Analyze

Insights

This module allows domain users to understand machine learning models in detail. Users can interact with the models for better understanding and decide on the appropriate business actions (and sometimes even data or machine learning actions) based on that understanding. Click here to view Model & Insights in detail.

Explore

In Explore tab, you can undertake thorough analysis of patterns and segments in input source with the help of Sunburst chart:

Explore tab will show you the analysis of the latest snapshot of your model. You can always click Show All Models button to select any snapshot and perform analysis on it.

Upon clicking, you will directed to the same window where we have seen multiple snapshots of your model:

Sunburst chart is a visualization type to represent and explain decision trees better. Hence, while we explain the chart, do note that decision tree related terms are used frequently. Now let’s explain sunburst chart further starting from the inner-most circle:

This circle represents the root node of our explanation model which is in fact a decision tree model fitted to our predictions on the actual input source. Total Population means total instances that sit right at the highest level of hierarchy of our explanation model which happens to be the root node.

When you click on the root node you’ll see a small window giving you some statistics of instances pertaining to the children node they are in. Since we are at the root node, it gives information about all of the instances.

Starting from the root node every ring in the chart represents a level in the decision tree. Outer the ring deeper the level of decision tree’s hierarchy instances reside at:

Each outer ring contains the children nodes of the inner ring's nodes. The number of instances captured by a node determine its arc length (or its size in radians). Hence, when you go down the tree, which is equilavent to going from inner to the outer ring in sunburst chart, a new condition is added to the decision path and number of instances belonging to a segment fullfilling the conditions gets smaller. The level we are at gets deepened as well. When there is no outer ring to transition to, we come to a end on our decision path meaning that a decision is made (whether the segment of instances belongs to either class).

The beauty of sunburst chart representation lies at its capability to explain why and when a prediction is made by the particular machine learning model. We can stop at any level in the chart before we get to a leaf node and interpret the information given by small window contaning the useful statistics. Let’s interpret results when we go down the tree from root node to an outer ring:

Following the path IF (job=admin) we have 2,501 instances of 28,831 whose feature job has value "admin". Our model predicts this pattern to be classified as "yes". However, when we look at Train Label Counts we can see that the majority of the instances that meets this condition are in class NO. Average decision score of instances falling under the segment of this path can be spotted as well. This help us understand how confident we are with the prediction instance classes. Now, following the same path from root node, let’s go down one level deeper:

Note that, with the new outer ring added, number of instances drops as the arc length of new ring decreases. Now we have a new segment within a previous segment. 1,114 of 28,831 instances have their feature job value as "admin" as well as marital value as married. This new path creates a refined and a smaller grouping with a decision score of 0.783.

At any level in the sunburst chart when we click on a pattern, a pop-up will open:

Pattern Actions

Copy Pattern: You can copy the pattern (explanation) you have selected to the clickboard to use later in the Search feature.

Plan Business Action: When you click here, you can see preset Business Actions and plan to take any action in the list upon the pattern you have selected:

Show Feature Histograms: You can see feature histograms of sample instances of the selected pattern upon clicking this.

Show Instances: Sample instances can be directly seen and downloaded as json or csv format from here:

Now, let’s explain other components of Explore tab further:

Filter: You can use the Filtering & Export option to see and dowload a report of instances meeting a certain criteria:

You can select the minimum number of instances to be in all of the nodes in Sunburst chart along with the minimum average decision score (confidence) they need to have. Class value can also be added to the filter criteria. When you hit Filter, Sunburst chart will be updated according to the filtered instances and paths. There is also an option to Export explanations to a csv file.

Search: Search functionality lets you search and show instances by feature or a previously copied pattern. You can search by feature to look for patterns that contain a feature of your choice.

By changing search type to pattern, you can type in a pattern or paste a pattern you have copied before.

Widen the model:

Feature Importances: Here, you can see how your model determines which features are important for you.

Model Insights

In Benefit Overview, you can see average and label based accuracies with the count numbers of each. In the above image, 2,038 instances out of 3,248 are labelled correctly as "yes" by TAZI. We can also examine that 19,584 instances out of 25,583 are correctly classified as "no". We would also see benefit feature’s accuracy if we chose one while configuring our model.

In Performance Slices, Top K % Table, ROC Curve, Lift Chart, Cumulative Gain Chart Tabs and Confusion Matrix can be further inspected.

After scrolling through performance slices, we head to a section called Most Coverage Nodes:

In a nutshell, the insights given by the explanation model after it has been fitted to our predictions can be seen here. Since explanation model is a decision tree, most decision tree related terms and metrics apply:

Node: This corresponds to a leaf node in decision tree. It shows the final outcome of a decision path. A node can be thought of as the final destination where all the branches taken by the model lead to (which is a decision)

Path: This is simply the way going towards the decision. Decision path or decision rule contains all of the conditions that the instance(s) must meet in order to be labelled as yes or no (or any other class, in general). The features that are important and relevant as a deciding factor in classifying an instance are used in conditions. Decision path starts from the root (start) node and ends up with the leaf (end) node. Root node comprises of all the instances since it is at the highest level of hierarchy. Starting from left to right, with each condition added, we go from more general segments to more specific groupings of instances. All of the paths shown in the path list above lead to yes by default. We can change it to no to see the instances that are labeled as no by clicking no label:

Now, let’s explain what we mean by a decision path or rule with an example. Path from the above path list;

root -> job=technician -> marital=married -> month__apr=0 -> day_of_week=wed

can be easily written as an if clause:

IF (job==technician)
AND (marital == married)
AND (month_apr=00)
AND (day_of_week == wed)

THEN yes

First, we start at the root node of our decision tree. All of the instances in our input source gets segmented based on their feature value of job (our first condition). Further, instances that the feature Channel Name has value technician face another condition (marital == married) while creating another refined segment in itself. Every time, the instances meet a condition their new segment becomes smaller than before as they go through all of the conditions. After the very last condition, decision is made as either yes or no. In our case from the above example yes, indeed.

% of Total: Percentage of instances that are labelled following a particular decision path

Confidence: Average decision score of all of the instances that fall under the segment of the decision path.

Sample Instances: Shows sample instances that are part of a certain segment following the decision path.

Sunburst Chart: Upon clicking, you will be directed to visual representation of the decision path on a sunburst chart. The instances that follow the path will be highlighted:

In Key Findings, nodes with the most coverage in both classes will be visualized in sunburst chart side by side along with the conditions that are met:

In Overall section, percentage of classes can be seen.

Create Test & Fit

Create Test

Here, you can give another input to your model. Then you can get predictions for your new input. There will be no training.

Fit to Explonation Model

It provides you to pass the input data through the model at selected save point. During this operation, train will be disabled, model won't be evolved and system will process all input data in test mode. Hence, you can see where all of instances in input data goes on the model.

Deploy

Export BM & Report

Export BM

It lets users download model and configuration files of the Business Model. A .tgz file will be downloaded. Then you can import this file from import business model section. If you select 'Export IO' in the opening screen, it also downloads DB connection information.

PDF BM Report

When you click Get BM Report a nicely formatted pdf file will be downloaded in which the information related to your Business Model is displayed.

Export Model Outcome

Salesforce

You can export your model outcome to Salesforce.

Download Model

It provides to download the model and output files. If Output is set as file, you can download it. Otherwise it will just download the latest model file.

Export Model as PMML

will be updated

Integrate

REST API

When you want to integrate your business model into other enterprise applications, you can create restful api here. In the opening section, you should give a name to your Restful service and select the appropriate Saved State to use. So your restful service will work based on this model. Now, with one click you have your REST API service ready to use.