Graph Neural Networks (GNNs) have emerged as powerful tools for representation learning on structured data. The graph convolutional filter (GCF) for aggregating neighbor information is shown to be the key factor that leads to GNNs' success. Various GCFs are designed but how to select the proper filter that can best benefit the data and the task remains an open problem. In this paper, we introduce the Adaptive Filter Selection (AdaFS) framework that addresses two critical issues: (1) defining a criterion to establish a strong base filter set; and (2) adaptively selecting filters for a specific task, even when labeled data is limited, by employing Laplacian score regularization. We further connect this multiple GCF learning process and the well-developed multiple kernel learning problem to provide a solid rationale for filter selection. With experiments on 9 datasets, AdaFS gets the best average performance.

Next-item prediction is a a popular problem in the recommender systems domain. As the name suggests, the task is to recommend subsequent items that a user would be interested in given contextual information and historical interaction data. In our paper, we model a general notion of context via a sequence of item interactions. We model the next item prediction problem using the Bayesian framework and capture the probability of appearance of a sequence through the posterior mean of the Beta distribution. We train two neural networks to accurately predict the alpha & beta parameter values of the Beta distribution. Our novel approach of combining black-box style neural networks, known to be suitable for function approximation with Bayesian estimation methods have resulted in an innovative method that outperforms various state-of-the-art baselines. We demonstrate the effectiveness of our method in two real world datasets. Our framework is an important step towards the goal of building privacy preserving recommender systems.

In this work, we address the problem of providing job recommendations in an online session setting, in which we do not have full user histories. We propose a recommendation approach, which uses different autoencoder architectures to encode sessions from the job domain. The inferred latent session representations are then used in a k-nearest neighbor manner to recommend jobs within a session. We evaluate our approach on three datasets, (1) a proprietary dataset we gathered from the Austrian student job portal Studo Jobs, (2) a dataset released by XING after the RecSys 2017 Challenge and (3) anonymized job applications released by CareerBuilder in 2012. Our results show that autoencoders provide relevant job recommendations as well as maintain a high coverage and, at the same time, can outperform state-of-the-art session-based recommendation techniques in terms of system-based and session-based novelty.

Most industrial recommender systems rely on the popular collaborative filtering (CF) technique for providing personalized recommendations to its users. However, the very nature of CF is adversarial to the idea of user privacy, because users need to share their preferences with others in order to be grouped with like-minded people and receive accurate recommendations. While previous privacy preserving approaches have been successful inasmuch as they concealed user preference information to some extent from a centralized recommender system, they have also, nevertheless, incurred significant trade-offs in terms of privacy, scalability, and accuracy. They are also vulnerable to privacy breaches by malicious actors. In light of these observations, we propose a novel selective privacy preserving (SP2) paradigm that allows users to custom define the scope and extent of their individual privacies, by marking their personal ratings as either public (which can be shared) or private (which are never shared and stored only on the user device). Our SP2 framework works in two steps: (i) First, it builds an initial recommendation model based on the sum of all public ratings that have been shared by users and (ii) then, this public model is fine-tuned on each user's device based on the user private ratings, thus eventually learning a more accurate model. Furthermore, in this work, we introduce three different algorithms for implementing an end-to-end SP2 framework that can scale effectively from thousands to hundreds of millions of items. Our user survey shows that an overwhelming fraction of users are likely to rate much more items to improve the overall recommendations when they can control what ratings will be publicly shared with others.

Association rules, one of the most useful constructs in data mining, can be exerted to capture interesting dependencies between variables in large datasets. Herawan and Dens initiated the investigation of mining association rules from transactional datasets using soft set theory. Unfortunately, some existing concepts in the literature were unable to realize properly Herawan and Deris's initial idea. This paper aims to offer further detailed insights into soft set based association rule mining. With regard to regular association rule mining using soft sets, we refine several existing concepts to improve the generality and clarity of former definitions. Regarding maximal association rule mining based on soft sets, we point out the drawbacks of some existing definitions and offer some way to rectify the problem. A number of new notions, such as transactional data soft sets, parameter-taxonomic soft sets, parameter cosets, realizations and M-realizations of parameter sets are proposed to facilitate soft set based association rule mining. Several algorithms are designed to find M-realizations of parameter sets or extract sigma-M-strong and gamma-M-reliable maximal association rules in parameter-taxonomic soft sets. We also present an example to illustrate potential applications of our method in clinical diagnosis. Moreover, two case studies are conducted to highlight the essentials of soft set based association rule mining approach. (C) 2016 Elievier B.V. All rights reserved.

