Approximation Complexity of Maximum A Posteriori Inference in Sum-Product Networks

About

We discuss the computational complexity of approximating maximum a posteriori inference in sum-product networks. We first show NP-hardness in trees of height two by a reduction from maximum independent set; this implies non-approximability within a sublinear factor. We show that this is a tight bound, as we can find an approximation within a linear factor in networks of height two. We then show that, in trees of height three, it is NP-hard to approximate the problem within a factor $2^{f(n)}$ for any sublinear function $f$ of the size of the input $n$. Again, this bound is tight, as we prove that the usual max-product algorithm finds (in any network) approximations within factor $2^{c \cdot n}$ for some constant $c < 1$. Last, we present a simple algorithm, and show that it provably produces solutions at least as good as, and potentially much better than, the max-product algorithm. We empirically analyze the proposed algorithm against max-product using synthetic and realistic networks.

Diarmaid Conaty, Denis D. Mauá, Cassio P. de Campos · 2017

Related benchmarks

Task	Dataset	Result
MAP Inference	Twenty Datasets	accidents1.6	15
MAP Inference	accidents 50% query size	Runtime (seconds)2.74	5
MAP Inference	adult 50% query size	Runtime (s)0.237	5
MAP Inference	baudio 50% query size	Runtime (s)7.2	5
MAP Inference	bnetflix 50% query size	Runtime (s)2.47	5
MAP Inference	book 50% query size	Runtime (s)5.02	5
MAP Inference	connect4 50% query size	Runtime (s)5.22	5
MAP Inference	dna 50% query size	Runtime (s)1.17	5
MAP Inference	jester 50% query size	Runtime (seconds)3.03	5
MAP Inference	kdd 50% query size	Runtime (s)0.638	5