Constrained Query Optimization in Differentiated Feature Spaces

Abstract In recent years, with the increasing complexity of user demands, there is a growing need for hybrid multi-modal queries with both structured and unstructured constraints. A lot of constrained hybrid query methods were proposed to retrieval objects with similar features to the query while satisfying given structured attribute constraints. However, due to the overlook of attribution dimension differences and attribute ambiguity, these methods may exhibit reduced accuracy and limited robustness in a differentiated feature space. To address this problem, we propose an optimized constrained query method specifically for differentiated feature spaces. Firstly, interaction relationships between attributes are captured by Factorization Machines model. Then, a novel fusion distance metric is developed, incorporating both the Euclidean distance between feature vectors and semantic distance of the differentiated attributes. Furthermore, approximate matching is introduced to find the most suitable constraint conditions, finally achieving constrained query optimization in differentiated feature spaces. Experiments on four large-scale benchmark tests show the significant performance gains of our method in differentiated feature spaces. Our method is faster than the state-of-the-art hybrid query methods and achieves an average 98% recall @10 within one second under thousands of attribute constraints.