活動議程

在一整天前瞻豐富的演講分享後，誠摯邀請資料科學同好參與今年度的交流晚宴。所有講師跟主持人信步其中，與年輕積極的學子們暢談知識與人生，在美食與音樂薰陶下，各位朋友無論交流言歡、或要拓廣人脈，必定更加得心應手。晚宴票種名額有限，請勿錯過！

Previous Next

Abstract :
在這個演講中，我將與大家分享廿餘年來一些研究、教學及產學合作的經驗。這些經驗有些頗為成功，但也有些不盡如人意，因此帶給我若干反思。希望這些回顧與反思，對一些正在或即將面臨類似情境的年輕同仁有參考價值。雖然這是一場名之為「傳薪」的研討會，但「火傳薪豈盡，漏響夜方深」，期待與大家一起再戰十年。

Abstract :
Causal inference has become a popular approach for investigating causality and mechanism in complex data format. In the first part, I will provide a general picture of the causal inference framework: defining a scientific question as a causal parameter based on counterfactual model, identifying this causal parameter as a statistical parameter, and then estimating (or testing) this statistical parameter by statistics inference. The mechanism investigation for the effect of religious behavior on survival rate will be used as the motivating example. In the second part, I will show how to define and identify the longitudinal total and mediation effects (two effects of interest in mechanism investigation) as analytic formula, and obtains parametric estimates by Monte-Carlo Simulation and bootstrapping method. In the Framingham Heart Study data, we apply this method to estimate the mediation effect of smoking behaviors sustained over a 10-year period on blood pressure when considering weight change as a time-varying mediator. Compared with non-smoking, smoking 20 cigarettes per day for 10 years was estimated to increase blood pressure by 1.18 (95 % CI: -0.68, 2.69) mm-Hg. The mediation effect was -0.34 (95 % CI: -0.52, -0.13) mm-Hg, which is negative because smoking leads to lower weight which leads to lower blood pressure. These results provide evidence that weight change in fact partially conceals the detrimental effects of cigarette smoking on blood pressure. In addition, I will also introduce an approach to conduct mediation analysis for survival data with time-varying exposures, mediators, and confounders. We identify certain interventional direct and indirect effects through a survival mediational g-formula and describe the required assumptions. We also provide a feasible parametric approach along with an algorithm and software to estimate these effects. We apply this method to analyze the Framingham Heart Study data to investigate the causal mechanism of smoking on mortality through coronary artery disease. The risk ratio of smoking 30 cigarettes per day for ten years compared with no smoking on mortality is 2.34 (95 % CI = (1.44, 3.70)). Of the overall effect, 7.91% (95% CI: = 1.36%, 19.32%) is mediated by coronary artery disease. Our work represents the first application of the parametric mediational g-formula in an epidemiologic cohort study.

Abstract :
Many studies measure the same type of information repeatedly on the same subject at multiple time points or locations, and each subject has a curve or an image. Clustering of such data has many important applications. We propose novel and easy to implement methods based on spline functions and random effect models. These methods handle data observed at regular or irregular points in the same way. We start with an example of one-dimensional curves, in which case our method takes into account the estimation uncertainty and can be used for outlier detection as well. We then proceed to another example for clustering two- or three- dimensional medical images, where quantitative imaging biomarkers are extracted and our proposed method can capture broader scale spatial patterns than existing techniques.

Abstract :
For high-reliability products, degradation analysis is a commonly used tool to predict the product's lifetime. For the degradation model, the gamma process (GP) model is widely applied to the monotonic degradation path. However, when the degradation path exhibits a non-decreasing situation, the GP model can’t be used to analyze this kind of data. In this work, we develop the non-decreasing gamma process model incorporated mixture probability distribution to deal with the non-decreasing degradation data. Furthermore, taking the fact that the degree of decay is violent with tested time into account, we consider the probability of zero-increment should be affected by the tested time. Therefore, we propose the non-homogeneously non-decreasing gamma process (NNGP) model incorporated assuming the probability of zero-increment to be a function of time. Based on the simplified method of multiple optimizations, the maximum likelihood estimates of the NNGP model parameters can be obtained as well as the reliability prediction. Simulation studies is carried out to examine the NNGP model. Finally, a real data is applied to the proposed NNGP model to reveal the benefits. (This is joint work with Shi- Hong Chen)

Abstract :
In this talk, I will focus on the regression analysis based on two types of survival data without survival time: prevalent data without follow-up as well as incident and prevalent covariate data. Under fully nonparametric consideration, the conditional survival function, and hence the absolute covariate effect, is not estimable due to fully right censoring. However, the central subspace can still be estimated based on these two types of data. Different from existing methods in the literature, this feature requires no parametric or semiparametric model assumptions. In addition, the basis matrix of central subspace can be interpreted as relative covariate effects and used to screen inactive covariates. According to these findings, we propose a penalized cross-validation type criterion to estimate the central subspace and select active covariates. This method requires no stringent distributional assumptions on covariates such as linearity condition or constant variance. Moreover, the proposed estimator is shown to have the oracle property of variable selection.

Abstract :
We consider the problem of selecting high-dimensional regression models with heteroscedastic errors, which have broad applications in biology, economics, industry, and many other fields. A high-dimensional dispersion function is employed to account for the heteroscedasticity. By making use of the orthogonal greedy algorithm and the high-dimensional information criterion, we propose a new model selection procedure that consistently chooses the relevant variables in both the regression and the dispersion functions. The finite sample performance of the proposed procedure is also illustrated via simulations and real data analysis.

