**4. Proposed Method**

Figure 3 shows the overall system architecture of the proposed method. It has a modular BN that infers the target activity node from a child node, which infers the low-level context, and simple decision trees that infer evidence nodes of the modular BN (see Sections 4.2 and 4.3). When the training process starts and the raw sensor data from nine channels and its class information are entered, the system learns and constructs its decision tree and conditional probability table, as described in the Section 4.3. For the recognition, the trained decision trees obtain raw sensor data continuously and make an inference of the probability of their evidence node, and the modular BN infers gradually from the evidence nodes to the query node, the eating activity. If the probability of the query node is larger than the predefined threshold, the recognition result becomes 'eating'.
