In my next instalment I plan to look a little more closely at the broader context of recommendation engines and how they might be used but foe now I will simply point out that they are broken into two broad types Content Based and Collaboration Based. The first refers to product knowledge and the second to behaviour of customers. Today I am only discussing the latter type
There’s two fundamental types of customer information used in recommendation engines and customisation strategies: 1. Real time knowledge of customer in context and 2. Historical knowledge of people with context.
I like to refer to them as 1. Personal and 2. Collective.
The first problem you run into when you start an ecommerce store is that you don’t have any historical data to start from. Making decisions about “people” is another way of describing prejudice. Of course all prejudice is not bad, we help old ladies across the road and we are kind to children. But when somebody hears my accent and offers me a particular brand, I am irritated.
What I find especially infuriating is when I simply cant find what I want on Google because it thinks it knows what I really want based on the fact that I looked for it twice last week.
The point I am making is that people are individuals, they change their minds, they like variety, they find what they are looking for and loose interest, in fact they really do not want to see what they just bought on offer at a better price.
The bottom line here is that relationships have always been difficult even when you are at the next desk or sharing a home, let alone seeing a cookie form the same machine create different click streams and view different content.
Personalisation and making recommendations based on algorithms can be high risk and must be done with care and forethought.
Context is a critical factor in storing and classifying interactions for recommendations and personalisation. When I go in our local store the shopkeeper who knows me very well takes into account the day of the week, time of day, look on my face, which shelves I’m looking on and what he remembers about my usual purchases and he then decides whether to let me walk out or to ask if he can help. He rarely gets it wrong.
If he did get it wrong very many times, I would be very tempted to walk a few yards extra and shop anonymously. Why should I have to explain myself?
Imagine if he wiped lots of stuff off the shelves on the basis that I won’t be interested in them today, but filled the shelves with what I bought on this day last week, or on my last visit. Not very good.
Imagine if I run a social network, I know a great deal about where you are going, what you like etc. If I provide your phone network, I know who you call, and receive calls form, when you use it and how long for, when you never use it and when it is off and where you are right now. Maybe I see your calendar and your emails too? Not to mention half a dozen spook outfits eavesdropping.
With this kind of information I could be very invasive, so it is critical to only use this type of data with the customer’s consent and to put that customer firmly in control in case they should change their minds. The way to achieve this is to offer them value in return for their data.
In a typical e-commerce implementation you will have access to the following types of information:
Purchases, wish list adds, cart adds, views, shares, favourites, downloads, shares, watches, likes, follows, where they were referred from and you may know about the search terms they used to get to your site.
You may also know about the sites these customers visited after they left your site when they have been active, their spending patterns, whether they buy for children, their gender, devices used, devices ordered from, devices emails were read on, where they have browsed from geographically, and a whole lot more. Once registered you could collect data on their social media activity.
If you are an OmniChannel retailer, you will have more data in your CRM, your ERP, and other places
This information exists in many places: In your customer database, in the web server log files, out there on social media sites, on Google, in an advertising network you work with,
In a typical solution, the information might be classified in three ways: 1. Customer , 2. info and context 3. Offer.
e.g. you create facts such as:
Customer, info, (info context), Offer.
John Brown, likes(3) (June 4 14, Saturday, mobile ), gizmo product
John brown, price search(June 6 14, Monday, PC), gizmo product
Each piece of information is about a specific offering or product and it relates, of course, to an individual customer. In addition to that basic information, it is invaluable to collect context.
Understanding that Mr Brown never orders on his mobile, but does look at info helps you decide what to present to him. Maybe the solution is to offer him a convenient app that makes ordering easier. Certainly if 80% of his browsing was on mobile devices and 100% of ordering was on a PC,, this would be a strong indication
The way you might arrive at this type of conclusion would be to query these stores of facts, summarise what we know and make inferences.
This is a lot like data warehouses where you store vast flat tables, analyse them once then stores summaries in a more accessible system.
e.g. John Brown, likes(3), gizmo product That weighted his interest at 3
Two days later he returned infer a higher interest, say 4
He asked for pricing info, this puts him at 8 or thereabouts
So we can summarise it as ; John Brown likes (June 6 14),gizmo product (8)
We know that he orders via PC, that was not a barrier.
We know that seeking price info is a sign of desire and conviction so why did he not order?
We might assume the price was too high. Perhaps we are all too aware that we are not the cheapest in this market and we rely on positioning to overcome price objections.
We may decide to send a once off special offer on Thursday morning, so he cant resist because he frequently browses then so he may have free time on Thursdays.
That is an example of generating a “Next action” on the basis of knowledge and intelligent inference.
Of course it could be that he planned to order on Friday at the full price, that is a price we either do or do not decide to pay for John Brown’s business. Eventually we will know enough to be able to predict this accurately.
If we knew that he was a highly influential blogger with a large following, I wonder would this influence our decision?.
Contrary to what is written by cloud vendors and Hadoop junkies, this kind of data and that kind of information is rarely dependent on petabytes of data spread across numerous clusters. It could be done with a basic RDBMS, though it would probably be easier with an RDF store downloaded for free.
Don’t get me wrong, there would still be some development to do and it wouldn’t work till a bit of data had built up, but you don’t have to be Google or Yahoo to benefit from this technique.
An interesting side effect of this approach is that the longer the customer keeps visiting the bigger investment you have in keeping her, because you know exactly how to twist her arm. Her customer lifetime value is infinitely greater even at a discount. Perhaps eventually, these relations will be reflected in the price each individual customer is asked to pay. There’s a thought to conjure with.
What if John turns up at your Branch in Oxford street with his decision made and ready to order? What if he orders online but wants to collect from your store in Paris next week? What if there’s an issue after delivery and he wants to get help form your store in Las Vegas while on his travels?
What if the cart tells him “sorry, we’re out of stock” and then he walks to the shops and sees one in your local store?
I know just how difficult these problems can be and how expensive it is to overcome them. It requires not just a technology rebirth, but a culture change like no other.