Predicting biomedical interactions is crucial for understanding various biological processes and drug discovery. Graph neural networks (GNNs) are promising in identifying novel interactions when extensive labeled data are available. However, labeling biomedical interactions is often time-consuming and labor-intensive, resulting in low-data scenarios. Furthermore, distribution shifts between training and test data in real-world applications pose a challenge to the generalizability of GNN models. Recent studies suggest that pre-training GNN models with self-supervised learning on unlabeled data can enhance their performance in predicting biomedical interactions. Here, we propose LukePi, a novel self-supervised pre-training framework that pre-trains GNN models on biomedical knowledge graphs (BKGs). LukePi is trained with two self-supervised tasks: topology-based node degree classification and semantics-based edge recovery. The former is to predict the degree of a node from its topological context and the latter is to infer both type and existence of a candidate edge by learning semantic information in the BKG. By integrating the two complementary tasks, LukePi effectively captures the rich information from the BKG, thereby enhancing the quality of node representations. We evaluate the performance of LukePi on two critical link prediction tasks: predicting synthetic lethality and drug-target interactions, using four benchmark datasets. In both distribution-shift and low-data scenarios, LukePi significantly outperforms 22 baseline models, demonstrating the power of the graph pre-training strategy when labeled data are sparse.

SUMMARY: Synthetic lethality (SL) refers to a type of genetic interaction in which the simultaneous inactivation of two genes leads to cell death, while the inactivation of a single gene does not affect cell viability. It significantly expands the range of potential therapeutic targets for anti-cancer treatments. SL interactions are primarily identified through experimental screening and computational prediction. Although various computational methods have been proposed, they tend to ignore providing evidence to support their predictions of SL. Besides, they are rarely user-friendly for biologists who likely have limited programming skills. Moreover, the genetic context specificity of SL interactions is often not taken into consideration. Here, we introduce a web server called SL-Miner, which is designed to mine the evidence of SL relationships between a primary gene and a few candidate SL partner genes in a specific type of cancer, and to prioritize these candidate genes by integrating various types of evidence. For intuitive data visualization, SL-Miner provides a range of charts (e.g. volcano plot and box plot) to help users get insights from the data. AVAILABILITY AND IMPLEMENTATION: SL-Miner is available at https://slminer.sist.shanghaitech.edu.cn.

MOTIVATION: Synthetic lethality (SL) is a type of genetic interaction in which the simultaneous inactivation of two genes leads to cell death, while the inactivation of a single gene does not affect the cell viability. It can effectively expand the range of anti-cancer therapeutic targets. SL interactions are identified mainly by experimental screening and computational prediction. Recent machine-learning methods mostly learn the representation of each gene individually, ignoring the representation of the pairwise interaction between two genes. In addition, the mechanisms of SL, the key to translating SL into cancer therapeutics, are often unclear. RESULTS: To fill the gaps, we propose a pairwise interaction learning-based graph neural network (GNN) named PiLSL to learn the representation of pairwise interaction between two genes for SL prediction. First, we construct an enclosing graph for each pair of genes from a knowledge graph. Secondly, we design an attentive embedding propagation layer in a GNN to discriminate the importance among the edges in the enclosing graph and to learn the latent features of the pairwise interaction from the weighted enclosing graph. Finally, we further fuse the latent features with explicit features extracted from multi-omics data to obtain powerful gene representations for SL prediction. Extensive experimental results demonstrate that PiLSL outperforms the best baseline by a large margin and generalizes well under three realistic scenarios. Besides, PiLSL provides an explanation of SL mechanisms via the weighted paths in the enclosing graphs by attention mechanism. AVAILABILITY AND IMPLEMENTATION: Our source code is available at https://github.com/JieZheng-ShanghaiTech/PiLSL.

