Data Science

Many a times need arises for deriving insights from structured or unstructured data. To provide meaningful information from this large data, one incorporates data science to interpret data for decision-making.

Data Science is a field that comprises everything related to data cleansing, preparation and analysis. It is the method used to extract insights and information from data.

Gathr incorporates data science technique for deriving meaningful information from data. Using Gathr ML, PMML and H2O processors one can train models as well as score models for streaming and batch data.

Making predictions on real-time data streams or on a batch of data involves building an offline model and applying it to a stream. Models incorporate one or more machine-learning algorithm that is trained using the collected data.

Supported Technologies

We support models that are created using ML, H2O, and PMML. In addition, we support some of the R models (created in PMML) to be used for scoring on data in Gathr.

ML

ML Analytics processor enables you to use predictive models built on top of ML package.

ML provides higher-level API built on top of data frames that helps you to create and tune practical machine learning pipelines.

The documentation of ML Models is divided in the following section:

• Training

• Prediction (Model Scoring)

Model Training

Models can be trained through Gathr with the help of ML processors. These models will be built on top of an ML package.

Model Training can only be performed on Batch Data.

You can connect multiple models of different or same algorithm and train model on the same message in a single pipeline.

Note: Intermediate columns calculated through transformations from one analytics processor will not be available on the next analytics processor if multiple analytics processor is trained in one pipeline.

Algorithms

You have eight algorithms under ML that supports Model Training and Scoring.

The data flow for all these models include a wizard-like flow, which is shown in the figure below:

Note: For the following models, Post-Processing, Model Evaluation and Hyper Parameters are not mandatory.

Isotonic Regression

Isotonic Regression belongs to the family of Regression algorithms. It gives an approximate series of one-dimensional observations with a non-decreasing function. Isotonic Regression Analytics processor is used to analyze data using ML Isotonic Regression Model.

To use an Isotonic Regression Model in Data Pipeline, drag and drop the model component to the pipeline canvas and right click on it to configure.

The Configuration window of every ML model is same.

After Configuration tab comes, Feature Selection tab. (which is also same for all the models except K Means).

Once Feature Selection is done, perform Pre-Processing on the data before feeding it to the Model. The configuration settings are same for all the ML models.

Then configure the Model using Model Configuration. Following tabs are generated for this Model:

After Model Configuration, Post-Processing is done, Model Evaluation can be performed.

Then, apply the Hyper Parameters on the model to enable tuning your configuration; after which you can simply add notes and save the Configuration.

Linear Regression

Regression is an approach for modeling the relationship between a scalar dependent variable and one or more explanatory variables (or independent variables).

Regression Analytics processor is used to analyze data using ML LinearRegressionModel.

To use a Linear Regression Model in Data Pipeline, drag and drop the model component to the pipeline canvas and right click on it to configure.

The Configuration window of every ML model is same.

After Configuration tab comes, Feature Selection tab. (which is also same for all the models except K Means).

Once Feature Selection is done, perform Pre-Processing on the data before feeding it to the Model. The configuration settings are same for all the ML models.

Then configure the Model using Model Configuration. Following tabs are generated for this Model:

After Model Configuration, Post-Processing is done, Model Evaluation can be performed.

Apply the Hyper Parameters on the model to enable tuning your configuration; after which you can simply add notes and save the Configuration.

Decision Tree

Description

Decision trees and their ensembles are popular methods for the machine learning tasks of classification and regression. Decision trees algorithms are easy to interpret, they handle categorical features, extend to the multi-class classification setting, do not require feature scaling, and are able to capture non-linearity and feature interactions.

Decision Tree Analytics processor is used to analyze data using ML’s DecisionTreeClassificationModel and DecisionTreeRegressionModel.

To use a Decision Tree Model in Data Pipeline, drag and drop the model component to the pipeline canvas and right click on it to configure:

The Configuration window of every ML model is same.

After Configuration tab comes, Feature Selection tab. (which is also same for all the models except K Means).

Once Feature Selection is done, perform Pre-Processing on the data before feeding it to the Model. The configuration settings are same for all the ML models.

Then configure the Model using Model Configuration. Following tabs are generated for this Model:

After Model Configuration, Post-Processing is done, Model Evaluation can be performed.

Apply the Hyper Parameters on the model to enable tuning your configuration; after which you can simply add notes and save the Configuration.

View Model

Decision Tree model support visualization of trained models. View Model feature is available in case of Prediction pipeline.

The trained tree model will be visualized as below:

Gradient Boosted Tree

Gradient-Boosted Trees (GBTs) are ensembles of decision trees. GBTs can be used for binary classification and for regression, using both continuous and categorical features.

Gradient-Boosted Trees Analytics processor is used to analyze data using ML’s GBTClassificationModel and GBTRegressionModel.

Note: GBTs does not support multi-label classification.