This paper introduces human curation signals and demonstrates incorporating human curation signals improves the relevance of state-of-art recommendation system models by up to 30% by experiments on a large-scale Pinterest dataset.

Twitter (and similar microblogging services) has become a central nexus for discussion of the topics of the day. Twitter data contains rich content and structured information on users' topics of interest and behavior patterns. Correctly analyzing and modeling Twitter data enables the prediction of the user behavior and preference in a variety of practical applications, such as tweet recommendation and followee recommendation. Although a number of models have been developed on Twitter data in prior work, most of these only model the tweets from users, while neglecting their valuable retweet information in the data. Models would enhance their predictive power by incorporating users' retweet content as well as their retweet behavior.In this paper, we propose two novel Bayesian nonparametric models, URM and UCM, on retweet data. Both of them are able to integrate the analysis of tweet text and users' retweet behavior in the same probabilistic framework. Moreover, they both jointly model users' interest in tweet and retweet. As nonparametric models, URM and UCM can automatically determine the parameters of the models based on input data, avoiding arbitrary parameter settings. Extensive experiments on real-world Twitter data show that both URM and UCM are superior to all the baselines, while UCM further outperforms URM, confirming the appropriateness of our models in retweet modeling.

Large online promotion activities are different from traditional online marketing activities. Issues about consumer behavior in large online promotion activities deserve further discussion. Via consumer surveys before and after the "Double 11" promotion on Taobao.com, we studied the unplanned consumption and its influencing factors (promotion, time pressure, social environment (SE), in-store slack) based on the SOR model and theories of self-regulation. The results show that consumers allocate in-store slack in their shopping budget as a strategy of controlling unplanned consumption. Promotion range, preparation time, SE and in-store slack have effects on unplanned consumption. Actual shopping time and preparation time moderate the relationship between in-store slack and unplanned consumption. Actual shopping time attenuates the relationship between in-store slack and unplanned consumption, while preparation time strengthens the relationship between in-store slack and unplanned consumption. Managerial implications are also discussed based on the research.

Over the past few years, major web search engines have introduced knowledge bases to offer popular facts about people, places, and things on the entity pane next to regular search results. In addition to information about the entity searched by the user, the entity pane often provides a ranked list of related entities. To keep users engaged, it is important to develop a recommendation model that tailors the related entities to individual user interests. We propose a probabilistic Three-way Entity Model (TEM) that provides personalized recommendation of related entities using three data sources: knowledge base, search click log, and entity pane log. Specifically, TEM is capable of extracting hidden structures and capturing underlying correlations among users, main entities, and related entities. Moreover, the TEM model can also exploit the click signals derived from the entity pane log. We further provide an inference technique to learn the parameters in TEM, and propose a principled preference learning method specifically designed for ranking related entities. Extensive experiments with two real-world datasets show that TEM with our probabilistic framework significantly outperforms a state of the art baseline, confirming the effectiveness of TEM and our probabilistic framework in related entity recommendation.

