Skip to content

Median aggregation of EM estimates can produce m probabilities that don’t sum to one #3199

Description

@lamaeldo

I noticed that _populate_m_u_from_trained_values() takes the median independently for each comparison level.

This can produce a final vector that no longer sums to one, even though every individual EM estimate does. For example, three valid estimates:

[0.90, 0.05, 0.05]
[0.05, 0.90, 0.05]
[0.05, 0.05, 0.90]

give level-wise medians of [0.05, 0.05, 0.05], which sum to 0.15.

Obviously this is an extreme example, but I encountered this in a one of my pipelines, with final m vectors summing to 0.735 and 0.787.

I am not sure if this is intentional, or acceptable for other reasons, but I wondered whether the final vector should be normalised, or at least produce a warning when it is materially different from one, hinting to a possibly degenerate training

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions