• Feeding historical data
• Training the model
• Determining future outcomes
• Periodically retraining the model with newer data and arriving on the model that is the best fit, based on your business needs

Prediction model prerequisites

You can build and train custom prediction models to predict outcomes tailored to suit your business needs. 

What skills do you need?

• You (admin/super-admin) can create and use prediction models, whereas developers and users can only consume the models you've created. 
• It is recommended that you should know enough about all your business requirements to determine the dataset that will be used for training the model. 
• It is preferred that you do an exploratory data analysis (EDA) while collecting data to train your prediction model.
• Creator is a low-code platform, so it doesn't require you and your users to have prior coding skills to create and consume the prediction models. You also need not know machine learning (ML) algorithms, as Creator has an AI modeler with ML capabilities.

What data do you need?

• You need a minimum of 50 -100 records in your training dataset, but for best results you should have at least 10000 records.

Which pricing plan must you be in?

• AI Models will be available for users in paid plans of C. Refer to our pricing page to know more.

Which version of Creator should you be using?

• You must be using Zoho Creator 6 (C6) to be able to create custom AI models whereas, the ready-to-use AI models are available in both C6 and C5.

Example 

Now that you're aware of the type of skills and data required, the next stage is to know how to narrow down and choose the values to be fed into training data. 

Let's assume that you've built an application named Zylker Insurance using Creator. You've added a form named Insurance Claim in which your customers submit claim requests. You want to find out whether or not the insurance claim requests raised by your customers are legitimate. As you can see, this is a form of binary prediction outcome model. 

Now, think of the factors that'll influence the outcomes that you want the prediction model to make. For example, for the question "Is the insurance claim request legitimate or not?" think about questions like these:

• What is the severity of the incident reported?
• How long (number of months/years) have they been a trusted customer?
• What is the total claim amount quoted?
• Are there unknowns in a column that might cause uncertainty?

You can use the above information to make your data selections. 

1. Identifying what value you want to predict. Here, you want to predict whether the claim request raised by your customer is legitimate or not.
   
2. Identifying the data that you want to use for prediction. training/historical/previous data set must cover these factors:
   
1. The number of months as a trusted customer
   
2. Severity of the incident
   
3. Total claim amount quoted
   

• Ensuring that your Insurance Claim form has as many records as possible. You can either use all the records or define a criteria to use specific records for your model training. For example, you can filter records for a particular time period.

Note: In case you have all your training data stored in a CSV file, ensure that it has as many data columns as possible. 

1. First, mapping the base field and dependent fields selected earlier to their relevant fields in the form. A prediction field (base field) will now be added to your form.
2. Next, accessing your application and entering values based on which the predicted value will be displayed in the prediction field.

Note:

1. In the case of adding training data via CSV files, field mapping is mandatory.
2. If you've added training data using your application records, you need not perform field mapping when you've deployed a prediction model in the same form from which the model was built using that form's data (fields). In this case, the fields will be auto-mapped.
   

Prediction Model Flow

Add training Data

Training data is the initial dataset that is used by the model to find patterns, make interpretations, and arrive at a prediction. Once you've finalized the data that you want to feed into your model, you can add the data in two ways: from the form fields in your application or from a CSV file. In prediction models, training data consists of base field and dependent fields in the first method of adding data and base column and dependent columns for the latter method.

You might wonder when to choose which of the above methods of adding training data. If you have sufficient records in your application to be used as training data, you can go with adding fields from your forms, whose records will be used as training data. Alternatively, if you do not have sufficient records in your application, but have your data stored in a file, you can go with the latter option of adding data from a CSV file. 

Note: 

1. The fields that you select from your application will act as base and dependent fields for training.
2. The columns that you select from this CSV will act as base and dependent columns for training.
   

The most crucial thing to consider here is whether a record/column that isn't your historical outcome column is indirectly impacted by the outcome.

Let's say you want to predict whether an order is going to be delayed. You may have the actual delivered date in your data. This date is present only after the order is delivered. If you include this column, the model will have close to 100 percent accuracy. The orders for which you want to predict the delivery date, won't have been delivered yet, so they won't have the delivered date column populated. In order to achieve accurate outcomes, you should deselect columns like this before training. 

Data from your form fields

You can select the data stored in your application fields to be fed into the model as the training data. 

Base and dependent fields

While selecting application fields as your training data, you require two types of field data:

• Base field is the field for which you want to predict the outcome. 
• Dependent fields are the fields that you want the prediction model to use for the prediction process. 

In the above example, the base field would be the "Is this a Fraud Claim", while the dependent fields could be"Months as Customer", "Incident Severity", and "Total Claim Amount" fields.

Record Selection

After you've chosen the records for your training data, by default, the data from all your records will be taken into consideration. Sometimes, you might want to focus on learning and making predictions on a specific set of records. You can define a criteria that filters a specific set of records to train your prediction model; this criteria can be set by laying out a set of conditions as per your need. You can use this step to filter your data if you're aware that the records you are using to train a model contain irrelevant information.

Let's assume an insurance company uses Creator to build an application that predicts whether an insurance claim request is fraudulent or not. To predict this accurately, the model should be trained with all the records found in the application form, thereby widening the model's comprehension. 

Now, let's say that the insurance company has declared that all claim requests with the incident date before the year 2018 cannot be processed. When this is entered as a criteria, only the requests raised after 2018 will be considered for the fraud claim. You can use the criteria in this case and filter the records accordingly.

Data from CSV File

Another way of selecting the training data is to use the data stored in your files of CSV format. In case your data has unsupported field columns, they won't be shown in the column selection page.

