POI Recommendation Method Based on Multi-Source Information Fusion Using Deep Learning in Location-Based Social Networks

1. Introduction

Owing to the expansion of the Internet, information resources have grown exponentially, resulting in information overload. This prevents users from quickly obtaining useful knowledge from massive amounts of information. A recommendation system [1-5] can generate a list of recommended items for users or predict users’ preference for a specific item without providing users with clear demand information.

Deep learning-based recommendation algorithms can take implicit, explicit, and auxiliary information data as input [6]. Deep network models and deep learning techniques are used to learn hidden feature representations of users and items. However, check-in point of interest (POI) are extremely sparse [7,8] in location-based social networks [9,10], hindering the learning of deep features and resulting in poor model effectiveness.

To address the problem of extremely sparse check-in points in social networks, a content-aware POI recommendation algorithm (POIRA) is proposed based on a convolutional neural network (CNN). The main contributions of this study are summarised as follows:

1) A new representation learning method is proposed. This method integrates heterogeneous multi-source data into the recommendation model to learn the features with higher discriminative power and improve the precision of the model.

2) A two-layer stacked sparse autoencoder (SAE) network is proposed. The network connects two SAEs training activation vectors and softmax and fine-tunes weights to optimise the loss function, thereby improving the precision of the model.

The remainder of this paper is organised as follows: related work on social network recommendation systems and the motivation for writing this paper are described in Section 2. The architecture and detailed process of our proposed method are presented in Section 3. Experimental verification using recommended system datasets and comparative analysis with other existing methods are detailed in Section 4. Conclusions and future work are presented in Section 5.

2. Related Works

Researchers have proposed numerous methods to address the problem of extremely sparse check-in points in social networks. For example, in [11], an adaptive POI recommendation method was proposed by combining user activities and spatial features. This method performed adaptive operations based on user activities, improving model efficiency. In [12], a recommendation algorithm was proposed based on a deep neural network that uses basic data of users and projects. This method improved the recognition rate of the model by fusing the user feature matrix and item feature matrix. In [13], a model based on neural networks (preference and context embedding) was proposed. The model smoothed the sparse data by using neighbouring users and POI and constructed a context information graph to jointly learn the embedding of users and POI, improving the precision of the model. In [14], a POIRA was proposed based on an edge computing environment. The algorithm analysed users’ personalised preference functions on the edge server and embedded geographical information into the framework to obtain POI, thereby improving the precision of the model. However, the latent features learned by these methods when the data are too sparse are not particularly effective.

In [15], a social recommendation algorithm was proposed based on stochastic gradient matrix decomposition. The algorithm improved the precision of the model by fusing social network information. In [16], a recommendation algorithm was proposed by combining probabilistic semantic clustering analysis and collaborative filtering. The algorithm improved the precision of the model by increasing the mining for the sequence of items with time dimension. In [17], a recommendation method was proposed based on group trust and user weight analysis. The algorithm improved the precision of the model by identifying trustworthy users and integrating rent-based trust and user-based trust. In [18], a collaborative filtering recommendation algorithm was proposed based on information theory and bi-clustering. The algorithm improved model efficiency by introducing information entropy and bi-clustering into collaborative work and extracting local dense scoring modules. However, these methods use a bag-of-words model to process comment information, ignoring the semantic context information of comment information.

In [19], a matrix factorisation recommendation algorithm (segmentation-based matrix factorisation [SPMF]) was proposed based on social trust and preference segmentation. The algorithm improved the precision of the model by distinguishing trust relationships and differences in preference domains. In [20], an autoencoder-based multi-criteria recommendation algorithm (AE-MCCF) was proposed. The algorithm used multi-criteria preferences to represent the relationship between users nonlinearly, improving the precision of the model. In [21], a location regularisation recommendation algorithm was proposed based on social networks. The algorithm improved the precision of the model by limiting the ranking loss and assuming that the nearest neighbour POI is more easily accessed by similar users. However, the comment information of these methods has not been fully utilised in related POI studies.

In [22], a context-specific sequence-aware POIRA with multi-gate cyclic units was proposed. The algorithm improved the precision of the model by separately processing the influence for the context of each category. In [23], a feature-space separated factorisation model (FSS-FM) was proposed. The algorithm improved the precision of the model by representing the POI feature space as separate slices, adding spatial and temporal information, and other contexts. However, when the data set is large, these methods are prone to overfitting.

