A5042921648- Recommender Systems and Techniques
- Topic Modeling
- Consumer Market Behavior and Pricing
- Image and Video Quality Assessment
- Advanced Graph Neural Networks
- Machine Learning in Healthcare
- Advanced Bandit Algorithms Research
- Sentiment Analysis and Opinion Mining
- Computational Drug Discovery Methods
- Advanced Text Analysis Techniques
- Hydrocarbon exploration and reservoir analysis
- Explainable Artificial Intelligence (XAI)
- Forecasting Techniques and Applications
- Drilling and Well Engineering
- Robot Manipulation and Learning
- Traffic Prediction and Management Techniques
- Data Stream Mining Techniques
- Reinforcement Learning in Robotics
- Data Management and Algorithms
- Hydraulic Fracturing and Reservoir Analysis
- Biomedical Text Mining and Ontologies
- Building Energy and Comfort Optimization
- Time Series Analysis and Forecasting
- Traditional Chinese Medicine Analysis
- Text and Document Classification Technologies
First Teaching Hospital of Tianjin University of Traditional Chinese Medicine
2024
Alibaba Group (China)
2023
Peking University
2019-2021
Ministry of Education
2019
Institute of Software
2018-2019
China University of Petroleum, Beijing
2017
Recommender systems have been playing an increasingly important role in our daily life due to the explosive growth of information. Accuracy and explainability are two core aspects when we evaluate a recommendation model become one fundamental trade-offs machine learning. In this paper, propose alleviate trade-off between accuracy by developing explainable deep that combines advantages learning-based models existing methods. The basic idea is build initial network based on hierarchy (e.g.,...
Learning user representation is a critical task for recommendation systems as it can encode preference personalized services. User generally learned from behavior data, such clicking interactions and review comments. However, less popular domains, the data insufficient to learn precise representations. To deal with this problem, natural thought leverage content-rich domains complement Inspired by recent success of BERT in NLP, we propose novel pre-training fine-tuning based approach U-BERT....
Diagnosis prediction, which aims to predict future health information of patients from historical electronic records (EHRs), is a core research task in personalized healthcare. Although some RNN-based methods have been proposed model sequential EHR data, these two major issues. First, they cannot capture fine-grained progression patterns patient conditions. Second, do not consider the mutual effect between important context (e.g., demographics) and diagnosis. To tackle challenges, we propose...
Drug-drug interactions (DDIs) are a major cause of preventable hospitalizations and deaths. Recently, researchers in the AI community try to improve DDI prediction two directions, incorporating multiple drug features better model pharmacodynamics adopting multi-task learning exploit associations among types. However, these directions challenging reconcile due sparse nature labels which inflates risk overfitting models when features. In this paper, we propose semi-supervised framework MLRDA...
Existing review-based recommendation models mainly learn long- term user and item representations from a set of reviews. Due to the ignorance rich side information reviews, these suffer two drawbacks: 1) they fail capture short-term changes preferences features reflected in reviews 2) cannot accurately model high-order user-item collaborative signals To overcome limitations, we propose multi-view approach named Set-Sequence-Graph (SSG), augment existing single-view (i.e., view set) methods...
Despite the development of ranking optimization techniques, pointwise loss remains dominating approach for click-through rate prediction. It can be attributed to calibration ability since prediction viewed as click probability. In practice, a CTR model is also commonly assessed with ability. To optimize ability, (e.g., pairwise or listwise loss) adopted they usually achieve better rankings than loss. Previous studies have experimented direct combination two losses obtain benefit from both...
Abstract Background In Traditional Chinese Medicine (TCM) theory, cold dampness obstruction is one of the common syndromes osteoarthritis. Therefore, in clinical practice, main treatment methods are to dispel wind, remove dampness, and dissipate cold, used treat knee osteoarthritis (KOA). This report describes a mulitercenter study assess Zhuifeng Tougu Capsule’s efficacy safety patients who syndrome KOA, provide evidence-based medical for rational use Capsules practice. Methods randomized,...
