Techniques for training machine learning models using actor data

Abstract

Systems, methods, and articles of manufacture are disclosed for learning models of movies, keywords, actors, and roles, and querying the same. In one embodiment, a recommendation application optimizes a model based on training data by initializing the mean and co-variance matrices of Gaussian distributions representing movies, keywords, and actors to random values, and then performing an optimization to minimize a margin loss function using symmetrical or asymmetrical measures of similarity between entities. Such training produces an optimized model with the Gaussian distributions representing movies, keywords, and actors, as well as shift vectors that change the means of movie Gaussian distributions and model archetypical roles. Subsequent to training, the same similarity measures used to train the model are used to query the model and obtain rankings of entities based on similarity to terms in the query, and a representation of the rankings may be displayed via, e.g., a display device.