Policies are the “rules/Guidelines” you can write to override a model prediction. This can be done by:

Policies (Main menu on the left) > Create Policy

Define the policy and define the feature (data point on which you want to write the policy on). Select the conditional operators (Viz. not equal to, equal to, greater than, less than) and current expression. 

Add the policy statement, select the input under ‘Decision’ and mention the decision value you want in the final prediction, and select ‘Save’.

All policies are displayed on the Policy dashboard. You can easily Activate/ Deactivate, edit or delete the policies from here.

When viewing cases, the ‘Policies’ tab (ML Explainability > View cases > ‘view’ under the Options column) will display the policy details for the particular case.

Here, ‘Model Prediction’ is the original model prediction and ‘Final prediction’ is the overridden prediction based on the custom rules defined.

Introduction:

'Similar cases' aka reference as explanations is a parallel method of 'citing references' to a prediction. 'Similar cases' extracts the top 15 most similar cases in the training data as compared to the 'prediction case'. The similarity algorithm varies depending on the plan. For the AryaXAI Developer version, we use the 'prediction probability' similarity method whereas for AryaXAI Enterprise, other methods like 'Feature Importance Similarity', and 'Data Similarity' is also available.

View Similar Cases:

This tab displays similar cases from the previous data where the prediction was similar or almost similar. The features are plotted in a graph where you can filter based on the data labels. Below the graph, all similar cases are listed, which can be filtered based on the Feature name. You can also view details of any of the similar cases listed from the ‘view’ option.

‍

‍

Introduction:

'Observations' provides the easiest and most effective way of estimating the correlation of industry knowledge vs model functioning. It allows the subject matter experts to be part of the explainability framework and provides easily understandable explainability notes to all stakeholders.

Creating/Editing Observations:

The ‘Observations’ section explains the reasoning behind the predictions made. If you want to see a causation correlation with the model prediction, you can easily define the conditions/ causes as ‘observations’. To access this, go to ML Explainability > Observations.

To create a new observation, select the ‘Create observation’ button on the right. 

Next, define the observation and define the feature (data point on which you want to write the observation). Select the conditional operators (Viz. not equal to, equal to, greater than, less than) and current expression to add multiple IFTT conditions.

Once the operation is written, link them to engineered features (actual features that are going into the model). You can select multiple features here and write an observation statement. You can call for the data in the observation statement using curly brackets, i.e. {

View observations:

Once saved, if any of the observations hold true for a case, it will be displayed below the case. This can be viewed at ML explainability > View cases > ‘View’ under the ‘Options’ column in the summary table.

Selecting the ‘Advanced view’ option provides additional details on the observations. The ‘Success’ column here displays whether the particular observation is running on the. ‘Triggered’ will show if the observation is relevant to the current case.

Introduction:

Feature importance is one of the standard ways to explain a model. It provides a very high-level overview of how the features are used by the model to arrive at the prediction output and also debug the issues in the model.

AryaXAI uses the feature defined in the data settings and creates an XAI model to derive feature importance. In developer version of the product, it uses the first file that is uploaded as the training data to build the XAI model. In the AryaXAI enterprise version, it can use the model directly to build the XAI model.

Data settings:

When you are defining the data settings, ensure that the features are matching with the features you actually used in your model for higher accuracy of explainability. These final features are mandatory in any new file uploaded in that project.

• ‍Model Type: Select the model type classification/regression‍
• UID: define the variable to be used as UID. This will be used to identify duplicate cases‍
• True value: Select the true value variable in the data‍
• Predicted Value: Select the predicted value variable in the data. If the true value is missing, AryaXAI will use the 'true' value to build the XAI model‍
• Feature to exclude: Exclude all the features that are not used in your modelling or are irrelevant to your model‍
• Exclude other UID: If there is any other UID, you can exclude them by checking this option

AryaXAI will deploy AutoML and builds an XAI model based on these settings, which will be used for deriving feature importance.

Here, you can view the Global explainability, observations and Case view. 

Note: When defining data features, specifically the data settings, it should be noted that these become the base for explainable model training. The feature selection that is done here should align with the final features that have been used in the model. 

‍

Local explanations: Case-wise

The ‘View cases’ tab displays all your data points. You can also filter among the data points using a Unique identifier, the data upload dates or the data tag. For in-depth insights into a particular case, click on ‘View’ under the ‘Options’ column in the cases summary table. This will lead you to the case view dashboard.

