Priporočila delotokov za Orange z učenjem iz zbirke posnetkov zaslonov delotokov

This thesis presents a recommender system that suggests relevant widgets for Orange Data Mining workflows by learning from a curated collection of workflow screenshots. The dataset consists of 819 screenshots sourced from 25 chapters of Data Mining tutorials, created by various authors over time. We detect widgets in these screenshots using a two-stage computer vision pipeline: the Circle Hough Transform locates circular widget regions, and the Scale Invariant Feature Transform (SIFT) matches these regions against the Orange Widget Catalogue. The detection results — counts of each widget present in each screenshot — form a widget–screenshot matrix, which serves as the input for collaborative filtering. We train a Biased Regularized Incremental Simultaneous Matrix Factorization (BRISMF) model to predict the most suitable next widget for a partially built workflow. In evaluation, the BRISMF-based recommender achieved an average reconstructed position of 5.04, outperforming a frequency-based baseline (5.76) by 12.5%. These findings suggest the potential for automated recommendations in workflow design within visual programming environments.