Algorithms have both advantages and drawbacks
Mathematical equations consistently follow rules set up by people. Numerical methods make decision processes transparent. This, however, is not true of certain machine learning techniques since the rules used by a trained algorithm are not always easily to explain.
Algorithms may compound a bias present in data which the algorithm uses as input. Any patterns created from such a bias are only magnified by the algorithm. We therefore need to give consideration to the data variables we use as input.
Biases can be revealed by testing
Human judgment is needed to assess the accuracy of output when testing the algorithm. Feedback is necessary for improvements to be made. In the case of Amazon mentioned in the previous article, the statistics showed that words in people’s resumes expressing confidence are used predominantly by males.
This discovery helped reveal a bias that had existed previously, though hiring managers had probably been unaware that this particular wording was influencing them. Or if they were consciously aware of it, then perhaps self-confidence is ultimately not such a useful indicator of competence as they thought.
In this case, Amazon might improve its hiring practices as follows:
- It can redact irrelevant words (if these are not informative) before resumes are reviewed.
- Programmers can stop using these words as predictive cues.
The conclusion of the Harvard Business Review is that existing input variables should be examined so as to avoid maintaining bias in algorithmic decision-making.
-jk-