Quantifying Availability and Discovery in Recommender Systems via Stochastic Reachability
Authors: Mihaela Curmei, Sarah Dean, Benjamin Recht
ICML 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We demonstrate evaluations of recommendation algorithms trained on large datasets of explicit and implicit ratings. Our results illustrate how preference models, selection rules, and user interventions impact reachability and how these effects can be distributed unevenly. |
| Researcher Affiliation | Academia | 1Department of Electical Engineering and Computer Sciences, University of California, Berkeley, USA. Correspondence to: Mihaela Curmei <mcurmei@berkeley.edu>, Sarah Dean <sarahdean@eecs.berkeley.edu>. |
| Pseudocode | No | The paper describes mathematical formulations and processes but does not include any clearly labeled pseudocode or algorithm blocks. |
| Open Source Code | Yes | 1Reproduction code available at github.com/ modestyachts/stochastic-rec-reachability |
| Open Datasets | Yes | Datasets We evaluate1 max ρ reachability in settings based on three popular recommendation datasets: Movie Lens 1M (ML-1M) (Harper & Konstan, 2015), Last FM 360K (Celma, 2010) and MIcrosoft News Dataset (MIND) (Wu et al., 2020). |
| Dataset Splits | No | For each dataset and recommender model we perform hyper-parameter tuning using a 10%-90% test-train split. |
| Hardware Specification | No | The paper mentions 'computational burden' but does not specify any details about the hardware (e.g., CPU, GPU models, memory) used for the experiments. |
| Software Dependencies | Yes | Ap S, M. MOSEK Optimizer API for Python Release 9.0.88, 2019. URL https://docs.mosek.com/9. 0/pythonapi.pdf. |
| Experiment Setup | Yes | We consider three types of user action spaces: History Edits, Future Edits, and Next K in which users can strategically modify the ratings associated to K randomly chosen items from their history, K randomly chosen unobserved items, or the top-K items according to the baseline scores of the preference model. For each of the action spaces we consider a range of K values. We further constrain actions to lie in an interval corresponding to the rating range, using [1, 5] for movies and [0, 10] for music and news. For each dataset and recommendation pipeline, we compute max reachability for soft-max selection rules parametrized by a range of β values. |