A tourist attraction recommendation method integrating attribute co-occurrence and interactive behavior features

Abstract

The present invention relates to the technical field of tourist attraction recommendation, and discloses a tourist attraction recommendation method that integrates attribute co-occurrence and interactive behavior features. First, an attribute co-occurrence graph of users and attractions is constructed, and homogeneous attribute co-occurrence features of users and attractions are learned through a graph neural network. Then, the homogeneous attribute co-occurrence features of users and attractions are matched to learn heterogeneous attribute co-occurrence features. The homogeneous and heterogeneous attribute co-occurrence features are integrated to obtain attribute co-occurrence feature representations. The interactive behavior feature representations are obtained through the interactive behavior of users and attractions. The attribute co-occurrence features and interactive behavior features are integrated to obtain feature representations of users and attractions. The user's rating of candidate attractions is calculated through inner product, and a recommended attraction list is generated based on the ratings. The reasons for recommendation are annotated for attractions based on the user's co-occurrence attributes and the attraction's co-occurrence attributes. The present invention explains users' preferences for attractions based on user co-occurrence attributes and attraction co-occurrence attributes, providing great support for mining potential tourists at tourist attractions.

Description

Technical Field

[0001] The present invention relates to the technical field of tourist attraction recommendation, and in particular to a tourist attraction recommendation method integrating attribute co-occurrence and interactive behavior features. Background Art

[0002] In tourism scenarios, tourists and scenic spots often possess different attribute types. For example, tourists possess social attributes, including gender, age, hobbies, and occupation; while scenic spots possess natural and social attributes, including resource type, location, quality rating, and ticket price. In tourist attraction recommendation systems, these social attributes reflect the user's potential interests to a certain extent, while the attraction's categorical attributes reflect the attraction's potential appeal. This combined user and attraction attribute information can be used to model and explain the relationship between user ratings and attraction ratings. Existing tourist attraction recommendation systems focus on improving recommendation performance, neglecting the reasons behind user preferences, making it difficult for users to achieve full satisfaction with recommended attractions. Therefore, it is crucial to effectively utilize the attribute information of users and attractions to achieve feature modeling of both users and attractions, thereby learning to predict user ratings and achieving accurate, personalized tourist attraction recommendations. Summary of the Invention

[0003] In view of the fact that existing tourism recommendation systems ignore the role of social attribute information of users and attractions in improving user satisfaction, the present invention provides a tourist attraction recommendation method that integrates attribute co-occurrence and interactive behavior feature information.

[0004] In order to achieve the above object, the present invention adopts the following technical solutions:

[0005] The present invention provides a tourist attraction recommendation method that integrates attribute co-occurrence and interactive behavior feature information, comprising the following steps:

[0006] Step S1: construct a user attribute graph based on the user's attribute set, and learn the co-occurrence feature representation of the user's homogeneous attributes based on the graph neural network;

[0007] Step S2: constructing a scenic spot attribute graph based on the attribute set of the scenic spot, and learning the co-occurrence feature representation of homogeneous attributes of the scenic spot based on the graph neural network;

[0008] Step S3, matching the user's homogeneous attribute co-occurrence features with the scenic spot's homogeneous attribute co-occurrence features through the attention mechanism, and modeling the user's heterogeneous attribute co-occurrence feature representation;

[0009] Step S4, fusing the user's homogeneous attribute co-occurrence feature representation and the heterogeneous attribute co-occurrence feature representation to model the user's attribute co-occurrence feature representation;

[0010] Step S5, learning the user's interactive behavior feature representation based on the user's interactive behavior towards the scenic spot;

[0011] Step S6: fusing the attribute co-occurrence feature representation and the interaction behavior feature representation to model the user's feature representation;

[0012] Step S7: Based on steps S1-S6, similarly fusing the attribute co-occurrence feature representation and the interaction behavior feature representation of the scenic spot to model the feature representation of the scenic spot;

[0013] Step S8, calculating the user's rating of the scenic spot by inner product based on the user's feature representation and the feature representation of the scenic spot;

[0014] Step S9, sorting the candidate attractions according to the user's ratings to generate a list of recommended attractions;

[0015] Step S10: annotate the recommended scenic spots with the reasons for recommendation based on the co-occurrence attributes of the user and the scenic spots.

[0016] Furthermore, in step S1, a user attribute graph is constructed based on the user's attribute set, and the co-occurrence feature representation of the user's homogeneous attributes is learned based on the graph neural network. The specific steps are:

[0017] Step 1.1: Users have attribute information such as gender, occupation, and age. These attributes are related to each other. Based on this information, the attribute graph U of each user is constructed. A , where the nodes represent the attributes of the users and the embedding is represented as u A ,edges represent the co-occurrence relationship between pairs of homogeneous attributes of users;

[0018] Step 1.2: Based on the graph attention mechanism, we learn different attribute feature combinations in the graph and model the co-occurrence correlation between user attributes. The co-occurrence correlation between user attributes i and j is expressed as:

[0019]

[0020] Among them, W hom , a is the weight parameter of the neural network, || represents splicing, and σ is the LeakyReLU activation function;

[0021] Step 1.3: In the user's attribute set, define the attention weight between user attributes i and j as:

[0022]

[0023] in, is the normalized attention weight of the co-occurrence correlation between attributes i and j in all attributes of the user, N u is the attribute set of user u;

[0024] In step 1.4, considering the correlation between user attribute i and all other attributes, the co-occurrence feature of user attribute i is defined as:

[0025]

[0026] in, is the normalized attention weight of the co-occurrence correlation between user attributes i and j in all user attributes, which describes the contribution of attribute j in all user attributes. u is the attribute set of user u.

[0027] Furthermore, in step S2, a scenic spot attribute graph is constructed based on the attribute set of the scenic spot, and the co-occurrence feature representation of homogeneous attributes of the scenic spot is learned based on the graph neural network. The specific steps are as follows:

[0028] Step 2.1: Attractions have attribute information such as location, level, and ticket price. These attributes are related to each other. Based on this information, the attribute graph V of each attraction is constructed. A , where nodes represent the attributes of attractions and embedding is represented as v A ,edges represent the co-occurrence relationship between scenic spot attributes;

[0029] Step 2.2: Based on the graph attention mechanism, we learn different attribute feature combinations in the graph and model the co-occurrence correlation between scenic spot attributes. The co-occurrence correlation between scenic spot attributes m and n is expressed as:

[0030]

[0031] Among them, W hom , a is the weight parameter of the neural network, σ is the LeakyReLU activation function;

[0032] Step 2.3, in the attribute set of the scenic spot, define the attention weight between the scenic spot attributes m and n as:

[0033]

[0034] in, is the normalized attention weight of the co-occurrence correlation between attributes m and n in all attributes of the scenic spot, N v is the attribute set of attraction v;

[0035] In step 2.4, considering the correlation between the attraction attribute m and all other attributes, the co-occurrence feature of the attraction attribute m is defined as:

[0036]

[0037] in, is the normalized attention weight of the co-occurrence correlation between the scenic spot attributes m and n in all the attributes of the scenic spot, which describes the contribution of attribute n in all the attributes of the scenic spot. v is the attribute set of attraction v.

[0038] Furthermore, in step S3, the user's homogeneous attribute co-occurrence features are matched with the homogeneous attribute co-occurrence features of the scenic spots through the attention mechanism to model the user's heterogeneous attribute co-occurrence feature representation. The specific steps are:

[0039] In step 3.1, considering that different attributes of users may prefer different attributes of attractions, the co-occurrence characteristics of heterogeneous attributes of users and attractions are matched through the attention mechanism; the co-occurrence correlation between user attribute i and attraction attribute m is expressed as:

[0040]

[0041] Among them, W het 、b het is the weight parameter of the neural network, ⊙ represents element-by-element multiplication, and σ is the LeakyReLU activation function;

[0042] Step 3.2: In the attribute set of scenic spots, define the attention weight between user attribute i and scenic spot attribute m as:

[0043]

[0044] Among them, ε im is the co-occurrence correlation between user attribute i and attraction attribute m, the normalized attention weight among all attributes of the attraction, N v is the attribute set of attraction v;

[0045] In step 3.3, considering the co-occurrence correlation between user attribute i and all attributes of scenic spots, the heterogeneous attribute co-occurrence feature of user attribute i is defined as:

[0046]

[0047] Among them, ε im is the attention weight between user attribute i and scenic spot attribute m, which describes the contribution of scenic spot attribute m to user attribute i. v is the attribute set of attraction v.

[0048] Furthermore, in step S4, the user's homogeneous attribute co-occurrence feature representation and the heterogeneous attribute co-occurrence feature representation are integrated to model the user's attribute co-occurrence feature representation. The specific steps are:

[0049] In step 4.1, considering the co-occurrence feature representation of homogeneous attributes and the co-occurrence feature representation of heterogeneous attributes of user attribute i, the co-occurrence feature representation of user attribute i is obtained by summing them up:

[0050]

[0051] In step 4.2, considering all attributes of user u, the attribute co-occurrence feature of user u is defined as:

[0052]

[0053] Furthermore, in step S5, the user's interactive behavior characteristics are learned based on the user's interactive behavior towards the scenic spot. The specific steps are:

[0054] According to the user's historical rating of the scenic spot collection R u , the interactive behavior characteristics of learning user u are expressed as:

[0055]

[0056] Among them, e l is the user's historical rating attraction set R u The initialization embedding vector representation of the lth scenic spot in .

