Jump to Content

Better Sliding Window Algorithms to Maximize Subadditive and Diversity Objectives

Michele Borassi
Proceedings of the 38th ACM SIGMOD-SIGACT-SIGAI Symposium on Principles of Database Systems (PODS 2019)
Google Scholar


The streaming computation model is a standard model for large-scale data analysis: the input arrives one element at a time, and the goal is to maintain an approximately optimal solution using only a constant, or, at worst, poly-logarithmic space. In practice, however, recency plays a large role, and one often wishes to consider only the last w elements that have arrived, the so-called sliding window problem. A trivial approach is to simply store the last w elements in a buffer; our goal is to develop algorithms with space and update time sublinear in w. In this regime, there are two frameworks: exponential histograms and smooth histograms, which can be used to obtain sliding window algorithms for families of functions satisfying certain properties. Unfortunately, these frameworks have limitations and cannot always be applied directly. A prominent example is the case of a submodular function with cardinality constraints. Some of these difficulties can be rectified, but often only on a case-by-case basis. Here, we describe an alternative approach to designing efficient sliding window algorithms for maximization problems. Then we instantiate this approach on a wide range of problems, yielding better algorithms for submodular function optimization, diversity optimization and general subadditive optimization. In doing so, we improve state-of-the art results obtained using problem-specific algorithms.