Causal Contextual Bandits with One-Shot Data Integration
Author Information
Author(s): Subramanian Chandrasekar, Ravindran Balaraman
Primary Institution: Indian Institute of Technology Madras
Hypothesis
How can causal side information be effectively integrated into contextual bandit settings to improve decision-making?
Conclusion
The proposed algorithm outperforms baseline methods in various experiments, demonstrating its effectiveness in utilizing causal information.
Supporting Evidence
- The algorithm shows improved performance in various experimental settings compared to baseline methods.
- It effectively integrates causal side information to enhance decision-making.
- The results indicate that as the budget increases, the algorithm's regret converges to zero.
Takeaway
This study introduces a new way to use causal information in decision-making processes, helping to make better choices faster.
Methodology
The study proposes a new algorithm that minimizes a novel entropy-like measure while allowing for one-shot data acquisition within a budget.
Potential Biases
The algorithm does not guarantee demographic parity, which may lead to biased outcomes.
Limitations
The algorithm's performance may degrade with mis-specified causal graphs.
Digital Object Identifier (DOI)
Want to read the original?
Access the complete publication on the publisher's website