To use a GBT Model in Data Pipeline, drag and drop the model component to the pipeline canvas and right click on it to configure.

The Configuration window of every ML model is same.

After Configuration tab comes, Feature Selection tab. (which is also same for all the models except K Means).

Once Feature Selection is done, you will have to perform Pre-Processing on the data before feeding it to the Model. The configuration settings are same for all the ML models.

Then configure the Model using Model Configuration. Following tabs are generated for this Model:

After Model Configuration, Post-Processing is done, Model Evaluation can be performed.

Apply the Hyper Parameters on the model to enable tuning your configuration; after which you can simply add notes and save the Configuration.

Random Forest Trees

Random forests are ensembles of decision trees. Random forests combine many decision trees to reduce the risk of over fitting. Random forests can be used for binary and multi-class classification and for regression, using both continuous and categorical features.

Random Forest Trees Analytics processor is used to analyze data using ML’s RandomForestClassificationModel and RandomForestRegressionModel.

To use a Random Forest Tress Regression Model in Data Pipeline, drag and drop the model component to the pipeline canvas and right click on it to configure.

The Configuration window of every ML model is same.

After Configuration tab comes, Feature Selection tab. (which is also same for all the models except K-Means).

Once Feature Selection is done, perform Pre-Processing on the data before feeding it to the Model. The configuration settings are same for all the ML models.

Then configure the Model using Model Configuration. Following tabs are generated for this Model:

After Model Configuration, Post-Processing is done, Model Evaluation can be performed.

Apply the Hyper Parameters on the model to enable tuning your configuration; after which you can simply add notes and save the Configuration.

Logistic Regression

Logistic regression is a popular method to predict a categorical response. A special case of generalized linear models predicts the probability of the outcomes. It can be used for both binary and multi-class classification problems.

Logistic Regression Analytics processor is used to analyze data using ML’s LogisticRegressionModel.

To us e a Logistic Regression Model in Data Pipeline, drag and drop the model component to the pipeline canvas and right-click on it to configure.

The Configuration window of every ML model is same.

After Configuration tab comes, Feature Selection tab. (which is also same for all the models except K-Means).

Once Feature Selection is done, perform Pre-Processing on the Model. The configuration settings are same for all the ML models.

Then configure the Model using Model Configuration. Following tabs are generated for this Model:

After Model Configuration, Post-Processing is done, Model Evaluation can be performed.

Apply the Hyper Parameters on the model to enable tuning your configuration; after which you can simply add notes and save the Configuration.

Naive Bayes

Naive Bayes are a family of simple probabilistic classifiers based on applying Bayes’ theorem with strong (naive) independence assumptions between the features. Currently, it supports both multinomial Naive Bayes and Bernoulli Naive Bayes.

Naive Bayes Analytics processor is used to analyze data using ML’s NaiveBayesModel.

To use a Naïve Bayes Model in Data Pipeline, drag and drop the model component to the pipeline canvas and right-click on it to configure.

The Configuration window of every ML model is same.

After Configuration tab comes, Feature Selection tab. (which is also same for all the models except K Means).

Once Feature Selection is done, perform Pre-Processing on the data before feeding it to the Model. The configuration settings are same for all the ML models.

Then configure the Model using Model Configuration. Following tabs are generated for this Model:

After Model Configuration, Post-Processing is done, Model Evaluation can be performed.

Apply the Hyper Parameters on the model to enable tuning your configuration; after which you can simply add notes and save the Configuration.

K-Means

K-Means is one of the most commonly used clustering algorithms that clusters the data points into a predefined number of clusters. K-Means Analytics processor is used to analyze data using ML’s K-means Model.

To use a K-Means Model in Data Pipeline, drag and drop the model component to the pipeline and right click on it to configure.

The Configuration window of every ML model is same.

After Configuration tab comes, Feature Selection tab. (which is also same for all the models except K Means).

Once Feature Selection is done, perform Pre-Processing on the data before feeding it to the Model. The configuration settings are same for all the ML models.

Then configure the Model using Model Configuration. Following tabs are generated for this Model:

Configuration Section

The configuration page is common for all the ML Models except Tree based, Naive Bayes and Logistic Regression models. For rest of the models, following are the properties with screenshot:

Note: Algorithm Type parameter is required in case of all the tree algorithms only i.e. Decision Trees, Gradient-Boosted Trees and Random Forests.

Feature Selection

For using analytics processor in both training and prediction mode, you have to explicitly specify Input Labels and Variables such as Continuous, Categorical and Text.

Note:

• In case of Isotonic Regression, specify Input Label and Continuous Variable.

• In case of K-Means, Input Label is not required, since it is used for clustering issues

.

Pre-Processing

In Pre-Processing, the data is transformed or consolidated so that the resulting mining process is more efficient, and the patterns found are easier to understand.

