Elsevier

Information Sciences

Volume 390, June 2017, Pages 54-69
Information Sciences

A visual analytical approach for transfer learning in classification

https://doi.org/10.1016/j.ins.2016.03.021Get rights and content

Abstract

Classification can be highly challenging when the dataset is extremely large, or when the training data in the underlying domain are difficult to obtain. One feasible solution to this challenge is transfer learning, which extracts the knowledge from source tasks and applies the knowledge to target tasks. Extant transfer learning schemes typically assume that similarities between the source task and the target task to some degree. This assumption does not hold in certain actual applications; analysts unfamiliar with the learning strategy can be frustrated by the complicated transfer relations and the non-intuitive transfer process. This paper presents a suite of visual communication and interaction techniques to support the transfer learning process. Furthermore, a pioneering visual-assisted transfer learning methodology is proposed in the context of classification. Our solution includes a visual communication interface that allows for comprehensive exploration of the entire knowledge transfer process and the relevance among tasks. With these techniques and the methodology, the analysts can intuitively choose relevant tasks and data, as well as iteratively incorporate their experience and expertise into the analysis process. We demonstrate the validity and efficiency of our visual design and the analysis approach with examples of text classification.

Introduction

Conventional data analysis approaches assume that the training and test data are drawn from the same feature space and the same distribution. Knowledge transfer, also known as transfer learning, seeks to fulfill a new task by leveraging the knowledge and insights from accomplished tasks. The mechanism of transfer learning is similar to the “learning by analogy” approach, which understands a new situation by leveraging familiar conditions [21]. The “source analog” is an existing situation that conveys information and reference in exploring a novel “target”, with its main concern found in the similarity between the source and the target. The concept of “learning by analogy” that comes from cognitive psychology is relatively abstract, whereas transfer learning algorithms provide a formal description of analogy between learned tasks and new tasks in machine learning. The new task can be in the same scenario with the source tasks or even a different scenario, given that a type of analogy exists between the tasks or the underlying datasets. As such, transfer learning can be used to analyze a large amount of data by first training a part of it and transferring the result to the remaining data [40]. Similarly, knowledge transfer can be performed between two tasks performed in two different applications.

Transfer learning is truly beneficial because it can significantly reduce the cost and burden of exploring unknown data and unknown scenarios, and help discover common features hidden in tasks and data from different application fields. One example is the classification of web documents, in which bundles of data from different sources (e.g., forums, news press, and websites) are handled. Each data source has a specific word distribution, and leads to a unique data classification model. For a new data source, directly applying existing classification models will cause a performance drop. Transfer learning has proven to be significantly useful [14].

Despite the progress of transfer learning, many challenges exist regarding its usage in real applications. The most important challenge is the estimation of the transferability between the task pairs or the data pairs. The transferability actually measures the capability of the knowledge transfer for a specific task. With an appropriate transferability, the analyst can easily choose existing analysis results analogous to the target task, and perform the transfer. Conversely, determining the reusability of data instances for the new task is non-trivial. Both challenges cannot be fully addressed with an automatic transfer learning process when handling a complicated scenario, and may be aggravated by the complicated transfer relations and the non-intuitive transfer process.

We argue that applying visualization techniques in the learning process would be an effective means for incorporating the human intelligence into complicated analysis tasks. In particular, visualization technologies that integrate the intelligences from both the machine and human within a visual information communication interface have achieved significant success on many areas, including classification [17], [23], summarization, and clustering [11], [25], [49]. These techniques could be certainly adapted to the transfer learning process to bridge the gap between the domain experts and the transfer learning approaches.

This paper presents a pioneering visual-assisted transfer learning scheme in the context of text classification. The core idea of our work lies in our design of a suite of visualization communication and interaction techniques to enhance the analyst’s understanding and the manipulability of the knowledge transfer process. By incorporating the domain experiences and expertise into the knowledge transfer process within a visual exploration loop, the analyst can intuitively study the relevances among various tasks and data, and identify similar tasks. For analogous tasks, the analyst can interactively choose the related data instances, and use the trained data for a new task, which is supported by a pioneering visual representation of the classification result and the data instance similarity. To our best knowledge, our work is the first effort to apply visualization methods to the knowledge transfer process, and leads to the following main contributions:

  • a suite of visual communication and interaction techniques that support the knowledge transfer process;

  • a pioneering visual analytics based transfer learning methodology capable of analyzing relevances between different tasks and knowledge transfer on the level of data instances;

  • an exploratory data classification prototype called TransXplorer that follows the proposed techniques and methodology, and demonstrates high efficiency for text classification.

