Sas Institute Inc.

Susan Edwards Haller

Ilknur Kaynar Kabul

Jorge Manuel Gomes Da Silva

Jared Langford Dean

Patrick Hall

Ralph Walter Abbey

Raymond Eugene Wright

Robert William Pratt

Ivan Borges Oliveira

Udo Sglavo

Manoj Keshavmurthi Chari

Wendy Ann Czika

David Rawlins Duling

Funda Gunes

A computing device selects a feature set and hyperparameters for a machine learning model to predict a value for a characteristic in a scoring dataset. A number of training model iterations is determined. A unique evaluation pair is selected for each iteration that indicates a feature set selected from feature sets and a hyperparameter configuration selected from hyperparameter configurations. A machine learning model is trained using each unique evaluation pair. Each trained machine learning model is validated to compute a performance measure value. An estimation model is trained with the feature set, the hyperparameter configuration, and the performance measure value computed for unique evaluation pair. The trained estimation model is executed to compute the performance measure value for each unique evaluation pair. A final feature set and a final hyperparameter configuration are selected based on the computed performance measure value.

System for automatic, simultaneous feature selection and hyperparameter tuning for a machine learning model

One or more embodiments may include techniques to computer generate one or more plots based on computational clustering performed by a system. Embodiments include performing clustering on a dataset to generate a number of clusters of data for the dataset. The clusters may be processed and used to generate the one or more plots. In some embodiments, the plots may include one or more variables plotted against a weighted average score associated with a cluster, the plot may visually indicate the effect that the one or more variables has on the predicted outcome. The one or more plots may be presented in a display on a display device. In some embodiments, the plots may be segmented and each segment may correspond with a number of individual curves. The segmented curves may be plotted and displayed on the display device.

Computer system to generate scalable plots using clustering

A computing device to compute clusters using random subsets of variables is provided. Each data point of a plurality of data points is associated with a variable to define a plurality of variables. A subset of the plurality of variables is randomly selected. The subset does not include all of the plurality of variables. A number of clusters into which to segment the received data is determined. Cluster data that defines each cluster of the determined number of clusters is determined by executing a clustering algorithm with the received data using only the plurality of data points defined for each observation that are associated with the randomly selected subset of the plurality of variables. The determined cluster data is stored to cluster second data into the determined number of clusters. The second data is different from the received data.

Cluster computation using random subsets of variables

A computing device to select decorrelated variables using a graph based method is provided. A correlation value is computed between each pair of a plurality of variables to define a correlation matrix. A binary threshold value is compared to each correlation value to define a binary similarity matrix from the correlation matrix. An undirected graph comprising a subgraph that includes one or more connected nodes is defined based on the binary similarity matrix to store connectivity information for the plurality of variables. Each node of the subgraph is pairwise associated with a unique variable of the variables. (a) A least connected node is selected from the undirected graph based on the connectivity information. (b) The selected least connected node is removed from the undirected graph. (c) The connectivity information for the undirected graph is updated based on the removed node. (d) (a)-(c) are repeated until a stop criterion is satisfied.

Graph based selection of decorrelated variables

A computing device to compute composite clusters is provided. A first and a second plurality of centroid locations are computed by executing a clustering algorithm with a first portion of data and a first input parameter and a second portion of the data and a second input parameter, respectively. The first portion is different from the second portion or the first input parameter is different from the second input parameter. A plurality of composite centroid locations is computed using the computed first and second plurality of centroid locations to define a composite set of clusters. An observation is selected. A cluster of the composite set of clusters to which to assign the observation is determined using the plurality of composite centroid locations. The selecting and the determining is repeated with each observation of the plurality of observations as the observation to define cluster assignments for the plurality of observations.

Determination of composite clusters

A computing device presents a cluster visualization based on a neural network computation. First centroid locations are computed for first clusters. Second centroid locations are computed for second clusters. Each centroid location includes a plurality of coordinate values where each coordinate value relates to a single variable of a plurality of variables. Distances are computed pairwise between each centroid location. An optimum pairing is selected based on a minimum distance of the computed pairwise distances where each pair is associated with a different cluster of a set of composite clusters. Noised centroid location data is created. A multi-layer neural network is trained with the noised centroid location data. A projected centroid location is determined in a multidimensional space for each centroid location as values of hidden units of a middle layer of the multi-layer neural network. A graph is presented for display that indicates the determined, projected centroid locations.

Neural network based cluster visualization that computes pairwise distances between centroid locations, and determines a projected centroid location in a multidimensional space

A computing device to assign observations to clusters based on a statistical probability is provided. A first cluster assignment is defined by assigning the plurality of observations to a first set of clusters. A second cluster assignment is defined by assigning the plurality of observations to a second set of clusters. A set of composite clusters is defined based on the defined first set of clusters and the defined second set of clusters. For each observation, a statistical probability value for assigning an observation to each composite cluster of the defined set of composite clusters is computed based on the first and second cluster assignments and a composite cluster assignment is defined by assigning the observation to a cluster of the set of composite clusters based on the computed statistical probability value. The defined composite cluster assignment is stored.

Probabilistic cluster assignment

A computing device to automatically cluster a dataset is provided. Data that includes a plurality of observations with a plurality of data points defined for each observation is received. Each data point of the plurality of data points is associated with a variable to define a plurality of variables. A number of clusters into which to segment the received data is repeatedly selected by repeatedly executing a clustering algorithm with the received data. A plurality of sets of clusters is defined based on the repeated execution of the clustering algorithm that resulted in the selected number of clusters. A plurality of composite clusters is defined based on the defined plurality of sets of clusters. The plurality of observations is assigned to the defined plurality of composite clusters using the plurality of data points defined for each observation.

Computerized cluster analysis framework for decorrelated cluster identification in datasets

Computer-implemented systems and methods are provided for generating bins for a scorecard. An approximate set of bins is generated by applying an optimization model to binning data. The optimization model includes an objective function, constraints, and surrogate weight of evidence metric(s). The approximated set of bins are then used in scorecard operations.

Growing community of inventors

Raleigh, NC, United States of America

Susan Edwards Haller