Recommendation Engines & Search

Links to external
academic references
(not from Paradigm Shift AI)

We are all familiar with modern recommendation engines:  they produce the product suggestion feeds on sites like Amazon, saying: "Based on your Purchase History", "Similar Items", or "Frequently Purchased Together". But did you know that between 30% - 60% of online retail sales are generated by a site's recommendation system? (see study to right). These systems also generate powerful off-site recommendations by email that take control of user engagement, allowing retail businesses to drive the arc of their customer's journey in ways that build satisfaction, trust, loyalty, and ultimately - of course - profitability.

At its heart, recommendation engine technology allows building highly personalized marketing for each user based on commonalities of their profile with other similar users, and based on commonalities of their past purchases (or indicators-of-interest) with other products or services. Sites like Amazon and Google Play have tens of millions of products to offer - and creating an algorithm that can choose (within milliseconds) the top ten most likely to generate a purchase is indeed no small feat, and a sales tool whose power cannot be overstated.

Recommender engines are closely related to search engines - as they are both intrinsically looking to find deep similarities between customers (usually called 'users') and products (called 'items'). Traditional recommender algorithms (for example, collaborative filtering) find implicit relationships between users or items without having to train on explicit metadata-like attributes. Newer algorithms, particularly deep learning-based, can find explicit attributes that are similar by comparing images of products, reviews or descriptions of products, or even sounds in musical tracks. 

Recent deep learning approaches to recommender systems
Reinforcement learning to optimize customer metrics

The ultimate business weapon in the battle to acquire & retain customers and increase Customer Lifetime Value (CLV) is being developed in the research teams of the most advanced recommender systems labs:  This is the ability to create algorithms trained to optimize not just across a SINGLE current customer interaction, but for the trajectory of multiple site visits, interactions, and purchases over a longer time.

For example, in some cases the placement of an ad view (called an "impression") that is most likely to maximize chances for a user to click through (CTR - or click through rate) or better yet to drive a purchase - is NOT necessarily the action most likely to maximize customer satisfaction or long term purchases (CLV). For example, some purchases may be likely but result in a certain type of customer not returning at a high rate due to predictable dissatisfaction. 

Reinforcement learning and related algorithms called multi-arm bandits are able to optimize to customer metrics factoring in the longer set of interactions in the arc of a "customer journey", metrics such a revenue over several months or years, frequency of return visits, lower rate of returns, and increasing value of purchases over time.

These advanced algorithms allow for creating the highest customer satisfaction & loyalty, while also generating the highest long-term profit and revenue.