[0057] Furthermore, in step S6, the attribute co-occurrence feature representation and the interaction behavior feature representation are integrated to model the user's feature representation. The specific steps are:

[0058] Considering the attribute co-occurrence feature representation and interaction behavior feature representation of user u, they are spliced ​​and fused to obtain the feature representation of user u as follows:

[0059] u=[u Att ,u Int ] (13).

[0060] Furthermore, in step S7, the attribute co-occurrence feature representation and the interactive behavior feature representation of the scenic spot are similarly integrated according to steps S1-S6 to model the feature representation of the scenic spot. The specific steps are:

[0061] In step 7.1, the co-occurrence correlation between the attraction attribute m and the user attribute i is expressed as:

[0062]

[0063] Among them, W het 、b het is the weight parameter of the neural network, ⊙ represents element-by-element multiplication, and σ is the LeakyReLU activation function;

[0064] Step 7.2: Define the attention weight between the attraction attribute m and the user attribute i as:

[0065]

[0066] Among them, ε mi is the co-occurrence correlation between attraction attribute m and user attribute i, the normalized attention weight among all user attributes, N u is the attribute set of user u;

[0067] In step 7.3, considering the co-occurrence correlation between the attraction attribute m and all user attributes, the heterogeneous attribute co-occurrence feature of the attraction attribute m is defined as:

[0068]

[0069] Among them, ε mi is the contribution of scenic spot attribute m to user attribute i, N u is the attribute set of user u;

[0070] In step 7.4, considering the homogeneous attribute co-occurrence feature representation and heterogeneous attribute co-occurrence feature representation of the scenic spot attribute m, sum them up to obtain the attribute co-occurrence feature representation of the scenic spot attribute m:

[0071]

[0072] In step 7.5, considering all the attributes of scenic spot v, the attribute co-occurrence feature of scenic spot v is defined as:

[0073]

[0074] Among them, N v is the attribute set of attraction v;

[0075] Step 7.6, considering the historical rating user set R of the scenic spot v , the interactive behavior characteristics of scenic spot v are defined as:

[0076]

[0077] Among them, e s is the initialization embedding vector representation of the sth user in the historical rating user set of the scenic spot;

[0078] In step 7.7, considering the co-occurrence feature representation of scenic spot attributes and the interactive behavior feature representation, they are spliced ​​and fused to obtain the feature representation of the scenic spot:

[0079] v=[v Att ,v Int ] (20).

[0080] Furthermore, in step S8, the user's rating of the scenic spot is calculated by inner product based on the user's feature representation and the feature representation of the scenic spot. The specific steps are:

[0081] For the user's feature representation u and the scenic spot's feature representation v, the user's rating of the scenic spot is calculated based on the inner product, that is, the rating prediction function is defined as:

[0082] y uv =u T v (21).

[0083] Furthermore, in step S9, the candidate attractions are sorted according to the user's ratings to generate a list of recommended attractions. The specific steps are as follows:

[0084] For a new attraction v' in the candidate set of attractions, the feature representation of user u and attraction v' is used to calculate the user's rating of the attraction through the rating prediction function, and the attractions are sorted according to the ratings, and the top k attractions with higher ratings are selected for recommendation.

[0085] Furthermore, in step S10, the recommended scenic spots are marked with recommendation reasons based on the co-occurrence attributes of the user and the scenic spot. The specific steps are as follows:

[0086] For the recommendation results, according to the attention weight calculation method of step S3, the co-occurrence weight of the matching of user attributes and scenic spot attributes is obtained, and the corresponding recommendation reason explanation is generated for the recommendation result based on the user attributes and scenic spot attributes with the highest co-occurrence weight.

[0087] The present invention further provides a computer-readable storage medium having a computer program stored thereon. When the computer program is executed, the method for recommending tourist attractions by integrating attribute co-occurrence and interactive behavior features is implemented.

[0088] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and runnable on the processor. When the processor executes the computer program, the method for recommending tourist attractions that integrates attribute co-occurrence and interactive behavior features is implemented.

