Enhancing Classification and Retrieval Performance by Mining Semantic Similarity Relation from Data



Journal Title

Journal ISSN

Volume Title




When describing unstructured data, e.g., images and texts, humans often resort to similarity defining the characteristics of these data in relative terms rather than absolute terms. The subtle differences between such data can be indicated by a human easily while completely describing a single instance of them is a challenging task. For example, in an image retrieval task, to determine if two images are describing the same object, humans may simply ignore the differences in illumination, scaling, background, occlusion, viewpoint and only pay attention to the object itself. On the other hand, describing an image with all its information is hard and unnecessary. Cognitive evidence also suggests that we interpret objects by relating them to prototypical examples stored in our brain. Thus, the similarity is a fundamental property and of great importance in classification and retrieval tasks alike. Metric learning is the process of determining a non-negative, symmetric, and subadditive distance function d(a, b) that aims to establish the similarity or dissimilarity between objects. It reduces the distance between similar objects and increases the distance between dissimilar objects. From the human’s perspective, metric learning can be viewed as determining a function that best matches the user interpretation of the similarity and dissimilarity relation between data items. In this dissertation, we explore the possibilities to enhance the classification and retrieval performance by mining semantic similarity relations in data via metric learning. Unfortunately, existing metric learning solutions have several drawbacks. First, most metric learning models have a fixed model capacity that cannot be changed for adaption to input data. Second, existing online metric learning models learn a linear metric function which limits the model’s expressiveness. Third, they usually require a user-specified margin sensitive to input data and ignore a lot of failure cases during learning. To address these drawbacks, we propose a novel online metric learning framework OAHU that automatically adjusts model capacity based on input data, and introduce an Adaptive Bound Triplet Loss (ABTL) to avoid failure cases during learning. On the other hand, as an important subarea of classification, imbalanced classification is critical to the success of many real-world applications, but few existing solutions have ever considered utilizing data similarity to assist imbalanced learning. Based on this observation, we introduce a novel framework named SetConv, which customizes the feature extraction process for each input sample by considering its semantic similarity relation to the minority class to alleviate the model bias towards the majority classes. We also incorporate metric/similarity learning into a novel open-world stream classifier SIM to handle classifications on open-ended data distribution. Based on our research, we demonstrate that mining semantic similarity relation in data is critical to improving the performance of real-world classification and retrieval tasks.



Data mining, Machine learning, Learning classifier systems