Learning Bipartite Graphs: Heavy Tails and Multiple Components

Authors: José Vinícius de Miranda Cardoso , Jiaxi Ying, Daniel Palomar

NeurIPS 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental The proposed estimators outperform state-of-the-art methods for bipartite graph learning, as evidenced by real-world experiments using financial time series data.
Researcher Affiliation Academia José Vinícius de M. Cardoso1, Jiaxi Ying2, Daniel P. Palomar1,3 Department of Electronic and Computer Engineering1 Department of Mathematics2 Department of Industrial Engineering and Decision Analytics3 The Hong Kong University of Science and Technology Clear Water Bay, Hong Kong SAR China {jvdmc, jx.ying}@connect.ust.hk, palomar@ust.hk
Pseudocode Yes Algorithm 1: Gaussian bipartite graph learning (GBG) Algorithm 2: Student-t bipartite graph learning (SBG) Algorithm 3: Student-t k-component bipartite graph learning (k SBG)
Open Source Code Yes The code, data, and instructions to reproduce the experiments are available in https: //github.com/mirca/bipartite.
Open Datasets Yes We perform experiments using publicly available, historical daily prices, queried from Yahoo! Finance TM, of r = 333 stocks listed in the S&P500 Index and q = 8 sectors indexes
Dataset Splits Yes More precisely, we use a window of length 504 days (2 years in terms of stock market days) and shift this window by 63 days (3 months in terms of stock market days).
Hardware Specification No The paper mentions that the algorithms were implemented using the R programming language but does not specify any hardware details like GPU models, CPU types, or memory.
Software Dependencies No The proposed algorithms and benchmarks were implemented using the R programming language. For SGA and SGLA, we relied on their official implementation via the package spectral Graph Topology [46]. (No version numbers are provided for R or the package).
Experiment Setup Yes In our Algorithm 3, we set the hyperparameter ρ = 1 and the relative tolerance ϵ = 10 5. An estimate for the degree of freedoms ν, required by Algorithms 2 and 3, is obtained by fitting a univariate Student-t distribution to the log-returns of the S&P500 index during the corresponding time period.