User-Creator Feature Polarization in Recommender Systems with Dual Influence

Authors: Tao Lin, Kun Jin, Andrew Estornell, Xiaoying Zhang, Yiling Chen, Yang Liu

NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We then investigate, both theoretically and empirically, approaches for mitigating polarization and promoting diversity in recommender systems. We also provide empirical results (Section 5) on both synthetic and real-world (Movie Lens) data.
Researcher Affiliation Collaboration Tao Lin Harvard University tlin@g.harvard.edu Kun Jin Google kunjin@google.com Andrew Estornell Byte Dance andrew.estornell@bytedance.com Xiaoying Zhang Byte Dance zhangxiaoying.xy@bytedance.com Yiling Chen Harvard University yiling@seas.harvard.edu Yang Liu University of California, Santa Cruz yangliu@ucsc.edu
Pseudocode Yes Algorithm 1 Real-world Recommendation with Dual Influence
Open Source Code Yes Provided in the supplemental file.
Open Datasets Yes We conduct experiments on the Movie Lens 20M dataset [19].
Dataset Splits No The paper uses the Movie Lens 20M dataset and mentions 'train' and 'validation' in the context of the two-tower model, but it does not specify explicit training/validation/test dataset splits (e.g., percentages, sample counts, or specific predefined splits) that are needed to reproduce the data partitioning.
Hardware Specification No The paper does not provide specific hardware details (e.g., exact GPU/CPU models, memory amounts, or detailed computer specifications) used for running its experiments. The NeurIPS checklist states 'Computer resources are not a limitation in our experiments.'
Software Dependencies No The paper describes the computational model and architecture (e.g., 'two-tower model') but does not specify any software dependencies with version numbers (e.g., specific Python, PyTorch, or TensorFlow versions, along with other libraries) used in the experiments.
Experiment Setup Yes The dynamics is initialized by randomly generating user and creator features on the unit sphere in Rd. We pick d = 10, number of creators n = 50, number of users m = 100. We use the softmax recommendation probability function (2). We simulate the dynamics for T = 1000 steps, repeated 100 times each with a new initialization. We choose the sign impact function g(uj, vi) = sign( uj, vi ) for creator updates. For user updates, we choose inner product f(vi, uj) = vi, uj . We set them to β = 1, ηc = ηu = 0.1, and change one parameter at a time to see its effect on the dynamics.