We are interested in organizing a continuous stream of sparse and noisy texts, known as "tweets", in real time into an ontology of hundreds of topics with measurable and stringently high precision. This inference is performed over a full-scale stream of Twitter data, whose statistical distribution evolves rapidly over time. The implementation in an industrial setting with the potential of affecting and being visible to real users made it necessary to overcome a host of practical challenges. We present a spectrum of topic modeling techniques that contribute to a deployed system. These include non-topical tweet detection, automatic labeled data acquisition, evaluation with human computation, diagnostic and corrective learning and, most importantly, high-precision topic inference. The latter represents a novel two-stage training algorithm for tweet text classification and a close-loop inference mechanism for combining texts with additional sources of information. The resulting system achieves 93% precision at substantial overall coverage.

Online user engagement optimization is key to many Internet business. Several research areas are related to the concept of online user engagement optimization, including machine learning, data mining, information retrieval, recommender systems, online A/B (bucket) testing and psychology. In the past, research efforts in this direction are pursued in separate communities and conferences, yielding potential disconnected and repeated results. In addition, researchers and practitioners are sometimes only exposed to a specific aspect of the topic, which might be incomplete and suboptimal to the whole picture. Here, we organize the first workshop on the topic of online user engagement optimization, explicitly targeting the topic as a whole and bring researchers and practitioners together to foster the field. We invite two leading researchers from industry to give keynote talks about online machine learning and online experimentations. In addition, several invited talks from industry and academic researchers have covered the topics of content personalization, online experimental platforms and recommender systems. Also, six novel submissions are included as short papers in the workshop such that new results are discussed and shared among the workshop.

\emphDiffusion network inference and \emphmeme tracking have been two key challenges in viral diffusion. This paper shows that these two tasks can be addressed simultaneously with a probabilistic model involving a mixture of mutually exciting point processes. A fast learning algorithms is developed based on mean-field variational inference with budgeted diffusion bandwidth. The model is demonstrated with applications to the diffusion of viral texts in (1) online social networks (e.g., Twitter) and (2) the blogosphere on the Web.

Mobile devices provide people with a conduit to the rich infor-mation resources of the Web. With consent, the devices can also provide streams of information about search activity and location that can be used in population studies and real-time assistance. We analyzed geotagged mobile queries in a privacy-sensitive study of potential transitions from health information search to in-world healthcare utilization. We note differences in people's health infor-mation seeking before, during, and after the appearance of evidence that a medical facility has been visited. We find that we can accu-rately estimate statistics about such potential user engagement with healthcare providers. The findings highlight the promise of using geocoded search for sensing and predicting activities in the world.

We study the problem of labeling the edges of a social network graph (e.g., acquaintance connections in Facebook) as either positive (i.e., trust, true friendship) or negative (i.e., distrust, possible frenemy) relations. Such signed relations provide much stronger signal in tying the behavior of online users than the unipolar Homophily effect, yet are largely unavailable as most social graphs only contain unsigned edges.

Feature selection is fundamental to knowledge discovery from massive amount of high-dimensional data. In an effort to establish theoretical justification for feature selection algorithms, this paper presents a theoretically optimal criterion, namely, the discriminative optimal criterion (DoC) for feature selection. Compared with the existing representative optimal criterion (RoC, [CHECK END OF SENTENCE]) which retains maximum information for modeling the relationship between input and output variables, DoC is pragmatically advantageous because it attempts to directly maximize the classification accuracy and naturally reflects the Bayes error in the objective. To make DoC computationally tractable for practical tasks, we propose an algorithmic framework, which selects a subset of features by minimizing the Bayes error rate estimated by a nonparametric estimator. A set of existing algorithms as well as new ones can be derived naturally from this framework. As an example, we show that the Relief algorithm [CHECK END OF SENTENCE] greedily attempts to minimize the Bayes error estimated by the k-Nearest-Neighbor (kNN) method. This new interpretation insightfully reveals the secret behind the family of margin-based feature selection algorithms [CHECK END OF SENTENCE], [CHECK END OF SENTENCE] and also offers a principled way to establish new alternatives for performance improvement. In particular, by exploiting the proposed framework, we establish the Parzen-Relief (P-Relief) algorithm based on Parzen window estimator, and the MAP-Relief (M-Relief) which integrates label distribution into the max-margin objective to effectively handle imbalanced and multiclass data. Experiments on various benchmark data sets demonstrate the effectiveness of the proposed algorithms.

While a user's preference is directly reflected in the interactive choice process between her and the recommender, this wealth of information was not fully exploited for learning recommender models. In particular, existing collaborative filtering (CF) approaches take into account only the binary events of user actions but totally disregard the contexts in which users' decisions are made. In this paper, we propose Collaborative Competitive Filtering (CCF), a framework for learning user preferences by modeling the choice process in recommender systems. CCF employs a multiplicative latent factor model to characterize the dyadic utility function. But unlike CF, CCF models the user behavior of choices by encoding a local competition effect. In this way, CCF allows us to leverage dyadic data that was previously lumped together with missing data in existing CF models. We present two formulations and an efficient large scale optimization algorithm. Experiments on three real-world recommendation data sets demonstrate that CCF significantly outperforms standard CF approaches in both offline and online evaluations.

A key challenge in recommender system research is how to effectively profile new users, a problem generally known as cold-start recommendation. Recently the idea of progressively querying user responses through an initial interview process has been proposed as a useful new user preference elicitation strategy. In this paper, we present functional matrix factorization (fMF), a novel cold-start recommendation method that solves the problem of initial interview construction within the context of learning user and item profiles. Specifically, fMF constructs a decision tree for the initial interview with each node being an interview question, enabling the recommender to query a user adaptively according to her prior responses. More importantly, we associate latent profiles for each node of the tree --- in effect restricting the latent profiles to be a function of possible answers to the interview questions --- which allows the profiles to be gradually refined through the interview process based on user responses. We develop an iterative optimization algorithm that alternates between decision tree construction and latent profiles extraction as well as a regularization scheme that takes into account of the tree structure. Experimental results on three benchmark recommendation data sets demonstrate that the proposed fMF algorithm significantly outperforms existing methods for cold-start recommendation.

Targeting interest to match a user with services (e.g. news, products, games, advertisements) and predicting friendship to build connections among users are two fundamental tasks for social network systems. In this paper, we show that the information contained in interest networks (i.e. user-service interactions) and friendship networks (i.e. user-user connections) is highly correlated and mutually helpful. We propose a framework that exploits homophily to establish an integrated network linking a user to interested services and connecting different users with common interests, upon which both friendship and interests could be efficiently propagated. The proposed friendship-interest propagation (FIP) framework devises a factor-based random walk model to explain friendship connections, and simultaneously it uses a coupled latent factor model to uncover interest interactions. We discuss the flexibility of the framework in the choices of loss objectives and regularization penalties and benchmark different variants on the Yahoo! Pulse social networking system. Experiments demonstrate that by coupling friendship with interest, FIP achieves much higher performance on both interest targeting and friendship prediction than systems using only one source of information.

Learning to rank arises in many information retrieval applications, ranging from Web search engine, online advertising to recommendation system. In learning to rank, the performance of a ranking function heavily depends on the number of labeled examples in the training set; on the other hand, obtaining labeled examples for training data is very expensive and time-consuming. This presents a great need for making use of available auxiliary data, i.e., the within-domain unlabeled data and the out-of-domain labeled data. In this paper, we propose a general framework for ranking with auxiliary data, which is applicable to various ranking applications. Under this framework, we derive a generic ranking algorithm to effectively make use of both the within-domain unlabeled data and the out-of-domain labeled data. The proposed algorithm iteratively learns ranking functions for target domain and source domains and enforces their consensus on the unlabeled data in the target domain.

The classical Bag-of-Word (BOW) model represents a document as a histogram of word occurrence, losing the spatial information that is invaluable for many text analysis tasks. In this paper, we present the Language Pyramid (LaP) model, which casts a document as a probabilistic distribution over the joint semantic-spatial space and motivates a multi-scale 2D local smoothing framework for nonparametric text coding. LaP efficiently encodes both semantic and spatial contents of a document into a pyramid of matrices that are smoothed both semantically and spatially at a sequence of resolutions, providing a convenient multi-scale imagic view for natural language understanding. The LaP representation can be used in text analysis in a variety of ways, among which we investigate two instantiations in the current paper: (1) multi-scale text kernels for document categorization, and (2) multi-scale language models for ad hoc text retrieval. Experimental results illustrate that: for classification, LaP outperforms BOW by (up to) 4% on moderate-length texts (RCV1 text benchmark) and 15% on short texts (Yahoo! queries); and for retrieval, LaP gains 12% MAP improvement over uni-gram language models on the OHSUMED data set.

This paper addresses Named Entity Mining (NEM), in which we mine knowledge about named entities such as movies, games, and books from a huge amount of data. NEM is potentially useful in many applications including web search, online advertisement, and recommender system. There are three challenges for the task: finding suitable data source, coping with the ambiguities of named entity classes, and incorporating necessary human supervision into the mining process. This paper proposes conducting NEM by using click-through data collected at a web search engine, employing a topic model that generates the click-through data, and learning the topic model by weak supervision from humans. Specifically, it characterizes each named entity by its associated queries and URLs in the click-through data. It uses the topic model to resolve ambiguities of named entity classes by representing the classes as topics. It employs a method, referred to as Weakly Supervised Latent Dirichlet Allocation (WS-LDA), to accurately learn the topic model with partially labeled named entities. Experiments on a large scale click-through data containing over 1.5 billion query-URL pairs show that the proposed approach can conduct very accurate NEM and significantly outperforms the baseline.

We propose Dirichlet-Bernoulli Alignment (DBA), a generative model for corpora in which each pattern (e.g., a document) contains a set of instances (e.g., paragraphs in the document) and belongs to multiple classes. By casting predefined classes as latent Dirichlet variables (i.e., instance level labels), and modeling the multi-label of each pattern as Bernoulli variables conditioned on the weighted empirical average of topic assignments, DBA automatically aligns the latent topics discovered from data to human-defined classes. DBA is useful for both pattern classification and instance disambiguation, which are tested on text classification and named entity disambiguation for web search queries respectively.

This paper presents a stagewise least square (SLS) loss function for classification. It uses a least square form within each stage to approximate a bounded monotonic nonconvex loss function in a stagewise manner. Several benefits are obtained from using the SLS loss function, such as: (i) higher generalization accuracy and better scalability than classical least square loss; (ii) improved performance and robustness than convex loss (e.g., hinge loss of SVM); (iii) computational advantages compared with nonconvex loss (e.g. ramp loss in ψ-learning); (iv) ability to resist myopia of Empirical Risk Minimization and to boost the margin without boosting the complexity of the classifier. In addition, it naturally results in a kernel machine which is as sparse as SVM, yet much faster and simpler to train. A fast online learning algorithm with an integrated sparsification procedure is also provided. Experimental results on several benchmarks confirm the advantages of the proposed approach.