Recommendation Systems have become an integral part of many businesses today. They have variety of applications. For example, on an e-commerce website like Amazon or Walmart, it would recommend you more relevant products to buy or in completely different setting they which would suggest you to add new people to your social network like that on Facebook or Twitter. Recommendation Systems are fueled and motivated by desire of Business growth and expansion. And just like other things backed by a desire to personal growth and earning profits, there is always a chance that it can cross ethical boundaries and act irresponsibly.
Meetup.com organized a talk given by Evan Estola, one of the company’s senior Machine Learning Engineer, to discuss the scenarios when “Recommendation Systems go Bad.” He talked about various aspects of recommendation systems, including the technical details as well as what would be an irresponsible behavior for a Recommendation system.
Evan described how Meetup’s recommendation system is bringing people closer to connect offline. Its goal is to find correct groups for the users and for suggest most interested members to the group owners.
Meetup’s recommendation system is currently using the K-Nearest Neighbors algorithm, where they use a similarity measure like cosine similarity to find the users which are closest to each other, in some feature space and generate recommendations based on this closeness. This is in some ways like Amazon’s collaborative filtering algorithm. Evan emphasized that recommendations these days is focused around generating a Ranking of the items to be recommended using Machine Learning models. This ranking is done in either of three approaches:
- Pointwise: algorithm focusing on items individually. A comparison between each item’s result gives a ranking among them.
- Pairwise: Items are compared in pairs and ranked.
- Listwise: optimizes on entire list. It can also usable to recommend diverse results.
Evan then shifted his focus from technical discussion to a more ethical one. He started by reiterating how Data Science has become involved in our lives. Machine Learning and Data Science is being used in generating news feeds, recommending job openings, generating loan credibility reports, search results on google, Ads on various websites etc. If we think deeply, some of these applications are very crucial and can change our perspective or our lifestyle. For example, a biased news feed, selectively focusing only on negative parts of a political candidate, can shape a neutral viewers opinion. Apart from this hypothetical scenarios, several real incidents or potential cases of biases and discrimination of data generated results have come up every now and then. These include:
- A Study showed lower paying job being recommended to women candidates.
- A news article mentioned, the possibility of expensive flight recommendations were made to MacBook owners.
- Google’s search results are affected by your previous search history and your user profile.
Such biases can be built into the systems to serve a company’s desire to earn more or a developer’s ego about making his system more and more distinguished. But, it’s necessary to realize, while working on these data based systems, that we have an obligation towards the customer and society. It is our duty to make recommendation systems free from biases, like those of gender and race.
Evan recommended some measures that can be taken in ensuring that recommendation systems and Data Science in general do not go to the “Dark Side”. Awareness of the issue tops the list. Using feature segregation is another thing that Meetup intentionally uses in a recommendation system. They keep gender and interest as part of different models and use Ensemble approach to generate combined results from these models. This is done to prevent the features of different natures from interacting and producing biased results. Lastly, a thorough testing is recommended on a diverse test set to catch any possibility of biased recommendations.
It was a great talk, covering several important aspects of recommendation systems. Evan clearly got the balls rolling to spread an awareness of ethics around the building of Data Science based systems.
An article by Rishabh Jain