This paper is based on the papers by S. Schlag et al [1], Liu et al [2], presented to us by HIT student Han Yang and discussed in October 2018 in Harbin.
Also we are aware about other papers on this topic, e.g. the survey by D. A. Papa and I. L. Markov [3].
The work by S. Schlaget al, “way Hypergraph Partitioning via Level Recursive Bisection”
In the paper [1], a multilevel algorithm for multigraph partitioning that contracts the vertices one at a time is developed. The running time is reduced by up to twoorders of magnitude compared to a naive level algorithm that would be adequate for ordinary graph partitioning. The overall performance is even better than the widely used hMetis hypergraph partitioner that uses a classical multilevel algorithm with few levels. Considerably larger improvements are observed for some instance classes like social networks, for bipartitioning, and for partitions with an allowed imbalance of 10%. The algorithm presented in this work forms the basis of the hypergraph partitioning framework KaHyPar (Karlsruhe Hypergraph Partitioning).
1. Introduction
A hypergraph is a generalization of a graph, where each (hyper)edge can connect more than two vertices. The way partitioning problem for a hypergraph generalizes the wellknown problem of graph partitioning:
How to divide the set of vertices into disjoint parts with sizes not exceeding of the average block size, while the cost function, i.e. the sum of wieghts of all hyperedges that connect different parts should be minimized.
Hypergraph partitioning (HGP) has a lot of applications. The two important areas of applications are VLSI circuit design and scientific calculations (e.g. speeding up sparse matrixvector multiplications)
[3]. While the first one provides an example, where minor optimization can give sufficient effect, in the second one, modelling based on hypergraphs is more flexible than that based on graphs [5], [6], [7], [8].As the hypergraph partitioning is an NPhard problem [9] and as it is NPhard even to find a good approximate solution for graphs [10]
, heuristic algorithms are usually used. The most often used heuristic algorithm is the multilevel paradigm, which consists of three phases: In the coarsening phase, the hypergraph is recursively coarsened to obtain a hierarchy of smaller hypergraphs that reflect the basic structure of the input. After applying an initial partitioning algorithm to the smallest hypergraph in the second phase, coarsening is undone and, at each level, a local search method is used to improve the partition induced by the coarser level.
2. Combinatorial formulation of the problem
Since the problem of hypergraph partitioning is formulated approximately, we suggest to replace the original problem by its approximation from the very beginning.
So, we start with a hypergraph , consisting of a finite number of vertices and a finite number of hyperedges. Each hyperedge is given by its ends, which are connected by this hyperedge, i.e. by a finite subset In particular, among the hyperedges, there may be simplest edges, that connect only two vertices, i.e. such hyperedges that .
The partitioning problem for a hypergraph can be formulated as follows: to find subsets … such that:

they are disjoint;

up to ;

the number of hyperedges that connect vertices from different subsets is minimal.
3. Reduction of the combinatorial problem to a continuous problem
Consider first the simplest case of the combinatoriaal problem, when the hypergraph is a classical onedimensional graph, i.e. all edges are onedimensional.
Consider then the simplex generated by vertices . Everything happens on the space . Each vertex is identified with the deltafunction on with this vertex being its support. Therefore, we may replace the set of vertices by the space of functions . If the graph is partitioned into two parts, then, instead of these parts, , , we consider two functions, , such that
The requirement can be replaced by the requirement
or, approximately, by
The size of a part is measured by the integral
which should be approximately equal to the average part size, i.e.
Each edge of the graph can be described as a function on the Cartesian product . This function should approximate the edge by using the support of the function . Then the number of edges connecting the two parts can be written as
Therefore, the problem is reduced to minimizing the integral
Summing up, the problem reduces to the following one: Find functions and , , satisfying the conditions:

The condition of disjointness:

The condition of almost equal sizes:

minimizing the integral
The formulation of the problem can be naturally transferred to the case of hypergraphs, where edges are replaced by hyperedges, and the number of parts can be greater than two.
4. Solution of the analytical problem
Note that the condition of disjointness is of different nature than the two other conditions, namely, it should be checked at each point of separately, while the two other conditions are integrals. We may replace the disjointness condition by a weaker one:
In this way we may get a few points, where both and are not small, but the number of such points cannot be too great.
Let us also replace by , and the problem reduces to that of finding a function such that satisfies the two conditions:
and minimizes the integral
This can be written in a matrix form. Let denote the matrix of , the vector with all coordinates equal to 1. To simplify the notation, let also Then the above conditions are:
This can be solved by using the Lagrange multipliers method. We have to minimize the functional
The critical points of this functional satisfy
for any .
When the matrix is symmetric (which is natural for adjacency matrices), we may rewrite this as
hence the critical points of the functional should satisfy
together with
and
References
 [1] S. Schlag, V. Henne, T. Heuer, H. Meyerhenke, P. Sanders, Ch. Schulz, way Hypergraph Partitioning via Level Recursive Bisection 2016 Proceedings of the Eighteenth Workshop on Algorithm Engineering and Experiments (ALENEX), p. 53–67.
 [2] H. Liu, P. LePendu, R. Jin, and D. Dou A Hypergraphbased Method for Discovering Semantically Associated Itemsets 2011 11th IEEE International Conference on Data Mining, p.398–406.
 [3] D. A. Papa and I. L. Markov Hypergraph Partitioning and Clustering, 2011 11th IEEE International Conference on Data Mining, p.398–406. In T. F. Gonzalez, editor, Handbook of Approximation Algorithms and Metaheuristics. Chapman and Hall/CRC, 2007.
 [4] C. Curino, E. Jones, Y. Zhang, and S. Madden, Schism: A Workloaddriven Approach to Database Replication and Partitioning, Proceedings VLDB Endow., 3(12): 48–57, September 2010.
 [5] B. Heintz and A. Chandra, Beyond graphs: Toward scalable hypergraph analysis systems. SIGMETRICS, Perform. Eval. Rev., 41(4):94–97, April 2014.
 [6] Ü. V. Catalyürek and C. Aykanat, Hypergraph PartitioningBased Decomposition for Parallel Sparse Matrix Vector Multiplication, IEEE Transactions on Parallel and Distributed Systems, 10(7):673–693, July 1999.
 [7] B. Hendrickson and T. G. Kolda. Graph partitioning models for parallel computing. Parallel Computing, 26(12):1519–1534, 2000.
 [8] S. Klamt, U. Haus, and F. Theis. Hypergraphs and Cellular Networks. PLoS Comput. Biol., 5(5): e1000385, 05 2009.
 [9] T. Lengauer. Combinatorial Algorithms for Integrated Circuit Layout. John Wiley & Sons, Inc., 1990.
 [10] Thang Nguyen Bui and Curt Jones. Finding Good Approximate Vertex and Edge Partitions is NPHard. Information Processing Letters, 42(3): 153–59, 1992.
Comments
There are no comments yet.