Once features are selected on Features selection tab, you can apply various transformations using Pre-Processing tab.

All ML models require feature column to be Vector Data type. For transforming raw input fields into type Vector, use Pre-Processing transformations.

Following are the descriptions of all the transformations/algorithms supported by Gathr over various analytics processor.

Binarizer

Binarizer thresholds numerical features to binary (0/1) features. Feature values greater than the threshold are binarized to 1.0; values equal to or less than the threshold are binarized to 0.0.

Enter values in the following parameters:

Bucketizer

Bucketizer transforms a column of continuous features to a column of feature buckets.

For configuration of Bucketizer Transformation, select algorithm Bucketizer.

Enter values in the following parameters:

Imputer

The Imputer transformer completes missing values in a dataset, either using the mean or the median of the columns in which the missing values are located. Imputer doesn’t support categorical features. By default, all null values in the input columns are treated as missing, and so are also imputed. The input columns should be of decimal type.

CountVectorizer

CountVectorizer aim to help convert a collection of text documents to vectors of token counts.

HashingTF

HashingTF is a transformer which takes sets of terms and converts those sets into fixed-length feature vectors.

IDF

The IDFModel takes feature vectors (generally created from HashingTF or CountVectorizer) and scales each column. Intuitively, it down-weights columns which appear frequently in a corpus.

MaxAbsScaler

MaxAbsScaler transforms a dataset of Vector rows, rescaling each feature to range [-1, 1] by dividing through the maximum absolute value in each feature. It does not shift/center the data, and thus does not destroy any sparsity.

MaxAbsScaler computes summary statistics on a data set and produces a MaxAbsScalerModel. The model can then transform each feature individually to range [-1, 1].

MinMaxScaler

MinMaxScaler transforms a dataset of Vector rows, rescaling each feature to a specific range (specified by parameter-min/max).

OneHotEncoder

One-hot encoding maps a column of label indices to a column of binary vectors, with at most one single value. This encoding allows algorithms which expect continuous features, such as Logistic Regression, to use categorical features.

NGram

NGram takes as input a sequence of strings (e.g. the output of a Tokenizer). The parameter n is used to determine the number of terms in each n-gram. The output will consist of a sequence of n-grams where each n-gram is represented by a space-delimited string of n consecutive words. If the input sequence contains fewer than n strings, no output is produced.

Normalizer

Normalizer is a transformer that transforms a dataset of Vector rows, normalizing each Vector to have unit norm. It takes parameter p, which specifies the p-norm used for normalization. (p=2 by default.) This normalization can help standardize your input data and improve the behavior of learning algorithms.

PCA

PCA is a statistical procedure that uses an orthogonal transformation to convert a set of observations of possibly correlated variables into a set of values of linearly uncorrelated variables called principal components. A PCA class trains a model to project vectors to a low-dimensional space using PCA.

StandardScaler

StandardScaler transforms a dataset of Vector rows, normalizing each feature to have unit standard deviation and or zero mean.

StopWordsRemover

Stop words are words which should be excluded from the input, characteristically because the words appear frequently and do not carry much meaning.

StopWordsRemover takes as input a sequence of strings (e.g. the output of a Tokenizer) and drops all the stop words from the input sequences. The list of StopWords is specified by the StopWords parameter.

Tokenizer

Tokenization is the process of taking text (such as a sentence) and breaking it into individual terms (usually words). A simple Tokenizer class provides this functionality.

VectorAssembler

VectorAssembler is a transformer that combines a given list of columns into a single vector column.

VectorIndexer

VectorIndexer helps index categorical features in datasets of Vectors. It can both automatically decide which features are categorical and convert original values to category indices.

For Configuration of VectorIndexer Transformation one has to select as algorithm VectorIndexer on the transformations tab. It asks for various configuration fields. Their description is as below:

Word2Vec

Word2Vec maps each word to a unique fixed-size vector.

For configuration of Word2Vec Transformation, one has to select as algorithm Word2Vec on the transformations tab. It asks for various configuration fields. Their description is as below:

StringIndexer

StringIndexer encodes a string column of labels to a column of label indices. The indices are in [0, numLabels), ordered by label frequencies, so the most frequent label gets index 0. The unseen labels will be put at index numLabels if user chooses to keep them. If the input column is numeric, we cast it to string and index the string values.

For configuration of StringIndexer Transformation, one has to select as algorithm StringIndexer on the transformations tab. It asks for various configuration fields. Their description is as below:

Feature Hasher

Feature Hasher projects a set of categorical or numerical features into a feature vector of specified dimension. This is done using a hashing trick to map features to indices in the feature vector. Null (missing) values are ignored (implicitly zero in the resulting feature vector).

For configuration of Feature Hasher, one has to select the algorithm on the transformations tab. It asks for various configuration fields. The description is as below:

Note: Model Training is not supported in streaming pipelines.

