Please use this identifier to cite or link to this item: http://localhost/handle/Hannan/206513
Title: Modeling Data Correlations in Recommendation
Authors: Yuan He;Cheng Wang;Changjun Jiang
Year: 2017
Publisher: IEEE
Abstract: In the field of recommender systems, the Beer & Nappies is a famous story, which reveals the latent relationships between different categories of items. Though matrix factorization (MF) has demonstrated its great effectiveness in most previous work, it neglects the co-occurrences of items selected by individuals. In most MF-based models, the latent preferences of users (or the latent categories of items) are assumed independent, which thereby leads to the weak correlation between the Beer and the Nappies. It also greatly limits the models' ability in recommendation. In this paper, we propose a pure probabilistic generative model, which applies a Gaussian prior to capture the semantic correlations between the latent factors. We also show that our model theoretically achieves better expressive power than traditional MF-based models. We derive efficient inference and learning algorithms based on variational EM methods. The effectiveness of our proposed model is comprehensively verified on three different public data sets. Experimental results show that our approach achieves significant improvements on prediction quality compared with the current state of the art.
Description: 
URI: http://localhost/handle/Hannan/206513
volume: 5
More Information: 11030,
11042
Appears in Collections:2017

Files in This Item:
File SizeFormat 
7941995.pdf4.37 MBAdobe PDF
Title: Modeling Data Correlations in Recommendation
Authors: Yuan He;Cheng Wang;Changjun Jiang
Year: 2017
Publisher: IEEE
Abstract: In the field of recommender systems, the Beer & Nappies is a famous story, which reveals the latent relationships between different categories of items. Though matrix factorization (MF) has demonstrated its great effectiveness in most previous work, it neglects the co-occurrences of items selected by individuals. In most MF-based models, the latent preferences of users (or the latent categories of items) are assumed independent, which thereby leads to the weak correlation between the Beer and the Nappies. It also greatly limits the models' ability in recommendation. In this paper, we propose a pure probabilistic generative model, which applies a Gaussian prior to capture the semantic correlations between the latent factors. We also show that our model theoretically achieves better expressive power than traditional MF-based models. We derive efficient inference and learning algorithms based on variational EM methods. The effectiveness of our proposed model is comprehensively verified on three different public data sets. Experimental results show that our approach achieves significant improvements on prediction quality compared with the current state of the art.
Description: 
URI: http://localhost/handle/Hannan/206513
volume: 5
More Information: 11030,
11042
Appears in Collections:2017

Files in This Item:
File SizeFormat 
7941995.pdf4.37 MBAdobe PDF
Title: Modeling Data Correlations in Recommendation
Authors: Yuan He;Cheng Wang;Changjun Jiang
Year: 2017
Publisher: IEEE
Abstract: In the field of recommender systems, the Beer & Nappies is a famous story, which reveals the latent relationships between different categories of items. Though matrix factorization (MF) has demonstrated its great effectiveness in most previous work, it neglects the co-occurrences of items selected by individuals. In most MF-based models, the latent preferences of users (or the latent categories of items) are assumed independent, which thereby leads to the weak correlation between the Beer and the Nappies. It also greatly limits the models' ability in recommendation. In this paper, we propose a pure probabilistic generative model, which applies a Gaussian prior to capture the semantic correlations between the latent factors. We also show that our model theoretically achieves better expressive power than traditional MF-based models. We derive efficient inference and learning algorithms based on variational EM methods. The effectiveness of our proposed model is comprehensively verified on three different public data sets. Experimental results show that our approach achieves significant improvements on prediction quality compared with the current state of the art.
Description: 
URI: http://localhost/handle/Hannan/206513
volume: 5
More Information: 11030,
11042
Appears in Collections:2017

Files in This Item:
File SizeFormat 
7941995.pdf4.37 MBAdobe PDF