ExCC

Exact partitioning method(s) for the Correlation Clustering (CC) problem

• Copyright 2020-21 Nejat Arınık

ExCC is free software: you can redistribute it and/or modify it under the terms of the GNU General Public License as published by the Free Software Foundation. For source availability and license information see the file LICENCE

• GitHub repo: https://github.com/arinik9/ExCC
• Contact: Nejat Arınık [email protected]

Description

ExCC aims at solving optimally the Correlation Clustering problem. It offers two different tasks: Obtaining a single vs. all optimal solution(s).

In both tasks, an ILP model to solve the CC problem can be constructed in two different ways: decision variables defined on 1) vertex-pair (Fv: vertex formulation type) or 2) edge (Fe: edge formulation type). If we denote n by the number of vertices in the graph and m by the number of edges, there are (n(n-1)/2) variables in Fv, whereas there are m variables in Fe. Nevertheles, the number of constraints in both methods is of polynomial and exponential orders, respectively.

Each mentioned task above can be performed with two resolutions methods: Branch&Bound (B&B) and Branch&Cut (B&C). In B&B, we just let Cplex solve it with B&B. In B&C, there are two successful applications of B&C in the literature:

1. Adding violated valid inequalities only at the root of the B&B tree through the Cutting Plane (CP) method and then proceeding to the branching phase as in B&B. In the literature, it is also called Cut&Branch.
2. Adding violated valid inequalities only for integer solutions during the branching phase.

According to the literature, the first B&C application is better suited for Fv (that we call B&C(Fv)), whereas the second one performs better for Fe (that we call B&C(Fv)). So, we have two successful resolutions methods, B&C(Fv) and B&C(Fe), to solve the CC problem. But, which one should be used?

In chapter 2 of [Arınık'21], some experiments are conducted based on unweighted random signed graphs to clarify this point. These experiments show that the choice of the formulation and its resolution method depends on the characteristics of the network at hand. When a network is sparse (resp. dense), the B&C(Fe) (resp. B&C(Fv)) better performs. Moreover, for a medium graph density, the B&C(Fv) type is less sensitive to increase in n than the other methods, hence preferable in this case.

To solve the CC problem we can either read a signed graph with a .G graph file format or import a Cplex LP file, where a ILP model is already recorded in a previous run. The advantage of doing the second option is that if we provide ExCC with a ILP model containing violated valid inequalities found during a CP method, then it amounts to skip the CP phase of the B&C method. So, it directly proceeds to the second phase: branching. This allows to gain a considerable amount of time.

Input Parameters

• formulationType: ILP formulation type. Either vertex for Fv or edge for Fe.
• inFile: Input file path. See in/net.G for the input graph format.
• outDir: Output directory path. Default "." (i.e. the current directory).
• cp: True if B&C (i.e. Cutting Plane method + branching) will be used. Default false.
• enumAll: True if enumerating all optimal solutions is desired. Default false. We call OneTreeCC this enumeration method in Chapter 5 of [Arınık'21], when the formulation type is Fv.
• tilim: Time limit in seconds for the whole execution process. Default -1, which means no time limit.
• tilimForEnumAll Time limit in seconds when enumerating all optimal solutions, except the first one. This is useful when doing a benchmarking with EnumCC for the OneTreeCC method.
• solLim Maximum number of optimal solutions to be discovered when OneTreeCC is called. This can be useful when there are a huge number of optimal solutions, e.g. 50,000. Default -1.
• MaxTimeForRelaxationImprovement: Max time limit for relaxation improvement in the first phase of the Cutting Plane method. This is independent of the time limit. If there is no considerable improvement for X seconds, it stops and passes to the 2nd phase, which is branching. This parameter can be a considerable impact on the resolution time. For medium-sized instances (e.g. 50,60), it might be beneficial to increase the value of this parameter (e.g. 1800 or 3600s). The default value is 600s. Moreover, it might be beneficial to decrease the default value to 30s or 60s if the graph is easy to solve or the number of vertices is below 28.
• lazyInBB: Used only for B&C method. True if adding lazily triangle constraints (i.e. lazy callback approach) in the branching phase. If it is False, the whole set of triangle constraints is added before branching. Based on our experiments, we can say that the lazy callback approach is not preferable over the default approach. Default false.
• userCutInBB: Used only for B&C method. True if adding user cuts during the branching phase of the B&C method or in B&B method is desired. Based on our experiments, we can say that it does not yield any advantage, and it might even slow down the optimization process. Default false.
• nbThread: number of threads.
• verbose: Default value is True. When True, it enables to display log outputs during the Cutting Plane method.
• initMembershipFilePath Default value is "". It allows to import an already known solution into the optimization process. Since we solve a minimization problem, the imbalance value of the imported solution is served as the upper bound. It is usually beneficial to use this option, when we possess some good-quality heuristics.
• LPFilePath Default value is "". It allows to import a LP file, corresponding to a ILP formulation. Remark: such a file can be obtained through Cplex by doing exportModel().
• onlyFractionalSolution Useful mostly for the Fv formulation type. Default value is False. It allows to run only the cutting plane method in B&C, so the program does not proceed to the branching phase
• fractionalSolutionGapPropValue Useful mostly for the Fv formulation type. It allows to limit the gap value to some proportion value during the cutting plane method in B&C. It can be useful when we solve an easy graph. Hence, we do not spent much time by obtaining very tiny improvement when the solution is already close to optimality. Default -1.
• triangleIneqReducedForm : Used only for the Fv formulation type. When it is set to true, this amounts to remove redundant triangle inequalities from the formulation. See [Miyaichi'18] for the definition of such redundancy. Default value is false, which keeps the whole set of triangle constraints. See Chapter 2 in [Arınık'21]. Note that removing redundant triangle inequalities can substantially accelerate the optimization process for finding a single optimal solution. However, if the goal is to enumerate all optimal solutions, then such removing can degrade the performance. This last point is briefly mentioned in Chapter 5 in [Arınık'21], but it needs to be investigated thoroughly in a follow-up work.

