Abstract: The present invention relates to a method for purifying key intermediates of Citalopram, i.e. 4-[4-(dimethylamino)-1-(4-fluorophenyl)-1-hydroxybutyl]-3-hydroxymethylbenzonitrile and a salt thereof. The method comprises dissolving crude 4-[4-(dimethylamino)-1-(4-fluorophenyl)-1-hydroxybutyl]-3-hydroxymethylbenzonitrile (compound of formula I containing formaldehyde impurity) with an organic solvent, adding a washing solution, controlling the temperature, stirring, leaving to stand for layering, and removing the aqueous layer, so as to obtain a purified organic solution of 4-[4-(dimethylamino)-1-(4-fluorophenyl)-1-hydroxybutyl]-3-hydroxymethylbenzonitrile. The method provided by the present invention can effectively remove aldehyde group-containing impurities in the intermediate. The method of the present invention has the advantages of simple operation, cheap raw materials and mild conditions, and is suitable for large-scale industrial production.

Abstract: A computing device trains a neural network machine learning model. A forward propagation of a first neural network is executed. A backward propagation of the first neural network is executed from a last layer to a last convolution layer to compute a gradient vector. A discriminative localization map is computed for each observation vector with the computed gradient vector using a discriminative localization map function. An activation threshold value is selected for each observation vector from at least two different values based on a prediction error of the first neural network. A biased feature map is computed for each observation vector based on the activation threshold value selected for each observation vector. A masked observation vector is computed for each observation vector using the biased feature map. A forward and a backward propagation of a second neural network is executed a predefined number of iterations using the masked observation vector.

Abstract: A computing device trains a neural network machine learning model. A forward propagation of a first neural network is executed. A backward propagation of the first neural network is executed from a last layer to a last convolution layer to compute a gradient vector. A discriminative localization map is computed for each observation vector with the computed gradient vector using a discriminative localization map function. An activation threshold value is selected for each observation vector from at least two different values based on a prediction error of the first neural network. A biased feature map is computed for each observation vector based on the activation threshold value selected for each observation vector. A masked observation vector is computed for each observation vector using the biased feature map. A forward and a backward propagation of a second neural network is executed a predefined number of iterations using the masked observation vector.

Abstract: A computing device trains a neural network machine learning model. A forward propagation of a first neural network is executed. A backward propagation of the first neural network is executed from a last layer to a last convolution layer of a plurality of convolutional layers to compute a gradient vector for first weight values of the last convolution layer using observation vectors. A discriminative localization map is computed for each observation vector with the gradient vector using a discriminative localization map function. A forward and a backward propagation of a second neural network is executed to compute a second weight value for each neuron of the second neural network using the discriminative localization map computed for each observation vector. A predefined number of iterations of the forward and the backward propagation of the second neural network is repeated.

Abstract: A computing device compresses a gradient boosting tree predictive model. A gradient boosting tree predictive model is trained using a plurality of observation vectors. Each observation vector includes an explanatory variable value of an explanatory variable and a response variable value for a response variable. The gradient boosting tree predictive type model is trained to predict the response variable value of each observation vector based on a respective explanatory variable value of each observation vector. The trained gradient boosting tree predictive model is compressed using a compression model with a predefined penalty constant value and with a predefined array of coefficients to reduce a number of trees of the trained gradient boosting tree predictive model. The compression model minimizes a sparsity norm loss function. The compressed, trained gradient boosting tree predictive model is output for predicting a new response variable value from a new observation vector.

Abstract: Machine-learning models and behavior can be visualized. For example, a machine-learning model can be taught using a teaching dataset. A test input can then be provided to the machine-learning model to determine a baseline confidence-score of the machine-learning model. Next, weights for elements in the teaching dataset can be determined. An analysis dataset can be generated that includes a subset of the elements that have corresponding weights above a predefined threshold. For each overlapping element in both the analysis dataset and the test input, (i) a modified version of the test input can be generated that excludes the overlapping element, and (ii) the modified version of the test input can be provided to the machine-learning model to determine an effect of the overlapping element on the baseline confidence-score. A graphical user interface can be generated that visually depicts the test input and various elements' effects on the baseline confidence-score.

Abstract: A computing device compresses a gradient boosting tree predictive model. A gradient boosting tree predictive model is trained using a plurality of observation vectors. Each observation vector includes an explanatory variable value of an explanatory variable and a response variable value for a response variable. The gradient boosting tree predictive type model is trained to predict the response variable value of each observation vector based on a respective explanatory variable value of each observation vector. The trained gradient boosting tree predictive model is compressed using a compression model with a predefined penalty constant value and with a predefined array of coefficients to reduce a number of trees of the trained gradient boosting tree predictive model. The compression model minimizes a sparsity norm loss function. The compressed, trained gradient boosting tree predictive model is output for predicting a new response variable value from a new observation vector.

Abstract: A computing device computes a quantile value. A maximum value and a minimum value are computed for unsorted variable values to compute an upper bin value and a lower bin value for each bin of a plurality of bins. A frequency counter is computed for each bin by reading the unsorted variable values a second time. A bin number and a cumulative rank value are computed for a quantile. When an estimated memory usage value exceeds a predefined memory size constraint value, a subset of the plurality of bins are split into a plurality of bins, the frequency counter is recomputed for each bin, and the bin number and the cumulative rank value are recomputed. Frequency data is computed using the frequency counters. The quantile value is computed using the frequency data and the cumulative rank value for the quantile and output.

