Performative Recommendation: Diversifying Content via Strategic Incentives
Authors: Itay Eilat, Nir Rosenfeld
ICML 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We provide analytic and empirical results that demonstrate when and how diversity can be incentivized, and experimentally demonstrate the utility of our approach on synthetic and semi-synthetic data. Using our proposed learning framework, we empirically demonstrate how properly accounting for strategic incentives can improve diversity and how neglecting to do so can lead to homogenization. We begin with a series of synthetic experiments, each designed to study a different aspect of our setup, such as the role of time, the natural variation in user preferences, and the cost of applying strategic updates. We then evaluate our approach in a semi-synthetic environment using real data (Yelp restaurants) and simulated responses. Our results demonstrate the ability of strategically-aware retraining to bolster diversity, and illustrate the importance of incentivizing the creation of diversity. |
| Researcher Affiliation | Academia | 1Faculty of Computer Science, Technion Israel Institute of Technology, Haifa, Israel. Correspondence to: Nir Rosenfeld <nirr@cs.technion.ac.il>. |
| Pseudocode | No | The paper describes algorithms through mathematical formulas and text (e.g., Eq. (11)), but does not include any explicitly labeled 'Pseudocode' or 'Algorithm' blocks, nor structured code-like procedures. |
| Open Source Code | Yes | All code is made publicly available at: https://github.com/itayeilat/Performative-Recommendation. |
| Open Datasets | Yes | Data. Our experimental setup is based on the restaurants portion of the Yelp dataset5, which includes user-submitted restaurant reviews. We focus on users having at least 100 reviewed restaurants. Our experiments use the Yelp dataset, which is publicly available at https://www.yelp.com/dataset/ download. |
| Dataset Splits | Yes | For training, we assume that at each round the system has access to 30 items per user, randomly selected (out of the 40) per round; of these, a random 20 are used for training, and the remaining 10 are added for validation (tuning λ and early stopping). |
| Hardware Specification | Yes | All experiments were run on a cluster of AMD EPYC 7713 machines (1.6 Ghz, 256M, 128 cores). |
| Software Dependencies | No | The paper mentions 'Adam' as an optimizer and 'pytorch' for implementation (e.g., 'pytorch implementation'), but it does not specify version numbers for these or any other software dependencies. |
| Experiment Setup | Yes | For optimizing predictive models f in each experimental condition, we use Adam and train for a maximum of 200 epochs with learning rate 0.1. For smoothing (see Sec. 3.1), we use temperatures τ = 0.1 for NDCG, τ = 1 for the permutation matrix approximation, and τ = 5 for the soft-k function; all were chosen to be the largest feasible values that permit smooth training. |