DIFM Prebuilt Machine Learning Recipes For Propensity-Based Customer Use Cases
Data continues to flood every organization, both in size and in speed. Sometimes more data is better, but the challenge can be that critical decision-making information gets buried and regretfully, never gets unearthed. Skilled analytical talent with domain experience in modern marketing is the key to move a brand from reactive to proactive. Thus, varying flavors of IP, technology features and automation are critically important in accelerating the equation of time-to-insight-to-value:
- Do-It-For-Me (DIFM) automation
- Do-It-Yourself (DIY) customization
- Prebuilt machine learning recipes for common customer use cases
Within the martech industry, there are several factors that contribute to the challenges surrounding brands and their decision-making processes. Obviously, customers and markets are more competitive and demanding. When you step back and reflect on this, it's a linear trend upward year-after-year when it comes to consumer expectations. This means, to satisfy that demand, it's well recognized that brands need to respond quicker, but it's often overlooked that accuracy and quality hold equivalent weights of importance.
In addition, every software vendor claims to be a CDP, do magic with AI, and activate across an ever-growing list of customer communication touchpoints. However, that doesn't mean every technology and solution provider is the same, and it's vital to look past high-level buzzwords displayed in large font sizes of slide-based presentations. Transparency is key in meeting a brand's data, analytical & activation requirements.
But don't take it from me. Fresh publications from Forrester Research released on the subjects of customer analytics (Q2, 2024) and real-time interaction management (Q1-2024) amplify the present day importance of brands leveraging technology that prioritizes performance, productivity, trust – the trifecta of undeniable AI platform value in delivering precision within martech, audience targeting and personalization.
In the context of customer propensity analysis, SAS strives to provide:
- Algorithms that produce better analytic scores and accuracy
- Automation of machine learning that aligns with greater productivity
- Embedded analytics, making AI more impactful and consumable
- Human-like interfaces, creating approachability
- Trust, which is critical when using propensity scoring effectively
As cited in an introductory article explaining how SAS Customer Intelligence 360 enables DIFM prebuilt machine learning recipes, there is a little secret in the customer analytics ecosystem that practitioners frequently will admit to when pressed for honest feedback. A massive proportion of customer & marketing analysts in 2024 continue to skew towards the wrong end of this workflow spectrum:
"I spend more than 80% of my time preparing data, and less than 20% actually performing analysis."
Speed bumps like this usually emerge when customer experience teams require advanced insights for propensity scoring, algorithmic segmentation, retention strategies or next-best-actions. Accelerating through this challenge has been a key area of interest at SAS, as we recognize customer experience management has an insatiable appetite for data intelligence.
Domain Expertise & Propensity-Based Use Cases
Think about the magnitude of requests that come in from customer experience and marketing teams to their supporting analysts. The wish list includes actionable propensity scoring for topics like:
- Lead scoring
- Acquisition
- Upsell
- Retention
- Win-back
- Supervised segmentation
- Next-best-action (or experience)
This myriad of desires stratifies further when considering industry context. For example, retail brands commonly desire to optimize their app's shopping experience and increase the efficiency of conversion rates. Alternatively, non-profit brands want to segment their digital traffic and personalize experiences based on one's likelihood to donate vs. research and/or engage.
Propensity-Based Use Cases & DIFM Prebuilt Recipes
An emerging trend to combat the ongoing analysis inefficiencies cited above involve Do-It-For-Me (DIFM) prebuilt recipes representing a specific ML/AI algorithm or model ensemble, processing logic, and configuration to auto-build and execute a trained solution that comprehensively solves (or improves efforts against) specific business problems. The propensity models and data engineering pipelines are ingredients of a broader recipe that get trained on data and parameter configurations to optimize the solution's ability to contribute significant value when pivoting to customer inference.
Training is the process of learning patterns and insights from labeled data. A trained model represents the actionable output of a model training process, in which a set of training data was applied to the model instance. The benefits of a trained model include scoring, inference and the opportunity to create an intelligent customer service.
