Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in [1].
Producers Equilibria and Dynamics in Engagement-Driven Recommender Systems
Authors: Krishna Acharya, Varun Vangala, Jingyan Wang, Juba Ziani
TMLR 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | In Section 5, we present experiments on synthetic data and three real-world datasets Movielens-100k, Amazon Music, and Rent the Runway (Harper and Konstan, 2015; Ni et al., 2019; Misra et al., 2018). First, we introduce a computationally efficient heuristic based on best-response dynamics (Algorithm 1) for computing pure Nash equilibria of our engagement game. We observe that this heuristic almost always converges to a pure Nash equilibrium within relatively few steps. Second, our experiments further characterize producer specialization, showing that it occurs even under more complex settings than those in Section 4. We then study the effect of the temperature parameter in the softmax content-serving rule on both producer specialization and utility. |
| Researcher Affiliation | Academia | Krishna Acharya EMAIL Georgia Institute of Technology Varun Vangala EMAIL Georgia Institute of Technology Jingyan Wang EMAIL Toyota Technological Institute at Chicago Juba Ziani EMAIL Georgia Institute of Technology |
| Pseudocode | Yes | Algorithm 1 Best Response Dynamics for Pure Equilibrium Computation |
| Open Source Code | Yes | Code available at https://github.com/krishnacharya/recsys_eq |
| Open Datasets | Yes | In Section 5, we present experiments on synthetic data and three real-world datasets Movielens-100k, Amazon Music, and Rent the Runway (Harper and Konstan, 2015; Ni et al., 2019; Misra et al., 2018). |
| Dataset Splits | No | The paper mentions generating synthetic data with "K = 10,000 users" and describes real-world datasets (e.g., Movielens-100k with "100k ratings, 943 users and 1682 movies"). However, it does not specify any training, validation, or test splits for these datasets, nor does it provide percentages or counts for such splits. |
| Hardware Specification | No | The paper mentions "CPU hours" for experiments and that "to parallelize workloads for embedding generation and for the following experiments we use Slurm job arrays" (Appendix H). However, it does not specify any particular CPU or GPU models, memory configurations, or other detailed hardware specifications. |
| Software Dependencies | No | The paper mentions using "the NMF implementation in the scikit-surprise package (Hug, 2020)". While it names a software package, it does not provide a specific version number for scikit-surprise itself, which is required for reproducibility. |
| Experiment Setup | Yes | Algorithm 1 Best Response Dynamics for Pure Equilibrium Computation Inputs: User embeddings (c1, . . . , c K). Utility ui(si, s i) for producer i. Max iterations Nmax. We consider 5 different values for the softmax temperature τ {0.01, 0.1, 1, 10, 100} and use the linear-proportional serving rule as a benchmark. |