In [24], a total time tensor decomposition model was proposed based on POI recommendations. The model used linear combination operators to aggregate time latent features on different time scales, improving the precision of the model. However, the model runs slowly.

In [25], an adaptive POIRA was proposed based on a multi-order Markov model. The algorithm improved the precision of the model by predicting users’ next-favourite POI. However, this method is not suitable for long-term system prediction.

Most of the existing recommendation methods are based on word bags or document topic models to process comment text, which can only provide a shallow understanding of user preferences. This study focuses on the content-aware POIRA based on a deep CNN, to fuse the users’ social relationships and geographical information factors of location POI and to optimise the objective function after fusing features.

3. Preliminaries

3.1 Text Expression

For each independent POI, we sort out the related text information. For [TeX:] $$\mathrm{POI}_{j}$$, all comment texts related to [TeX:] $$\mathrm{POI}_{j}$$ and all description texts for [TeX:] $$\mathrm{POI}_{j}$$ are stored in the same document, [TeX:] $$d_{j}$$. We consider it as text information on [TeX:] $$\mathrm{POI}_{j}$$. For the [TeX:] $$\mathrm{POI}_{j}$$ document, the s word, [TeX:] $$w_{j s}$$, in [TeX:] $$d_{j}$$ is applied with the embedded function of words to obtain a n-dimensional word vector expression, [TeX:] $$\eta\left(w_{j s}\right)$$, regarding word [TeX:] $$w_{j s}$$. Therefore, we let the embedded vector matrix describing all words in document [TeX:] $$d_{j}$$ related to [TeX:] $$\mathrm{POI}_{j}$$ be denoted by [TeX:] $$d_{j}$$ to represent the initial feature vector for text semantics of [TeX:] $$\mathrm{POI}_{j}$$, as shown in formula (1).

Here, [TeX:] $$\eta()$$ is an embedded function of words and is called the embedding function. It matches each word to a fixed-length feature space. This is a vectorisation process. The embedded expression result of the s word, [TeX:] $$w_{j s}$$, is located in the s column of matrix [TeX:] $$\stackrel{\perp }{d_{j}}$$. [TeX:] $$\oplus$$ is the connection operator, and is the number of words in document [TeX:] $$d_{j}$$. The result of the embedded function for words represents different features of words in different dimensions. Therefore, the values of projection for words in different dimensions represent different semantics. Hence, the initial feature vector expression, [TeX:] $$\stackrel{\perp }{d_{j}}$$, for the semantics of texts can be obtained.

First, the convolutional layer is used to extract the semantic features of texts. The convolutional layer includes multiple neurons, which can use the convolution operator to extract new features from the initial feature vector, [TeX:] $$\stackrel{\perp }{d_{j}}$$, of text semantics obtained previously. To extract features locally, we set an sized convolution kernel (set as ), where a=m,n. Because each convolution kernel has fixed parameters, the extracted features will be relatively single. Thus, the model sets multiple different convolution kernels to obtain different features. Subsequently, the output obtained after each convolution kernel can be regarded as each feature obtained after the feature extraction of texts. Convolution kernel [TeX:] $$K_{c}, c=\{0,1 K k\}$$, is then used to extract feature [TeX:] $$z_{c}$$. Thus, the convolution feature generated by the [TeX:] $$d_{j}$$ document under the action of convolution kernel [TeX:] $$K_{c} \text { is } z_{c}^{j}$$, asshown in formula (2).

Here, f( ) is the activation function, * is the convolution operator, and [TeX:] $$b_{c}$$ is the deviation term of convolution. [TeX:] $$K_{c} \in R^{h \times t}$$; the window size is [TeX:] $$h \times t$$.

Subsequently, rectified linear units (ReLUs) are selected to further enhance the expressive power of CNNs. Compared with other activation functions, the ReLUs function is simple to operate. The function has suitable sparse activation, wide excitability boundaries, and unilateral inhibition. Furthermore, ReLUs are more computationally efficient for deep neural networks and are more suitable for deep network learning and training. The specific ReLU function is shown in formula (3).

To reduce the dimension of features, reduce the number of parameters, and avoid overfitting, the next step is to undergo a process of pooling a layer. In the pooling layer, the maximum pooling is selected to obtain the required features, as shown in formula (4).

Here, [TeX:] $$r_{s}$$ represents the size of the pooling window area, and [TeX:] $$a_{i}$$ represents the i feature in window [TeX:] $$r_{s}$$. represents the i feature in window r_s. Thus, the new features obtained become

Next, the features generated by all convolution kernels after passing through the maximum pool are

where is the number of convolution kernels.

