Deep dynamic co-clustering of count data streams: application to pharmacovigilance

Co-clustering is a widely used technique that allows the analysis of complex and high-dimensional data in various domains. However, existing models mostly concentrate on continuous and dense data in fixed time situations, where cluster assignments remain unchanged over time. For example, in the field of pharmacovigilance, it is crucial to cluster in real time drugs and adverse effects simultaneously, facilitating the automation of safety signal detection processes. However, traditional co-clustering methods require all the data to be loaded into memory, which can be a challenge for large datasets or even impossible in certain scenarios. The proposed online co-clustering model is designed to overcome this challenge by processing the data incrementally, one step at a time. This work introduces a novel inference process for the latent block model that addresses the challenge of online co-clustering of sparse data matrices. To properly model this type of data, we assume that the observations follow a time and block dependent mixture of zero-inflated distributions, thus combining stochastic processes with the time-varying sparsity modeling. To detect abrupt changes in the dynamics we make use of a Bayesian online change point detection method on both cluster memberships and data sparsity estimations. The inference relies on an original variational procedure whose maximization step trains a LSTM neural network in order to solve the dynamical systems. Numerical experiments on simulated datasets demonstrate the effectiveness of the proposed methodology in the context of count data streams. Then, we fit the model to a large-scale dataset supplied by the Regional Center of Pharmacovigilance of Nice (France), providing meaningful online segmentation of drugs and adverse drug reactions.