Instructions & Use

• Install IBM CPlex. The default installation location is: /opt/ibm/ILOG/CPLEX_Studio<YOUR_VERSION>. Tested with Cplex 12.8 and 20.1.
• Put /opt/ibm/ILOG/CPLEX_Studio<YOUR_VERSION>/cplex/lib/cplex.jar into the lib folder in this repository.
• Change the value of the Cplex installation path in line 34 of the file build.xml according to your Cplex installation.
• Compile and get the jar file for ExCC: ant -v -buildfile build.xml clean compile jar.
• Run one of the scripts .sh available in this repository.

Examples

See run-bb.sh, run-bb-enum-all.sh, run-cp-bb.sh, run-cp-bb-enum-all.sh, run-cp-only.sh, run-lp-bb.sh and run-lp-bb-enum-all.sh for more execution scenarios.

• run-bb.sh: Branch&Bound for finding a single optimal solution. It does not include any valid inequalities that can be obtained through Cutting Plane.
• run-bb-enum-all.sh: The same as run-bb.sh, but for enumerating all optimal solutions.
• run-cp-bb.sh: It corresponds to one of two successful B&C applications mentioned abovemethod depending on the formulation type. It is for finding a single optimal solution.
• run-cp-bb-enum-all.sh: The same as run-cp-bb.sh, but for enumerating all optimal solutions.
• run-cp-only.sh: only Cutting Plane method, i.e. strengthing the initial LP model with tight valid inequalities without proceeding to the branching phase. It is only for the Fv formulation type.
• run-lp-bb.sh: The same as run-bb.sh, but it reads the ILP formulation from a Cplex LP file (rather than a signed graph file). It is for finding a single optimal solution.
• run-lp-bb-enum-all.sh: The same as run-lp-bb.sh, but for enumerating all optimal solutions.

Example for B&C(Fv)

ant clean compile jar; ant -DinFile=in/net.G -DoutDir=out/net -DformulationType="vertex" -Dcp=true -DenumAll=false -DMaxTimeForRelaxationImprovement=120 -DfractionalSolutionGapPropValue=0.01 -DnbThread=4 -Dverbose=true -Dtilim=300 -DtriangleIneqReducedForm=true run

ant clean compile jar; ant -DinFile=in/net.G -DoutDir=out/net -DformulationType="vertex" -Dcp=true -DenumAll=false -DMaxTimeForRelaxationImprovement=120 -DfractionalSolutionGapPropValue=0.01 -DnbThread=4 -Dverbose=true -Dtilim=300 -DtriangleIneqReducedForm=true run

Example for B&C(Fe)

ant clean compile jar; ant -DinFile=in/net.G -DoutDir=out/net -DformulationType="edge" -DenumAll=false -Dcp=true -DMaxTimeForRelaxationImprovement=120 -DlazyCB=true -DuserCutCB=false -DinitMembershipFilePath="" -DLPFilePath="" -DonlyFractionalSolution=false -DfractionalSolutionGapPropValue=-1.0 -DnbThread=4 -Dverbose=true -Dtilim=300 run

Output

• The names of the optimal solutions are in the following form: 'solXX.txt', where XX is the solution id.
• The ILP model strengthed during the optimization process is exported into a LP file, if the underlying task is to find a single optimal solution. The name of such file is strengthedModel.lp. Moreover, if the Cutting Plane (CP) method is used (i.e. the input parameter cp), then an additional LP file is also exported just after the end of the CP (and before the branching phase) method and it is called strengthedModelAfterRootRelaxation.lp.

Remarks

• In the task of enumerating all optimal solutions, the number of optimal solutions found by B&C(Fv) and B&C(Fe) can be different (similarly, by B&B(Fv) and B&B(Fe)), if the underlying signed graph does not contain a single connected component defined on positive edges (e.g. the presence of a singleton vertex).

Acknowledgement

I thank Zacharie Ales for providing me with his code (for a similar problem), which constitutes an early version of this code.

References

• [Arınık'21] N. Arınık, Multiplicity in the Partitioning of Signed Graphs. PhD thesis in Avignon Université (2021).
• [Miyauchi'18] A. Miyauchi, T. Sonobe, and N. Sukegawa, Exact Clustering via Integer Programming and Maximum Satisfiability, in: AAAI Conference on Artificial Intelligence 32.1 (2018).