Post-Processing

The post-processing tab enables you to perform transformations on model output before displaying the final result.

Currently, Gathr supports only one algorithm i.e IndexToString for post-processing.

IndexToString

IndexToString maps a column of label indices back to a column containing the original labels as strings. A common use case is to produce indices from labels with StringIndexer, train a model with those indices and retrieve the original labels from the column of predicted indices with IndexToString. However, you are free to supply your own labels.

Model Evaluation

Evaluate models on the metrics available for the ML algorithm.

Model Evaluation is configured using the following three properties:

Note: Elastic Search will be used in the background to store actual label of the data and model’s output to evaluate the model metrics.

Hyper Parameters

Using this tab, you are able to optimize hyper parameters of algorithms used in the transformation chain.

Note: Connection Name and HDFS Path are populated when Select Click on the Add Notes tab. Enter the notes in the specified area.

Click on the SAVE button after entering all the information.

Prediction/Model Scoring

Using Models Trained in Gathr

Once the model is trained using training pipelines, it is registered to be used for scoring in any pipeline.

To use a trained model in a pipeline for scoring, drag and drop the analytics processor and change the mode of analytics processor from training to prediction.

Note:

Operations, Message Name Model Name- Common for models

Anomaly options are only available for K means and Algorithm Type is available for Decision Tree, GBT ands Random Forest models.

Save the analytics processor and connect an emitter, for verifying the output.

Once the pipeline is saved, run it for predicting the output.

Using Externally Trained Models

For using externally trained model into Gathr, register the model from Register Entities tab.

Registering Trained Models

For registering a model,

1.​Click on the register entities section on the left pane.

2.​After clicking on register, entities section move to register models tab.

3.​Click on the (+) icon shown in the top-right corner of register models tab. A new window will open, enter model related parameters such as name of the model, whether pipeline model or not, API used for creating model and the algorithm type of the model. In case of Tree Ensemble based models, select, if the model is for classification problem or regression one.

4.​After configuring the above fields, upload model for registration or provide HDFS path wherein the model is saved.

5.​Validate the model by clicking validate model button, if it fails then the trained model is incorrect and if it is valid, register the model for prediction by clicking on the register model button.

6.​Once the model is registered, you can use it for Prediction.

Note: You can only register models trained using ML and H20 API.

PMML

PMML stands for “Predictive Model Markup Language”. It is the de facto standard to represent predictive solutions. A PMML file may contain a myriad of data transformations (pre and post-processing) as well as one or more predictive models.

Its structure follows a set of pre-defined elements and attributes which reflect the inner structure of a predictive workflow: data manipulations followed by one or more predictive models.

Below mentioned is the list of PMML Models available in Gathr.

Logistic Regression

Logistic Regression is a classification algorithm. It is used to describe data and to explain the relationship between one dependent binary variable and one or more independent variables. It is used to predict a binary outcome like 1 or 0, yes or no, true or false given a set of independent variables.

Logistic Regression analytics processor will be used to perform prediction on the incoming data by using a PMML.

An existing PMML file can uploaded to the processor or a PMML can by created using the processor.

Configuring Logistic Regression Model

To add Logistic Regression model into your pipeline, drag Logistic Regression model to the canvas and right click on it to configure.

The Configuration Settings of Logistic Regression model are as follows:

Logistic Regression can be processed in two ways:

Import PMML model

Choose Import Model as Yes and click the Import button to load the PMML file.

Once the model is imported, following checks will be performed on the Model:

• The imported file must be a valid PMML file and it must have Logistic Regression model.

• The features and output defined in the model must be present in the selected message with the data type expected by model.

If the check fails, then the model will not be loaded in the processor and save features will be disabled. If above checks are met, then model can be saved.

Note:

• If the PMML model imported is not as per the analytics processor chosen, then validation of the model will throw an error.

a) All the values on screen will be read-only since an imported model cannot be edited.

b) If model variables are not defined in the message, they appear in red. This is to highlight that these fields need to be defined in the message, to consume the model in Gathr, as shown below:

Since PMML model is imported, the Variables cannot be edited.

Note: If there is any Variable Type that is not defined in the message, the variables will appear in red.

If you want to classify the model output i.e. probability, then specify the threshold parameter, Low and High Classifiers under Variable Type. Threshold parameter takes a numeric value. The output value of the model is compared with the threshold, if the output is greater than threshold then the High classifier appear as output otherwise the Low classifier will appear.

After viewing the Variable Type, Click Next and Model Coefficients Page will open.

Click on Load Defined Variables to load the Variables and provide values against variables.

Now you can Test your Model with Values.

Once you have tested the mode, provide your notes in the add notes section and save the configuration.

Other option is to create your own PMML model.

Create PPML Model

If you want to create a PMML Logistic Regression model, follow the steps mentioned below:

Select a message from Message and select Import PMML as No.

