Ruoran Liu
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Previous Organizations / Institutes
The organizations / institutes from the scholar's other articles in the past ten years
Research Interests
Ruoran Liu has a total of 41 co-authors. Here are the top co-authors by number of shared publications:
- Daniel Dajun Zeng
State Key Laboratory of Multimodal Artificial Intelligence Systems, Institute of Automation, Chinese Academy of Sciences, Beijing 100190, China
6 shared publications - Qiudan Li
Institute of Automation, Chinese Academy of Sciences, Beijing, China
6 shared publications - Scott James Leischow
College of Health Solutions, Arizona State University, Phoenix, AZ, United States
3 shared publications - Can Wang
The State Key Laboratory of Management and Control for Complex Systems, Institute of Automation, Chinese Academy of Sciences, Beijing, China
2 shared publications - Lei Wang
Chinese Academy of Sciences
2 shared publications - Riheng Yao
The State Key Laboratory of Management and Control for Complex Systems, Institute of Automation, Chinese Academy of Sciences, Beijing, China
2 shared publications
Latest Publications
- Open Access
- 8 August 2026
- Published by Association for Computing Machinery (ACM)
- p. 7477-7488
GPR: Towards a Generative Pre-trained One-Model Paradigm for Large-Scale Advertising Recommendation
Online advertising recommendation is the economic engine of the digital ecosystem, yet traditional multi-stage cascading systems suffer from objective misalignment and error propagation. Conversely, emerging generative models often struggle with the strict latency and revenue alignment requirements of industrial applications. To address these issues, we propose GPR (Generative Pre-trained Recommender). Instead of the complex retrieval and pre-ranking cascade, GPR employs a unified generative framework to directly generate candidate sets, which are then refined by a lightweight fine-ranking stage for precise business alignment. Unlike prior works, GPR introduces three key technical innovations to ensure industrial feasibility: (1) Unified Representation: We utilize RQ-KMeans+ to map heterogeneous user behaviors and ad content into a shared semantic ID space; (2) Decoupled Architecture: We design a Heterogeneous Hierarchical Decoder (HHD) to decouple user intent modeling from ad generation, enabling efficient long-horizon inference; (3) Value-Aligned Generation: We propose a Value-Guided Trie-Constrained Beam Search and a hierarchical training strategy (MTP + HEPO) to strictly enforce business constraints and maximize commercial value. GPR has been fully deployed in Tencent Weixin Channels, serving over 500 million daily active users. Extensive online A/B testing demonstrates that GPR delivers substantial gains in GMV and CTCVR over highly optimized cascading baselines, validating the generative paradigm as a superior solution for large-scale advertising systems.
- 8 August 2026
- Published by Association for Computing Machinery (ACM)
- p. 7477-7488
Open AccessGPR: Towards a Generative Pre-trained One-Model Paradigm for Large-Scale Advertising Recommendation
Online advertising recommendation is the economic engine of the digital ecosystem, yet traditional multi-stage cascading systems suffer from objective misalignment and error propagation. Conversely, emerging generative models often struggle with the strict latency and revenue alignment requirements of industrial applications. To address these issues, we propose GPR (Generative Pre-trained Recommender). Instead of the complex retrieval and pre-ranking cascade, GPR employs a unified generative framework to directly generate candidate sets, which are then refined by a lightweight fine-ranking stage for precise business alignment. Unlike prior works, GPR introduces three key technical innovations to ensure industrial feasibility: (1) Unified Representation: We utilize RQ-KMeans+ to map heterogeneous user behaviors and ad content into a shared semantic ID space; (2) Decoupled Architecture: We design a Heterogeneous Hierarchical Decoder (HHD) to decouple user intent modeling from ad generation, enabling efficient long-horizon inference; (3) Value-Aligned Generation: We propose a Value-Guided Trie-Constrained Beam Search and a hierarchical training strategy (MTP + HEPO) to strictly enforce business constraints and maximize commercial value. GPR has been fully deployed in Tencent Weixin Channels, serving over 500 million daily active users. Extensive online A/B testing demonstrates that GPR delivers substantial gains in GMV and CTCVR over highly optimized cascading baselines, validating the generative paradigm as a superior solution for large-scale advertising systems.
- 7 February 2023
- Published by Wiley in Laser & Photonics Reviews
- Vol. 17 (5)
High‐Performance Mode‐Multiplexing Device with Anisotropic Lithium‐Niobate‐on‐Insulator Waveguides
Thin‐film lithium‐niobate photonics has attracted intensive attention due to its superior material properties. In particular, multimode lithium‐niobate photonics is becoming attractive for the applications of high‐capacity optical interconnects and high‐efficiency nonlinear optics. However, multimode manipulation in lithium‐niobate‐on‐insulator (LNOI) waveguides is not easy because there might occur some mode hybridness due to the vertical asymmetry as well as the material anisotropy. In this paper, high‐performance mode‐multiplexing devices with anisotropic LNOI waveguides are proposed, consisting of a multi‐channel mode (de)multiplexer and a sharp multimode waveguide bend (MWB). The mode (de)multiplexer is designed with Z‐propagation LNOI waveguides to avoid the mode hybridness. Meanwhile, the sharp MWB is designed to connect with a Y‐propagation waveguide, enabling access to the highest electro‐optic coefficient and second‐order nonlinear coefficient. The fabricated mode (de)multiplexer has a low excess loss less than 0.2 dB as well as low crosstalk less than ‐19 dB in the wavelength range of 1520–1600 nm. These high‐performance multimode photonic devices are useful for developing multimode LN photonics.
- 7 February 2023
- Published by Wiley in Laser & Photonics Reviews
- Vol. 17 (5)
High‐Performance Mode‐Multiplexing Device with Anisotropic Lithium‐Niobate‐on‐Insulator Waveguides
Thin‐film lithium‐niobate photonics has attracted intensive attention due to its superior material properties. In particular, multimode lithium‐niobate photonics is becoming attractive for the applications of high‐capacity optical interconnects and high‐efficiency nonlinear optics. However, multimode manipulation in lithium‐niobate‐on‐insulator (LNOI) waveguides is not easy because there might occur some mode hybridness due to the vertical asymmetry as well as the material anisotropy. In this paper, high‐performance mode‐multiplexing devices with anisotropic LNOI waveguides are proposed, consisting of a multi‐channel mode (de)multiplexer and a sharp multimode waveguide bend (MWB). The mode (de)multiplexer is designed with Z‐propagation LNOI waveguides to avoid the mode hybridness. Meanwhile, the sharp MWB is designed to connect with a Y‐propagation waveguide, enabling access to the highest electro‐optic coefficient and second‐order nonlinear coefficient. The fabricated mode (de)multiplexer has a low excess loss less than 0.2 dB as well as low crosstalk less than ‐19 dB in the wavelength range of 1520–1600 nm. These high‐performance multimode photonic devices are useful for developing multimode LN photonics.
- 24 April 2020
- Published by Institute for Operations Research and the Management Sciences (INFORMS) in INFORMS Journal on Computing
- Vol. 32 (4) , 996-1011
Understanding and Predicting Users’ Rating Behavior: A Cognitive Perspective
Online reviews are playing an increasingly important role in understanding and predicting users’ rating behavior, which brings great opportunities for users and organizations to make better decisio...
- 24 April 2020
- Published by Institute for Operations Research and the Management Sciences (INFORMS) in INFORMS Journal on Computing
- Vol. 32 (4) , 996-1011
Understanding and Predicting Users’ Rating Behavior: A Cognitive Perspective
Online reviews are playing an increasingly important role in understanding and predicting users’ rating behavior, which brings great opportunities for users and organizations to make better decisio...
- Open Access
- 4 March 2019
- Published by JMIR Publications Inc. in JMIR Preprints
A Deep Learning Method to Analyze and Classify Happy Moments: A Comparative Analytic Study (Preprint)
BACKGROUND Happiness is considered as an important indicator of users’ mental and physical health. Fostering happiness has gained increasing public attention as one of the ways to decrease health costs in the long run. Understanding what makes users feel happy may help policy makers develop policies and methods that steer users towards behaviors identified to promote happiness. OBJECTIVE This paper aimed to investigate the use of deep learning methods to analyze happy moments and compare them with the traditional machine learning methods, which may provide a mechanism to accurately classify happy moments and help understand why users feel happy. METHODS A crowdsourced corpus of happy moments, HappyDB, was used. The dataset contained 14,125 posts with category labels that described sources and reasons for happy feelings: Achievement, Affection, Bonding, Enjoy the moment, Leisure, Nature and Exercise. We compared the performance of deep learning methods such as the convolutional neural network (CNN), bidirectional long-short term memory (Bi-LSTM), and attention Bi-LSTM with that of the traditional machine learning methods including logistic regression, SVM, and naïve Bayes. Standard measures including precision, recall, and F1 were adopted for each category. Macro-precision, macro-recall, and macro-F1 were used to evaluate the overall performance of the models. RESULTS We found that CNN achieved the best performance on macro-precision, macro-recall, and macro-F1, with values of 80.8, 79.3, and 80.0, respectively. Among the traditional machine learning methods, logistic regression performed the best, with macro-precision of 80.6, macro-recall of 71.1, and macro-F1 of 75.5. A detailed comparison of CNN and logistic regression on each category showed that CNN was able to improve F1 score for all categories. Specifically, F1 improved by at least 1.8% on the Bonding category and up to 11.3% on Nature. Performance improvements mainly depended on significant improvements on recall, especially for minor categories. For example, the recall of CNN was 80.9 and 70.9 for Nature and Exercise, which was an improvement of 28.5% and 11.6% compared with logistic regression. The reason was that CNN explicitly modeled the relationship between word features and the categories of happy moments by extracting important word features through convolution and pooling operations. CONCLUSIONS This is the first study to analyze happy moments based on deep learning methods. Compared with the traditional machine learning methods, deep learning methods, especially CNN, showed superiority on classifying the happy moments, which would facilitate understanding of the reasons why users feel happy and thus help policy makers formulate targeted policies to promote happiness.
- 4 March 2019
- Published by JMIR Publications Inc. in JMIR Preprints
Open AccessA Deep Learning Method to Analyze and Classify Happy Moments: A Comparative Analytic Study (Preprint)
BACKGROUND Happiness is considered as an important indicator of users’ mental and physical health. Fostering happiness has gained increasing public attention as one of the ways to decrease health costs in the long run. Understanding what makes users feel happy may help policy makers develop policies and methods that steer users towards behaviors identified to promote happiness. OBJECTIVE This paper aimed to investigate the use of deep learning methods to analyze happy moments and compare them with the traditional machine learning methods, which may provide a mechanism to accurately classify happy moments and help understand why users feel happy. METHODS A crowdsourced corpus of happy moments, HappyDB, was used. The dataset contained 14,125 posts with category labels that described sources and reasons for happy feelings: Achievement, Affection, Bonding, Enjoy the moment, Leisure, Nature and Exercise. We compared the performance of deep learning methods such as the convolutional neural network (CNN), bidirectional long-short term memory (Bi-LSTM), and attention Bi-LSTM with that of the traditional machine learning methods including logistic regression, SVM, and naïve Bayes. Standard measures including precision, recall, and F1 were adopted for each category. Macro-precision, macro-recall, and macro-F1 were used to evaluate the overall performance of the models. RESULTS We found that CNN achieved the best performance on macro-precision, macro-recall, and macro-F1, with values of 80.8, 79.3, and 80.0, respectively. Among the traditional machine learning methods, logistic regression performed the best, with macro-precision of 80.6, macro-recall of 71.1, and macro-F1 of 75.5. A detailed comparison of CNN and logistic regression on each category showed that CNN was able to improve F1 score for all categories. Specifically, F1 improved by at least 1.8% on the Bonding category and up to 11.3% on Nature. Performance improvements mainly depended on significant improvements on recall, especially for minor categories. For example, the recall of CNN was 80.9 and 70.9 for Nature and Exercise, which was an improvement of 28.5% and 11.6% compared with logistic regression. The reason was that CNN explicitly modeled the relationship between word features and the categories of happy moments by extracting important word features through convolution and pooling operations. CONCLUSIONS This is the first study to analyze happy moments based on deep learning methods. Compared with the traditional machine learning methods, deep learning methods, especially CNN, showed superiority on classifying the happy moments, which would facilitate understanding of the reasons why users feel happy and thus help policy makers formulate targeted policies to promote happiness.
