Neural networks for personalized item rankings

doi:10.1016/j.neucom.2018.10.083

Neurocomputing

Volume 342, 21 May 2019, Pages 60-65

https://doi.org/10.1016/j.neucom.2018.10.083 Get rights and content

Abstract

Most users typically interact with products only through implicit feedback such as clicks or purchases rather than explicit user-provided information like product ratings. Learning to rank products according to individual preferences using only implicit feedback can be helpful to make useful recommendations. In this paper, a neural network architecture to solve collaborative filtering problems for personalized rankings on implicit feedback datasets is presented. It is shown how a layer of constant weights forces the network to learn pairwise item rankings. Additionally, similarities between the proposed neural network and a matrix factorization model trained with the Bayesian Personalized Ranking optimization criterion are proven. The experiments indicate state-of-the-art performance for the task of personalized ranking.

Section snippets

Introduction and related work

Users can be overwhelmed by the number of available items to choose from in today’s ecommerce sites. Helping the user by limiting this selection with personalized item recommendations can therefore substantially increase the probability of a purchase. However, selecting the most appropriate items is a difficult task. Especially, if there is no individual explicit feedback, like ratings, available.

Implicit feedback is one of the most available sources of information about user preferences. With

Preliminaries

Let $U = {1, \dots, N}$ be a set of users and $I = {1, \dots, M}$ a set of items with $N, M \in N$ . We have a dataset of observed interactions S ⊆ U × I, where (u, i) ∈ S means that user u interacted with item i in some way. Each observation (u, i) ∈ S is regarded as positive feedback and (u, i) ∈ (U × I)∖S as negative feedback. We are calling an item positive or negative if it is associated with positive or negative feedback, respectively, and define $I_{u}^{+}$ as the set of all positive items of user u ∈ U and $I_{u}^{-}$ as the set

Overview and notation

Our proposed model is a modified feedforward neural net with four specific layers L: an user layer L¹ with N units, a hidden layer L² with K units, an item layer L³ with M units and a ranking layer L⁴ with one unit (see Fig. 1). Therefore, the user layer L¹ has as many units as there are users and the item layer L³ has as many units as there are items. The parameter K determines the size of the user and item representations [5].

The following short notations are used in this paper: let $W^{l}$ be the

Connections to Bayesian personalized ranking

The optimization criterion BPR was proposed by Rendle et al. in [3]. It was designed to directly optimize AUC and the authors also applied it to a matrix factorization model. In this section, we will now show that a simplified version of our proposed network is equivalent to a matrix factorization model trained with the BPR criterion. This simplified version uses only identity activations functions and no bias layer.

First of all, the network proposed in this paper is similar to a matrix

Setting

The MovieLens 1M dataset [11], the Netflix Prize dataset [12] and the Jester dataset (version 2) [13] are used to evaluate our model. Since all datasets contain explicit ratings, they are converted to implicit ones by only keeping samples with a rating above 3 for the Netflix Prize and Movielens 1M dataset and a score above 0 for the Jester dataset. A random sample of 10,000 users is used for the Netflix Prize dataset to speed up the evaluation. Each dataset is split randomly into two sets of

Summary

In this paper, we have designed a neural network architecture to achieve a matrix factorization which directly optimizes the AUC metric to learn personalized item rankings. It was shown how to train this network with the BPR optimization criterion using a layer of constant weights and the common cross entropy cost function. We have proven that a simplified version of our network is similar to a matrix factorization model trained with BPR. Due to the transition to neural networks we were able to

Josef Feigl received a diploma in business mathematics from the University of Leipzig in 2011. Since 2014, he is pursuing his Ph.D. at the Department of Computer Engineering at the same university. His main areas of interest and research are the use of neural networks for recommender systems.

References (26)

HuY. et al.
Collaborative filtering for implicit feedback datasets
Proceedings of International Conference on Data Mining Workshops, ICDM
(2008)
B. Sarwar et al.
Incremental singular value decomposition algorithms for highly scalable recommender systems
Proceedings of the Fifth International Conference on Computer and Information Science
(2002)
S. Rendle et al.
BPR: Bayesian personalized ranking from implicit feedback
J. Feigl et al.
Neural networks for implicit feedback datasets
Proceedings of 26th European Symposium on Artificial Neural Networks, Computational Intelligence and Machine Learning, ESANN
(2018)
J. Feigl et al.
Collaborative filtering with neural networks
Proceedings of 25th European Symposium on Artificial Neural Networks, Computational Intelligence and Machine Learning, ESANN
(2017)
C.M. Bishop
Pattern Recognition and Machine Learning (Information Science and Statistics)
(2006)
D.E. Rumelhart et al.
Neurocomputing: Foundations of Research
(1988)
S. Rendle et al.
Improving pairwise learning for item recommendation from implicit feedback
Proceedings of the 7th ACM International Conference on Web Search and Data Mining, WSDM ’14
(2014)
G. Takács et al.
Scalable collaborative filtering approaches for large recommender systems
J. Mach. Learn. Res.
(2009)
Y. Koren et al.
Matrix factorization techniques for recommender systems
IEEE Comput.
(2009)

F.M. Harper et al.

The movielens datasets: history and context

ACM Trans. Interact. Intell. Syst.

(2015)

J. Bennett et al.

The netflix prize

Proceedings of KDD Cup and Workshop in Conjunction with KDD

(2007)

K. Goldberg et al.

Eigentaste: a constant time collaborative filtering algorithm

Inf. Retr.