Finally, the fully connected layers in the CNN are used to integrate the feature vectors obtained after multiple convolutional layers and pooling layers. In addition, the high-level semantic output, including the described object and its comment information in the text information, is obtained. That is, the expression learns the result of [TeX:] $$\text { POI }_{j}$$ for the text semantics. This is also the output of the CNN, as shown in formula (7).

Here, [TeX:] $$\stackrel{\perp}d_{j}$$ denotes the [TeX:] $$P O I_{j}$$ initial feature vector of text semantics, and [TeX:] $$W_{d}$$ is the weight matrix above the fully connected layer related to the text. [TeX:] $$L_{d}$$ is the feature vector generated before being fully connected, and [TeX:] $$g_{d}$$ represents the skew associated with texts in the fully connected layer. The characteristic dimension of [TeX:] $$\boldsymbol{C N N}_{d}\left(\stackrel{\perp}d_{j}\right) \text { is } d^{\prime}$$.

3.2 Image Expression

Image [TeX:] $$p_{j}$$ describing the size of [TeX:] $$P O I_{j} \text { as } m \times n \text { comprises } m \times n \text { pixels.}$$ A scale-invariant feature transform [26] is used to obtain several “key point” sets on image [TeX:] $$p_{j}$$. Each “key point” can be expressed as a 128-dimensional vector. For the entire image set, the k-means results of these “key points” can be used to generate several clusters, and each “cluster” serves as a “visual word.” Finally, all the “clusters” are used to create a “visual word dictionary.” In different images, each “key point” is assigned to the cluster closest to it. In this way, each image can be represented by a vector of size in the end. The value in vectors indicates the number of times the ‘visual word’ in each image appears in the image. At this time, image [TeX:] $$p_{j}$$ can be expressed as a vector expression of [TeX:] $$\stackrel{\perp}d_{j}=\left(a_{1}, a_{2}, \mathrm{~L} \quad a_{z}\right)$$. The obtained image vector, [TeX:] $$p_{j}$$, is used as the initial feature vector of image features to perform deeper feature extraction as input in the deep learning process in the next step.

After processing the entire CNN, the output of image [TeX:] $$p_{j} \text { of } P O I_{j}$$ can be obtained, that is, the expression learns the result of [TeX:] $$P O I_{j}$$ for the image features, as shown in formula (8).

Here, [TeX:] $$\stackrel{\perp}{p}_{j}$$ represents the initial feature vector of [TeX:] $$\mathrm{POI}_{j}$$ for image features. [TeX:] $$W_{p}$$ is the weight matrix on the fully connected layer related to images, and [TeX:] $$L_{p}$$ is the feature vector generated after passing through the pooling layer related to images. [TeX:] $$g_{p}$$ is the image-related skew in the fully connected layer, and the feature dimension of [TeX:] $$C N N_{p}\left(\begin{array}{l}\stackrel{\perp}{p}_{j}\end{array}\right) \text { is } p^{\prime}$$.

In summary, the initial feature vector, [TeX:] $$\stackrel{\perp}{d}_{j}$$, of text semantics and the initial feature vector, [TeX:] $$\stackrel{\perp}{p}_{j}$$, of image features are added to the CNN as inputs. This includes not only the feature extraction results of text and image information but also the cognition and understanding of the association between texts, images, and users’ behaviour. These initial feature vector expressions of text semantics and image features contain richer features and information than text semantics.

3.3 Bayesian Personalized Ranking Model

Assumption: User u prefers item [TeX:] $$I_{i}$$ over item [TeX:] $$I_{j}$$. While observing the [TeX:] $$\left(u, I_{i}\right)$$rating pair, it is impossible to observe rating pair [TeX:] $$\left(u, I_{j}\right)$$; U and I represent the user set and item set, respectively. This definition assumes the following:

where [TeX:] $$\Lambda$$ represents the parameter set of the ranking function, and [TeX:] $$\hat{\delta}_{u, I_{i}}(\Lambda) \text { and } \hat{\delta}_{u, I_{j}}(\Lambda)$$ are the predicted scores of the ranking function.

