A major assumption in many machine learning and data mining algorithms is that the training and future data must be in the same feature space and have the same distribution. However, in many real-world applications, this assumption may not hold. For example, we sometimes have a classification task in one domain of interest, but we only have sufficient training data in another domain of interest, where the latter data may be in a different feature space or follow a different data distribution. In such cases, knowledge transfer, if done successfully, would greatly improve the performance of learning by avoiding much expensive data labeling efforts. In recent years, transfer learning has emerged as a new learning framework to address this problem. This survey focuses on categorizing and reviewing the current progress on transfer learning for classification, regression and clustering problems. In this survey, we discuss the relationship between transfer learning and other related machine learning techniques such as domain adaptation, multitask learning and sample selection bias, as well as co-variate shift. We also explore some potential future issues in transfer learning research.

Traditional machine learning makes a basic assumption: the training and test data should be under the same distribution. However, in many cases, this identicaldistribution assumption does not hold. The assumption might be violated when a task from one new domain comes, while there are only labeled data from a similar old domain. Labeling the new data can be costly and it would also be a waste to throw away all the old data. In this paper, we present a novel transfer learning framework called TrAdaBoost, which extends boosting-based learning algorithms (Freund & Schapire, 1997). TrAdaBoost allows users to utilize a small amount of newly labeled data to leverage the old data to construct a high-quality classification model for the new data. We show that this method can allow us to learn an accurate model using only a tiny amount of new data and a large amount of old data, even when the new data are not sufficient to train a model alone. We show that TrAdaBoost allows knowledge to be effectively transferred from the old data to the new. The effectiveness of our algorithm is analyzed theoretically and empirically to show that our iterative algorithm can converge well to an accurate model.

Domain adaptation is an important problem in natural language processing (NLP) due to the lack of labeled data in novel domains. In this paper, we study the domain adaptation problem from the instance weighting perspective. We formally analyze and characterize the domain adaptation problem from a distributional view, and show that there are two distinct needs for adaptation, corresponding to the different distributions of instances and classification functions in the source and the target domains. We then propose a general instance weighting framework for domain adaptation. Our empirical results on three NLP tasks show that incorporating and exploiting more information from the target domain through instance weighting is effective. 1

The effectiveness of knowledge transfer using classification algorithms depends on the difference between the distribution that generates the training examples and the one from which test examples are to be drawn. The task can be especially difficult when the training examples are from one or several domains different from the test domain. In this paper, we propose a locally weighted ensemble framework to combine multiple models for transfer learning, where the weights are dynamically assigned according to a model’s predictive power on each test example. It can integrate the advantages of various learning algorithms and the labeled information from multiple training domains into one unified classification model, which can then be applied on a different domain. Importantly, different from many previously proposed methods, none of the base learning method is required to be specifically designed for transfer learning. We show the optimality of a locally weighted ensemble framework as a general approach to combine multiple models for domain transfer. We then propose an implementation of the local weight assignments by mapping the structures of a model onto the structures of the test domain, and then weighting each model locally according to its consistency with the neighborhood structure around the test example. Experimental results on text classification, spam filtering and intrusion detection data sets demonstrate significant improvements in classification accuracy gained by the framework. On a transfer learning task of newsgroup message categorization, the proposed locally weighted ensemble framework achieves 97 % accuracy when the best single model predicts correctly only on 73 % of the test examples. In summary, the improvement in accuracy is over 10 % and up to 30 % across different problems.

Multi-task learning is the problem of maximizing the performance of a system across a number of related tasks. When applied to multiple domains for the same task, it is similar to domain adaptation, but symmetric, rather than limited to improving performance on a target domain. We present a more principled, better performing model for this problem, based on the use of a hierarchical Bayesian prior. Each domain has its own domain-specific parameter for each feature but, rather than a constant prior over these parameters, the model instead links them via a hierarchical Bayesian global prior. This prior encourages the features to have similar weights across domains, unless there is good evidence to the contrary. We show that the method of (Daumé III, 2007), which was presented as a simple “preprocessing step, ” is actually equivalent, except our representation explicitly separates hyperparameters which were tied in his work. We demonstrate that allowing different values for these hyperparameters significantly improves performance over both a strong baseline and (Daumé III, 2007) within both a conditional random field sequence model for named entity recognition and a discriminatively trained dependency parser. 1

