How to Build a Recommendation Engine without Machine Learning Experts?

Want to serve relevant content? YES! 

Have machine learning expert? NO.

Now what?

Before the machine learning era and personalized content, old timers pop the page with a list of most popular items of all time.

The common issues you'd notice with this approach:

1. Content Gets Old - Users finding same content every time they sign onto your service.
2. New Items Never Get Surfaced - Since the leader board may not normalize the popularity by time elapsed since publishing, great new contents have a hard time making it.
3. You Don't Know How to Pick your Battle - If you decide to let the new content take over, your collapsed conversion may give you a heart attack, especially if you have a less active user base.

How do you build a simple recommendation engine with the highest take rate and decent user experience? The answer lies in understanding your user behavior, in this particular case: Usage Pattern. 

Here is step by step to build a simple but high performing recommendation engine

1. Show Only Convertible Items - Remove items from the recommendation list if the user has already taken it in the past, or if a user has clicked through very recently and abandoned the flow thereafter.
2.  Segment your User Base - Analyze the distribution of user activity across a reasonable time period, say 4 weeks. Finding a cut off to group the users into high activity cohort and low activity cohort. For example people who are active once every other day will go to segment A and those who are active only once a week or longer go to segment B. There is no universal formula for that because it varies by the business you are in, use industry benchmark and your own data to make a decision.
3. Balance your Recommendations - Test and observe how segment B respond to best-selling catalog items. Since they don't log in as often, you make sure they see the most bullet proof offerings when they do. On the other hand, offer newer popular contents to users who frequent your shop in general as they have likely consumed fair share of the most popular content.

The above three steps to build an effective recommendation engine is for a catalog of moderate number of genres and/or relatively homogeneous user base. For users that may range from young teenagers to business users that could have zero overlap in their consumption, the above approach is not the best - well I highly doubt you want to serve them with one product anyway.

Follow me on twitter (@qindizhang) or add me to your Google+ circle, my friend! Let me know in the comment what other topics you want to read about!