Stanford Center for Professional Development, Stanford University, Palo Alto, CA 94305, USA; Department of Medicine, Jiangsu University School of Medicine, Zhenjiang, Jiangsu 212013, China; Institute of Radio Frequency & Optical Electronics-Integrated Circuits, School of Information and Engineering, Southeast University, Nanjing, Jiangsu 210096, China; Wenjing Technologies, Shanghai 200020, China; Dawnlight Technologies, Palo Alto, CA 94304, USA; Department of Respiratory and Critical Care Medicine, The First Affiliated Hospital, Nanjing Medical University, Nanjing, Jiangsu 210029, China.

As a widely used vital sign within cardiology, Electrocardiography (ECG) provides the basis for assessing heart function and diagnosing cardiovascular diseases. Automated anomaly detection for ECG plays an important role in improving patient diagnosis efficiency and reducing healthcare costs. Practically, due to the limits of electronics support or the medical system setting, image is a more common format for large-scale ECG storage in most clinical institutions. To guarantee an automated ECG detection model's scalability and practicality in clinical applications, taking good advantage of ECG images is crucial. However, existing time digital-based discriminative models fail to learn from images effectively for two reasons. First of all, the signals recorded on images have much lower resolution and higher noise, which makes it impractical to extract precise ECG signals following existing techniques. Meanwhile, the differences between abnormal signals are usually subtle, and they may be overwhelmed by the noises in the images as well. Towards this end, we design a novel neural framework that can be directly applied to massive ECG images determining various types of cardiology abnormalities. It classifies fine-grained ECG images based on weakly supervised strategy, in which case only image-level labeling is required. By eliminating the need for part annotations, the proposed method can result in significant savings in annotation time and cost. The effectiveness of the method is demonstrated by experimental results on two real ECG datasets.

Recently, increasingly more methods are proposed to automatically detect the abnormalities in Electrocardiography (ECG). Despite their success on public golden standard datasets, two challenges hinder the adoption of existing methods on real-world clinical ECG data in practice. To start with, most methods are designed based on digital signal data while most ECG data in the hospital are stored as images. Additionally, they ignore the correlation among different abnormal cardiac patterns and hence cannot detect multiple abnormalities at the same time. To practically address these challenges, we propose a Fine-grained Multi-label ECG (FM-ECG) framework to effectively detect the abnormalities from the real clinical ECG data in the following two aspects. Firstly, we propose to directly detect the abnormalities on the ECG images via a weakly supervised fine-grained classification mechanism, which can discover the potential discriminative parts and adaptively fuse them via image-level annotations only. Secondly, we take the ECG label dependencies into consideration by inferencing with a recurrent neural network (RNN). Experimental results on two real-world large-scale ECG datasets prove the capability of FM-ECG comparing with other state-of-the-art methods in ECG abnormally detection. Moreover, visualization analyses on attention parts show that meaningful spatial attention can be effectively learned by FM-ECG.

Factorization Machines (FM) have been recognized as an effective learning paradigm for incorporating complex relations to improve item recommendation in recommender systems. However, one open issue of FM lies in its factorized representation (latent factors) for each feature in the observed feature space, a characteristic often resulting in a large parameter space. Therefore, training FM (in other words, learning a large number of parameters in FM) is a computationally expensive task. Our work targets to improve the scalability of FM by building it in a distributed environment. We propose a new system framework that integrates Parameter Server (PS) with the Map/Reduce (MR) framework. In addition to the data parallelism achieved via MR, our framework particularly benefits from PS for model parallelism, a critical characteristic for learning with a large number of parameters in FM. We further address two specific challenges in our system, namely, communication cost and parameter update collision. Through both offline and online experiments on recommendation tasks, we demonstrate that the proposed system framework succeeds in scaling up FM for very large datasets, while it also maintains competitive performance on recommendation quality compared to alternative baselines.

