Personalization is hot these days, widely seen ([1] [2] [3] [4]) as a key battleground in the battle between Microsoft, Yahoo, and Google. But what is personalization?
I'd define web personalization as delivering different and unique content to each individual customer based on the customer's interests. It's your own version of the web site, a site just for you.
To distinguish personalization and customization, I would argue that personalization uses implicit interests and customization uses explicit interests. Personalization learns what you like from your actions; you are what you click on and what you buy. Customization requires you to explicitly specify what you want; you are what you say you are. My Yahoo is an example of customization. You tell the site what you want. Amazon.com is an example of personalization. The site learns your interests and adapts.
In a world with a glut of information, personalization offers a way to find focus. It doesn't waste your time showing you what everyone else sees. It learns what you like, shows you what you want to see, and filters out the rest. That's personalization.
Saturday, June 05, 2004
Subscribe to:
Post Comments (Atom)
7 comments:
I think the key challenge is to provide a service that is an optimal mix of personalization and customization. We want the user to be able, but not have to, modify preferences to suit his current tastes. Since sometimes the underlying personalization model might be too simplistic.
Great point, Seun. A mix of customization and personalization can work well. Customization can allow users to correct mistaken guesses and assumptions by the personalization engine. It also gives power users the ability to tinker with the system. But it's important to realize that most people won't bother dealing with customization features, so they'll always only be applicable to a small percentage of your customers.
The vast majority of people do not use customization on web sites because the discovery of such features is low and the cognative load of using such features is high. This is an interface issue, not a problem with the concept of customization.
Humans use customization all the time. We decorate our living spaces, we order our kitchens etc.
When we walk into a store what is the first thing the sales person is likely to say?
"Hi, How can I help you."
We need interfaces that understand humans. Not the otherway around. If we create these interfaces the concept of "infered intent" will become a joke.
You make an interesting point, but let's be careful with our terms. From the examples you gave, asking a sales person for directions and a well organized room, I'd think the closest parallels would be search and categorized directories. Clearly, when someone knows what they want, it's important to get out of there way and let them find it.
But customization is not search and browse. Customization is manually specifying your individual interests to modify the way information is displayed on the site.
Given my definition, I don't agree that customization is merely an interface issue. The basic problem with customization is that it requires effort from the user. Users will not devote time to doing something that where the reward isn't obvious to them. Most will simply use the default.
Better interfaces may reduce the effort required and may be able to make the gain more obvious. But there will always be a large group of users, often the vast majority, who will not be willing to put in the effort.
I totaly agree with you that the app should get out of the user's way when the user knows what she wants and can express it. On the other hand, consider the following:
SP: "Hi, how may I help you?"
Cust: "I'm looking for a new digital camera."
SP: "Great, did you have a model in mind?"
Cust: "Uh, no, I'm not sure which model I need. I was hoping you could help me."
SP: "Great! Well, let me ask you a few questions ..."
Often times the user has a desire but is unsure of how to satisfy that desire. It is the job of the sales person to figure out a satisfactory solution in an efficient way. Notice how the interface of SP (recently upgraded to version 2.3) adapted to allow Cust to input the relevant information in an efficient manner.
Humans have been doing this for awhile now and we have gotten quite good at it by using speech.
The difference between human and machine interfaces is that humans engage in dialog, current machine attempts at this are kludgy at best; limited typically to query boxes, drop down menus radio buttons and result sets.
In part because it is difficult to allow the user to directly enter their desires or intents, inference techniques are employed in an attempt to exploit the minimal amount of information present in an "optimal" manner.
Of course, the other side of the coin is that users sometimes don't know they would want something until they see it. This is where inference shines (or anoys if you are a parent in the grocery store checkout line).
Those involved in developing these inference techniques should pat them selves on the back. They have enabled wonderful applications. But there is so much more that can be done. Inference and statistics alone will not get us there.
If we figure out ways to let the users explicitly state their desires in a natural way they will be more likely to get what they want, and when used, inference will become more accurate.
How can the application get out of the users way and let them express their intent? Even if they are not sure what they want, they know they want something and given the appropriate interface they should be able to discover how best to satisfy that desire. Humans often use dialog to adress this issue. Maybe computer interfaces can grow in that direction as well (without using anoying wizards).
Ultimately, personalization is about giving the person what she wants.
We're certainly in agreement on the end goal. The holy grail would be an automated expert system that could converse on various topics in natural language. But a system capable of this, capable of the kind of dialogue you describe, is incredibly challenging and many years away.
You did mention that inference techniques aren't the way toward this goal and that the user should be able to explicitly state desires and get answers. I think it's important to realize how ambiguous natural language dialogues are.
For example, in the conversation example you gave, the automated sales person has to make a lot of guesses about what the customer wants. Does "looking" mean that they want to buy a camera? Or that they lost a camera? Seems obvious to you, perhaps, but the only reason it's obvious is that you're applying knowledge to determine the meaning of the word from the context. Does "model" refer back to camera? What does "model" mean in this context? Does "help" refer back to "looking for a digital camera"? What would be most helpful? What information is necessary for this decision? What are the customer's goals? How can I match my information about various cameras to best satisfy (my wild guess at) the customer's goals?
Everyday conversation is ambiguous. Nothing is explicit or fully specified. Meaning isn't clear, information is incomplete and noisy, goals are unknown, models of the world differ. Inference and assumptions are how we through our day.
I know I am a preachin to the choir, but...
Speech is certainly full of ambiguities, much more than many people realize, yet humans have evolved the machinery to effectively deal with these ambiguities (most of the time). However, humans do not rely on rear-view statistical inference alone to resolve the ambiguities. In addition, humans consider the given information within the context of their own expierinces. You could argue that to some extent a form of statistical inference is used to inform the experience. However, statistical inference alone can not account for this ability.
You can not squeeze blood out of a turnip and you can not squeeze induction out of a Bayesian zealot.
I certainly agree that solutions to these problems are far off. Talk about job security!
I guess my point is that every once in a while we should look up from our daily achievements and see how far we have yet to go.
Post a Comment