Transaction composition graph node embedding
View Patent ↗A computer-implemented process for transaction composition graph node embedding comprising traversing a data flow of transactions to convert a full graph to multiple directed acyclic subgraphs/paths in spanning trees, taking one-by-one nodes as input to a predetermined neural network, generating a set of one-hot vectors for all nodes, computing an embedding vector from a corresponding one-hot vector, computing a probability that an output node is nearby, and embedding the node to a latent feature vector.
1. A computer-implemented method (CIM) comprising:
receiving a plurality of transaction composition graphs with each transaction composite graph including a plurality of nodes including a starting node;
for each given transaction composite graph of the plurality of transaction composition graphs, traversing the given transaction composite graph from the starting node and following a data flow of transactions to convert the given transaction composition graph into a respectively corresponding directed acyclic subgraph representing a spanning tree corresponding to nodes and edges of the given transaction composite graph;
for each given node of each given plurality of nodes of each given transaction composite graphs of the plurality of transaction composite graphs, taking one-by-one nodes surrounding the given node within a predetermined number of edge connections with respect to the given node as input to a predetermined neural network;
for each given node of each given plurality of nodes of each given transaction composite graphs of the plurality of transaction composite graphs, generating a set of one-hot vectors for the given node;
for each given node of each given plurality of nodes of each given transaction composite graphs of the plurality of transaction composite graphs, computing an embedding vector based on the one-hot vector corresponding to the given node using a hidden layer of the predetermined neural network, with the computation of the embedding vector including application of a weight matrix to the one-hot vector; and
using the embedding vector to analyze areas of interest within the plurality of transaction composite graphs.
2. The CIM of claim 1 further comprising:
computing a probability value corresponding to a probability that an output node is nearby an input node using a softmax output layer of the predetermined neural network;
determining that the probability value exceeds a predetermined threshold value; and
in response to a determination that the probability value exceeds the threshold value, embedding the input node to a latent feature vector.
3. The CIM of claim 2 wherein
each transaction composition graph of the plurality of transaction composition graphs represents a plurality of database updating transactions.