POI Recommendation Method of Neural Matrix Factorization Integrating Auxiliary Attribute Information

Abstract

Point-of-interest (POI) recommendation is the prevalent personalized service in location-based social networks (LBSNs). A single use of matrix factorization (MF) or deep neural networks cannot effectively capture the complex structure of user–POI interactions. In addition, to alleviate the data-sparsity problem, current methods primarily introduce the auxiliary information of users and POIs. Auxiliary information is often judged to be equally valued, which will dissipate some of the valuable information. Hence, we propose a novel POI recommendation method fusing auxiliary attribute information based on the neural matrix factorization, integrating the convolutional neural network and attention mechanism (NueMF-CAA). First, the k-means and term frequency–inverse document frequency (TF-IDF) algorithms are used to mine the auxiliary attribute information of users and POIs from user check-in data to alleviate the data-sparsity problem. A convolutional neural network and an attention mechanism are applied to learn the expression of auxiliary attribute information and distinguish the importance of auxiliary attribute information, respectively. Then, the auxiliary attribute information feature vectors of users and POIs are concatenated with their respective latent feature vectors to obtain the complete latent feature vectors of users and POIs. Meanwhile, generalized matrix factorization (GMF) and multilayer perceptron (MLP) are used to learn the linear and nonlinear interactions between users and POIs, respectively, and the last hidden layer is connected to integrate the two parts to alleviate the implicit feedback problem and make the recommendation results more interpretable. Experiments on two real-world datasets, the Foursquare dataset and the Weibo dataset, demonstrate that the proposed method significantly improves the evaluation metrics—hit ratio (HR) and normalized discounted cumulative gain (NDCG).

1. Introduction

With the popularity of the Internet and the maturity of geolocation technology, location-based social networks (LBSNs) have emerged, such as Foursquare, Gowalla, Yelp, and Weibo [1]. As a vital service of LBSNs, point-of-interest (POI) recommendation has gradually become a research hotspot in the fields of geographic information systems and web information retrieval [2]. POI recommendation is used to mine users’ various life patterns and personal preferences, hidden behind large-scale data, to help users find the most interesting personalized places, thereby enriching users’ life experience [3]. Concurrently, it also helps relevant service providers provide intelligent, precise, and personalized services for potential users, which greatly enhances the users’ loyalty to the LBSNs platform, thereby improving economic benefits [4].

Collaborative filtering (CF) [5], which calculates users’ preference score for an unvisited POI through a user-POI check-in matrix, is a principal method for POI recommendation. However, users only visit a small fraction of POIs in LBSNs. Therefore, POI recommendation is confronted with a high data-sparsity problem, making it difficult for CF to play an effective role. Among many CF methods, matrix factorization (MF) is the most extensively used [6] because of its excellent scalability and accurate prediction ability in processing large-scale data. Cheng et al. [7] integrated users’ social relationships and location information into probability MF. He et al. [8] proposed a POI recommendation algorithm based on weighted MF, integrating social geographic location information that models social geographic information to mine users’ preferences for unvisited locations and improves the objective function of weighted MF in the form of implicit feedback. Lian et al. [9] presented a POI recommendation method based on the weighted MF model that captures spatial clustering phenomena from the perspective of two-dimensional kernel density estimation and integrates it into the MF model. However, all the above studies used latent feature vectors to represent users and POIs, and the interaction function of users and POIs was simply modeled as the inner product of their corresponding latent feature vectors. In this way, the complex structure of user–POI interactions could not be captured effectively, and POI recommendation can cause ranking errors.