Abstract :
With the development of sequencing techniques, increased numbers of multi-omics data and single-cell molecular profiling data from large cohorts are made available. Due to the large case-to-case or sell-to-sell heterogeneity, many associations between molecular and clinical features are expected to be presented only in some subpopulations. How to decipher the heterogeneity from the complex association became one of the major challenges in big bio-data analysis. In this talk, I will share my experiences in bioinformatics research. I will talk about the challenges biologists are facing and the opportunities for statisticians in this big bio-data era. Lastly, I will also discuss the prerequisites and the benefits of interdisciplinary research for statisticians.

Abstract :
本次演講會先跟大家分享，我從生物領域跨到資訊領域，再跳到統計領域的一些心得感想。
最初我在陽明唸書，是一個畢業後理論上要去醫院工作的系，但到生資所後，發現自己更喜愛理論，所以做了一點演算法理論的研究。之後覺得許多生資使用的方式沒有很穩固的數學支持，所以又跳進統計領域摸索。
再來會跟大家介紹一下因果模型，因果分析在哲學上的一些根源，以及我在黃彥棕老師家學習因果模型的心得。
最後跟大家分享一下最近的研究興趣跟進展。

Abstract :
In this talk, we propose forward variable selection procedures for feature screening in ultra-high-dimensional generalized varying coefficient models. We employ regression spline to approximate coefficient functions and then maximize the log-likelihood to select an additional relevant covariate sequentially. If we decide we do not significantly improve the log-likelihood any more by selecting any new covariates from our stopping rule, we terminate the forward procedures and give our estimates of relevant covariates. The effect of the size of the current model has been overlooked in stopping rules for sequential procedures for high-dimensional models. Our stopping rule takes into account the size of the current model suitably. Our forward procedures have screening consistency and some other desirable properties under regularity conditions. We also present the results of numerical studies to show their good finite sample performances.

Abstract :
In this article, we analyze informatively-censored recurrence processes using general semiparametric latent intensity regression models. When the distributions of the subject-specific latent variable and the censoring time are left unspecified, the distinct distributional features of the recurrent event times are found to be linked to the shape parameter, which, hence, merits the development of estimation and testing procedures. In light of this finding, two contrasting estimation methods are proposed for shape-dependent and -independent models. Especially, the estimation criteria are useful in building test rules to distinguish between competing rate regression models without the need to specify a significance level. Under very mild conditions, we establish large-sample properties of the estimators and test statistics. Comprehensive simulations are further conducted to assess their finite-sample performance. Moreover, our methodology is demonstrated in applications to recurrent event samples of intravenous drug users needing inpatient care and patients with chronic granulomatous disease.

Abstract :
As the structure prediction of 3D protein structure becomes mature after the release of AlphaFold2 from DeepMind last December, the obvious next step is to predict the movement and flexibility of the underlying protein which will provide crucial medical insight for drug or vaccine development. Among the tools available in structure biology, Cryogenic Electron Microscopy (cryo-EM) is a promising and powerful technique with high efficiency that can perform conformation analysis. Remarkably, the first dynamic movie of 2019-nCoV Spike trimer structure is derived from cryo- EM within 12 days using computational methods. However, the data characteristics include strong noise, huge dimension, large sample size and high heterogeneity with unknown orientations have made analysis very challenging. In this talk, I will first introduce the related background and principles of cryo-EM image processing. Second, I will discuss our proposed dimension reduction strategy called two-stage dimension reduction (2SDR) and clustering approach called DRMRA which alleviate the computation burden and improves performance over existing methods. Specifically, I will discuss how to use the 2SDR to improve the workflow and perform 3D conformation analysis. Finally, I will discuss future directions in 3D conformation analysis.

Abstract :
在資料分析的範圍裡，如何視覺化高維度的資料，一直是一門重要的課題。尤其在現今的時代裡，不僅僅是資料變數多，在一些領域當中，如生物、醫療、基因，資料數量也多。因此如果可以將一份高維度資料合理的呈現出來，對整體資料分析與探討是大有幫助的。目前已有許多視覺化的方法，如PCA、MDS，但是這些方法較適用於可以被線性區分的資料，一旦處理高維度資料，且無法用現性區分開來，則視覺化的效果不佳。Maaten and Hinton (2008) 提出的t-SNE，是是一種用非線性轉換，將高維度資料以二維或是三維視覺化呈現的非監督式(unsupervised) 方法，同時也是首個可以將MNIST成功呈現視覺化的方法。但是處理巨量化資料上，亦面臨到無法將資料一次分析完畢的難題，因此我們提出了Distributed t-SNE，利用多次分割再重組的過程，每次只需要處理部分資料，就可以將巨量資料視覺化。最後我們也證明Dt-SNE的大樣本性質，在起始值及迭代次數一樣的條件下，Dt-SNE結果會近似t-SNE的結果。

Abstract :
Count data typically exhibit an unexpected preponderance of zeros in various fields including species abundance, clinical trials, social science, traffic accidents, manufacturing, etc. We consider a problem of estimation in the presence of missing data. We develop several semiparametric double robustness M-type estimators under the excess zeros count regression model framework in the presence of missing data. The proposed estimators are roots of the estimating equations that use the nonparametric nuisance parameters to estimate the selection probability and the augmentation in the preliminary stage. Here, the nonparametric nuisance parameters are of the kernel smoothing and generalized additive models types. Since the resulting M-estimations at the second stage are all semiparametric types, so are the proposed estimators. Moreover, using the parametric estimator as the benchmark, we examine the relative efficiency of the proposed estimators and compare their performances by conducting a simulation study and exploring a real example.