This repository contains the code for both measuring faultlines according to our definition (which we present shortly) as well as the implementation of our algorithms for creating teams with low faultline potential.
Given a team of "k" people each with "f" different attributes, we define the faultline score to be the total number of bad traingles formed by the team. A bad triangle (with respect to a feature) refers to three individuals where two of them are similar with respect to the feature and one is dissimilar to the other two. For instance, if we consider the occupation as a feature, then two programmers and one graphic designer form a bad triangle.
Bad triangles serve as a good proxy for faultines as they practically measure to what extent similar groups of people can oppose the others who are different from them. Notice that this definition allows for diversity. For instance, a programmer, a graphic designer, and a sales person do not form a conflict trianlge.
Note that the number of conflict traingles can be computed by considering all triangles. However, this is time-consuming in practice. We present a more efficient way for counting such triangles. The code for computing the faultline score for a given team can be found in here.
Our second contribution is building an algorithm that partitions a group of individuals into a set of teams that have small faultline scores. We also present two baseline algorithms. Here is an overview of the methods presented here:
-
Faultline-Splitter is our proposed algorithm for solving this problem which outperforms th eother two baselines. Faultline-Splitter works similar to the k-means algorithm. Starting from a random partition of individuals into teams, the algorithms starts by measuring to what extent moving a person to another group lowers the faultline score. Then, the algorithm re-assigns the indivduals to minimize the fautlines. This process is repeated untill convergence.
-
Greedy algorithm works by creating a team at each step. It deos so by selecting the next best available individual that can join a team while keeping its faultline score as low as possible. While the first few teams created by this algorithm have a very small faultline score, the final teams usually end up with a high-score.
-
Clustering simply refers our usual technique for clustering data-points. That is, the algorithm prefers to put similar individuals in the same groups. Note that while creating a team of similar individuals results in a low faultline score, this method fails to recognize that diverse teams can also have a good faultline score, and thus has far more limited options to lower the faultline score.
An implementation of all these techniques can be found here.