In recent years, with deep learning’s rapid advancement, it has been applied to the field of recommendation systems on a large scale [10]. Deep learning can effectively capture nonlinear and special user–item relations and express more complex and abstract data to overcome the problem of data sparsity [2]. Therefore, a growing number of studies combine deep learning with CF [11], hoping to improve the performance of traditional recommendation algorithms. Meng et al. [12] learned the association between users and POIs in their respective attributes by constructing a convolutional neural network model. Feng et al. [13] proposed a mixed POI recommendation model based on deep learning that uses a convolutional neural network to learn the feature representations of comment information and then uses the extended MF model to fuse the feature of comment information. From the above literature, scholars have explored deep learning models based on implicit feedback, which primarily uses deep learning networks to model auxiliary information. To explore the high-order nonlinear relationship between users and items more effectively, He et al. [11] presented neural collaborative filtering (NCF), which directly employs multilayer perceptron (MLP) to model the interactions between users and items on the basis of MF. Guo et al. [14] proposed DeepFM as an end-to-end model that uses a deep neural network to model the low-order interaction relationship of FM and the interaction relationships with high-order characteristics. Deng et al. [15] proposed a DeepCF model framework that combines feature learning with a matching function to enhance the performance of the model. The above studies all apply deep learning networks to the interactions between users and items, which improves the predictive ability of the model. Inspired by the NCF framework, we consider mining the high-order nonlinear relationships between users and POIs based on MF in POI recommendation. MLP is often used to mine higher-order nonlinear relationships between users and items. While using MLP to directly learn the matching function gives the model great flexibility, the learning process can be inefficient without introducing the human experience. Therefore, it is necessary to explore low-order linear and high-order nonlinear relationships between users and POIs.

To relieve the pressure of data sparsity, many researchers try to solve this problem by introducing auxiliary information [16,17,18,19]. The CoupledCF model [20] uses a convolutional neural network to extract the attribute features of users and items and recommends items for users by combining the historical interactions between users and items. Wei et al. [21] subdivided the spatial relationship into the spatial distance relationship and spatial topology relationship to model the relationships between users and POIs. Chen et al. [22] used a neural network to extract the semantic features of comment text and images related to users and POIs, modeled the user–text semantic feature relationship and POI semantic feature relationship, and then fused them into a unified probability generation model. Wang [23] fused the POI with its auxiliary information using the graph-embedding method, obtained a deep-level POI vector with rich information, and then input it into the neural network. However, auxiliary information is equivalently treated without distinguishing its value. In fact, the importance of various auxiliary information is different; therefore, the impact on the POI recommendation results is different.

To mitigate data sparsity and implicit feedback in POI recommendation, this paper employs an MLP in the deep neural network to express the users’ preference for POI nonlinearly and utilizes the generalized matrix factorization (GMF) method to linearly describe the users’ preference for POI. To further enhance the accuracy of POI recommendation, the k-means and term frequency–inverse document frequency (TF-IDF) algorithms are used to construct the auxiliary attribute information of users and POIs from user check-in data. A convolutional neural network and an attention mechanism are used to learn the expression of auxiliary attribute information and distinguish the importance of auxiliary attribute information, respectively. To obtain the complete latent feature vectors of users and POI, we integrate the auxiliary attribute information expression of users and POIs into their respective latent feature vectors.

The main contributions of this paper are as follows:

• Employing the k-means and TF-IDF algorithms to construct the auxiliary attribute information of users (including user preference category and user high activity location) and the auxiliary attribute information of POIs (including category of POI, popularity POI, and region of POI) from user check-in data.
• For the optimum use of auxiliary attribute information, a convolutional neural network is used to learn the expression of auxiliary attribute information, and an attention mechanism is introduced to distinguish the importance of auxiliary attribute information. The complete latent feature vectors of users and POIs are expressed as the integration of the auxiliary attribute information feature vectors of users and POIs with the latent feature vectors of users and POIs.
• Based on the NCF framework, a novel neural matrix factorization method (NueMF-CAA) for POI recommendation is proposed. The method incorporates the auxiliary attribute information of users and POIs and uses GMF and MLP to deeply explore the interactions between users and POIs.
• Based on the Foursquare dataset and Weibo dataset, the feasibility and effectiveness of NueMF-CAA for POI recommendation are verified.

2. Method

2.1. Problem Definition

In LBSNs, $U=\left\{u_1,u_2,\dots ,u_m\right\}$ and $V=\left\{v_1,v_2,\dots ,v_n\right\}$ are the set of users and POIs, respectively, where $m$ is the number of users and $n$ is the number of POIs. $v\in V$ and $v$ contains latitude $lat$ and longitude $lon$. The interactions between users and POIs can be defined as check-in matrix $R^{m\times n}$. If the user $u$ has visited the POI $v$, the corresponding position of the check-in matrix is 1; otherwise, it is 0. Each element in the check-in matrix reflects the interaction between the user and the POI. In addition to the check-in matrix mentioned above, to alleviate the problem of data sparsity, we construct auxiliary attribute information of users and POIs from user check-in data to enrich the latent feature vectors of users and POIs. POI recommendation is used to recommend top-K POIs to a given user.

2.2. Overall Framework

Aiming at the problem of data sparsity and implicit feedback, this paper proposes a personalized recommendation method based on the classical NCF framework. The method framework is shown in Figure 1. The k-means and TF-IDF algorithm are used to mine the auxiliary attribute information of users and POIs from user check-in data. A convolutional neural network is used to learn latent representation from auxiliary attribute information, and an attention mechanism is introduced to distinguish the importance of auxiliary attribute information, thereby enriching the expression of the latent feature vectors of users and POIs. The NeuMF-CAA method integrates GMF and MLP to model the complete latent feature vectors of users and POIs. GMF makes use of linear kernels to model the interactions between users and POIs, and MLP employs nonlinear kernels to learn the interaction function. The data delivered by the above two parts are input into the sigmoid function in the form of vector concatenation, and finally, the prediction result is outputted.

2.3. Constructing Auxiliary Attribute Information

Data-sparsity is one of the main challenges for POI recommendations. A series of studies exploit different contextual information to alleviate this problem [24]. The main assumption of this type of work is that the combination of auxiliary information can better extract users’ behavior patterns [25]. The more information we have about the factors that influence users’ choices, the better we can model their behavior patterns and provide more appropriate recommendations. Therefore, this paper employs user check-in data to mine the auxiliary attribute information of users and POIs.

2.3.1. Constructing Auxiliary Attribute Information of POIs

The category of POI, popularity of POI, and region of POI constitute the auxiliary attribute information of POIs. The POI category is an inherent attribute of POI, and no further mining is required to construct it. The calculation of POI popularity and POI region are explained below.

The popularity of POIs refers to the popularity of POIs by users, which is evaluated by user check-in data. User check-in data include information about not only the inherent properties of the POI (e.g., longitude, latitude, and category), but also the number of check-ins at the POI. The number of check-ins at a POI is a satisfactory indicator to measure the popularity of POI, but it is not enough to directly use the number of POI check-ins to calculate the POI’s popularity. Research [26] has pointed out that the category information of POI plays an essential role in the POI recommendation because the category information of the POI can be used to characterize the POI’s features. Therefore, this paper calculated the popularity of the POI by introducing the TF-IDF algorithm. The algorithm considers not only the check-in frequency of a POI but also the popularity of the category to which the POI belongs [27]. Based on the user check-in data, POI popularity can be calculated by Formula (1):

$$popularity(v_i)=\frac{sizeof(v_i)}{sizeof(S{v}_i)}\times \log \frac{n}{sizeof(C{v}_i)}$$

(1)

where $sizeof(v_i)$ indicates the check-in times of POI $v_i$, $sizeof(S{v}_i)$ represents the number of check-ins of all POIs in the same category as $v_i$, and $sizeof(C{v}_i)$ represents the number of all POIs in the same category as $v_i$.

