Home  >   Blog  >   Ethics in Recommendation Pipeline—A First Look at RecSys 2022 Papers


Ethics in Recommendation Pipeline—A First Look at RecSys 2022 Papers

  Support by donation   Gift a cup of coffee

RecSys 2022, the 2022 edition of a top academic conference in the field of recommender systems, happened last month in Seattle. Even though I didn't attend the event like the last couple of years, it's still fun to check out the papers and see how the research trends, as well as the industry's focuses, are shifting.


At a glance, while many studies continuously try to fulfill the gap between theory and practice by applying state-of-the-art neural network-based techniques to their own problem spaces (e.g., news, music), more research also started casting "deeper" problem statements that cannot be easily addressed by the conventional optimization problems: privacy, fairness, and offline evaluation vs. online performance. In fact, these topics are not new, and we've seen the presence of relevant work over the last few years as I summarized the 2021 trend as user-centricity. However, I feel the number of papers incorporating these perspectives into a key message has increased at a steady pace.

In my view, such a trend surfaces an underlying challenge about ethics in recommender systems. As Big Tech popularized the power of personalization in their services, we see not only the benefits of these capabilities but a flip side of them such as privacy threats, polarization, and loss of autonomy at large. To address these concerns, developers need to make conscious design choices at every step of recommender system development, ranging across UI/UX design, data collection, model training, evaluation, deployment, and monitoring.

With that in mind, here is my reading list from the proceeding of RecSys 2022:

Unlike the previous years, where the fairness aspects were primarily associated with the evaluation phase in the form of metrics, we can see further consideration about how to incorporate fairness "by design" at the pre-/post-evaluation phases of a recommendation pipeline.

Speaking of recommender pipeline, this year's best demo paper, RecPack: An(other) Experimentation Toolkit for Top-N Recommendation using Implicit Feedback Data, was a good read. The authors nicely extracted common patterns every recommender will equally require, and they strongly highlighted how crucial improving the reusability of recommender implementation is [to fulfill the offline evaluation vs. online performance gap]. A recent podcast episode of Recsperts invited the authors and had insightful conversations about the package and motivation.

My immediate question after reading the paper was Where do the non-accuracy metrics fit into RecPack's evaluation module? and it was answered in the podcast. They mentioned the difficulty in standardizing the interfaces of these metrics when it comes to modularization, and I share the sense based on my experience of creating a recommender package in Julia and implementing diversity & serendipity metrics in it. I think we need a good "design pattern" of recommender systems outside of the machine learning domain; replicating scikit-learn's fit interface isn't enough to make recommender implementation widely (re)usable because recommendation is not always the same as machine learning in many cases.

Since 2021-2022 was a special period for me working closely with real-world recommenders at Amazon's industry-leading personalization team, I realized that how I read the list of accepted papers changed a lot in a meaningful way. One interesting observation is that I've become less curious about the industry talks. Maybe I read too much into it, but, although they are talking about different products with cool technologies and algorithms, I see the objectives are still very much flattened and boring—Maximizing user engagement and downstream profit. Hopefully, the reality of the industry's recommender development is (will be) the same as much as the research trend indicates. Ultimately, I'd like to take the situation more seriously and turn the concerns into actions whenever I interact with recommender systems both as a user and developer.



Recommender Systems

  See also

Cross Validation for Recommender Systems in Julia
User-Centricity Matters: My Reading List from RecSys 2021
Reviewing Ethical Challenges in Recommender Systems


Last updated: 2022-10-14

  Author: Takuya Kitazawa

Takuya Kitazawa is a freelance software developer, previously working at a Big Tech and Silicon Valley-based start-up company where he wore multiple hats as a full-stack software developer, machine learning engineer, data scientist, and product manager. At the intersection of technological and social aspects of data-driven applications, he is passionate about promoting the ethical use of information technologies through his mentoring, business consultation, and public engagement activities. See CV for more information, or contact at [email protected].

  Support by donation   Gift a cup of coffee


  • Opinions are my own and do not represent the views of organizations I am/was belonging to.
  • I am doing my best to ensure the accuracy and fair use of the information. However, there might be some errors, outdated information, or biased subjective statements because the main purpose of this blog is to jot down my personal thoughts as soon as possible before conducting an extensive investigation. Visitors understand the limitations and rely on any information at their own risk.
  • That said, if there is any issue with the content, please contact me so I can take the necessary action.