Feature Importance

Selecting the ‘View’ option for a particular case provides a complete overview of the parameters AryaXAI is using for explainability. You can view the local features and the feature importance plot.

The feature importance plot displays the top 20 features, and you can select the ‘Show more’ tab to view all the features in your data that positively and negatively impact the prediction.  

Note: The surrogate XAI model (parallel model) uses the features (or variables) configured in data settings to build the explainability model.

‍

Raw data

The ‘Raw data’ tab displays the details of the data uploaded, where you can verify if the data upload was done correctly.

Global Feature importance

This global feature importance dashboard displays the aggregation of features and feature importance across all the baseline data. 

Retraining the XAI model

To retrain the explainability model, you can simply modify the data settings. This can be done by selecting the ‘Update config’ option present in ‘Data settings’. Whenever the settings are modified, the explainability model is retrained again. XAI model can be retrained as many times as needed to achieve the best correlation between model prediction and the model functioning.

‍

Introduction:

ML Explainability allows you to explain your models using multiple ways - feature importance, observations & similar cases. To know more about these techniques, please go through our 'Resources' section.

Explainable AI has been taking a lot of attention in recent times, while AI adoption is increasing. The sophistication and complexity of AI systems have evolved to an extent that they are difficult for humans to comprehend. There is a commonality between the complexity of the model and the complexity of explanations - the more complex a model is, the tougher it is to explain.

From a regulatory standpoint, it becomes imperative to understand how the model reached a particular decision, and whether or not the decision was a result of any bias. It's not just from a regulation standpoint, but also from a fundamental model-building standpoint, especially for the ML team- if one can understand how the model is functioning and how it is working behind the scenes, it gets easier to find ways for improving the model.

Whereas from a business user standpoint, there needs to be confidence in the system that can be built by providing a clear understanding of how the model is working, and what would be the scope boundaries for the model. There is also a need to validate the product before it can be used in production.

Basic concepts: 

• ‍Baseline: Users can define the baseline basis on ‘Tag’ or segment of data based on ‘date’
• ‍Frequency: Users can define how frequently they want to calculate the monitoring metrics
• ‍Alerts frequency: Users can configure how frequently they want to be notified about the alerts

‍

Model performance

The model performance dashboard lets you can analyze your model's performance over time or between model versions. This analysis is displayed across various parameters for the predicted and actual performance.

The model performance report will display various metrics like accuracy, precision, recall and also quality metrics.

Alerts and Monitors

From here you can easily create and view customized alerts for Data drift, target drift and model performance through the alerts dashboard. For this, select ‘Create alerts' in the 'Monitors' tab and define the baseline and current data parameters like we did above and set the frequency of alerts, which can be daily, weekly, monthly, quarterly or yearly. 

To create new alerts, go to:

 ML Monitoring (Main menu on left) > select ‘Monitoring’ (from the sub-tabs) > click ‘Create Alerts’

All the newly created and existing alerts are displayed on this dashboard, along with details of trigger creator, name, type and options. 

‍

Model Performance Monitor:

Be informed about your model performance proactively using 'monitors'.

• ‍Select model type: Classification/Regression. ‍
• Select model performance metrics: You can define any of the following performance metrics - accuracy, f1, Auc-roc, precision and recall. ‍
• Select the baseline and current: Use the tags to define the baseline and current. 'Current' is your production data if you are tracking drift in your production data. ‍
• Select predicted &true label: Map the appropriate feature for 'Baseline predicted/true label' & 'Current predicted/true label'. ‍
• Segmenting the baseline or current: You can use date features to further segment your baseline. You can also use 'Time period in days' to dynamically select the recent 'n' days as the current data. If you have added 'Time period in days', it'll use that value as the time period the day it calculated the drift as the end date.

‍

Tip: If you deployed multiple model versions, you can append the model prediction in the same dataset as new features instead of creating duplicate dataset copies. You can use these to track the model performance.

Alert Report

The ‘Alert’ tab (beside the Monitoring sub-tab) displays the list of alerts that have been triggered. Clicking ‘View trigger info’ displays the Trigger details, such as the current data size, data drift triggered, drift percentage, etc.

Notifications

If there is an identified drift, you'll get the alert for the same in both the web app and email at the specified frequency.