Location information is the basic attribute of POIs, and it is also a key factor that POI recommendation algorithms must consider. In this paper, the k-means is used to cluster POIs according to the location information of POIs to obtain appropriate region blocks so that each POI is assigned a location label for its corresponding region. The method for processing POI clustering based on k-means algorithm is as follows:

• Randomly select $k$ POIs from the set of POIs as the initial cluster center;
• Calculate the Euclidean distance $\rho$ from the remaining POIs to the cluster center, and put the closest POIs into the corresponding class to form a new class. The calculation of $\rho$ is shown in Formula (2):

  $$\rho =\sqrt{{(la{t}_i-la{t}_j)}^2+{(lo{n}_i-lo{n}_j)}^2}$$

  (2)
• Take the mean of all the latitudes and longitudes of POIs in the current cluster as the new center point and update the POIs closest to the cluster center;
• Until the objective function converges or the cluster center remains unchanged, it will transfer to 2;
• Output POI clustering results.

2.3.2. Constructing Auxiliary Attribute Information of Users

User preference category and user high-activity location (region) constitute the auxiliary attribute information of users. The user preference category refers to the category of POI that the user checks in most often. In the real world, the user’s high activity location may be the user’s residential region. Therefore, the POIs most frequently checked in by users are used to infer high activity location for users [28].

2.4. Learning Linear Interactions between Users and POIs

The traditional MF utilizes latent feature vectors to represent users and POIs, maps users and POIs to the same latent space, and estimates their interactions as the inner product of latent feature vectors. However, GMF estimates their interactions as the element-wise product of latent feature vectors. We now explain how to integrate the auxiliary attribute information into the GMF.

First, one-hot encoding on user ID and POI ID is performed. To apply the auxiliary attribute information of users and POIs in the model, the numerical variables are normalized and the categorical variables are one-hot encoded. The user latent feature vector $U_u$, the POI latent feature vector $V_v$, the user auxiliary attribute information $U_a$, and the POI auxiliary attribute information $V_a$ are sent to the embedding layer. The calculation of the latent feature vectors of $U_u$, $V_v$, $U_a$, and $V_a$ is shown in Formula (3):

$$\left\{\begin{array}{c}{H}_{U_u}=f(U_u)\\ G_{V_v}=g(V_v)\\ H_{U_a}=h(U_a)\\ G_{V_a}=l(V_a)\end{array}\right.$$

(3)

where, $f$, $g$, $h$, and $l$ denote representing functions.

Due to the widespread use of a convolutional neural network and its excellent performance, a convolutional neural network is used to learn the latent representation from the auxiliary attribute information of users and POIs. A convolutional neural network consists of a convolutional layer and a pooling layer, which can extract deeper features. Its execution mode is shown in Formula (4):

$$\left\{\begin{array}{c}{H}_{U_a}^{(1)}=pooling(g(w\ast H_{U_a}+b))\\ G_{V_a}^{(1)}=pooling(g(w\ast G_{V_a}+b))\end{array}\right.$$

(4)

where $\ast$ is the convolution operator, $w$ is the filter, $b$ is the bias of $w$, $g$ is the nonlinear activation function, and pooling is the pooling function (for example, maximum pooling or average pooling).

On this basis, an attention mechanism is introduced to distinguish the importance of auxiliary attribute information. The attention mechanism uses the $soft\max (·)$ function to compute the score as a probability distribution over the output of the auxiliary attribute information. Finally, the output is combined with $H_{U_a}$ and $G_{V_a}$, and the final output of the attention mechanism is obtained by the element-wise product of $(\otimes )$. The implementation of attention mechanism is shown in Formula (5):

$$\left\{\begin{array}{c}{H}_{U_a}^{(2)}=soft\max (H_{U_a}^{(1)})\otimes H_{U_a}^{}\\ G_{V_a}^{(2)}=soft\max (G_{V_a}^{(1)})\otimes G_{V_a}\end{array}\right.$$

(5)

