Search

This paper concerns a deep learning approach to relevance ranking in information retrieval (IR). Existing deep IR models such as DSSM and CDSSM directly apply neural networks to generate ranking scores, without explicit understandings of the relevance. According to the human judgement process, a relevance label is generated by the following three steps: 1) relevant locations are detected, 2) local relevances are determined, 3) local relevances are aggregated to output the relevance label. In this paper we propose a new deep learning architecture, namely DeepRank, to simulate the above human judgment process. Firstly, a detection strategy is designed to extract the relevant contexts. Then, a measure network is applied to determine the local relevances by utilizing a convolutional neural network (CNN) or two-dimensional gated recurrent units (2D-GRU). Finally, an aggregation network with sequential integration and term gating mechanism is used to produce a global relevance score. DeepRank well captures important IR characteristics, including exact/semantic matching signals, proximity heuristics, query term importance, and diverse relevance requirement. Experiments on both benchmark LETOR dataset and a large scale clickthrough data show that DeepRank can significantly outperform learning to ranking methods, and existing deep learning methods., Comment: Published as a conference paper at CIKM 2017, CIKM'17, November 6--10, 2017, Singapore TextNet (https://github.com/pl8787/textnet-release) PyTorch (https://github.com/pl8787/DeepRank_PyTorch)

Rare efforts have been devoted to generating the structured Navigation Box (Navbox) for Wikipedia articles. A Navbox is a table in Wikipedia article page that provides a consistent navigation system for related entities. Navbox is critical for the readership and editing efficiency of Wikipedia. In this paper, we target on the automatic generation of Navbox for Wikipedia articles. Instead of performing information extraction over unstructured natural language text directly, an alternative avenue is explored by focusing on a rich set of semi-structured data in Wikipedia articles: linked entities. The core idea of this paper is as follows: If we cluster the linked entities and interpret them appropriately, we can construct a high-quality Navbox for the article entity. We propose a clustering-then-labeling algorithm to realize the idea. Experiments show that the proposed solutions are effective. Ultimately, our approach enriches Wikipedia with 1.95 million new Navboxes of high quality.

This paper attempts to conduct analysis for one certain type of user reviews; that is, the reviews on a super-entity (e.g., restaurant) involve descriptions for many sub-entities (e.g., dishes). To deal with such analysis, we propose a text embedding framework for ranking sub-entities from user reviews of a given super-entity. Experiments on two real-world datasets show that our method outperforms three baselines by a statistically significant amount. Intriguing cases from the experiments are discussed in the paper.

In this paper we focus on the problem of optimizing email volume for maximizing sitewide engagement of an online social networking service. Email volume optimization approaches published in the past have proposed optimization of email volume for maximization of engagement metrics which are impacted exclusively by email; for example, the number of sessions that begin with clicks on links within emails. The impact of email on such downstream engagement metrics can be estimated easily because of the ease of attribution of such an engagement event to an email. However, this framework is limited in its view of the ecosystem of the networking service which comprises of several tools and utilities that contribute towards delivering value to members; with email being just one such utility. Thus, in this paper we depart from previous approaches by exploring and optimizing the contribution of email to this ecosystem. In particular, we present and contrast the differential impact of email on sitewide engagement metrics for various types of users. We propose a new email volume optimization approach which maximizes sitewide engagement metrics, such as the total number of active users. This is in sharp contrast to the previous approaches whose objective has been maximization of downstream engagement metrics. We present details of our prediction function for predicting the impact of emails on a user's activeness on the mobile or web application. We describe how certain approximations to this prediction function can be made for solving the volume optimization problem, and present results from online A/B tests.

In this paper, we presented a new study for Web query entity disambiguation (QED), which is the task of disambiguating different candidate entities in a knowledge base given their mentions in a query. QED is particularly challenging because queries are often too short to provide rich contextual information that is required by traditional entity disambiguation methods. In this paper, we propose several methods to tackle the problem of QED. First, we explore the use of deep neural network (DNN) for capturing the character level textual information in queries. Our DNN approach maps queries and their candidate reference entities to feature vectors in a latent semantic space where the distance between a query and its correct reference entity is minimized. Second, we utilize the Web search result information of queries to help generate large amounts of weakly supervised training data for the DNN model. Third, we propose a two-stage training method to combine large-scale weakly supervised data with a small amount of human labeled data, which can significantly boost the performance of a DNN model. The effectiveness of our approach is demonstrated in the experiments using large-scale real-world datasets.

Display Advertising has generated billions of revenue and originated hundreds of scientific papers and patents, yet the accuracy of prediction technologies leaves much to be desired. Conversion rates (CVR) predictions can often be formulated as a matrix or tensor completion problem where each dimension consists of thousands or even hundreds of thousands of levels. Observed entries are typically extremely sparse, comprising only 0.01% to 1% of the entire matrix or tensor with highly unevenly distributed conversion as well as impression sizes. To deal with these issues, we propose an extension of matrix factorization, namely Bayesian Heteroscedastic Matrix Factorization (BHMF), with three key features. First, BHMF accounts for the fact that each observed entry of a matrix has different magnitude of errors depending on the corresponding impression sizes. We extend the previous research on empirical instance-wise weighted matrix factorization with rigorous probabilistic modelling framework. Second, BHMF is amenable to an efficient Bayesian inference algorithm that is scalable to high dimensional matrices. Compared to the optimization based training, it is more robust to the choices of dimensions of the latent factors as well as regularization parameters. Last, the Bayesian approach provides predictive uncertainty estimations for unseen entries that is capable of dealing with cold-start problems. This can potentially affect a good amount of revenue in the real time bidding (RTB) environment. We focus on matrix CVR predictions in this paper but the proposed BHMF can be naturally extended and applied to higher dimensional tensors. We demonstrate the substantial improvement of our model in predictive capabilities on Yahoo! demand side platform (DSP) BrightRoll.

