Understanding personalization engines

Our research shows that 91% of consumers are more likely to make purchases online when the content is personalized for them. 

Customer experience personalization is a powerful tool that allows retailers and brands to improve the performance of their digital and physical touchpoints. Retailers can leverage personalization by targeting coupons to be printed at points of sale based on customer propensity to try new product categories, switch to a different brand or private label, or increase consumption of a private brand product.

They can also use personalized product recommendations on retail websites that leverage customer-to-product affinity scores to improve conversion rates and enable cross-selling, issue dynamic-offers targeting on retail websites based on cart abandonment scores, and run personalized email campaigns based on estimated product-replenishment cycles and customer-to-product affinity scores. These are all ways to use personalization to lift a private brand’s success rate online.

The above capabilities can be used both to improve manufacturer-sponsored campaigns and to promote private label products. Although retailers widely use machine learning methods to create decision-making models for these and other personalization use cases, the sophistication of customer behavior and diversity of communication channels are so high that most companies have a lot of room for improvement. 

Below, I want to share how advancements in machine learning enable the development of new types of personalization models and algorithms. We at Grid Dynamics have identified five important trends based on projects we completed for a number of Fortune 1000 companies between 2017-2020.

Trend No. 1: Prescriptive Models
Many traditional marketing analytics and personalization methods use predictive modeling to score customers. For example, customers can be scored according to their probability of converting on a website or switching to a different brand. The typical problem with this approach is that marketers struggle to choose the right action when they have multiple scores. There is a lot of interest in prescriptive models that recommend a specific optimal action with regard to a certain outcome, such as customer retention, and we expect companies to develop an increasing number of similar models.

Trend No. 2: Strategic Optimization
Traditional personalization methods focus on immediate (myopic) outcomes, such as improving click-through rates. In reality, most companies are interested in building long-term relationships with customers, so they take optimal marketing actions within this strategic context. This problem can be addressed using reinforcement learning methods that optimize sequences of actions rather than individual actions. Adopting such techniques is challenging on a practical level, but more and more companies are experimenting with the process.

Trend No. 3: Plug & Play Platforms
Another benefit of reinforcement learning is the ability to create plug-and-play personalization platforms that learn directly from production event streams. This sharply reduces the engineering and data-science effort associated with the development and productization of personalization models, but the industry lacks mature open-source platforms that enable this approach. Several reasonably good frameworks exist that a few companies already use in production, and we expect this trend to continue.

Trend No. 4: Hybrid Data
Many personalization and recommendation algorithms use only one type of data, such as behavioral histories or textual product descriptions. In practice, it is beneficial to combine multiple heterogeneous data sources, such as clickstream events, product images, and textual descriptions. This is challenging to accomplish efficiently using traditional methods, but companies are increasingly adopting deep learning methods that allow them to create hybrid-input recommendation and personalization systems.

Trend No. 5: Event-Level Models
Traditional modeling techniques rely on aggregated customer statistics, such as the total number of purchases over the last month. The disadvantage of this approach is that the temporal dynamics of customer activity are usually lost in aggregation. Deep learning models can consume raw event sequences and thus avoid the limitations of aggregation. We expect these methods to be widely adopted in the next couple of years.