Twitter: 1, Robots: 0

July 29th, 2009

Types of Recommendation

Currently, systems deliver recommendations (like search results, products, or compelling content) based on two types of user input: explicit and implicit.

What You Tell the Robots

Explicit input is what you do to purposefully tune a system. You might opt out of tech news and opt in to sports. You might type in a couple of your key stock symbols. Explicit input is configuration. The cynical idea behind configurable services is that users will interact more with a tool they’ve spent time customizing. This can be true, but it can also alienate non-tinkerer audiences. More importantly, forcing someone to create their own portal removes any kind of authority from the service in question. It exposes the robot behind the curtain.

The Robots are Watching You

Implicit input is what you do without knowing you’re tuning a system. Where you click, what you search for, and what you ignore all contribute. Coming into a magazine site from a particular Google search might change the content you see on the page. Banner ads are usually served based on what you’re reading and what the ad network knows about you. On Amazon, anything you look at feeds into the recommendation system that shows you products that (may) interest you. In fact, I love pranking people by sending links to enema kits on Amazon. It pollutes their recommendations for a while. This is implicit input at work.

The Robots are Watching Everyone

An additional way that sites make recommendations is based on the stuff you tell them you want (explicit input), the stuff you show them you want (implicit input), and the input from other people like you. There are some sweet algorithms out there for determining that people are alike and using the tendencies of members in a group to recommend things to other members in the group.

In theory, the idea is compelling. Unfortunately, most of these algorithms end up turning individuals into caricatures of themselves or of their demographic group. It’s infeasible for a system to know enough about an individual to make consistently great recommendations. No computer will ever be able to recommend a book or a date or a pair of shoes with the same accuracy as a friend. More importantly, they can’t deliver the same accuracy as a casual acquaintance.

Humans can fill in gaps and infer rich meaning from subtle body language, tone, and personal experience. Those factors are noticeably absent in most online recommendation, search, and content delivery systems.

The Perfect System

I don’t think people want to interact much with systems to find information. Weird view, I know.

Some systems are certainly popular (Google) and and very successful (Amazon), but the affinity people feel for those systems is a form of Stockholm Syndrome.

What we all want is to have our grandfather (or an equivalent symbol of infinite wisdom and no-bullshit-taking) sitting on our shoulder, pointing us in the right direction at all times. We don’t need confidence in recommendations (of content, purchases, people) as much as faith. We want to Ask Jeeves, but we want Jeeves to have a clue – not a hundred page slush pile of possible hits.

Bad or boring recommendations lead to a lack of faith in recommendation systems in general and contribute to an overall fatigue. Recommendations need to inspire either confidence or faith. Confidence comes when you’re 100% accurate at all times. Faith comes when I feel like you understand me. Your “you might like…” recommendations are something I care about because I empathize with you. Confidence is for robots. Faith is for humans.

Enter Social Networks

There is a lot of potential in the networks of people we curate online, the reviews we write, the reservations we make, the choices we make on the spot in stores. The data doesn’t say who we are and how we feel, and no machine can step in to fill in the gaps and make the proper inferences.

The best system will have hints of the “#lazyweb” approach to Twitter. I ask my curated audience for advice. Even if I don’t know them all personally, and thus have little confidence in their individual experience and knowledge of my personality, I have high faith in the mechanism. I ask, and I receive. From people.

It’s a model as old as language. Until computers can empathize, algorithms need to shift their focus. Instead of attempting to make recommendations, technology should attempt to facilitate recommendations.

Twitter is on the right track with the new focus on search. I’ve long maintained that the best source of purchase advice is Twitter search. Twitter is a barometer. Nobody searches Twitter to find out definitive, high-confidence information. Instead, we search to get a feel for what people seem to know. We read short statements and either empathize with them automatically or shut them out. We have faith in our ability to read between the lines and derive answers. We use personal brands as clues to authority, bias, and brilliance.

It’s very personal and all too human.

With the insanity around the Yahoo and Bing/Microsoft relationship, I wonder if Twitter – or an app built around it – won’t become the search and recommendation tool we all really want.