The label aggregation for crowdsourced classification datasets consists in presenting a set of $n_{task}$ training tasks to classify to a crowd. The label given for task $i$ by worker $j$ is denoted $y_i^{(j)}$. Given an aggregation strategy $\texttt{agg}$ (like majority voting or Dawid and Skene's model), we look at the recovery of the underlying ground truth labels $y_i^*$:

$$ \mathrm{AccTrain} = \frac{1}{n_{task}} \sum_{i=1}^{n_{task}} 1\!\!1(\hat{y}_i^{\texttt{agg}}=y_i^*).$$

Other objectives as the F1 score can also be considered.

This benchmark can be run using the following commands:

$ pip install -U benchopt
$ git clone https://github.com/benchopt/benchmark_crowdsourcing
$ benchopt run benchmark_crowdsourcing

Apart from the problem, options can be passed to benchopt run, to restrict the benchmarks to some solvers or datasets, e.g.:

$ benchopt run benchmark_crowdsourcing -s solver1 -d dataset2 --max-runs 10 --n-repetitions 10

Use benchopt run -h for more details about these options, or visit https://benchopt.github.io/api.html.