Select Next, it will take you to Variable Type tab.

Select the Input Variables, which is Continuous and Categorical Variables.

Provide a name to the Predicted Variable (Output) and provide Class labels for output values and a Threshold value.

Click Next to view the Model Coefficients. All the model features defined on Variables Type screen can be used in Model Coefficients.

The screenshot above represents the generic formula for Logistic Regression, where P0 represents model Intercept and P_iX_i represents combination of Model Coefficient and Model feature respectively.

You can load all defined model features using Load Defined Variables link or choose one at a time from the list. You can specify the respective coefficients of the feature under coefficients column (P_O-P₈). If you are aware of the number of model variables to be used in model, specify the number in formula (above Sigma symbol) and those many rows will be automatically loaded on screen. In each row, you must specify the coefficient next to Pi and respective feature next to Xi, where i is the row number. Apart from features, a combination of continuous model features i.e. Interaction terms can also be defined as X_i.

Provide values to Probability and proceed to testing the Model.

Test the Model

Once the model is loaded, you can test the model with Model Test tab. Click Load Model to load and view the model.

Specify value of all model features and perform a single record test. The system will evaluate the model for this input and will show all the output parameters on screen.

Add Notes and save the Configuration.

Regression

Regression analytics processor is used to analyze data through Regression model. Regression analysis is used for estimating the relationship among variables. It helps to identify how the value of dependent variable changes when any one of the independent variable is changed, while other independent variables are fixed. It is used for prediction and forecasting.

Configuring Regression Model

To add Regression model into your pipeline, drag Regression model to the canvas and right click on it to configure.

The configuration settings of Regression model are as follows:

Regression can be processed in two ways:

Import PMML Model

Choose Import Model as Yes and click the Import button to load the PMML file.

Once the model is imported, a message will be displayed validating if the model is valid or not.

Following checks will be performed:

• The imported file must be a valid PMML file and it must have Regression model. If this check fails, then the model will not be loaded in the processor, View Model and Save features are also disabled.

• The features and output defined in the model must be present in the selected message with the data type expected by model. If the message does not have certain attributes defined in model, then the View Model feature will be enabled but the model still cannot be saved. The error message in this case will explain which fields need to be defined.

When you click on next, Variable Type Tab opens the Variables as shown below:

Since PMML model is imported, the Variables cannot be edited.

Note: If there is any Variable Type that is not defined in the message, the variables will appear in red.

After viewing the Variable Type, click Next and Model Coefficients page will open.

Click on Load Defined Variables to load the variables and provide values against variable

Now, test your Model with Values.

After the Model is tested, provide your notes in the Add Notes section and save the configuration.

Other option is to create your own PMML model.

Create PMML Model

If you want to create a PMML Regression model, choose Import PMML as No.

Variable Type

Variable Type and Model Coefficients tabs are enabled.

You must select the message field this variable corresponds to, along with the possible categories.

You can also use upload CSV option to populate categories for categorical variables under Categorical Variables via Add Variables, as shown below:

Model Coefficients

All the model features defined on Variable Type can be used on Model Coefficients page. The screenshot below represents the generic formula for Regression.

When you click on Next, you can test your model with values.

Once you have tested the mode, provide your notes in the Add Notes section and save the configuration.

ClusterModel

Cluster Model Analytics processor is used to analyze data through Cluster model. It is commonly known as data clustering.

Data clustering is the task of diving a dataset into subset of similar items. Applying data clustering to a dataset generates group of similar data items. These groups are called clusters i.e. collection of similar data items.

Data clustering can help you identify, learn or predict the nature of new data items especially how new data can be linked for making predictions.

For example, in pattern recognition, analyzing pattern in the data like buying pattern in particular region or age group can assist you develop predictive analysis.

Configuring Cluster Model

To add ClusterModel into your pipeline, drag ClusterModel to the canvas and right click on it to configure.

ClusterModel can be processed in two ways:

Import an existing model

Choose import model as Yes and click the Import button to load the PMML file.

Once the model is imported, click on Validate Model.

Checks will be performed on the model and if the defined variables does not match with the model then you will receive an error.

Create your own model through Gathr UI

If you have a ClusterModel definition and wish to create the model in Gathr. Choose import model as No and then clustering tab will be enabled.

You can add new cluster by clicking on add cluster link (plus icon)

Test the model

Once the model is loaded (through UI or import) then you can test the model from Model Test screen. Click Load Model to load and view the Model.

Specify value to all model features and perform a single record test through Test Single Record. The system will evaluate the model for this input and will show all the output parameters on screen.

SupportVectorMachine

SupportVectorMachine analytics processor is used to analyze data through SupportVectorMachine model. SupportVectorMachine is a machine-learning algorithm that can be used for both classification and regression challenges. However, it is mostly used in classification problems.