[0089] Compared with the prior art, the present invention has the following advantages:

[0090] The method provided by this invention distinguishes itself from existing methods by designing a semantic embedding representation method for users and attractions that simultaneously integrates user and attraction attributes and interactive behavior feature information; establishing a rating prediction model that enhances user and attraction attributes; proposing a tourist attraction recommendation strategy that integrates attribute co-occurrence and interactive behavior features; and interpreting recommendation results based on user and attraction attributes. This method integrates homogeneous attribute co-occurrence feature representations of user and attraction attributes with heterogeneous attribute co-occurrence feature representations to learn user interest preferences and the attractiveness of items. It also explains user rating behavior for attractions based on user and attraction attributes, improving the interpretability of attraction recommendations and providing significant support for online recommendation predictions of popular tourist attractions. BRIEF DESCRIPTION OF THE DRAWINGS

[0091] Figure 1 Schematic diagram of the overall model architecture of the method of the present invention.

[0092] Figure 2 A visualization of the co-occurrence correlation weights of scenic spot attributes. DETAILED DESCRIPTION

[0093] The method for recommending tourist attractions by integrating attribute co-occurrence and interactive behavior features of the present invention is implemented by a computer program. The specific implementation of the technical solution proposed by the present invention will be described in detail below according to the process.

[0094] like Figure 1 As shown, the tourist attraction recommendation method integrating attribute co-occurrence and interactive behavior features of the present invention includes the following steps:

[0095] Step S1: construct a user attribute graph based on the user's attribute set, and learn the co-occurrence feature representation of the user's homogeneous attributes based on the graph neural network. Specifically, the following steps are included:

[0096] Step 1.1: Users have attribute information such as gender, occupation, and age. These attributes are related to each other. Based on this information, the attribute graph U of each user is constructed. A , where the nodes represent the attributes of the users and the embedding is represented as u A ,edges represent the co-occurrence relationship between pairs of homogeneous attributes of users;

[0097] Step 1.2: Based on the graph attention mechanism, we learn different attribute feature combinations in the graph and model the co-occurrence correlation between user attributes. The co-occurrence correlation between user attributes i and j is expressed as:

[0098]

[0099] Among them, W hom , a is the weight parameter of the neural network, || represents splicing, and σ is the LeakyReLU activation function;

[0100] Step 1.3: In the user's attribute set, define the attention weight between user attributes i and j as:

[0101]

[0102] in, is the normalized attention weight of the co-occurrence correlation between attributes i and j in all attributes of the user, N u is the attribute set of user u;

[0103] In step 1.4, considering the correlation between user attribute i and all other attributes, the co-occurrence feature of user attribute i is defined as:

[0104]

[0105] in, is the normalized attention weight of the co-occurrence correlation between user attributes i and j in all user attributes, which describes the contribution of attribute j in all user attributes. u is the attribute set of user u.

[0106] Step S2: constructing a scenic spot attribute graph based on the attribute set of the scenic spot, and learning the co-occurrence feature representation of homogeneous attributes of the scenic spot based on the graph neural network, which specifically includes the following steps:

[0107] Step 2.1: Attractions have attribute information such as location, level, and ticket price. These attributes are related to each other. Based on this information, the attribute graph V of each attraction is constructed. A , where nodes represent the attributes of attractions and embedding is represented as v A ,edges represent the co-occurrence relationship between scenic spot attributes;

[0108] Step 2.2: Based on the graph attention mechanism, we learn different attribute feature combinations in the graph and model the co-occurrence correlation between scenic spot attributes. The co-occurrence correlation between scenic spot attributes m and n is expressed as:

[0109]

[0110] Among them, W hom , a is the weight parameter of the neural network, σ is the LeakyReLU activation function;

[0111] Step 2.3, in the attribute set of the scenic spot, define the attention weight between the scenic spot attributes m and n as:

[0112]

[0113] in, is the normalized attention weight of the co-occurrence correlation between attributes m and n in all attributes of the scenic spot, N v is the attribute set of attraction v;

[0114] In step 2.4, considering the correlation between the attraction attribute m and all other attributes, the co-occurrence feature of the attraction attribute m is defined as:

[0115]

[0116] in, is the normalized attention weight of the co-occurrence correlation between the scenic spot attributes m and n in all the attributes of the scenic spot, which describes the contribution of attribute n in all the attributes of the scenic spot. v is the attribute set of attraction v.

