Bias in Data Science – the Good, the Bad and the Avoidable !?

In recent years, there have been a few prominent examples of accidental bias in machine-learning applications, such as smart phones’ beauty filters (that essentially ended up whitening skin) [1] or Microsoft’s from-innocent-teen-to-racist-in-24-hours chatbot [2,3]. Examples such as these obviously fell victim to inherently biased data being fed into algorithms too complex to allow for much transparency. Hidden bias continues to be an issue on ubiquitous social media platforms, such as Instagram, whose curators appear to profess themselves both regretful AND baffled [4]. Unfortunately, any model will somewhat regurgitate what it has been fed and interventions at this level of model complexity may prove tricky.  

Interestingly, bias itself does not need to be harmful and is often built into a model’s design on purpose, either to address only a subset of the overall population or to model a real-world state, for instance when predicting house prices from its size and number of bedrooms, the model’s bias parameter often represents the average house price in the data set. Thus, we need to distinguish between conscious and unconscious bias in data analysis. Additionally, there is the factor of intent, i.e. is the person conducting the analysis well-intentioned to follow good scientific method or trying to manipulate it to achieve a particular outcome.

In the following, I will only discuss aspects of unintentional and unconscious bias, meaning bias hidden from the researcher or data scientist introducing it. This is by no means an exhaustive discussion, but merely a highlight of some pervasive aspects:

A. Data availability bias

B. Coherent theory bias

C. Method availability/popularity bias

A. Data availability bias

A. The problem of scientists’ selecting their data out of convenience rather than suitability or representativeness for the current task has been around for a while [4], e.g. the ideal data set may not be available in machine-readable format or would require higher costs and more time for processing, in short, several obstacles to doing an analysis quickly. For instance, in the area of Natural Language Processing, the major European languages, like English, French and German etc. tend to receive more attention because both data and tools to analyze them are widely available. Similarly, psychology research has mostly focused on so-called WEIRD societies (White, Educated, Industrialized, Rich, Democratic) [5] and out of convenience often targets the even smaller population of “North American college students” that unsurprisingly have been found to not represent human populations at large.

B. Coherent theory bias

B. Various studies suggest that we as people strongly favour truths that fit into our pre-existent world view, and why would scientists be exempt from this? Thus, it appears when people analyze data they are often biased by their underlying beliefs about the outcome and are then less likely to yield to unexpected non-significant results [6]. This does not include scientists disregarding new evidence because of conflicting interests [7]. This phenomenon is commonly referred to as confirmation bias or, more fittingly, “myside” bias.

C. Method availability/popularity bias

C. There is a tendency of hailing new trendy algorithms as one-fits-all solutions for whatever task or application. The solution is presented before examining the problem and its actual requirements. While more complex models are often more powerful, this comes at the cost of interpretability, which in some cases is not advisable. Additionally, some methods, both simple and complex ones, enjoy popularity primarily because they come ready-to-use in one’s favourite programming language.

 

Going forward… 

We as data scientists should:

a. Select our data carefully with our objective in mind. Get to know our data and its limitations.

b. Be honest with ourselves about possible emotional investment in our analyses’ outcomes and resulting conflicts.

c. Examine the problem and its (theoretical) solutions BEFORE making any model design choices.

 

 

References:

[1] https://www.theguardian.com/technology/2017/apr/25/faceapp-apologises-for-racist-filter-which-lightens-users-skintone(last accessed 21.10.2020)

 [2] https://www.bbc.com/news/technology-35902104(last accessed 21.10.2020)

 [3] https://spectrum.ieee.org/tech-talk/artificial-intelligence/machine-learning/in-2016-microsofts-racist-chatbot-revealed-the-dangers-of-online-conversation(last accessed 21.10.2020)

[4] https://www.theguardian.com/technology/2020/aug/09/instagrams-censorship-of-black-models-photo-shoot-reignites-claims-of-race-bias-nyome-nicholas-williams(last accessed 21.10.2020)

[5] Joseph Rudman (2003) Cherry Picking in Nontraditional Authorship Attribution Studies, CHANCE, 16:2,26-32, DOI: 10.1080/09332480.2003.10554845

[6] Henrich, Joseph; Heine, Steven J., and Norenzayan, Ara. The Weirdest People in the World? Behavioral and Brain Sciences, 33(2-3):61–83, 2010. doi: 10.1017/S0140525X0999152X.

[7] Hewitt CE, Mitchell N, Torgerson DJ. Heed the data when results are not significant. BMJ. 2008;336(7634):23-25. doi:10.1136/bmj.39379.359560.AD

[8] Ioannidis JP. Why most published research findings are false. PLoS Med. 2005;2(8):e124. doi:10.1371/journal.pmed.0020124