Welcome to the latest One Model product release update. This article provides an overview of the product innovations and improvements to be delivered on 6 April 2022. You can see the full 2022 release schedule here: https://help.onemodel.co/en/articles/3134972-one-model-product-release-calendar.
User Experience
Deep linking from One Model to other systems in your ecosystem
An exciting new feature we are delivering in this release is the ability to configure specific data elements within One Model to include a dynamic hyperlink. The new deep linking feature is available within Storyboard tables and in the drill-through table from charts and tables within Storyboards.
Example uses for this capability we have heard from customers is to be able to click on employee names and link back to the HRIS and/or Talent apps to an individual Employee Profile, Compensation Form, Learning Goals, etc. or clicking Hiring Candidates, or Requisitions and opening these in the source Applicant Tracking System. (ref 10765)
Example of a table column that includes an embedded deep link.
Example of a drill-through table that includes an embedded deep link.
Deep link columns are configured within the One Model Table & Column Label Editor screen. [NB: Usually only Admin users have access to this page, which uses the Application Access Permissions “CanEditDataWarehouseTableAndColumnLabels”.]
In addition to the existing option to rename a column label, you can see that there is now a new field where you can define a hyperlink. This is a text field where you can type in the relevant standard link information for the target system and you are able to include within this text reference to another field within the same One Model table. These reference fields should use the original field name and be bound with {curly brackets}. The screenshot below contains a simple example using Google Search as the target system and it is dynamically appending the employee’s first name.
One AI Innovations
This release represents a significant step forward in One AI’s mission to democratize transparent, ethical machine learning in HR & people analytics. With this release, our team has wholly re-hauled the output from One AI. These updates will allow for richer storyboards that help consumers understand the quality, drivers, and methodology used to generate each prediction. Along with the output changes we added probability calibration to our workflow so the probability estimations coming out of One AI will work better in storyboard content.
We have also released a new advanced forecasting capability. Like our existing forecaster, the new advanced forecaster is embedded within Storyboards and represents a foundational move for One AI with even more exciting features coming over the next year. The new Advanced Forecaster brings seasonality and changepoints along with substantial improvements under the hood.
On the algorithm front, we have removed Catboost in favor of LightGBM , another gradient boosting algorithm, and updated our default feature selection method(s). The new feature selection uses wrapper and filter methods to arrive at an optimal and more robust set of features with less user input than our previous version.
What does all this mean? One AI users will have access to more transparent & rich data sets to build storyboards for their stakeholders via One AI outputs. The new forecaster will allow for more accurate forecasts and richer historical analysis via change points and other features. Finally, the default behaviors of One AI’s predictive features will be more stable and robust with less user input.
Detailed Notes
-
One AI V2 Output: After this release goes out, a much richer set of data will be generated and fed back into your One Model instance when you deploy a classification or regression augmentation. This data includes but is not limited to:
-
Estimator details
-
Model performance indicators
-
SHAP
-
Feature and label information
-
Data transformation details
The team at One Model is working on new storyboard content from this data, enabling visibility into your machine learning data that was never possible in the past.
This data will be loaded via a new set of data sources. The new section is called "One AI - Prediction" and will be automatically created upon your first augmentation deployment after the release.
If you have existing deployed classification or regression augmentations, the team at One Model will do our best to proactively update your pipeline script to accommodate the new data. Nothing will break in your One Model instance if this is not done, but the newly deployed data also will not be visible in storyboards. Please reach out to the One Model team if you notice this happening.
If you are interested in learning how the data joins, you can also reach out to the One Model team, and an entity-relationship diagram can be provided.
Here are some screenshots of the new files via data sources and the file download option from the Results summary page:
-
Probability Calibration: In the advanced configuration section, Probability Calibration can now be turned on for augmentations. Probability Calibration will adjust the output probabilities such that they are more evenly distributed into risk buckets in alignment with users’ expectations. The next release will include a GUI configuration that does not involve YML configurations in the advanced section.
-
Example Advanced Configuration
-
Example Output from the Results Summary Report
-
New Feature Selection: Historically, OneAI has supported several different feature selection and dimensionality reduction techniques. Our default approach has been to use different filter methods. With the release of OneAI V2, we are using both filter and wrapper methods by default. A filter method is used to select a maximum possibility of features ( ex: 15 ). Then a wrapper method ( ex: recursive feature elimination ) is used in conjunction with a minimum number of features ( ex: 3 ) to select the optimum number and composition of features between the maximum and the minimum ( 15 - 5 ). This has consistently resulted in better, more stable models with less user configuration. If you do not want to use the new feature selection method, you can always use the edit page's overrides to change these behaviors or manually select the specific features you want to use.
-
Advanced Forecasting: With the release of OneAI V2 we have added a new forecaster embedded within Storyboards. Currently, the feature is behind a flag that needs to be turned on per client so please contact your One Model Customer Success contact if you are interested in giving it a try. The new Advanced Forecaster exists alongside the existing Forecaster as each forecasting type has its own use case. The Advanced Forecaster is meant for forecasts with larger data sets and the potentiality of seasonal trends. The existing Forecaster is intended for smaller data sets that are unlikely to have seasonality. Because the Advanced Forecaster is configuring and measuring much more than the existing Forecaster, it can take a bit longer to run. Both forecasters can be accessed from the lightbulb dropdown on a time series line chart, shown below.
-
Changepoints: With the Advanced Forecaster, you have the option to render changepoints within the chart as well. Changepoints occur in the first 80% of the historical data ( not the forecasted / future data ). Changepoints can help you understand when the Advanced Forecaster thinks meaningful changes in trends occurred over time. We don't display the changepoints by default, but after running the Advanced Forecaster, a new show changepoint option appears in the lightbulb menu, shown below. Clicking on any of the forecasted data points will pop up an explanation of change points as well as a list of the detected changepoints.
-
Variance Inflation Factor: The Variance Inflation Factor ( VIF ) indicates how much each column relates to the other columns in the dataset. The lowest value is 1, while a value of 8 or more indicates a high degree of correlation with another column. Many of the estimators used by OneAI are robust against collinearity. However, the VIF score can be used to understand how much collinearity occurs within a feature set if you need/want to reduce collinearity. The VIF scores have been added to the Feature Analysis section of the Results Summary Report for each augmentation. You can also find some early-stage SHAP graphs in this section.
-
Changes to Supported Algorithms: Given the current state of events in Ukraine, we have deprecated our use and support of Catboost. In Catboost's place, we have implemented LightGBM, which is another stellar gradient boosting algorithm. The Catboost option has been removed from the GUI, and OneAI will produce an error for any persisted models that use Catboost. You can retrain the persisted model using LightGBM if this occurs; however, if you have any trouble or concerns, please reach out to our team, and we will help you with any retraining. LightGBM is included in the default set of estimators that OneAI uses, and it can also be manually configured via the GUI, as shown below. Finally, we've removed Decision Tree Classifiers from the default set of estimators that OneAI uses. After working with many clients, we've found this estimator to be less stable than we want. You can still add it to a run via checking the override button in the estimator configuration section of the edit screen for each augmentation.
Minor Improvements & Bugs Squashed
-
The One Model application is built with responsive design in mind, but we encountered a bug that would cause Storyboards to crash on a mobile device (phone/ tablet). We are looking forward to delivering more improvements to how the application works on a mobile web browser in coming releases. (ref 11059)
Comments
0 comments
Please sign in to leave a comment.