Web app alerts: Any alert triggered will be displayed as a notification on the top right corner. You can view all notifications from the tab and clear them. 

Email Alerts: The admin of the workspace will get an email if there is an identified drift.

Basic concepts: 

• ‍Baseline: Users can define the baseline basis on ‘Tag’ or segment of data based on ‘date’.
• ‍Frequency: Users can define how frequently they want to calculate the monitoring metrics
• ‍Alerts frequency: Users can configure how frequently they want to be notified about the alerts

‍

Target Drift

It has similar inputs to 'Data Drift'. But unlike 'Data drift', you need to define the Baseline and Current data parameters, along with the True label, predicted label and the model type.

The dashboard report provides a detailed analysis of the target distribution by feature.

Drift Metrics: These statistical tests are available to analyze data drifts, namely the Chi-square test, Jensen-Shannon distance, Kolmogorov-Smirnov (K-S) test, Population Stability index (Psi), and Z-test.

You can learn more about these tests in our wiki section. 

Selecting dates: If you are selecting the dates, then the entire data under that tag will be used for calculating the drift. When you are selecting the date variable, ensure that there is data within these dates.

Mixing multiple tags: If you want to merge data from different tags, you can simply select multiple tags in the segment(baseline/current).

Alerts and Monitors

From here you can easily create and view customized alerts for Data drift, target drift and model performance through the alerts dashboard. For this, select ‘Create alerts' in the 'Monitors' tab and define the baseline and current data parameters like we did above and set the frequency of alerts, which can be daily, weekly, monthly, quarterly or yearly. 

To create new alerts, go to:

 ML Monitoring (Main menu on left) > select ‘Monitoring’ (from the sub-tabs) > click ‘Create Alerts’

All the newly created and existing alerts are displayed on this dashboard, along with details of trigger creator, name, type and options. 

‍

Target Drift Monitor

You can not only track target drift but can get notified if there is an identified drift in your data. To set up a 'Target drift', select 'Target drift' under 'Monitor type', post which you can the specific details.

• Select model type: Classification/Regression.
• Select drift calculation metrics: You can choose drift calculation metrics from the list provided, and set the threshold for data drift and dataset drift (when the dataset itself was drifting).
• Select the baseline and current: Use the tags to define the baseline and current. 'Current' is your production data if you are tracking drift in your production data.
• Select Baseline/Current true label: Map the appropriate feature for 'Baseline true label' & 'Current true label'.
• Segmenting the baseline or current: You can use date features to further segment your baseline. You can also 'Time period in days' to dynamically select the recent 'n' days as the current data. If you have added 'Time period in days', it'll use that value as the time period the day it calculated the drift as the end date.

Alert Report

The ‘Alert’ tab (beside the Monitoring sub-tab) displays the list of alerts that have been triggered. Clicking ‘View trigger info’ displays the Trigger details, such as the current data size, data drift triggered, drift percentage, etc.

Notifications

If there is an identified drift, you'll get the alert for the same in both the web app and email at the specified frequency.

Web app alerts: Any alert triggered will be displayed as a notification on the top right corner. You can view all notifications from the tab and clear them. 

Email Alerts: The admin of the workspace will get an email if there is an identified drift.

Basic concepts: 

• ‍Baseline: Users can define the baseline basis on ‘Tag’ or segment of data based on ‘date’.
• ‍Frequency: Users can define how frequently they want to calculate the monitoring metrics
• ‍Alerts frequency: Users can configure how frequently they want to be notified about the alerts

‍

Data Drift

You can monitor your models for data drift by using AryaXAI ML monitoring. You can create a new dashboard by defining the Baseline $ Current metrics, pick the statistical method to calculate drift and customize thresholds if needed. Your dashboard will get generated. You can create or modify the dashboard any number of times.

Drift Metrics: These statistical tests are available to analyze data drifts, namely the Chi-square test, Jensen-Shannon distance, Kolmogorov-Smirnov (K-S) test, Kullback-Leibler Divergence, Population Stability index (Psi), Wasserstein distance and Z-test.

You can learn more about these tests in our wiki section. 

Selecting dates: If you are selecting the dates, then the entire data under that tag will be used for calculating the drift. When you are selecting the date variable, ensure that there is data within these dates.