Formula (6) is used to integrate the user feature vector $H_{U_u}$ and the user auxiliary attribute information feature vector $H_{U_a}^{(2)}$ to obtain a complete user feature vector $H^U$, and integrate the POI feature vector $G_{V_v}$ and the POI auxiliary attribute information feature vector $G_{V_a}^{(2)}$ to obtain a complete POI feature vector $G^V$.

$$\left\{\begin{array}{c}{H}^U=\left. [\begin{array}{c}{H}_{U_u}\\ H_{U_a}^{(2)}\end{array}\right]\\ G^V=\left. [\begin{array}{c}{G}_{V_v}\\ G_{V_a}^{(2)}\end{array}\right]\end{array}\right.$$

(6)

After obtaining the complete latent feature vectors of users and POIs, the mapping function of the prediction vector is defined, as shown in Formula (7):

$${\varphi }^{GMF}(H^U,G^V)=H^U\odot G^V$$

(7)

where, $\odot$ denotes the element-wise product of vectors. Then, the vector is projected to the output layer, as shown in Formula (8):

$${\widehat{y}}_{ui}=a_{out}(h^T(H^U\odot G^V))$$

(8)

where $a_{out}$ and $h$ represent the activation function and edge weights of the output layer, respectively. If $a_{out}$ and $h$ are vectors of 1, it will be a standard MF model. If a nonlinear function is used for $a_{out}$, it generalizes MF to a nonlinear setting, which may be more expressive than a linear MF model. In this paper, the sigmoid function $\sigma (x)=\frac{1}{(1+e^{-x})}$ is used as $a_{out}$.

2.5. Learning Nonlinear Interactions between Users and POIs

The input and auxiliary attribute information of the MLP is processed in the same way as the MF. After obtaining the complete latent feature vectors of users and POIs, current methods usually implement an element-wise product or concatenation of the two latent feature vectors. To capture the interactions between users and POIs, we concatenate the complete latent feature vector of the users with the complete latent feature vector of the POIs. However, simple concatenation cannot explore latent feature interactions between users and POIs. Affected by implicit feedback, the interactions between users and POIs become increasingly complicated. In this context, it is crucial to extend the interaction function characterizing the users and POIs to a nonlinear space. The benefit of MLP is that it enables the extension of user and POI modeling to more complicated nonlinear spaces, leveraging its powerful learning capability. Therefore, it is ideal for learning the interactions between users and POIs. The MLP is defined as Formula (9):

$$\left\{\begin{array}{c}{Z}_1={\varphi }_1(H^U,G^V)=\left. [\begin{array}{c}{H}^U\\ G^V\end{array}\right]\\ {\varphi }_2{}^{MLP}(Z_1)=a_2(W_2^T{Z}_1+b_2)\\ \dots \dots \\ {\varphi }_L{}^{MLP}(Z_{L-1})=a_L(W_L^T{Z}_{L-1}+b_L)\end{array}\right.$$

(9)

where $W_x$, $b_x$, and $a_x$ denote the weight matrix, bias vector, and activation function for the x-th layer’s perceptron, respectively.

At the output layer, the MLP and GMF are processed in the same way, as shown in Formula (10). Following NCF [11], we adopt a rectified linear unit (ReLU) function as the activation of the MLP layer function. The ReLU function is more biologically plausible and proven to be non-saturated [29]. Moreover, it supports sparse activation, being well-suited for sparse data and making the model less likely to be overfitting.

$${\widehat{y}}_{ui}=\sigma (h^T{\varphi }_L(Z_{L-1})+c)$$

(10)

