Ci Gao
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Research Subjects
Ci Gao has a total of 11 co-authors. Here are the top co-authors by number of shared publications:
- Aoyang Zhang
ByteDance, Shanghai, China
1 shared publication - Dan Zhao
Peng Cheng Laboratory, Beijing, China
1 shared publication - Dingmin Wang
University of Oxford, Oxford, United Kingdom
1 shared publication - Lingjia Gu
College of Electronic Science and Engineering, Jilin University, Changchun 130012, China
1 shared publication - Mingda Jiang
Jilin Univ. (China)
1 shared publication - Qian Ma
ByteDance Inc., Beijing, China
1 shared publication
Latest Publications
- Open Access
- 4 August 2023
- Published by Association for Computing Machinery (ACM)
- p. 4894-4903
Counterfactual Video Recommendation for Duration Debiasing
Duration bias widely exists in video recommendations, where models tend to recommend short videos for the higher ratio of finish playing and thus possibly fail to capture users' true interests. In this paper, we eliminate the duration bias from both data and model. First, based on the extensive data analysis, we observe that play completion rate of videos with the same duration presents a bimodal distribution. Hence, we propose to perform threshold division to construct binary labels as training labels for alleviating the drawback of finish playing labels overly biased towards short videos. Algorithmically, we resort to causal inference, which enables us to inspect causal relationships of video recommendations with a causal graph. We identify that duration has two kinds of effect on prediction: direct and indirect. Duration bias lies in the direct effect, while the indirect effect benefits prediction. To this end, we design a model-agnostic Counterfactual Video Recommendation for Duration Debiasing (CVRDD) framework, which incorporates multi-task learning to estimate different causal effect during training. In the inference phase, we perform counterfactual inference to remove the direct effect of duration for unbiased prediction. We conduct experiments on two industrial datasets, and in addition to achieving highly promising results on traditional top-k recommendation metrics, CVRDD also improves the user watch time.
- 4 August 2023
- Published by Association for Computing Machinery (ACM)
- p. 4894-4903
Open AccessCounterfactual Video Recommendation for Duration Debiasing
Duration bias widely exists in video recommendations, where models tend to recommend short videos for the higher ratio of finish playing and thus possibly fail to capture users' true interests. In this paper, we eliminate the duration bias from both data and model. First, based on the extensive data analysis, we observe that play completion rate of videos with the same duration presents a bimodal distribution. Hence, we propose to perform threshold division to construct binary labels as training labels for alleviating the drawback of finish playing labels overly biased towards short videos. Algorithmically, we resort to causal inference, which enables us to inspect causal relationships of video recommendations with a causal graph. We identify that duration has two kinds of effect on prediction: direct and indirect. Duration bias lies in the direct effect, while the indirect effect benefits prediction. To this end, we design a model-agnostic Counterfactual Video Recommendation for Duration Debiasing (CVRDD) framework, which incorporates multi-task learning to estimate different causal effect during training. In the inference phase, we perform counterfactual inference to remove the direct effect of duration for unbiased prediction. We conduct experiments on two industrial datasets, and in addition to achieving highly promising results on traditional top-k recommendation metrics, CVRDD also improves the user watch time.
- 30 September 2022
- Published by SPIE-Intl Soc Optical Eng
- Vol. 1 2, 52-477
Deep learning combined with topology and channel features for road extraction from remote sensing images
Using remote sensing images to automatically extract road network information has become an imposing method because manually labeling roads is a time-consuming and laborious task. Traditional road extraction is regarded as a pixel-based segmentation method, which predicts the probability of each pixel on the road. It often neglects the pixel’s neighboring information and the topology property of the road network, so some branches of roads cannot be recognized by this method. To mitigate the problem, this paper proposes a new road extraction method with attention and topology modules, TSELinkNet. It extracts channel features and topology features from the images and then integrates them with spatial features to acquire a comprehensive feature. This method is conducted on two different types of road datasets including urban and rural areas. Also, we compare the predicted results from TSE-LinkNet with other results from existing methods using Precision, Recall, F1 and mean intersection over union (mIoU) evaluation metrics. In both datasets, TSE-LinkNet improves Precision metrics by 0.61% and 2.58% respectively. In the urban road dataset, other metrics arise greatly as well, where Recall, F1 and mIoU increase1.28%, 14.18% and 4.46% sharply. The extracted roads from TSE-LinkNet have better connectivity compared with other methods in qualitative results. Experimental results showed that this method has a satisfying ability to extract roads from complex-topology urban areas.
- 30 September 2022
- Published by SPIE-Intl Soc Optical Eng
- Vol. 1 2, 52-477
Deep learning combined with topology and channel features for road extraction from remote sensing images
Using remote sensing images to automatically extract road network information has become an imposing method because manually labeling roads is a time-consuming and laborious task. Traditional road extraction is regarded as a pixel-based segmentation method, which predicts the probability of each pixel on the road. It often neglects the pixel’s neighboring information and the topology property of the road network, so some branches of roads cannot be recognized by this method. To mitigate the problem, this paper proposes a new road extraction method with attention and topology modules, TSELinkNet. It extracts channel features and topology features from the images and then integrates them with spatial features to acquire a comprehensive feature. This method is conducted on two different types of road datasets including urban and rural areas. Also, we compare the predicted results from TSE-LinkNet with other results from existing methods using Precision, Recall, F1 and mean intersection over union (mIoU) evaluation metrics. In both datasets, TSE-LinkNet improves Precision metrics by 0.61% and 2.58% respectively. In the urban road dataset, other metrics arise greatly as well, where Recall, F1 and mIoU increase1.28%, 14.18% and 4.46% sharply. The extracted roads from TSE-LinkNet have better connectivity compared with other methods in qualitative results. Experimental results showed that this method has a satisfying ability to extract roads from complex-topology urban areas.