One key challenge in talent search is to translate complex criteria of a hiring position into a search query, while it is relatively easy for a searcher to list examples of suitable candidates for a given position. To improve search efficiency, we propose the next generation of talent search at LinkedIn, also referred to as Search By Ideal Candidates. In this system, a searcher provides one or several ideal candidates as the input to hire for a given position. The system then generates a query based on the ideal candidates and uses it to retrieve and rank results. Shifting from the traditional Query-By-Keyword to this new Query-By-Example system poses a number of challenges: How to generate a query that best describes the candidates? When moving to a completely different paradigm, how does one leverage previous product logs to learn ranking models and/or evaluate the new system with no existing usage logs? Finally, given the different nature between the two search paradigms, the ranking features typically used for Query-By-Keyword systems might not be optimal for Query-By-Example. This paper describes our approach to solving these challenges. We present experimental results confirming the effectiveness of the proposed solution, particularly on query building and search ranking tasks. As of writing this paper, the new system has been available to all LinkedIn members.

Employees in companies can be divided into di erent communities, and those who frequently socialize with each other will be treated as close friends and are grouped in the same community. In the enterprise context, a large amount of information about the employees is available in both (1) o ine company internal sources and (2) online enterprise social networks (ESNs). Each of the information sources also contain multiple categories of employees' socialization activities at the same time. In this paper, we propose to detect the social communities of the employees in companies based on the broad learning se ing with both these online and o ine information sources simultaneously, and the problem is formally called the "Broad Learning based Enterprise Community Detection" (BL-Ecd) problem. To address the problem, a novel broad learning based community detection framework named "HeterogeneoUs Multi-sOurce ClusteRing" (Humor) is introduced in this paper. Based on the various enterprise social intimacy measures introduced in this paper, Humor detects a set of micro community structures of the employees based on each of the socialization activities respectively. To obtain the (globally) consistent community structure of employees in the company, Humor further fuses these micro community structures via two broad learning phases: (1) intra-fusion of micro community structures to obtain the online and o ine (locally) consistent communities respectively, and (2) inter-fusion of the online and o ine communities to achieve the (globally) consistent community structure of employees. Extensive experiments conducted on real-world enterprise datasets demonstrate our method can perform very well in addressing the BL-Ecd problem., 10 Pages, 12 Figures. Full paper has been accepted by CIKM 2017,In: Proceedings of the 2017 International Conference on Information and Knowledge Management

This paper addresses the task of cross-domain social emotion classification of online documents. The cross-domain task is formulated as using abundant labeled documents from a source domain and a small amount of labeled documents from a target domain, to predict the emotion of unlabeled documents in the target domain. Although several cross-domain emotion classification algorithms have been proposed, they require that feature distributions of different domains share a sufficient overlapping, which is hard to meet in practical applications. This paper proposes a novel framework, which uses the emotion distribution of training documents at the cluster level, to alleviate the aforementioned issue. Experimental results on two datasets show the effectiveness of our proposed model on cross-domain social emotion classification.

CNN (Convolution Neural Network) is widely used in visual analysis and achieves exceptionally high performances in image classification, face detection, object recognition, image recoloring, and other learning jobs. Using deep learning frameworks, such as Torch and Tensorflow, CNN can be efficiently computed by leveraging the power of GPU. However, one drawback of GPU is its limited memory which prohibits us from handling large images. Passing a 4K resolution image to the VGG network will result in an exception of out-of-memory for Titan-X GPU. In this paper, we propose a new approach that adopts the BSP (bulk synchronization parallel) model to compute CNNs for images of any size. Before fed to a specific CNN layer, the image is split into smaller pieces which go through the neural network separately. Then, a specific padding and normalization technique is adopted to merge sub-images back into one image. Our approach can be easily extended to support distributed multi-GPUs. In this paper, we use neural style network as our example to illustrate the effectiveness of our approach. We show that using one Titan-X GPU, we can transfer the style of an image with 10,000×10,000 pixels within 1 minute.

Network alignment is becoming an active topic in network data analysis. Despite extensive research, we realize that efficient use of topological and attribute information for large attributed network alignment has not been sufficiently addressed in previous studies. In this paper, based on Stochastic Block Model (SBM) and Dirichlet-multinomial, we propose "divide-and-conquer" models CAlign that jointly consider network alignment, community discovery and community alignment in one framework for large networks with node attributes, in an effort to reduce both the computation time and memory usage while achieving better or competitive performance. It is provable that the algorithms derived from our model have sub-quadratic time complexity and linear space complexity on a network with small densification power, which is true for most real-world networks. Experiments show CAlign is superior to two recent state-of-art models in terms of accuracy, time and memory on large networks, and CAlign is capable of handling millions of nodes on a modern desktop machine.

Pointwise and pairwise collaborative ranking are two major classes of algorithms for personalized item ranking. This paper proposes a novel joint learning method named alternating pointwise-pairwise learning (APPL) to improve ranking performance. APPL combines the ideas of both pointwise and pairwise learning, and is able to produce a more effective prediction model. The extensive experiments with both explicit and implicit feedback settings on four real-world datasets demonstrate that APPL performs significantly better than the state-of-the-art methods.

