object,
generating a comprehensive graph representation by concatenating the graph convolutional representation and the image-based convolutional representation, and
processing, using the feedforward neural network, the comprehensive graph representation to identify the portion of the cross-entity relationship graph data object for processing by the post-prediction system based at least in part on the comprehensive graph representation; and
providing, by the one or more processors, data associated with the portion of the cross-entity relationship graph data object to the post-prediction system.
2 . The computer-implemented method of claim 1 , wherein the comprehensive graph representation comprises: (i) one or more extracted relational features determined based at least in part on a related edge subset for the primary predictive entity that comprises the group of directed entity relationship edges whose source predictive entity is the primary predictive entity, (ii) one or more extracted pairwise features determined based at least in part on a related pair subset of a group of entity pairs that comprise the primary predictive entity, and (iii) one or more extracted node features determined based at least in part on a related node subset of the plurality of entity nodes whose entity node attributes satisfy an entity node attribute threshold.
3 . The computer-implemented method of claim 2 , wherein:
(i) the one or more extracted pairwise features comprise D extracted distance features, and
(ii) an extracted distance feature of the D extracted distance features describes a cross-entity distance measure for a corresponding entity pair whose corresponding cross-entity distance measure is among D highest cross-distance entity measures associated with the related pair subset.
4 . The computer-implemented method of claim 2 , wherein:
(i) the one or more extracted relational features comprise C extracted contribution features, and
(ii) an extracted contribution feature of the C extracted contribution features describes a normalized historical contribution measure for a corresponding directed relationship edge whose corresponding normalized historical distribution measure is among top C highest normalized historical contribution measures associated with the related edge subset.
5 . The computer-implemented method of claim 1 , further comprising:
allocating one or more computing entities to the post-prediction system, responsive to a number of a plurality of currently-allocated computing entities not satisfying a number of allowed computing entities; and
deallocating the one or more computing entities from the plurality of currently-allocated computing entities, responsive to the number of the plurality of currently-allocated computing entities exceeding the number of allowed computing entities.
6 . The computer-implemented method of claim 1 , wherein the graph convolutional neural network and the image-based convolutional neural network are trained end to end.
7 . A system comprising:
one or more processors; and
at least one memory storing processor-executable instructions that, when executed by any one or more of the one or more processors, causes the one or more processors to perform operations comprising:
processing feature data from a cross-entity relationship graph data object, using a hybrid graph-based processing machine learning framework that comprises (a) a graph convolutional neural network, (b) an image-based convolutional neural network, and (c) a feedforward neural network, to identify a portion of the cross-entity relationship graph data object for processing by a post-prediction system by:
identifying the cross-entity relationship graph data object associated with a plurality of predictive entities relative to a primary predictive entity, wherein the cross-entity relationship graph data object comprises:
(i) a plurality of entity nodes associated with the plurality of predictive entities,
(ii) an entity node attribute, associated with an entity node of the plurality of entity nodes, that is determined based at least in part on an entity size measure associated with a predictive entity, of the plurality of predictive entities, that is associated with the entity node, and
(iii) a group of directed entity relationship edges, wherein a directed entity relationship edge of the group of directed entity relationship edges is associated with a source entity node and a destination entity node of the plurality of entity nodes,
generating, using the graph convolutional neural network and based at least in part on the cross-entity relationship graph data object, a graph convolutional representation of the cross-entity relationship graph data object,
converting, using a node visualization schema, the cross-entity relationship graph data object to a graph image data object by:
(i) converting the entity node of the cross-entity relationship graph data object to a node icon comprising a defined geographic shape centered in a defined region of the graph image data object, and
(ii) mapping a directed relationship edge of the cross-entity relationship graph data object to an edge line in the graph image data object that stretches from (a) a first node icon corresponding to the source entity node of the directed relationship edge to (b) a second node icon corresponding to the destination entity node of the directed relationship edge,
generating, using the image-based convolutional neural network and based at least in part on the graph image data object of the cross-entity relationship graph data object, an image-based convolutional representation of the cross-entity relationship graph data object,
generating a comprehensive graph representation by concatenating the graph convolutional representation and the image-based convolutional representation, and
processing, using the feedforward neural network, the comprehensive graph representation to identify the portion of the cross-entity relationship graph data object for processing by the post-prediction system based at least in part on the comprehensive graph representation; and
providing data associated with the portion of the cross-entity relationship graph data object to the post-prediction system.
