Recommender Systems Should Not Be Black Boxes

In an interview with Patrik Dokoupil, recipient of the James Chen Best Student Paper Award presented at the ACM UMAP conference, we take a look behind the scenes of modern recommender systems research. Read about how his path led from fixing computers as a child and game development to top-tier science, what challenges the integration of artificial intelligence into peer-review processes brings, and why it is crucial in academia to set firm boundaries before approaching conference deadlines completely consume you.

What first sparked your interest in computer science/informatics, and what current technological trend or societal challenge in the field excites you the most right now?

I have been interested in computers since my late childhood. This is probably a cliché, but my early motivation came partly from computer games and from trying to fix various issues on my own computer.

I became more interested in programming probably around the age of 15, and a bit later, I started working more seriously with C++, including some basic game development using SDL.

As I progressed to more advanced topics like OpenGL, shaders, and related mathematical concepts, I realized I was missing deeper theoretical foundations. That was the point when I became much more motivated to study both mathematics and computer science more seriously.

Speaking about current trends, over the last few years, we have seen a growing impact of AI-based systems on our daily lives —from large language model-based chatbots improving productivity to recommendation algorithms shaping how we consume information on a daily basis. Observing this evolution is something I find very exciting, especially in terms of how these systems will continue to scale and what new technical challenges this will bring.

Could you briefly introduce your work?

I work on recommender systems – the algorithms behind the scenes on platforms like streaming apps, news sites, and e-commerce stores. Because these systems have a massive impact on dictating what information and content people actually see, getting them right is incredibly important. My specific focus is on multi-objective systems. Instead of just optimizing for a single metric like accuracy, I look at how we can balance competing goals, such as diversity, novelty, fairness, and exploration.

A significant part of my research is about user control. I don't want these algorithms to be black boxes. I explore how we can give users the ability to actively tune their own feeds – like choosing whether they want more diverse suggestions or highly relevant ones. Ultimately, it's about making sure these systems are transparent and actually align with what users expect.

What sparked your interest in doing research in this field?

My interest was sparked when I was working on my Master’s thesis. At that time, I was able to work for a longer period on something that was both novel and technically challenging, and I found that I enjoy working on problems without a clear solution, as well as spending time on experimentation.

At the end of my master’s studies, I had the opportunity to publish the results of my thesis, together with an additional extension, and to attend my first conference, which I would say is the main conference in the field – ACM RecSys. Seeing other presentations back then and being exposed to the community early on played a significant role for me.

Lastly, I was motivated by the idea of contributing meaningfully to the community, as it seemed interesting to give something back after spending years studying the work of others.

What specific research or project are you currently working on? What is the biggest technical or methodological challenge you're tackling?

I am currently working on algorithms and evaluation methods for controllable multi-objective recommender systems.

One of the key challenges we are tackling is scalability. Many existing approaches rely on post-processing or re-ranking techniques that become computationally expensive at larger scales. We are therefore exploring approaches such as sparse autoencoders and their applicability to calibration and multi-objective recommendation, with the goal of making these methods more efficient and more suitable for real-world deployment.

Illustration from a current ACM UMAP 2026 conference paper

What key methods, tools, and data sources do you use most often in your research?

My work combines several complementary approaches: machine learning methods and re-ranking algorithms for recommendation, user studies to evaluate perception, satisfaction, and interaction and also statistical analysis to rigorously compare different approaches. In terms of tools, I typically use Python-based stacks (NumPy, Pandas, Matplotlib/Seaborn) for data processing and evaluation, and frameworks such as TensorFlow or PyTorch for model development.

For user studies, we developed our own framework, EasyStudy, built on Flask, SQLite, and Redis, which allows us to deploy and evaluate recommender systems in realistic experimental settings.

Our data comes both from public datasets (e.g., movies or product domains) and from user interaction data collected through controlled experiments (e.g., via Prolific).

How is AI affecting your day-to-day work? Where do you see the biggest advantages and, conversely, the biggest risks associated with AI in the research environment?

I’ll start with the risks, mainly connected to low-quality peer reviews. Some of the reviews I have seen lately seemed surprisingly shallow or too generic. This could negatively affect the entire peer review process and potentially also risk the leaking of information. Recently, some conferences and venues have started to either prohibit such use or regulate it in different ways.

On a similar note, but from a different perspective, equipping authors with AI tools may increase the number of submissions to each venue, which could lead to a shortage of human reviewers able to review all submissions without compromising quality.

