User Guide¶
ML cube Platform is built following the API first principle, by which it can be used both via Web Application and REST API. In this first guide, we explain the basic concepts in ML cube Platform. In order to find easily every information we suggest you to use the search bar, moreover, you can go to Glossary to find the key concepts and definition we use along this site.
Example Company
Every time you find a section like this you will see a examples with a fictional company that will help you to better understand the concepts and the entities in ML cube Platform. We will use a fictional company called Delta Energy that is a producer of Photovoltaic Modules that own Photovoltaic fields and trades the produced energy to the market.
Company¶
To use ML cube Platform you need to belong to a Company that is created during your first login in the Web Application. Everything will be done inside it, for example, if you want your colleagues to work with you in ML cube Platform you need to create new Users inside the company. Users has assigned a specific Role that defines their privileges and what they are able to do (refer to User Roles for more information).
Project¶
Your work on ML cube Platform is organized through Projects, Task and Models. A Project is an artificial intelligence application that uses a set of algorithms to reach a common set of KPIs, usually a Project contains several Tasks.
Delta Energy inc
Delta Energy, created the Energy Revenue Project to enhance their revenue from energy trading. They invested in four AI algorithms:
- Fault detection
- Fault diagnosis
- Soiling detection
- Trading
In ML cube Platform, these four algorithms define four Tasks inside the Project Energy Revenue. They have been placed into the same Project because they share the business goal i.e., the net revenue, the data and are interdependent.
Task¶
As you may have guessed, in ML cube Platform a Task corresponds to an AI Algorithm. To be more precise, a Task is an AI Problem with a dataset composed of input features and a target. A Task can have more than one AI Model that uses the input features estimates the target.
Delta Energy inc
In out example company, the Fault Detection Task has as input features the PV modules and weather data and the target is the presence of a fault. The Task has two Models: logistic regression and random forest. Both models use the same data and predict the same target but are different in the techniques used to perform the estimation.
A Task is specified by several attributes, the most important are:
type: regression, classification, object detection ...data structure: tabular data, image data, ...optional target: if the target is not always available. This happen when input samples are labeled and the most part of production data do not have a labeldata schema: specifies the inputs and the target of the task, see Data Schema section for more detailscost info: information about the economic costs of the error on the target
Delta Energy inc
The Task of Fault Detection has clear costs due to false positives or false negatives. Every time a fault is not detected the false negative cost is proportional to the power reduction. While, false positives determine costs for the maintenance team that will go on the field for nothing.
Model¶
The last entity is the Model that is the actual AI model that predicts the target. A Model has a Version that defines the training data used to train it. All model's data will be uploaded specifying the model version in order to track each prediction with the right model instance. The model version is updated whenever a new training of the model is done.
A Model has the following attributes:
name: uniquely identifies the model inside the Taskversion: specifies different trained instance of the modelmetric name: metric to use to measure the model performancesuggestion type: type of retraining action to use for that model, see Retraining section for more detailsretraining cost: how much model training cost. This information is used by the Business module
Note
It's worth nothing to note that in ML cube Platform you do not actually need to upload the model on the application. We just need to know its training data and its predictions for the production data. In this way, ML cube Platform is considered as model agnostic.
Now that you have clear the basic concepts we invite you to explore the other ML cube Platform pages:
-
Modules
Explore available features in ML cube Platform
-
Integrations
We are part of MLOps ecosystem and natively integrated with other solutions.
-
Automation rules
Discover how to setup automation rules to increase your reactivity.
-
Roles and access
Get more info about RBAC inside ML cube Platform