As more systems and sectors are driven by predictive analytics, there is increasing awareness of the possibility and pitfalls of algorithmic discrimination. In what ways do you think Recommender Systems reinforce human bias? Reflecting on the techniques we have covered, do you think recommender systems reinforce or help to prevent unethical targeting or customer segmentation? Please provide one or more examples to support your arguments.

The presentation by Evan Estola puts the possibilities of algorithmic discrimination in focus. In one instance, Estola explained how an algorithm he wrote created negativity within the media. An article, “On Orbitz, Mac Users Steered to Pricier Hotels”, caused many to believe the priciest hotels would be become their primary options. He noted . That’s what the existing data shows that females are much less likely to go to tech meetups and if their algorithms were to run on on that data directly, the results would show the same trend. In order to negate this effect they run their algorithms by separating the gender variable completely from the interests variable and combine them in a separate model, to ensure that no gender assumptions are carried over, perpetuating the idea that women dont go to tech meetups.

Recommender systems may not be able to prevent unethical targeting on purpose and a lack of representation, sparse data, and antiquated techniques can continue to widen gaps. In Estola’s presentation he mentioned ethical issues with the admittance to college, college entrance exams have also been seen as culturally biased exams which computer systems tend to rely on. The Google recommends ads regarding rehabiliation or laywers (ads that are suggestive of criminal activity) if they have a “Black Sounding Name”. In order to circumvent biases within recommender systems, We need to account and reverse for these biases in the systems which produces this data in the first place. Similar to the techniques mentioned by Evan where would would do better at removing these biases by seggregating the daata in different segments which would minimize the bias in the data.