User Model

To improve the user model, we first plan to add more types of preferences. As we discussed in Section 3.1, combination and diversity preferences can capture more complex user behavior than our current model, and we plan to incorporate both into the next version of our system. Combination preferences can help it better predict either values or acceptable attributes, based on previously provided constraints. The PLACE ADVISOR can model value combination preferences by learning association rules [3] or extending to a Bayesian network, either of which would then influence the query, in turn influencing the similarity calculation and the case base. For preferences about acceptable attribute combinations, the system can learn conditional probabilities based on past interactions and use this to influence attribute ranking. While ``drifting'' preferences are not likely to cause problems in item selection applications as they might in ones like news updates, our model could be extended to handle within-user diversity. One way to do this is to capture the user's desired time interval between the suggestion of a particular item or value. We can calculate this by determining the mean time interval between a user's explicit selection or rejection of a value (value diversity preferences) or item (item diversity preferences). We will incorporate both these diversity preferences into the similarity calculation from Section 3.2 by extending R_I and P(V_j) in that equation to incorporate time effects. We define R_D(I) and P_D(V_j) as:

$$R_D(I) = R_I \times {1 \over {1 + e^{-k_I(t-t_I-t_{ID})}}}$$

$$P_D(V_j) = P(V_j) \times {1 \over {1 + e^{-k_{V}(t-t_{V}-t_{VD})}}} ~ ,$$

where t is the current time, t_I and t_V are the time when the item or value was last selected, and t_ID and t_VD are the time differences the user wants to have between having the item or value suggested again. R_D and P_D are in form of a sigmoid function where k_I and k_V determine the curve's slope. One empirical question is whether users also have attribute diversity preferences. We hypothesize that diversity preferences differ for each value of each attribute, and that this implicitly overrides attribute diversity. For example, a user may have different preferences about the frequency with which expensive restaurants versus cheap ones are suggested, but may not care about how often questions about price are asked. We plan to investigate this hypothesis. There are other improvements we might add to our user modeling technique. For example, the system may learn more quickly if it updates the user model in dialogue situations other than the current three. Also, using collaborative user models to initialize individual models could speed up the learning process. A more explicit combination of collaborative and individual user models [57,41] is also a viable direction to explore.