A basic assumption in traditional machine learning is that the training and test data distributions should be identical. This assumption may not hold in many situations in practice, but we may be forced to rely on a different-distribution data to learn a prediction model. For example, this may be the case when it is expensive to label the data in a domain of interest, although in a related but different domain there may be plenty of labeled data available. In this paper, we propose a novel transfer-learning algorithm for text classification based on an EM-based Naive Bayes classifiers. Our solution is to first estimate the initial probabilities under a distribution Dℓ of one labeled data set, and then use an EM algorithm to revise the model for a different distribution Du of the test data which are unlabeled. We show that our algorithm is very effective in several different pairs of domains, where the distances between the different distributions are measured using the Kullback-Leibler (KL) divergence. Moreover, KL-divergence is used to decide the trade-off parameters in our algorithm. In the experiment, our algorithm outperforms the traditional supervised and semi-supervised learning algorithms when the distributions of the training and test sets are increasingly different.

Supervised sequence-labeling systems in natural language processing often suffer from data sparsity because they use word types as features in their prediction tasks. Consequently, they have difficulty estimating parameters for types which appear in the test set, but seldom (or never) appear in the training set. We demonstrate that distributional representations of word types, trained on unannotated text, can be used to improve performance on rare words. We incorporate aspects of these representations into the feature space of our sequence-labeling systems. In an experiment on a standard chunking dataset, our best technique improves a chunker from 0.76 F1 to 0.86 F1 on chunks beginning with rare words. On the same dataset, it improves our part-of-speech tagger from 74 % to 80 % accuracy on rare words. Furthermore, our system improves significantly over a baseline system when applied to text from a different domain, and it reduces the sample complexity of sequence labeling. 1

A common assumption in supervised learning is that the input points in the training set follow the same probability distribution as the input points that will be given in the future test phase. However, this assumption is not satisfied, for example, when the outside of the training region is extrapolated. The situation where the training input points and test input points follow different distributions while the conditional distribution of output values given input points is unchanged is called the covariate shift. Under the covariate shift, standard model selection techniques such as cross validation do not work as desired since its unbiasedness is no longer maintained. In this paper, we propose a new method called importance weighted cross validation (IWCV), for which we prove its unbiasedness even under the covariate shift. The IWCV procedure is the only one that can be applied for unbiased classification under covariate shift, whereas alternatives to IWCV exist for regression. The usefulness of our proposed method is illustrated by simulations, and furthermore demonstrated in the brain-computer interface, where strong non-stationarity effects can be seen between training and test sessions. c2000 Masashi Sugiyama, Matthias Krauledat, and Klaus-Robert Müller.

The problem of transfer learning, where information gained in one learning task is used to improve performance in another related task, is an important new area of research. In this paper we address the subproblem of domain adaptation, in which a model trained over a source domain is generalized to perform well on a related target domain, where these two domains ’ data are distributed similarly, but not identically. Previous work has studied the supervised version of this problem in which labeled data from both source and target domains are available for training. In this work, however, we study the more challenging problem of unsupervised transductive transfer learning, where no labeled data from the target domain are available at training time, but instead, unlabeled target test data are available during training. We describe some current state-of-the-art inductive and transductive approaches involving three popular learning models, namely the maximum entropy, support vector machines and naive Bayes models. We then adapt these models to the problem of transfer learning for protein name extraction. In the process, we introduce a novel maximum entropy based technique, Iterative Feature Transformation (IFT), and show that it achieves comparable performance with state-of-the-art transductive SVMs. Finally, we compare the relative strengths and weaknesses of these models across the various learning settings, shedding light both on the algorithms examined and the difficulty of the respective problems. In addition, we show how simple relaxations, such as providing additional information like the proportion of positive examples in the test data, can significantly improve the performance of some of the transductive transfer learners. 1

We describe a new approach to SMT adaptation that weights out-of-domain phrase pairs according to their relevance to the target domain, determined by both how similar to it they appear to be, and whether they belong to general language or not. This extends previous work on discriminative weighting by using a finer granularity, focusing on the properties of instances rather than corpus components, and using a simpler training procedure. We incorporate instance weighting into a mixture-model framework, and find that it yields consistent improvements over a wide range of baselines. 1