HGSMAP: a novel heterogeneous graph-based associative percept framework for scenario-based optimal model assignment

The escalating pervasiveness of big data applications has incited the development of a multitude of models for the same objectives within the identical scenarios and datasets. Existing methods primarily focus on the allocation of tasks to workers, recommendation of items to users, and model selection in a fixed scenario. However, there currently lacks a unified framework for assigning optimal models to different scenarios. Furthermore, integrating heterogeneous information, uncovering their implicit associations, and conducting comprehensive evaluations of models remain significant challenges. To address this, a novel heterogeneous graph-based scenario and model associative percept framework (HGSMAP) is proposed. This framework incorporates a scoring function with three components: graph embedding, feature embedding, and the cross-feature relationship learning network (CFRLN). The graph embedding transforms the heterogeneous graph into lower-dimensional vectors by utilizing the metapath2vec++ method. The feature embedding learns vectors pertaining to model performance, as well as the node features of datasets and models in the graph by exploiting an embedding block. CFRLN integrates feature vectors derived from the feature embedding with their corresponding graph embedding vectors obtained from the graph embedding, extracts explicit and implicit dependencies among heterogeneous data, and employs an attention fusion block to intelligently fuse them. Six popular traffic scenarios are chosen as study cases and extensive experiments are conducted on a dataset to verify the effectiveness and efficiency of HGSMAP and the score function.