2.6. Neural Matrix Factorization Integrating Auxiliary Attribute Information

After obtaining the complete latent feature vector of users and POIs, GMF employs linear kernels to model the latent feature interactions between users and POIs, and MLP utilizes nonlinear kernels to learn the interactions function between users and POIs. To effectively integrate the GMF and MLP so that they can reinforce each other and learn complex user–POI interactions, let the GMF and MLP learn individual embedding, and combine the two parts by connecting the last hidden layer, as shown in Formula (11):

$$\left\{\begin{array}{c}{\varphi }^{GMF}=H_{}^U\odot G_{}^V\\ {\varphi }^{MLP}=a_L(W_L^T(a_{L-1}(\dots a_2(W_2^T\left. [\begin{array}{c}{H}_{}^U\\ G_{}^V\end{array}\right]+b_2)\dots ))\\ {\widehat{y}}_{ui}=\sigma (h^T\left. [\begin{array}{c}{\varphi }^{GMF}\\ {\varphi }^{MLP}\end{array}\right])\end{array}\right.$$

(11)

where $\sigma$ represents the sigmoid activation function. The model combines the linearity of GMF and the nonlinearity of DNN to simulate the latent user–POI structure.

3. Experimental Results and Analysis

3.1. Dataset

In this experiment, we use two real-world datasets, Foursquare check-in data [30] in New York and Weibo check-in data in Shanghai, to evaluate the performance of our method. The Foursquare dataset collected user check-in data from 3 April 2012 to 16 February 2013. The Foursquare dataset contains information such as user ID, POI ID, time, latitude and longitude, and POI category. The Weibo dataset collected user check-in data from 1 January 2019 to 10 November 2021. The Weibo dataset contains information such as user ID, POI ID, time, latitude, and longitude. Since there is no category information for POI in the Weibo dataset, Python is used to call the AutoNavi map API [31] to obtain it. To alleviate the impact of data sparsity, the users who visited fewer than five POIs and the POIs visited by users fewer than five times are removed from the Foursquare dataset and Weibo dataset. The basic statistics of the dataset are shown in

Table 1. Dataset Statistics.

Statistical Accounts	Foursquare	Weibo
Number of users	1083	11,436
Number of POIs	10,250	17,565
Number of check-ins	172,838	731,044
Category of POI	251	130
Average number of POI check-ins by users	159	64
Sparsity/%	98.44	99.64

.

3.2. Comparison Method

To evaluate the recommendation performance of the NeuMF-CAA method, it is compared with the following recommendation methods:

• MF [32]: the method is a traditional recommendation model in recommender systems. It maps users and POIs into a latent low-dimensional space and computes the similarity between the two for recommendation results.
• NeuMF [11]: the model is implemented based on the NCF framework. Two models are proposed under the NCF framework, namely generalized matrix factorization and multilayer perceptron. NeuMF is a fusion model of these two models.
• CoupledCF [20]: this work builds on non-IID learning to propose a neural user–item coupling learning for collaborative filtering. CoupledCF jointly learns explicit and implicit couplings between users and items.

At the same time, ablation experiments were conducted to verify the influence of the GMF-CAA module, the MLP-CAA module, and the NeuMF-A module on the POI recommendation method proposed in this paper.

• GMF-CAA: the nonlinear kernels are not considered to model the interactions between users and POIs.
• MLP-CAA: the linear kernels are not considered to model the interactions between users and POIs.
• NeuMF-A: user and POI auxiliary attribute information is not processed using the convolutional attention mechanism.

3.3. Experimental Settings

To evaluate the performance of POI recommendation, we adopted the leave-one-out evaluation, which has been broadly used in the literature [33,34]. For each user, we selected a user’s latest interaction as the test data and employed the remaining data for training. In addition, a random sample of 99 POIs that the user had not interacted with was used to form the user’s test set with the above selected test data [35]. The NeuMF-CAA method was used to rank the 100 POIs of each user, and then the recommendation algorithm performance was evaluated. Similarly to the literature [11], hit ratio (HR) and normalized discounted cumulative gain (NDCG) were used as evaluation metrics in this experiment. The larger the value of the metrics, the better the recommendation effect.

