Tsung-Ting Kuo

Weakly supervised learning of biomedical information extraction from curated data

BackgroundNumerous publicly available biomedical databases derive data by curating from literatures. The curated data can be useful as training examples for information extraction, but curated data usually lack the exact mentions and their locations in the text required for supervised machine learning. This paper describes a general approach to information extraction using curated data as training examples. The idea is to formulate the problem as cost-sensitive learning from noisy labels, where the cost is estimated by a committee of weak classifiers that consider both curated data and the text.ResultsWe test the idea on two information extraction tasks of Genome-Wide Association Studies (GWAS). The first task is to extract target phenotypes (diseases or traits) of a study and the second is to extract ethnicity backgrounds of study subjects for different stages (initial or replication). Experimental results show that our approach can achieve 87 % of Precision-at-2 (P@2) for disease/trait extraction, and 0.83 of F1-Score for stage-ethnicity extraction, both outperforming their cost-insensitive baseline counterparts.ConclusionsThe results show that curated biomedical databases can potentially be reused as training examples to train information extractors without expert annotation or refinement, opening an unprecedented opportunity of using “big data” in biomedical text mining.

Unsupervised link prediction using aggregative statistics on heterogeneous social networks

The concern of privacy has become an important issue for online social networks. In services such as Foursquare.com, whether a person likes an article is considered private and therefore not disclosed; only the aggregative statistics of articles (i.e., how many people like this article) is revealed. This paper tries to answer a question: can we predict the opinion holder in a heterogeneous social network without any labeled data? This question can be generalized to a link prediction with aggregative statistics problem. This paper devises a novel unsupervised framework to solve this problem, including two main components: (1) a three-layer factor graph model and three types of potential functions; (2) a ranked-margin learning and inference algorithm. Finally, we evaluate our method on four diverse prediction scenarios using four datasets: preference (Foursquare), repost (Twitter), response (Plurk), and citation (DBLP). We further exploit nine unsupervised models to solve this problem as baselines. Our approach not only wins out in all scenarios, but on the average achieves 9.90% AUC and 12.59% NDCG improvement over the best competitors. The resources are available at http://www.csie.ntu.edu.tw/~d97944007/aggregative/

Blockchain distributed ledger technologies for biomedical and health care applications

Abstract Objectives To introduce blockchain technologies, including their benefits, pitfalls, and the latest applications, to the biomedical and health care domains. Target Audience Biomedical and health care informatics researchers who would like to learn about blockchain technologies and their applications in the biomedical/health care domains. Scope The covered topics include: (1) introduction to the famous Bitcoin crypto-currency and the underlying blockchain technology; (2) features of blockchain; (3) review of alternative blockchain technologies; (4) emerging nonfinancial distributed ledger technologies and applications; (5) benefits of blockchain for biomedical/health care applications when compared to traditional distributed databases; (6) overview of the latest biomedical/health care applications of blockchain technologies; and (7) discussion of the potential challenges and proposed solutions of adopting blockchain technologies in biomedical/health care domains.

Modeling and evaluating information propagation in a microblogging social network

Microblogging platforms, such as Twitter and Plurk, allow users to express feelings, discuss ideas, and share interesting things with their friends or even strangers with similar interests. With the popularity of microblogs, there are growing data and opportunities in understanding information propagation behaviors in online social networks. Though some influence models had been proposed based on certain assumptions, most of them are based on the simulation approach (not data driven). This paper aims at designing a framework to model, measure, evaluate, and visualize influence propagation in a microblogging social network. Considering how information contents are spread in a social network, we devise two influence propagation models from the views of messages posted and responded. Based on the proposed models, we are able to measure the influence capability of an individual with respect to a user-given topic. Our design of influence measures consider (a) the number of people influenced, (b) the speed of propagation, and (c) the geographic distance of the propagation. To test the effectiveness of our influence model, we further propose a novel evaluation framework that predicts the propagation links and influential nodes in a real-world microblogging social network. Finally, we develop an online visualization system allowing users to explore the information propagation with the functions of displaying propagation structures, influence scores of individuals, timelines, and the geographical information for any user-query terms.

