Computing Reviews
Today's Issue Hot Topics Search Browse Recommended My Account Log In
Review Help
Search
Deep item-based collaborative filtering for top-N recommendation
Xue F., He X., Wang X., Xu J., Liu K., Hong R. ACM Transactions on Information Systems37 (3):1-25,2019.Type:Article
Date Reviewed: Nov 19 2019

Recommender systems are an essential component in digital platforms, nudging consumers toward more efficient decision making by predicting and presenting products and services in a personalized ranking order that is of top interest. By filtering out items of less interest to the user, the recommender system alleviates information overload akin to a search engine.

Recommender systems can be classified into two major types: 1) content-based recommendations that use the item’s characteristics (features) to find similar items, and 2) collaborative filtering-based systems that rely on user-item interactions. Collaborative filtering includes user-based collaborative filtering (UCF) and item-based collaborative filtering (ICF): UCF uses the user’s past behaviors, such as direct or implicit ratings for items, to build a user preference model to predict their ratings or a set of K-top items; ICF recommends items based on similarity to a user’s past purchased items and ratings.

The ICF approaches include 1) “statistical measures such as a Pearson correlation and cosine similarity to quantify the similarity between two [purchased] items,” and 2) data-driven methods that learn item-to-item similarity from data, for example, the sparse linear method (SLIM), the factored item similarity model (FISM), and the neural attentive item similarity (NAIS) model, which uses attention networks to identify important item-to-item similarities. The authors claim that these approaches capture the pairwise relations between two items; however, they lack higher-order item relations among multiple items, which are used to estimate user preferences and can thus potentially improve recommender systems.

The proposed approach, DeepICF, uses a neural network consisting of an input embedding layer that vectorizes the user’s historical items and the target item, a pairwise interaction layer that captures pairwise similarities between two vectors, and a pooling layer that generates a fixed-size vector with two variants (DeepICF uses weighted average pooling and DeepICF+a uses attention-based pooling). Subsequently, the deep interaction layer consists of a multilayer perceptron (MLP) to capture the nonlinear higher-order relations, followed by the prediction layer that generates the prediction score.

The DeepICF and DeepICF+a models are tested using MovieLens and Pinterest datasets; results show improvements in prediction performance and interpretability compared to other approaches. These improvements are attributed to the effective learning of higher-order relations and the attention mechanism that differentiates the importance of the user’s historical items.

Industry-level recommendation systems require efficient and high-performance recommendation systems. This study tries to model not only pairwise similarity between two items, but also higher-order nonlinear relations among multiple items. In this new era of deep-learning-based recommender systems, industries and researchers may have to replicate and devise more interpretable deep learning models, thus verifying the findings in the study.

Reviewer:  Soon Ae Chun Review #: CR146785 (2004-0082)
Bookmark and Share
  Featured Reviewer  
 
General (H.0 )
 
 
General (H.3.0 )
 
Would you recommend this review?
yes
no
Other reviews under "General": Date
Introduction to database and knowledge-base systems
Krishna S., World Scientific Publishing Co., Inc., River Edge, NJ, 1992. Type: Book (9789810206192)
Nov 1 1993
An introduction to information science
Flynn R., Marcel Dekker, Inc., New York, NY, 1987. Type: Book (9789780824775087)
Apr 1 1988
Acta Informatica 19, 4 (Sept. 1983)
  Acta Informatica 44:1983. Type: Journal
Mar 1 1986
more...

E-Mail This Printer-Friendly
Send Your Comments
Contact Us
Reproduction in whole or in part without permission is prohibited.   Copyright 1999-2024 ThinkLoud®
Terms of Use
| Privacy Policy