The NeuMF-CAA model was implemented in Python based on the Keras framework [36]. A binary cross-entropy loss function was chosen to learn the model parameters, and Adam was used as the optimizer. The parameters were initialized with a random normal distribution (mean and standard deviation were 0 and 0.01, respectively). We evaluated the effect of different number of clusters (10, 20, 30, 40, 50, 60, 70, 80, 90, 100). For the Foursquare dataset, when the number of clusters was 50, the recommendation effect was the best; for the Weibo dataset, when the number of clusters was 70, the recommendation effect was the best. We tested with batch sizes (64, 128, 256, 512), learning rates (0.0001, 0.0005, 0.001, 0.005), and number of negatives (1, 2, 3, 4). Since the last hidden layer of NeuMF-CAA determines the performance of the model, we call it the predictive factor. We evaluated the predictive factors (8, 16, 32, 64). For the Foursquare dataset and the Weibo dataset, when the batch size was 256, the learning rate was 0.001, the number of negatives was 4, and the predictive factor was 32, giving the best recommendation effect.

3.4. Analysis of Experimental Results

3.4.1. Cluster Analysis

The NeuMF-CAA method needs to first determine $k$ in the POI clustering based on the k-means algorithm. We refer to the $k$ calculated by the Elbow Algorithm [37]. As shown in Figure 2, the Foursquare dataset had better clustering performance when $k$ was approximately 40–60; however, the Weibo dataset had better clustering performance when $k$ was approximately 40–70. This “elbow” cannot always be unambiguously identified, making this method very subjective and unreliable. Therefore, we referred to the relevant literature [12] for the selection of $k$ and performed a series of experiments. Figure 3 shows the changes in HR and NDCG of NeuMF-CAA with $k$ when the POI recommendation list is 10. As shown in Figure 3, for the Foursquare dataset, the best recommendation performance was achieved when $k=50$; for the Weibo dataset, the best recommendation performance was achieved when $k=70$. The $k$ values selected by the experiment are within the range of $k$ determined by the Elbow Algorithm. Consequently, $k=50$ and $k=70$ are fixed in Foursquare dataset and Weibo dataset, respectively.

3.4.2. Predictive Factors Analysis

The predictive factors consist of two parts, the latent feature vectors of users and POIs, and the corresponding auxiliary attribute information feature representation. We distinguish the importance of auxiliary attribute information, so in this part, we mainly focus on the predictive factors of the latent feature vectors of users and POIs. As shown in Figure 4a, for the Foursquare dataset, when the predictive factors are 32, NeuMF-CAA achieves the best performance on the metrics. As shown in Figure 4b, for the Weibo dataset, when the predictive factors are 32, NeuMF-CAA achieves the best performance on the metrics. The predictive factors of the latent feature vectors of users and POIs have a certain impact on the performance of the model. The recommendation effect produced by choosing different parameter models is also different.

3.4.3. Algorithm Comparison and Analysis

Figure 5 illustrates the comparison experiment based on the Foursquare dataset. Compared with several comparison methods, the method proposed in this paper has improved on evaluation metrics. Figure 5a shows that when K = 10, the HR of the NeuMF-CAA model is 0.8486, which is about 5.93% higher than that of the better-performing NCF. Figure 5b shows that when K = 10, the NDCG of the NeuMF-CAA model is 0.7423, and the NDCG of the NeuMF-CAA model is improved by approximately 3.25~8.49% compared with other baseline methods. When K = 8, the NDCG of the NeuMF-CAA model is improved by about 11.43% compared with the traditional MF.

Figure 6 shows the results of the comparison experiment based on the Weibo dataset. The performance of the POI recommendation method proposed in this paper was shown to be significantly superior than other comparison algorithms on the evaluation metrics. Figure 6a shows that when K = 10, the HR of the NeuMF-CAA model is 0.8853, which is about 3.53% higher than that of the better-performing NCF. Compared with traditional MF and CoupledCF, it is improved by 7.31% and 6.34%, respectively. Figure 6b shows that when K = 10, the NDCG of the NeuMF-CAA model is 0.7884, and the NDCG of the NeuMF-CAA model is improved by approximately 3.74~16.8% compared with other baseline methods.