Mixing multiple tags: If you want to merge data from different tags, you can simply select multiple tags in the segment(baseline/current).

‍

Tip: If you want to see the drift in only one feature, you can simply select that feature under 'Features to select' and calculate the drift.

Alerts and Monitors

From here you can easily create and view customized alerts for Data drift, target drift and model performance through the alerts dashboard. For this, select ‘Create alerts' in the 'Monitors' tab and define the baseline and current data parameters like we did above and set the frequency of alerts, which can be daily, weekly, monthly, quarterly or yearly. 

To create new alerts, go to:

 ML Monitoring (Main menu on left) > select ‘Monitoring’ (from the sub-tabs) > click ‘Create Alerts’

All the newly created and existing alerts are displayed on this dashboard, along with details of trigger creator, name, type and options. 

‍

Data Drift Monitors:

You can not only track drift but can get notified if there is an identified drift in your data. To set up a 'Data drift', select 'Data drift' under 'Monitor type', post which you can the specific details.

Select drift calculation metrics: You can choose drift calculation metrics from the list provided, and set the threshold for data drift and dataset drift (when the dataset itself was drifting).

Select the baseline and current: Use the tags to define the baseline and current. 'Current' is your production data if you are tracking drift in your production data.

Select features: You can either create 1 monitor to track all the features in your data or you can select the specific feature for which you are tracking the drift.

Segmenting the baseline or current: You can use date features to further segment your baseline. You can also 'Time period in days' to dynamically select the recent 'n' days as the current data. If you have added 'Time period in days', it'll use that value as the time period the day it calculated the drift as the end date.

Alert Report

The ‘Alert’ tab (beside the Monitoring sub-tab) displays the list of alerts that have been triggered. Clicking ‘View trigger info’ displays the Trigger details, such as the current data size, data drift triggered, drift percentage, etc.

Notifications

If there is an identified drift, you'll get the alert for the same in both the web app and email at the specified frequency.

Web app alerts: Any alert triggered will be displayed as a notification on the top right corner. You can view all notifications from the tab and clear them. 

Email Alerts: The admin of the workspace will get an email if there is an identified drift.

ML monitoring in machine learning is the method of tracking the performance metrics of a model from development to production. Monitoring encompasses establishing alerts on key model performance metrics such as accuracy and drift.  Initially, the success of such ML projects was dependent on successful model deployment. However, it is important to note that Machine learning models are dynamic in nature - their performance needs to be monitored, or it degrades over time.

ML monitoring helps identify precisely when the model performance starts diminishing, and you can proactively work on resolving it quickly. Monitoring the automated workflows helps to maintain the required accuracy and keeps transformations error-free.

Basic concepts: 

• Baseline: Users can define the baseline basis on ‘Tag’ or segment of data based on ‘date’. 
• Frequency: Users can define how frequently they want to calculate the monitoring metrics. 
• Alerts frequency: Users can configure how frequently they want to be notified about the alerts. 

‍

Data addition from GUI

Upon accessing the project dashboard, the first thing you need to do is to upload the data. This can be the data used for training, testing, validation, production data or any other data that you used in your project. 

While uploading the data for the first time (even if you want to use an API), you must first upload at least one sample data from the dashboard and define the data settings. The rest of the data can also be uploaded through the API. 

‍

To start with this, select ‘Upload Data File’, which directs you to the data settings page.  Select ‘Upload file’. 

Here, to classify your data, you will see the ‘Upload Type’ dropdown, where you can specify the data type to 'Data' or 'Data description'. Next, you will see the ‘Upload Tag’ dropdown, where you can specify the data type - Training, Testing, Validation, or you can choose to add a custom tag as well. 

Select the file to be uploaded.

‍Note: You can only upload one file at a time, and the file can only be in CSV format.

‍

Once the upload is complete, you will be directed to ‘Project Config’ to configure the details.  

Data addition from API

First, Get the API Token for the Project. This is accessible at Workspace > Projects > Documentation. 

The project token (and Client Id) is accessible only to the Admins of the particular project. You can refresh your API token through the ‘Refresh token’ option provided beside the Client Id.

Below this, the project URL for uploading the data is displayed. The Python script is present, which can be used directly in your compiler.