The rest of the paper is organized as follows. Section 2 summarizes related works. Section 3 provides the overview of the entire transfer learning framework. Sections 4 and 5 present detailed descriptions of its two stages. Section 6 presents experimental results and analysis. Section 7 present discussion and Section 8 draws conclusions.

Section snippets

Related work

A large body of literature is devoted to machine learning and text classification. Below, we briefly review the most relevant ones.

Text classification: Text classification deals with text feature representation and classification rules. First, feature representation plays an important role in text classification. Each text classification task requires specific text feature measurements. Successful text classification methods should choose the classification criteria relatively suitable for the

Approach overview

For clarity, we first explain the related concepts involved in transfer learning and text classification.

Background: For simplicity, we will describe the knowledge transfer framework in the context of binary classification of text. First, the bag of word (BOW) model is employed to vectorize each document, yielding a vector representation.

Task, domain, and model: Compared with the conventional classification approaches, multiple datasets(document collections) and associated trained models are

Interactive task analysis

Conventional transfer learning approaches assume that the source task/data are highly analogous to the target task/data. In practical applications, what makes the knowledge transfer infeasible is the uncertainty of the analogy degree between the source and target parts. If the transfer is performed between two non-analogous tasks, the transfer performance can be extremely low, and can even negatively influence the analysis of the new task. Thus, a generalized and model-free metric of task

Interactive transfer and classification

Efficiency of the target task not only depends on the selected source tasks, but is further influenced by the data instances used for training its classification model. Thus, one challenging problem in this stage is to find appropriate data instances from the source tasks, and transfer them to the target task. With them, the target task is accomplished. This is an iterative process, which runs until the analyst collects adequate data instances, and then accomplishes the target task.

Case studies

Two case studies were conducted with TransXplorer, namely, the Amazon product reviews dataset and the 20 newsgroups dataset, which are widely used in studying transfer learning.

Discussion

Case study: The case study demonstrates the applicability of our system in visualizing the task relevance and data instances. Our current visualization technique depicts the relations of tasks with the force-directed layout. The layout of task nodes may be slightly changed because of the instability of the iterative force-directed graph layout algorithm. The results demonstrate that it does not influence recognition of similar tasks because we use the distance to identify the task relevance.

Conclusion

This paper presents a suite of visualization techniques and a novel visual-assisted transfer learning methodology for performing binary classification. The key idea is to first assign a source task in the task exploration view, and then, select useful data instances from the source task by using the classifier transfer view. Experimental results on two real datasets demonstrate the validity and efficiency of our approach.

Acknowledgments

This work was supported by National 973 Program of China (2015CB352503), Major Program of the National Natural Science Foundation of China (61232012), National Natural Science Foundation of China (61422211), Zhejiang Provincial Natural Science Foundation of China (LR13F020001), and the Fundamental Research Funds for the Central Universities.

References (50)

  • D. Kifer et al.

    Detecting change in data streams

    Proceedings of the Thirtieth International Conference on Very Large Data Bases

    (2004)
  • Y. Liu et al.

    Design and evaluation of visualization support to facilitate decision trees classification

    Int. J. Humam-Comput. Study

    (2007)
  • B. Alsallakh et al.

    Visual methods for analyzing probabilistic classification data

    IEEE Trans. Vis. Comput. Graph.

    (2014)
  • S. Amershi et al.

    Regroup: interactive machine learning for on-demand group creation in social networks

    Proceedings of ACM SIGCHI Conference on Human Factors in Computing Systems

    (2012)
  • G. Andrienko et al.

    Interactive visual clustering of large collections of trajectories

    Proceedings of IEEE Conference on Visual Analytics Science and Technology

    (2009)
  • M. Ankerst et al.

    Visual classification: an interactive approach to decision tree construction

    Proceedings of ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, KDD’99

    (1999)
  • S. Ben-David et al.

    Analysis of representations for domain adaptation

    Adv. Neural Inf. Process. Syst.

    (2006)
  • E. Bertini et al.

    Surveying the complementary role of automatic data analysis and visualization in knowledge discovery

    Proceedings of ACM SIGKDD Workshop on Visual Analytics and Knowledge Discovery: Integrating Automated Analysis with Interactive Exploration, VAKD’09

    (2009)
  • J. Blitzer et al.

