Ubiquitous computing becomes a more fascinating research area since it may offer us an unobtrusive way to help users in their environments that integrate surrounding objects and activities. To date, there have been numerous studies focusing on how user's activity can be identified and predicted, without considering motivation driving an action. However, understanding the underlying motivation is a key to activity analysis. On the other hand, user's desires often generate motivations to engage activities in order to fulfill such desires. Thus, we must study user's desires in order to provide proper services to make the life of users more comfortable.

In this study, we present how to design and implement a computational model for inference of user's desire. First, we devised a hierarchical desire inference process based on the Bayesian Belief Networks (BBNs), that considers the affective states, behavior contexts and environmental contexts of a user at given points in time to infer the user's desire. The inferred desire of the highest probability from the BBNs is then used in the subsequent decision making.

Second, we extended a probabilistic framework based on the Dynamic Bayesian Belief Networks (DBBNs) which model the observation sequences and information theory. A generic hierarchical probabilistic framework for desire inference is introduced to model the context information and the visual sensory observations. Also, this framework dynamically evolves to account for temporal change in context information along with the change in user's desire.

Third, we described what possible factors are relevant to determine user's desire. To achieve this, a full-scale experiment has been conducted. Raw data from sensors were interpreted as context information. We observed the user's activities and get user's emotions as a part of input parameters. Throughout the experiment, a complete analysis was conducted whereas 30 factors were considered and most relevant factors were selectively chosen using correlation coefficient and delta value. Our results show that 11 factors (3 emotions, 7 behaviors and 1 location factor) are relevant to inferring user's desire.

Finally, we have established an evaluation environment within the Smart Home Lab to validate our approach. In order to train and verify the desire inference model, multiple stimuli are provided to induce user's desires and pilot data are collected during the experiments. For evaluation, we used the recall and precision methodology, which are basic measures. As a result, average precision was calculated to be 85% for human desire inference and 81% for Think-Aloud.