Traditional MF only utilizes linear kernels to learn the interaction function. Therefore, the experimental results are not as good as NeuMF-CAA. This reflects the positive effect of using a neural network model to learn the interaction function of users and POI with respect to recommendation performance. NCF only considers the respective intrinsic features of users and POIs, while NeuMF-CAA constructs the auxiliary attribute information of users and POIs using TF-IDF and k-means to obtain more accurate vector representation of users and POIs. CoupledCF handles auxiliary attribute information equivalently, while NeuMF-CAA introduces an attention mechanism to distinguish the importance of auxiliary attribute information, which can more accurately capture users’ preferences.

NeuMF-CAA had excellent performance on the Foursquare public dataset and the crawled Weibo dataset, showing that the method proposed in this paper has good generalization. Due to the large number of Weibo datasets and the diversity of POIs checked in by users being slightly more singular than that of the Foursquare dataset, the performance of NeuMF-CAA on Weibo datasets was generally better than that of Foursquare datasets.

3.4.4. The Influence of Various Factors in the NeuMF-CAA Method on the Experimental Results

To better reflect the impact of GMF-CAA, MLP-CAA, and NeuMF-A on model performance, we conducted a set of ablation experiments. Figure 7 and Figure 8 show that using the convolutional attention mechanism to process the auxiliary attribute information can improve the performance of the recommender system. For the Foursquare dataset, when K = 2, 4, 6, 8, and 10, compared with NeuMF-A, the HR of NeuMF-CAA with the convolutional attention mechanism was improved by approximately 4.15~16.79% and the NDCG is improved by approximately 5.69~24.10%. For the Weibo dataset, HR was improved by approximately 1.28~15.34% and NDCG was improved by approximately 6.72~18.35%. Compared with GMF-CAA and MLP-CAA, NeuMF-CAA, which integrates linear GMF and nonlinear MLP models, has a higher expressive ability in both the Foursquare dataset and the Weibo dataset. This result further demonstrates the superiority of combining linear and nonlinear kernels to model user preferences.

4. Discussion and Conclusions

This paper employs the k-means and TF-IDF algorithms to mine the auxiliary attribute information of users and POIs from user check-in data. A convolutional neural network is used to learn latent representation from the auxiliary attribute information of users and POIs, and an attention mechanism is introduced to distinguish the importance of the constructed auxiliary attribute information of users and POIs. Finally, the auxiliary attribute information expressions of users and POIs are integrated into their respective latent feature vectors, and the complete feature vectors of users and POIs are obtained. A neural matrix factorization model is built. The linear interactions of users’ and POIs’ feature vectors is captured by GMF, and the nonlinear interactions of user and POI feature vectors is captured by MLP. Finally, the two parts are fused by concatenating the last hidden layer. Experiments were carried out on the Foursquare dataset and the Weibo dataset, and the results show that (1) the performance of NeuMF-CAA is better than that of the current commonly used recommendation algorithms; (2) further processing of auxiliary attribute information using convolutional attention mechanism can enhance the accuracy of POI recommendation to some extent; (3) by incorporating auxiliary attribute information into the neural matrix factorization model and learning the linear and nonlinear interactions between users and POIs, the accuracy of the recommendation algorithm can be more effectively improved.

POI recommendation is a typical application of location-based services and has great research value. The POI recommendation method proposed in this paper has a certain flexibility and generality to incorporate other linear and nonlinear models. Moreover, it can flexibly and effectively learn the low-order linear interactions and the high-order nonlinear interactions between users and POIs. In future work, we will attempt to add factors such as time and social relations into the NeuMF-CAA model to study the influence patterns of factors such as spatio-temporal relations and social relations.