RACMF: robust attention convolutional matrix factorization for rating prediction

Matrix factorization is widely used in collaborative filtering, especially when the data are extremely large and sparse. To deal with the scale and sparsity problem of data, several recommender models adopt users and items’ side information to improve the recommendation results. However, some existi...

Celý popis

Uloženo v:
Podrobná bibliografie
Vydáno v:Pattern analysis and applications : PAA Ročník 22; číslo 4; s. 1655 - 1666
Hlavní autoři: Zeng, Biqing, Shang, Qi, Han, Xuli, Zeng, Feng, Zhang, Min
Médium: Journal Article
Jazyk:angličtina
Vydáno: London Springer London 01.11.2019
Springer Nature B.V
Témata:
Artificial neural networks
Computer Science
Datasets
Factorization
Industrial and Commercial Application
Model accuracy
Noise reduction
Pattern Recognition
Sparsity
Attention mechanism
Additional stacked denoising autoencoder
Convolutional neural network
Rating prediction
ISSN:1433-7541, 1433-755X
On-line přístup:Získat plný text
Tagy: Přidat tag
Žádné tagy, Buďte první, kdo vytvoří štítek k tomuto záznamu!
Popis
Shrnutí:Matrix factorization is widely used in collaborative filtering, especially when the data are extremely large and sparse. To deal with the scale and sparsity problem of data, several recommender models adopt users and items’ side information to improve the recommendation results. However, some existing works do not perform well enough for they are not effectively use the side information. For example, using bag-of-words model, topic model to gain the latent representation of words or merely utilizing items or users’ side information, leads to the result that the performance deteriorates, especially when rating dataset is extremely large and sparse. To overcome the data sparsity problem, we present a hybrid model named robust attention convolutional matrix factorization (RACMF) model, which is composed of attention convolutional neural network (ACNN) and additional stacked denoising autoencoder (aSDAE); ACNN and aSDAE are used to extract the items’ and users’ latent factors, respectively. The experimental results show that our RACMF model has good prediction ability, even when the rating data are sparse or the scale of rating data is large. What’s more, compared with the state-of-the-art model PHD, the present model RACMF increased the accuracy rate on ML-100k, ML-1m, ML-10m and AIV-6 datasets by 4.80%, 0.57%, 1.98% and 3.67%, respectively.
Bibliografie:ObjectType-Article-1
SourceType-Scholarly Journals-1
ObjectType-Feature-2
content type line 14
ISSN:1433-7541
1433-755X
DOI:10.1007/s10044-019-00814-2