In this paper, we study an important yet largely under-explored setting of graph embedding, i.e., embedding communities instead of each individual nodes. We find that community embedding is not only useful for community-level applications such as graph visualization, but also beneficial to both community detection and node classification. To learn such embedding, our insight hinges upon a closed loop among community embedding, community detection and node embedding. On the one hand, node embedding can help improve community detection, which outputs good communities for fitting better community embedding. On the other hand, community embedding can be used to optimize the node embedding by introducing a community-aware high-order proximity. Guided by this insight, we propose a novel community embedding framework that jointly solves the three tasks together. We evaluate such a framework on multiple real-world datasets, and show that it improves graph visualization and outperforms state-of-the-art baselines in various application tasks, e.g., community detection and node classification.

Information gathered from multiple sources on the Web often exhibits conflicts. This phenomenon motivates the need of truth discovery, which aims to automatically find the true claim among multiple conflicting claims. Existing truth discovery methods are mainly based on iterative updates or probabilistic models. In particular, iterative methods specify rules that govern how credibility flows from sources to claims and then back to sources. However, these manually-defined rules tend to be ad hoc and are difficult to adapt and analyze. Probabilistic methods model a few latent factors that impact how sources make claims, such as randomly choosing, guessing, or mistaking. However, these manually-defined factors may not well reflect the underlying data distributions. Given these limitations, we propose a new, unsupervised model for truth discovery in this paper. Our model first constructs a heterogenous network that exploits both source-claim and source-source relationships. It then embeds the network into a low dimensional space through a principled algorithm such that trustworthy sources and true claims (meanwhile, unreliable sources and false claims) are close. In this way, truth discovery can be conveniently performed in the embedding space. Compared with existing methods, our model does not need manually-defined rules or factors. Rather, it learns the embeddings automatically from data. Experiments on two real-world datasets demonstrate that our model outperforms existing state-of-the-art methods for truth discovery.

A graph database D is a collection of graphs. To speed up subgraph query answering on graph databases, indexes are commonly used. State-of-the-art graph database indexes do not adapt or scale well to dynamic graph database use; they are static, and their ability to prune possible search responses to meet user needs worsens over time as databases change and grow. Users can re-mine indexes to gain some improvement, but it is time consuming. Users must also tune numerous parameters on an ongoing basis to optimize performance and can inadvertently worsen the query response time if they do not choose parameters wisely. Recently, a one-pass algorithm has been developed to enhance the performance of frequent subgraphs based indexes by using the algorithm to update them regularly. However, there are some drawbacks, most notably the need to make updates as the query workload changes. In this paper, we propose a new index based on graph-coarsening to speed up subgraph query answering time in dynamic graph databases. Our index is parameter-free, query-independent, scalable,small enough to store in the main memory, and is simpler and less costly to maintain for database updates. Experimental results show that our index outperforms hybrid-indexes (i.e. indexes updated with one-pass) for query answering time in the case of social network databases, and is comparable with these indexes for frequent and infrequent queries on chemical databases. Our index can be updated up to 60 times faster in comparison to one-pass on dynamic graph databases. Moreover, our index is independent of the query workload for index update and is up to 15 times faster after hybrid-indexes are attuned to query workload.

Entity resolution (ER) is the task of identifying different representations of the same real-world object across datasets. Designing and tuning ER algorithms is an error-prone, labor-intensive process, which can significantly benefit from data-driven, automated learning methods. Our focus is on "big data'' scenarios where the primary challenges include 1) identifying, out of a potentially massive set, a small subset of informative examples to be labeled by the user, 2) using the labeled examples to efficiently learn ER algorithms that achieve both high precision and high recall, and 3) executing the learned algorithm to determine duplicates at scale. Recent work on learning ER algorithms has employed active learning to partially address the above challenges by aiming to learn ER rules in the form of conjunctions of matching predicates, under precision guarantees. While successful in learning a single rule, prior work has been less successful in learning multiple rules that are sufficiently different from each other, thus missing opportunities for improving recall. In this paper, we introduce an active learning system that learns, at scale, multiple rules each having significant coverage of the space of duplicates, thus leading to high recall, in addition to high-precision. We show the superiority of our system on real-world ER scenarios of sizes up to tens of millions of records, over state-of-the-art active learning methods that learn either rules or committees of statistical classifiers for ER, and even over sophisticated methods based on first-order probabilistic models.

Prior work on using retrievability measures in the evaluation of information retrieval (IR) systems has laid out the foundations for investigating the relation between retrieval performance and retrieval bias. While various factors influencing retrievability have been examined, showing how the retrieval model may influence bias, no prior work has examined the impact of the index (and how it is optimized) on retrieval bias. Intuitively, how the documents are represented, and what terms they contain, will influence whether they are retrievable or not. In this paper, we investigate how the retrieval bias of a system changes as the inverted index is optimized for efficiency through static index pruning. In our analysis, we consider four pruning methods and examine how they affect performance and bias on the TREC GOV2 Collection. Our results show that the relationship between these factors is varied and complex - and very much dependent on the pruning algorithm. We find that more pruning results in relatively little change or a slight decrease in bias up to a point, and then a dramatic increase. The increase in bias corresponds to a sharp decrease in early precision such as NDCG@10 and is also indicative of a large decrease in MAP. The findings suggest that the impact of pruning algorithms can be quite varied - but retrieval bias could be used to guide the pruning process. Further work is required to determine precisely which documents are most affected and how this impacts upon performance.

Speaker change detection (SCD) is an important task in dialog modeling. Our paper addresses the problem of text-based SCD, which differs from existing audio-based studies and is useful in various scenarios, for example, processing dialog transcripts where speaker identities are missing (e.g., OpenSubtitle), and enhancing audio SCD with textual information. We formulate text-based SCD as a matching problem of utterances before and after a certain decision point; we propose a hierarchical recurrent neural network (RNN) with static sentence-level attention. Experimental results show that neural networks consistently achieve better performance than feature-based approaches, and that our attention-based model significantly outperforms non-attention neural networks., Comment: In Proceedings of the ACM on Conference on Information and Knowledge Management (CIKM), 2017

With the development of large-scale knowledge bases, people are building systems which give simple answers to questions based on consolidate facts. In this paper, we focus on simple questions, which ask about only a subject and relation in the knowledge base. Observing that certain parts of a question usually overlap with names of its corresponding subject and relation in the knowledge base, we argue that a question is formed by a mixture of copying and generation. To model that, we propose a sequence-to-sequence (seq2seq) architecture which encodes a candidate subject-relation pair and decodes it into the given question, where the decoding probability is used to select the best candidate. In our decoder, the copying mode points the subject or relation and duplicates its name, while the generating mode summarizes the meaning of the subject-relation pair and produces a word to smooth the question. Realizing that although sometimes a subject or relation is pointed, different names or keywords might be used, we also incorporate a paraphrasing mode to supplement the copying mode using an automatically mined lexicon. Extensive experiments on the largest dataset exhibit our better performance compared with the state-of-the-art methods.

Given a large graph, the densest-subgraph problem asks to find a subgraph with maximum average degree. When considering the top-$k$ version of this problem, a na\"ive solution is to iteratively find the densest subgraph and remove it in each iteration. However, such a solution is impractical due to high processing cost. The problem is further complicated when dealing with dynamic graphs, since adding or removing an edge requires re-running the algorithm. In this paper, we study the top-$k$ densest-subgraph problem in the sliding-window model and propose an efficient fully-dynamic algorithm. The input of our algorithm consists of an edge stream, and the goal is to find the node-disjoint subgraphs that maximize the sum of their densities. In contrast to existing state-of-the-art solutions that require iterating over the entire graph upon any update, our algorithm profits from the observation that updates only affect a limited region of the graph. Therefore, the top-$k$ densest subgraphs are maintained by only applying local updates. We provide a theoretical analysis of the proposed algorithm and show empirically that the algorithm often generates denser subgraphs than state-of-the-art competitors. Experiments show an improvement in efficiency of up to five orders of magnitude compared to state-of-the-art solutions., Comment: 10 pages, 8 figures, accepted at CIKM 2017

The problem of recommending bibliographic citations to an author who is writing an article has been well-studied. However, different researchers have used different datasets to evaluate proposed techniques, and have sometimes reported contradictory findings regarding the relative effectiveness of various approaches. In addition, these datasets are problematic in one way or another (e.g., in terms of size or availability), precluding the possibility of adopting one (or some) of them as standard benchmarks. A recently created test collection that makes use of data from CiteSeerx is large, heterogenous, and publicly available, but has certain other limitations. In this paper, we propose a way to modify this test collection to address these limitations. We also use the improved test collection to establish a set of baseline results using elementary content-based techniques, as well as reference directed indexing.

Today's video streaming market is crowded with various content providers (CPs). For individual CPs, understanding user behavior, in particular how users migrate among different CPs, is crucial for improving users' on-site experience and the CP's chance of success. In this paper, we take a data-driven approach to analyze and model user migration behavior in video streaming, i.e., users switching content provider during active sessions. Based on a large ISP dataset over two months (6 major content providers, 3.8 million users, and 315 million video requests), we study common migration patterns and reasons of migration. We find that migratory behavior is prevalent: 66% of users switch CPs with an average switching frequency of 13%. In addition, migration behaviors are highly diverse: regardless large or small CPs, they all have dedicated groups of users who like to switch to them for certain types of videos. Regarding reasons of migration, we find CP service quality rarely causes migration, while a few popular videos play a bigger role. Nearly 60% of cross-site migrations are landed to 0.14% top videos. Finally, we validate our findings by building an accurate regression model to predict user migration frequency, and discuss the implications of our results to CPs.

Network Centrality is one of the core concepts in network analysis, which ranks the importance of a node in a network. A considerably extensive range of centrality measures exist that serve the purpose of quantifying the importance of a node according to its application and domain. One such measure is the Betweenness Centrality (BC) which computes the importance of a node in terms of total number of shortest paths that pass through that node. However, these computations are very expensive and pose different challenges for large scale networks. With an attempt to deal with these challenges, our paper presents an approximate algorithm for BC maximization problem, which tries to find a set of nodes with largest BC. The core of our algorithm is the estimation technique, which is based on progressive sampling with early stopping conditions. The reduction in sample size results not only in small computations overhead, but also scales well with large networks. We experimentally evaluate our technique using different datasets to confirm the performance of the developed techniques.

Advanced processor architectures have been driving new designs, implementations and optimizations of main-memory hash join algorithms recently. The newly released Intel Xeon Phi many-core processor of the Knights Landing architecture (KNL) embraces interesting hardware features such as many low-frequency out-of-order cores connected on a 2D mesh, and high-bandwidth multi-channel memory (MCDRAM). In this paper, we experimentally revisit the state-of-the-art main-memory hash join algorithms to study how the new hardware features of KNL affect the algorithmic design and tuning as well as to identify the opportunities for further performance improvement on KNL. Our experiments show that, although many existing optimizations are still valid on KNL with proper tuning, even the state-of-the-art algorithms have severely underutilized the memory bandwidth and other hardware resources.

Promotional listing of products or advertisements is a major source of revenue for online retail companies. These advertisements are often sold in the guaranteed delivery market, serving of which critically depends on the ability to predict supply or potential impressions from a target segment of users. In this paper, we study the problem of predicting user visits or potential ad-impressions to online retail websites, based on historical time-stamps. We explore the time-series and temporal point process models. We find that a successful model must encompass three properties of the data: (1) temporally non-homgeneous rates, (2) self excitation and (3) handling special events. We propose a novel non-homogeneous Hawkes process based model for the same, and new algorithm for fitting this model without overfitting the self-excitation part. We validate the proposed model and algorithm using mulitple large scale ad-serving dataset from a top online retail company in India.

Information Extraction by Text Segmentation (IETS) aims at segmenting text inputs to extract implicit data values contained in them.The state-of-art IETS approaches mainly rely on machine learning techniques, either supervised or unsupervised.However, while the supervised approaches require a large labelled training data, the performance of the unsupervised ones could be unstable on different data sets.To overcome their weaknesses, this paper introduces CNN-IETS, a novel unsupervised probabilistic approach that takes the advantages of pre-existing data and a Convolution Neural Network (CNN)-based probabilistic classification model. While using the CNN model can ease the burden of selecting high-quality features in associating text segments with attributes of a given domain, the pre-existing data as a domain knowledge base can provide training data with a comprehensive list of features for building the CNN model.Given an input text, we do initial segmentation (according to the occurrences of these words in the knowledge base) to generate text segments for CNN classification with probabilities. Then, based on the probabilistic CNN classification results, we work on finding the most probable labelling way to the whole input text.As a complementary, a bidirectional sequencing model learned on-demand from test data is finally deployed to do further adjustment to some problematic labelled segments.Our experimental study conducted on several real data collections shows that CNN-IETS improves the extraction quality of state-of-art approaches by more than 10%.

Given e-commerce scenarios that user profiles are invisible, session-based recommendation is proposed to generate recommendation results from short sessions. Previous work only considers the user's sequential behavior in the current session, whereas the user's main purpose in the current session is not emphasized. In this paper, we propose a novel neural networks framework, i.e., Neural Attentive Recommendation Machine (NARM), to tackle this problem. Specifically, we explore a hybrid encoder with an attention mechanism to model the user's sequential behavior and capture the user's main purpose in the current session, which are combined as a unified session representation later. We then compute the recommendation scores for each candidate item with a bi-linear matching scheme based on this unified session representation. We train NARM by jointly learning the item and session representations as well as their matchings. We carried out extensive experiments on two benchmark datasets. Our experimental results show that NARM outperforms state-of-the-art baselines on both datasets. Furthermore, we also find that NARM achieves a significant improvement on long sessions, which demonstrates its advantages in modeling the user's sequential behavior and main purpose simultaneously.

Most of the existing recommender systems assume that user's visiting history can be constantly recorded. However, in recent online services, the user identification may be usually unknown and only limited online user behaviors can be used. It is of great importance to model the temporal online user behaviors and conduct recommendation for the anonymous users. In this paper, we propose a list-wise deep neural network based architecture to model the limited user behaviors within each session. To train the model efficiently, we first design a session embedding method to pre-train a session representation, which incorporates different kinds of user search behaviors such as clicks and views. Based on the learnt session representation, we further propose a list-wise ranking model to generate the recommendation result for each anonymous user session. We conduct quantitative experiments on a recently published dataset from an e-commerce company. The evaluation results validate the effectiveness of the proposed method, which can outperform the state-of-the-art significantly.

Answer selection for question answering is a challenging task, since it requires effective capture of the complex semantic relations between questions and answers. Previous remarkable approaches mainly adopt general Compare-Aggregate framework that performs word-level comparison and aggregation. In this paper, unlike previous Compare-Aggregate models which utilize the traditional attention mechanism to generate corresponding word-level vector before comparison, we propose a novel attention mechanism named Dynamic-Clip Attention which is directly integrated into the Compare-Aggregate framework. Dynamic-Clip Attention focuses on filtering out noise in attention matrix, in order to better mine the semantic relevance of word-level vectors. At the same time, different from previous Compare-Aggregate works which treat answer selection task as a pointwise classification problem, we propose a listwise ranking approach to model this task to learn the relative order of candidate answers. Experiments on TrecQA and WikiQA datasets show that our proposed model achieves the state-of-the-art performance.

An effective way to visualize the prediction of deep neural networks on an image is to decompose the prediction into the contribution of units (pixels or patches). In the existing works, these units are largely considered independently, thus limiting the performance of visualization. In this paper, we propose a new predication visualization method that uses super-pixel as a contribution unit. Moreover, our method takes into consideration of the interaction of adjacent super-pixels. We implement our technique and evaluate its performance with various images. Our results show its excellent performance.

Online social media remarkably facilitates the production and delivery of information, intensifying the competition among vast information for users' attention and highlighting the importance of predicting the popularity of information. Existing approaches for popularity prediction fall into two paradigms: feature-based approaches and generative approaches. Feature-based approaches extract various features (e.g., user, content, structural, and temporal features), and predict the future popularity of information by training a regression/classification model. Their predictive performance heavily depends on the quality of hand-crafted features. In contrast, generative approaches devote to characterizing and modeling the process that a piece of information accrues attentions, offering us high ease to understand the underlying mechanisms governing the popularity dynamics of information cascades. But they have less desirable predictive power since they are not optimized for popularity prediction. In this paper, we propose DeepHawkes to combat the defects of existing methods, leveraging end-to-end deep learning to make an analogy to interpretable factors of Hawkes process --- a widely-used generative process to model information cascade. DeepHawkes inherits the high interpretability of Hawkes process and possesses the high predictive power of deep learning methods, bridging the gap between prediction and understanding of information cascades. We verify the effectiveness of DeepHawkes by applying it to predict retweet cascades of Sina Weibo and citation cascades of a longitudinal citation dataset. Experimental results demonstrate that DeepHawkes outperforms both feature-based and generative approaches.

