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].

Adversarial Attacks on Online Learning to Rank with Stochastic Click Models

Authors: Zichen Wang, Rishab Balasubramanian, Hui Yuan, chenyu song, Mengdi Wang, Huazheng Wang

TMLR 2024 | Venue PDF | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experimental results based on synthetic and real-world data further validate the effectiveness and cost-efficiency of the proposed attack strategies. [...] In the experiment section, we apply the proposed attack methods against the OLTR algorithms listed in Table 1 with their corresponding click models. We compare the effectiveness of our attack on synthetic data and real-world Movie Lens dataset.
Researcher Affiliation Academia Zichen Wang EMAIL Department of Electrical Computer Engineering & Coordinated Science Laboratory University of Illinois Urbana-Champaign; Rishab Balasubramanian EMAIL Department of Computer Science Oregon State University; Yuan Hui EMAIL Department of Electrical Computer Engineering Princeton University; Chenyu Song EMAIL Department of Computer Science Oregon State University; Mengdi Wang EMAIL Department of Electrical Computer Engineering Princeton University; Huazheng Wang EMAIL Department of Computer Science Oregon State University
Pseudocode Yes Algorithm 1 Generalized List Poisoning Attack (GA); Algorithm 2 Attack-Then-Quit (ATQ) Algorithm; Algorithm 3 Cascade UCB1 (Kveton et al., 2015a); Algorithm 4 Batch Rank (Zoghi et al., 2017); Algorithm 5 Display Batch; Algorithm 6 Collect Clicks; Algorithm 7 Update Batch; Algorithm 8 The Top Rank (Lattimore et al., 2018)
Open Source Code Yes Our code and data can be accessed publicly for reproducibility.1. [Footnote 1: https://github.com/rishabbala/Online-Learning-to-Rank-for-Stochastic-Click-Models]
Open Datasets Yes We also evaluate the proposed attacks on Movie Lens dataset (Harper and Konstan, 2016).
Dataset Splits No The paper states: 'We first split the dataset into train and test data subsets.' for the Movie Lens dataset, but does not provide specific percentages, sample counts, or a detailed methodology for these splits. For synthetic data, it describes generation, not splitting of an existing dataset.
Hardware Specification No The paper does not provide specific hardware details (e.g., GPU/CPU models, memory amounts) used for running its experiments.
Software Dependencies No The paper does not provide specific software dependencies with version numbers (e.g., Python version, library versions like PyTorch, TensorFlow, scikit-learn).
Experiment Setup Yes For all our experiments, we use L = 50, K = 5 (the set up of Zoghi et al. (2017); Lattimore et al. (2018) is L = 10 and K = 5) and T = 10^5. For ATQ, we set the T1 in Algorithm 2 by Theorem 3 and Theorem 4. [...] We generate a size-L item set D, in which each item ak is related to a unique attractiveness score α(ak). Each attractiveness score α(ak) is drawn from a uniform distribution U(0, 1).