Trustworthy-ML

This is a repository for the data and analysis code for the paper "Human Reliance on Machine Learning Models When Performance Feedback is Limited: Heuristics and Risks".

Repository Structure

• data contains 1: the cleaned data including the predictions of tasks and answers to the questionnaire given by subjects 2:datingData the raw data of dating profiles and the instances selected for the pilot study.

• analysis contains the analysis code implemented by python and R via jupyter notebook and R markdown.

Each file in data and analysis starting with "experimentOne","experimentTwo", "experimentThree", are each corresponding to the Experiment 1, 2, 3 in the paper.

Data Sample

Here attached a data sample of predictions made by subjects in Experiment 1. Each row is a piece of record on a subject making predictions on one task.

1 indicates "The dating participant wants to see the date again." 0 indicates "The dating participant wants to see the date again."

workerID	idpAgreement	taskId	profile	globalId	decision	selfPrediction	finalPrediction	selfCorrect	finalCorrect	agreement	prediction	finalAgreement	mlCorrect	switch
94	100	4	26	734	0	1	1	FALSE	FALSE	1	1	TRUE	FALSE	FALSE

• workerId Int. The unique ID for each subject assigned in each experiment.
• idpAgreement Int. The three different agreement levels of each experimental group, indicating the ML model in this group was designed to agree with the majority of people on idpAgreement percent (e.g., 40%).
• taskId Int. The order number of the task among the experiment.
• profile Int. The unique profile ID of the dating profile observed by the subject in each task, corresponding to each dating profile.
• globalId Int. Another unique profile ID of the dating profile.
• decision Int. The ground truth of the dating profile used in the task.
• selfPrediction Int. The prediction given by the subject before she observe the ML model's prediction.
• finalPrediction Int. The prediction given by the subject after she observe the ML model's prediction.
• finalCorrect Bool. Whether the final prediction given by the subject is correct or not.
• agreement Bool. Whether the initial prediction given by the subject is the same as the ML model's prediction.
• prediction Int. The prediction given by the ML model in this task.
• finalAgreement Bool. Whether the final prediction given by the subject is the same as the ML model's prediction.
• mlCorrect Bool. Whether the ML model's prediction is correct on this task.
• switch Bool. Whether the subject switched her prediction after observing the ML model's prediction.

Subjects' predictions in Experiment 2 and 3 are similar to those in Experiment 1. Due to the different experiment designs, there can be additional columns in the data as follow:

In Experiment 2: * acc Int. The designed accuracy of the ML model's predictions in Phase 1 of the experiment.

In Experiment 3: * treatment Int. The designed experimental group of the subject. 0 indicates high confidence agreement - high confidence disagreement group, 1 indicates high confidence agreement - low confidence disagreement group, 2 indicates low confidence agreement - high confidence disagreement group, 0 indicates low confidence agreement - low confidence disagreement group.

Here attached a sample of the data of answers made by subjects to the questionnaire during the Experiment 3. Each row is a piece of record on a subject answering questionnairs.

1 indicates "The dating participant wants to see the date again." 0 indicates "The dating participant wants to see the date again."

workerID	humanPredictAccuracy	humanPredictSelfAccuracy	reliability	faith	competence	understandability	treatment	surveyTrust
94	70	80	5	3	5	7	0	5

• workerId Int. The unique ID for each subject assigned in each experiment.
• humanPredictAccuracy Int. How accurate did the subject think the model was between 0 to 100.
• humanPredictSelfAccuracy Int. How accurate did the subject think she was between 0 to 100.
• reliability Int. The subject's rating on the ML model's reliability between 1 to 7.
• faith Int. The subject's faith in the ML model between 1 to 7.
• competence Int. The subject's rating on the ML model's competence between 1 to 7.
• understandability Int. The subject's rating on the ML model's understandability between 1 to 7.
• surveyTrust Int. The subject's rating on her overall trust in the ML model between 1 to 7.
• treatment Int. Same as previously mentioned in prediction records.

For questionnaires in Experiment 1 and 2, they are also similar to those in experiment 3, except that on a 7-point Likert scale, we considered subjects who provided a rating of 1–3 asdisagreeing with the statement and encode them as 1. For subjects who provided a rating of 5–7 as agreeing with the statement and encoded them to 1. Due to a mistake in Experiments 1 and 2, for subject’s evaluationson the competence, reliability, understandability, and faith statements, ratings of 6 and 7 were all recorded as 5.

Additionally, some perceptions are not measured in Experiment 1 and 2 due to the experiment design and acc and idpAgreement are used to indicate different experimental groups, which is similar as mentioned in prediction records.

Licensing & Citing

When using or building upon the data in an academic publication, please consider citing as follows:

Lu, Z., & Yin, M. (2021, May). Human Reliance on Machine Learning Models When Performance Feedback is Limited: Heuristics and Risks. In Proceedings of the 2021 CHI Conference on Human Factors in Computing Systems

[DOI:10.1145/3411764.3445562]