Nikol Pushkash

Hi everyone. I’m a Bachelor student of Computer Science at Faculty of Mathematics and Physics, Charles University. I like to figure out how things work, math, music and volleyball.

Contact me

Office: CoRE building, room 419

Email: pushnikol@gmail.com

At REU 2025, I work on Classical, Computational, and AI-Powered Social Choice. My mentor in Rutgers University is Professor Lirong Xia, and Professor Farhad Mohsin

Project Description

Week log

Week 1: 05/27–06/01
- I came to the US with two other students from Charles Universty, Hana Salavcová and Martin Černý. On my second day here, I attended an orientation meeting held mostly by Lazaros Gallos and other members of DIMACS family. On third one, I also attended a website creating workshop held by Larry Frolov (special thanks to him for organazing meals on first day, making an office tour and rather funny workshop :))
- Regarding my project, I've got one, something to be proud of already :). I read an introductory article on Social choice and got a basic idea of what I might be working with. At least for now, I do like the topic I have. I had a meeting with Professor Xia and after that with Professor Mohsin, who introduced me to the project and first steps that I should do to get started.
- I also created this website, which I kindly borrowed from a former student of DIMACS REU program. Hope he either would not be mad or would not find out :).
Week 2: 06/02-06/08
- Early in the week, I have prepared and gave a presentation about the problem I am going to try to solve this summer. To sum up, I will try to train a model using machine learning such that it, given different types of data, can find Plackett-Luce parameters of distribution of objects. I also listened to projects of other students. I do like the variety of topics DIMACS have provided, even though there is relatively smaller number of participants this year.
- After that, I have completed two courses on Coursera platform, ML specialization and Deep Learning specialization, both by Andrew Ng and DeepLearning.AI. I made myself a bit more comfortable with basic concepts of supervised learning and structure of neural networks.
Week 3: 06/09-06/15
- This week I started reading the Lirong Xia's book on Ranking Data and learning from it. I have read first 4 chapters, which gave me a better understanding of Plackett-Luce ranking model which I will be using in my further experiments.
- At the end of the week, I managed to implement my first model, which learns Plackett-Luce parameters of the ranking distribution. I used one of the easiest and most common loss functions in machine learning, and results quite surprised me. With no complicated techniques, they were pretty nice.
Week 4: 06/16-06/22
- This week I implemented 2 other models, which required a bit more of computations for training, which made those slower to train. But the results were slightly better on small number of alternatives. I think they should be better for higher number of alternatives.
- What I also realized was that my method of evaluating the models was incorrect, hence the results might not be that accurate. I will fix this first thing next week.
Week 5: 06/23-06/29
- I have changed the way of evaluation of the model to RMSE (Root Mean Square Error) between tedicted and ground Plackett-Luce parameters. This one is also used by researchers in Luce Spectral Ranking paper. The good thing is that the results are still good.
- I thought about improving robustness of models and adding extra few distributions to each PL parameter vector in the datasets. So I tried. And it turned out that it was good idea with not that good results. I hoped it would at least have the same accuracy as my previous models, but apparantly the Universe did not cate much about my expectations. The results became even worse somehow, so I abandoned this direction for now.
- Up to now, I have been working only with full rankings disrtibutions of one number of voters. Now I tried to create a joint model which will have different number of voters in distributions in training dataset. But this also did not work that well.
Week 6: 06/30-07/06
- I should have normalized the input. I should have normalized the input. I should have normalized the input.
- As it turns out, my models sizes were linearly increasing as the number of voters increased because I have not been normalizing the input properly. Actually it makes sence because the model should scale down the input and then compute the weights for PL parameters prediction. But now even small neural networks are able to predict PL parameters of data up to n_candidates = 6. 7 and 8 have already too large input :(.
- It also makes sence from the mathematical point of view: when we are gathering full rankings and then normalizing the frequency vector of those, we are basically getting the probabilities of full ranking to appear. And this is actually the reason why everyone cares about Plackett-Luce ranking model at the first place: it makes it easier to work with probabilities of rankings in distributions.
- So I have also discovered that the joint model now works. Since we have just normalized vectors on the input and we don't have to scale down anything. Good news.
Week 7: 07/07-07/13
- So I started evaluating and gathering data for the final report. This is such an ungrateful job. Just sitting there and figuting out how to organize the code and pipelines such that it does everything for you. Programming is hard. This takes much longer than I have expected.
- I have finally moved from full rankings to partial rankings. Now I am able to increase the size of the candidates set up to 20-30 at least. And this is actually the only thing I have time for in this program. But at least I wrote a code and tried it.
- Turns out partial rankings also work somehow nicely. But LSR method does not work on these larger sets for small number of voters since the comparison graph of candidates is unlikely to be stronly connected. So I found one mode package which computes Plackett-Luce parameters given distributions, called FPR (Fast Partial Rankings). First thing next week I will runt he experiments and gather the performance of it as well.
Week 8: 07/14-07/20
- So as I mentioned at the end of last week's section, I ran the package. Turns out my models are both faster and extimate PL parameters better. I am pretty happy about it.
- I started wrining a paper. Now I even more appreciate the work of everyone who writes those. It's so hard to gather all your thoughts and describe your work. New day new challenge.

Acknowledgments

I would like to thank

My supervisors for agreeing working with me
Charles University, expecially Department of Applied Mathematics and Informatics Institute of the University, and also PSJ Foundation for funding my work here
Rutgers University for hosting DIMACS REU 2025 program and giving such a nice opportunity for academic, cultural and social personal and professional development

References