    Biographies, Bollywood, boom-boxes and blenders: domain adaptation for sentiment classification

    Proceedings of Annual Meeting of the Association For Computational Linguistics

    (2007)
  • H. Bosch et al.

    ScatterBlogs2: real-time monitoring of microblog messages through user-guided filtering

    IEEE Trans. Vis. Comput. Graph.

    (2013)
  • E.T. Brown et al.

    Dis-function:learning distance functions interactively

    Proceedings of IEEE Conference on Visual Analytics Science and Technology

    (2012)
  • A. Buja et al.

    Interactive data visualization with multidimensional scaling

    J. Comput. Graph. Stat.

    (2008)
  • N. Cao et al.

    DICON: interactive visual analysis of multidimensional clusters

    IEEE Trans. Vis. Comput. Graph.

    (2011)
  • J. Choo et al.

    iVisClassifier: an interactive visual analytics system for classification based on supervised dimension reduction

    Proceedings of IEEE Conference on Visual Analytics Science and Technology

    (2010)
  • Q. Cui et al.

    Measuring data abstraction quality in multiresolution visualization

    IEEE Trans. Vis. Comput. Graph.

    (2006)
  • W. Dai et al.

    Co-clustering based classification for out-of-domain documents

    Proceedings of ACM SIGKDD International Conference on Knowledge Discovery and Data Mining

    (2007)
  • E. Eaton et al.

    Modeling transfer relationships between learning tasks for improved inductive transfer

    Proceedings of European Conference on Machine Learning and Knowledge Discovery in Databases

    (2008)
  • E. Eaton et al.

    Interactive learning using manifold geometry

    Proceedings of AAAI Fall Symposium on Manifold Learning and its Applications

    (2009)
  • S. van den Elzen et al.

    BaobabView: Interactive construction and analysis of decision trees

    Proceedings of IEEE VAST’11

    (2011)
  • A. Endert et al.

    Observation-level interaction with statistical models for visual analytics

    Proceedings of IEEE Conference on Visual Analytics Science and Technology

    (2011)
  • J. Fogarty et al.

    CueFlik: interactive concept learning in image search

    Proceedings of ACM SIGCHI Conference on Human Factors in Computing Systems

    (2008)
  • T.M. Fruchterman et al.

    Graph drawing by force-directed placement

    Softw.: Pract. Exp.

    (1991)
  • D. Gentner et al.

    Reasoning and learning by analogy: introduction.

    Am. Psychol.

    (1997)
  • J. Han et al.

    Data Mining: Concepts and Techniques

    (2000)
  • F. Heimerl et al.

    Visual classifier training for text document retrieval

    IEEE Trans. Vis. Comput. Graph.

    (2012)
  • Cited by (16)

    • Visual ranking of academic influence via paper citation

      2018, Journal of Visual Languages and Computing
      Citation Excerpt :

      After the training process, a vectorized representation is obtained for each articles in the citation network. The cosine coefficients can be employed to measure the similarity between high-dimensional vectors [23], which will be further used to optimize traditional PageRank model. In addition to citation count, our weighted PageRank model also takes network structure and their implicit semantic relationship into account.

    • Exploring linear projections for revealing clusters, outliers, and trends in subsets of multi-dimensional datasets

      2018, Journal of Visual Languages and Computing
      Citation Excerpt :

      In this section, we demonstrate the usability of our system through three use cases, including visual cluster analysis, visual outlier analysis, and visual trend analysis. To evaluate the ability of cluster analysis, we use the 12-D synthetic dataset used in [28,38–40]. There are 750 points distributed in six Gaussian clusters.

    • Scalable transfer support vector machine with group probabilities

      2018, Neurocomputing
      Citation Excerpt :

      Different from conventional machine learning algorithms which assume that the training data should have the same distribution as that of the test data, transfer learning is able to utilize knowledge from data which follows a different distribution. Up to now, many transfer learning algorithms have been successfully applied in many areas, such as speech recognition, computer vision, information retrieval and natural language processing [12–15]. In this work, we construct a novel transfer learning framework with group probabilities using labeled samples in the source domain and the group probabilities in the target domain.

    • Transfer Learning with Deep Neural Networks for Image Classification in the E-commerce Industry

      2022, 2022 IEEE 7th International conference for Convergence in Technology, I2CT 2022
    View all citing articles on Scopus
    View full text