In this paper we study skyline queries in the distributed computational model, where we have s remote sites and a central coordinator; each site holds a piece of data, and the coordinator wants to compute the skyline of the union of the s datasets. The computation is in terms of rounds, and the goal is to minimize both the total communication cost and the round cost. We first give an algorithm with a small communication cost but potentially a large round cost; we show information-theoretically that the communication cost is optimal even if we allow an infinite number of communication rounds. We next give algorithms with smooth communication-round tradeoffs. We also show a strong lower bound for the communication cost if we can only use one round of communication. Finally, we demonstrate the superiority of our algorithms over existing ones by an extensive set of experiments on both synthetic and real world datasets.

Recent AI research has witnessed increasing interests in automatically generating image descriptions in text, which is coined as theimage captioning problem. Significant progresses have been made in domains where plenty of labeled training data (i.e. image-text pairs) are readily available or collected. However, obtaining rich annotated data is a time-consuming and expensive process, creating a substantial barrier for applying image captioning methods to a new domain. In this paper, we propose a cross-domain image captioning approach that uses a novel dual learning mechanism to overcome this barrier. First, we model the alignment between the neural representations of images and that of natural languages in the source domain where one can access sufficient labeled data. Second, we adjust the pre-trained model based on examining limited data (or unpaired data) in the target domain. In particular, we introduce a dual learning mechanism with a policy gradient method that generates highly rewarded captions. The mechanism simultaneously optimizes two coupled objectives: generating image descriptions in text and generating plausible images from text descriptions, with the hope that by explicitly exploiting their coupled relation, one can safeguard the performance of image captioning in the target domain. To verify the effectiveness of our model, we use MSCOCO dataset as the source domain and two other datasets (Oxford-102 and Flickr30k) as the target domains. The experimental results show that our model consistently outperforms previous methods for cross-domain image captioning.

In recent years, while deep neural networks have shown impressive performance to solve various recognition and classification problems, collaborative filtering (CF) received relatively little attention to utilize deep neural networks. Because of inherent data sparsity, it remains a challenging problem for deep neural networks. In this paper, we propose a new CF model, namely the imputation-boosted denoising autoencoder (IDAE), for top-N recommendation. Specifically, IDAE consists of two steps: imputing positive values and learning with imputed values. First, it infers and imputes positive user feedback from missing values. Then, the correlation between items is learned by using the denoising autoencoder (DAE) with imputed values. Unlike the existing DAE that randomly corrupts the input, the key characteristic of IDAE is that original user values are taken as the input, and imputed values are reflected as the corrupted output. Our experimental results demonstrate that IDAE significantly outperforms state-of-the-art CF algorithms using autoencoders (by up to 5%) on the MovieLens datasets.

In recent years, we have witnessed a rapid increase of text con- tent stored in digital archives such as newspaper archives or web archives. Many old documents have been converted to digital form and made accessible online. Due to the passage of time, it is however difficult to effectively perform search within such collections. Users, especially younger ones, may have problems in finding appropriate keywords to perform effective search due to the terminology gap arising between their knowledge and the unfamiliar domain of archival collections. In this paper, we provide a general framework to bridge different domains across-time and, by this, to facilitate search and comparison as if carried in user's familiar domain (i.e., the present). In particular, we propose to find analogical terms across temporal text collections by applying a series of transformation procedures. We develop a cluster-biased transformation technique which makes use of hierarchical cluster structures built on the temporally distributed document collections. Our methods do not need any specially prepared training data and can be applied to diverse collections and time periods. We test the performance of the proposed approaches on the collections separated by both short (e.g., 20 years) and long time gaps (70 years), and we report improvements in range of 18%-27% over short and 56%-92% over long periods when compared to state-of-the-art baselines.

Missing data is a severe problem for algorithms that operate over knowledge graphs (KGs). Most previous research in KG completion has focused on the problem of inferring missing entities and missing relation types between entities. However, in addition to these, many KGs also suffer from missing entity types (i.e. the category labels for entities, such as /music/artist). Entity types are a critical enabler for many NLP tasks that use KGs as a reference source, and inferring missing entity types remains an important outstanding obstacle in the field of KG completion. Inspired by recent work to build a contextual KG embedding model, we propose a novel approach to address the entity type prediction problem. We compare the performance of our method with several state-of-the-art KG embedding methods, and show that our approach gives higher prediction accuracy compared to baseline algorithms on two real-world datasets. Our approach also produces consistently high accuracy when inferring entities and relation types, as well as the primary task of inferring entity types. This is in contrast to many of the baseline methods that specialize in one prediction task or another. We achieve this while preserving linear scalability with the number of entity types. Source code and datasets from this paper can be found at (https://github.ncsu.edu/cmoon2/kg).

Many modern database-backed web applications are built upon Object Relational Mapping (ORM) frameworks. While such frame- works ease application development by abstracting persistent data as objects, such convenience comes with a performance cost. In this paper, we studied 27 real-world open-source applications built on top of the popular Ruby on Rails ORM framework, with the goal to understand the database-related performance inefficiencies in these applications. We discovered a number of inefficiencies rang- ing from physical design issues to how queries are expressed in the application code. We applied static program analysis to identify and measure how prevalent these issues are, then suggested techniques to alleviate these issues and measured the potential performance gain as a result. These techniques significantly reduce database query time (up to 91%) and the webpage response time (up to 98%). Our study provides guidance to the design of future database engines and ORM frameworks to support database application that are performant yet without sacrificing programmability.