- Open Access
- 15 August 2018
- Published by JMIR Publications Inc. in Journal of Medical Internet Research
- Vol. 20 (8) , e252
Understanding Users’ Vaping Experiences from Social Media: Initial Study Using Sentiment Opinion Summarization Techniques
Journal of Medical Internet Research - International Scientific Journal for Medical Research, Information and Communication on the Internet
- 15 August 2018
- Published by JMIR Publications Inc. in Journal of Medical Internet Research
- Vol. 20 (8) , e252
Open AccessUnderstanding Users’ Vaping Experiences from Social Media: Initial Study Using Sentiment Opinion Summarization Techniques
Journal of Medical Internet Research - International Scientific Journal for Medical Research, Information and Communication on the Internet
- Open Access
- 10 November 2017
- Published by JMIR Publications Inc. in JMIR Preprints
Understanding Users Electronic Nicotine Delivery Systems Vaping Experience from Social Media Using Sentiment Opinion Summarization Techniques (Preprint)
BACKGROUND E-liquid is one of the main components in electronic nicotine delivery systems (ENDS). ENDS review comments could serve as an early warning on use patterns, and even function to serve as an indicator of problems or adverse events pertaining to the use of specific e-liquids – much like the types of responses that are tracked by the Food and Drug Administration (FDA) regarding medications. OBJECTIVE This paper aims to understand users’ vaping experience using sentiment opinion summarization techniques, which can help characterize how consumers, ENDS vendors, and others think about specific e-liquids and their characteristic (e.g. flavor, throat hit, vapor production, etc.). METHODS We collected e-liquid reviews on JuiceDB from June 27, 2013 to December 31, 2017 by using its public application programming interface (API). The dataset contains 27070 reviews for 8058 e-liquid products. Each review is accompanied by an overall rating and a set of four aspect ratings of an e-liquid, each on a scale of 1-5: flavor accuracy, throat hit, value and cloud production. An Iterative Dichotomiser 3 (ID3) based influential aspect analysis model was adopted to learn the key elements that impact e-liquid use. Then, fine-grained sentiment analysis was employed to mine opinions on various aspects of the vaping experience related to e-liquids. RESULTS We found that: (1) Flavor accuracy and value were the two most important aspects that influence users’ sentiments towards e-liquids. (2) 67.83% (18362/27070) of the reviews in JuiceDB were positive, while 12.67% (3430/27070) were negative. This indicates that users generally hold positive attitudes towards e-liquids. (3) Among the nine flavors, fruity and sweet were the two most popular flavors. (4) Great and sweet tastes, reasonable value, and strong throat hit made users satisfied with the fruity and sweet flavors, whereas weird tastes made users dislike these flavors. Meanwhile, users complained about steep/expensive price, bad quality and harsh throat hit of some e-liquids. (5) There were 2342 fruity e-liquids and 2049 sweet e-liquids. Positive sentiments towards fruity e-liquids and sweet e-liquids measured 55.81% and 59.83% respectively; negative sentiments were 13.62% and 12.88% respectively. Great flavors and good vapors contributed to positive reviews of fruity and sweet products. However, bad tastes such as “sour” or “bitter” resulted in negative reviews. These findings can help businesses and policy makers further improve product quality and formulate effective policy. CONCLUSIONS This is the first study to provide an effective mechanism to analyze users’ ENDS vaping experience based on sentiment opinion summarization techniques. Sentiment opinions on aspect and products can be found using our method, which is of great importance to monitor e-liquid products and improve work efficiency.
- 10 November 2017
- Published by JMIR Publications Inc. in JMIR Preprints
Open AccessUnderstanding Users Electronic Nicotine Delivery Systems Vaping Experience from Social Media Using Sentiment Opinion Summarization Techniques (Preprint)
BACKGROUND E-liquid is one of the main components in electronic nicotine delivery systems (ENDS). ENDS review comments could serve as an early warning on use patterns, and even function to serve as an indicator of problems or adverse events pertaining to the use of specific e-liquids – much like the types of responses that are tracked by the Food and Drug Administration (FDA) regarding medications. OBJECTIVE This paper aims to understand users’ vaping experience using sentiment opinion summarization techniques, which can help characterize how consumers, ENDS vendors, and others think about specific e-liquids and their characteristic (e.g. flavor, throat hit, vapor production, etc.). METHODS We collected e-liquid reviews on JuiceDB from June 27, 2013 to December 31, 2017 by using its public application programming interface (API). The dataset contains 27070 reviews for 8058 e-liquid products. Each review is accompanied by an overall rating and a set of four aspect ratings of an e-liquid, each on a scale of 1-5: flavor accuracy, throat hit, value and cloud production. An Iterative Dichotomiser 3 (ID3) based influential aspect analysis model was adopted to learn the key elements that impact e-liquid use. Then, fine-grained sentiment analysis was employed to mine opinions on various aspects of the vaping experience related to e-liquids. RESULTS We found