In part one of this blog, we dis­cussed the require­ments for the foun­da­tion of an effec­tive rec­om­men­da­tions sys­tem: the abil­ity to make deci­sions based on con­tex­tual sim­i­lar­ity and also affin­ity between items based on other fac­tors, such as user behav­ior, all dri­ven by rich user data. Today, we’ll talk about how lever­ag­ing machine learn­ing in rec­om­men­da­tions can gen­er­ate sig­nif­i­cant rev­enue lift by pit­ting algo­rithms against each other and learn­ing to apply those that yield opti­mal results.

There are a lot of ways to power rec­om­men­da­tions beyond sim­ply rely­ing on con­tent sim­i­lar­ity. You can trig­ger them based on what else the customer—or other cus­tomers in that segment—viewed, bought, or liked (cus­tomer X bought item Y and also viewed or liked item Z, etc.). You can take past behav­ior into account or limit the deci­sion to the cur­rent ses­sion as well.

It’s not a “one size fits all” approach, and one algo­rithm is not likely to work for every use case. For instance, the women’s apparel mar­ket is dif­fer­ent than the user intent behav­ior for car buy­ers, which is in turn dif­fer­ent from peo­ple look­ing for finan­cial ser­vices prod­ucts.  Each unique shop­ping ses­sion has a dif­fer­ent user mind­set asso­ci­ated with it, which is depen­dent on a com­plex set of char­ac­ter­is­tics rang­ing from the ver­ti­cal of the orga­ni­za­tion, nature of the offers being pre­sented, sense of urgency that is asso­ci­ated with the cam­paign, and other factors.

To opti­mize rec­om­men­da­tions, Adobe Tar­get applies mul­ti­ple algo­rithms to dif­fer­ent sam­ples of your vis­i­tors, and iden­ti­fies which algo­rithm per­forms the best by seg­ment and other vari­ables, and then auto­mat­i­cally favors the best-performing algo­rithms. It also con­tin­ues to explore the user behav­ior by ongo­ing exper­i­men­ta­tion. This is the goal of per­son­al­iza­tion: to show vis­i­tors the most rel­e­vant con­tent that has the most poten­tial to increase your rev­enue per vis­i­tor or con­ver­sion ratio.

Using machine learn­ing with strate­gic test­ing in this man­ner can help you increase rev­enue even from vis­i­tors who haven’t pre­vi­ously pur­chased. Some users are sim­ply wait­ing for the right product—by using opti­mized rec­om­men­da­tions, you can uncover what they’re most inter­ested in.

When eval­u­at­ing rec­om­men­da­tions solu­tions, look for one that will give you the flex­i­bil­ity to test algo­rithms against each other and let them “fight it out,” instead of just spec­i­fy­ing rules. You’ll be glad you did; we’ve seen cus­tomers gen­er­ate a 10% rev­enue lift based on rec­om­men­da­tions alone.

0 comments