Skip to contents

A list containing triplet judgments for 6 participants on 32 icons showing faces and buildings. These triplets were used to compute embeddings of the 32 items separately for each participant (icon_emb_ind) as well as a single group embedding (icon_emb_group).

Usage

icon_triplets

Format

icon_triplets

A named list, each containing a dataframe with 11 columns:

head, winner, loser

Integer indices for items in a given triplet.

worker_id

Random identifier for each participant.

rt

Response time on triplet (in miliseconds).

Center

The target item.

Left, Right

The option items appearing on the left and right.

Answer

The option item chosen by the participant.

sampleAlg

The algorithm used to sample the item.

sampleSet

Which set the sampled item belongs to.

Source

Colon et al., in prep

Details

Each element of the list contains the triplet data for one participant in the study. Rows of a dataframe correspond to a single trial. The elements of the full list are named by the random participant ID number.

Each triplet can be sampled in one of four ways indicated by sampleAlg:

  1. random: Sampled randomly with uniform probability from all triplets.

  2. validation: Sampled randomly from a fixed, pre-specified set of possible triplets.

  3. check: Sampled from a small set of items where the answer is obvious, used to check attention and data quality.

  4. adaptive: Sampled according to some adaptive sampling algorithm.

The column sampleSet indicates how the triplet is to be used in computing and evaluating embeddings:

  1. train: Triplets used to fit embedding.

  2. test: Triplets held out from training, used to evaluate embedding quality.

  3. NA: Triplets used for checking attention / data quality.

Validation trials can be used to measure the extent of inter-subject agreement: since these are drawn from a common pool, these triplets will appear repeatedly across participants. Thus for each such triplet one can compute the majority vote across participants who received the triplet, and the proportion of those participants who agree with the majority vote. This gives an estimate of how consistent judgments are across participants.