Keywords

1 Introduction

With the widespread deployment of the Internet and mobile Internet, billions of people have experienced online shopping. In online shopping applications like Amazon, one of the most important intelligent systems is the recommendation system, that is, the system recommends potential products to users or expands users’ interests in other areas; recommendation systems are also widely used in social networks to automate the social process of recommending friends or news to users [1].

One kind of recommendation system connect two different areas together, Zero-Shot learning (ZSL) and Cold-Start Recommendation (CSR) use their own Low-rank Linear Auto-Encoder (LLAE) [2]. The important challenge faced by online recommendation systems is the well-known cold start problem: how to provide advice to the new user? The embedded Influential-context Aggregation Unit (ICAU) as their ways to solve the problem for CSR. Their ICAU-based Heterogeneous Relations for Sparse model was presented in the passage to learn the user’s behaviour and give appropriate recommendations [3]. In the recommendation system, a MAML-based user preference estimator for movie recommendation. The MeLU model was separated into several layers that could be constantly updated to suit for new users based on its fast-learning speed. When user plug in their basic information, the model will adjust the movies for users to evaluate based on their ages and work previously collected by the system, then give the recommended movies based on the ratings the user gives. The feature or advantage of the model could give better results than regular methods, such as PPR and Wide & Deep, when encounter new users or new items [4]. Another approach of meta-learning to deal with CSR questions. This model proposed in the paper has the features of fast-learning speed and offers satisfying results just based on small datasets. Another unique feature of this model is its adaptive learning based on HINs to cope with different tasks easily. The result of the researcher’s experiments shows that, in both normal and new conditions, the HIN-based meta-learning model gives better results than regular models used in previous researches [5].

The recommendation complete current condition of the CSR problems and proposes their two separate solutions. The first solution is the framework of investigating the CF approach and machine learning algorithms to improve the performance for CS items. Then the second solution proposed is based on the first solution’s general framework. The original timeSVD++ model was presented by researchers to deal with the problem. This model make uses of CCS items with non-CS items’ similarity, and make use of different biases predictors to fully demonstrate the ability of the model. The results show that the timeSVD++ based IRCD-ICS model has the best performance of the five tested model [7]. The paper [9] proposed one linear-based model to deal with the CSR problems. To begin their researches, this paper analyzes three popular models that commonly used in solving CSR recommendations, and leads to the result that they are all the special case of the linear content-based model. Based on this results, the researchers gives their own model, the Low-Rank Linear CSR model.

This paper proposes a parallel network structure based on user interaction. The parallel graph neural network structure is used to process a matrix containing user interaction information, social media information and comment information at the same time. The purpose is to form a unified information among the three. The embedded structure fully captures the high-level relevance of the three, and reduces the computational dimension through parallel GNN. Experiments based on standard data sets prove that this method is better than baseline in standard measures and has a certain improvement in efficiency.

The rest of this paper is: the part II gives the general method of cold start of the recommendation system; the part III introduces the parallel network structure based on user interaction; the fourth part is the score results on the dataset; the last part gives the conclusion.

2 Cold-Start Recommendation Structure

In the recommendation system application, there are two types of entities, which we call users and items. The main purpose of the recommendation system is to filter based on the user’s preference for a certain item (such as a movie or book), generally using content-based item features or user social data based on collaborative filtering. The general structure of the recommendation system is shown in Fig. 1. In the past ten years, due to the popularization of the Internet, the massive amount of data generated has provided a rapid development opportunity for the recommendation system. The increasing demand for recommendation systems has caused many difficulties and challenges. Methods similar to cluster filtering and enhanced collaborative filtering have been proposed as a rich research field, recommendation system still needs continuous improvement.

Bi-clustering and Fusion [12] is a method that combines clustering and scoring to provide accurate recommendations for social recommendation systems. It tries to construct dense areas of the item-user rating matrix to solve the cold start problem. First, the method determines the popular items and extracts the scores into the item-user rating matrix; next, the role of Bi-clustering is to reduce the sparseness problem, smooth the ratings and aggregate similar users/items to form clusters, so that the items can be recommended to the classified customers Bi-clustering and Fusion. Its advantage is that it improves the accuracy of the recommendation while further reducing the dimension of the item-user rating matrix. In addition, the solution to the cold start problem is to remove the impact of sparseness and cluster users/items for smoothing.

Fig. 1.
figure 1

Recommendation system framework [10].

Full size image

The starting point for the design of neural networks is that computers learn to a certain extent similar to the way the human brain processes information. For the cold-start problem of the recommendation system, neural network [13] could optimize the similarity scoring process, which especially in the hybrid recommendation system by using neural network to learn user parameters or in the cluster recommendation system to learn voting information, such as Widrow-Hoff and other methods are used to learn user/item information to refine user parameter granularity.

The mathematical description of the cold start problem is as follows [8]: U is the group of users and \(\mathcal{P} \) is the group of products. \(a_{{u^{^{\prime}} }}\) represents whether current user purchased p. Each u ∈ \(\mathcal{U} \) connected with \(\mathcal{P} \) and has a timestamp. A small number of U linked to their social media content. \(\mathcal{A} \) denote the social media features and each account has a |\(\mathcal{A} \)| size vector. The social media account u ∉ \(\mathcal{U} \) is a new user to the e-commerce platform because it has no record of purchasing on the platform. In order to generate a unique product purchase recommendation ranking for each account from its social media account, due to the heterogeneous problem of social media and product purchase, the information from the social media account cannot be directly useful for product recommendation. Change the user’s social account information to feature Vu′, where the purpose of u is to make platform recommendations.