Learning to improve area-under-FROC for imbalanced medical data classification using an ensemble method

This paper presents our solution for KDD Cup 2008 competition that aims at optimizing the area under ROC for breast cancer detection. We exploited weighted-based classification mechanism to improve the accuracy of patient classification (each patient is represented by a collection of data points). Final predictions for challenge 1 are generated by combining outputs from weighted SVM and AdaBoost; whereas we integrate SVM, AdaBoost, and GA to produce the results for challenge 2. We have also tried location-based classification and model adaptation to add the testing data into training. Our results outperform other participants given the same set of features, and was selected as the joint winner in KDD Cup 2008.

A Content-Based Matrix Factorization Model for Recipe Recommendation

This paper aims at bringing recommendation to the culinary domain in recipe recommendation. Recipe recommendation possesses certain unique characteristics unlike conventional item recommendation, as a recipe provides detailed heterogeneous information about ingredients and cooking procedure. Thus, we propose to treat recipes as an aggregation of features, which are extracted from ingredients, categories, preparation directions, and nutrition facts. We then propose a content-driven matrix factorization approach to model the latent dimension of recipes, users, and features. We also propose novel bias terms to incorporate time-dependent features. The recipe dataset is available at http://mslab.csie.ntu.edu.tw/~tim/recipe.zip

Enriching Cold Start Personalized Language Model Using Social Network Information

We introduce a generalized framework to enrich the personalized language models for cold start users. The cold start problem is solved with content written by friends on social network services. Our framework consists of a mixture language model, whose mixture weights are es- timated with a factor graph. The factor graph is used to incorporate prior knowledge and heuris- tics to identify the most appropriate weights. The intrinsic and extrinsic experiments show significant improvement on cold start users.

Exploiting Temporal Information in a Two-Stage Classification Framework for Content-Based Depression Detection

Depression has become a critical illness in human society as many people suffer from the condition without being aware of it. The goal of this paper is to design a system to identify potential depression candidates based on their write-ups. To solve this problem, we propose a two-stage supervised learning framework. The first stage determines whether the user possesses apparent negative emotion. Then the positive cases are passed to the second stage to further evaluate whether the condition is clinical depression or just ordinary sadness. Our training data are generated automatically from Bulletin Board Systems. The content and temporal features are designed to improve the classification accuracy. Finally we develop an online demo system that takes a piece of written text as input, and outputs the likelihood of the author currently suffering depression. We conduct cross-validation and human study to evaluate the effectiveness of this system.

Assessing the Quality of Diffusion Models Using Real-World Social Network Data

Recently, there has been growing interest in understanding information cascading phenomenon on popular social networks such as Face book, Twitter and Plurk. The numerous diffusion events indicate huge governmental and commercial potential. People have proposed several diffusion and cascading models based on certain assumption, but until now we do not know which one is better in predicting information propagation. In this paper, we propose a novel framework that utilizes the micro-blog data to evaluate which model is better under different circumstances. In our framework, we devise two schemes for evaluation: the direct and the indirect schemes. We conduct experiments using three diffusion models on Plurk data. The results show Independent Cascade model outperforms other diffusion models using direct scheme, while Linear Threshold model, Degree, In-Degree and Page Rank perform best using indirect scheme. The main contribution is to provide a general evaluation framework for various diffusion models.

Learning-based concept-hierarchy refinement through exploiting topology, content and social information

Concept hierarchies, such as the ACM Computing Classification Scheme and InterPro Protein Sequence Classification, are widely used in categorization and indexing applications. In the Internet and Web 2.0 era, new concepts and terms are emerging on an almost daily basis, so it is essential that such hierarchies maintain up-to-date records of concepts. This paper proposes a mechanism to identify the most suitable position to insert new terms into an existing concept hierarchy. The problem is challenging because there are hundreds or even thousands of candidate positions for insertion. Furthermore, usually there is no training instance available for an insertion; nor is it practical to assume the availability of a detailed description of the target concept, except in the hierarchy itself. To resolve the problem, we exploit the topology, content and social information, and apply a learning approach to identify the underlying construction criteria of the concept hierarchy. We utilize three metrics (namely, accuracy, taxonomic closeness, and ranking) to evaluate the proposed learning-based approach on the ACM CCS, the DOAJ and the InterPro datasets to evaluate the proposed learning-based approach. The results demonstrate that, in all three metrics, our approach outperforms similarity-based approaches, such as the Normalized Google Distance, by a significant margin. Finally, we propose a level-based recommendation scheme as a novel application of our system. The source code, dataset, and other related resources are available at http://www.csie.ntu.edu.tw/~d97944007/refinement/.