Updating the User Model

Our main contribution is the addition of personalization to the above conversational recommendation model. The user model (Section 3.1) represents this personalization, but the ADAPTIVE PLACE ADVISOR must update it appropriately. While some adaptive recommendation systems [74,52,58,49] require the user to provide direct feedback to generate the user model, our basic approach is to unobtrusively derive the user preferences. Thus, the system does not introduce unnecessary interactions, but learns from the interactions needed to support item recommendation. We describe here how the system gathers item, attribute, and value preferences. As described more fully below, the system modifies feature and value weights [33,82,14,80] for the latter two, and increases the counts in the ratio of accepted to presented items. When determining the points in the dialogue at which to update the user model, we considered several factors. We wanted to enable the system to acquire information quickly, but to discourage it from making erroneous assumptions. We thought that users might explore the search space the most while constraining attributes, so we decided not to have the system update value preferences after each user-specified constraint. However, if we instead chose to only allow model updates after an item suggestion, the learning process might be too slow. The choices described below are, we feel, a good tradeoff between the extremes. The three circumstances that we chose for user model update were (1) after the user's ACCEPT speech acts in a SUGGEST-RELAX situation, (2) after the user's ACCEPT speech acts in a RECOMMEND-ITEM situation, and (3) after the user's REJECT speech act after a RECOMMEND-ITEM speech act by the system. First, we assume that when a user accepts an item, he is indicating: (1) a preference for the item itself, (2) preferences for the attributes he constrained to find this item, and (3) preferences for the values he provided for those attributes. Thus, when a user accepts an item presented by the system, the probabilities for the appropriate item, attributes, and values are increased. For the item preference, the system simply adds one to the presentation and acceptance counts. For attribute and value preferences, the system increases the probability of the appropriate weight by a small amount proportional to its current weight, then renormalizes all weights. Thus attribute and value preferences are biased measures that avoid zero counts for values that the user never chooses, as is typical for this type of probabilistic representation. Second, when a user rejects an item presented by the system, we only assume that he has a dislike for the particular item. We do not assume anything about the characteristics of that item, since the user has specified some of those characteristics. Therefore, for rejected items the system simply adds one to the presentation count. The third situation in which the system updates the user model is when, after the query has become over-constrained, it presents an attribute for relaxation and the user accepts that relaxation. In this situation, we assume that, had there been a matching item, the user would have been satisfied with it, since the characteristics specified in the conversation so far were satisfactory. Therefore, before the relaxation occurs, the system increases the attribute preferences for the constrained attributes and increases the value preferences for user-specified values, in a manner similar to an ACCEPT situation after a RECOMMEND-ITEM. This enables the ADAPTIVE PLACE ADVISOR to more quickly make inferences about the user's preferences.