We propose a neural embedding approach to identify temporally like-minded user communities, i.e., those communities of users who have similar temporal alignment in their topics of interest. Like-minded user communities in social networks are usually identified by either considering explicit structural connections between users (link analysis), users' topics of interest expressed in their posted contents (content analysis), or in tandem. In such communities, however, the users' rich temporal behavior towards topics of interest is overlooked. Only few recent research efforts consider the time dimension and define like-minded user communities as groups of users who share not only similar topical interests but also similar temporal behavior. Temporal like-minded user communities find application in areas such as recommender systems where relevant items are recommended to the users at the right time. In this paper, we tackle the problem of identifying temporally like-minded user communities by leveraging unsupervised feature learning (embeddings). Specifically, we learn a mapping from the user space to a low-dimensional vector space of features that incorporate both topics of interest and their temporal nature. We demonstrate the efficacy of our proposed approach on a Twitter dataset in the context of three applications: news recommendation, user prediction and community selection, where our work is able to outperform the state-of-the-art on important information retrieval metrics.

Bots are regarded as the most common kind of malwares in the era of Web 2.0. In recent years, Internet has been populated by hundreds of millions of bots, especially on social media. Thus, the demand on effective and efficient bot detection algorithms is more urgent than ever. Existing works have partly satisfied this requirement by way of laborious feature engineering. In this paper, we propose a deep bot detection model aiming to learn an effective representation of social user and then detect social bots by jointly modeling social behavior and content information. The proposed model learns the representation of social behavior by encoding both endogenous and exogenous factors which affect user behavior. As to the representation of content, we regard the user content as temporal text data instead of just plain text as be treated in other existing works to extract semantic information and latent temporal patterns. To the best of our knowledge, this is the first trial that applies deep learning in modeling social users and accomplishing social bot detection. Experiments on real world dataset collected from Twitter demonstrate the effectiveness of the proposed model.

Community detection or graph clustering is crucial to understanding the structure of complex networks and extracting relevant knowledge from networked data. Latent factor model, e.g., non-negative matrix factorization and mixed membership block model, is one of the most successful methods for community detection. Latent factor models for community detection aim to find a distributed and generally low-dimensional representation, or coding, that captures the structural regularity of network and reflects the community membership of nodes. Existing latent factor models are mainly based on reconstructing a network from the representation of its nodes, namely network decoder, while constraining the representation to have certain desirable properties. These methods, however, lack an encoder that transforms nodes into their representation. Consequently, they fail to give a clear explanation about the meaning of a community and suffer from undesired computational problems. In this paper, we propose a non-negative symmetric encoder-decoder approach for community detection. By explicitly integrating a decoder and an encoder into a unified loss function, the proposed approach achieves better performance over state-of-the-art latent factor models for community detection task. Moreover, different from existing methods that explicitly impose the sparsity constraint on the representation of nodes, the proposed approach implicitly achieves the sparsity of node representation through its symmetric and non-negative properties, making the optimization much easier than competing methods based on sparse matrix factorization.

Public transportation is essential in people's daily life and it is crucial to understand how people move around the city. Some prior works have exploited GPS, Wi-Fi or bluetooth to collect data, in which extra sensors or devices were needed. Other works utilized data from smart card systems. However, some public transportation systems have their own smart card system and the smart card data cannot include all kinds of transportation modes, which makes it unsuitable for our study.Nowadays, each user has his/her own mobile phones and from the cellular data of mobile phone service providers, it is possible to know the uses' transportation mode and the fine-grained crowd flows. As such, given a set of cellular data, we propose a system for public transportation mode detection, crowd density estimation, and crowd flow estimation. Note that we only have cellular data, no extra sensor data collected from users' mobile phones. In this paper, we refer to some external data sources (e.g., the bus routing networks) to identify transportation modes. Users' cellular data sometimes have uncertainty about user location information. Thus, we propose two approaches for different transportation mode detection considering the cell tower properties, spatial and temporal factors. We demonstrate our system using the data from Chunghwa Telecom, which is the largest telecommunication company in Taiwan, to show the usefulness of our system.

This paper presents Storyfinder, an application which consists of a browser plugin and a web server backend with the goal to highlight and manage the information contained in web pages by combining techniques from natural language processing and visual analytics. Webpages are analyzed while visiting them by means of natural language processing components, and metadata in the form of named entities and keywords are extracted and stored for further reference. The extracted information is instantaneously highlighted in the web page and stored in a graph of entities and relations. The graph can be inspected and modified. The investigational scope can be set to a single web page, multiple web pages, or the complete set of analyzed web pages in a user's history. The graph view is designed to adhere to standards of visual analytics and information visualization. Storyfinder is available as an open source application. Its benefit for information access is evaluated in a small user study.

Hyper-local pricing data, e.g., about foods and commodities, exhibit subtle spatiotemporal variations that can be useful as crucial precursors of future events. Three major challenges in modeling such pricing data include: i) temporal dependencies underlying features; ii) spatiotemporal missing values; and iii) constraints underlying economic phenomena. These challenges hinder traditional event forecasting models from being applied effectively. This paper proposes a novel spatiotemporal event forecasting model that concurrently addresses the above challenges. Specifically, given continuous price data, a new soft time-lagged model is designed to select temporally dependent features. To handle missing values, we propose a data tensor completion method based on price domain knowledge. The parameters of the new model are optimized using a novel algorithm based on the Alternative Direction Methods of Multipliers (ADMM). Extensive experimental evaluations on multiple datasets demonstrate the effectiveness of our proposed approach.

