Bayesian Deep Collaborative Matrix Factorization

Authors: Teng Xiao, Shangsong Liang, Weizhou Shen, Zaiqiao Meng5474-5481

AAAI 2019 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experiments conducted on two sparse datasets show that BDCMF can significantly outperform the state-of-the-art CF methods. and Experiments Experimental Setup Research Questions. The research questions guiding the remainder of the paper are: (RQ1) How does our proposed BDCMF-1 compare to state-of-the-art MF methods for CF on sparsity matrix? (RQ2) How do different parameter settings (e.g., the social parameter λq and content parameter λv) affect BDCMF-1? (RQ3) Does the Bayesian posterior estimation outperform Bayesian point estimation? Datasets. In order to answer our research questions, we conduct experiments on two real-world datasets from Lastfm 1 (lastfm-2k) and Delicious2 (delicious-2k) collected by Brusilovsky et al. (2010).
Researcher Affiliation Academia Teng Xiao,1,2 Shangsong Liang,1,2, Weizhou Shen,1,2 Zaiqiao Meng1,2 1School of Data and Computer Science, Sun Yat-sen University, China 2Guangdong Key Laboratory of Big Data Analysis and Processing, Guangzhou 510006, China {cstengxiao, liangshangsong}@gmail.com, shenwzh3@mail2.sysu.edu.cn, zqmeng@aliyun.com
Pseudocode Yes Algorithm 1 BDCMF-1 inference algorithm. Require: user-item matrix R, item content matrix X, parameters λv, λu, λg, and λq, and P = 100 and L = 1. 1: Randomly initialize variational parameters ˆui, ˆvj, ˆgk and network parameters θ and φ. 2: while not converged do...
Open Source Code No The paper does not contain any explicit statements or links indicating that the source code for the methodology is openly available.
Open Datasets Yes Datasets. In order to answer our research questions, we conduct experiments on two real-world datasets from Lastfm 1 (lastfm-2k) and Delicious2 (delicious-2k) collected by Brusilovsky et al. (2010). and 1http://www.lastfm.com. 2http://www.delicious.com.
Dataset Splits No To evaluate our model performance on extreme sparse matrix, we use 90% of dataset to train our model and the remainder for testing. For fair comparisons, We first set the parameters for PMF, So Rec, CTR, CTR-SMF, CDL, Neu MF via five-fold cross validation. The paper specifies a 90% training split and "the remainder for testing", but does not explicitly define a separate validation set split for their own model. While it mentions 5-fold cross-validation for baselines, it doesn't specify a validation split for their primary model's reproduction.
Hardware Specification No The paper does not provide specific details on the hardware used, such as GPU/CPU models or other computing specifications.
Software Dependencies No The paper mentions general software concepts like neural networks but does not provide specific version numbers for any programming languages, libraries, or solvers used in the experiments.
Experiment Setup Yes Settings. For fair comparisons, We first set the parameters for PMF, So Rec, CTR, CTR-SMF, CDL, Neu MF via five-fold cross validation. For PMF, we set D, λu and λv as 50, 0.01 and 0.001, respectively. The parameter settings for So Rec are λc = 10, λu = λv = λz = 0.001. For CTR, we find it achieve best performance when D=50, λu=0.1, λv=100, a=1 and b=0.01. CTR-SMF yields the best performance when D=75, λu=0.1, λv=100, λq=100, a=1 and b=0.01. For CDL, it achieves the best performance when we set a = 1, b = 0.01, D = 50, λu=1, λv=10, λn = 1000, and λw=0.0001. For Neu MF, we set D=50, and the last hidden layer is 16. For Poisson MF-CS, we set λc=0.1 and λs=0.1. For our model BDCMF, we set D=50. We set λv=1, λq=10 for dataset Lastfm, and set λv=0.1, λq=10 for dataset Delicious. ... The dimensions of the layers of the network for BDCMF are (L+N)-200-100(zj)-100-100-(L+N), where L and N are the dimension of xj and R j.