Trigger-action programming (TAP) in the Internet of Things (IoT) enables users to easily customize the desired behaviors of connected entities, such as smart devices and online services, by creating trigger-action rules, also known as recipes. An example of such a rule is "If Sunset, Then Turn on lights". As potential trigger-action combinations grow exponentially, there has been significant interest in automatically generating recipes based on users' natural language instead of relying on manual creation. In this article, we present TAPFuser, an IoT data fusion framework designed to encode natural language and recipes, aiming to map human demand and recipes using a heterogeneous information network (HIN) embedding approach. Firstly, we divide TAP data into machine and human levels based on their sources. We then model these two levels of data as an IoT-HIN and employ HIN embedding techniques to learn vector representations of nodes, which is beneficial for recipe recommendation. To enhance the adaptive capabilities of TAPFuser, we propose a specific metagraph-guided random walk method that captures the level-aware heterogeneity of the IoT-HIN. Finally, we formulate the mapping between natural language generated by the human level and recipes generated by the machine level as a multi-label classification problem, where each node represents a natural language description, and the label corresponds to the category of the entity included in the recipe. We conducted multi-label classification experiments using the IFTTT dataset, and the results demonstrate the effectiveness of our proposed framework.