Configuring SupportVectorMachine Model

To add SupportVectorMachine model into your pipeline, drag SupportVectorMachine model to the canvas and right click on it to configure.

SVM Processor can be processed in the following way:

Import PMML Model

If you have a PMML file representing SupportVectorMachine then import it. Click on Import button to load the file.

Once the model is imported, model validation with variable checks will be performed.

The next tab that is enabled is Test Model.

Test Model

You can test the model by clicking on Model Test Tab. Click Load Model to load and view the Model.

Specify value to all model features and perform a single record test through Test Single Record. The system will evaluate the model for this input and will show all the output parameters on screen.

NaiveBayes

NaiveBayes analytics processor is used to analyze data through NaïveBayes model. NaiveBayes model is easy to build and particularly useful for large data sets.

NaiveBayes classifier assumes that the presence of a particular feature in a class is unrelated to the presence of any other feature.

For example, a fruit may be considered an apple if it is red, round, and about 4 inches in diameter. Even if these features are dependent on each other or upon the existence of the other features, all of these attributes individually, contribute to the probability that this fruit is an apple and that is why it is known as ‘Naive’.

A Naive Bayesian model is simple to build, with no complex iterative parameter estimation that makes it specifically useful for very large datasets.

Input to this processor can only be provided as:

1. If you have a PMML file representing NaiveBayes, then you can directly import it. To do so click on Import button to load the file.

Once the model is imported, it should be validated. Following checks will be performed:

a) The imported file must be a valid PMML file and it must have NaiveBayes model. If this check fails, then the model will not be loaded in the processor and Save feature will be disabled.

b) The features and output defined in the model must be present in the selected message with the data type expected by model.

If the message does not have certain attributes defined in model then model cannot be saved, since without these features in message you will not able to execute the processor. The error message in this case will explain which fields needs to be defined.

Save the model once, the checks are met.

2. You can Download PMML file that you have imported.

3. Once the model is loaded, test the model using Model Test tab.

Test Model

Click on the Model Test tab. Click on Load Model button for loading the model fields.

Specify value for all the model features and perform a single record test by clicking on TEST SINGLE RECORD button. The system will evaluate the model for this input and will show all the output parameters on screen.

Add notes and save the configuration.

Ensemble

This analytics processor is used to analyze data through Ensemble model. Ensemble modeling is the process of running two or more related but different analytical models and then combining the results into a single score, which helps in improving the accuracy of predictive analytics and data mining applications.

Ensemble modeling offers one of the most convincing way to build highly accurate predictive models. Ensemble model combines multiple models together and delivers superior prediction power.

Configuring Ensemble Model

To add Ensemble model into your pipeline, drag Ensemble model to the canvas and right click on it to configure.

Input to this processor can only be provided as:

If you have a PMML file representing Ensemble, then you can import it. Click on Import button to load the file.

Once the model is imported, it is validated. Following checks will be performed:

The imported file must be a valid PMML file and it must have Ensemble model. If this check fails, then the model will not be loaded in the processor and Save features will be disabled.

The features and output defined in the model must be present in the selected message with the data type expected by model. In case the message does not have certain attributes defined in model, the error message will explain you, which fields needs to be defined.

If above checks are met, then model can be saved for execution.

Once the model is loaded, you can test the model by clicking on Model Test tab.

If you click Load Model link, model features will appear on screen.

Add notes and save the configuration.

Neural Network

This analytics processor is used to analyze data through Neural Network model. A neural network is a powerful computational data model that is able to capture, represent complex input and output relationships.

Neural Networks are widely used for data classification, process past and current data to estimate future values.

Configuring Neural Network Model

To add Neural Network model into your pipeline, drag Neural Network model to the canvas and right click on it to configure.

Input to this processor can only be provided as:

• If you have a PMML file, representing Neural Network then you can directly import it. Click the Import button to load the file.

Once the model is imported, it shall be validated.

Once it is validated, you can click on View Model.

The Neural Network tab shows the graphical representation of the model. If you click on any weight or edge, you will be able to view the corresponding data.

We are able to visualize the Neutral Network. Weight specific to every edge is highlighted when you click on it.

Test Model

You can test the model by clicking on Model Test Tab. Click on Load Model button for loading the model fields.

Specify value for all the model features and perform a single record test by clicking on TEST SINGLE RECORD button. The system will evaluate the model for this input and will show all the output parameters on screen.

Tree Model

This analytics processor is used to analyze data through Tree model. Tree based learning algorithms are considered to be one of the best learning methods.

This model provides high accuracy, stability and ease of interpretation. They are adaptable at solving either classification or regression problems.

Configuring Tree Model

To add Tree Model into your pipeline, drag Tree Model to the canvas and right click on it to configure.

Input to this processor can only be provided as:

1. If you have a PMML file, representing Tree then you can directly import it. To do so you must choose Import Model as Yes and then click the Import button to load the file.

