SDNN: Symmetric deep neural networks with lateral connections for recommender systems
Introduction
With the rapid development of Internet technology, we have entered the era of information explosion, where recommender system (RS) plays a vital role in helping consumers to find the most interesting and relevant items (e.g., products, service, movies) amongst a huge variety of options. Modern recommender systems are usually based on collaborative filtering (CF), which relies on historical behavioral data, such as ratings or clicks, to lean user preferences [30]. Among various CF techniques, matrix factorization (MF), which represents users and items by learning a latent space, provides superior performance over others and has become a de facto standard for latent factor model-based recommendation [1]. Earlier MF algorithms mainly rely on explicit ratings from users on items [28], [17]. Later, much research effort has been devoted to improving MF from different aspects, such as considering user or item biases [18], integrating extra data like social relations [32], and exploiting implicit feedback [25], [15].
In the past few years, deep learning has achieved great success on computer vision, speech recognition and nature language processing [4], [9]. Due to its good performance at representation learning, deep learning models have also been widely used in recommender systems, such as Restricted Boltzmann Machines [29], Autoencoders and the denoising autoencoders [19], [31]. With explicit historical ratings as the input, these methods tend to learn hidden structures that can reconstruct a user’s ratings. In addition to explicit ratings, implicit feedback has also been utilized in deep learning recommendation models. For example, Wu et al. [36] proposed a collaborative denoising autoencoder (CADE) for CF with implicit feedback, He et al. [11] proposed neural collaborative filtering (NCF) that models the interaction between user and item latent factors with a multi-layer perceptron neural network. Note that the above two works only exploit implicit feedback, in [37], by utilizing both explicit ratings and implicit feedback, a deep matrix factorization (DMF) model was proposed, which uses two distinct multi-layer perceptron (MLP) networks to learn user and item latent vectors and then models the user-item interaction by cosine similarity of the latent vectors. Recently, Deng et al. [5] categorized CF models into two types, i.e., representation learning-based CF and matching function learning-based CF, and proposed a Deep Collaborative Filtering (DeepCF) framework, which combines the strengths of these two types of CF models to achieve better performance.
The matching function learning-based CF method learns the relationship between user-item representation and matching score by stacked neural layers. As deep neural networks have proven to be capable of approximating any continuous function [13], such kind of method can capture complex relationship well. Nevertheless, deep neural networks applied on recommender systems are a pyramidal structure in general, which maps user and item latent vector into low-dimensional space to learn abstract and invariant feature. However, it may lead to partial features hard to be separated. Mapping to higher dimension may make it easier to classify those features that are complex in low-dimensional space. For example, as shown in Fig. 1, data represented in two-dimensional space are linearly indivisible, but when it is mapped into three dimensional-space, it can be divided by a plane. To better understand the usefulness of high-dimensional space representation, consider the example in Fig. 2, where the recommendation systems needs to rank three movies for a user for later recommendation. Suppose that a single dimension simply represents a kind of factor. The feature vector [0.7, 0.1, 0.9] indicates the user’s explicit preferences, including Prefer Detective Comics, Insensitive to Role and Like Adventure. Each movie is embed by a feature vector with different factors. When considering only two factors, Studio and Role, a ranking list can be obtained by conducting a dot product between user feature vector and item feature vector. It can be observed that the scores of The Dark Knight (0.96) and Superman Returns (0.95) are very close to each other, which makes it hard to make a proper recommendation (suppose Top-1 recommendation here). If we take one more factor, such as Genres, into consideration, another ranking list can be obtained and the score of different movies become easily distinguishable. We can simply choose the movie Superman Returns with the highest score of for recommendation. This example illustrates that high-dimensional space representation can help better differentiate data that are hard to be separated in low-dimensional space.
In this paper, we propose a novel Symmetric Deep Neural Networks (SDNN) with lateral connections to learn the complex relationship between user-item representation and matching score. SDNN contains two parts, a forward MLP is used to capture the invariant relations in low-dimensional space, while a reverse MLP combined with lateral connections is used to capture relations in higher dimension. Besides, with lateral connections, the reverse MLP can obtain the output of the hidden layer from the forward MLP layer by which allows the reverse MLP to include more information. Moreover, considering that deep neural network is very inefficient in catching low-rank relations between users and items, following the spirit of DeepCF, we combine SDNN with an improved deep matrix factorization model [5], to learn complex matching relations and low-rank relations between users and items jointly. We term the joint model as Dual Collaborative Filtering (DualCF). Our main contributions are summarized as follows,
- •
We propose a novel Symmetric Deep Neural Networks (SDNN) that can learn the mapping relations between user-item representation and matching score from two perspectives, i.e., high-dimensional space and low-dimensional space, to help improve the recommendation performance.
