Model deployment

PreviousModels and variables statistics NextABM API

Last updated 21 days ago

Model deployment

In ABM you can deploy your model in three ways:

Deploying mode as a web-service
By scoring a new dataset
By generating scoring code and using it to score data in your custom process

Deploying model as a we-service

The model can be deployed as a REST API service – this is the basic method for applications that require near real-time responses.

To deploy the model as an API service, use the Deploy button

We can choose from two implementation methods:

Directly in AMB (ABM account)

If you choose an ABM implementation, the implementation is performed automatically and then you get an example of a query. We use an authorization token from the ABM account

In the case of implementation in Scoring.one, it is required to provide:

Server address
Authorization key (available for download in the Scoring.one account settings)
And the name of the model under which the model will be implemented in Scoring.one

As in the case of the implementation in ABM, we get an example of a query to the model

Scoring a new data

After your predictive model is built, you can calculate scores for the given dataset.

For instance, if the goal of your project was to select the best target group for your marketing campaign, you would like to know the response probability for each customer in your database. In this task, ABM will assign the response probability (calculated based on the predictive model built by ABM) to your customers (identified by an ID variable which must be unique for each customer).

To start the scoring data task:

From the Homepage level, open the selected project by clicking on its name or click the Open project button

Click the Score data button (from the Project’s reports level)

The scoring data task is divided into six steps for classification projects and five steps for approximation projects.

Step 1: Select a file for scoring

Step 2: Input detailed information about the file selected for scoring

Step 3: Select threshold (this step is omitted for approximation projects)

Step 4: Choose the ID variable(s) to which score values will be assigned (e.g. customer’s id)

Step 5: Start the data scoring process

Step 6: Download scoring result

And that’s it!

Step 1: Selecting file for scoring

Remark: It’s crucial that the file for scoring and for model construction have the same structure:

identical number of variables
variables should have the same names and types

You can also add a new file by clicking the Browse for files button and by selecting the file from your computer. Upload the file with the Upload button and wait until it’s sent to the ABM server.

Choose the file by clicking its name and the Next button if you want to proceed to Step 2.

Step 2: Entering file settings

In Step 2 you will be asked to indicate what separators were used in the chosen file, including:

Columns separators. For instance, if columns in CSV file are separated by a comma (e.g. name, email, age), write , in the first field
Decimal separators. For instance, if a decimal character is a dot (e.g. 4.25) in CSV file, then write . in the second field
Text separators. For instance, if a quotation mark is used to indicate the beginning and end of the text (e.g.’Baker Street 221b’), write ‘ in the third field

Click the Next button if you want to proceed to Step 3 or the Back button if you want to change the previous project settings.

Step 3: Selecting threshold (classification projects)

In step 3 you can set the classification threshold for scoring. ABM suggests the optimal threshold based on the final model and the quality measure chosen when building the model. The user can manually change the threshold (for instance based on the model quality measures such as KS Score, Equal TPR TNR Score, Min Distance Score or the values on the ROC curve).

Step 4: Selecting identity variable

In Step 4 you will be asked to indicate the ID variable(s) name(s). The ID variable is a variable with unique values which is used for identification of observations (e.g. customer ID). The target predicted value will be assigned to this variable.

Click the Next button if you want to proceed or the Back button if you want to change the previous project settings.

Step 5: Running data scoring process

You’re almost there. Check if the information provided is correct. Click the Run button to start the scoring process or the Back button if you want to change the scoring process settings.

Step 6: Downloading scoring results

When the scoring process is finished you will see the following screen:

Click the Download button to see the scoring result. For classification projects the output file is a .csv file containing three columns divided by commas:

PositiveTargetProb: the probability of a positive value of the target variable (the situation when the event occurred, e.g. 1, the customer churned)
Identity variable name: name of the chosen ID variable to which the probability will be assigned (e.g. customer ID)
PredictedTarget: the predicted value of the target variable. By default, if the threshold = 0.5, PredictedTarget will be equal to ‘1’ if the PositiveTargetProb >= 0.5, and to ‘0’ otherwise

For approximation projects the output file is a .csv file containing two columns divided by commas:

PredictedTargetValue: the predicted value of your target variable
Identity variable name: name of the chosen ID variable to which the predicted value will be assigned (e.g. product id)

Generating scoring code

You can also download the scoring code for the final model. The scoring code is a code (in case of ABM: SQL or Java code) that given the input variables, calculates the model’s output. It includes the code to perform all the necessary variable transformations and formulas to produce the final output variables (e.g. PositiveTargetProb, PredictedTarget for classification). In order to generate the scoring code, follow the below steps:

Open project and click the Scoring code button.

Select what language your scoring code should be generated in. The following options are available: