In unsupervised domain adaptation (UDA), classifiers for the target domain (TD) are trained with clean labeled data from the source domain (SD) and unlabeled data from TD. However, in the wild, it is hard to acquire a large amount of perfectly clean labeled data in SD given limited budget. Hence, we consider a new, more realistic and more challenging problem setting, where classifiers have to be trained with noisy labeled data from SD and unlabeled data from TD---we name it wildly UDA (WUDA). We show that WUDA provably ruins all UDA methods if taking no care of label noise in SD, and to this end, we propose a Butterfly framework, a panacea for all difficulties in WUDA. Butterfly maintains four models (e.g., deep networks) simultaneously, where two take care of all adaptations (i.e., noisy-to-clean, labeled-to-unlabeled, and SD-to-TD-distributional) and then the other two can focus on classification in TD. As a consequence, Butterfly possesses all the necessary components for all the challenges in WUDA. Experiments demonstrate that under WUDA, Butterfly significantly outperforms existing baseline methods.
Unsupervised domain adaptation (UDA) trains with clean labeled data in source domain and unlabeled data in target domain to classify target-domain data. However, in real-world scenarios, it is hard to acquire fully-clean labeled data in source domain due to the expensive labeling cost. This brings us a new but practical adaptation called wildly-unsupervised domain adaptation (WUDA), which aims to transfer knowledge from noisy labeled data in source domain to unlabeled data in target domain. To tackle the WUDA, we present a robust one-step approach called Butterfly, which trains four networks. Specifically, two networks are jointly trained on noisy labeled data in source domain and pseudo-labeled data in target domain (i.e., data in mixture domain). Meanwhile, the other two networks are trained on pseudo-labeled data in target domain. By using dual-checking principle, Butterfly can obtain high-quality target-specific representations. We conduct experiments to demonstrate that Butterfly significantly outperforms other baselines on simulated and real-world WUDA tasks in most cases.
This paper uses the weather forecasting as an application background to illustrate the technique of \textit{deep uncertainty learning} (DUL). Weather forecasting has great significance throughout human history and is traditionally approached through numerical weather prediction (NWP) in which the atmosphere is modelled as differential equations. However, due to the instability of these differential equations in the presence of uncertainties, weather forecasting through numerical simulations may not be reliable. This paper explores weather forecasting as a data mining problem. We build a deep prediction interval (DPI) model based on sequence-to-sequence (seq2seq) that predicts spatio-temporal patterns of meteorological variables in the future 37 hours, which incorporates the informative knowledge of NWP. A big contribution and surprising finding in the training process of DPI is that training by mean variance error (MVE) loss instead of mean square error loss can significantly improve the generalization of point estimation, which has never been reported in previous researches. We think this phenomenon can be regarded as a new kind of regularization which can not only be on a par with the famous Dropout but also provide more uncertainty information, and hence comes into win-win situation. Based on single DPI, we then build deep ensemble. We evaluate our method on dataset from 10 realistic weather stations in Beijing of China. Experimental results shown DPI has better generalization than traditional point estimation and deep ensemble can further improve the performance. The deep ensemble method also achieved top-2 online score ranking in the competition of AI Challenger 2018. It can dramatically decrease up to 56\% error compared with NWP.
The cooperative hierarchical structure is a common and significant data structure observed in, or adopted by, many research areas, such as: text mining (author-paper-word) and multi-label classification (label-instance-feature). Renowned Bayesian approaches for cooperative hierarchical structure modeling are mostly based on topic models. However, these approaches suffer from a serious issue in that the number of hidden topics/factors needs to be fixed in advance and an inappropriate number may lead to overfitting or underfitting. One elegant way to resolve this issue is Bayesian nonparametric learning, but existing work in this area still cannot be applied to cooperative hierarchical structure modeling. In this paper, we propose a cooperative hierarchical Dirichlet process (CHDP) to fill this gap. Each node in a cooperative hierarchical structure is assigned a Dirichlet process to model its weights on the infinite hidden factors/topics. Together with measure inheritance from hierarchical Dirichlet process, two kinds of measure cooperation, i.e., superposition and maximization, are defined to capture the many-to-many relationships in the cooperative hierarchical structure. Furthermore, two constructive representations for CHDP, i.e., stick-breaking and international restaurant process, are designed to facilitate the model inference. Experiments on synthetic and real-world data with cooperative hierarchical structures demonstrate the properties and the ability of CHDP for cooperative hierarchical structure modeling and its potential for practical application scenarios.
Nonnegative Matrix Factorization (NMF) aims to factorize a matrix into two optimized nonnegative matrices appropriate for the intended applications. The method has been widely used for unsupervised learning tasks, including recommender systems (rating matrix of users by items) and document clustering (weighting matrix of papers by keywords). However, traditional NMF methods typically assume the number of latent factors (i.e., dimensionality of the loading matrices) to be fixed. This assumption makes them inflexible for many applications. In this paper, we propose a nonparametric NMF framework to mitigate this issue by using dependent Indian Buffet Processes (dIBP). In a nutshell, we apply a correlation function for the generation of two stick weights associated with each pair of columns of loading matrices, while still maintaining their respective marginal distribution specified by IBP. As a consequence, the generation of two loading matrices will be column-wise (indirectly) correlated. Under this same framework, two classes of correlation function are proposed (1) using Bivariate beta distribution and (2) using Copula function. Both methods allow us to adopt our work for various applications by flexibly choosing an appropriate parameter settings. Compared with the other state-of-the art approaches in this area, such as using Gaussian Process (GP)-based dIBP, our work is seen to be much more flexible in terms of allowing the two corresponding binary matrix columns to have greater variations in their non-zero entries. Our experiments on the real-world and synthetic datasets show that three proposed models perform well on the document clustering task comparing standard NMF without predefining the dimension for the factor matrices, and the Bivariate beta distribution-based and Copula-based models have better flexibility than the GP-based model.
Traditional Relational Topic Models provide a way to discover the hidden topics from a document network. Many theoretical and practical tasks, such as dimensional reduction, document clustering, link prediction, benefit from this revealed knowledge. However, existing relational topic models are based on an assumption that the number of hidden topics is known in advance, and this is impractical in many real-world applications. Therefore, in order to relax this assumption, we propose a nonparametric relational topic model in this paper. Instead of using fixed-dimensional probability distributions in its generative model, we use stochastic processes. Specifically, a gamma process is assigned to each document, which represents the topic interest of this document. Although this method provides an elegant solution, it brings additional challenges when mathematically modeling the inherent network structure of typical document network, i.e., two spatially closer documents tend to have more similar topics. Furthermore, we require that the topics are shared by all the documents. In order to resolve these challenges, we use a subsampling strategy to assign each document a different gamma process from the global gamma process, and the subsampling probabilities of documents are assigned with a Markov Random Field constraint that inherits the document network structure. Through the designed posterior inference algorithm, we can discover the hidden topics and its number simultaneously. Experimental results on both synthetic and real-world network datasets demonstrate the capabilities of learning the hidden topics and, more importantly, the number of topics.
Incorporating the side information of text corpus, i.e., authors, time stamps, and emotional tags, into the traditional text mining models has gained significant interests in the area of information retrieval, statistical natural language processing, and machine learning. One branch of these works is the so-called Author Topic Model (ATM), which incorporates the authors's interests as side information into the classical topic model. However, the existing ATM needs to predefine the number of topics, which is difficult and inappropriate in many real-world settings. In this paper, we propose an Infinite Author Topic (IAT) model to resolve this issue. Instead of assigning a discrete probability on fixed number of topics, we use a stochastic process to determine the number of topics from the data itself. To be specific, we extend a gamma-negative binomial process to three levels in order to capture the author-document-keyword hierarchical structure. Furthermore, each document is assigned a mixed gamma process that accounts for the multi-author's contribution towards this document. An efficient Gibbs sampling inference algorithm with each conditional distribution being closed-form is developed for the IAT model. Experiments on several real-world datasets show the capabilities of our IAT model to learn the hidden topics, authors' interests on these topics and the number of topics simultaneously.