Partial VAE for Hybrid Recommender System

Chao Ma; Wenbo Gong; José Miguel Hernández-Lobato; Noam Koenigstein; Sebastian Nowozin; Cheng Zhang

Partial VAE for Hybrid Recommender System

Chao Ma
, Wenbo Gong
, José Miguel Hernández-Lobato
, Noam Koenigstein
, Sebastian Nowozin
, Cheng Zhang

University of the Cambridge, UK
Microsoft Research, Cambridge, UK
Allan Turing Institute, UK
Microsoft R&D, Israel

Research output: Chapter in Book/Report/Conference proceeding › Conference contribution › peer-review

Abstract

We propose a novel hybrid recommender system method that treats missing data in
a principled manner and that uses amortized inference for fast predictions. We name
this method, the Partial Variational Autoencoder (p-VAE). P-VAE uses a novel
probabilistic generative model to handle varying numbers of user ratings in a principled way. Using the proposed amortized partial inference technique in p-VAEs,
learning and inference can be efficiently performed by minimizing the so-called
partial variational upper bound, without making ad-hoc assumptions on the values
of missing ratings. Empirical experiments on the MovieLens dataset demonstrate
the state-of-the-art performance of our method for movie recommendations.

Original language	English
Title of host publication	Bayesian Deep Learning
Subtitle of host publication	NIPS 2018 Workshop
Number of pages	7
State	Published - 2018
Externally published	Yes
Event	NIPS 2018 Workshop: Bayesian Deep Learning - Palais des Congrès de Montréal, Montréal, Canada Duration: 7 Dec 2018 → 7 Dec 2018 http://bayesiandeeplearning.org/2018/

Workshop

Workshop	NIPS 2018 Workshop
Abbreviated title	NIPS Workshop
Country/Territory	Canada
City	Montréal
Period	7/12/18 → 7/12/18
Internet address	http://bayesiandeeplearning.org/2018/

Access to Document

Cite this

@inproceedings{a4e986ca69a34f0e9f9173dfd2e68426,

title = "Partial VAE for Hybrid Recommender System",

abstract = "We propose a novel hybrid recommender system method that treats missing data in a principled manner and that uses amortized inference for fast predictions. We name this method, the Partial Variational Autoencoder (p-VAE). P-VAE uses a novel probabilistic generative model to handle varying numbers of user ratings in a principled way. Using the proposed amortized partial inference technique in p-VAEs, learning and inference can be efficiently performed by minimizing the so-called partial variational upper bound, without making ad-hoc assumptions on the values of missing ratings. Empirical experiments on the MovieLens dataset demonstrate the state-of-the-art performance of our method for movie recommendations.",

author = "Chao Ma and Wenbo Gong and Hern{\'a}ndez-Lobato, \{Jos{\'e} Miguel\} and Noam Koenigstein and Sebastian Nowozin and Cheng Zhang",

year = "2018",

language = "אנגלית",

booktitle = "Bayesian Deep Learning",

note = "NIPS 2018 Workshop ; Conference date: 07-12-2018 Through 07-12-2018",

url = "http://bayesiandeeplearning.org/2018/",

}

TY - GEN

T1 - Partial VAE for Hybrid Recommender System

AU - Ma, Chao

AU - Gong, Wenbo

AU - Hernández-Lobato, José Miguel

AU - Koenigstein, Noam

AU - Nowozin, Sebastian

AU - Zhang, Cheng

PY - 2018

Y1 - 2018

N2 - We propose a novel hybrid recommender system method that treats missing data in a principled manner and that uses amortized inference for fast predictions. We name this method, the Partial Variational Autoencoder (p-VAE). P-VAE uses a novel probabilistic generative model to handle varying numbers of user ratings in a principled way. Using the proposed amortized partial inference technique in p-VAEs, learning and inference can be efficiently performed by minimizing the so-called partial variational upper bound, without making ad-hoc assumptions on the values of missing ratings. Empirical experiments on the MovieLens dataset demonstrate the state-of-the-art performance of our method for movie recommendations.

AB - We propose a novel hybrid recommender system method that treats missing data in a principled manner and that uses amortized inference for fast predictions. We name this method, the Partial Variational Autoencoder (p-VAE). P-VAE uses a novel probabilistic generative model to handle varying numbers of user ratings in a principled way. Using the proposed amortized partial inference technique in p-VAEs, learning and inference can be efficiently performed by minimizing the so-called partial variational upper bound, without making ad-hoc assumptions on the values of missing ratings. Empirical experiments on the MovieLens dataset demonstrate the state-of-the-art performance of our method for movie recommendations.

M3 - ???researchoutput.researchoutputtypes.contributiontobookanthology.conference???

BT - Bayesian Deep Learning

T2 - NIPS 2018 Workshop

Y2 - 7 December 2018 through 7 December 2018

ER -

Partial VAE for Hybrid Recommender System

Abstract

Workshop

Access to Document

Fingerprint

Cite this