MOTIVATION: Detecting synthetic lethality (SL) is a promising strategy for identifying anti-cancer drug targets. Targeting SL partners of a primary gene mutated in cancer is selectively lethal to cancer cells. Due to high cost of wet-lab experiments and availability of gold standard SL data, supervised machine learning for SL prediction has been popular. However, most of the methods are based on binary classification and thus limited by the lack of reliable negative data. Contrastive learning can train models without any negative sample and is thus promising for finding novel SLs. RESULTS: We propose NSF4SL, a negative-sample-free SL prediction model based on a contrastive learning framework. It captures the characteristics of positive SL samples by using two branches of neural networks that interact with each other to learn SL-related gene representations. Moreover, a feature-wise data augmentation strategy is used to mitigate the sparsity of SL data. NSF4SL significantly outperforms all baselines which require negative samples, even in challenging experimental settings. To the best of our knowledge, this is the first time that SL prediction is formulated as a gene ranking problem, which is more practical than the current formulation as binary classification. NSF4SL is the first contrastive learning method for SL prediction and its success points to a new direction of machine-learning methods for identifying novel SLs. AVAILABILITY AND IMPLEMENTATION: Our source code is available at https://github.com/JieZheng-ShanghaiTech/NSF4SL. SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.

The COVID-19 pandemic has spread globally for several months. Because its transmissibility and high pathogenicity seriously threaten people's lives, it is crucial to accurately and quickly detect COVID-19 infection. Many recent studies have shown that deep learning (DL) based solutions can help detect COVID-19 based on chest CT scans. However, most existing work focuses on 2D datasets, which may result in low quality models as the real CT scans are 3D images. Besides, the reported results span a broad spectrum on different datasets with a relatively unfair comparison. In this paper, we first use three state-of-the-art 3D models (ResNet3D101, DenseNet3D121, and MC3\_18) to establish the baseline performance on three publicly available chest CT scan datasets. Then we propose a differentiable neural architecture search (DNAS) framework to automatically search the 3D DL models for 3D chest CT scans classification and use the Gumbel Softmax technique to improve the search efficiency. We further exploit the Class Activation Mapping (CAM) technique on our models to provide the interpretability of the results. The experimental results show that our searched models (CovidNet3D) outperform the baseline human-designed models on three datasets with tens of times smaller model size and higher accuracy. Furthermore, the results also verify that CAM can be well applied in CovidNet3D for COVID-19 datasets to provide interpretability for medical diagnosis. Code: https://github.com/HKBU-HPML/CovidNet3D.

Deep learning has been successfully applied to learn local feature descriptors in recent years. However, most of existing methods are supervised methods relying on a large number of labeled training patches, which are also proposed for learning real valued descriptors. In this paper, we propose a novel unsupervised deep learning method for binary descriptor learning. The binary descriptors are much more compact and efficient than the real valued descriptors and unsupervised leaning is highly required in many applications due to its label-free characteristic as the annotations are sometimes expensive to obtain. The core idea of our method is to explore the locality consistency in the descriptor space as well as to distinguish different patches while maintaining the ability to match a patch with its geometric transformed ones. We also give a theorical analysis about the role of batch normalization in learning effective binary descriptors. Benefited from this analysis, there is no need to append two additional losses on minimizing the quantization error and maximizing the entropy to the final learning objective like previous works did, thus simplifying our network training. Experiments on four benchmarks demonstrate that the proposed method is able to learn binary descriptors significantly outperforming previous unsupervised binary descriptors, even superior to most supervised ones. Especially, it obtains 21.2% of improvement on the UBC Phototour dataset, and 19.8%, 26.7%, 26.0% of improvements for patch verification, matching, retrieval tasks respectively on the HPatches dataset compared to the previous best unsupervised method.

3D points distribute in a continuous 3D space irregularly, thus directly adapting 2D image convolution to 3D points is not an easy job. Previous works often artificially divide the space into regular grids, yet it could be suboptimal to learn geometry. In this paper, we propose SPCNN, namely, Space Probing Convolutional Neural Network, which naturally generalizes image CNN to deal with point clouds. The key idea of SPCNN is learning to probe the 3D space in an adaptive manner. Specifically, we define a pool of learnable convolutional weights, and let each point in the local region learn to choose a suitable convolutional weight from the pool. This is achieved by constructing a geometry guided index-mapping function that implicitly establishes a correspondence between convolutional weights and some local regions in the neighborhood (Fig. 1). In this way, the index-mapping function learns to adaptively partition nearby space for local geometry pattern recognition. With this convolution as a basic operator, SPCNN, a hierarchical architecture can be developed for effective point cloud analysis. Extensive experiments on challenging benchmarks across three tasks demonstrate that SPCNN achieves the state-of-the-art or has competitive performance.

