📝 SummarXiv

Abstract

Learning the influence graph G of a high-dimensional Markov process is a challenging problem. Prior work has addressed this task when the process has finite memory. However, the more general regime in which the system probabilistically "jumps back in time" - so that the state at t+1 depends on a sample from a distant past t-d - remains unexplored. The process with probabilistic resets can be modeled as a Markov process with memory, but estimations become computationally expensive. To tackle this, we introduce PIMRecGreedy, a modification of the RecGreedy algorithm originally designed for i.i.d. samples. The proposed method does not assume memory, requires no prior knowledge of d, and recovers G with high probability even without access to the specific time indices at which such temporal jumps occur, and without imposing any constraints on the graph structures.