Here are the key concepts you need to know when working with Predictions, in alphabetic order.
In the context of the Predictions module, this term refers to a custom prediction model type. Classification is an example of supervised learning algorithm, which is used to work with the labeled data sets. Classification should be used whenever you want to predict binary values or classes like 0/1, true/false (for example, churn, purchase probability, customer will/will not perform a specific event, and so on).
A feature is an individual, measurable property of an object (for example, of a customer) you would like to analyze. A feature is also referred to as a variable or attribute. Features are automatically built statistics (for example, sums, averages, trends, deltas) based on the events already stored within the Synerise platform in order to be taken ultimately as data inputs for predictions.
Global feature importance gives you an idea how strong the influence of certain features was on overall model outputs. They have positive value and sum to an entity.
Local feature importance gives you an understanding which features influenced the most a single prediction in comparison to an average model prediction. They can have positive and negative values and they are scaled in the prediction values.
Lookalike predictions allow you to discover customers sharing the same preferences or any other qualities. Lookalikes let compare two segmentations called source (for example, segmentation of customers who converted in a campaign in the past) and target. The prediction lets you search the target segmentation for customers similar to the ones in the source segmentation. This, in turn, opens up opportunities to use the lookalike segments to address personalized communication, extending your reach or other actions. Lookalikes take advantage of our state-of-the-art AI/ML algorithms and frameworks such as Cleora and EMDE and are built upon behavioral profiles.
Percentile is a score below which a specified percentage of customers from an analyzed group falls. For instance, the 50th percentile means that 50% of the customers have lower score than this one. Percentiles are especially useful whenever you direct communication only to a certain number of customers. Let’s assume that you want to make a prediction for a segment that is made up of 100 000 customers, and you want to send a message only to 10 000 of them. In this case, you can pick TOP 10% of customers with the highest score: in terms of percentiles this value is expressed as “over the 90th percentile”.
A prediction (score) refers to a single output of a model after it has been trained, and then scored based on a selected data set (that contains chosen features). Predictions are probable values generated by a model, created for every customer who belongs to the selected audience. Depending on the type of the problem you want to solve, predictions can be perceived as likelihood or forecasted values. The output of the prediction is a
snr.prediction.score event generated on the profile of customers who were selected as a target of prediction.
Prediction result scale
The result of a prediction can be presented on a 2-point or 5-point (default) scale.
- 2-point: low, high
- 5-point: very low, low, medium, high, very high
You can predict the propensity for buying particular products, brands, categories, but you can also use any other item attributes for the prediction. The generated predictions can be easily used to build segments for further use in actions in Automation or Analytics. As with the other prediction types, propensity predictions are based on our state-of-the-art AI/ML algorithms and frameworks such as Cleora and EMDE and are built upon behavioral profiles.
In the context of the Predictions module, this term refers to a custom prediction model type. Regression is an example of supervised learning algorithm, which is used to work with the labeled data sets. Regression should be used whenever you want to predict the continuous value (for example, customer lifetime value, number of transactions in the next 90 days, total transaction value in the next 30 days, and so on).