On short-term behaviors

What does it take to identify visitors on an e-commerce site who are most likely to buy in that visit? And do so while they are still on the site, using just their first few clicks. Or pinpoint a dissatisfied customer just as they show the first sign of leaving the mobile app?

Not only are these types of predictions based on forecasting outcomes in short time windows, but they are also highly influenced by “in-the-moment” activity or short-term behaviors. These scenarios represent the next battleground for personalizing the customer experience, and it’s a blend of anticipation (machine learning models) and actions (content or other information to influence outcomes).

What are short-term behaviors?

Let us expand on the problem of identifying site visitors who will buy. In particular, we want to recognize visitors who will not buy under normal circumstances but can be influenced to buy. This problem can be framed as building an ML model that understands in-session behaviors, in order to predict outcomes within the same visit. Figure 1 illustrates this goal: to identify “persuadables” or those who are unlikely to buy unless influenced. These are the so-called on-the-fence shoppers.

We also need to identify those who are likely to buy, “sure things” and “do not disturbs,” so that we can leave them alone. Finally, knowing “lost causes” will help so we don’t try to convert them in this visit.

Figure 1. Using short-term behaviors for near-term outcomes

Why are short-term behaviors important?

The quest to anticipate consumer behavior is not new. Learning from past data has helped marketers address the “what” of consumer engagement — for example, matching audiences with product recommendations or advertising content. While past behavior is a good predictor of future outcomes, it also presents some challenges.

Consumer state changes rapidly, so predictions can easily get outdated.

Consumer behavior can vary significantly across multiple visits to the same site, depending on their purpose, the urgency of that need, and the time or place of accessing the site. Customer behavior today is also influenced by real-time trends and what’s popular. However, incorporating historical data such as past transactions and profile data such as demographics remains useful and must continue to be used for predictions. The combination of the two requires accounting more significantly for more recent data and behaviors —perhaps by emphasizing what’s happening in-session the most and deemphasizing activity from the past.

What good are AI predictions, if we don’t act.

We need to think about the actionability of intelligence. Businesses want to evaluate prediction-based actions in terms of objectives most important to them. Longitudinal time analysis addresses the what, and less so the when. From an actionability standpoint, it is entirely possible that actions based on only long-term predictions can be out of context.

Actionability requires timely predictions.

It’s important to make predictions soon after a trend sets in. Or understand the likelihood to purchase as soon as a customer arrives on your site. For first-time visitors, all you have is the opportunity within the first few clicks of the clickstream. By making a prediction early in the session, you gain the ability to show more visitors a persuasive offer before they are likely to leave the site. However, early prediction of purchase propensity is not trivial. We also need to take instant action on early predictions. Consumer behavior changes rapidly, so purchase predictions for a session must be used immediately to trigger actions, and need to be computed for every subsequent visit separately.

Sources of short-term behavioral data

Short-term user behaviors can be derived from 4 key contributions: in-session clickstream data, cross-session customer activity, aggregate site activity representing popular trends, and contextual data about users (like loyalty points balance, or cart contents) as well as site pages (like page metadata such as category, price, quantities on hand, etc.).

Figure 2. Key sources of short-term user behavior

Data democratization means we need to rethink how customer data needs to be collected and organized. What’s needed is a shift from CRM-style, centralized storage of customer data to a fundamentally decentralized and federated way of storing data. Also, in addition to understanding all the attributes that are part of a customer profile, it is important that their activity data be analyzed using data profiling techniques in order to extract insights from raw customer data.

Predictions using short-term behaviors

The application of machine learning to understand and analyze short-term behaviors along with historical data is key to create in-the-moment customer experiences. By recognizing short-term behaviors that are driving in-session outcomes, businesses can adapt their offerings using personalization and communications to what is happening with each consumer, in real-time.

In a manner similar to the mapping of the human genome, ZineOne has developed a new way to model short-term behaviors as an ordered sequence, we call this sequence the Customer DNA. As the encoding and analysis of human DNA revolutionized scientific insight, our patent-pending algorithms have sequenced spatial and temporal consumer traits, fundamentally changing our ability to anticipate behavior.

Figure 3. Real-time customer activity represented as an ordered sequence

ZineOne combines these machine learning-powered sequence models by industry to create highly effective “Industry Genomes.” This framework is then fed by real-time data streams and edge-based intelligence to detect patterns that match optimal, or suboptimal, activities by the consumer and in turn to make a continuous stream of determinations on how best to advance their journey. ZineOne’s architecture enables it to operate at both high speed and high scale so that the brands it is powering can deliver an unparalleled real-time experience to their consumers.

This can be used for instance to identify on-the-fence shoppers — site visitors who are less likely to buy — while they are still on the site. You can then focus on least likely buyers and give them a personalized nudge like a 30-min personalized offer based on their propensity to purchase. We have seen that these kinds of AI-driven offers create high impact leading to incremental revenue lifts of 20% to 30%.

Anticipating key outcomes in short timeframes provides new opportunities for taking a range of in-session actions to impact business gains. These range from promotional incentives to informational reminders and loyalty rewards, all of which can be delivered as personalized in-app or as post-visit messages. Some examples of these actions are:

• Personalized offers
• Personalized free shipping threshold — cart-based, department-specific, white-gloved, or flat rate
• Personalized loyalty rewards — bonus earnings or lowered redemption threshold

Brands are using short-term behaviors to enhance their customer experiences, removing friction points for both online and omnichannel shoppers, and focusing on tailored help, increased phone customer service support, and added flexibility to their return policies.

To see Speed to Sense in action, or learn more about AI-driven personalization, visit ZineOne.