A panel that does not quite agree | Rank confidence sets | Probability of membership in the top k | Selection with a guarantee | Resampling the data instead of the panel | How much is a 95% set worth? | A note on cost | Related work

The portfolio | The panel | What the panel agrees on | What survives resampling | Where the disagreement lives | What a consensus cannot do for you

Sparsity: keep every value above a threshold | Landmarks: the Nystrom approximation | What the approximation is for | How many trees, and how much does the answer move? | Choosing between them | Streaming: never allocate it at all | The shape this feature was promised in was wrong | What it costs, which is not nothing | Two things it refuses

1. Are there segments, and are they risk segments? | 2. Does the structure survive a shift in the window? | 3. What is the protected attribute worth to the model that never saw it? | What this does not show

Two forests to compare | Correlation between two proximity matrices | Alignment of representations | Out-of-bag matrices have to be repaired first | Superimposing the two configurations | Partitioning one matrix | The order of the terms is part of the question | Undefined pairs are refused rather than dropped | Still to come

Extracting a proximity matrix | Diagnostics | In-bag against out-of-bag | The out-of-bag proximity is not a kernel | Other engines | The views | What comes next

The consensus | Starting from importance scores | Building the panel from fitted models | Ties are information | Not every judge deserves an equal vote | What comes next

The pipeline | Which function do I want? | Getting importance out of a model | Turning scores into rankings | Building a panel | Weighting the judges | The consensus | Asking who disagrees | Uncertainty | Turning uncertainty into a decision | Plots | Under the hood | Where to go next

Two panels | What set stability sees | What rank consensus sees | Where the two agree | The honest summary

The problem | Why rankings and not scores | The Kemeny median | Ties are part of the answer | Measuring agreement | Uncertainty about a rank | What the intervals look like, and why | Testing whether a panel is one population | What none of this protects you from | Sources

A dataset built to cause it | Building the panel | The two methods do disagree | But the rankings agree | Is there one panel, or two? | How often does it divide? | What a rank-based diagnostic can and cannot see