Online learning for imbalanced streaming data is an important and challenging problem for many classification tasks in the machine learning research field. Traditional online learning algorithms are mainly focused on classification tasks with balanced data, and with little consideration about the characteristics of imbalanced streaming data. In this paper, we propose a novel online learning algorithm called Truncated Gradient Confidence-Weighted (TGCW), which integrate the truncated gradient algorithm with the confidence weighted algorithm together to improve the feature selection ability while reducing the dimensions of imbalanced streaming data effectively. We study a number of classification tasks with various imbalance data ratio including the pedestrian detection application and compare the performance of the TGCW algorithm with traditional online learning algorithms, and empirical results show that the TGCW algorithm can achieve better performance consistently than other baseline approaches.

We advocate for and presentTourSense, a framework for tourist identification and preference analytics using city-scale transport data (bus, subway, etc.). Our work is motivated by the observed limitations of utilizing traditional data sources (e.g., social media data and survey data) that commonly suffer from the limited coverage of tourist population and unpredictable information delay.TourSensedemonstrates how the transport data can overcome these limitations and provide better insights for different stakeholders, typically including tour agencies, transport operators, and tourists themselves. Specifically, we first propose a graph-based iterative propagation learning algorithm to recognize tourists from public commuters. Taking advantage of the trace data from the identified tourists, we then design a tourist preference analytics model to learn and predict their next tour, where an interactive user interface is implemented to ease the information access and gain the insights from the analytics results. Experiments with real-world datasets (from over 5.1 million commuters and their 462 million trips) show the promise and effectiveness of the proposed framework: the Macro and Micro F1 scores of the tourist identification system achieve 0.8549 and 0.7154, respectively, whereas the tourist preference analytics system improves the baselines by at least 23.53 and 11.44 percent in terms of precision and recall.

Social recommender systems exploit users' social relationships to improve recommendation accuracy. Intuitively, a user tends to trust different people regarding with different scenarios. Therefore, one main challenge of social recommendation is to exploit the most appropriate dependencies between users for a given recommendation task. Previous social recommendation methods are usually developed based on pre-defined user dependencies. Thus, they may not be optimal for a specific recommendation task. In this paper, we propose a novel recommendation method, named probabilistic relational matrix factorization (PRMF), which can automatically learn the dependencies between users to improve recommendation accuracy. In PRMF, users' latent features are assumed to follow a matrix variate normal (MVN) distribution. Both positive and negative user dependencies can be modeled by the row precision matrix of the MVN distribution. Moreover, we also propose an alternating optimization algorithm to solve the optimization problem of PRMF. Extensive experiments on four real datasets have been performed to demonstrate the effectiveness of the proposed PRMF model.

In this paper, we propose a generic framework to learn context-aware latent representations for context-aware collaborative filtering. Contextual contents are combined via a function to produce the context influence factor, which is then combined with each latent factor to derive latent representations. We instantiate the generic framework using biased Matrix Factorization as the base model. A Stochastic Gradient Descent (SGD) based optimization procedure is developed to fit the model by jointly learning the weight of each context and latent factors. Experiments conducted over three real-world datasets demonstrate that our model significantly outperforms not only the base model but also the representative context-aware recommendation models.

By leveraging crowdsourcing, Web credibility evaluation systems (WCESs) have become a promising tool to assess the credibility of Web content, e.g., Web pages. However, existing systems adopt a passive way to collect users' credibility ratings, which incurs two crucial challenges: (1) a considerable fraction of Web content have few or even no ratings, so the coverage (or effectiveness) of the system is low; (2) malicious users may submit fake ratings to damage the reliability of the system. In order to realize a highly effective and robust WCES, we propose to integrate recommendation functionality into the system. On the one hand, by fusing Matrix Factorization and Latent Dirichlet Allocation, a personalized Web content recommendation model is proposed to attract users to rate more Web pages, i.e., the coverage is increased. On the other hand, by analyzing a user's reaction to the recommended Web content, we detect imitating attackers, which have recently been recognized as a particular threat to WCES to make the system more robust. Moreover, an adaptive reputation system is designed to motivate users to more actively interact with the integrated recommendation functionality. We conduct experiments using both real datasets and synthetic data to demonstrate how our proposed recommendation components significantly improve the effectiveness and robustness of existing WCES.

