Deep Heterogeneous Autoencoders for Collaborative Filtering

Abstract

This paper leverages heterogeneous auxiliary information to address the data sparsity problem of recommender systems. We propose a model that learns a shared feature space from heterogeneous data, such as item descriptions, product tags and online purchase history, to obtain better predictions. Our model consists of autoencoders, not only for numerical and categorical data, but also for sequential data, which enables capturing user tastes, item characteristics and the recent dynamics of user preference. We learn the autoencoder architecture for each data source independently in order to better model their statistical properties. Our evaluation on two MovieLens datasets and an ecommerce dataset shows that mean average precision and recall improve over state-of-the-art methods.

Deep learning techniques for rating prediction: a survey of the state-of-the-art

“…Further, some works such as Seo et al (2017) and TARMF (Lu et al 2018) adopted attention network in deep neural models to find more representative features. Furthermore, feature combination (Wu et al 2019) and heterogeneous feature learning Dong et al (2017), Ma et al (2018) is studied. In the-state-of-the-art models such as Liu et al (2019), joint learning of features is adopted.…”

“…The well known recent hybrid models in the problem domain are TARMF (Lu et al 2018), aSDAE (Dong et al 2017), DHA (Ma et al 2018), RMR (Ling et al 2014), top-icMF (Bao et al 2014), CDL (Wang et al 2015), HTF (McAuley and Leskovec 2013), DAML (Liu et al 2019), CARL (Wu et al 2019) etc.…”

“…Denoising autoencoders (DAE) are a special class of AE where we use corrupted input, whereas stacked denoising autoencoders (SDAE) are stacked DAEs that are usually employed to extract more hidden representative features. DAEs are usually effective in predicting missing values in recommender systems from corrupted data (Wu et al 2016b), while SDAEs are helpful in finding out the denser form of the input data (matrix) and also useful in integrating auxiliary data in recommender system from a variety of data sources (Wang et al 2015;Dong et al 2017;Ma et al 2018;Li et al 2015;Strub et al 2016;Ying et al 2016;Wei et al 2017). Li et al (2015) leverages from marginalized DAEs and Wang et al (2015) exploits Bayesian SDAE for integrating auxiliary information into ratings in their recommender systems.…”

“…4 layers used) and about the selection of dimensionality (64) of learned latent factors for user and item. Ma et al (2018) proposed a deep model using two heterogeneous autoencoders to address the data sparsity problem. The proposed model, called DHA, learns a shared feature space for heterogeneous data i.e.…”

“…Further, a stacked version of AEs produces more accurate predictions then AE (simple) as stacking improves feature learning (Li et al 2015). Furthermore, AEs are useful for feature extraction, dimensionality reduction, and of course rating predictions and recommendations as is revealed by the state-of-the-art AEbase methodologies (Wang et al 2015;Dong et al 2017;Ma et al 2018).…”