Confidence-Aware Matrix Factorization for Recommender Systems

Authors: Chao Wang, Qi Liu, Runze Wu, Enhong Chen, Chuanren Liu, Xunpeng Huang, Zhenya Huang

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

Reproducibility Variable Result LLM Response
Research Type Experimental Finally, extensive experiments on three real-world datasets demonstrate the effectiveness of our framework from multiple perspectives.
Researcher Affiliation Academia Anhui Province Key Laboratory of Big Data Analysis and Application, University of Science and Technology of China, {wdyx2012, wrz179, hxpsola, huangzhy}@mail.ustc.edu.cn, {qiliuql, cheneh}@ustc.edu.cn Decision Science and Management Information Systems Department, Drexel University, chuanren.liu@drexel.edu
Pseudocode Yes Algorithm 1 Gibbs sampling for CBPMF
Open Source Code No The paper does not provide any concrete access to source code for the methodology described, nor does it explicitly state that the code will be made open source.
Open Datasets Yes We conduct experiments on three real-world datasets, i.e., Movie Lens1and Netflix2 and Jester3. 1http://www.grouplens.org/node/73 2www.netflixprize.com 3http://eigentaste.berkeley.edu/dataset/
Dataset Splits No In our experiments, we randomly select 80% to 90% data as training sets according to different datasets and the rest part as test sets for five times. The paper specifies training and test splits, but does not explicitly mention a separate validation split for hyperparameter tuning.
Hardware Specification No The paper does not specify the hardware used to run the experiments, such as CPU or GPU models, or memory specifications.
Software Dependencies No The paper does not specify any software dependencies with version numbers.
Experiment Setup Yes In the proposed CBPMF, we introduced variance parameters γU and γV which obey Gamma distribution. As discussed in Section 3, the shape and rate of Gamma distribution is initialized via a = b so that the mean value would always be 1. Here, a and b are of great importance to statistical dispersion of Gamma distribution. Larger a and b would lead to more compact sampling results. We tune the value of a and b from the candidate set {5, 10, 20, 30, 50, 100, 200}. In PMF, following Mnih and Salakhutdinov(2008), we divide the datasets into mini-batches and update parameters after every mini-batch. In BPMF, following salakhutdinov and Mnih(2008), we set μ0 = 0 by symmetry, ν0 = D where D is the dimension of latent factors and W0 = ID D which is the identity matrix. To speed up the training process, the Gibbs sampler is initialized with PMF s output estimate. All parameters are tuned following the authors.