Augmented feature enablement of machine learning models via graph neural networks

1 . A method comprising:

retrieving properties associated with a plurality of machine learning (ML) models;

building a graph data structure using the properties, the graph data structure including nodes corresponding to the properties and edges representing connections between the properties;

training a multi-stack message-passing graph neural network (MSMPGNN) using the graph data structure, wherein training the MSMPGNN comprises iteratively updating a state of each node in the graph data structure by aggregating hidden states of neighboring nodes using a message-passing function, and computing, for each node, a positional embedding based on a current positional embedding for the node, a combination of positional embeddings for all neighbors in a neighborhood of the node, and edge vectors between the node and neighbors of the node;

inputting a graph fragment into the MSMPGNN, wherein inputting the graph fragment comprises attaching the graph fragment to the graph data structure and initializing node representations of the graph fragment;

generating a prediction based on the graph fragment using the MSMPGNN by performing message-passing iterations on the graph data structure with the attached graph fragment and applying a readout function to node-level representations; and

training a new ML model using the prediction.

2 . The method of claim 1 , wherein retrieving the properties comprises retrieving one or more of:

raw features used by the plurality of ML models;

derivative features used by the plurality of ML models;

categories of the plurality of ML models;

feature significance values associated with the plurality of ML models;

category feature significance values associated with the plurality of ML models; and

performance statuses associated with the plurality of ML models.

3 . The method of claim 1 , wherein training the MSMPGNN comprises, for a given ML model in the plurality of ML models:

initializing a state of the given ML model to a default state;

acquiring at least one hidden state of a second ML model in the graph data structure; and

updating the state of the given ML model based on the at least one hidden state.

4 . The method of claim 3 , wherein training the MSMPGNN further comprises computing, for each node, a positional embedding based on an aggregation of positional embeddings for all neighbors in a neighborhood of a respective node.

5 . The method of claim 1 , further comprising re-training the MSMPGNN in response to a trigger comprising one of:

detecting a new ML model added to the plurality of ML models;

detecting a new feature added to an ML model in the plurality of ML models;

detecting a performance change in one of the plurality of ML models;

detecting a new category of ML model; and

receiving a request for re-training from a user.

6 . The method of claim 1 , wherein inputting a graph fragment into the MSMPGNN comprises inputting an ML model category and wherein generating a prediction based on the graph fragment comprises generating a graph including a set of predicted features for the ML model category.

7 . The method of claim 1 , wherein inputting a graph fragment into the MSMPGNN comprises inputting an ML model category and a performance indicator and wherein generating a prediction based on the graph fragment comprises generating a graph including a set of predicted features for the ML model category.

8 . The method of claim 1 , wherein inputting a graph fragment into the MSMPGNN comprises inputting a feature and wherein generating a prediction based on the graph fragment comprises generating a graph including a set of model categories and derivative features.

9 . The method of claim 1 , wherein inputting a graph fragment into the MSMPGNN comprises inputting an ML model category and a feature and wherein generating a prediction based on the graph fragment comprises generating a graph including a set of derivative features.

10 . A non-transitory computer-readable storage medium for tangibly storing computer program instructions capable of being executed by a computer processor, the computer program instructions defining steps of:

retrieving properties associated with a plurality of machine learning (ML) models;

building a graph data structure using the properties, the graph data structure including nodes corresponding to the properties and edges representing connections between the properties;

training a multi-stack message-passing graph neural network (MSMPGNN) using the graph data structure, wherein training the MSMPGNN comprises iteratively updating a state of each node in the graph data structure by aggregating hidden states of neighboring nodes using a message-passing function, and computing, for each node, a positional embedding based on a current positional embedding for the node, a combination of positional embeddings for all neighbors in a neighborhood of the node, and edge vectors between the node and neighbors of the node;

inputting a graph fragment into the MSMPGNN, wherein inputting the graph fragment comprises attaching the graph fragment to the graph data structure and initializing node representations of the graph fragment;

generating a prediction based on the graph fragment using the MSMPGNN by performing message-passing iterations on the graph data structure with the attached graph fragment and applying a readout function to node-level representations; and

training a new ML model using the prediction.

11 . The non-transitory computer-readable storage medium of claim 10 , wherein retrieving the properties comprises retrieving one or more of:

raw features used by the plurality of ML models;

derivative features used by the plurality of ML models;

categories of the plurality of ML models;

feature significance values associated with the plurality of ML models;

category feature significance values associated with the plurality of ML models; and

performance statuses associated with the plurality of ML models.

12 . The non-transitory computer-readable storage medium of claim 10 , wherein training the MSMPGNN comprises, for a given ML model in the plurality of ML models:

initializing a state of the given ML model to a default state;

acquiring at least one hidden state of a second ML model in the graph data structure; and

updating the state of the given ML model based on the at least one hidden state.

13 . The non-transitory computer-readable storage medium of claim 12 , wherein training the MSMPGNN further comprises computing, for each node, a positional embedding based on an aggregation of positional embeddings for all neighbors in a neighborhood of a respective node.

14 . The non-transitory computer-readable storage medium of claim 10 , wherein inputting a graph fragment into the MSMPGNN comprises inputting an ML model category and wherein generating a prediction based on the graph fragment comprises generating a graph including a set of predicted features for the ML model category.

15 . The non-transitory computer-readable storage medium of claim 10 , wherein inputting a graph fragment into the MSMPGNN comprises inputting an ML model category and a performance indicator and wherein generating a prediction based on the graph fragment comprises generating a graph including a set of predicted features for the ML model category.

16 . The non-transitory computer-readable storage medium of claim 10 , wherein inputting a graph fragment into the MSMPGNN comprises inputting a feature and wherein generating a prediction based on the graph fragment comprises generating a graph including a set of model categories and derivative features.

17 . The non-transitory computer-readable storage medium of claim 10 , wherein inputting a graph fragment into the MSMPGNN comprises inputting an ML model category and a feature and wherein generating a prediction based on the graph fragment comprises generating a graph including a set of derivative features.

18 . A device comprising:

a processor configured to:

retrieve properties associated with a plurality of machine learning (ML) models,

build a graph data structure using the properties, the graph data structure including nodes corresponding to the properties and edges representing connections between the properties,

train a multi-stack message-passing graph neural network (MSMPGNN) using the graph data structure, wherein training the MSMPGNN comprises iteratively updating a state of each node in the graph data structure by aggregating hidden states of neighboring nodes using a message-passing function, and computing, for each node, a positional embedding based on a current positional embedding for the node, a combination of positional embeddings for all neighbors in a neighborhood of the node, and edge vectors between the node and neighbors of the node,

input a graph fragment into the MSMPGNN, wherein inputting the graph fragment comprises attaching the graph fragment to the graph data structure and initializing node representations of the graph fragment,

generate a prediction based on the graph fragment using the MSMPGNN by performing message-passing iterations on the graph data structure with the attached graph fragment and applying a readout function to node-level representations, and

train a new ML model using the prediction.

19 . The device of claim 18 , wherein retrieving the properties comprises retrieving one or more of:

raw features used by the plurality of ML models;

derivative features used by the plurality of ML models;

categories of the plurality of ML models;

feature significance values associated with the plurality of ML models;

category feature significance values associated with the plurality of ML models; and

performance statuses associated with the plurality of ML models.

20 . The device of claim 18 , wherein training the MSMPGNN comprises, for a given ML model in the plurality of ML models:

initializing a state of the given ML model to a default state;

acquiring at least one hidden state of a second ML model in the graph data structure; and

updating the state of the given ML model based on the at least one hidden state.