Superseded federated learning

A non-transitory computer readable medium (CRM) comprising computer readable program code, which when executed by a computer processor, enables the computer processor to perform a method for inferring outcomes using superseded federated learning on a system, the method comprising:

• identifying, on a node of a plurality of nodes of the system, a set of common features of data sets stored on different nodes of a plurality of nodes by implementing a private set intersection (PSI) between the plurality of nodes to ascertain an intersection of the data sets of the different nodes without exposing respective data sets of the different nodes to other nodes;
• identifying a set of common sample identifiers (IDs) based on a common feature of the set of common features, wherein the PSI is applied to varying feature value vectors, representing the common feature of the set of common features, to identify the set of common sample IDs;
• obtaining a local domain dataset mapped to the set of common sample IDs, wherein the local domain dataset is a portion of the local domain data and wherein the local domain data comprises data of the node;
• initializing, after obtaining the local domain dataset, a local learning model by configuring a learning state with respective default values;
• producing a trained local learning model from the local learning model by training on the local domain dataset, wherein training entails iterative adjustments to the learning state until an acceptable performance condition is reached;
• initializing, after producing the trained local learning model, a local generative adversarial network (GAN), wherein the local GAN comprises a local generator model and a local discriminator model and wherein the local generator model and the local discriminator model are pit against each other to generate synthetic data;
• producing a trained local GAN from the local GAN based on a local generator model training data set and a local discriminator model training data set, wherein the local generator model training set comprises the data of the node mapped to the set of common features and wherein the local discriminator model training data set comprises second data sampled from the local domain data;
• extracting a trained local generator model from the trained local GAN;
• sharing the trained local generator model and the local learning model with at least one other remote node of the plurality of nodes via a network, wherein the at least one other remote node is separate from the node;
• receiving, after sharing and from the at least one other remote node of the system, one or more trained remote generator models and one or more trained remote learning models via the network;
• after receiving the one or more trained remote generator models and the one or more trained remote learning models:
• obtaining a new sample at the node of a system;
• processing the new sample using the trained local learning model to obtain a local model outcome;
• deriving a remote generator model input data set from the new sample;
• processing the remote generator model input data set using the one or more trained remote generator models of the plurality of nodes to obtain at least one synthesized remote learning model input data set, wherein the at least one synthesized remote learning model input data set is created to simulate sample data from separate nodes of the plurality of nodes
• processing the at least one synthesized remote learning model input data set using the one or more trained remote learning models on the node to obtain at least one remote model outcome;
• and
• aggregating the local model outcome and the at least one remote model outcome using a local ensemble learning model of the node to produce an aggregated outcome, wherein the local ensemble learning model comprises a machine learning model configured to aggregate the local model outcome and the at least one remote model outcome using an ensemble learning technique.

A system, comprising:

• a plurality of nodes operatively connected to one another through a network, wherein each node of the plurality of nodes comprises a computer processor configured to perform a method for inferring outcomes using superseded federated learning, the method comprising:
• identifying a set of common features of data sets stored on different nodes of the plurality of nodes by implementing a private set intersection (PSI) between the plurality of nodes to ascertain an intersection of the data sets of the different nodes without exposing respective data sets of the different nodes to other nodes;
• identifying a set of common sample identifiers (IDs) based on a common feature of the set of common features, wherein the PSI is applied to varying feature value vectors, representing the common feature of the set of common features, to identify the set of common sample IDs;
• obtaining a local domain dataset mapped to the set of common sample IDs, wherein the local domain dataset is a portion of the local domain data and wherein the local domain data comprises data of the node;
• initializing, after obtaining the local domain dataset, a local learning model by configuring a learning state with respective default values;
• producing a trained local learning model from the local learning model by training on the local domain dataset, wherein training entails iterative adjustments to the learning state until an acceptable performance condition is reached;
• initializing, after producing the trained local learning model, a local generative adversarial network (GAN), wherein the local GAN comprises a local generator model and a local discriminator model and wherein the local generator model and the local discriminator model are pit against each other to generate synthetic data;
• producing a trained local GAN from the local GAN based on a local generator model training data set and a local discriminator model training data set, wherein the local generator model training set comprises the data of the node mapped to the set of common features and wherein the local discriminator model training data set comprises second data sampled from the local domain data;
• extracting a trained local generator model from the trained local GAN;
• sharing the trained local generator model and the local learning model with at least one other remote node of the plurality of nodes via a network, wherein the at least one other remote node is separate from the node;
• receiving, after sharing and from the at least one other remote node of the system, one or more trained remote generator models and one or more trained remote learning models via the network;
• after receiving the one or more trained remote generator models and the one or more trained remote learning models;
• obtaining a new sample;
• processing the new sample using the trained local learning model to obtain a local model outcome;
• deriving a remote generator model input data set from the new sample;
• processing the remote generator model input data set using the one or more trained remote generator models of the plurality of nodes to obtain at least one synthesized remote learning model input data set, wherein
• the at least one synthesized remote learning model input data set is created to simulate sample data from separate nodes of the plurality of nodes;
• processing the at least one synthesized remote learning model input data set using the one or more trained remote learning models on the node to obtain at least one remote model outcome;
• and
• aggregating the local model outcome and the at least one remote model outcome using a local ensemble learning model of the node to produce an aggregated outcome, wherein the local ensemble learning model comprises a machine learning model configured to aggregate the local model outcome and the at least one remote model outcome using an ensemble learning technique.