[0117] Step S3, matching the user's homogeneous attribute co-occurrence features with the scenic spot's homogeneous attribute co-occurrence features through the attention mechanism, and modeling the user's heterogeneous attribute co-occurrence feature representation, including the following steps:

[0118] In step 3.1, considering that different attributes of users may prefer different attributes of attractions, the co-occurrence characteristics of heterogeneous attributes of users and attractions are matched through the attention mechanism; the co-occurrence correlation between user attribute i and attraction attribute m is expressed as:

[0119]

[0120] Among them, W het 、b het is the weight parameter of the neural network, ⊙ represents element-by-element multiplication, and σ is the LeakyReLU activation function;

[0121] Step 3.2: In the attribute set of scenic spots, define the attention weight between user attribute i and scenic spot attribute m as:

[0122]

[0123] Among them, ε im is the co-occurrence correlation between user attribute i and attraction attribute m, the normalized attention weight among all attributes of the attraction, N v is the attribute set of attraction v;

[0124] In step 3.3, considering the co-occurrence correlation between user attribute i and all attributes of scenic spots, the heterogeneous attribute co-occurrence feature of user attribute i is defined as:

[0125]

[0126] Among them, ε imis the attention weight between user attribute i and scenic spot attribute m, which describes the contribution of scenic spot attribute m to user attribute i. v is the attribute set of attraction v.

[0127] Step S4, fusing the user's homogeneous attribute co-occurrence feature representation and the heterogeneous attribute co-occurrence feature representation to model the user's attribute co-occurrence feature representation, includes the following steps:

[0128] In step 4.1, considering the co-occurrence feature representation of homogeneous attributes and the co-occurrence feature representation of heterogeneous attributes of user attribute i, the co-occurrence feature representation of user attribute i is obtained by summing them up:

[0129]

[0130] In step 4.2, considering all attributes of user u, the attribute co-occurrence feature of user u is defined as:

[0131]

[0132] Among them, N u is the attribute set of user u.

[0133] Step S5, based on the user's interactive behavior towards the scenic spot, learning the user's interactive behavior feature representation, includes the following steps:

[0134] According to the user's historical rating of the scenic spot collection R u , the interactive behavior characteristics of learning user u are expressed as:

[0135]

[0136] Among them, e l is the user's historical rating attraction set R u The initialization embedding vector representation of the lth scenic spot in .

[0137] Step S6, fusing the attribute co-occurrence feature representation and the interaction behavior feature representation to model the user's feature representation, includes the following steps:

[0138] Taking into account the co-occurrence feature representation of user attributes and the feature representation of interactive behavior, they are spliced ​​and fused to obtain the user feature representation as follows:

[0139] u=[u Att ,u Int ] (13).

[0140] Step S7, based on steps S1-S6, similarly fuses the attribute co-occurrence feature representation and the interaction behavior feature representation of the scenic spot to model the feature representation of the scenic spot, including the following steps:

[0141] In step 7.1, the co-occurrence correlation between the attraction attribute m and the user attribute i is expressed as:

[0142]

[0143] Among them, W het 、b het is the weight parameter of the neural network, ⊙ represents element-by-element multiplication, and σ is the LeakyReLU activation function;

[0144] Step 7.2: Define the attention weight between the attraction attribute m and the user attribute i as:

[0145]

[0146] Among them, ε mi is the co-occurrence correlation between attraction attribute m and user attribute i, the normalized attention weight among all user attributes, N u is the attribute set of user u;

[0147] In step 7.3, considering the co-occurrence correlation between the attraction attribute m and all user attributes, the heterogeneous attribute co-occurrence feature of the attraction attribute m is defined as:

[0148]

[0149] Among them, ε mi is the contribution of scenic spot attribute m to user attribute i, N u is the attribute set of user u;

[0150] In step 7.4, considering the homogeneous attribute co-occurrence feature representation and heterogeneous attribute co-occurrence feature representation of the scenic spot attribute m, sum them up to obtain the attribute co-occurrence feature representation of the scenic spot attribute m:

[0151]

[0152] In step 7.5, considering all the attributes of scenic spot v, the attribute co-occurrence feature of scenic spot v is defined as:

[0153]

[0154] Among them, N v is the attribute set of attraction v;

[0155] Step 7.6, considering the historical rating user set R of the scenic spot v , the interactive behavior characteristics of scenic spot v are defined as:

[0156]

[0157] Among them, e sis the initialization embedding vector representation of the sth user in the historical rating user set of the scenic spot;

[0158] In step 7.7, considering the co-occurrence feature representation of scenic spot attributes and the interactive behavior feature representation, they are spliced ​​and fused to obtain the feature representation of the scenic spot:

[0159] v=[v Att ,v Int ] (20).