Once the model is imported, it is validated.

Tree Model tab is enabled and you can view Tree diagram of the model, as shown below.

After the Tree Model view, you can then test the Model.

Test Model

You can test the model by clicking on Model Test tab.

Specify value for all the model features and perform a single record test by clicking on TEST SINGLE RECORD button. The system will evaluate the model for this input and will show all the output parameters on screen.

H2O

Algorithms Supported

Following is list of Algorithms of which PoJos created in H2O as models are supported for scoring over Gathr:

• Deep Learning

• Distributed Random Forest

• Gradient Boosting Machine

• K-Means

• Generalized Linear Modeling

• Naive Bayes

Scikit

Scikit-learn provides a range of supervised and unsupervised learning algorithms via a consistent interface in Python.

The Algorithm types supported by Scikit are:

• Regression

• Classification

• Clustering

• Pipeline

Score the Model

To use a Scikit model for scoring, drag Scikit processor from analytics section on the pipeline canvas and right click on it for further configuration.

The configuration of SCIKIT is explained below:

Score the Model Using H2O

For scoring using H2O model, drag H2O processor from analytics section on the pipeline canvas and right click on it for further configurations.

Configuring H2O Processor:

Models

Models Page lists all the models which are trained through Gathr and models registered in Projects.

The Models home page consists of the list of models and brief of their configuration.

Model Version

On clicking the Model Version link, the model version details are displayed.

Once you land on the model versions page, following are the properties displayed:

Create Model Version

You can create different versions of the H2O MOJO model and use them in prediction pipelines.

To do so, enter the Models page, click on the View model versions (eye icon) under the Actions column. Here, you can view the existing versions of the model.

Click (+)Create Version button on the top of the screen.

The Create Version window pops up. Choose one of the model types:

• Distributed Random Forest

• Gradient Boosting Machine

• Generalized Linear Modeling

• Isolation Forest

There are two options to select the model source. These are:

• Upload local zip file

• Mention the HDFS connection and zip file location on the HDFS server

After mentioning the model source click Validate.

Once the model is successfully validated, click create for version creation.

Click on the link under the Active column to activate the model version. You can now use this version in the prediction pipeline.

View Model

To open a model’s version, you can click on the Open icon, under Actions, as shown below:

When you open the model version, depending on the Model Type-Regression or Classification (-Binary and Multi-class), the following properties are defined:

Not all Classification models will have all of the above mentioned properties but a few and same goes for Regression. Below explained is each property and the Model- Type under which it will be show.

Model Configuration

Model Configuration lists the model configuration’s parameters (Key) and Values.

This tab is common for Classification and Regression models.

Model Details

This tab enables you to visualize th Model. Depending on the type of models, below tabs are shown:

Pipeline Stages: The algorithm stages of the pipeline.

Intercept: The intercept is the expected mean value of Y when all x=0.

Coefficients: The coefficient for a feature represents the change in the mean response associated with a change in that feature, while the other features in the model are held constant. The sign of the coefficient indicates the direction of the relationship between the feature and the response.

Note: For Isotonic Regression, Naive Bayes, Model Details page shows only Pipeline Stages. Intercept and Coefficients are available for Linear and Logistic Regression Model.

For Tree Based Models:

Feature Importance

This graph shows the estimation of importance of each feature used to train the model. Y-axis shows the feature names and X-axis shows the feature importance in percentage.

Metrics

The metrics window will display all the performance indicators of the trained model.

For Isotonic Regression and Linear Regression Model, Evaluation Metrics are also generated, as shown below:

For Logistic Regression Model, Naive Bayes and Tree Based Models, performance indicators on which this classification model is evaluated such as Area under ROC, Area Under PR, Precision, Recall, Accuracy, FMeasure are generated, as shown below:

Confusion Matrix

A confusion matrix is a table that is often used to describe the performance of a classification model (or “classifier”) on a set of test data for which the true values are known. Each row of the matrix represents the instances in an Actual class while each column represents the instances in an Predicted class.

Terms associated with Confusion matrix:

True Positives (TP): Number of instances where the model correctly predicts the positive class.

True Negatives (TN): Number of instances where the model correctly predicts the negative class.

False Positives (FP): Number of instances where the model incorrectly predicts the positive class.

False Negative (FN): Number of instances where model incorrectly predicts the negative class.

Advanced metrics

Recall, Precision, Specificity and Accuracy are calculated from the confusion matrix

Recall – TP/(TP + FN)

Precision – TP / (TP + FP)

Specificity – TN / (TN + FP)

Accuracy – TP + FN / (TP + FP + TN + FN)

Precision Recall/ROC

ROC Curve:

ROC curves summarize the trade-off between the true positive rate and false positive rate for a predictive model using different probability thresholds.

