cFilter & pSALSA – Two Approaches to Product Recommendations

Learn Data Science
Teradata Employee

At Teradata Aster, we continue to bring the power of advanced analytics at scale to a broader base of users. Market basket analysis was one of the use cases we addressed early on when Aster Data was a standalone startup, and it’s an area where we continue to excel.

For example, let’s look at the analytic techniques behind the Product Recommender solution we recently rolled out for business leaders at retailers and e-commerce companies. Collaborative filtering (cFilter) and the personalized stochastic approach to link structure analysis (pSALSA) are two approaches to data-driven, customer-centric product recommendations that we are putting in the hands of businesspeople.

We’ll start with cFilter, as this is a great recommendation technique we’ve used for years. cFilter explores pairwise affinity. In the case of product recommendations, we use the function to explore all our historical market baskets and determine how frequently items are purchased together. Then we can calculate with statistical confidence the likelihood of a specific customer to purchase an item, based on the customer’s historical baskets.

That approach is great for companies like grocery chains with relatively stable product lines. But what about companies with a large and constantly changing product line (and thus limited pairwise history)? Consider clothing retailers that regularly evolve the items they carry based on seasons and trends.

pSALSA can provide a powerful approach in this case. The most popular implementation of pSALSA is perhaps as the analytic muscle behind Twitter’s “who to follow” feature. pSALSA leverages Aster’s graph engine to look for similar customers and makes recommendations based on what they’ve purchased. It considers each person a node and the products they’ve purchased connectors. It walks those connectors to find a set of customers similar to a specific customer, looks at what else the connected customers have purchased, and makes recommendations based on that.

So if you work for a large clothing retailer, and you have a new style of pants you are carrying but little pairwise purchase history for those pants, pSALSA will: recognize that Susan has purchased those pants; determine that Stephanie, Kendra and Stacie are similar to Susan; and recommend the pants to Stephanie, Kendra and Stacie.

Even better, there are some use cases where a blended approach using both cFilter and pSALSA may yield the best results. Give us a call if you want to discuss this.

Did I mention that we built a product recommender app for Teradata Aster AppCenter? Now business users can run these recommendations in two or three clicks and easily integrate the results with systems such as marketing automation or CRM solutions. Check out this recent Q&A to learn more.