Therefore, based on all partial order pairs [TeX:] $$f_{u}$$ and the Bayesian theory, the posterior distribution of all partial order paris [TeX:] $$f_{u}$$ is obtained. The likelihood function is as follows:

where

[TeX:] $\hat{z}_{u, I_{i}, I_{j}}(\Lambda)$$ represents any real-valued function concerning the model parameter, [TeX:] $$\Lambda$$. It also reflects the preference relationship of user u for items [TeX:] $$I_{i} \text { and } I_{j} . \eta$$ is defined as follows:

Combining the above formulas, the final objective function is

where [TeX:] $$\lambda_{\Lambda}$$ is the model-specific regularisation parameter, and [TeX:] $$D_{x}=\left\{\left(u, I_{i}, I_{j}\right) \mid I_{i}, I_{j} \in I ; u \in U\right\}$$ is the training dataset. The Bayesian personalized ranking model uses stochastic gradient descent (SGD) to optimise the objective function.

4. Proposed Method

The framework of the proposed content-aware POIRA model based on a CNN is shown in Fig. 1. It first fuses geographical information, user social relations, and comment content information to construct the objective function. Subsequently, a two-layer stacked sparse autoencoder network is used for training to generate hidden features and item hidden features. Finally, a recommendation list is generated.

4.1 Geographical Information Modelling

The preference of user [TeX:] $$u_{i}$$ for several neighbouring positions of position [TeX:] $$l_{j}$$ represents the preference of user [TeX:] $$u_{i}$$ for position [TeX:] $$l_{j}$$. The geographic location weighting strategy [27] is used to complete the missing geographical information in the matrix decomposition model. The objective function minimisation problem can be expressed by the following formula.

is a weight parameter that controls the influence of adjacent positions. [TeX:] $$\operatorname{sim}\left(l_{j}, l_{k}\right)$$ represents the geographical weight of adjacent position [TeX:] $$l_{k}$$ at position [TeX:] $$l_{j}$$. [TeX:] $$z\left(l_{j}\right)$$ is a regularisation term defined as

[TeX:] $$\operatorname{sim}\left(l_{j}, l_{k}\right)$$ uses a Gaussian function, as shown in formula (19):

To distinguish the geographical range, a geographical area distance variable, F, is proposed because the user is less likely to check-in in the area farther from the current location. [TeX:] $$E\left(l_{j}\right)$$ represents the adjacent position of [TeX:] $$l_{j}$$. In the experiment, we set according to experience. If the location to be recommended is not in the user's current location, [TeX:] $$E\left(l_{j}\right)$$, the location is not considered.

4.2 Image Emotional Classification

For the image-based emotional analysis method in social media, a double SAE, as shown in Fig. 2, is constructed. First, an SAE network is trained [28]. When a new sample, x, is input into the network, vector [TeX:] $$h_{1}$$ composed of the activation values of the units for the hidden layer can represent x, as shown in formula (20), where i=1. The number of hidden layer nodes of the first SAE is set to 1000, and 1000 dimensional features are used as the new feature vector, [TeX:] $$h_{1}$$. This feature is called a first-order feature. This first-order feature is used as the input of the second SAE. The hidden layer sets the number of nodes to 256, and the 256-dimensional features trained are regarded as second-order features([TeX:] $$\left(h_{2}\right)$$). Two of these SAE networks use the nonlinear activation function, ReLUs, as shown in formula (21).

Next, we connect [TeX:] $$h_{1} \text { and } h_{2}$$ trained in the two SAE networks with a Softmax classifier to fine-tune weights. In this way, the output of the last fully connected layer can be converted into the emotion classification probability of images:

where [TeX:] $$h_{i}$$ is the output of the last fully connected layer. The loss function of the recognition probability is a multiclass cross-entropy loss function:

where [TeX:] $$y_{i}$$ is the true label of images.

4.3 User Preference and Location POI Attribute Modelling

The CNN can use the pre-trained word embedding model to more deeply understand the content of the position interest review paper. The text features are expressed as [TeX:] $$\operatorname{CNN}\left(W, C_{S}\right)$$, where [TeX:] $$C_{S}$$ is the set of review texts and W is the internal weight of the CNN. [TeX:] $$u_{i}$$ denotes users, and [TeX:] $$C_{S}$$ indicates posted comments. [TeX:] $$P=\left(f_{i s}=1 \mid u_{i}, C_{S}\right)$$, where [TeX:] $$f_{i s}$$ indicates whether [TeX:] $$C_{S}$$ is issued by [TeX:] $$u_{i}$$. Then, [TeX:] $$P=\left(f_{i s}=1 \mid u_{i}, C_{S}\right)$$ is specifically defined as follows:

Here, [TeX:] $$C \in R^{K \times d}$$ is the interaction matrix between the content features of comment texts and the potential features of users. The matrix can be used to distinguish users’ latent feature [TeX:] $$u_{i}$$, that is, whether comment [TeX:] $$C_{S}$$ is issued by user [TeX:] $$u_{i}$$. The Softmax function is used to normalise weights [29], which has the meaning of probabilistic interpretation and can realise the output of the network and interpret it as the posterior probability of the classification target variable. To extract the users’ latent feature vector, [TeX:] $$u_{i}$$, we convert formula (24) to the objective function.

Similarly, probability [TeX:] $$P\left(h_{j p}=1 \mid l_{j}, c_{\mathrm{P}}\right)$$ that comment [TeX:] $$c_{P}$$ is related to position [TeX:] $$l_{j}$$ is defined as

Here, [TeX:] $$h_{j p}$$ indicates whether comment [TeX:] $$c_{P}$$ is associated with position [TeX:] $$l_{j}$$. [TeX:] $$P \in R^{K \times d}$$ represents the interaction matrix between the features of comment content and the potential features of location POI. The interaction matrix can be used to distinguish the potential features, [TeX:] $$l_{j}$$, of location POI, that is, whether comment [TeX:] $$C_{P}$$ is associated with location [TeX:] $$l_{j}$$. Formula (26) is converted into the following objective function:

4.4 Constructing the Objective Loss Function

The comprehensive formula of the target loss function of the POIRA algorithm is obtained as follows:

Here, the regularisation term, [TeX:] $$\lambda_{\Omega}\|\Omega\|^{2}$$, is used to avoid overfitting during the learning process, and [TeX:] $$\lambda_{\Omega}$$ is the regularisation parameter.

Finally, according to the matrix decomposition technique of predicting [TeX:] $$\hat{y}$$, the following formula is obtained:

Here, [TeX:] $$Q_{u_{v}}$$ represents the latent feature matrix of user v. [TeX:] $$H_{l_{v}}, H_{l_{s}}, H_{l_{g}}, H_{l_{w}}$$ represent the latent feature matrix of POI [TeX:] $$l_{v}, l_{s}, l_{g}, l_{w}$$, respectively 30. [TeX:] $$b_{l_{v}}, b_{l_{s}}, b_{l_{g}}, b_{l_{w}}$$ represent the deviation items of POI, [TeX:] $$l_{v}, l_{s}, l_{g}, l_{w}$$, respectively. [TeX:] $$\varpi$$ denotes the dimension of the matrix decomposition. Row vector [TeX:] $$Q_{u_{v}}$$ represented by each row of matrix [TeX:] $$q_{u_{v}}$$ corresponds to the feature vector of each user. Row vectors [TeX:] $$H_{l_{v}}, H_{l_{s}}, H_{l_{g}}, H_{l_{w}}$$ represented respectively by each row of matrices [TeX:] $$h_{l_{v}}, h_{l_{s}}, h_{l_{q}}, h_{l_{w}}$$ correspond to the feature vector of each POI.

4.5 Loss Function Optimisation and Learning

Using SGD to optimise formula (28), we obtain the following formula:

Here, [TeX:] $$\lambda_{r, q}$$ represents the regularisation parameter of [TeX:] $$Q_{u_{v}}, \lambda_{r, h}$$ denotes the regularisation parameter of [TeX:] $$H_{l_{v}}, H_{l_{s}}, H_{l_{g}}, H_{l_{w}}$$, and [TeX:] $$\lambda_{r, b}$$ represents the regularisation parameter of [TeX:] $$b_{l_{v}}, b_{l_{s}}, b_{l_{g}}, b_{l_{w}}$$. The algorithm parameters are updated to obtain the following formula based on the aforementioned gradient:

Here, is the learning rate. The pseudocode of the POIRA is shown in Algorithm 1.

Spatial coding is mainly used in the POIRA algorithm, and the storage space does not change with the size of the data; therefore, the spatial complexity of the POIRA algorithm is O(1).

5. Experimental Results and Analysis

To verify the effectiveness of our proposed content-aware POIRA based on CNNs, sufficient experimental evaluations are performed on the Instagram-NewYork (NY) and Instagram-Chicago (CHI) datasets. The SPMF proposed in [19], AE-MCCF proposed in [20], FSS-FM proposed in [23], and POIRA are compared through experiments. These methods are implemented on Python 3.0.