The header XAI token needs to be defined, whereas the Client Id and project name are automatically defined. 

 
headers = 
{    
"x-api-token": your project access secret token 
} 
base_format = 
{
"client_id": test_user_arya,
"project_name": Risk-monitoring_FW6FSKJQRE
}
 

Next, prepare Data in Format of Dictionary (you can upload multiple data points in the list of dictionary format)

Define the unique identifier for the data:

 
 "unique_identifier": 
 

‍

A single data set can be passed in string format. If multiple data sets are uploaded, you need to pass a list of unique identifiers. 

Similarly, a single data point (with one unique identifier and 3-4 columns) can be directly passed through the API. However, for multiple data points, a list of unique identifiers and column needs to be created. 

Finally, you will need to give the post request:

 
resp = requests.post(url,headers= headers json=base_format)

For every post request, data successful responses and acknowledgements are provided, so you are updated on the status.   

‍

Data Settings

1. Select the Project type - which can be a classification or regression problem. 
2. Define the ‘Unique identifier’ - The identifier for every unique data point 
3. Select the true label - The variable target you are trying to predict (Eg. for a data set from Real estate industry, the true label can be ‘Sale price’ of a house)
4. If your data has a predicted label, choose the label from the dropdown (This applies when you already have a model and you only want to evaluate the predictions of your model)
5. Select the features (data points) to be excluded such the XAI model uses the same features used in your model.

There might be multiple features within your project. You can exclude the features that might not be relevant to your project from the ‘Features exclude’ option. You can see all the features included and excluded on the right. 

Note: Your data can have some duplicate unique identifiers, which can be dropped by selecting the checkbox. 

‍

True and Predicated label

The predicted label is required if you want the XAI model to explain model the predictions. If the predicted label is not defined, AryaXAI will pick the true label to build the XAI model.

Once the above steps are completed, select ‘Submit’.

At this point, you can see the overview of the data submitted. The total data volume, Unique features (data points) and alerts are displayed. 

Note: When defining data features, specifically the data settings, it should be noted that these become the base for explainable model training. The feature selection that is done here should align with the final features that have been used in the model. 

‍

The ‘Features’ section displays the data type. The platform starts analyzing the data and creates an explainability model for you.  

Note: Until the XAI model is not trained, the explanations (Feature importance) will show ‘nan’ as the values, you can upload any new file or open case view pages. Once the model is trained, the XAI model results can be seen in all these pages.

‍

Project Summary

Once you submit the Project Config, the ‘Project overview’ dashboard displays the total data volume, Unique features, volume graph, and data summary (Features). 

Overview displays the details of total data volume, Unique features, and alerts. 

Volume displays the volume graph, which provides an overview of the data upload activity over time. You can investigate different parameters and plot the data activity, based on data label, feature name, date (of the feature name or date of creation),  range and plot type,  for writing codes with ease.

Data Summary table

The section provides you with a summary of data features and displays the data type. You can easily navigate to the different data tags by selecting them from the dropdown list on the right. 

Note: If the feature table displays ‘NA’ under ‘Feature importance’, it means that the particular feature is not used in the explainability model. (This setting comes from project config. Mentioned in data settings).

Note: Using the ‘Refresh Data’ option will provide the most latest view of the data. The loading time will differ based on the data volume.

‍

Data Setting

Here, the data upload details are displayed, from where you can upload data or delete uploaded data files.

Note: It is important to define the first file that is uploaded. This file which is uploaded in any of the categories (viz. training, testing, validation or custom), becomes the training data for the explainability model.

When defining data features, specifically the data settings, it should be noted that these become the base for explainable model training. The feature selection that is done here should align with the final features that have been used in the model.

Modify Data Settings

To modify the data settings, select ‘Update config’ option present in ‘Data settings’. Whenever the settings are modified, the explainability model is retrained again.

Whenever these data settings are updated, it triggers training for the XAI model.

1. Implementing Arya-xAI on IMAGENET data

We publish white papers from time to time taking. You can find the details of the white paper here:

Product White Paper:

Arya-xAI: Accelerating the path to ML transparency:

We recently published white paper on AI explainability. Get insights on the AI explainability imperative, tangible business benefits of XAI, overview of current XAI methods, their challenges and details on the functioning of Arya-XAI framework.

Industry White Paper

AI Explainability Framework in Financial Services: The Trust Imperative

This white paper explores the current AI adoption in financial services, the ‘black box’ problem with AI and how explainability helps resolve the trade-off between accuracy, automation and being compliant.