[0160] Step S8, calculating the user's rating of the scenic spot by inner product based on the user's feature representation and the scenic spot's feature representation, includes the following steps:

[0161] For the user's feature representation u and the scenic spot's feature representation v, the user's rating of the scenic spot is calculated based on the inner product, that is, the rating prediction function is defined as:

[0162] y uv =u T v (21).

[0163] Step S9, sorting the candidate attractions according to the user's ratings, and generating a list of recommended attractions, includes the following steps:

[0164] For a new attraction v' in the candidate set of attractions, the feature representation of user u and attraction v' is used to calculate the user's rating of the attraction through the rating prediction function, and the attractions are sorted according to the ratings, and the top k attractions with higher ratings are selected for recommendation.

[0165] Step S10, based on the co-occurrence attributes of the user and the scenic spot, annotating the recommended scenic spot with a recommendation reason, includes the following steps:

[0166] For the recommendation results, according to the attention weight calculation method of step S3, the co-occurrence weight of the matching of user attributes and scenic spot attributes is obtained, and the corresponding recommendation reason explanation is generated for the recommendation result based on the user attributes and scenic spot attributes with the highest co-occurrence weight.

[0167] This embodiment models the characteristics of users and attractions based on their rich attribute information, thereby predicting users' preferences for attractions and explaining the reasons for their preference behavior for attractions based on the co-occurrence attributes of users and attractions. At the same time, users' interactive behavior in rating attractions reveals their direct rating preferences for attractions, and based on historical rating interactions, the characteristics between users and attractions can also be learned. In addition, graph attention is used in deep neural networks to focus on specific input features, which can analyze the importance of various aspects of input features and improve the expressive power of the model. Based on the graph attention mechanism, the contribution of the attribute semantic features of users and attractions in the rating prediction of the tourist attraction recommendation system can be analyzed, and the reasons for users' ratings of attractions can be explored to achieve accurate attraction recommendations.

[0168] To verify the effectiveness of the proposed method, we conducted experiments on the Yelp dataset (https: / / www.kaggle.com / yelp-dataset / yelp-dataset). We filtered out attractions of the data set, such as 'Zoos', 'Museums', 'Amusement Parks', 'Temples', and 'Botanical Gardens', and their users who interacted with them. We then processed the interaction records, ensuring that each user had at least five reviews and each attraction had been reviewed at least five times. The filtered dataset information is shown in Table 1:

[0169] Table 1 Dataset

[0170]

[0171] The user and attraction attributes used include user ID, project ID, attraction city, attraction score, and attraction type. Experiments were conducted using a dataset split into training, validation, and test sets (6:2:2). Evaluation metrics used included AUC, LogLoss, NDCG@5, and NDCG@10. To verify the effectiveness and advancement of the proposed technical solution, several existing prediction model methods were selected for comparison: WDL, DeepFM, DCN, and AFM. The experimental results are shown in Table 2.

[0172] Table 2 Experimental results of different prediction models

[0173]

[0174] It can be seen from the results in Table 2 that the technical solution of the present invention can obtain detection results with better accuracy and reliability than the existing methods when predicting users' attractions.

[0175] At the same time, the present invention randomly selects 32 scenic spots and visualizes the co-occurrence correlation weights of scenic spot attributes, such as Figure 2As shown in the figure, the attention weights between pairs of attraction attributes show that the co-occurrence correlations of different attributes contribute differently to learning users' attraction preferences. For example, the weights between the attributes in the fifth column, namely the high co-occurrence between amusement parks and children, can generate corresponding recommendation explanations, such as "This attraction has amusement park attribute and may be of interest to children." The weights between the attribute pairs in the ninth column, namely the high co-occurrence between the humanities type and the elderly type, can be interpreted as "This attraction belongs to the humanities type and may be attractive to the elderly." In the field of actual travel recommendations, we can generate corresponding explanations based on user attributes (gender, age, occupation) and attraction attributes (type, ticket price, location), such as "Recommendations based on your age and attraction type" or "Recommendations based on your occupation and attraction ticket price."

[0176] Example 2

[0177] This embodiment provides a computer-readable storage medium having a computer program stored thereon. When the computer program is executed, the method for recommending tourist attractions by integrating attribute co-occurrence and interactive behavior features of the above-mentioned embodiment 1 is implemented.

[0178] Example 3

[0179] This embodiment provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the method for recommending tourist attractions by integrating attribute co-occurrence and interactive behavior features of the above-mentioned embodiment 1 is implemented.

[0180] The above embodiments are preferred implementations of the present invention, but the implementation of the present invention is not limited to the above embodiments. For ordinary technicians in this field, several modifications and improvements can be made without departing from the principles of the present invention, which are all included in the scope of protection of the present invention.

