Publications

MV-HAN: A Hybrid Attentive Networks based Multi-View Learning Model for Large-scale Contents Recommendation

Published in Journal 31, 2022

Industrial recommender systems usually employ multi-source data to improve the recommendation quality, while effectively sharing information between different data sources remain a challenge. In this paper, we introduce a novel Multi-View Approach with Hybrid Attentive Networks (MV-HAN) for contents retrieval at the matching stage of recommender systems. The proposed model enables high-order feature interaction from various input features while effectively transferring knowledge between different types. By employing a well-placed parameters sharing strategy, the MV-HAN substantially improves the retrieval performance in sparse types. The designed MV-HAN inherits the efficiency advantages in the online service from the two-tower model, by mapping users and contents of different types into the same features space. This enables fast retrieval of similar contents with an approximate nearest neighbor algorithm. We conduct offline experiments on several industrial datasets, demonstrating that the proposed MV-HAN significantly outperforms baselines on the content retrieval tasks. Importantly, the MV-HAN is deployed in a real-world matching system. Online A/B test results show that the proposed method can significantly improve the quality of recommendations.

Recommended citation: Fan, Ge, et al. "MV-HAN: A Hybrid Attentive Networks based Multi-View Learning Model for Large-scale Contents Recommendation ." In 37th IEEE/ACM International Conference on Automated Software Engineering (ASE ’22), 2022. [PDF]

Field-aware Variational Autoencoders for Billion-scale User Representation Learning

Published in Journal 31, 2022

User representation learning plays an essential role in Internet applications, such as recommender systems. Though developing a universal embedding for users is demanding, only few previous works are conducted in an unsupervised learning manner. The unsupervised method is however important as most of the user data is collected without specific labels. In this paper, we harness the unsupervised advantages of Variational Autoencoders (VAEs), to learn user representation from large-scale, high-dimensional, and multi-field data. We extend the traditional VAE by developing Field-aware VAE (FVAE) to model each feature field with an independent multinomial distribution. To reduce the complexity in training, we employ dynamic hash tables, a batched softmax function, and a feature sampling strategy to improve the efficiency of our method. We conduct experiments on multiple datasets, showing that the proposed FVAE significantly outperforms baselines on several tasks of data reconstruction and tag prediction. Moreover, we deploy the proposed method in real-world applications and conduct online A/B tests in a look-alike system. Results demonstrate that our method can effectively improve the quality of recommendation. To the best of our knowledge, it is the first time that the VAE-based user representation learning model is applied to real-world recommender systems.

Recommended citation: Fan, Ge, et al. "Field-aware Variational Autoencoders for Billion-scale User Representation Learning." 2022 IEEE 38th International Conference on Data Engineering (ICDE), 2022. [PDF]

PPPNE: Personalized Proximity Preserved Network Embedding

Published in Journal 31, 2021

After being proved extremely useful in many applications, the network embedding has played a critical role in the network analysis. Most of recent works usually model the network by minimizing the joint probability that the target node co-occurs with its neighboring nodes. These methods may fail to capture the personalized informativeness of each vertex. In this work, we propose a method named \emph{Personalized Proximity Preserved Network Embedding} (PPPNE) to adaptively capture the personalization of vertices based on the personalized ranking loss. Our theoretical analysis shows that PPPNE generalizes prior work based on matrix factorization or neural network with single layer, and we argue that preserving personalized proximity is the key to learning more informative representations. Moreover, to better capture the network structure in multiple scales, we exploit the distance ordering of each vertex. Our method can be efficiently optimized with a vertex-anchored sampling strategy. The results of extensive experiments on five real-world networks demonstrate that our approach outperforms state-of-the-art network embedding methods with a considerable improvement on several common tasks including link prediction and vertex classification. Additionally, PPPNE is efficient and can be easily accelerated by parallel computing, which enables PPPNE to work on large scale networks.

