Overview of Building Machine Learning Models

Here’s an overview of the high-level steps for building your machine learning models. Note that you must have the Manage Sales Machine Learning (ZCA_MANAGE_SALES_MACHINE_LEARNING_PRIV) privilege to build sales machine learning models.

1. You've a choice of either duplicating an existing model and changing the copy, or you can create a model of your own from scratch.
2. You add some basic information. If you're creating a model from scratch, you select the object and type of model you're building: either a prediction model or a model to identify similar records. For example, select the model type you want to build such as:
   • Predict outcome - This model type is used for making predictions such as predicting if an opportunity is likely to get won. It can also be used to predict deal size (number) or time to close (days) or account score (number).
   • Identify similar records - This model type is used for finding similar records such as finding other customers similar to the one being viewed. It can also be used to find similar opportunities to the one being worked on.
3. You select the attributes you want to use for your model.
4. You can add data set filters to define your training data set for both prediction and similar records models
   • Prediction models - This is the data set where machine learning models are trained. For example, if a model needs to make a prediction on an outcome for an opportunity, it needs to learn the patterns for both a win and loss outcome by learning from past closed opportunities. During model training, the model learns from the patterns that are different for won and lost opportunities.
   • Similar records models - This is the data set where similar records are found. For example, if active accounts are used as part of the training data set, for each account, it will find only similar records from active accounts. Any inactive accounts will not surface as similar records.
5. You can specify the attributes that stores prediction outcomes for the predict outcome model type. Select what outcome would be considered a successful outcome for this use case. For example, for an opportunity prediction, an outcome of Won might be considered as success criteria.
6. You click Prepare Data to check if there's enough data for each attribute.
7. If your model results in the Error status, you don't have enough data for some attributes and you must select different attributes or fix the data.
8. If your model completes with the Prepared status, then you can add data categories for the different attributes in the Add Features step.
9. Click Submit to build the model.
10. Analyze the model results. You get tips on how to improve the model.
11. If the model needs improvement, you can go back to the previous steps to tweak it.

12. You can deploy the model to automatically run on a daily, weekly, or a monthly schedule.

    Tip:

    For prediction models, you should rebuild your model based on how often the patterns being predicted would change. For example, the training recurrence for opportunity win probability models can be determined by how often the pattern of winning versus losing opportunities change.

    For similar records models, you should rebuild model based on how soon you'd expect new or updated records to show up as similar records. For example, any newly created account should show up as similar within a day or within a week.

Here's a sample flowchart that shows how to build a machine learning model.