While matrix factorisation models are ubiquitous in large scale recommendation and search, real time application of such models requires inner product computations over an intractably large set of item factors. In this manuscript we present a novel framework that uses the inverted index representation to exploit structural properties of sparse vectors to significantly reduce the run time computational cost of factorisation models. We develop techniques that use geometry aware permutation maps on a tessellated unit sphere to obtain high dimensional sparse embeddings for latent factors with sparsity patterns related to angular closeness of the original latent factors. We also design several efficient and deterministic realisations within this framework and demonstrate with experiments that our techniques lead to faster run time operation with minimal loss of accuracy.

Content recommendation systems are typically based on one of the following paradigms: user based customization, or recommendations based on either collaborative filtering or low rank matrix factorization methods, or with systems that impute user interest profiles based on content browsing behavior and retrieve items similar to the interest profiles. All of these systems have a distinct disadvantage, namely data sparsity and cold-start on items or users. Furthermore, very few content recommendation solutions explicitly model the wealth of information in implicit negative feedback from the users. In this paper, we propose a hybrid solution that makes use of a latent factor model to infer user interest vectors. The hybrid approach enables us to overcome both the data sparsity and cold-start problems. Our proposed method is learned purely on implicit user feedback, both positive and negative. Exploiting the information in the negative feedback allows the user profiles generated to be discriminative. We also provide a Map/Reduce framework based implementation that enables scaling our solution to real-world recommendation problems. We demonstrate the efficacy of our proposed approach with both offline experiments and A/B tests on live traffic on Yahoo properties.

Despite having various attractive qualities such as high prediction accuracy and the ability to quantify uncertainty and avoid over-fitting, Bayesian Matrix Factorization has not been widely adopted because of the prohibitive cost of inference. In this paper, we propose a scalable distributed Bayesian matrix factorization algorithm using stochastic gradient MCMC. Our algorithm, based on Distributed Stochastic Gradient Langevin Dynamics, can not only match the prediction accuracy of standard MCMC methods like Gibbs sampling, but at the same time is as fast and simple as stochastic gradient descent. In our experiments, we show that our algorithm can achieve the same level of prediction accuracy as Gibbs sampling an order of magnitude faster. We also show that our method reduces the prediction error as fast as distributed stochastic gradient descent, achieving a 4.1% improvement in RMSE for the Netflix dataset and an 1.8% for the Yahoo music dataset.

Many internet companies, such as Yahoo, Facebook, Google and Twitter, rely on content recommendation systems to deliver the most relevant content items to individual users through personalization. Delivering such personalized user experiences is believed to increase the long term engagement of users. While there has been a lot of progress in designing effective personalized recommender systems, by exploiting user interests and historical interaction data through implicit (item click) or explicit (item rating) feedback, directly optimizing for users' satisfaction with the system remains challenging. In this paper, we explore the idea of using item-level dwell time as a proxy to quantify how likely a content item is relevant to a particular user. We describe a novel method to compute accurate dwell time based on client-side and server-side logging and demonstrate how to normalize dwell time across different devices and contexts. In addition, we describe our experiments in incorporating dwell time into state-of-the-art learning to rank techniques and collaborative filtering models that obtain competitive performances in both offline and online settings.

Friendship prediction is an important task in social network analysis (SNA). It can help users identify friends and improve their level of activity. Most previous approaches predict users' friendship based on their historical records, such as their existing friendship, social interactions, etc. However, in reality, most users have limited friends in a single network, and the data can be very sparse. The sparsity problem causes existing methods to overfit the rare observations and suffer from serious performance degradation. This is particularly true when a new social network just starts to form. We observe that many of today's social networks are composite in nature, where people are often engaged in multiple networks. In addition, users' friendships are always correlated, for example, they are both friends on Facebook and Google+. Thus, by considering those overlapping users as the bridge, the friendship knowledge in other networks can help predict their friendships in the current network. This can be achieved by exploiting the knowledge in different networks in a collective manner. However, as each individual network has its own properties that can be incompatible and inconsistent with other networks, the naive merging of all networks into a single one may not work well. The proposed solution is to extract the common behaviors between different networks via a hierarchical Bayesian model. It captures the common knowledge across networks, while avoiding negative impacts due to network differences. Empirical studies demonstrate that the proposed approach improves the mean average precision of friendship prediction over state-of-the-art baselines on nine real-world social networking datasets significantly.