Supervised learning algorithms are trained using labeled examples (conversion vs. non-conversion), such as an input where the desired output is known. The learning algorithm receives a set of inputs along with the corresponding correct outputs, and the algorithm learns by comparing its actual output with correct outputs to find errors. It then modifies the model accordingly. Supervised learning is commonly used in applications where historical data predict likely future events (another way of describing the value of propensity scoring). For example, supervised learning can anticipate when an insurance customer is likely to file a claim, or identify a patron has a higher likelihood to be interested in a particular excursion (versus alternatives) for their cruise experience.
SAS supports two types of supervised learning problems:
Classification – When the data are being used to predict a categorical target, supervised learning is called classification. This is the case when assigning a label or indicator (for example, a customer selects to subscribe or not subscribe). When there are only two labels, this is called binary classification. When there are more than two categories, the problems are called multi-nominal classification.
Regression – When the data are being used to predict interval targets (for example, the donation amount as opposed to whether someone will or will not donate), the problems are referred to as regression.
For more details on how SAS supports supervised learning in martech, please go here. Now, with that backdrop, we can transition our thinking about customer propensity-based recipes through a stepwise (or itemized set of workflow steps) that can be templated, shared and if a user desires, customized further. Let's walk through a use case where brands can leverage SAS Customer Intelligence 360 and Viya to benefit from this concept.
SAS Customer Intelligence 360 + Viya for DIFM Prebuilt Propensity Recipes
The structured data model available to users of SAS Customer Intelligence 360 provides 1st party digital customer data views suitable for a variety of customizable analysis purposes. The data model and it's associated schema is an opportunity to leverage propensity use case-driven recipes.
Here are examples of what an analyst can do with data originating from SAS Customer Intelligence 360 when instrumented on a brand's website, mobile app and/or any channel-based communications/campaigns.
- Leverage session-based behavior where structured information about all visitors (identifiable and/or anonymous) including device type, browser, date, time, location and other dimensions/metrics is accessible to feed into a propensity-scoring recipe.
- Use journey-based customer behavior where a consolidated view of all sessions, attributes and activities across all 1st-party cookies and devices is available to derive propensity scores related to any important macro- or micro-goals.
- Detect bias and mitigate (or remove) such negative and unintended effects on customer propensity scores.
- Build a product offer engine rank-ordered by propensities for data-at-rest or data-in-motion use cases leveraging 1st party customer behaviors.
- Benefit from a variety of supervised learning techniques embedded within propensity-oriented recipes.
The list above is simply a sampling of ideas on how the SAS Customer Intelligence 360 data model can be leveraged for customer propensity use cases. Next, let's take a high-level look at the types of tables that reside within the data model's schema.
As analysts within SAS, the concept of accessible data of any type within SAS Libraries applies to users of both SAS 9 and SAS Viya. We are accustomed to creating connections to where data resides within on-prem and cloud-based locations. Using this same concept, SAS Customer Intelligence 360 makes 1st party digital interaction data between consumers and a brand available to SAS analysts in the same exact manner.
The screenshot below shows an instance of SAS Studio & Flows within Viya 4. On the left-side, SAS analysts will immediately recognize a library with data tables. One specific table, entitled FIN_ACTIVITY_CONVERSION is opened to highlight the available dimensions, metrics, date and time-based variables. The Pre-Process Code node in the first swim lane of the Flow contains a short snippet of code to enable the analyst to create a data connection. It simply contains the instructions/credentials highlighting how SAS Customer Intelligence 360 data residing within a Snowflake environment can be accessed in a repeatable fashion. After running the node, the connection to the server residing in Snowflake is established and the library containing all of the data model's tables are available for analysis. Outside of Snowflake, readers should know the same approach can be used for SAS Customer Intelligence 360 data landed for long-term storage in other platforms (hyperscalers, on-prem databases, etc.).