8 . The system of claim 7 , wherein the comprehensive graph representation comprises: (i) one or more extracted relational features determined based at least in part on a related edge subset for the primary predictive entity that comprises the group of directed entity relationship edges whose source predictive entity is the primary predictive entity, (ii) one or more extracted pairwise features determined based at least in part on a related pair subset of a group of entity pairs that comprise the primary predictive entity, and (iii) one or more extracted node features determined based at least in part on a related node subset of the plurality of entity nodes whose entity node attributes satisfy an entity node attribute threshold.
9 . The system of claim 8 , wherein:
(i) the one or more extracted pairwise features comprise D extracted distance features, and
(ii) an extracted distance feature of the D extracted distance features describes a cross-entity distance measure for a corresponding entity pair whose corresponding cross-entity distance measure is among D highest cross-distance entity measures associated with the related pair subset.
10 . The system of claim 8 , wherein:
(i) the one or more extracted relational features comprise C extracted contribution features, and
(ii) an extracted contribution feature of the C extracted contribution features describes a normalized historical contribution measure for a corresponding directed relationship edge whose corresponding normalized historical distribution measure is among top C highest normalized historical contribution measures associated with the related edge subset.
11 . 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:
process feature data from a cross-entity relationship graph data object, using a hybrid graph-based processing machine learning framework that comprises (a) a graph convolutional neural network, (b) an image-based convolutional neural network, and (c) a feedforward neural network, to identify a portion of the cross-entity relationship graph data object for processing by a post-prediction system by:
identifying the cross-entity relationship graph data object associated with a plurality of predictive entities relative to a primary predictive entity, wherein the cross-entity relationship graph data object comprises:
(i) a plurality of entity nodes associated with the plurality of predictive entities,
(ii) an entity node attribute, associated with an entity node of the plurality of entity nodes, that is determined based at least in part on an entity size measure associated with a predictive entity, of the plurality of predictive entities, that is associated with the entity node, and
(iii) a group of directed entity relationship edges, wherein a directed entity relationship edge of the group of directed entity relationship edges is associated with a source entity node and a destination entity node of the plurality of entity nodes,
generating, using the graph convolutional neural network and based at least in part on the cross-entity relationship graph data object, a graph convolutional representation of the cross-entity relationship graph data object,
converting, using a node visualization schema, the cross-entity relationship graph data object to a graph image data object by:
(i) converting the entity node of the cross-entity relationship graph data object to a node icon comprising a defined geographic shape centered in a defined region of the graph image data object, and
(ii) mapping a directed relationship edge of the cross-entity relationship graph data object to an edge line in the graph image data object that stretches from (a) a first node icon corresponding to the source entity node of the directed relationship edge to (b) a second node icon corresponding to the destination entity node of the directed relationship edge,
generating, using the image-based convolutional neural network and based at least in part on the graph image data object of the cross-entity relationship graph data object, an image-based convolutional representation of the cross-entity relationship graph data object,
generating a comprehensive graph representation by concatenating the graph convolutional representation and the image-based convolutional representation, and
processing, using the feedforward neural network, the comprehensive graph representation to identify the portion of the cross-entity relationship graph data object for processing by the post-prediction system based at least in part on the comprehensive graph representation; and
provide data associated with the portion of the cross-entity relationship graph data object to the post-prediction system.
12 . The one or more non-transitory computer-readable storage media of claim 11 , wherein the comprehensive graph representation comprises: (i) one or more extracted relational features determined based at least in part on a related edge subset for the primary predictive entity that comprises the group of directed entity relationship edges whose source predictive entity is the primary predictive entity, (ii) one or more extracted pairwise features determined based at least in part on a related pair subset of a group of entity pairs that comprise the primary predictive entity, and (iii) one or more extracted node features determined based at least in part on a related node subset of the plurality of entity nodes whose entity node attributes satisfy an entity node attribute threshold.
13 . The one or more non-transitory computer-readable storage media of claim 12 , wherein:
(i) the one or more extracted pairwise features comprise D extracted distance features, and
(ii) an extracted distance feature of the D extracted distance features describes a cross-entity distance measure for a corresponding entity pair whose corresponding cross-entity distance measure is among D highest cross-distance entity measures associated with the related pair subset.
14 . The one or more non-transitory computer-readable storage media of claim 12 , wherein:
(i) the one or more extracted relational features comprise C extracted contribution features, and
(ii) an extracted contribution feature of the C extracted contribution features describes a normalized historical contribution measure for a corresponding directed relationship edge whose corresponding normalized historical distribution measure is among top C highest normalized historical contribution measures associated with the related edge subset.