5.1 Experimental Dataset

NY and CHI are datasets of two different regions on Instagram, which allows each user to associate geotags with pictures and related comments by posting pictures using their mobile phones. To improve the performance of this recommendation model, we first filter out the selfies and unclear virtual images. We then remove text and images that have no actual content or meaning. A maximum of 50 pictures and 100 text comments are retained. Users who have less than 10 check-in POI are also filtered out. We use 70% of the dataset for training, 10% for verification, and 20% for testing. The details of the two datasets are shown in Table 1.

5.2 Evaluation Criteria

The prediction rate and F1 value are used as evaluation criteria, and the detailed definitions are shown in formulas (32) and (34).

Here, TP represents the number of samples predicted to be positive and actually positive. FP represents the number of samples predicted to be positive but actually negative. FN represents the number of samples predicted to be negative but actually positive.

5.3 Effect Verification of Geographical Information, Users’ Social Relationship, and Comment Content Information

Precision is used in the experiment to measure the effectiveness of the model, abbreviated as [TeX:] $$\mathrm{P} @ \mathrm{k}$$. For a target user, [TeX:] $$u_{i}, \mathrm{P} @ \mathrm{k}$$ indicates the proportion of test access POI that will be included in the top k recommended POI. [TeX:] $$Q\left(u_{i}\right)$$ represents the POI that user represents the POI that user [TeX:] $$u_{i}$$ has checked in, and [TeX:] $$E\left(u_{i}\right)$$ represents the top k POI recommended. [TeX:] $$\text { P@k }$$ is defined as follows:

where V represents the number of users in the test data. To verify the effectiveness of the content-aware POIRA based on CNNs, experiments are performed on the NY and CHI datasets using several existing methods. In the experiment, three components—geographical information, users’ social relationship, and comment content information—are analysed and compared with those of the POIRA. The experimental results are listed in Table 2.

As can be seen in Table 2, the first three components are all critical to POI recommendations. The fusion of the third component is helpful in improving recommendation precision. Users are affected by various contextual information in real life. Therefore, the user's preference prediction cannot be modelled based on one aspect. The POIRA fully utilises the context information of various POI to solve the problem of sparse data in POI recommendation.

5.4 Impact Analysis of Parameter Changes

There are three important parameters in the POIRA: (1) control comment parameter, (2) control social relationships affect parameter, and (3) geographical neighbourhood relational weighting parameter. When studying and analysing these parameters, we change the value of one of the parameters while fixing the other parameters to analyse its impact on the final recommendation results and the sensitivity of POIRA to parameters. The experimental results are shown in Figs. 3–5.

As can be seen in Figs. 3 and 4, when the comment parameter is 0.05 and the social relationships affect parameter is 0.001, this model achieves the best recommendation effect. This indicates that users may mention only some potential topics, but not all in one review. As can be seen in Fig. 5, a good result is obtained when the geographical neighbourhood relational weighting parameter is 0.5. This shows the importance of geographical neighbourhood relational weighting parameters in measuring users’ preference for the recommended POI and the characteristics of the geographical neighbourhood.

5.5 Comparison with Other Methods

To verify the superiority of our algorithm, it is comprehensively compared with existing methods on the NY and CHI datasets. We ensure that the experiment on the training object is conducted under conditions unrelated to those of the test object. The experimental results are presented in Table 3.

As can be seen in Tables 3 and 4, compared with several other existing POI recommendation algorithms, the POIRA algorithm shows the best recommendation performance in terms of the F1 value and precision rate. As the number of POI increases, the precision rate and F1 value continue to decline. This is because the POIRA algorithm recommends more POI to users, which helps users find more POI. This will encourage users to be more willing to check in POI, thereby improving the precision and F1 of the model.

6. Conclusion

In this study, a new content-aware POIRA based on a CNN is proposed. Based on the matrix decomposition model, the algorithm uses CNNs to extract the content features of comment texts and model the comment information of POI. Furthermore, it integrates users’ social relationships and geographical information to recommend location POI. Experimental results show that the content-aware POIRA based on a CNN improves the precision of the recommendation effectively.

In practical applications, the interest preferences of users often have an impact on POI owing to changes in time. However, this factor is not considered in the proposed method. Therefore, to the method of considering the time factor and modelling the dynamic changes of user interest preferences will be an important research direction in future work.