Analysts are not limited to simply selecting ALL the data. The next screenshot shown below opens up a second swim lane within this Flow which provides no-code user features to select:
- Specific area(s) of the data model (dimension, event or reporting)
- Acute categories of data
- Data-at-rest snapshots filtered on time windows
The beauty of this example is enabled in SAS Viya 4 using a Custom Step feature. For readers unfamiliar with this, SAS Studio ships with several SAS Steps, which are available from the SAS Steps tab in the Steps pane (see screenshot below). For example, one use case for a Custom Step enables analysts to create an interface for no-code users at your organization to complete a specific task. Custom Steps can be saved to your SAS Server or in SAS Content.
The SAS Server is any file system that SAS can access. Users can save Custom Steps on a local network, in Git, and other shared mounted systems. Analysts can access these steps from the Explorer pane within SAS Studio on Viya. Steps that are saved can be shared with other SAS users at your organization.
The Shared tab shown on the left-side in the screenshot above lists all the Custom Steps available in my demo environment. It includes Steps for simplifying the process of accessing data originating from SAS Customer Intelligence 360. This list includes any Steps that an analyst authored, any steps that are saved at a location where users have access, and any Steps that have been shared. This is a critical moment in this article's reading.
For every brand that selects to use SAS Customer Intelligence 360 and Viya 4, the deployment of the technology can include Custom Steps to remove any high-code requirements and users can accelerate their access of important 1st party digital customer data originating from interactions with a website, app, marketing channels and paid media. Removing the friction of working with SAS APIs (although high-code users are still welcome to leverage this alternative) is the value proposition to accelerate usage and amplify analysis efforts.
For brands who have this no-code preference, simply communicate this request to your supporting SAS account/support team members, and the Custom Steps highlighted in this article can be shared with your user team.
Moving on, now that an analyst has selected the relevant data from SAS Customer Intelligence 360 to work with, we can explore the value propositions of customer propensity recipes and the associated ingredients for leveraging ABTs (analytic base tables) & supervised learning algorithms. The benefit of prebuilt machine learning recipes for propensity scoring can be very helpful to data scientists and developers so they don’t have to start from scratch. If users prefer, they can adapt the proposed prebuilt recipe described below to their needs (or use it as inspiration to start from scratch to build your own custom recipe). Once analysts train/tune a recipe, creating an intelligent activation layer doesn’t require a developer - just a few clicks and marketers are enabled to build targeted, personalized customer experiences using the SAS Customer Intelligence 360 Audiences API. More on this in a moment...
The concept of a recipe may be a new term to some readers. Let's break this down:
|
Recipe |
A recipe is a term for a data-driven solution for a particular use case and is a holistic asset representing a specific ML/AI algorithm or ensemble, processing logic, and configuration required to build and execute a trained model and hence help solve specific business problems via inference. |
|
Model |
A model (prediction/ML/AI) is a recipe ingredient that is trained using historical data and configurations to solve for a specific use case. |
|
Training |
Training is the process of learning patterns and insights from labeled data. |
|
Trained Model |
A trained model is a recipe ingredient representing the executable output, in which a set of training data was applied to the algorithmic solution. The trained model is suitable for scoring and creating a customer treatment strategy. |
|
Scoring |
Scoring is a recipe ingredient representing the process of generating actionable insights (inference) from data using a trained model. |
|
CI360 Audiences Service |
A deployed service is a recipe ingredient which exposes functionality of an advanced algorithm (originating from SAS and/or open-source) through an API so that it can be consumed and activated by SAS Customer Intelligence 360. |
With that said, let's dive into an example. We will start with a recipe for generating customer propensities, where SAS strives to accelerate how our users can make intelligent decisions regarding customer treatments.
The customer propensity recipe begins with input tables originating from SAS Customer Intelligence 360's data model. For this example (screenshot above), we are using five tables:
- Product Views: The PRODUCT_VIEWS table provides information about what products visitors view. The table is sourced from the product view event in SAS Customer Intelligence 360.
