Feature-Level Domain Adaptation
Authors: Wouter M. Kouw, Laurens J.P. van der Maaten, Jesse H. Krijthe, Marco Loog
JMLR 2016 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Our empirical evaluation of flda focuses on problems comprising binary and count data in which the transfer can be naturally modeled via a dropout distribution, which allows the classifier to adapt to differences in the marginal probability of features in the source and the target domain. Our experiments on several real-world problems show that flda performs on par with state-of-the-art domain-adaptation techniques. |
| Researcher Affiliation | Academia | Wouter M. Kouw EMAIL Laurens J.P. van der Maaten EMAIL Department of Intelligent Systems Delft University of Technology Mekelweg 4, 2628 CD , the Netherlands Jesse H. Krijthe EMAIL Department of Intelligent Systems Delft University of Technology Mekelweg 4, 2628 CD Delft, the Netherlands Department of Molecular Epidemiology Leiden University Medical Center Eindhovenweg 20, 2333 ZC Leiden, the Netherlands Marco Loog EMAIL Department of Intelligent Systems Delft University of Technology Mekelweg 4, 2628 CD Delft, the Netherlands The Image Group, University of Copenhagen Universitetsparken 5, DK-2100, Copenhagen , Denmark |
| Pseudocode | Yes | Algorithm 1 Binary flda with dropout transfer model and quadratic loss function. procedure flda-q(S, T) for d=1,. . . , m do ηd = |S| 1 P xi S 1xid =0 ζd = |T| 1 P zj T 1zjd =0 θd = max n 0, 1 ζd / ηd o end for w = XX + diag θ 1 θ XX 1Xy Element-wise product return sign(w Z) end procedure |
| Open Source Code | No | The paper discusses the source code of a third-party tool, libsvm, that the authors used: 'We made use of the libsvm package by Chang and Lin (2011) with a radial basis function kernel and we performed cross-validation to estimate the kernel bandwidth and the ℓ2-regularization parameter. All multi-class classification is done through an one-vs-one scheme. This method can be readily compared to subspace alignment (sa) and transfer component analysis (tca) to evaluate the effects of the respective adaptation approaches.' and 'Software available at http://www.csie.ntu.edu.tw/ cjlin/libsvm.' There is no explicit statement about releasing the source code for their own methodology (flda). |
| Open Datasets | Yes | We have collected six data sets from the UCI machine learning repository (Lichman, 2013) with missing data: Hepatitis (hepat.), Ozone (ozone; Zhang and Fan, 2008), Heart Disease (heart; Detrano et al., 1989), Mammographic masses (mam.; Elter et al., 2007), Automobile (auto), and Arrhythmia (arrhy.; Guvenir et al., 1997). We created a domain adaptation setting by considering two handwritten digit sets, namely MNIST (Le Cun et al., 1998) and USPS (Hull, 1994). The Office-Caltech data set (Hoffman et al., 2013). The Internet Movie Database (IMDb) (Pang and Lee, 2004). two data sets from the UCI machine learning repository: one containing 4205 emails from the Enron spam database (Klimt and Yang, 2004) and one containing 5338 text messages from the SMS-spam data set (Almeida et al., 2011). We performed a similar experiment on the Amazon sentiment analysis data set of product reviews (Blitzer et al., 2007). |
| Dataset Splits | Yes | The source training and validation data was generated from the same bivariate Poisson distributions as in Figure 2. The target data was constructed by generating additional source data and dropping out the first feature with a probability of 0.5. Each of the four data sets contained 10, 000 samples. The experiment was repeated 50 times for every sample size to calculate the standard error of the mean. In the experiments, we construct the training set (source domain) by selecting all samples with no missing data, with the remainder as the test set (target domain). In all experiments, we estimate the hyperparameters, such as ℓ2-regularization parameters, via cross-validation on held-out source data. |
| Hardware Specification | No | The paper does not provide any specific details about the hardware (e.g., GPU/CPU models, memory) used to run the experiments. It only describes the methodology and datasets. |
| Software Dependencies | No | The paper mentions 'libsvm package by Chang and Lin (2011)' and provides a link, but it does not specify a version number for libsvm that was used in their experiments. No other specific software dependencies with version numbers are provided. |
| Experiment Setup | Yes | In the first experiment, we generate binary features by drawing 100, 000 samples from two bivariate Bernoulli distributions. The marginal distributions are 0.7 0.7 for class one and 0.3 0.3 for class two. The source data is transformed to the target data using a dropout transfer model with parameters θ = 0.5 0 . In the second experiment, we generate count features by sampling from bivariate Poisson distributions. Herein, we used rate parameters λ = 2 2 for the first class and λ = 6 6 for the second class. |