Watch time is widely used as a proxy for user satisfaction in video recommendation platforms. However, raw watch times are influenced by confounding factors such as video duration, popularity, and individual user behaviors, potentially distorting preference signals and resulting in biased recommendation models. We propose a novel relative advantage debiasing framework that corrects watch time by comparing it to empirically derived reference distributions conditioned on user and item groups. This approach yields a quantile-based preference signal and introduces a two-stage architecture that explicitly separates distribution estimation from preference learning. Additionally, we present distributional embeddings to efficiently parameterize watch-time quantiles without requiring online sampling or storage of historical data. Both offline and online experiments demonstrate significant improvements in recommendation accuracy and robustness compared to existing baseline methods.

In current classification tasks, an important method to improve accuracy is to pre-train the model using a large-scale domain-specific dataset. However, many tasks such as writer identification (writerID) lack suitable large-scale datasets in practical scenarios. To address this issue, this paper proposes a method that can improve prediction accuracy without relying on significant pre-training but leveraging the diversity of probability distributions predicted by multiple networks, and enhancing the top-1 accuracy through complementary post-processing. Specifically, top-k distributions are sampled from the multiple probability mass functions separately. When the distribution differences of top-k are maximized, the intersection other than the correct category can be narrowed down. Finally, the correct target with suboptimal probability can be rectified by the only intersection. Furthermore, our method has exhibited an intriguing trait during experimentation. Its prediction accuracy enhances concurrently with the incorporation of novel SOTA methods, ultimately surpassing the performance of these new methods.

Recently, image enhancement approaches yield impressive progress. However, most methods are still based supervised-learning, which requires plenty of paired data. Meanwhile, owing to the complex illumination condition in a real-world scenario, those methods trained on synthetic images cannot restore details in extremely dark or bright areas and lead to exposure errors. The traditional losses that deem all pixels the same in training also produce blurry edges in the result. To handle these problems, in this article, we present an effective semi-supervised framework for severely underexposed image enhancement. Our network consists of a supervised and an unsupervised branch, which shares weights and can make full use of paired data and plenty of unpaired data. Meanwhile, a multi-exposure fusion module is designed to adaptively fuse the corrected images to address the low contrast and color bias issues occurring in some extreme situations. Moreover, we propose a supervised context attention module to better use the edge information as supervision to recover fine image details. Extensive experiments have proved that the proposed method outperforms state-of-the-art approaches in enhancing exposure images.

Image dehazing is an important task since it is the prerequisite for many downstream high-level computer vision tasks. Previous dehazing methods depend on either the hand-designed priors/assumptions or supervised learning with plenty of data, which are not easy to implement in practice. Meanwhile, synthesizing hazy images is also significant in many scenes like multi-weather image generation. In this paper, we change the viewpoint of this task to image translation and develop a weakly supervised framework to achieve it. Instead of simply considering the hazy image as the source domain and the haze-free image as the target domain for translation, we design a feature representation scheme that generates a domain indicator, and embed it into the decoder to achieve both hazing and dehazing within one network. This design significantly reduces the complexity of network and can be more easily extended to multi-domain translation tasks than the previous methods, which need one pair of generator-discriminator for each direction of the translation. Meanwhile, aiming at solving the haze-relevant task, we design a haze attention module, which takes the local entropy map as the input. Unlike the previous weakly supervised dehazing methods, our approach only requires unpaired hazy and haze-free images rather than any intermediate supervising data like the transmission map or atmospheric light defined in the atmospheric scattering model. Experimental results on synthetic datasets show our method can achieve competitive results when compared with the state-of-the-art methods and yield more appealing dehazing and hazing results on real-world images.

We study the task of learning entity and relation embeddings in knowledge graphs for predicting missing links. Previous translational models on link prediction make use of translational properties but lack enough expressiveness, while the convolution neural network based model (ConvE) takes advantage of the great nonlinearity fitting ability of neural networks but overlooks translational properties. In this paper, we propose a new knowledge graph embedding model called ParamE which can utilize the two advantages together. In ParamE, head entity embeddings, relation embeddings and tail entity embeddings are regarded as the input, parameters and output of a neural network respectively. Since parameters in networks are effective in converting input to output, taking neural network parameters as relation embeddings makes ParamE much more expressive and translational. In addition, the entity and relation embeddings in ParamE are from feature space and parameter space respectively, which is in line with the essence that entities and relations are supposed to be mapped into two different spaces. We evaluate the performances of ParamE on standard FB15k-237 and WN18RR datasets, and experiments show ParamE can significantly outperform existing state-of-the-art models, such as ConvE, SACN, RotatE and D4-STE/Gumbel.

Recover drawing orders from a Chinese handwriting image is a challenge issue. Most of English drawing order recovery(DOR) methods perform unsatisfactorily in Chinese. This paper proposes a novel image-to-sequence algorithm to deal with Chinese DOR problem. The proposed method utilizes two regression convolution neural network(CNN) models to generate two corresponding pen-tip movement heat-maps. To estimate pen-tip movement for most of the normal states in writing process, the algorithm analyzes the above two heat-maps with a specifically designed framework. Then the drawing order is restored through a simple iteration process based on the proposed framework. Experiments on public online handwriting database show that our method have got a remarkable result for Chinese DOR tasks. In addition, for English tasks, our method performs superiorly among state-of-the-art methods.

Pen Tip Motion Prediction (PTMP) is the key step for Chinese handwriting order recovery (DOR), which is a challenge topic in the past few decades. We proposed a novel algorithm framework using Convolutional Neural Network (CNN) to predict pen tip movement for human handwriting pictures. The network is a regression CNN model, whose inputs are a series of part-drawn handwriting images and output is a vector that represents the probability of next stroke point position. The predicted output vector is utilized by an iteration framework to generate pen movement sequences. Experiments on public Chinese and English online handwriting database have indicated that the proposed model performs competitively in multi-writer handwriting PTMP and DOR tasks. Furthermore, the experiment demonstrated that characters belong to different languages shares some common writing patterns and the proposed method could learn these laws effectively.