Feature Engineering

In Machine Learning, a feature is an individual measurable property that serves as an input to the Model. A feature can be a field of the Customer table, such as "customer_since", or a processed or combined form of several fields of the data tables. The process of taking available data fields and transforming them into useful Features for the Model is known as Feature Selection or Feature Engineering.

CrossEngage automatically retrieves data from the Customer, Transaction and Activity tables, and processes them into useful Features, which are then fed into the Model.

The importance of a specific feature depends on the data; hence for different data with the same / different layout, the importance of features changes. The most important features are automatically selected in any case.

Basic Features

Customers

All uploaded fields (columns in tables) in the Customers Table are prepared. However, only the additional Features that are explicitly mentioned in the Model Builder go into the model, in order to ensure reliability of the Model.

Transactions

The transaction data is aggregated to calculate a summary of transactions for each customer. This can then be combined with Customer data to train the Model.

• Recency: The time period (in weeks) between t0 (the time of modelling) and the most recent transaction

• Duration: The time period (in weeks) between t0 (the time of modelling) and the oldest transaction

• Frequency: Number of different invoice_ids

• Item_count: Number of invoice items (number of records)

• Quantity: Sum of the Quantity column

• Revenue: Total revenue ( sum of (price*quantity) )

• Value: Sum of sales according to the target definition specified in the Model Builder

• Any: Boolean indicator whether transactions exists

For calculating Value for a customer:

• Positive order_types are positive → price = abs(price)

• Negative order_types are received negatively → price = -abs(price)

• Neutral order_types are included without modification→ price = price

• Additionally, only transactions that take into account the other target definitions that have been set, e.g. B. Filter for certain product groups or order channels

Activities

Similar to transaction data, activity data is aggregated to calculate a summary of transactions for each customer.

• Recency: The time period (in weeks) between the most recent uploaded data and the most recent activity

• Count: Number of activity records

• Stock: Time-weighted sum of activity records. It can be calculated with the formula:

$$Stock_{c,interval} = \sum_{n=0}^{N} 0.8^{t_n}$$

where N is the number of interactions of Customer c in an interval, and t(n) is the number of days between the start date of the interval, and the activity n

• Any: Boolean indicator whether activities exist

Advanced Features

The Features of transactions and activity can also be restricted in two dimensions; intervals (e.g "within the last 30 days") and categories (e.g "from product group A”). It is also possible to combine these restrictions, e.g "within the last 30 days from Product Group A".

This means that additional features can be created by calculating features with restrictions. Crossengage's automatic data preparation already provides extended basic features for a certain set of categorical properties (always assuming that the information is available in the data!).

By category

The following additional Features can be calculated from the Transactions data.

• The top 8 Categories of the field: productgroup_id

• The top 5 Categories of the field: order_type

• The top 8 Categories of the field: order_channel

• The top 3 Categories of the field: mapped_order_type

The following additional Features can be calculated from the Activity data.

• The top 5 Categories of the field: wtr_type

• The top 5 Categories of the field: activity_type

Understanding Features of a Model

Choose the model you wish to see the features for. Click on Model Report, and scroll down to the section: "Weight of variables for Conversion / Value model".

The "weight" of a feature is its relative importance of the feature in making the prediction.

Here you can see a list of features for this model. In this example, the top 6 features combined provide roughly 2/3rds of the predictive power of the model.

Naming Convention

Features follow the following naming scheme:

[SPECIAL] (FEATURE)_IN_(INTERVAL) [AND_(CATEGORY)_IS_(ELEMENT)]

Here, square brackets denote optional parts of the name, while round brackets contain placeholders.

• Special: If this contains the word FILTER, this means that the entire history of the field up to t0 is used.

• Feature: This is the name of the field the feature is created from, e.g transaction_recency, revenue.

• Interval: The time interval filter, with respect to t0. You can find more about t0 below.

• Category: Category filter, e.g order_type or productgroup_id.

• Element: Value of the category, e.g sale, return or Shoes.

Time with respect to t0

The data preparation is based on points in time (t0) set in the past. Customer behavior before these points is used to draw conclusions about behavior (according to the target variable) after the respective point in time.

Generally, you can understand t0 as the time of the last transaction in the dataset. It is the point in time where the available data ends, and after which the behavior of the User is predicted.