- •
We propose a novel framework, Dual Collaborative Filtering (DualCF), which combines SDNN with an improved deep matrix factorization model to learn complex mapping relations and low-rank relations between users and items jointly.
- •
Extensive experiments are conducted on three publicly available datasets and the experimental results demonstrate the effectiveness of our methods over several state-of-the-art models.
The remainder of this paper is organized as follows. Section 2 discusses related work; Section 3 formally defines the problem to be addressed in this work; Section 4 details our proposed methods SDNN and DualCF; Section 5 presents and analyzes the experimental results; Finally, Section 6 draws a conclusion.
Section snippets
Related work
Collaborative filtering (CF) is one of the widely used method in recommender systems. CF methods can be categorized into two types: neighborhood based methods and latent factor based methods. Neighborhood based methods generate recommendations based on user’s or item’s neighborhood and can be classified as user-oriented and item-oriented accordingly [30]. As for latent factor based model, it mainly contains two types: matrix factorization (MF) and factorization machine (FM).
FM-based methods try
Problem definition
We consider a recommender system with M users and N items and attempt to learn user-item interactions from implicit feedback. Compared to explicit feedback, e.g., ratings, like or dislike, implicit feedback is more easy to obtain. Also, although explicit feedback directly reflect user’s preference, i.e., the degree a user liking an item, implicit feedback is more representative of user’s real thought in practical applications. For example, a user may be not very supportive of Donald Trump’s
Symmetric deep neural networks
Deep neural networks focus on learning abstract and invariant relations, which makes it possible to generalize the invariance to new and unseen interactions. As mentioned earlier, in many cases, it is useful to learn relations in high-dimensional space which can ease classification. Also, inspired by [23], we utilize the lateral connections between two deep neural networks to transfer intermediate results when relations are mapped into high-dimensional space, which we think can help improve the
Experiments
In this section, we conduct experiments to evaluate the performance of the proposed models SDNN and DualCF versus several state-of-the-art models. We first introduce the experimental setup, and then present and discuss the experimental results. We have implemented our models based on Keras1 and Tensorflow2.
Conclusion
In this paper, we design a new model named Symmetric Deep Neural Networks (SDNN) that contains two MLPs with lateral connections. By decomposing the relationship learning process into the forward MLP and the reverse MLP, SDNN learns relations in high-dimensional and low-dimensional space simultaneously. The lateral connections allow the invariant patterns learned from the forward MLP to modulate the reverse MLP and transfer information, and finally help improve the model performance. Moreover,
CRediT authorship contribution statement
Runzhi Xu: Conceptualization, Methodology, Software, Writing - original draft. Jianjun Li: Supervision, Resources, Writing - review & editing. Guohui Li: Supervision, Project administration. Peng Pan: Supervision. Quan Zhou: Data curation, Validation. Chaoyang Wang: Software, Validation.
Declaration of Competing Interest
The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.
Acknowledgement
The work was partially supported by the National Natural Science Foundation of China under Grant No. 61672252, and the Fundamental Research Funds for the Central Universities under Grant No. 2019kfyXKJC021.
References (40)
- et al.
A recommender system for component-based applications using machine learning techniques
Knowl.-Based Syst.
(2019) - et al.
Multilayer feedforward networks are universal approximators
Neural Networks
(1989) - et al.
An endorsement-based trust bootstrapping approach for newcomer cloud services
Inf. Sci.
(2020) - et al.
Intelligent recommender system based on unsupervised machine learning and demographic attributes
Simul. Model. Pract. Theory
(2021) Recommender Systems: The Textbook
(2016)- et al.
Wide & deep learning for recommender systems
- et al.
Multimodal trust based recommender system with machine learning approaches for movie recommendation
Int. J. Inform. Technol.
(2021) - et al.
A unified architecture for natural language processing: deep neural networks with multitask learning
- et al.
Deepcf: A unified framework of representation learning and matching function learning in recommender system
- et al.
A hybrid collaborative filtering model with deep structure for recommender systems
Deepfm: a factorization-machine based neural network for ctr prediction
Deep residual learning for image recognition
Neural factorization machines for sparse predictive analytics
Neural collaborative filtering
Field-aware factorization machines for ctr prediction
Convolutional matrix factorization for document context-aware recommendation
Adam: A method for stochastic optimization
Factorization meets the neighborhood: A multifaceted collaborative filtering model
Cited by (9)
Providing prediction reliability through deep neural networks for recommender systems
2023, Computers and Industrial EngineeringA novel graph-based feature interaction model for click-through rate prediction
2023, Information SciencesKnowledge distillation-enhanced multitask framework for recommendation
2023, Information SciencesGraphs get personal: learning representation with contextual pretraining for collaborative filtering
2023, Applied Intelligence