Recommender systems represent an important area of data science and can be a great source of value creation.  Most of us are already familiar with recommendation engines through the product recommendation engines used by online retailers.

But recommendation systems are not just for helping your customers discover new products. Recommendation engines can be used to help grow wallet share through repeat purchases, as well as make marketing efforts more efficient and impactful.

In this article we will cover some of the key applications for recommender systems as well as underlying approaches.

Wallet share growth

One key application for recommendation systems is wallet share growth. There are two key aspects: product discovery and renewals.

Product discovery focuses on introducing customers to new products that they might be interested in purchasing and have not purchased in the past.

Renewals focus on making recommendations for products that a customer has purchased in the past and is likely to purchase again. With renewals, recommendations have a key timing component. A recommendation for renewals identifies those products that are most likely to be purchased again within a turn certain time frame – typically measured in days or weeks.

Once created, recommendations may be leveraged in a variety of ways from sorting products used in customer product searches to the selection and timing of advertisements and special offers

Key approaches for recommendation

The science of recommender systems includes a variety of algorithmic approaches. Some of the key approaches for recommendation are collaborative filtering, Buy Till You Die (BTYD) models, and Deep Reinforcement Learning (DRL). Each approach has different strengths and weaknesses as well as applications.

Collaborative filtering

Collaborative filtering represents a group of algorithms used in product discovery. In collaborative filtering, comparisons are made on an item-to-item or customer-to-customer basis to infer the likelihood that a customer will like or buy a new product that they have not purchased in the past.

Collaborative filtering example. Learn more about personalized recommenders here.

In the customer-to-customer comparison, the logic goes like this: the purchase history of Customer A is most like Customers B and C. This type of similarity may be quantified by a distance metric such as cosine similarity. Since Customers B and C have similar product preferences to Customer A, there is a good chance that Customers B or C have purchased products that Customer A would also like but has not purchased yet. Those products become recommendations for Customer A.

Buy Till You Die (BTYD)

Buy Till You Die (BTYD) models represent a group of algorithms that are useful for promoting product repurchases. BTYD models are based on previous customer behavior – specifically the Recency, Frequency, and Monetary Value of past purchases.

BTYD models are statistical in nature and can be used to estimate the likelihood that a customer will make a purchase within a window of time. They can also be leveraged to make predictions on a product level, providing insights into likely future purchasing behavior.

Both collaborative filtering and BTYD models are trained using data. In the case of collaborative filtering, models may be based on customer ratings or previous purchases. BTYD models are based primarily on previous purchase behavior.

Deep Reinforcement Learning (DRL)

When data is limited or not available, BTYD models and collaborative filtering become less effective. In such cases Deep Reinforcement Learning (DRL) may be a viable alternative. DRL agents are trained through interaction with a simulator or live environment.

To illustrate the concept, consider a DRL agent designed to optimize advertisements shown on a website with the goal of maximizing click-through rate. The DRL agent starts by naively selecting ads to show customers. The agent receives feedback by customers clicking on the most relevant ads. That feedback is fed into a reward function that the DRL agent is trained to optimize. Over time, the DRL agent improves its ability to select ads that are more attractive to customers on the website.

Depending on the problem DRL agents can require tens of thousands to millions of interactions to fully train. As such they are most suitable for scenarios where the cost of poor decisions is low. With the right investment, DRL agents can achieve state of the art performance.

Choosing the best recommendation system for your business

Recommendation systems represents a versatile set of algorithms and approaches that can have a significant positive impact on business outcomes. They can help increase wallet share, marketing effectiveness, and strengthen relationships with end customers. For a discussion on how recommendation engines can help improve your business please contact us.

Further reading

• How to build a personalized recommender system using data science
• “Counting Your Customers” the Easy Way: An Alternative to the Pareto/NBD Model
• Overview of collaborative filtering algorithms
• DRN: A Deep Reinforcement Learning Framework for News Recommendation