Multiview Triplet Embedding: Learning Attributes in Multiple Maps

Authors: Ehsan Amid, Antti Ukkonen

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

Reproducibility Variable Result LLM Response
Research Type Experimental Our contributions: We propose the Multiview Triplet Embedding (MVTE) algorithm for learning multiple maps from a given set of triplets. We propose a number of applications of the algorithm, and conduct experiments that show how the method can be used to identify tasks and items that are confusing to the workers, as well as to identify the attribute each worker mainly uses when comparing the items. We conduct the experiments on a set of artificial as well as real-world datasets.
Researcher Affiliation Academia Ehsan Amid EHSAN.AMID@AALTO.FI Aalto University and Helsinki Institute for Information Technology HIIT, Finland Antti Ukkonen ANTTI.UKKONEN@TTL.FI Finnish Institute of Occupational Health, Helsinki, Finland
Pseudocode Yes The pseudocode for the algorithm is shown in Algorithm 1.
Open Source Code Yes Our MATLAB implementation of the algorithm is publicly available online5. 5https://github.com/eamid/mvte
Open Datasets Yes We use a subset of 2000 datapoints from the MNIST dataset (Le Cun & Cortes, 1999). We first build a map using the t-STE algorithm by considering a set of 20,000 strongly satisfied synthetic triplets (see Figure 2(a)). We select three features, corresponding to three different measurements in the data: plasma glucose concentration, diastolic blood pressure, and 2-hour serum insulin. Each feature represents a different view for each instance. We generate 100 triplets for each datapoint in each view. The confusion matrix on the training signal and the generalization error6 on a 10-fold cross-validation are shown in Figure 3(a) and Figure 3(b), respectively.
Dataset Splits Yes The confusion matrix on the training signal and the generalization error6 on a 10-fold cross-validation are shown in Figure 3(a) and Figure 3(b), respectively.
Hardware Specification No The paper does not provide any specific hardware details such as GPU models, CPU models, or memory specifications used for running the experiments.
Software Dependencies No The paper states 'Our MATLAB implementation of the algorithm is publicly available online5.', but it does not specify the version number of MATLAB or any other software dependencies with their versions.
Experiment Setup Yes We select three features, corresponding to three different measurements in the data: plasma glucose concentration, diastolic blood pressure, and 2-hour serum insulin. Each feature represents a different view for each instance. We generate 100 triplets for each datapoint in each view. We use a subset of 2000 datapoints from the MNIST dataset... We first build a map using the t-STE algorithm by considering a set of 20,000 strongly satisfied synthetic triplets... We consider a subset of 156 objects out of 12 object categories. We form M = 2 spaces, corresponding to shapes and colors of the objects. Figure 5 shows the results of the MVTE method with M = 2.