Common inputs in collaborative filtering include user set \(\mathcal{U} \) = u1, u2, …, un and item \(\mathcal{J} \) = v1, v2,…, vm. The recommendation level in the system can be represented by a matrix Y ∈ Rm×n that each item yij corresponds to the score of i by j. The general CF matrix decomposition is based on the rank Y ≈ UV form, where U ∈ ℝm×k and V ∈ ℝk×n characterization matrices represent potential factors, and the error is mainly obtained by minimizing reconstruction [11].

3 Parallel Network Structure Based on User Interaction

The latent factor model for users is one of the useful methods of the user recommendation system [6], but the interaction between users is often sparse, that is, there is a cold start problem, which limits the role of the latent factor model. The improved methods include normalized matrix decomposition for more relationship information similar to those embodied on social media, which to establish a standardized user-comment similarity evaluation model, and the use of word2vec to build an embedded model.

Graph representation is a method of describing data structure objects and their relationships in the form of nodes and edges [14, 15]. In recent years, many researchers have used machine learning to achieve graph representation, that is, graphs can be used to represent data structures in complex systems such as social networks for classification, Prediction and clustering operations. The graph neural network based on deep learning has interpretability and good performance. GNN draws on the ability of convolutional neural networks to express multi-scale spatial features, but CNN can only process European data (Fig. 2).

Fig. 2.
figure 2

User interations and expected social connections [6].

Full size image

Aiming at the problem of data sparseness caused by cold start, this paper proposes a parallel network cold start recommendation method based on user interaction information, social media information, and comment information, which is shown in Fig. 3. The purpose is to extract the embedded structure between the three types of information at the same time and obtain more information of high-level correlation inference. The purpose of the parallel structure is to compress further sparse data to achieve the purpose of reducing the computational dimension.

Fig. 3.
figure 3

Parrallel GNN structure based on three information.

Full size image

In the input part, the user interaction information, social media information and comment information are combined to extract the embedded structure and form an embedded matrix. In the parallel GNN, multiple Spatio-Temporal GCN parallel methods are mainly used to divide the matrix into multiple sub-matrices through the connection structure, where each most sub-matrix is adjusted to achieve parallel compression of sparse data and reduce the amount of calculation. Finally, loss optimization is performed and the recommended ranking result is output.

4 Experimental Results

E-commerce platforms like amazon can provide a large amount of user and product data. Founded in 2004, Yelp is a well-known merchant review website in the United States, covering merchants in restaurants, shopping malls, hotels, tourism, etc. from all over the world. Users use the Yelp website to rate merchants and submit reviews.

This paper selects Yelp’s 2014 dataset [16], which has more than 40k business items and 110k text comments from Phoenix and other regions. Yelp Reviews format is divided into two types: JSON and SQL, which contains user/check-in/business/tip/review saved in JSON files with specified ID. Comments for different business categories maybe very different in their contents. Therefore, it is necessary to clean and preprocess the data set to ensure the consistency of the data distribution.

First, we selected 100,000 reviews and converted the JSON format of these reviews into CSV format. From these reviews, we selected Cold-start users, that is, users with less than 5 user-item interactions. The model we used was pre-trained on the adjusted 2014 dataset training set. In order to verify the performance of this network structure, we compared and evaluated the baseline and the method proposed in this article on the above data set, and then selected part of the data for parameter fine-tuning, and the number of iterations in the fine-tuning stage is determined based on experience, and finally tested on the test dataset (Table 1).

Table 1. NDCG/HR score and average improvement of two methods.
Full size table

Normalized Discounted Cumulative Gain is the evaluation index of the sorting result to evaluate the accuracy of the sorting, where Gain represents the relevance score of each item in the list, and Cumulative Gain represents the accumulation of the gain of K items. The calculation formula is nDCGp = DCGp/IDCGp. Here for p < 0.05, the improvement is statistically significant compared to all other methods.

The baseline in this paper uses Neural collaborative filtering [17], which is a collaborative filtering method in recommendation systems. Unlike other algorithms that use neural networks to extract auxiliary features, user and item are still calculated using matrix inner products.

Table 1 shows the NDCG and HR scores obtained under the condition of cold start under Neural collaborative filtering (NeuMF) and the structure proposed in this paper. On the sparse Yelp dataset selected based on the cold start problem, the percentage improvements on the NDCG@10 and HR@10 indicators were 3.0% and 4.8%, respectively. This result shows that the proposed structure obtains better scores than the classical NeuMF method.

5 Conclusion

In online recommendation systems, products are recommended based on a large amount of user information. The cold start problem has always been one of the thorny issues that commercial recommendation platforms need to solve. Commonly used collaborative filtering methods are very unsuccessful for new users who do not have a lot of information. This paper proposes a parallel graph neural network based on user interaction, and extracts the embedded information of the user interaction letter/social media/comment information matrix to obtain high-level correlation. The parallel method further reduces the computational cost. Experiments based on the yelp data set prove that the standard index of this method under cold start conditions has certain advantages compared with NeuMF.