Probabilistic topic models, such as PLSA and LDA, are gaining popularity in many fields due to their high-quality results. Unfortunately, existing topic models suffer from two drawbacks: (1) model complexity and (2) disjoint topic groups. That is, when a topic model involves multiple entities (such as authors, papers, conferences, and institutions) and they are connected through multiple relationships, the model becomes too difficult to analyze and often leads to in-tractable solutions. Also, different entity types are classified into disjoint topic groups that are not directly comparable, so it is difficult to see whether heterogeneous entities (such as authors and conferences) are on the same topic or not (e.g., are Rakesh Agrawal and KDD related to the same topic?). In this paper, we propose a novel universal topic framework (UniZ) that addresses these two drawbacks using "prior topic incorporation." Since our framework enables representation of heterogeneous entities in a single universal topic space, all entities can be directly compared within the same topic space. In addition, UniZ breaks complex models into much smaller units, learns the topic group of each entity from the smaller units, and then propagates the learned topics to others. This way, it leverages all the available signals without introducing significant computational complexity, enabling a richer representation of entities and highly accurate results. In a widely-used DBLP dataset prediction problem, our approach achieves the best prediction performance over many state-of-the-art methods. We also demonstrate practical potential of our approach with search logs from a commercial search engine.

Query suggestions have been a valuable feature for e-commerce sites in helping shoppers refine their search intent. In this paper, we develop an algorithm that helps e-commerce sites like eBay mingle the output of different recommendation algorithms. Our algorithm is based on "Thompson Sampling" --- a technique designed for solving multi-arm bandit problems where the best results are not known in advance but instead are tried out to gather feedback. Our approach is to treat query suggestions as a competition among data resources: we have many query suggestion candidates competing for limited space on the search results page. An "arm" is played when a query suggestion candidate is chosen for display, and our goal is to maximize the expected reward (user clicks on a suggestion). Our experiments have shown promising results in using the click-based user feedback to drive success by enhancing the quality of query suggestions.

Peripheral blood (PB) is known as a source of mesenchymal stem cells (MSCs), as is bone marrow (BM), and is acquired easily. However, it is difficult to have enough MSCs, and their osteogenic capacity with dental implantations is scarce. Therefore, we characterized peripheral blood mesenchymal stem cells (PBMSCs) cultured on a bone marrow-derived mesenchymal stem cell (BMMSC) natural extracellular matrix (ECM) and demonstrated the osteogenic capability in an experimental chamber implant surgery model in rabbits. We isolated PBMSCs from rabbits by culturing on a natural ECM-coated plate during primary culture. We characterized the PBMSCs using a fluorescence-activated cell scanner, cell proliferation assay, and multiple differentiation assay and compared them with BMMSCs. We also analyzed the osteogenic potential of PBMSCs mixed with hydroxyapatite/tricalcium phosphate (HA/TCP) by transplanting them into immunocompromised mice. Then, the mixture was applied to the canals. After 3 and 6 wk, we analyzed new bone (NB) formation inside the chambers using histological and histomorphometric analyses. The PBMSCs had a similar rate of BrdU-positive cells to BMMSCs, positively expressing CD90 but negative for CD14. The PBMSCs also showed osteogenic, adipogenic, and chondrogenic ability in vitro and osteogenic ability in vivo. Histological and histomorphometric results illustrated that the PBMSC and BMMSC groups showed higher NB than the HA/TCP and defect groups in the upper and lower chambers at 6 wk and in the upper canal at 3 wk; however, there was no difference in NB among all groups in the lower canal at 3 wk. The PBMSCs have characteristics and bone regeneration ability similar to BMMSCs both in vitro and in vivo. ECM was effective for obtaining PBMSCs. Therefore, PBMSCs are a promising source for bone regeneration for clinical use.

Social influence analysis on microblog networks, such as Twitter, has been playing a crucial role in online advertising and brand management. While most previous influence analysis schemes rely only on the links between users to find key influencers, they omit the important text content created by the users. As a result, there is no way to differentiate the social influence in different aspects of life (topics). Although a few prior works do support topic-specific influence analysis, they either separate the analysis of content from the analysis of network structure, or assume that content is the only cause of links, which is clearly an inappropriate assumption for microblog networks. To address the limitations of the previous approaches, we propose a novel Followship-LDA (FLDA) model, which integrates both content topic discovery and social influence analysis in the same generative process. This model properly captures the content-related and content-independent reasons why a user follows another in a microblog network. We demonstrate that FLDA produces results with significantly better precision than existing approaches. Furthermore, we propose a distributed Gibbs sampling algorithm for FLDA, and demonstrate that it provides excellent scalability on large clusters. Finally, we incorporate the FLDA model in a general search framework for topic-specific influencers. A user freely expresses his/her interest by typing a few keywords, the search framework will return a ranked list of key influencers that satisfy the user's interest.

