A5044098687- Recommender Systems and Techniques
- Topic Modeling
- Multimodal Machine Learning Applications
- Domain Adaptation and Few-Shot Learning
- Face recognition and analysis
- Data Management and Algorithms
- Data Stream Mining Techniques
- Radiomics and Machine Learning in Medical Imaging
- Machine Learning and ELM
- Advanced Image and Video Retrieval Techniques
- Face and Expression Recognition
- Bayesian Modeling and Causal Inference
- Advanced Bandit Algorithms Research
- Remote-Sensing Image Classification
- Caching and Content Delivery
- Sparse and Compressive Sensing Techniques
- Machine Learning and Algorithms
- Data Quality and Management
- Data Mining Algorithms and Applications
- Sentiment Analysis and Opinion Mining
- Image and Video Quality Assessment
- Biometric Identification and Security
- Image Retrieval and Classification Techniques
Google (United States)
2020-2024
University of California, Los Angeles
2012-2015
Zhejiang University
2011
Large scale recommender models find most relevant items from huge catalogs, and they play a critical role in modern search recommendation systems. To model the input space with large-vocab categorical features, typical learns joint embedding through neural networks for both queries user feedback data. However, millions to billions of corpus, users tend provide very small set them, causing power-law distribution. This makes data long-tail extremely sparse.
Highly skewed long-tail item distribution is very common in recommendation systems. It significantly hurts model performance on tail items. To improve tail-item recommendation, we conduct research to transfer knowledge from head items items, leveraging the rich user feedback and semantic connections between Specifically, propose a novel dual learning framework that jointly learns both model-level item-level: 1. The builds generic meta-mapping of parameters few-shot many-shot model. captures...
We explore in this paper efficient algorithmic solutions to robustsubspace segmentation. propose the SSQP, namely SubspaceSegmentation via Quadratic Programming, partition data drawnfrom multiple subspaces into clusters. The basic idea ofSSQP is express each datum as linear combination of otherdata regularized by an overall term targeting zero reconstructioncoefficients over vectors from different subspaces. derivedcoefficient matrix solving a quadratic programming problem istaken affinity...
Modern recommender systems have evolved rapidly along with deep learning models that are well-optimized for overall performance, especially those trained under Empirical Risk Minimization (ERM). However, a recommendation algorithm focuses solely on the average performance may reinforce exposure bias and exacerbate "rich-get-richer" effect, leading to unfair user experience. In simulation study, we demonstrate such gap among various groups is enlarged by an ERM-trained in long-term. To...
Embedding learning of categorical features (e.g. user/item IDs) is at the core various recommendation models. The standard approach creates an embedding table where each row represents a dedicated vector for every unique feature value. However, this method fails to efficiently handle high-cardinality and unseen values new video ID) that are prevalent in real-world systems. In paper, we propose alternative framework Deep Hash (DHE), replacing tables by deep network compute embeddings on fly....
Industry recommender systems usually suffer from highly-skewed long-tail item distributions where a small fraction of the items receives most user feedback. This skew hurts quality especially for slices without much While there have been many research advances made in academia, deploying these methods production is very difficult and few improvements industry. One challenge that often hurt overall performance; additionally, they could be complex expensive to train serve.
Large scale recommender models find most relevant items from huge catalogs, and they play a critical role in modern search recommendation systems. To model the input space with large-vocab categorical features, typical learns joint embedding through neural networks for both queries user feedback data. However, millions to billions of corpus, users tend provide very small set them, causing power-law distribution. This makes data long-tail extremely sparse. Inspired by recent success...
Multi-Task Learning (MTL) is a powerful learning paradigm to improve generalization performance via knowledge sharing. However, existing studies find that MTL could sometimes hurt generalization, especially when two tasks are less correlated. One possible reason hurts spurious correlation, i.e., some and not causally related task labels, but the model mistakenly utilize them thus fail such correlation changes. In setup, there exist several unique challenges of correlation. First, risk having...
Industrial recommendation systems process billions of daily user feedback which are complex and noisy. Efficiently uncovering preference from these signals becomes crucial for high-quality recommendation. We argue that those not inherently equal in terms their informative value training ability, is particularly salient industrial applications with multi-stage processes (e.g., augmentation, retrieval, ranking). Considering that, this work, we propose a novel self-auxiliary distillation...
Highly skewed long-tail item distribution is very common in recommendation systems. It significantly hurts model performance on tail items. To improve tail-item recommendation, we conduct research to transfer knowledge from head items items, leveraging the rich user feedback and semantic connections between Specifically, propose a novel dual learning framework that jointly learns both model-level item-level: 1. The builds generic meta-mapping of parameters few-shot many-shot model. captures...
Embedding learning of categorical features (e.g. user/item IDs) is at the core various recommendation models including matrix factorization and neural collaborative filtering. The standard approach creates an embedding table where each row represents a dedicated vector for every unique feature value. However, this method fails to efficiently handle high-cardinality unseen values new video ID) that are prevalent in real-world systems. In paper, we propose alternative framework Deep Hash...
Industry recommender systems usually suffer from highly-skewed long-tail item distributions where a small fraction of the items receives most user feedback. This skew hurts quality especially for slices without much While there have been many research advances made in academia, deploying these methods production is very difficult and few improvements industry. One challenge that often hurt overall performance; additionally, they could be complex expensive to train serve. In this work, we aim...