The skyline query is important in database community. In recent years, the researches on incomplete data have been increasingly considered, especially for the skyline query. However, the existing skyline definition on incomplete data cannot provide users with valuable references. In this paper, we propose a novel skyline definition utilizing probabilistic model on incomplete data where each point has a probability to be in the skyline. In particular, it returnsK points with the highest skyline probabilities. Meanwhile, it is a big challenge to compute probabilistic skyline on incomplete data. We propose an efficient algorithm PISkyline, which utilizes two pruning strategies to reduce the number of points and adopts two optimizations to accelerate probability computation for each point. Nevertheless, PISkyline is susceptible to the order of input data and there is still a great deal of room for optimization. We develop a point-level sorting technique by adjusting the order of accessing points to further improve the efficiency of PISkyline. Our experimental results demonstrate that our algorithms are tens of times faster than the naive algorithm on both synthetic and real datasets.

With the proliferation of GPS-enabled smart devices and increased availability of wireless network, spatial crowdsourcing (SC) has been recently proposed as a framework to automatically request workers (i.e., smart device carriers) to perform location-sensitive tasks (e.g., taking scenic photos, reporting events). In this paper we study a destination-aware task assignment problem that concerns the optimal strategy of assigning each task to proper worker such that the total number of completed tasks can be maximized whilst all workers can reach their destinations before deadlines after performing assigned tasks. Finding the global optimal assignment turns out to be an intractable problem since it does not imply optimal assignment for individual worker. Observing that the task assignment dependency only exists amongst subsets of workers, we utilize tree-decomposition technique to separate workers into independent clusters and develop an efficient depth-first search algorithm with progressive bounds to prune non-promising assignments. Our empirical studies demonstrate that our proposed technique is quite effective and settle the problem nicely.

Conventional spatial keyword queries confront the difficulty of returning desired objects that are synonyms but morphologically different to query keywords. To overcome this flaw, this paper investigates the interactive spatial keyword querying with semantics. It aims to enhance the conventional queries by not only making sense of the query keywords, but also refining the understanding of query semantics through interactions. On top of the probabilistic topic model, a novel interactive strategy is proposed to precisely infer the latent query semantics by learning from user feedbacks. In each interaction, the returned objects are carefully selected to ensure effective inference of user intended query semantics. Query processing is carried out on a small candidate object set at each round of interaction, and the whole querying process terminates when the latent query semantics learned from user feedback becomes explicit enough. The experimental results on real check-in dataset demonstrates that the quality of results has been significantly improved through limited number of interactions.

The incorporation of contextual information is an important part of context-aware recommendation. Many context-aware recommendation systems adopt tensor completion to include contextual information. However, the symmetries between dimensions of a tensor induce an unreasonable assumption that users, items and contexts should be treated equally in recommender systems. In this paper, we address this by using matrices over finite abelian group algebra (AGA) to model context-aware interactions between users and items. Specifically, we formulate context-aware recommendation as a low-rank matrix completion problem over AGA (MC-AGA) and derive a new algorithm using the inexact augmented Lagrange multiplier method. We then test MC-AGA on two real-world datasets: one containing implicit feedback and one with explicit feedback. Experiment results show that MC-AGA outperforms not only existing tensor completion algorithms but also recommendation systems with other context-aware representations.

Network embedding leverages the node proximity manifested to learn a low-dimensional node vector representation for each node in the network. The learned embeddings could advance various learning tasks such as node classification, network clustering, and link prediction. Most, if not all, of the existing works, are overwhelmingly performed in the context of plain and static networks. Nonetheless, in reality, network structure often evolves over time with addition/deletion of links and nodes. Also, a vast majority of real-world networks are associated with a rich set of node attributes, and their attribute values are also naturally changing, with the emerging of new content patterns and the fading of old content patterns. These changing characteristics motivate us to seek an effective embedding representation to capture network and attribute evolving patterns, which is of fundamental importance for learning in a dynamic environment. To our best knowledge, we are the first to tackle this problem with the following two challenges: (1) the inherently correlated network and node attributes could be noisy and incomplete, it necessitates a robust consensus representation to capture their individual properties and correlations; (2) the embedding learning needs to be performed in an online fashion to adapt to the changes accordingly. In this paper, we tackle this problem by proposing a novel dynamic attributed network embedding framework - DANE. In particular, DANE first provides an offline method for a consensus embedding and then leverages matrix perturbation theory to maintain the freshness of the end embedding results in an online manner. We perform extensive experiments on both synthetic and real attributed networks to corroborate the effectiveness and efficiency of the proposed framework., Comment: 10 pages

One of the main challenges in large-scale data integration for relational schemas is creating an accurate mediated schema, and generating accurate semantic mappings between heterogeneous data sources and this mediated schema. Some applications can start with a moderately accurate mediated schema and mappings and refine them over time, which is referred to as the pay-as-you-go approach to data integration. Creating the mediated schema and mappings automatically to bootstrap the pay-as-you-go approach has been extensively studied. However, refining the mediated schema and mappings is still an open challenge because the data sources are usually heterogeneous and use diverse and sometimes ambiguous vocabularies. In this paper, we introduce UFeed, a system that refines relational mediated schemas and mappings based on user feedback over query answers. UFeed translates user actions into refinement operations that are applied to the mediated schema and mappings to improve their quality. We experimentally verify that UFeed improves the quality of query answers over real heterogeneous data sources extracted from the web.

The proliferation of Electronic Health Records (EHRs) challenges data miners to discover potential and previously unknown patterns from a large collection of medical data. One of the tasks that we address in this paper is to reveal previously unknown effects of drugs on laboratory test results. We propose a method that leverages drug information to find a meaningful list of drugs that have an effect on the laboratory result. We formulate the problem as a convex non smooth function and develop a proximal gradient method to optimize it. The model has been evaluated on two important use cases: lowering low-density lipoproteins and glycated hemoglobin test results. The experimental results provide evidence that the proposed method is more accurate than the state-of-the-art method, rediscover drugs that are known to lower the levels of laboratory test results, and most importantly, discover additional potential drugs that may also lower these levels.