- Session Details: The SESSION_DETAILS table provides information about web and mobile sessions. This information includes a wide range of data about your users. For example, you can use this table to determine which browsers are used to access your content, identify the geographic location of web or mobile app users, and identify the traffic sources that brought a user to your website.
- Identity Map: The IDENTITY_MAP table stores the associations between anonymous users and identified customers, and is updated when an anonymous user is identified by SAS Customer Intelligence 360. An anonymous user can be identified through an identity event or a data import. These examples illustrate when an entry is added to this table:
-
When an anonymous user signs in to your site (or app) and triggers an event that captures the login_id or customer_id, that anonymous user becomes an identified user.
-
When the same person is identified by SAS Customer Intelligence 360 across different devices, an entry is added. For example, a user first signs in to your mobile application, then the same user navigates to your site without signing in. The user is identified on the mobile app, but anonymous on the site. After the user signs in to your site, SAS Customer Intelligence 360 associates the anonymous user with the identified user that logged in to the mobile application.
-
If you import data that has information in multiple ID columns and one of those columns is already known in SAS Customer Intelligence 360, the system merges the IDs and their attributes into one ID. For example, SAS Customer Intelligence 360 has a customer_id for a user from an identity event and a subject_id for the same user from an external event. To associate these two IDs, you can import data that contains all of the identifying information for this user, and SAS Customer Intelligence 360 merges all the known information about this user to a single customer identity.
-
When an entry is added to the Identity Map table, any attributes that are associated with anonymous user become associated with the identified user.
-
- Identity Attributes: The IDENTITY_ATTRIBUTES table contains information about a person’s identity. Data is stored in this table when any of these actions occur:
-
Identity events created in SAS Customer Intelligence 360 are triggered, and the events contain a customer_id or login_id. This includes identity events sent through our JavaScript API and the mobile SDK.
-
External events are sent to SAS Customer Intelligence 360 using either the external API gateway or the on-premises API gateway. The events must contain a subject_id, login_id, or customer_id.
-
1st party customer data is imported into SAS Customer Intelligence 360.
-
- Custom Events: The CUSTOM_EVENTS table captures user-defined custom events for which SAS Customer Intelligence 360 does not have a standard definition.
For more information about the SAS Customer Intelligence 360 data model, please visit the documentation pages here. The next couple of steps in the Propensity Recipe involve the use of native data transformation capabilities within SAS.
Keep in mind, a Flow is a sequence of operations on data. Data and operations are represented in SAS by steps that users can access from the Steps section of the left-navigation pane. Each step in a flow is represented by a node on the flow canvas. The nodes on this Flow canvas above represent some of the steps that are available in SAS Studio. We are simply using Query node capabilities to perform various types of deterministic table joins to connect customer identities with behaviors like product viewing, session activities and custom events.
SAS Studio is shipped with many predefined steps that include queries and data transformations. The steps are organized into categories that indicate the function that they perform. With respect to the Query step acutely, users can leverage this node to select, join, filter, and sort columns from a table in a Flow.
Once the input tables have collectively been mapped to customer identities, we proceed with transposing the products customers viewed and custom event definitions they met. The Transpose Data step turns selected columns of an input table into the rows of an output table. Our desire is to manipulate the input data and reshape them as predictors in anticipation of performing supervised learning and propensity modeling.
After these nodes complete processing, the formation of a model-ready ABT is now ready for algorithmic propensity modeling within this demo example. The reason this milestone is important is because if users can accelerate to this step of the process, it reduces the time-to-insight issue (80-20 trend) cited at the beginning of this article. In addition, this is an example of how SAS provides robust CDP+ capabilities, matching the typical requirements of a CDP solution today in martech while also extending incremental benefits of a CDP to the data science, analyst and marketing communities.
For simplicity, we will use imputation and a tree-based algorithm in the next two steps.