[0181] In addition, it should be understood that although this specification is described in terms of implementation methods, not every implementation method contains only one independent technical solution. This narrative method of the specification is only for the sake of clarity. Those skilled in the art should regard the specification as a whole. The technical solutions in each embodiment can also be appropriately combined to form other implementation methods that can be understood by those skilled in the art.

Claims

1. A tourist attraction recommendation method that integrates attribute co-occurrence and interactive behavior features, characterized by: The following steps are involved: Step S1: construct a user attribute graph based on the user's attribute set, and learn the co-occurrence feature representation of the user's homogeneous attributes based on the graph neural network; Step S2: constructing a scenic spot attribute graph based on the attribute set of the scenic spot, and learning the co-occurrence feature representation of homogeneous attributes of the scenic spot based on the graph neural network; Step S3: Match the user's attribute co-occurrence features with the scenic spot's attribute co-occurrence features through the attention mechanism to model the user's heterogeneous attribute co-occurrence feature representation; Step S4, fusing the user's homogeneous attribute co-occurrence feature representation and the heterogeneous attribute co-occurrence feature representation to model the user's attribute co-occurrence feature representation; Step S5, learning the user's interactive behavior feature representation based on the user's interactive behavior towards the scenic spot; Step S6: fusing the attribute co-occurrence feature representation and the interaction behavior feature representation to model the user's feature representation; Step S7: Based on steps S1-S6, similarly fusing the attribute co-occurrence feature representation and the interaction behavior feature representation of the scenic spot to model the feature representation of the scenic spot; Step S8, calculating the user's rating of the scenic spot by inner product based on the user's feature representation and the feature representation of the scenic spot; Step S9, sorting the candidate attractions according to the user's ratings to generate a list of recommended attractions; Step S10: annotate the recommended scenic spots with the reasons for recommendation based on the co-occurrence attributes of the user and the scenic spots.

2. The method for recommending tourist attractions by integrating attribute co-occurrence and interactive behavior features according to claim 1, characterized in that: In step S1, a user attribute graph is constructed based on the user's attribute set, and the co-occurrence feature representation of the user's homogeneous attributes is learned based on the graph neural network. The specific steps are: Step 1.1: Users have attribute information such as gender, occupation, and age. These attributes are related to each other. Based on this information, the attribute graph U of each user is constructed. A , where the nodes represent the attributes of the users and the embedding is represented as u A ,edges represent the co-occurrence relationship between pairs of homogeneous attributes of users; Step 1.2: Based on the graph attention mechanism, we learn different attribute feature combinations in the graph and model the co-occurrence correlation between user attributes. The co-occurrence correlation between user attributes i and j is expressed as: Among them, W hom , a is the weight parameter of the neural network, || represents splicing, and σ is the LeakyReLU activation function; Step 1.3: In the user's attribute set, define the attention weight between user attributes i and j as: in, is the normalized attention weight of the co-occurrence correlation between attributes i and j in all attributes of the user, N u is the attribute set of user u; In step 1.4, considering the correlation between user attribute i and all other attributes, the co-occurrence feature of user attribute i is defined as: in, is the normalized attention weight of the co-occurrence correlation between user attributes i and j in all user attributes, which describes the contribution of attribute j in all user attributes. u is the attribute set of user u.

3. The method for recommending tourist attractions by integrating attribute co-occurrence and interactive behavior features according to claim 2, characterized in that: In step S2, a scenic spot attribute graph is constructed based on the attribute set of the scenic spot, and the co-occurrence feature representation of homogeneous attributes of the scenic spot is learned based on the graph neural network. The specific steps are as follows: Step 2.1: Attractions have location, level, and ticket attribute information. These attributes are related to each other. Based on this information, the attribute graph V of each attraction is constructed. A , where nodes represent the attributes of attractions and embedding is represented as v A ,edges represent the co-occurrence relationship between scenic spot attributes; Step 2.2: Based on the graph attention mechanism, we learn different attribute feature combinations in the graph and model the co-occurrence correlation between scenic spot attributes. The co-occurrence correlation between scenic spot attributes m and n is expressed as: Among them, W hom , a is the weight parameter of the neural network, σ is the LeakyReLU activation function; Step 2.3, in the attribute set of the scenic spot, define the attention weight between the scenic spot attributes m and n as: in, is the normalized attention weight of the co-occurrence correlation between attributes m and n in all attributes of the scenic spot, N v is the attribute set of attraction v; In step 2.4, considering the correlation between the attraction attribute m and all other attributes, the co-occurrence feature of the attraction attribute m is defined as: in, is the normalized attention weight of the co-occurrence correlation between the scenic spot attributes m and n in all the attributes of the scenic spot, which describes the contribution of attribute n in all the attributes of the scenic spot. v is the attribute set of attraction v.