Real world large-scale recommender systems are always dynamic: new users and items continuously enter the system, and the status of old ones (e.g., users' preference and items' popularity) evolve over time. In order to handle such dynamics, we propose a recommendation framework consisting of an online component and an offline component, where the newly arrived items are processed by the online component such that users are able to get suggestions for fresh information, and the influence of longstanding items is captured by the offline component. Based on individual users' rating behavior, recommendations from the two components are combined to provide top-N recommendation. We formulate recommendation problem as a ranking problem where learning to rank is applied to extend upon matrix factorization to optimize item rankings by minimizing a pairwise loss function. Furthermore, to better model interactions between users and items, Latent Dirichlet Allocation is incorporated to fuse rating information and textual information. Real data based experiments demonstrate that our approach outperforms the state-of-the-art models by at least 61.21% and 50.27% in terms of mean average precision (MAP) and normalized discounted cumulative gain (NDCG) respectively.

In many large‐scale distributed systems and on the Web, agents need to interact with other unknown agents to carry out some tasks or transactions. The ability to reason about and assess the potential risks in carrying out such transactions is essential for providing a safe and reliable interaction environment. A traditional approach to reason about the risk of a transaction is to determine if the involved agent is trustworthy on the basis of its behavior history. As a departure from such traditional trust models, we propose a generic, trust framework based on machine learning where an agent uses its own previous transactions (with other agents) to build a personal knowledge base. This is used to assess the trustworthiness of a transaction on the basis of the associated features, particularly using the features that help discern successful transactions from unsuccessful ones. These features are handled by applying appropriate machine learning algorithms to extract the relationships between the potential transaction and the previous ones. Experiments based on real data sets show that our approach is more accurate than other trust mechanisms, especially when the information about past behavior of the specific agent is rare, incomplete, or inaccurate.

With the prosperity of cloud computing, an increasing number of Small and Medium-sized Enterprises (SMEs) move their business to public clouds such as Amazon EC2. To help tenants deploy services in the cloud, researchers either conduct performance evaluations or design mechanisms and software on seeking virtual machines of better performance. However, few studies have investigated the impact of instance seeking strategies on resource allocation in clouds if every tenant starts to apply the same method to find the better performing virtual machine. In this paper, we propose a cloud and a tenant model in order to simulate the process of tenants' seeking better-performing instances in the cloud. We discuss, implement and evaluate six cloud resource allocation strategies and five instance seeking strategies. We perform the evaluation via simulation based on real data traces. Our results show that instance seeking strategies can cause the exhaustion of better-performing instances and significant request growth in the cloud. Furthermore, we find that tenants could save time and budget through collaborative seeking strategies. Finally, we discuss the implications of our findings from perspectives of both tenants and providers.

Location-based social networks (LBSNs) offer researchers rich data to study people's online activities and mobility patterns. One important application of such studies is to provide personalized point-of-interest (POI) recommendations to enhance user experience in LBSNs. Previous solutions directly predict users' preference on locations but fail to provide insights about users' preference transitions among locations. In this work, we propose a novel category-aware POI recommendation model, which exploits the transition patterns of users' preference over location categories to improve location recommendation accuracy. Our approach consists of two stages: (1) preference transition (over location categories) prediction, and (2) category-aware POI recommendation. Matrix factorization is employed to predict a user's preference transitions over categories and then her preference on locations in the corresponding categories. Real data based experiments demonstrate that our approach outperforms the state-of-the-art POI recommendation models by at least 39.75% in terms of recall.

Contexts and social network information have been proven to be valuable information for building accurate recommender system. However, to the best of our knowledge, no existing works systematically combine diverse types of such information to further improve recommendation quality. In this paper, we propose SoCo, a novel context-aware recommender system incorporating elaborately processed social network information. We handle contextual information by applying random decision trees to partition the original user-item-rating matrix such that the ratings with similar contexts are grouped. Matrix factorization is then employed to predict missing preference of a user for an item using the partitioned matrix. In order to incorporate social network information, we introduce an additional social regularization term to the matrix factorization objective function to infer a user's preference for an item by learning opinions from his/her friends who are expected to share similar tastes. A context-aware version of Pearson Correlation Coefficient is proposed to measure user similarity. Real datasets based experiments show that SoCo improves the performance (in terms of root mean square error) of the state-of-the-art context-aware recommender system and social recommendation model by 15.7% and 12.2% respectively.