With the rapid growth of Web 2.0, a variety of content sharing services, such as Flickr, YouTube, Blogger, and TripAdvisor etc, have become extremely popular over the last decade. On these websites, users have created and shared with each other various kinds of resources, such as photos, video, and travel blogs. The sheer amount of user-generated content varies greatly in quality, which calls for a principled method to identify a set of authorities, who created high-quality resources, from a massive number of contributors of content. Since most previous studies only infer global authoritativeness of a user, there is no way to differentiate the authoritativeness in different aspects of life (topics). In this paper, we propose a novel model of Topic-specific Authority Analysis (TAA), which addresses the limitations of the previous approaches, to identify authorities specific to given query topic(s) on a content sharing service. This model jointly leverages the usage data collected from the sharing log and the favorite log. The parameters in TAA are learned from a constructed training dataset, for which a novel logistic likelihood function is specifically designed. To perform Bayesian inference for TAA with the new logistic likelihood, we extend typical Gibbs sampling by introducing auxiliary variables. Thorough experiments with two real-world datasets demonstrate the effectiveness of TAA in topic-specific authority identification as well as the generalizability of the TAA generative model.

Categorical (topic) similarity between a web page and an advertisement (ad) text has long been used for contextual advertising. In this paper, we explore the use of the categorical similarity score, referred to as Category Match Score (CMS), in the context of search advertising. In particular, we explore the effect of CMS on various ad-effectiveness prediction tasks, including user-judgment prediction, ad click-through-rate prediction (CTR), and revenue-per-impression prediction. Our extensive experiments on two editorial datasets and one live traffic dataset demonstrate that CMS is one of the strongest features in the judgment prediction task and that CMS-based filtering is very effective in improving revenue per impression as well as CTR. We believe that our analyses can be extremely effective in helping web service providers serve more relevant and profitable ads to users.

OF THE DISSERTATION Probabilistic Topic Models for Graph Mining

We always delineated internal target volume (ITV) observing tumor movement by 4-dimensional computed tomography (4D CT) in lung tumor. However, when we saw the mass outside ITV during verified bone matching using megavoltage CT (MVCT) in 1st treatment. This study evaluated the fast adaptive planning method using commercially deformable image registration (DIR) tool and graphic processing unit (GPU) based radiation treatment planning system from MVCT images. From November 2013 to January 2014 we selected lung cancer patients who have irregular breathing pattern with concurrent chemoradiation therapy. Treatment dose was 48 Gy per total in 4 fractions using 3d helical tomotherapy. We got planning CT images and 4D images through 4D CT. We delineated gross tumor volume (GTV) included a gross tumor and ITV included a gross tumor through 4D CT. In addition, planning target volume (PTV) included a 3 to 5 mm safety margin around ITV. Every treatment was verified MVCT because mass can detected by MVCT. If we saw the mass outside ITV during verified MVCT in each treatment. We would perform respiration education as simulation and re_MVCT. Although we have performed re_MVCT, we saw the mass outside the ITV and we would select resimulation and replanning. Most of patients delayed 1st treatment times. If patients have treated chemotherapy before radiation therapy, we could select replanning using MVCT. We sent the MVCT images to Mim_vista for replanning. In addition, we added redelineated GTV and ITV for comparing MVCT and planning 4D CT on Mim_vista. First we evaluated replanning time using GPU based RTP and DIR program. Second we checked dosimetric parameters such as mean lung dose (MLD) and V5, V10, V20. Planning time including redelineated ITV and replanned helical IMRT plan takes less than 30 minutes. MLD increased 1∼1.5 Gy. V5, V10, V20 increased 4∼5 %, 2∼3 %, 1∼2 % respectively. In less than 30 minutes fast adaptive planning is possible through using DIR program and GPU based RTP. Dosimetric parameters were in our criteria. The fast adaptive planning method would need patients to have irregular breathing patterns.