Rectangular Tiling Process

Authors: Masahiro Nakano, Katsuhiko Ishiguro, Akisato Kimura, Takeshi Yamada, Naonori Ueda

ICML 2014 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We used the following three real relational data sets: (a) Animal feature (50 85 binary data). We used a animalfeature matrix for 50 animals with 85 features (Kemp et al., 2006). (b) Donations (14 111 binary data). We used a political dataset for 14 countries with 111 binary features (Kemp et al., 2006). (c) Cities (55 46 categorical ( {0, 1, 2, 3}) data). This dataset consists of the distribution of offices for 46 service firms over 55 world cities. Service values for a firm in a city are given as 3, 2, 1 or 0 (Beaverstock et al., 2000).
Researcher Affiliation Industry NTT communication science laboratories, Morinosato Wakamiya 3-1, Atsugi-shi, Kanagawa
Pseudocode Yes Algorithm 1 LOCAL GROWTH ALGORITHM (LGA), Algorithm 2 REPEATED LGA (RLGA)
Open Source Code No The paper does not explicitly state that the source code for the methodology is available or provide a link to a repository.
Open Datasets Yes We used the following three real relational data sets: (a) Animal feature (50 85 binary data). We used a animalfeature matrix for 50 animals with 85 features (Kemp et al., 2006). (b) Donations (14 111 binary data). We used a political dataset for 14 countries with 111 binary features (Kemp et al., 2006). (c) Cities (55 46 categorical ( {0, 1, 2, 3}) data). This dataset consists of the distribution of offices for 46 service firms over 55 world cities. Service values for a firm in a city are given as 3, 2, 1 or 0 (Beaverstock et al., 2000).
Dataset Splits No The paper states, "For model comparison, we held out 20% of the data for testing," which implies a training split but does not mention a separate validation split.
Hardware Specification No The paper does not specify any hardware details (e.g., CPU, GPU models, memory) used for running the experiments.
Software Dependencies No The paper does not provide specific software dependencies with version numbers.
Experiment Setup Yes For inference, we set the real variable p = 0.7, set the Mondrian budget λ = 1, and let the intensity of the PPs and the MP be Lebesgue measures. In practice, we found that it was better to increase the frequency of the Metropolis-Hastings (MH) updates for rectangular partitioning since MH has lower acceptance rate than Gibbs (100% acceptance) for row and column entries. Thus, we performed one MH update (for rectangular partitioning) and one Gibbs update (for one row and one column) per iteration. To examine the influence of MCMC initialization, we also employed 3 types of manually-generated regular grid partitionings as initialization: (7 7) (referred as RTPs), (15 15) (RTPm), and (30 30) (RTPl).