analyse_experts

Analyses expert input weights to determine which features are most
important per subgroup. For two groups, returns a signed
difference plot (GroupB - GroupA). For &gt;2 groups, performs a
non-inferential analysis: per-group mean importances, pairwise
(B - A) difference tables for all pairs, and a multi-group plot
for the features with the largest max-min spread across groups.

Tools for training and analysing fairness-aware gated neural
networks for subgroup-aware prediction and interpretation in clinical datasets.
Methods draw on prior work in mixture-of-experts neural networks by
Jordan and Jacobs (1994) <doi:10.1007/978-1-4471-2097-1_113>,
fairness-aware learning by Hardt, Price, and Srebro (2016) <doi:10.48550/arXiv.1610.02413>,
and personalised treatment prediction for depression by Iniesta, Stahl, and McGuffin (2016)
<doi:10.1016/j.jpsychires.2016.03.016>.

Rhys Holland

fairGATE

Fair Gated Algorithm for Targeted Equity

analyse_experts function

<dl><dt>gnn_results</dt>
<dd>A list from <code>train_gnn()</code>.</dd>
<dt>prepared_data</dt>
<dd>A list from <code>prepare_data()</code> (used to retrieve group mappings if not provided).</dd>
<dt>group_mappings</dt>
<dd>Optional named mapping from group codes (names) to labels (values).</dd>
<dt>top_n_features</dt>
<dd>Integer; number of top features to visualise.</dd>
<dt>verbose</dt>
<dd>Logical; print progress messages.</dd></dl>

Arguments

Analyse and Visualise Expert Network Specialisation — analyse_experts

<dl>

<dt>gnn_results</dt>
<dd>A list from <code>train_gnn()</code>.</dd>


<dt>prepared_data</dt>
<dd>A list from <code>prepare_data()</code> (used to retrieve group mappings if not provided).</dd>


<dt>group_mappings</dt>
<dd>Optional named mapping from group codes (names) to labels (values).</dd>


<dt>top_n_features</dt>
<dd>Integer; number of top features to visualise.</dd>


<dt>verbose</dt>
<dd>Logical; print progress messages.</dd>

</dl>

analyse_experts: Analyse and Visualise Expert Network Specialisation

Description

Usage

Value

Arguments