Einstein Prediction Builder : Improving Predictions through choice of fields

In my last blog on Einstein Prediction Builder, we created a new numeric prediction. Link to last blog – Build Numeric Predictions with Einstein Prediction Builder

We saw that we got a “Good” prediction score of 44. But is our score good enough to to get quality predictions?

It is really important to understand that even if a score is good, there are chances that our predictions might not predict what we want to see.

To ensure quality, we need to do 2 level of checks :-

1. The fields we are using.
2. The data in our example set.

In this Blog, we will cover the improvement we gain from changing the fields.

STEP 1 : When we choose the fields, we have to first eliminate “Hindsight Bias“. These are fields where outcome occurs post our use case eg: In our scenario, we are checking the probability of case escalation, but have included “Escalated” field in our prediction set. This means when the field is true, the case is escalated. So if we are predicting chances of escalations, then we can’t have a field in prediction that tells us whether escalation has happened already. Thus it is better to remove this field.

Sometimes it is easy to locate as such fields as they are having High Impact on predictions. Removing such fields can sometimes drop the scores drastically. In my use case, the score dropped to 17.

Now we see Case Origin field having an impact and “LiveAgent” having most impact. However a quick check suggests that this field is not falling under the category. Thus we can eliminate the Hindsight Bias towards the prediction by carefully checking the report and fields being used.

Having got a low score at the moment, shouldn’t be a concern as we have taken a step towards quality prediction.

STEP 2 : We should now review the fields where we have no / minimal data in example set. These fields can’t help much in quality predictions and should be removed from the field set. In my use case, I chose to eliminate the fields with no / minimal data.

Checking the score with the change resulted in increase of score to 24.

STEP 3 : Remove the fields which have low impact / variance. This can be found when we open the Top Predictors report (link on Scorecard).

As we can see Account Name has no impact so removing the field can help in improving our results.

We have now reached the Good range again with a score of 47.

So we started with a score of 44 with low quality of predictions and ended with 47 but having a higher quality of prediction. The improvement we did is a step in the right direction to get quality predictions and would further help the score from Good to Great (which I will cover in the next blog).