Predicting values for a multitude of time series with target and input variables connected in a graph
A computer-implemented method for training a machine learning-artificial intelligence model for multiple prediction tasks includes inputting data for tasks and additional data sources through a common trainable task representation function to obtain a data representation for each. Each resulting data representation is input through two individual trainable linear functions to obtain a corresponding prediction and adversarial prediction. A prediction error for the tasks, an adversarial error across edges of a graph, an auxiliary error for the additional data sources, and a graph error are determined. Parameters of the common trainable task representation function and the trainable linear functions are trained based on a comparison against a weighted sum of the errors. The present invention can be used in a variety of applications including, but not limited to, several anticipated use cases in drug development, material synthesis, and medical/healthcare.
1 . A computer-implemented method for training one or more machine learning-artificial intelligence (ML-AI) models for multiple prediction tasks, the method comprising:
inputting data for each task of a plurality of tasks and data for each additional data source of a plurality of additional data sources through a common trainable task representation function to obtain a resulting data representation for each task and each additional data source;
inputting each resulting data representation through at least two individual trainable linear functions to obtain a corresponding prediction and a corresponding adversarial prediction;
determining a prediction error for the plurality of tasks, an adversarial error across edges of a graph comprising the plurality of tasks and the plurality of additional data sources, an auxiliary error for the plurality of additional data sources, and a graph error between a given input graph and a graph induced by model weights; and
training parameters of the common trainable task representation function and parameters of the two individual trainable linear functions based on a comparison against a weighted sum of the prediction error, the adversarial error, the auxiliary error, and the graph error.
2 . The method of claim 1 , wherein inputting each resulting data representation through the at least two individual trainable linear functions to obtain the corresponding prediction and the corresponding adversarial prediction further comprises:
inputting the resulting representation of the data for each task to a first linear function of the at least two individual trainable linear functions to obtain the corresponding prediction; and
inputting the resulting representation of the data for each additional data source to a second linear function of the at least two individual trainable linear functions to obtain the corresponding adversarial prediction.
3 . The method of claim 1 , wherein determining the prediction error for the plurality of tasks further comprises:
subtracting an output of each task from each corresponding prediction to achieve a prediction error result for each task and corresponding prediction; and
performing a summation of every prediction error result.
4 . The method of claim 1 , the method further comprising mapping each task and each additional data source to an embedding space, wherein determining the adversarial error across the edges of the graph further comprises determining the difference between:
a first distance in the embedding space between the corresponding prediction for each task and the corresponding adversarial prediction for each task; and
a second distance in the embedding space between the corresponding prediction for each additional data source and the corresponding adversarial prediction for each additional data source.
5 . The method of claim 1 , wherein the data for each task of the plurality of tasks and data for each additional data source of the plurality of additional data sources comprises historical time series data, and wherein determining the auxiliary error for the plurality of additional data sources further comprises:
subtracting an output of each additional data source from each corresponding prediction to achieve an auxiliary error result for each additional data source and corresponding prediction; and
performing a summation of the auxiliary error results.
6 . The method of claim 1 , wherein determining the graph error between the given input graph and the graph induced by the model weights further comprises:
determining a difference between the graph a weighted induced graph from learned weights of the at least two individual trainable linear functions.
7 . The method of claim 1 , the method further comprising making a prediction using the one or more ML-AI models based on the training parameters of the common trainable task representation function.
8 . The method of claim 1 , wherein the plurality of additional data sources comprise a plurality of different types of data, the different types of data comprising data of one or more sensors or devices measuring different signals of an individual, wherein the graph comprises edges between a target medical condition and the data of the one or more sensors or devices, and wherein the one or more ML-AI models contribute to a target medical condition prediction of the individual.
9 . The method of claim 1 , wherein data for each task of the plurality of tasks comprises vehicle traffic indicators, and the plurality of additional data sources comprise a plurality of different types of data, the different types of data comprising data of street noise, and air quality, wherein the graph comprises edges between the vehicle traffic indicators and the street noise and air quality data, and wherein the one or more ML-AI models contribute to a traffic condition prediction.
10 . The method of claim 1 , wherein the data for each task of the plurality of task comprises manufacturing conditions of a material, and wherein the plurality of additional data sources comprise a plurality of different types of data, the different types of data comprising applied heating conditions and applied pressure conditions, wherein the graph comprises edges between the manufacturing conditions of the material and the applied heating conditions and the applied pressure conditions, and wherein the one or more ML-AI models contribute to a physical or chemical property prediction of the material.
11 . A tangible, non-transitory computer-readable medium having instructions thereon which, upon being executed by one or more hardware processors, alone or in combination, provide for training one or more machine learning-artificial intelligence (ML-AI) models for multiple prediction tasks according to claim 1 .
12 . A computer system comprising one or more hardware processors for training one or more machine learning-artificial intelligence (ML-AI) models for multiple prediction tasks which, alone or in combination, are configured to provide for execution of the following steps:
inputting data for each task of a plurality of tasks and data for each additional data source of a plurality of additional data sources through a common trainable task representation function to obtain a resulting data representation for each task and each additional data source;
inputting each resulting data representation through at least two individual trainable linear functions to obtain a corresponding prediction and a corresponding adversarial prediction;
determining a prediction error for the plurality of tasks, an adversarial error across edges of a graph comprising the plurality of tasks and the plurality of additional data sources, an auxiliary error for the plurality of additional data sources, and a graph error between a given input graph and a graph induced by model weights; and
training parameters of the common trainable task representation function and parameters of the two individual trainable linear functions based on a comparison against a weighted sum of the prediction error, the adversarial error, the auxiliary error, and the graph error.