that: (1) Flavor accuracy and value were the two most important aspects that influence users’ sentiments towards e-liquids. (2) 67.83% (18362/27070) of the reviews in JuiceDB were positive, while 12.67% (3430/27070) were negative. This indicates that users generally hold positive attitudes towards e-liquids. (3) Among the nine flavors, fruity and sweet were the two most popular flavors. (4) Great and sweet tastes, reasonable value, and strong throat hit made users satisfied with the fruity and sweet flavors, whereas weird tastes made users dislike these flavors. Meanwhile, users complained about steep/expensive price, bad quality and harsh throat hit of some e-liquids. (5) There were 2342 fruity e-liquids and 2049 sweet e-liquids. Positive sentiments towards fruity e-liquids and sweet e-liquids measured 55.81% and 59.83% respectively; negative sentiments were 13.62% and 12.88% respectively. Great flavors and good vapors contributed to positive reviews of fruity and sweet products. However, bad tastes such as “sour” or “bitter” resulted in negative reviews. These findings can help businesses and policy makers further improve product quality and formulate effective policy. CONCLUSIONS This is the first study to provide an effective mechanism to analyze users’ ENDS vaping experience based on sentiment opinion summarization techniques. Sentiment opinions on aspect and products can be found using our method, which is of great importance to monitor e-liquid products and improve work efficiency.
- Open Access
- 20 January 2017
- Published by JMIR Publications Inc. in Journal of Medical Internet Research
- Vol. 19 (1) , e24
Identifying Topics for E-Cigarette User-Generated Contents: A Case Study From Multiple Social Media Platforms
Journal of Medical Internet Research - International Scientific Journal for Medical Research, Information and Communication on the Internet
- 20 January 2017
- Published by JMIR Publications Inc. in Journal of Medical Internet Research
- Vol. 19 (1) , e24
Open AccessIdentifying Topics for E-Cigarette User-Generated Contents: A Case Study From Multiple Social Media Platforms
Journal of Medical Internet Research - International Scientific Journal for Medical Research, Information and Communication on the Internet
- 3 November 2016
- Published by Institute of Electrical and Electronics Engineers (IEEE)
- p. 2568-2575
Predicting user's multi-interests with network embedding in health-related topics
With the rapid growth of Web 2.0, social media has become a prevalent information sharing and seeking channel for health surveillance, in which users form interactive networks by posting and replying messages, providing and rating reviews, attending multiple discussion boards on health-related topics. Users' behaviors in these interactive networks reflect users' multiple interests. To provide better information service for users, it is necessary to analyze the user interactions and predict users' multi-interests. Most existing work in predicting users' multi-interests based on multi label network classification focuses on using approximate inference methods to leverage the dependency information to improve classification results. Inspired by deep learning techniques, DEEPWALK learns label independent latent representations of vertices in a network using local information obtained from truncated random walks, which provides an efficient way for predicting users multi-interests from user interactions. In this paper, we develop a user's multi-interests prediction model based on DEEPWALK, weight information of user interactions is considered when modeling a stream of short constrained random walks and SkipGram is employed to generate more accurate representations of user vertices, which help identify users' interests. Experimental results on two real world health-related datasets show the efficacy of the proposed model.
- 3 November 2016
- Published by Institute of Electrical and Electronics Engineers (IEEE)
- p. 2568-2575
Predicting user's multi-interests with network embedding in health-related topics
With the rapid growth of Web 2.0, social media has become a prevalent information sharing and seeking channel for health surveillance, in which users form interactive networks by posting and replying messages, providing and rating reviews, attending multiple discussion boards on health-related topics. Users' behaviors in these interactive networks reflect users' multiple interests. To provide better information service for users, it is necessary to analyze the user interactions and predict users' multi-interests. Most existing work in predicting users' multi-interests based on multi label network classification focuses on using approximate inference methods to leverage the dependency information to improve classification results. Inspired by deep learning techniques, DEEPWALK learns label independent latent representations of vertices in a network using local information obtained from truncated random walks, which provides an efficient way for predicting users multi-interests from user interactions. In this paper, we develop a user's multi-interests prediction model based on DEEPWALK, weight information of user interactions is considered when modeling a stream of short constrained random walks and SkipGram is employed to generate more accurate representations of user vertices, which help identify users' interests. Experimental results on two real world health-related datasets show the efficacy of the proposed model.