Frame-based Data Factorizations

Authors: Sebastian Mair, Ahcène Boubekki, Ulf Brefeld

ICML 2017 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We provide theoretical and empirical results for our proposed approach and make use of the frame to accelerate Archetypal Analysis. The novel method yields similar reconstruction errors as baseline competitors but is much faster to compute.
Researcher Affiliation Academia 1Leuphana University, L uneburg, Germany 2German Institute for Educational Research, Frankfurt am Main, Germany.
Pseudocode Yes Algorithm 1 Archetypal Analysis (AA)
Open Source Code No The paper mentions using third-party libraries like 'pymf' and 'scikit-learn' with their respective URLs, but does not provide concrete access to the source code for the methodology developed in this paper.
Open Datasets Yes For the experiments in this section we want to control the frame density. Hence, use the same synthetic data1 as in Dul a and L opez (2012) which was generated according to a procedure described in L opez (2005). The data consists of n = 2500, 5000, 7500, 10000 data points in d = 5, 10, 15, 20 dimensions with a frame density of 1, 15, 25, 50, 75 percent respectively. 1http://www.people.vcu.edu/ jdula/ Frames Algorithms/
Dataset Splits Yes For the hyperparameter search we use a hold out set consisting of another 500 images per class that are disjoint from the training set.
Hardware Specification No No specific hardware details (exact GPU/CPU models, processor types, or memory details) used for running the experiments were provided.
Software Dependencies No The paper mentions using 'Python', 'pymf', and 'scikit-learn' but does not specify their version numbers.
Experiment Setup Yes The number of iterations executed per algorithm is fixed to 100 in order to obtain fair results. We use random initializations for all algorithms and report on averages over 36 repetitions. [...] For every obtained embedding, an SVM is trained on p = 2, 4, . . . , 16 archetypes/filters meaning that every image is now represented by a vector of size m = 98, 196, . . . , 784. Note that the last one is identical to the original image size. We deploy an RBF kernel and the parameters C and γ are optimized on a 2 7, . . . , 27 grid respectively.