Column Selection

The Prediction model supports training data from CSV in the following data types: number, text, date and date-time. Data from these two types can be added as the base column and dependent columns. 

Any columns with unwanted data that don't fall into one of these two data types below are unsupported in CSV prediction. If your data has unsupported field column types, they won't be shown in the field selection page.

Base and dependent columns

The columns that you select from this CSV will act as base and dependent columns for training.

• Base column is the column for which you want to predict the outcome. 
• Dependent columns are the columns that you want the prediction model to use for the prediction process. 

In the above example, the base column would be the "Fraud Claim", whereas the dependent columns could be "Months as Customer", "Incident Severity", and "Total Claim Amount" fields.

Model guidelines

• A base field/column is required to predict the outcome.
• A maximum of 20 dependent fields/columns can be added to your prediction model.
• You can add up to 20 criteria per prediction model.
• 4MB is the maximum file size limit for CSV files.
• You need a minimum of 50 records for training the model. To achieve higher accuracy, you can add more than 10000 records to make the data richer.

Model Summary

After adding the training data, you can review the model details such as Model Name, Base Field/Column, and Dependent Fields/Columns.  If required, you can modify the Model Name, Base Field/Column, and Dependent Fields/Columns by going back. Otherwise, you can proceed to train the model. 

• Train model 
• View and manage model details
• Test model
• Publish and use model

Train model

Before you can actually use your prediction model in your application, you have to train it to perform and produce the expected outcomes. After you've selected and reviewed your data fields/columns, click Train to train your model.

View and manage model details

Once training your model is complete, you can view the model details, the model's versions, and its deployment details, if any. The model is now ready to be published and deployed to your apps.

Note: 

1. The retrain option is available only when you add training data via form fields. This is because you can train your model as and when new records are continuously added to your application.
2. When adding data via CSV file, you can delete the model, upload a new file, then train the model again.

You can manage your model in the following ways:

• Retrain - Model retraining refers to updating a deployed AI model with new data. Because new data is always being generated, we recommend that you periodically retrain your model. This helps in improving the prediction model's reliability and accuracy. 

• You can click Retrain and your model will be retrained.
• After each retraining is over, a new version of the model will be created. You can switch between different versions according to your needs.
• If you want to delete a version that is currently used, you need to switch to another version before proceeding to delete the current version. 
• If the model training fails, the previous working model will be used for prediction.
  

• Rename - You can rename your model if required.
• Delete - If you want to delete your model due to inconsistent, wrongly-added data, you can use the Delete option.

Test model

After training, you can test your model to check and know how your model works and if the training is satisfactory, before deploying it in your applications. You can upload test data and after testing your model; you'll get the predicted outcome along with an accuracy score. 

• If the accuracy score (%) is high (85-97), you can proceed to publish your model.
• If the accuracy score (%) is good (70-84) or fair (51-69, you can retrain your model with newer values.
• If the accuracy score (%) is poor (<50), you need to check for inconsistencies in your data, refine them and train your model again

Accuracy score

The prediction model calculates the accuracy score for your trained model based on the prediction results of your test dataset. For example: if your dataset has 500 records, and the model correctly predicts 492 of them, then an accuracy score of 96 percent is shown.

Improve your prediction model performance

After you've trained, tested, and evaluated your model, if you find that the model outcomes aren't as expected, you can edit (optional) your model to improve its performance. Here are some things you can try to help improve your model's accuracy score.

• If model training has failed, check your model details and retrain the model.
• If there are no errors but your accuracy is low, check the model training details. 
• Try to address as many issues as possible, then retrain the model. Some of the issues might arise due to:

• When all the mandatory model guidelines are not met 
• Incorrect data is fed into the model

• In the case of adding training data from application records, sometimes the model training might fail when the respective base and dependent fields aren't present in your form. In this case, you need to first create then in your form, then train your model.
  

• You need at least a minimum of 50 - 100 records of data to train your model. In order to train a highly predictive model with higher accuracy, you need to have 10,000 or more correctly-labeled records.
• Sometimes, you might have a lot of correctly labeled training data. However, the model might still not perform well, as you might have selected fields/columns that are irrelevant to make the required prediction. Ensure that you only select columns that are relevant to influence what you want to predict, and deselect irrelevant columns.
• It is recommended to include a range of possibilities in your data that generally corresponds to the range of options you might expect to see as the predicted outcomes. For example, if you've two historical outcomes of Success or Failure, and most of your data records only have Success in this column, it's hard for your model to learn from this type of imbalanced data distribution.
• Ensure that your data doesn't have a high rate of missing values. You can populate the missing values with default data or remove the data from model training.
• If a field/column has a high correlation with prediction outcome, remove it, then proceed to train the model.

Cleaning data is the process of removing inaccurate, incorrectly formatted, duplicate, or insufficient information from a training dataset. Combining two or more data sources increases the risk of data duplication or labeling errors. Even when prediction results appear to be accurate, data mistakes can make them unreliable.

Before feeding the training data, it is crucial to clean up your data, as this will help improve your model performance. Taking the time to carefully review each row of data for typos, missing numbers, spelling mistakes, and other errors, is the best way to cleanup faulty data. By doing this, you can get rid of data that is obviously unsuitable for model training. 

Publish and use model

After you train your model, you can publish it to make it available to your users and start making predictions. Learn how

To use your prediction model in an environment-enabled Creator application, you must have at least one version of that application published in the production environment. After deploying your model in the application, you can filter between different stages of the environment to check which stage is the model deployed in. Learn how

Get started with sample data