Domain Adaptation And User-Transfer Based Personalization Applications

This article focus on domain-adaptation research and user-transfer based person-alization applications. It not only presents three creative algorithms for domain adapta-tion, but also analyzes the perspective of user-transfer based machine-learning applica-tions (e.g. brain-computer interface (BCI) system).During recent years, domain adaptation becomes a hot topic. It arises when the data distributions in the training and test domain are different to each other. The need for domain adaptation research is prevalent in many real-world application problems. For example, training data collected from different user groups can have different but related patterns.The first algorithm that this article presents is called as transferable discriminative dimensionality reduction (TDDR). In domain adaptation scenarios, previous discrimi-native dimensionality reduction methods tend to perform poorly owing to the difference between source and target distributions. In such cases, it is unsuitable to only consider discrimination in the low-dimensional source latent space since this would generalize badly to target domains. By resolving an objective function that encourages the sepa-ration of the domain-merged data and penalizes the distance between source and target distributions, TDDR can find a low-dimensional latent space which guarantees not only the discrimination of projected samples, but also the transferability to enable later clas-sification or regression models constructed in the source domain to generalize well to the target domain. In the experiments, we firstly verify the effectiveness of TDDR on a synthetic data. Then, we test TDDR on two real datasets from BCI applications. The experimental results show that the TDDR method can learn a low-dimensional latent feature space where the source models can perform well in the target domain.The second algorithm in this article is dynamical domain adaptation with adaptive and robust base-models and local-neighborhood-structure based weighting strategy. It is motivated by combining source domains with ensemble-learning framework. In the base-model construction step, it uses the1-norm regularization technology and differ-ent performance criterions to obtain two base-model subsets that ensure robustness and adaptiveness properties. In the weighting step, our method assigns dynamic weights to these base models according to the local structure of a given test sample. These weights represent the prediction consistency of the corresponding base model. This framework, as validated by experimental results, can achieve positive knowledge transfer for im- proving the performance in the domain-adaptation scene.This article finally proposes a novel ensemble-based approach for domain adap-tation based on the dynamical weighting idea. Concretely speaking, firstly, a model group is constructed with datasets from source domains. The models in the group will be different to each other owing to the distribution differences among source domains. Then, for each base model, a model-friendly classifier will be trained for predicting whether a test example should be classified by this model. This is achieved by construct a model friendly training set whose positive examples are those that the model can pre-dict rightly and negative examples are those that the model predicts wrongly. With the model-friendly training set, a classifier can be trained to point out which example the model can predict rightly. Finally, for each test example, the ensemble learner can ob-tain dynamical weights according to the outputs of the model-friendly classifiers. The experimental results show that our method can learn a final model performing well in the target domain.