Most existing collaborative filtering models only consider the use of user feedback (e.g., ratings) and meta data (e.g., content, demographics). However, in most real world recommender systems, context information, such as time and social networks, are also very important factors that could be considered in order to produce more accurate recommendations. In this work, we address several challenges for the context aware movie recommendation tasks in CAMRa 2010: (1) how to combine multiple heterogeneous forms of user feedback? (2) how to cope with dynamic user and item characteristics? (3) how to capture and utilize social connections among users? For the first challenge, we propose a novel ranking based matrix factorization model to aggregate explicit and implicit user feedback. For the second challenge, we extend this model to a sequential matrix factorization model to enable time-aware parametrization. Finally, we introduce a network regularization function to constrain user parameters based on social connections. To the best of our knowledge, this is the first study that investigates the collective modeling of social and temporal dynamics. Experiments on the CAMRa 2010 dataset demonstrated clear improvements over many baselines.

Transfer learning as a new machine learning paradigm has gained increasing attention lately. In situations where the training data in a target domain are not sufficient to learn predictive models effectively, transfer learning leverages auxiliary source data from related domains for learning. While most of the existing works in this area are only focused on using the source data with the same representational structure as the target data, in this paper, we push this boundary further by extending transfer between text and images. We integrate documents , tags and images to build a heterogeneous transfer learning factor alignment model and apply it to improve the performance of tag recommendation. Many algorithms for tag recommendation have been proposed, but many of them have problem; the algorithm may not perform well under cold start conditions or for items from the long tail of the tag frequency distribution. However, with the help of documents, our algorithm handles these problems and generally outperforms other tag recommendation methods, especially the non-transfer factor alignment model.

The Yahoo! music rating data set in KDD Cup 2011 raises several interesting challenges: (1) The data covers a lengthy time period of more than eight years. (2) Not only are training ratings associated date and time information, so are the test ratings. (3) The items form a hierarchy consisting of four types of items: genres, artists, albums and tracks. To capture the rich temporal dynamics within the data set, we design a class of time-aware matrix/tensor factorization models, which adopts time series based parameterizations and models user/item drifting behaviors at multiple temporal resolutions. We also incorporate the taxonomical structure into the item parameters by introducing sharing parameters between ancestors and descendants in the taxonomy. Finally, we have identified some conditions that systematically affect the effectiveness of different types of models and parameter settings. Based on these findings, we designed an informative ensemble framework, which considers additional meta features when making predictions for a particular pair of user and item. Using these techniques, we built the best single model reported officially, and our final ensemble model got third place in KDD Cup 2011.

As the popularity of the social media increases, as evidenced in Twitter, Facebook and China's Renren, spamming activities also picked up in numbers and variety. On social network sites, spammers often disguise themselves by creating fake accounts and hijacking normal users' accounts for personal gains. Different from the spammers in traditional systems such as SMS and email, spammers in social media behave like normal users and they continue to change their spamming strategies to fool anti spamming systems. However, due to the privacy and resource concerns, many social media websites cannot fully monitor all the contents of users, making many of the previous approaches, such as topology-based and content-classification-based methods, infeasible to use. In this paper, we propose a novel method for spammer detection in social networks that exploits both social activities as well as users' social relations in an innovative and highly scalable manner. The proposed method detects spammers following collective activities based on users' social actions and relations. We have empirically tested our method on data from Renren.com, which is the largest social network in China, and demonstrated that our new method can improve the detection performance significantly.

Recommender systems have to deal with the cold start problem as new users and/or items are always present. Rating elicitation is a common approach for handling cold start. However, there still lacks a principled model for guiding how to select the most useful ratings. In this paper, we propose a principled approach to identify representative users and items using representative-based matrix factorization. Not only do we show that the selected representatives are superior to other competing methods in terms of achieving good balance between coverage and diversity, but we also demonstrate that ratings on the selected representatives are much more useful for making recommendations (about 10% better than competing methods). In addition to illustrating how representatives help solve the cold start problem, we also argue that the problem of finding representatives itself is an important problem that would deserve further investigations, for both its practical values and technical challenges.