Mitigating the Harm of Recommender Systems

Answer

In their respective articles YouTube, the Great Radicalizer by Zeynep Tufekci and Up Next: A Better Recommendation System by Renee DiResta, both seem to vent a feeling of entrapment in themes proposed by recommender systems (RS) of YouTube or Pinterest. The RS senses the users’ patterns of searches and recommends themes of “similar” nature. What I read with a smidgen of humor is, on changing the topic, RS persists the user with a new theme — pursues the user, as it were. Tufekci calls this radicalizing effect. Perhaps her rationale is, by pigeonholing a user to his/her set of choices, it reinforces the bias. Based on Google’s conscientious engineer, Mr. Chaslot’s testimony, we further learn about Google’s penchant for Trump, in the 2016 elections. Tufekci’s next theme is the expression “fake news” that gained traction in 2016, but is still current. The fake news reaches a climax when we learn about YouTube videos that depict Newtown shootings as a hoax. So, DiResta contends that RS is the biggest threat to social cohesion. We further learn about some constructive approaches like Google Jigsaw’s Project Redirect, about Choice Architecture and so forth.

In week 4’s discussion, the question at hand is, how to counter the radicalizing effects of RS, or ways to prevent algorithmic discrimination.

Even before embarking upon my personal views on the basic premise of the article, let me try to understand the problem from a technical standpoint. The term radicalizing effect was coined by Tufekci. She seems to complain that RS senses her searches and narrowly constrains her to a theme. Perhaps she is asking for serendipity, novelty and diversity. If that’s her complain, then I think the solution is to build in some entropy in the RS. Based on her user experience, what Tufekci calls radicalizing effect, is to my eyes — a budding Data Scientist — an effect of lack of entropy. So, if I were to design the RS, I would add below snippet of code in my R program:

                     Random_model <- Recommender (train, method = “RANDOM”)  

I am a humble beginner in RS, so this is what I can propose, with my little experience.

So far about a potential solution to the technical problem. But I see a conflict of interests here — a conflict of interests, between a set of users, represented by the authors, and another set of users, who are alleged consumers of “Extreme” material. Couched in the language of ethics and didacticism, the authors seem to suggest some sort of censorship of content — filtering out content, based on falsity or other criteria. That’s their position.

Whether one likes it or not, there are consumers of many stripes — consumers of truths, academic literature, fictions, half-truths, outright lies, humor, humor at the expense of others, inflammatory material, aspersions, religious material, dogmatic material, skepticism, science and many more. It takes many types to make the world. Personally, I may not enjoy all forms of expression, but I believe that between tolerance and subversion of unpleasant speech, the hard-earned value of tolerance enables truth to get revealed, and mirror what actually exists. In other words, I uphold the First Ammendment.

DiResta mentioned about Google Jigsaw’s Project Redirect, which redirects certain types of users, searching certain type of content, to Wikis to fact-checking websites. YouTube is Google’s product (through purchase). If Google wants to disappoint one kind of consumers and sustain consequent losses, for their perceived social cohesion, that’s their business priority. I don’t have anything to add here. But, as a user, as a curious person, as person who enjoys fun and variety and laughter, I would not only like to have access to all types of content, but also the freedom to create them.

Marker: 612-04_d