Modeling trust in complex dynamic environments is an important yet challenging issue since an intelligent agent may strategically change its behavior to maximize its profits. In thispaper, we propose a context aware trust model to predict dynamic trust by using a Hidden Markov Model (HMM) to model an agent's interactions. Although HMMs have already been applied in the past to model an agent's dynamic behavior to greatly improve the traditional static probabilistic trust approaches, most HMM based trust models only focus on outcomes of the past interactions without considering interaction context, which we believe, reflects immensely on the dynamic behavior or intent of an agent. Interaction contextual information is comprehensively studied and integrated into the model to more precisely approximate an agent's dynamic behavior. Evaluation using real auction data and synthetic data demonstrates the efficacy of our approach in comparison with previous state-of-the-art trust mechanisms.

Reputation systems deployed in popular online auction sites simply aggregate feedback about a seller's past transactions. By studying a real auction site dataset, we infer that a non-negligible fraction of unsatisfactory transactions involve sellers with high reputation. Such a phenomenon can be interpreted by motivation theory from behaviorial science: A seller with high reputation has more business opportunities. Bad feedback for latest transactions do not immediately affect his reputation adequately to hurt business, hence he may not be as prudent as before. In this work, we propose the concept of imprudence to study and detect the inappropriate behavior of a 'reliable' seller (i.e., the one with high reputation computed using conventional approaches). Specifically, we first identify and verify the features that influence a seller's imprudence behavior. We then design a novel intelligent buying agent to combine these factors using logistic regression for predicting and studying the probability of imprudence of a target seller. We validate our approach using real datasets driven experiments.

Cloud computing has emerged as a popular paradigm that offers computing resources (e.g. CPU, storage, bandwidth, software) as scalable and on-demand services over the Internet. As more players enter this emerging market, a heterogeneous cloud computing market is expected to evolve, where individual

Using peer-to-peer overlays to notify users whenever a new update occurs is a promising approach to support Web based publish subscribe systems like really simple syndication (RSS). Such a peer-to-peer approach can scale well by reducing load at the source and also guarantee timeliness of notifications. Several such overlay based approaches have been proposed in recent years. However, malicious peers may pretend to relay but actually not, and thus deny service, or even propagate counterfeit updates - thus rendering a peer-to-peer mechanism not only useless, but even harmful (e.g., by false updates). We propose overlay independent randomized strategies to mitigate these ill-effects of malicious peers at a marginal overhead, thus enjoying the benefits of peer-to-peer dissemination, along with the assurance of content integrity in RSS like Web-based publish-subscribe applications without altering currently deployed server infrastructure.

Trust plays important roles in diverse decentralized environments, including our society at large. Computational trust models help to, for instance, guide users' judgements in online auction sites about other users; or determine quality of contributions in web 2.0 sites. Most of the existing trust models, however, require historical information about past behavior of a specific agent being evaluated - information that is not always available. In contrast, in real life interactions among users, in order to make the first guess about the trustworthiness of a stranger, we commonly use our "instinct" - essentially stereotypes developed from our past interactions with "similar" people. We propose StereoTrust, a computational trust model inspired by real life stereotypes. A user forms stereotypes using her previous transactions with other agents. A stereotype contains certain features of agents and an expected outcome of the transaction. These features can be taken from agents' profile information, or agents' observed behavior in the system. When facing a stranger, the stereotypes matching stranger's profile are aggregated to derive his expected trust. Additionally, when some information about stranger's previous transactions is available, StereoTrust uses it to refine the stereotype matching. According to our experiments, StereoTrust compares favorably with existing trust models that use different kind of information and more complete historical information. Moreover, because evaluation is done according to user's personal stereotypes, the system is completely distributed and the result obtained is personalized. StereoTrust can be used as a complimentary mechanism to provide the initial trust value for a stranger, especially when there is no trusted, common third parties.