Abstract: Machine-learning models and behavior can be visualized. For example, a machine-learning model can be taught using a teaching dataset. A test input can then be provided to the machine-learning model to determine a baseline confidence-score of the machine-learning model. Next, weights for elements in the teaching dataset can be determined. An analysis dataset can be generated that includes a subset of the elements that have corresponding weights above a predefined threshold. For each overlapping element in both the analysis dataset and the test input, (i) a modified version of the test input can be generated that excludes the overlapping element, and (ii) the modified version of the test input can be provided to the machine-learning model to determine an effect of the overlapping element on the baseline confidence-score. A graphical user interface can be generated that visually depicts the test input and various elements' effects on the baseline confidence-score.

Abstract: A computing device computes a quantile value. A maximum value and a minimum value are computed for unsorted variable values to compute an upper bin value and a lower bin value for each bin of a plurality of bins. A frequency counter is computed for each bin by reading the unsorted variable values a second time. A bin number and a cumulative rank value are computed for a quantile. When an estimated memory usage value exceeds a predefined memory size constraint value, a subset of the plurality of bins are split into a plurality of bins, the frequency counter is recomputed for each bin, and the bin number and the cumulative rank value are recomputed. Frequency data is computed using the frequency counters. The quantile value is computed using the frequency data and the cumulative rank value for the quantile and output.

Abstract: A computing device computes a quantile value. A maximum value and a minimum value are computed for unsorted variable values. An upper bin value and a lower bin value are computed for each bin of a plurality of bins using the maximum and minimum values. A frequency counter is computed for each bin by reading the unsorted variable values a second time. Each frequency counter is a count of the variable values within a respective bin. A bin number and a cumulative rank value are computed for a quantile. The bin number identifies a specific within which a quantile value associated with the quantile is located. The cumulative rank value identifies a cumulative rank for the quantile value associated with the quantile. Frequency data is computed using the frequency counters. The quantile value is computed using the frequency data and the cumulative rank value for the quantile and output.

Abstract: In accordance with the teachings described herein, systems and methods are provided for estimating or determining quantiles for data stored in a distributed system. In one embodiment, an instruction is received to estimate or determine a specified quantile for a variate in a set of data stored at a plurality of nodes in the distributed system. A plurality of data bins for the variate are defined that are each associated with a different range of data values in the set of data. Lower and upper quantile bounds for each of the plurality of data bins are determined based on the total number of data values that fall within each of the plurality of data bins. The specified quantile is estimated or determined based on an identified one of the plurality of data bins that includes the specified quantile based on the lower and upper quantile bounds.

Abstract: In accordance with the teachings described herein, systems and methods are provided for estimating or determining quantiles for data stored in a distributed system. In one embodiment, an instruction is received to estimate or determine a specified quantile for a variate in a set of data stored at a plurality of nodes in the distributed system. A plurality of data bins for the variate are defined that are each associated with a different range of data values in the set of data. Lower and upper quantile bounds for each of the plurality of data bins are determined based on the total number of data values that fall within each of the plurality of data bins. The specified quantile is estimated or determined based on an identified one of the plurality of data bins that includes the specified quantile based on the lower and upper quantile bounds.

Abstract: Techniques for providing interactive decision trees are included. For example, a system is provided that stores data related to a decision tree, wherein the data includes one or more data structures and one or more portions of code. The system receives input corresponding to an interaction request associated with a modification to the decision tree. The system determines whether the modification requires multiple-processing iterations of the distributed data set. The system generates an application layer modified decision tree when the generating requires no multiple-processing iterations of the distributed data set. The system facilitates server layer modification of the decision tree when the modification requires multiple-processing iterations of the distributed data set. The system generates a representation of the application layer modified decision tree or the server layer modified decision tree.

Abstract: Techniques for providing interactive decision trees are included. For example, a system is provided that stores data related to a decision tree, wherein the data includes one or more data structures and one or more portions of code. The system receives input corresponding to an interaction request associated with a modification to the decision tree. The system determines whether the modification requires multiple-processing iterations of the distributed data set. The system generates an application layer modified decision tree when the generating requires no multiple-processing iterations of the distributed data set. The system facilitates server layer modification of the decision tree when the modification requires multiple-processing iterations of the distributed data set. The system generates a representation of the application layer modified decision tree or the server layer modified decision tree.

Abstract: Systems and methods for determining an optimal splitting scheme for a node in a classification decision tree. A computing system may receive input data related to a decision tree to be generated from a data set. The input data identifies a target attribute of the data set and a set of candidate attributes of the data set to be used as nodes in the decision tree. The computing system may determine, using a clustering algorithm and the set of candidate attributes, a number of potential splitting schemes to be used to split a node in the decision tree. The computing system may calculate a splitting measurement for each of the plurality of potential splitting schemes. The computing system may select an optimal splitting scheme from the plurality of potential splitting schemes for each node in the decision tree based on the splitting measurement.

Abstract: Systems and methods for data reduction of a data set are included. A computing system may group data points in a data set into a number of data point bubbles represented by a number of representative points. A data point bubble may include a one or more data points from the data set and a representative point from the data set. The computing system may calculate a cluster assignment for the representative point by executing a clustering algorithm using the number of representative points.