X-axis shows false positive rate (False Positive/ (False Positive + True Negative)

Y-axis shows true positive rate (True Positive/ (True Positive + False Negative)

ROC curves are appropriate when the observations are balanced between each class.

Precision/Recall Curve:

Precision-Recall curves summarize the trade-off between the true positive rate (i.e. Recall) and the positive predictive value for a predictive model using different probability thresholds.

X-axis shows recall (True Positive/(True Positive + False Negative)

Y-axis shows precision (True Positive/(True Positive + False Positive)

Precision-recall curves are appropriate for imbalanced datasets.

Decision Chart

You can generate the Decision charts for binary classification models to understand a clear picture on the performance of the model.

The decision chart is a plot which is made by varying threshold and computing different values of the precision, recall and FMeasure scores.

Density Chart

Plots the probability vs probability density of different Classes. The minimum the overlapping area between the two curves, the better the model is.

Actual vs Predicted

Line Chart:

Two lines plotted on the graph i.e. one for the actual data and another one for the predicted data. This graph provides us information about how accurate model predictions are. All the predicted data points should overlap the predicted data points.

Scatter Plot:

This graph is plotted between the actual and predicted variables.

The regression line represents the linear line learned by the model.

Cumulative Gain Charts

Cumulative Gain charts are used to evaluate performance of classification model. They measure how much better one can expect to do with the predictive model comparing without a model.

X-axis shows the estimated probability in descending order, splits into ten deciles.

Y-axis shows the percentage of cumulative positive responses in each decile i.e. cumulative positive cases captured in each decile divided by total number of positive cases.

Green dotted line denotes the random model.

Blue line is for the predictive model

Residuals

Residual error is the difference between the actual value and the predicted values for the output variable.

Shown below are the two plotted graphs to visualize the residuals error for the training model. In the first graph, a line chart is plotted between the residual error on the y axis and the count of the test data rows on the x axis. In the second plot, a histogram is made where the residuals are plotted on the x axis and the record count on the y axis. Histogram helps in providing intuition about how many records are having a particular range of residuals error:

Model Deployment as Rest Service

Model as Service (H2O MOJO, Scikit, Spark)

We can now deploy (H2O MOJO, Scikit, Spark) models as REST endpoints on Gathr. In the models page, under the Actions column click on the eye icon to view model versions. You will be redirected to the models version page. Under the Actions column, click the Deploy as service option. Select the Local option. As you click Local, the window appears with fields of model name, version number and deployment port. Mention the port where the model will be deployed. Click Deploy.

Once the model is successfully deployed, a message will be prompted on screen with the end point URL that can be copied on the clipboard.

The deployment indicator will also turn green specifying that the model has been locally deployed. As you hover over the button, you can see the end point URL of the locally deployed model.

Under the Actions column, the two new options that get visible are:

1.Terminate model service

2.Test model

Click the Test model. You may add headers in the test request.

In this window, an editable sample request is provided containing the features list for this model.

If you want to test a single record, mention the feature values in the request body or else you may upload a csv file to test multiple records.

The uploaded file must contain the header row to map the feature values.

Click Send.

Terminate: You can terminate the model service deployed on the local server by clicking on the terminate button under the Actions column.

Note – To deploy H2O MOJO model as REST end point, make sure the H2O server is running. To start the embedded H2O server, please refer the installation guide.

Data Drift

After a model is deployed into production, the statistical properties of the data may change over time in unpredicted ways. Due to this, the predictions become less accurate. In Gathr, the user can monitor the data drift patterns in the deployed model on regular intervals.

Note: Data drift feature is supported only for the H2O models.

Enable Drift Detection

In the models page under actions column, click on the view model versions eye icon.

Under the Actions column, click on the Enable Drift Detection icon to configure drift detection. Upon clicking the icon, the Drift Detection configuration window pops up.

You may select the data source from which the model was trained from the below mentioned options:

1. Use Existing Dataset

2. Upload Sample Data

Note: If you are using the existing dataset its profile must be successfully run on the respective dataset. If you are uploading the sample data, the accuracy won’t be high as data drift would be calculated on the sample data.

Select the Dataset and choose its version if any. Click OK.

Note: To enable drift detection on a H2O model, make sure the H2O server is running. To start the embedded H2O server, please refer the installation guide.

View Data Drift

To view the data drift for the last 7 days, click on View Data Drift icon under the Actions column.

In the data drift window, choose the pipeline for which you want to see the data drift. You can select the features for which you want to view the initial data stats, current data stats, Mean Drift% and IQR Drift%

The View Trends depicts the trend of data drift for the selected features:

Data Drift Detector Configuration

To configure data drift with H20 (MOJO) model create a data pipeline with a source, H20 processor and data drift detector processor. Data drift detector processor is available under Analytics within the components palette of the data pipeline canvas.

Note: The data drift feature for streaming use case will not be supported where Spark version is less than 2.4.