Online Clustering of Contextual Cascading Bandits

Authors: Shuai Li, Shengyu Zhang

AAAI 2018 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We conduct experiments on both synthetic and real data, and demonstrate the effectiveness of our algorithm and the advantage of incorporating online clustering method.
Researcher Affiliation Academia Shuai Li, Shengyu Zhang The Chinese University of Hong Kong, Sha Tin, Hong Kong {shuaili, syzhang}@cse.cuhk.edu.hk
Pseudocode Yes Algorithm 1 CLUB-cascade
Open Source Code No The paper does not provide any explicit statement or link indicating that the source code for the described methodology is publicly available.
Open Datasets Yes restaurant recommendations with Yelp dataset1. The dataset contains user ratings for several businesses. 1http : //www.yelp.com/dataset challenge. In this experiment, we compare our algorithm, CLUBcascade, with C3-UCB/Cascade Lin UCB on the real dataset Movie Lens (Harper and Konstan 2016). We use the processed 20m dataset3, in which there are 20 million ratings for 27k movies by 138k users. 3https : //grouplens.org/datasets/movielens/20m/
Dataset Splits No The paper describes the construction of feature vectors from historical data and the use of 'future' matrices for online experiments, but it does not specify traditional train/validation/test dataset splits (e.g., percentages, explicit counts, or predefined standard splits) for reproducibility of data partitioning in a supervised learning context.
Hardware Specification No The paper does not provide any specific details about the hardware (e.g., GPU models, CPU types, memory) used to run the experiments.
Software Dependencies No The paper does not specify any software dependencies with version numbers (e.g., specific libraries, frameworks, or programming language versions) used for the experiments.
Experiment Setup Yes Input: λ, α, β > 0. Let λ K, β = λd + 2 ln(4mn) + λ and α = 4 d/λx, where d, m, u denotes the feature dimension, the number of clusters and the number of users, respectively. For each cluster j [m], we fix a weight vector θj with norm 1. In each round, a random user comes and the algorithm recommends K = 4 items to the user. In all the four settings, we randomly choose a content set with L = 200 items, each of which has a feature vector x Rd with x 2 1 and d = 20. We use u = 40 users and assign them randomly to m = 2, 5 clusters. To accelerate our algorithm, we use a sparse initialization instead of the complete graph initialization, similar in (Gentile, Li, and Zappella 2014).