Search

Fraud detection for peer-to-peer (P2P) lending is an important and challenging problem in both real application and research area. Different from existing methods which are mainly based on user demographic information, in this paper we study if other information such as user relationship represented by graph and transaction description information are helpful for the fraud detection problem. Meanwhile, attention mechanism is widely employed to explain how deep learning model works. However, existing studies don’t discriminate the importance of neighbors and different edge features in graph. In this paper, we propose a new graph attentive network model called ‘FDNE’ for P2P lending fraud detection based on text information and/or user relationship information. We design a novel attention method called edge-feature attention and use a global normalization operation to identify influential edge feature. Experiments conducted on a real dataset demonstrate that our model significantly outperforms other baselines and can make reasonable explanations simultaneously.

Matrix Factorization (MF) is a latent factor model, which has been one of the most popular techniques for recommendation systems. Performance of MF-based recommender models degrades as the sparseness of user-item rating data increases. MF-based models map each user and each item into a low dimensional space, where either of them is represented by a point in the space. While a point is a concise and simple representation of a user’s preference or an item’s characteristics, it is hard to learn the precise position of the point, especially when the data is very sparse. In this paper we propose an alternative latent space model, Latent Path Connected Space model (LSpace), to address this issue. In this model, users and items are both represented by path connected space described by different latent dimensions and spatial intersection between user space and item space reflects their matching degree. Extensive evaluations on four real-world datasets show that our approach outperforms the Matrix Factorization model on rating prediction task especially when the rating data is extremely sparse.

Extreme multi-label classification is an important data mining technique, which can be used to label each unseen instance with a subset of labels from a large label set. It has wide applications and many methods have been proposed in recent years. Existing methods either seek to compress label space or train a classifier based on instances’ features, among which tree-based classifiers enjoy the advantages of better efficiency and accuracy. In many real world applications, instances are not independent and relationship between instances is very useful information. However, how to utilize relationship between instances in extreme multi-label classification is less studied. Exploiting such relationship may help improve prediction accuracy, especially in the circumstance that feature space is very sparse. In this paper, we study how to utilize the similarity between instances to build more accurate tree-based extreme multi-label classifiers. To this end, we introduce the utilization of relationship between instances to state-of-the-art models in two ways: feature engineering and collaborative labeling. Extensive experiments conducted on three real world datasets demonstrate that our proposed method achieves higher accuracy than the state-of-the-art models.

Deep learning has been successfully introduced to collaborative filtering for recommendation systems in recent years. In these studies, autoencoder models are usually used to extract latent features of items or users, and related researches facilitate the learning techniques using item and user latent factors in matrix factorization models. Inputs of autoencoder models are usually side information of items or profile information of users. However, in many real world applications, we do not have users’ profile information. Moreover, results of both matrix factorization and deep learning models are difficult to interpret. To solve these issues, in this paper, we propose a deep collaborative filtering model with attention mechanism. With this model, learning of latent factors of users and items can be facilitated by deep processing of items’ tag information. Furthermore, user preferences learned are interpretable. Experiments conducted on a real world dataset demonstrate that our model can significantly outperform the state-of-the-art deep collaborative filtering models.

Glaucoma is a group of eye diseases that damage the optic nerves progressively and lead to deterioration in vision irreversibly. Diagnosing glaucoma based on retinal images automatically is meaningful both in practice and research area. While deep learning models have achieved superior performance in natural images recognition and have been also used for medical image diagnosis recently, the models usually rely on large dataset and expensive computing resources, thus limiting the wider use in medical areas. So how to train a deep learning model with relatively small amount of medical data is challenging. In this paper, we propose to incorporate domain knowledge to construct a two-branch Convolutional Neural Networks (CNN) to learn a classifier for glaucoma diagnosis based on the retinal image. Our two-branch CNN framework can analyze the whole image and pay special attention to discriminative local region of image at the same time. Experiments conducted on real medical dataset demonstrate the advantages of our method over traditional computer vision algorithm and classical CNN.

Readmission is often an important indicator of care quality, which also accounts for a major proportion of medical expenses. In this paper, we study the relationship between out-of-pocket ratio in medical cost and readmission. Apart from out-of-pocket ratio, we also consider other factors such as demographic features of patients, medical expenses, and variables about diseases. As there are a large number of diseases and operations, for better interpretation, we adopt data mining method to identify discriminative features. Our study is based on the front sheet of inpatient medical record from 150 hospitals in Beijing from year 2012 to 2014. In the records with primary diagnoses being rectal malignant tumor, we find out that when out-of-pocket ratio is low or high, the readmission is relatively high. Meanwhile, discriminative features found based on group difference mining method are helpful to gain understanding of the results.

Finding hidden topics and latest topic influential papers in a corpus can help researchers get a quick overview and recent development of a scientific research field. Existing work focused on finding milestone papers which are usually published many years ago. Finding latest influential papers is a more challenging problem due to lack of enough citation information of newly published papers. In this paper, we study this problem and propose a novel way of modeling citation links with a probabilistic generative model. The key idea is to consider two views of citation, both citing and being cited of each paper. Through this idea, we can not only model topic dependence between cited and citing papers but also incorporate latest papers into our model. Based on these ideas, we jointly model the two views with an extension of topic model, Bi-Citation-LDA model, which can not only find previous important papers but also find newly published influential papers in each topic. Experiments on real dataset and comparison with existing methods indicate that our model can effectively find latest topic influential papers.

This study compares the influence of different kinds of highlighting words beneath icons on visual search performance on the interface of tablets through response time and accurate rates of participants completing a search task. The results indicate highlighting words below icons could improve performance of searching a target icon; and when the icons are gray, under the condition of color words and flicker words is more effective for visual search. When the icons are colorful, under the condition of flicker words is more effective for visual search.

The volume of unstructured data has been growing sharply as the era of Big Data arrives. Decision tree is one of the most widely used classification models designed for structured data. Unstructured data such as text need to be converted to structured format before being analyzed using decision tree model. In this paper, we discuss how to construct decision trees for datasets containing unstructured data. For that purpose, a decision tree construction algorithm called CUST was proposed, which can directly tackle unstructured data. CUST introduces the use of splitting criteria formed by unstructured attribute values, and reduces the number of scans on datasets by designing appropriate data structures. Experiments on real-world datasets show that CUST improves the efficiency of building classifiers for unstructured data and performs as well as, if not better than existing solutions in classification accuracy.

Meaning extension is an important mechanism by virtue of which new idioms are produced. Guided by the Principle of Economy, language users reanalyze the meanings of original idiomatic expressions, their semantic focus and the construction of their components, thus get a different explanation of the original idioms based on a new understanding of their literal meanings, which eventually lead to the semantic evolution.

The boom of online social platforms of all kinds has triggered tremendous research interest in using social network data for user profiling, which refers to deriving labels for users that characterize their various aspects. Among different kinds of user profiling approaches, one line of work has taken advantage of the high level of label similarity that is often observed among users in one’s friendship network. In this work, we identify one critical point that has been so far neglected — different users in one’s friendship network play different roles in user profiling. In particular, we categorize all users in one’s friendship network into (I) close friends whom the user knows in real life and (II) online friends with whom the user forms connection through online interaction. We propose an algorithm that is affinity-aware in inferring users’ labels through network propagation. Our divide-and-conquer framework makes the proposed method scalable to large social network data. The experiment results in three real-world datasets demonstrate the superiority of our algorithm over baselines and support our argument for affinity-awareness in label profiling.