Summary Lithology identification is one of the keys to understand nature hydrocarbon reservoir. Deep learning has become a popular and reliable method for image classification in other fields. Instead using ordinary neural networks conventional logging curves, this paper developed deep methods showed that it possible identify lithology, results from borehole logs. In work, Convolutional Neural Network (CNN), which consists two convolutional layers, pooling layers fully-connected layer,...
Click-Through Rate (CTR) prediction serves as a fundamental component in online advertising. A common practice is to train CTR model on advertisement (ad) impressions with user feedback. Since ad are purposely selected by the itself, their distribution differs from inference and thus exhibits sample selection bias (SSB) that affects performance. Existing studies SSB mainly employ re-weighting techniques which suffer high variance poor calibration. Another line of work relies costly uniform...
In modern cities, complaining has become an important way for citizens to report emerging urban issues governments quick response. For ease of retrieval and handling, government officials usually organize citizen complaints by manually assigning tags them, which is inefficient cannot always guarantee the quality assigned tags. This work attempts solve this problem recommending complaints. Although there exist many studies on tag recommendation textual content, few them consider two...
Cascading architecture has been widely adopted in large-scale advertising systems to balance efficiency and effectiveness. In this architecture, the pre-ranking model is expected be a lightweight approximation of ranking model, which handles more candidates with strict latency requirements. Due gap capacity, models usually generate inconsistent ranked results, thus hurting overall system The paradigm score alignment proposed regularize their raw scores consistent. However, it suffers from...
Drug-drug interactions (DDIs) are a major cause of preventable hospitalizations and deaths. Predicting the occurrence DDIs helps drug safety professionals allocate investigative resources take appropriate regulatory action promptly. Traditional DDI prediction methods predict based on similarity between drugs. Recently, researchers revealed that predictive performance can be improved by better modeling pairs with bilinear forms. However, shallow models leveraging forms suffer from limitations...
We study two user demands that are important during the exploitation of explanations in practice: 1) understanding overall model behavior faithfully with limited cognitive load and 2) predicting accurately on unseen instances. illustrate correspond to major sub-processes human process propose a unified framework fulfill them simultaneously. Given local explanation method, our jointly learns number groupwise interpret most instances high fidelity specifies region where each applies....
Cascading architecture has been widely adopted in large-scale advertising systems to balance efficiency and effectiveness. In this architecture, the pre-ranking model is expected be a lightweight approximation of ranking model, which handles more candidates with strict latency requirements. Due gap capacity, models usually generate inconsistent ranked results, thus hurting overall system The paradigm score alignment proposed regularize their raw scores consistent. However, it suffers from...
Click-Through Rate (CTR) prediction serves as a fundamental component in online advertising. A common practice is to train CTR model on advertisement (ad) impressions with user feedback. Since ad are purposely selected by the itself, their distribution differs from inference and thus exhibits sample selection bias (SSB) that affects performance. Existing studies SSB mainly employ re-weighting techniques which suffer high variance poor calibration. Another line of work relies costly uniform...
Time series prediction is of great significance in many applications and has attracted extensive attention from the data mining community. Existing work suggests that for problems, shape current time may correlate an upcoming same or another series. Therefore, it a promising strategy to associate two recurring patterns as rule's antecedent consequent: occurrence can foretell consequent, learned consequent will give accurate predictions. Earlier employs symbolization methods, but symbolized...
Despite the development of ranking optimization techniques, pointwise loss remains dominating approach for click-through rate prediction. It can be attributed to calibration ability since prediction viewed as click probability. In practice, a CTR model is also commonly assessed with ability. To optimize ability, (e.g., pairwise or listwise loss) adopted they usually achieve better rankings than loss. Previous studies have experimented direct combination two losses obtain benefit from both...
Reinforcement learning often suffer from the sparse reward issue in real-world robotics problems. Learning demonstration (LfD) is an effective way to eliminate this problem, which leverages collected expert data aid online learning. Prior works assume that agent and aim accomplish same task, requires collecting new for every task. In paper, we consider case where target task mismatched but similar with of expert. Such setting can be challenging found existing LfD methods not effectively...