The VARIMPUTE procedure performs numeric variable imputation in SAS Viya. Imputation is a common step in data preparation. The VARIMPUTE procedure can replace numeric missing values with a specified value, with the mean or median of the non-missing values, or with some random value between the minimum value and the maximum value of the non-missing values.
The TREESPLIT procedure builds tree-based statistical models for classification and regression in SAS Viya. The procedure produces a classification tree, which models a categorical response, or a regression tree, which models a continuous response. Both types of trees are called decision trees, because the model is expressed as a series of if-then statements.
The predictor variables for tree models can be categorical or continuous. The set of all possible combinations of the predictor variables is called the predictor space. The model is based on partitioning the predictor space into nonoverlapping segments, which correspond to the terminal nodes (called leaves) of the tree. Partitioning is done repeatedly, starting with the root node, which contains all the data, and continuing until a stopping criterion is met. At each step, the parent node is split into child nodes by selecting a predictor variable and a split value for that variable that minimize the variability, according to a specified measure, in the response variable across the child nodes. Various measures, such as the Gini index, entropy, and residual sum of squares, can be used to assess candidate splits for each node. The selected predictor variable and its split value are called the primary splitting rule.
Tree models are built from training data for which the response values are known, and these models are subsequently used to score (classify or predict) response values for new data. For classification trees, the most frequent response level of the training observations in a leaf is used to classify observations in that leaf. For regression trees, the average response of the training observations in a leaf is used to predict the response for observations in that leaf. The splitting rules that define the leaves provide the information that is needed to score new data.
The process of building a decision tree begins with growing a large, full tree. The full tree can overfit the training data, resulting in a model that does not adequately generalize to new data. To prevent overfitting, the full tree is often pruned back to a smaller subtree that balances the goals of fitting training data and predicting new data. Two commonly applied approaches for finding the best subtree are cost-complexity pruning (Breiman et al. 1984) and C4.5 pruning (Quinlan 1993). Compared with other regression and classification methods, tree models have the advantage that they are easy to interpret and visualize. Tree-based methods scale well to large data, and they offer various methods of handling missing values, including surrogate splits.
Readers should be reminded that any supervised learning algorithm (or ensemble of algorithms) appropriate for propensity-driven use cases could be leveraged in SAS for the modeling ingredient of this recipe.
Although comprehensive use case-driven recipes can be shared between SAS and our user community, if an analyst wanted to author their own custom propensity model, SAS enables no/low-code users (not just high-code users) to leverage a GUI interface to assign the relevant predictors, parameters, auto tuning method, and other criterion properties. As these inputs are made, the right-side of the screenshot above highlights how SAS auto-scripts the programming language to run the custom model.
Users can leverage the code-to-flow feature to then map in the custom authored analysis into the Flow as a Step.
The result completes the Flow's propensity recipe ingredient for running the tree-based model (shown below).
Now, the big question that has been posed to analytical technology companies year-after-year from our customers is whether data-driven insights can bring positive momentum to mission-critical KPIs. This brings us to another important recipe ingredient because I have a message for my data science and analyst brothers and sisters:
There is more to activation than just scoring your model!
You want to see your analytical assets bring rewarding impact to your brand, right? You want to observe your efforts making a significant positive difference in customer journeys, correct? Then let's complete this by discussing the last recipe ingredient that ties into destinations, journey orchestration and prescriptive activation. In the screenshot below, the last Swimlane of the Propensity Recipe Flow is highlighted.
The first node provides us a view into the propensity scoring that resulted from running the tree-based model step. The first column entitled Subject_ID is the unique (and cloud-secure) identifier that enables SAS Customer Intelligence 360 users to communicate, target or personalize experiences on websites, apps and channels with individuals or audiences. The propensity scoring is embodied in the second column entitled P_CE_InterestedInCreditCards (or the propensity score associated with a customer's likelihood to be interested in a credit card offer from a financial services brand). For an overview on Identities in SAS Customer Intelligence 360, please go here to learn more.
