June 18, 2018
Avinash Bharadwaj published in AI Informed

Customer Micro-Segmentation with Time-Series Analysis in Convolutional Neural Networks

By Avinash Bharadwaj, AI Solution Architect, Razorthink, Inc.

Traditionally, customer segmentation was based on static sources of historic data. Largely comprised of geographic and demographic data, such data points changed slowly over time for general, fairly obvious customer segmentation.

In our postmodern world, however, such approaches are no longer sufficient. People’s preferences change daily, requiring new methods and sources for targeted visibility into customer behavior.

Convolutional neural networks (CNNs) are able to perform predictive, time-series analysis on rich transaction data to identify customer behaviors. These behaviors are used for customer micro-segmentation, a much more preferable means of categorizing customers that leads to higher success rates for marketing campaigns and product development purposes.

Static vs. Dynamic

Customer segmentation for demographic and geographic data usually relies on clustering algorithms and feature engineering. However, the static nature of these methods yields only basic information for segmentation. CNNs can find much more subtle patterns in far greater amounts of data that dynamically changes. These neural nets have a lengthy history in image recognition, in which they identify patterns in images for classification purposes.

For example, CNNs can look at images of cats and determine features such as tails, four legs, whiskers, and so forth. They can also provide these advantages for time-series data. The latter’s data points are pixels, while the former’s are points—often involving values and time. By capturing the changes of values over time, CNNs predict possible outcomes of those values. For example, these neural networks can monitor changes in customers’ bank accounts to model the way those accounts are increasing. That information can form the basis for micro-segmentation to pinpoint specific offers and services of interest based on these behaviors.


Filters are an important concept that enables CNNs to perform time-series analyses. Filters are able to detect small patterns inside data that contribute to determining features—the relevant characteristics of machine learning models for specific business purposes. CNNs can issue several filters over extended periods of time, each of which find feature-related patterns at the point in time they are distributed. This filter process results in a stack of many features.

By aggregating those features, data modelers get high-level features which identify specific customer behaviors. Those behaviors, in turn, can be used to micro-segment customers across multiple factors other than just demographic and geographic data, resulting in extremely specific marketing measures. These behaviors are also influential for product development, providing insight into opportunities for organizations to create additional products and services based on demonstrated customer behaviors.

Real World Use Cases

The applications of customer micro-segmentation due to time-series analysis in convolutional neural networks are practically limitless. In the financial services industry, banks can use this approach to determine which customers are making large investments, then offer them specific opportunities to purchase life insurance based on their particular behaviors. This method is also valuable for identifying behaviors resulting in churn, enabling organizations to take action before customers reduce their services.

Micro-segmentation is especially effective in the telecommunications industry, in which companies have hundreds of thousands of customers demonstrating any number of behaviors. Deploying time-series analysis with CNNs is essential for categorizing those behaviors in ways that make sense over large segments of customer populations. For instance, this method could determine new plans for data or call usage based on customer behavior.

Similarly, it could indicate additional features or services that customers might want by illustrating which ones they currently use the most. Those within this industry could leverage micro-segmentation measures to pitch these value added services to all mobile subscribers, relying on micro-segmentation to understand, and possibly influence, the industry as a whole.

Over Time

CNNs are invaluable to the enterprise for their ability to perform time-series analysis on large sets of data. This method is critical to micro-segmenting customer behaviors and capitalizing on the knowledge gained from those behaviors in any number of industries.