Probabilistic matrix factorization for automated machine learning

A method, comprising:

• generating a matrix using at least two different data sets to define columns of the matrix and at least two different workflows to define rows of the matrix, each of the different workflows comprising one or more preprocessing steps for a machine learning model, the machine learning model, and at least one parameter for the machine learning model or the one or more preprocessing steps;
• selecting a subset of the different data sets and for each data set in the subset of data sets:
• selecting one or more of the different workflows to be trained and tested on the data set;
• applying the selected one or more workflows to the data set;
• and
• generating a populated matrix by recording a workflow performance metric for each workflow applied to the data set in a cell of the generated matrix;
• calculating a low-rank decomposition of the populated matrix;
• and
• determining, for a new data set, a hybrid workflow using the calculated low-rank decomposition of the populated matrix, the hybrid workflow comprising one of the different workflows that provides a maximum performance metric for the new data set the hybrid workflow including a combination of portions of at least two of the different workflows, corresponding predicted performances of the at least two different workflows determined based on the low-rank decomposition, a predicted performance of the hybrid workflow determined by interpolating between the corresponding predicted performances, and the hybrid workflow not included in either of the populated matrix of the at least two different workflows.

A non-transitory computer-readable storage memory comprising stored instructions that are executable and, responsive to execution of the stored instructions by a computing system, the computing system performs operations comprising:

• generating a matrix using at least two different data sets to define columns of the matrix and at least two different workflows to define rows of the matrix, each of the different workflows comprising one or more preprocessing steps for a machine learning model, the machine learning model, and at least one parameter for the machine learning model or the one or more preprocessing steps;
• selecting a subset of the different data sets and for each data set in the subset of data sets:
• selecting one or more of the different workflows to be trained and tested on the data set;
• applying the selected one or more workflows to the data set;
• and
• generating a populated matrix by recording a workflow performance metric for each workflow applied to the data set in a cell of the generated matrix;
• calculating a low-rank decomposition of the populated matrix;
• and
• determining, for a new data set, a hybrid workflow using the calculated low-rank decomposition of the populated matrix, the hybrid workflow comprising one of the different workflows that provides a maximum performance metric for the new data set, the hybrid workflow including a combination of portions of at least two of the different workflows, corresponding predicted performances of the at least two different workflows determined based on the low-rank decomposition, a predicted performance of the hybrid workflow determined by interpolating between the corresponding predicted performances, and the hybrid workflow not included in either of the populated matrix of the at least two different workflows.