Recommended citation: Fan, Ge, et al. "PPPNE: Personalized Proximity Preserved Network Embedding." Neurocomputing, 2021. [PDF]

Collaborative filtering via heterogeneous neural networks

Published in Journal 1, 2021

Over the last few years, the deep neural network is utilized to solve the collaborative filtering problem, a method of which has achieved immense success on computer vision, speech recognition as well as natural language processing. On one hand, the deep neural network can be used to capture the side information of users and items. On the other hand, it is also capable of modeling interactions between users and items. Most of existing approaches exploit the neural network with solo structure to model user-item interactions such that the learning representation may be insufficient over the extremely sparse rating data. Recently, a large number of neural networks with mixed structures are devised for learning better representations. A carefully designed hybrid network is able to achieve considerable accuracy but only requires a small amount of extra computation. In order to model user-item interactions, we elaborate a hybrid neural network consisting of the global neural network and several local neural blocks. The multi-layer perceptron is adopted to build the global neural network and the residual network is used to form the local neural block which is inserted into two adjacent global layers. The hybrid network is further combined with the generalized matrix factorization to capture both the linear and nonlinear relationships between users and items. It is verified by experimental results on benchmark datasets that our method is superior to certain state-of-the-art approaches in terms of top-n item recommendation.

Recommended citation: Wei Zeng, Ge Fan, et al. Collaborative filtering via heterogeneous neuralnetworks." Applied Soft Computing: 1-15. [PDF]

Predicting ratings in multi-criteria recommender systems via a collective factor model

Published in Journal 1, 2021

In a multi-criteria recommender system, users are allowed to give an overall rating to an item and provide a score on each of its attribute. Finding an effective method to exploit a user s multi-criteria ratings to predict the overall rating becomes one of the most important challenges. Among traditional solutions, most of the architectures are not designed in an end-to-end manner. These approaches initially estimate a user s multi-criteria scores, and train a separate model to predict the user s overall rating. This introduces extra training overhead, and the overall prediction accuracy is usually sensitive to its multi-criteria ratings models. In this paper, we propose a collective model to predict user s overall rating by automatically weighting each of the predicted multi-criteria sub-scores. The proposed architecture integrates the multi-criteria ratings and the overall rating models in a unified system, which allows to train and perform multi-criteria recommendation in an end-to-end manner. Experiments on 3 real datasets show that our proposed architectures achieve up to 13.14% lower prediction error over baseline approaches.

Recommended citation: Ge Fan, et al. "Predicting ratings in multi-criteria recommender systems via a collective factor model." Companion Proceedings of the Web Conference 2021: 1-6. [PDF]

Preference modeling by exploiting latent components of ratings

Published in Journal 1, 2018

Understanding user preference is essential to the optimization of recommender systems. As a feedback of user’s taste, the rating score can directly reflect the preference of a given user to a given product. Uncovering the latent components of user ratings is thus of significant importance for learning user interests. In this paper, a new recommendation approach was proposed by investigating the latent components of user ratings. The basic idea is to decompose an existing rating into several components via a cost-sensitive learning strategy. Specifically, each rating is assigned to several latent factor models and each model is updated according to its predictive errors. Afterward, these accumulated predictive errors of models are utilized to decompose a rating into several components, each of which is treated as an independent part to further retrain the latent factor models. Finally, all latent factor models are combined linearly to estimate predictive ratings for users. In contrast to existing methods, our method provides an intuitive preference modeling strategy via multiple component analysis at an individual perspective. Meanwhile, it is verified by the experimental results on several benchmark datasets that the proposed method is superior to the state-ofthe-art methods in terms of recommendation accuracy.

Recommended citation: Chen, Junhua, et al. "Preference modeling by exploiting latent components of ratings." Knowledge and Information Systems (2018): 1-27. [PDF]

Discriminant Method of Mushroom Toxicity Based on Support Vector Machine

Published in Journal 31, 2015

he resemblance between edible mushroom and poisonous mushroom in appearance makes it hard to distinguish them from each other by conventional methods. In order to achieve the automation of judgment and strengthen the reliability, this paper proposed a method to measure the toxicity of mushroom based on support vector machine. To begin with, collection and pre-processing of the sample data were conducted. Then C-SVM model was built up and trained in accordance with one-to-one principle to further achieve multiclassification by support vector machine. At last, constant step length method was applied to obtain the optimum parameters of the model. By comparing accuracy of SVM classification in diverse sample sizes and parameters, the feasibility was verified in simulation experiments. SVM was more accurate, easy-conducting and practical comparing with neural network and decision tree.

Recommended citation: Fan, Ge, et al. "Discriminant Method of Mushroom Toxicity Based on Support Vector Machine." Chinese Agricultural Science Bulletin 31(19):232-236, 2015. [PDF]