Fixed-point algorithms for learning determinantal point processes

Authors: Zelda Mariet, Suvrit Sra

ICML 2015 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We present experimental results on both real and simulated data to illustrate the numerical performance of our technique. 3. Experimental results We compare performance of our algorithm, referred to as Picard iteration4, against the EM algorithm presented in Gillenwater et al. (2014). We experiment on both synthetic and real-world data.
Researcher Affiliation Academia Zelda Mariet ZELDA@CSAIL.MIT.EDU Suvrit Sra SUVRIT@MIT.EDU Massachusetts Institute of Technology, Cambridge, MA 02139 USA
Pseudocode Yes Pseudocode of our resulting learning method is presented in Algorithms 1 and 2.
Open Source Code No The paper states: "we used the code of Gillenwater et al. (2014) for our EM implementation". This indicates they used third-party code for comparison, but there is no explicit statement about releasing their own code for the Picard iteration.
Open Datasets Yes For real-world data, we use the baby registry test on which results are reported in (Gillenwater et al., 2014). This dataset consists in 111, 006 sub-registries describing items across 13 different categories; this dataset was obtained by collecting baby registries from amazon.com, all containing between 5 and 100 products, and then splitting each registry into subregistries according to which of the 13 categories (such as feeding , diapers , toys , etc.) each product in the registry belongs to. (Gillenwater et al., 2014) provides a more in-depth description of this dataset.
Dataset Splits No The paper only specifies a train/test split: "70% of the baby registries in the product category were used for training; 30% served as test." No explicit mention of a separate validation set.
Hardware Specification Yes These experiments were run with MATLAB, on a Linux Mint system, using 16GB of RAM and an i7-4710HQ CPU @ 2.50GHz.
Software Dependencies No The paper mentions "MATLAB" and "Linux Mint system", but does not provide specific version numbers for these or any other software libraries or dependencies used in the experiments.
Experiment Setup Yes We used a tighter convergence criterion for our algorithm (εpic = 0.5 εem) to account for the fact that the distance between two subsequent log-likelihoods tends to be smaller for the Picard iteration than for EM. The parameter a for Picard was set at the beginning of each experiment and never modified as it remained valid throughout each test. In EM, the step size was initially set to 1 and halved when necessary, as per the algorithm described in (Gillenwater et al., 2014); we used the code of Gillenwater et al. (2014) for our EM implementation.