Canadian Government Executive - Volume 25 - Issue 02

: : bias : : : : model bias : : : : data bias : : HISTORIAL BIAS Data is an accurate reflection of reality but that reality reflects lon�-standin� biases in society, which then �et entrenched and amplified by machine learnin�. AGGREGATION BIAS The labels used to describe the data may not be evenly applicable. A one-size-fits-all model will be biased and overall performance of the model will suffer. EVALUATION BIAS If benchmarks used to evaluate the model are not repre- sentative of the broader reality, any tunin� of the model will only be optimized for a limited set of circumstances. HUMAN-IN-THE-LOOP BIAS Humans may evaluate results for error or label ambi- �uous data. Slapdash checks may add bias. Some even use trickery (e.�., are dis�uised as bot tests). INDUCTIVE BIAS Trainin� data does not fully reflect the lar�er reality. That bias is inherent to �eneralization but is worse if the model tracks too closely or loosely to sparse trainin� data. ADVERSARIAL BIAS Hostile actors can deliberately contaminate the data- base with biased data. Uncurated data fed directly into the a�ent is a bi� (and common) source of vulnerability. SAMPLING BIAS The database does not accurately capture the lar�er reality because the collection method did not offer an equal opportunity for inclusion. MEASUREMENT BIAS In practice, “Bi� Data” is full of quality problems. Some machine learnin� applications even rely on data sources known to contain dubious data and data of uneven quality. Bias happens when the quality or �ranularity of the data varies from one �roup of people to another. Bias can also be baked into the data when the classification schemes used are oversimplifications that favour or disfavour a particular �roup. BIASED PROBLEM-FRAMING What is considered a problem and how it is under- stood may be biased. Task performance may involve trade-offs that are resolved in biased ways. Unfair, unanticipated downstream consequences may result. Re�ardless, task performance will not be perfect. It may be a demonstrable improvement over the altern- atives. Thus, some expectations have to be mana�ed. We want people in different �roups treated fairly and equitably, especially as a�ents have a �reater bearin� on our lives. What counts as unfair and inequitable bias is a matter of dispute. Nonetheless, it helps to know where differential treatment can creep in. SELECT SQUARES WITH A CROSSWALK? April/May 2019// Canadian Government Executive / 17