LMLFM: Longitudinal Multi-Level Factorization Machines
share
Selecting important variables and learning predictive models from high-dimensional longitudinal data is challenging due to the need to account for complex data correlation and expensive computation. In this work, we propose an extension of factorization machines, LMLFM, to deal with such longitudinal data. LMLFM is efficient, sparse, provably convergent and explainable. Specifically, LMLFM is the first multi-level model that can simultaneously select fixed effects and random effects while accounting for complex correlations in the data and non-linear interactions among variables. Experimental results with both simulated and real-world longitudinal data show that LMLFM outperforms the state-of-the-art longitudinal methods in terms of prediction accuracy with significantly lower false positive, using substantially less computational resources.
READ FULL TEXT