Thanks to their flexibility and scalability, collaborative embedding-based models are widely employed for the top-N recommendation task. Their goal is to jointly represent users and items in a common low-dimensional embedding space where users are represented close to items for which they expressed a positive preference. The training procedure of these techniques is influenced by several sources of randomness, that can have a strong impact on the embeddings learned by the models. In this paper we analyze this impact on Matrix Factorization (MF). In particular, we focus on the effects of training the same model on the same data, but with different initial values for the latent representations of users and items. We perform several experiments employing three well known MF implementations over five datasets. We show that different random initializations lead the same MF technique to generate very different latent representations and recommendation lists. We refer to these inconsistencies as instability of representations and instability of recommendations, respectively. We report that stability of item representations is positively correlated to the accuracy of the model. We show that the stability issues affect also the items for which the recommender correctly predicts positive preferences. Moreover, we highlight that the effect is stronger for less popular items. To overcome these drawbacks, we present a generalization of MF called Nearest Neighbors Matrix Factorization (NNMF). The new framework learns the embedding of each user and item as a weighted linear combination of the representations of the respective nearest neighbors. This strategy has the effect to propagate the information about items and users also to their neighbors and allows the embeddings of users and items with few interactions to be supported by a higher amount of information. To empirically demonstrate the advantages of the new framework, we provide a detailed description of the NNMF variants of three common MF techniques. We show that NNMF models, compared to their MF counterparts, largely improve the stability of both representations and recommendations, obtain a higher and more stable accuracy performance, especially on long-tail items, and reach convergence in a fraction of epochs.

Analyzing and improving stability of matrix factorization for recommender systems

Bernardis C.;Cremonesi P.
2022

Abstract

Thanks to their flexibility and scalability, collaborative embedding-based models are widely employed for the top-N recommendation task. Their goal is to jointly represent users and items in a common low-dimensional embedding space where users are represented close to items for which they expressed a positive preference. The training procedure of these techniques is influenced by several sources of randomness, that can have a strong impact on the embeddings learned by the models. In this paper we analyze this impact on Matrix Factorization (MF). In particular, we focus on the effects of training the same model on the same data, but with different initial values for the latent representations of users and items. We perform several experiments employing three well known MF implementations over five datasets. We show that different random initializations lead the same MF technique to generate very different latent representations and recommendation lists. We refer to these inconsistencies as instability of representations and instability of recommendations, respectively. We report that stability of item representations is positively correlated to the accuracy of the model. We show that the stability issues affect also the items for which the recommender correctly predicts positive preferences. Moreover, we highlight that the effect is stronger for less popular items. To overcome these drawbacks, we present a generalization of MF called Nearest Neighbors Matrix Factorization (NNMF). The new framework learns the embedding of each user and item as a weighted linear combination of the representations of the respective nearest neighbors. This strategy has the effect to propagate the information about items and users also to their neighbors and allows the embeddings of users and items with few interactions to be supported by a higher amount of information. To empirically demonstrate the advantages of the new framework, we provide a detailed description of the NNMF variants of three common MF techniques. We show that NNMF models, compared to their MF counterparts, largely improve the stability of both representations and recommendations, obtain a higher and more stable accuracy performance, especially on long-tail items, and reach convergence in a fraction of epochs.
File in questo prodotto:
File Dimensione Formato  
DAmico2022_Article_AnalyzingAndImprovingStability.pdf

Accesso riservato

Descrizione: Articolo principale
: Publisher’s version
Dimensione 1.6 MB
Formato Adobe PDF
1.6 MB Adobe PDF   Visualizza/Apri

I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.

Utilizza questo identificativo per citare o creare un link a questo documento: http://hdl.handle.net/11311/1210410
Citazioni
  • ???jsp.display-item.citation.pmc??? ND
  • Scopus 0
  • ???jsp.display-item.citation.isi??? 0
social impact