Analysts should consider this table as the "prescription" for our marketing counterparts. The marketer has a need or desire for intelligent scoring of the customers they want to target with a treatment (as well as exclude those who are deemed irrelevant). The recipe scoring is the prescription (or conduit) between data science and marketing for the given use case. Using the SAS Customer Intelligence 360 Audiences API, this last Swimlane contains two remaining steps:
- Query and share the relevant attributes between Viya and Customer Intelligence 360. In the screenshot below, we select to include the Subject_ID (unique & encrypted identifier for a customer) and the associated propensity scoring. All other attributes are removed since they are only relevant to data science and the associated quality of the propensity analysis. The removal of this information does not impact the marketer's activation workflow.
- Push the scored data to the correct cloud-based tenant using the Audiences API where your brand's instance of SAS Customer Intelligence 360 lives.
A large reason we are excited to share recipes with our user community is validated again in this last Custom Step of the demo. For those readers who have worked with APIs, you understand they typically require high-coding skills. In essence, what we have shown here removes the friction of an analyst having to author this code themselves and simply provide inputs in a few clicks. So, what is the result after running this final Swimlane? Users of SAS Customer Intelligence 360 will see the scored Audience within the software available for activation.
Based on a brand's SAS environment, this process will take either seconds or minutes. Once the processing is completed, users will see an updated status with additional details (as exemplified below).
From here, how a brand takes advantage of analytically-derived audiences for any recipe's use case (not just propensity scoring) can be activated across one or multiple channels. Here is a sampling of what is possible:
Once a touchpoint (or task) is selected (we will leverage Facebook Ads for this example), users can leverage 1st party customer data from a variety of options, including Audiences sourced from analytical activities originating from SAS Viya to support the best practices of responsible marketing.
After clicking on the Audiences tile button, users have the option to observe the Propensity Recipe Audience metadata.
In addition, users can preview the the audience data itself.
From here, users can select to target one or multiple segments that originated from the propensity analysis. For example, a gentle touch in delivering prescriptive Audiences to marketers can be to filter the propensity values associated with a Subject ID to only the customers that should be targeted. Instead of using the default output values (which include propensities both above and below a desired threshold), an analyst can simplify the experience for the marketer. For example, after interpreting the propensity analysis, an analyst may determine to filter out the low scores. Prior to uploading the Propensity Recipe Audience from SAS Viya to Customer Intelligence 360, an analyst would simply apply a filter to isolate the Subject IDs that are relevant for inclusion.
After applying the filter, an analyst can run the Audiences API to only load the High Propensity Audience.
Using descriptive Audience naming conventions relevant to a marketer's workflow can further minimize adoption and activation friction.
Advancing to the next step of the marketer's workflow, the user would set up the High Propensity Audience for Facebook Ads targeting in the same manner they would target any other customer segment or group.
Building upon this vision, Activity Maps for customer journey orchestration across multiple touchpoints is also "in-scope" as a value proposition on leveraging use case-driven recipes and activating attractive customer audiences.
In conclusion, the important takeaways from this article include:
- The introduction of Propensity Recipes in SAS.
- A detailed walkthrough of using/adapting a propensity recipe across SAS Viya and Customer Intelligence 360.
- Other recipes exist across the areas of acquisition, upsell, retention, next-best-action (or experience), recommendations, lifetime value, pricing personalization, attribution and more.
- If you are interested in leveraging any of these proposed recipes, please reach out to your SAS support team and the sharing can begin!
Remember, there is more to activation than just scoring your model!
Our vision at SAS is to serve as the market leader in advanced audience creation & targeting, independent of channel, for enterprise customers leveraging complex, disparate data sources and wishing to consistently deliver superior understanding into their customer journeys. In other words, we want to empower brands to practice responsible marketing.
Learn more about how SAS can be applied for customer analytics, journey personalization and integrated marketing here. For those who want to dive deeper into the current state of the customer analytics technologies ecosystem, check out fresh (and unbiased) research here.