(2001)

Cited by (6)

MulSimNet: A multi-branch sub-interest matching network for personalized recommendation
2022, Neurocomputing
Personalized recommendation serves as an indispensable functionality in many online services, where the key is to model the user’s preference based on past user-item interactions. Many well-known models learn a latent vector for each user and each item. Nevertheless, a single latent vector might not be able to well represent a user’s multifaceted interests revealed by the wide variety of items consumed by him/her. In this paper, we present a flexible framework named Multi-branch Sub-interest matching Network (MulSimNet) for personalized recommendation. Each user is explicitly represented by multiple sub-interest vectors as well as a general interest vector, and an item is matched against these vectors through different branches, where each branch is realized by appropriate neural networks. The compatibility scores output by multiple sub-interest matching branches are first aggregated via the max operator, which is then fused with the output of the general interest matching branch to estimate the user’s affinity with the item. A comprehensive set of experiments on benchmark datasets demonstrate the effectiveness of the proposed framework in fully capturing the user’s multifaceted interests and yielding better top-N recommendations.
A fuzzy recommendation system for predicting the customers interests using sentiment analysis and ontology in e-commerce
2021, Applied Soft Computing
Citation Excerpt :
Hongtao Liu et al. [25] developed another Hybrid recommendation model (HRDR) and this model provides the recommendations by referring the information from both ratings and explicit feedback obtained through user reviews by using deep learning approach. The rating matrix suffers when the number of users and products is huge [25]. This model preferred to use implicit feedbacks instead of referring to explicit feedback ratings.
In Electronic commerce, customer reviews play a significant role in purchase making decision. Most of the existing recommendation systems consider the customer reviews, user purchase history and product rating for predicting the recommended product. Since the users interest are varying over time, the existing recommendation systems lack in finding the current relevant items to the customers. To overcome this problem, this article proposes a new fuzzy logic-based product recommendation system which dynamically predicts the most relevant products to the customers in online shopping according to the users’ current interests. A novel algorithm has been proposed in this paper for computing the sentimental score of the product with associated end user target category. Finally, the proposed fuzzy rules and ontology-based recommendation system uses ontology alignment for making decisions that are more accurate and predict dynamically based on the search context. The experimental results of the proposed recommendation system show better performance than the existing product recommendation systems in terms of prediction accuracy of the relevant products for target users and in the time taken to provide such recommendations.
Discrete-time high-order neural network identifier trained with high-order sliding mode observer and unscented Kalman filter
2021, Neurocomputing
This paper presents a method to identify an unknown discrete-time nonlinear system, using high-order neural networks and high-order sliding mode algorithms, which may be subject to internal and external disturbances. Based on the information obtained from available system states, a high-order neural network model is proposed to approximate the system dynamics. Neural network weights are trained by means of the unscented Kalman filter and high-order sliding mode observer. A simulation example is included to illustrate effectiveness of the proposed scheme.
Learning the User's Deeper Preferences for Multi-modal Recommendation Systems
2023, ACM Transactions on Multimedia Computing, Communications and Applications
Computational Intelligence Enabled Cybersecurity for the Internet of Things
2020, IEEE Transactions on Emerging Topics in Computational Intelligence
Antiperiodic solutions to delayed inertial quaternion-valued neural networks
2020, Mathematical Methods in the Applied Sciences

Martin Bogdan In 1987 Martin Bogdan started his education in communication engineering at the Fachhochschule Offenburg. From 1991 to 1993 he studied Informatique Industrielle et Instrumentation in parallel with his studies at Offenburg resulting in a double diploma in 1993. Within his thesis he proved the feasibility of controlling a commercial limb prosthesis using recorded nerve signals obtaining his doctorate in 1999. From 2000 until 2015 he heads the research group for Neural Networks and Machine Learning at the University of Tübingen redirecting the main focus of the group towards biomedical applications, especially towards the communication with the nervous system and bio inspired information processing. Herein non-invasive and invasive Brain-Computer-Interfaces are playing a main role. In 2008 he received a full professorship for Computer Engineering at the Leipzig University. His department still deals with biomedical signal processing focused on neuroinformatics regarding neuro-inspired information processing and cognitive science (cognitive computing) as well as machine learning. In addition, his department works on Mainframe Computer and cyber security, respectively, cyber forensics.

View full text

Neural networks for personalized item rankings

Abstract

Section snippets

Introduction and related work

Preliminaries

Overview and notation

Connections to Bayesian personalized ranking

Setting

Summary

Collaborative filtering for implicit feedback datasets

Proceedings of International Conference on Data Mining Workshops, ICDM

Incremental singular value decomposition algorithms for highly scalable recommender systems

Proceedings of the Fifth International Conference on Computer and Information Science

BPR: Bayesian personalized ranking from implicit feedback

Neural networks for implicit feedback datasets

Proceedings of 26th European Symposium on Artificial Neural Networks, Computational Intelligence and Machine Learning, ESANN

Collaborative filtering with neural networks

Proceedings of 25th European Symposium on Artificial Neural Networks, Computational Intelligence and Machine Learning, ESANN

Pattern Recognition and Machine Learning (Information Science and Statistics)

Neurocomputing: Foundations of Research

Improving pairwise learning for item recommendation from implicit feedback

Proceedings of the 7th ACM International Conference on Web Search and Data Mining, WSDM ’14

Scalable collaborative filtering approaches for large recommender systems

J. Mach. Learn. Res.

Matrix factorization techniques for recommender systems