The User Model

Our focus on personalized conversation suggests a fine-grained model of user preferences, emphasizing the questions a user prefers to answer and the responses he tends to give, in addition to preferences about entire items. We now describe that model in more detail. In later sections, we will describe how it influences item ranking and question ordering, which in turn determine how quickly the system can stop asking questions and start presenting items. In general, a user may tend to:

• answer questions about some attributes more often than others,
• provide some attribute values more often than others,
• choose some items more often than others,
• provide certain combinations of values more often than their independent distribution would predict, and
• accept either large or small amounts of value and item diversity.

All of these tendencies are influenced by the user's preferences, which in turn are captured by our user model. Attribute preferences represent the relative importance a user places on attributes (e.g., cuisine vs. price) while selecting an item. Preferred values show the user's bias towards certain item characteristics (e.g., Italian restaurants vs. French restaurants). Item preferences are reflected in a user's bias for or against a certain item, independent of its characteristics. Combination preferences represent constraints on the combined occurrence of item characteristics (e.g., accepts restaurants in San Francisco only if they have valet parking). Diversity preferences model the time that needs to pass between an item or characteristic being suggested again or the user's tolerance for unseen values or items. Item preferences are related to single items, whereas attribute, value, and combination preferences are applicable to the search for those items in general. Diversity preferences relate to both the items and the search. Currently, the ADAPTIVE PLACE ADVISOR models preferences that the user may have about attributes, values, and items, but not combination or diversity preferences.

 

Table 1: Example of a user model.

User Name		Homer
Attributes	wi	Values and probabilities
Cuisine	0.4	Italian	French	Turkish	Chinese	German	English
		0.35	0.2	0.25	0.1	0.1	0.0
Price Range	0.2	one	two	three	four	five
		0.2	0.3	0.3	0.1	0.1
...	...	...
Parking	0.1	Valet		Street		Lot
		0.5		0.4		0.1
Item Nbr.		0815	5372	7638			6399
Accept/Present		23 / 25	10 / 19	33 / 36			12 / 23

 

The former are easily captured by either probability distributions or counts, as illustrated in Table 1. The PLACE ADVISOR maintains a probability distribution to represent attribute preferences and independent probability distributions to represent preferences for each attribute's set of values. For attribute preferences, the system uses domain knowledge to initialize the weights; for example, price is usually considered as more important than parking. In the absence of such information, as is the case with value preferences, the system begins with a uniform distribution. The system represents item preferences as a ratio of the number of times an item was accepted to the number of times it was presented; this is initialized by assuming that all items have been presented and then accepted a large percentage (nine out of ten, or 90%) of the time. While this may cause updates (see below) to have a small effect and undesirable items to be suggested more than once, it has the effect of not quickly discounting alternatives early in the learning process. This in turn encourages the user to explore alternatives, allowing the system to learn more about additional items. In sum, item preferences represent the probability of the user accepting a particular item after it is presented, rather than representing a probability distribution over all items.