Multimodal Data Fusion in High-Dimensional Heterogeneous Datasets Via Generative Models

The commonly used latent space embedding techniques, such as Principal Analysis, Factor Analysis, and manifold learning techniques, are used for learning effective representations of homogeneous data., they do not readily extend to heterogeneous data that are a of numerical and categorical variables, e.g., arising from linked and text data. In this paper, we are interested in learning probabilistic models from high-dimensional heterogeneous data in an unsupervised. The learned generative model provides latent unified representations capture the factors common to the multiple dimensions of the data, and enable fusing multimodal data for various machine learning tasks. a Bayesian approach, we propose a general framework that combines data types through the natural parameterization of the exponential of distributions. To scale the model inference to millions of instances thousands of features, we use the Laplace-Bernstein approximation for computations involving nonlinear link functions. The proposed is presented in detail for the commonly encountered heterogeneous with real-valued (Gaussian) and categorical (multinomial) features. on two high-dimensional and heterogeneous datasets (NYC Taxi and-10M) demonstrate the scalability and competitive performance of the algorithm on different machine learning tasks such as anomaly, data imputation, and recommender systems.