Overcoming Algorithm Aversion: A Comparison between Process and Outcome Control

Title of the Paper

Overcoming Algorithm Aversion: A Comparison of Process Control and Outcome Control

Paper Information

• Field of Study: Research on Human-Computer Interaction and Algorithm Usage Behavior
• Keywords: Algorithm aversion, process control, outcome control, human-computer interaction, decision support systems, machine learning model customization, transparency, fairness
• Publication Year: 2023
• Conference Name: CHI Conference on Human Factors in Computing Systems (CHI ’23)

Research Background and Problem

• Issues and Challenges:

  • Algorithm aversion refers to the phenomenon where people are reluctant to use algorithms, even when they outperform human predictions.
  • Existing studies suggest that granting users control over model prediction outcomes (referred to as outcome control) can mitigate algorithm aversion.
  • Compared to outcome control, the impact of process control (control over model training input variables and algorithm types) on algorithm aversion has not been fully explored.

• Significance:

  • Algorithm aversion may reduce the efficiency of data-driven algorithms, undermining the scientific basis of decision-making.
  • Investigating the effects of outcome control and process control is crucial for optimizing algorithm design and improving user acceptance.

• Research Motivation:

  • The authors question whether process control can mitigate algorithm aversion as effectively as outcome control.
  • Building on existing research, they further investigate whether the interaction effects of process control and outcome control can provide additional mitigation of algorithm aversion.

Solution

• Proposed Approach:

  • Design and conduct randomized experiments to study the independent and combined effects of process control and outcome control.
  • Outcome control is implemented by allowing users to adjust model prediction outcomes.
  • Process control is divided into two forms:
    1. Users select input variables for model training.
    2. Users select the algorithm for model training (e.g., linear regression, decision trees).

• Innovations:

  • Operationalizing process control into two distinct measures—user selection of input variables and algorithms—and quantifying their impact on algorithm aversion.
  • Exploring whether the combination of process control and outcome control produces effects greater than either approach alone.

• Implementation Steps and Key Techniques:

  1. Recruit participants using two online research platforms (MTurk and Prolific).
  2. Set experimental conditions, including whether users are allowed to adjust model predictions or participate in model design.
  3. Provide real datasets and model testing tasks, measuring participants' behaviors and opinions.
  4. Collect quantitative data on participants' model usage choices, prediction error rates, and differences from the model, as well as subjective evaluations of model transparency and fairness.

Research Findings

• Specific Results:

  • Outcome control (adjusting model prediction outcomes) significantly reduced prediction error rates and increased model usage.
  • The effects of process control varied by form:
    • Selecting training algorithms significantly mitigated algorithm aversion, improving usage rates and prediction performance.
    • Selecting input variables did not show significant effects.
  • Combining outcome control and process control (allowing both prediction adjustment and algorithm design) did not further reduce algorithm aversion, and the effects were not additive.

• Comparison with Existing Solutions:

  • Process control (particularly selecting training algorithms) demonstrated similar effectiveness to outcome control in reducing algorithm aversion.
  • Compared to prior research focusing on outcome control, process control showed advantages under specific conditions.

• Experimental Results and Evaluation:

  • Platform differences: Significant variations were observed between platforms (MTurk and Prolific) and across experimental batches, highlighting issues of reproducibility and external validity.
  • User perceptions of models: Subjective ratings of model transparency, fairness, and trust did not significantly change based on the type of control.

• Limitations and Future Directions:

  • Limitations:
    • Study participants were online platform workers who lack domain-specific expertise as actual algorithm users.
    • The study lacked higher-level user interaction designs relevant to real-world work scenarios.
  • Future Directions:
    • Investigate the impact of allowing users to select target variables or outcome metrics on algorithm aversion.
    • Explore the psychological mechanisms behind the effects of process control, such as psychological ownership or effort justification.
    • Enhance reproducibility testing and extend research to real-world work environments.

Conclusion

This study reveals that process control (particularly algorithm selection) is an effective measure for reducing algorithm aversion. However, the combined effects of process control and outcome control were not significantly greater than using either approach alone. The findings provide valuable insights into user participation in algorithm design while also highlighting the limitations and considerations of conducting experiments with general populations. Future research should address reproducibility issues and optimize participation designs to better reflect real-world scenarios.