IMAGE SIMILARITY USING LOGISTIC REGRESSION

Many machine learning algorithms, such as kernel machines, nearest neighbors, clustering, and anomaly detection, rely on distances or similarities to identify patterns in data. Before using these similarities to train a model, it is crucial to ensure they reflect meaningful relationships within the data. In this paper, we propose enhancing the interpretability of these similarities by augmenting them with explanations. To achieve this, we introduce Logistic Regression & Haar Cascade, a scalable and theoretically sound method designed to systematically decompose the output of pre-trained deep similarity models for pairs of input features. Our approach can be viewed as a composition of regression- based explanations, which previous research has shown to effectively scale to highly nonlinear models. Through extensive experiments, we demonstrate that Logistic Regression consistently provides robust explanations for complex similarity models. We also apply our method to a challenge in digital humanities: evaluating the similarity between historical documents, such as astronomical tables. In this context, Logistic Regression & Haar Cascade offers valuable insights and enhances the interpretability of a specialized, highly engineered similarity model.