As for advantages, grammar fixing and text rephrasing are definitely very useful from a non-native speaker’s perspective. Nowadays, I also often find myself using these tools as a replacement for a search engine, at least in some cases. Besides that, I have found them to be helpful for programming as well. Overall, AI definitely affects my work by increasing my productivity.

Do you collaborate with colleagues from other universities, research institutions, or industry partners on your research? If so, what does that collaboration look like?

Yes, the most notable collaboration has been with Dr. Ludovico Boratto from the University of Cagliari, Italy, with whom we have collaborated since 2022/2023.

The collaboration started as an idea to co-author a paper between my PhD advisor Dr. Ladislav Peška, and Dr. Boratto, as we were working in the same domain of multi-objective recommendation. It has significantly expanded since then, and as a result, we have already published seven papers together, with others currently in progress.

Due to being in different locations, we usually collaborate virtually, with regular meetings to discuss new ideas, future projects, and progress on ongoing work. We also meet in person at conferences in our field.

In addition, I have collaborated with colleagues from Maastricht University and the University of Sarajevo.

What would be the ideal scenario for applying your research results in the real world, and what steps would be needed to make that happen? How would you ultimately measure the success of such an implementation?

As a broader goal, I advocate for recommender systems that are not only optimized, but also interpretable, controllable, and aligned with user expectations.

In practice, this could mean systems where users can directly influence recommendation behavior, or systems that optimize not only for short-term engagement (e.g., clicks), but also for longer-term satisfaction and trust.

I see my research contributions as incremental steps toward this goal. For example, the RLProp algorithm we introduced enables direct control over the proportion of different objectives in recommendation outputs. A key next step is improving scalability and robustness so that such methods can be deployed in large-scale industrial systems.

Ultimately, success should be measured not only by traditional metrics, but also by long-term user satisfaction, trust, and alignment between system behavior and user expectations.

What technical skills (hard skills) and what interpersonal skills (soft skills) will be, in your opinion, most crucial for young computer science researchers five years from now?

I don’t feel I can answer this broadly for computer science in general, so I will focus on my field. In terms of hard skills, this is always very difficult to predict, as we often see shifts in both technologies (e.g., PyTorch vs. TensorFlow) and approaches. We have seen strong emphasis on deep learning-based recommender models, then a shift toward shallow or linear autoencoders, then LLMs, and now increasing interest in sparse autoencoders.

While these trends change frequently, I believe that data preprocessing, evaluation techniques (including statistical analysis), and reproducibility practices will remain crucial.

For soft skills, I would highlight critical thinking (being willing to question whether current approaches still make sense and to challenge the status quo if they do not) and presentation skills. Doing good research is one thing, but being able to present it in a clear and engaging way is equally important, if not more.

Overall, the field is very dynamic, so it is difficult to predict precisely what will matter most. I might revisit my answer in five years to see how wrong I was.

Do you have any strategies for maintaining a good work-life balance? And what's your perspective on workplace flexibility and mental well-being in the Czech academic sector?

In terms of workplace flexibility and mental well-being, my experience has been quite positive. The people I have worked with have generally been very passionate about their research.

I probably don’t have any particularly strong strategies for maintaining work-life balance, as my main “bad habit” is the tendency to meet deadlines, especially for conferences and journals. On the positive side, I feel that we operate with a good level of autonomy and flexibility.

The challenge is that with autonomy, it is up to us how we use it. If someone is highly motivated or ambitious, it is easy to put unnecessary pressure on themselves. I think it is important to recognize this early and set boundaries, especially during intense periods such as conference deadlines. In the long term, maintaining sustainability is just as important as short-term productivity.

What advice would you give to someone interested in pursuing a research career in computer science? And what are some common myths about a scientific career you'd like to debunk?

I would paraphrase a recommendation I got from someone early on: do not over-specialize yourself too early. Looking at this retrospectively, I think it is extremely hard to predict what is going to be important in the next X years, so having a reasonably broad topic could help you to steer a direction a bit, as needed. Also, don’t wait for predefined topics: bringing your own ideas is often the best way to stay motivated.

A common myth, or at least a myth I had before starting my PhD was that research is about working alone on problems and only publishing groundbreaking results. In reality, it is highly collaborative, and you don’t need to produce groundbreaking results to contribute meaningfully. A lot of research is incremental, but still valuable. Over time, these incremental contributions can have a significant impact.