Creating Workspace and project

Workspace

On accessing the platform, you can either set up a new workspace or access an existing workspace.

Creating a workspace

To set up a workspace, select ‘Add workspace’. Define a name for the workspace and submit. The workspace will be visible on the dashboard with details of the workspace owner, creation date and time.

You can also invite/add users in the workspace through the ‘Add User’ option and define their role for the workspace. The role can either be of an Owner, User or Manager. Each role has specific use access criteria as defined below. You can also revoke access through the workspace settings or modify the user role.

User role wise accessibility criteria:

Accessing an existing workspace

To access an already created workspace, the workspace owner has to send an invite through the ‘Add User’ option. This invitation to the workspace is shared to the invitee's email address provided.  

The ‘Settings’ option under Actions column displays the workspace details. Here, the 'User list' tab displays the user's list for the particular workspace. This list shows all the users you have invited to the workspace and the status of their invitations. The 'Usage' tab shows the usage details, such as, User subscription details, number of projects, users, data points, etc. and the usage cap on each.

Start using workspace

To go into the workspace, select ‘Show’ under the ‘Actions’ column. You can create multiple projects within a workspace.

Project

As mentioned earlier, your workspace can have multiple projects. For example, if you created a fraud detection model, that can be one project. 

Creating a project

To create a new project, select the ‘Add project’ option within the desired workspace. Define a name for the project and submit it. The project details will be visible on the dashboard with details of the project owner, creation date and time.

‍

Note: If you want a user to have access to a particular project and not the entire workspace, you can select the ‘Add user’ option under ‘Actions’. 

‍

You can define the roles of new users as an Owner/ Admin, User or Manager.

Note: Workspace access overrides the project access. For e.g. If you have given a user ‘Owner’ access at the Workspace level and a ‘User’ access at the project level, the Owner access overrides the User access. 

‍

User role wise accessibility criteria:

‍

‍

Welcome to AryaXAI!

The ML Observability platform for mission-critical AI solutions

Introducing AryaXAI

With AryaXAI, Data science and ML teams can monitor their models in production as well as gain reliable & accurate explainability,

AryaXAI offers reliable & accurate explainability, offering evidence that can support regulatory diligence, managing AI uncertainty by providing advanced policy controls and ensuring consistency in production by monitoring data or model drift and alerting users with root cause analysis.

AryaXAI also acts as a common workflow and provides insights acceptable by all stakeholders - Data Science, IT, Risk, Operations and compliance teams, making the rollout and maintenance of AI/ML models seamless and clutter-free.

‍

How to signup for AryaXAI

In just a few easy steps, users can sign up for AryaXAI with an invitation.

To get started, users who already have access to AryaXAI can invite others to sign up for the platform. This invitation is sent to the email address of the invitee.

From here, you can set up your account in 2 easy steps:

• 1. Setting your basic profile details and password
• 2. Setting your work profile and  industry

Once these steps are completed, select ‘Finish’ to complete the account verification. A confirmation of account verification is sent to your inbox, through which your workspace can be accessed.

AI and ML technologies have found their way into core processes of industries like financial services, healthcare, education, etc. Even with multiple use cases already in play, the opportunities with AI are unparalleled and its potential is far from exhausted.

However, with increasing use of AI and ML among AI-driven organizations, ML engineers and decision makers who rely on AI outcomes, are now faced with explaining and justifying the decisions by AI models. Decisions have already been made, with the formation of various regulatory compliance and accountability systems, legal frameworks, requirements of Ethics and Trustworthiness. Ultimately, an AI model will be deemed trustworthy only if its decisions are explainable, comprehensible and reliable.

Today, multiple methods make it possible to understand these complex systems, but they come with several challenges to be considered.

While ‘intelligence’ is the primary deliverable of AI, ‘Explainability’ has become the fundamental need of a product. It helps to serve important purposes like:

• Accountability
• Trust and transparency
• Better Model Pruning
• Better AI controls

Arya.ai has innovated a state-of-the-art framework, ‘AryaXAI’ to offer transparency, control and Interpretability on Deep learning models. In this documentation, we will explores the explainability imperative, tangible business benefits of XAI, overview and challenges with current methods, and details on the functioning of AryaXAI framework.