Propensity to Buy Enrichment

A machine learning (ML) based enrichment that easily allows a Wondaris user to configure and enrich their customer data with the likelihood of each user making another purchase, the purchase criteria can be specified for more granular predictions. The propensity to buy predictions can then be used to segment your customer data and activate these customers into your marketing and customer experience platforms.

Instructions

Click ‘Create Enrichment’ from the side panel
Selecting 'Propensity to Buy” card
Step through the configuration:

3.1. Step 1 - Setup

Model’s name: Name your model anything you like. This is a required field.
Description: Though optional, it's a best practice to describe the model for future reference.
Data source: Select a data source you want to build and train model from. The default data source has been prefilled and you can choose any other available data source.
Model Prediction scope: This is to limit the data that is being enriched once the model is trained - this is useful if you have varying customer cohorts which have differing behaviours (eg: business vs non-business customers). You can either select:
- All Customers; or
- Filter specific data segment you want to enrich

3.2. Step 2 - Model configuration

Select a column which indicates the purchase (or conversion) of your customers, then set what value indicates a “successful” completion of the event.

The options for defining success for each model can be customised to your needs as per the following (available options will change based on the column data type - ie: if the column is not numeric, the less than, greater than and range will not appear):

Exists - just by virtue of there being an event (eg: transaction id), indicates a successful “conversion”
Equals - if the column’s value equals a specific value (useful for perhaps a category of a product in a purchase)
Greater than - this can be used to refine the model towards higher value conversions eg: the propensity to purchase something with a value greater than $100
Less than - this can be used to refine the model towards lower value conversions eg: the propensity to purchase something with a value less than than $100
Range - this allows you to set a range for the value.

3.3. Step 3 - Training configuration

Set the expected accuracy and training scope for your model.

For accuracy, please note that:

If the result is below the minimum line, it will not run and require re-configuration.
If the result is in between the minimum accuracy and target accuracy, you can decide to run it or not, but beware that it can affect the effectiveness of your activation.
Any result that is above the target accuracy will be available automatically for applying to audiences and activations.
Ideally for machine learning, you should be thinking about what a human would guess and what sort of accuracy they would achieve and then expect a predictive model to do better than that.

Accuracy in the context of Machine Learning has a specific meaning, however Wondaris optimises towards various metrics which measure the performance of Machine Learning Models, including:

Accuracy
Precision
Recall
F1 Score
Log loss
ROC AUC

We expose these values once the model is trained so that your data science team can validate the quality of the model.

Training Scope

Set the training scope to one of the following:

All time: This will use all data available for the training & evaluation of the model
Time range: Here you can use a simple method to refine how far back Wondaris will “look” to train your model - this can be good to use if your products or offering changed at some point in the past (ie: exclude data before the change)
Advanced Scope: this allows you to use our standard filtering mechanism to filter the training data - this includes, existing segments, customer attributes and behaviours.

This is an important step to ensure the data you use to train your model is a good representation of what you are trying to predict. If you have an imbalance in the number of varying conversions, you should refine your training data to be more balanced.

ie: if you have a lot of one type of transaction but you want to predict general transactions, you should filter the training data to have a more even count of each.

Reach out to your reseller or support@wondaris.com if you require some help with this advanced feature.

3.4. Step 4 - Training & enrichment schedule

You can set schedule to train your model, either:

Now: The model training will run immediately with current available data
Choose a date/time: The model will run at the time you choose - this is helpful if you need to bring in more data or expect a process to complete prior to doing the training.

Enrichment Schedule

This allows you to choose when the model will predict the values - not all options may be present in your licence.

Options include:

Every Day / Week: Predictions will be run at a time between 2am & 5am daily or weekly
Advanced scheduling allows you to choose a date & time, as well as the repetition of the predictions.
Advanced Cron: This is a technical way of choosing a schedule and allows for even more granular scheduling.
Never: this allows you to manually run the prediction by clicking an “Enrich Now” button in the model once it has been trained & evaluated.

For legacy mode, you are just able to select Now or after next CDP run to train your model

3.5. Step 5 - Review

Review your enrichment configuration before clicking Finish to create the Propensity to Buy enrichment.

After successful creation, you can check the Propensity to Buy result in the saved Enrichment and select any segment in the grid chart to either build an audience/segment or activate to a destination.

Over time, your model will continue to be evaluated, if it slips below (or above) your chosen accuracy, then you will be notified and recommendation actions (re-train or continue using) will be provided to you in the platform.

Required Data

Must have:
- unique customer / user id (ideally persistent over time so that the predictions can be used are more accurate)
- event date/time (eg: transaction date)
- event value (if you want to maximise towards a value)
Good to have:
- As much other data about customers & behavioural event data as possible

Technical Details

The following are some added technical details for the propensity to buy model.

Algorithm: Boosted Tree Classifier
Feature Selection (which columns are used to inform the prediction): automated, based on Wondaris' analysis of your data. Wondaris will determine if data is biased or too sparse and not use these columns.
Label Selection: The user needs to choose what happens in the data when someone "converts" / "completes an action" / "purchases" - this is what is optimised towards with the model.
Model parameters: chosen automatically based on hyper-parameter tuning
Training & Evaluation Data: Chosen automatically using random algorithms to ensure a good spread of data from the input training data.
Quality score: We use all the standard quality scores for Machine Learning (Accuracy, Precision, Recall, F1 Score, Log loss, ROC AUC), to ensure it reaches a set quality rating that the user sets (should be better than what a human can do)
- If the model is evaluated to fit into the quality score range, then it can be applied, where it will be continually evaluated and predictions made.
- The rate of evaluation & predictions should be based on the rate the data changes (daily / weekly, etc)
- If the quality drops over time because the behaviours of the user/s / customers diverge from what the model understands, then a warning will be provided to the user to perhaps re-model the data / create a new model.
- Over-fitting: Wondaris recognises if a model is overfit and will recommend AGAINST using it of this is the case.