4. The method for recommending tourist attractions by integrating attribute co-occurrence and interactive behavior features according to claim 3, characterized in that: In step S3, the user's attribute co-occurrence features are matched with the attribute co-occurrence features of the scenic spot through the attention mechanism to model the user's heterogeneous attribute co-occurrence feature representation. The specific steps are: Step 3.1: Considering that different attributes of users may prefer different attributes of attractions, the attention mechanism is used to match the co-occurrence features of heterogeneous attributes of users and attractions; The co-occurrence correlation between user attribute i and attraction attribute m is expressed as: Among them, W het 、b het is the weight parameter of the neural network, ⊙ represents element-by-element multiplication, and σ is the LeakyReLU activation function; Step 3.2: In the attribute set of scenic spots, define the attention weight between user attribute i and scenic spot attribute m as: Among them, ε im is the co-occurrence correlation between user attribute i and attraction attribute m, the normalized attention weight among all attributes of the attraction, N v is the attribute set of attraction v; In step 3.3, considering the co-occurrence correlation between user attribute i and all attributes of scenic spots, the heterogeneous attribute co-occurrence feature of user attribute i is defined as: Among them, ε im is the attention weight between user attribute i and scenic spot attribute m, which describes the contribution of scenic spot attribute m to user attribute i. v is the attribute set of attraction v.

5. The method for recommending tourist attractions by integrating attribute co-occurrence and interactive behavior features according to claim 4, characterized in that: In step S4, the user's homogeneous attribute co-occurrence feature representation and the heterogeneous attribute co-occurrence feature representation are integrated to model the user's attribute co-occurrence feature representation. The specific steps are: In step 4.1, considering the co-occurrence feature representation of homogeneous attributes and the co-occurrence feature representation of heterogeneous attributes of user attribute i, the co-occurrence feature representation of user attribute i is obtained by summing them up: In step 4.2, considering all attributes of user u, the attribute co-occurrence feature of user u is defined as: Among them, N u is the attribute set of user u.

6. The tourist attraction recommendation method integrating attribute co-occurrence and interactive behavior features according to claim 5 is characterized by: In step S5, based on the user's interactive behavior towards the scenic spot, the user's interactive behavior feature representation is learned, and the specific steps are as follows: According to the user's historical rating of the scenic spot collection R u , the interactive behavior characteristics of learning user u are expressed as: Among them, e l is the user's historical rating attraction set R u The initialization embedding vector representation of the lth scenic spot in .

7. The tourist attraction recommendation method integrating attribute co-occurrence and interactive behavior features according to claim 6 is characterized by: In step S6, the attribute co-occurrence feature representation and the interaction behavior feature representation are integrated to model the user's feature representation. The specific steps are: Considering the attribute co-occurrence feature representation and interaction behavior feature representation of user u, they are spliced ​​and fused to obtain the feature representation of user u as follows: in=[in Att ,in Int ] (13).

8. The tourist attraction recommendation method integrating attribute co-occurrence and interactive behavior features according to claim 7 is characterized by: In step S7, the attribute co-occurrence feature representation and the interaction behavior feature representation of the scenic spot are similarly integrated according to steps S1-S6 to model the feature representation of the scenic spot. The specific steps are: In step 7.1, the co-occurrence correlation between the attraction attribute m and the user attribute i is expressed as: Among them, W het 、b het is the weight parameter of the neural network, ⊙ represents element-by-element multiplication, and σ is the LeakyReLU activation function; Step 7.2: Define the attention weight between the attraction attribute m and the user attribute i as: Among them, ε mi is the co-occurrence correlation between attraction attribute m and user attribute i, the normalized attention weight among all user attributes, N u is the attribute set of user u; In step 7.3, considering the co-occurrence correlation between the attraction attribute m and all user attributes, the heterogeneous attribute co-occurrence feature of the attraction attribute m is defined as: Among them, ε mi is the contribution of scenic spot attribute m to user attribute i, N u is the attribute set of user u; In step 7.4, considering the homogeneous attribute co-occurrence feature representation and heterogeneous attribute co-occurrence feature representation of the scenic spot attribute m, sum them up to obtain the attribute co-occurrence feature representation of the scenic spot attribute m: In step 7.5, considering all the attributes of scenic spot v, the attribute co-occurrence feature of scenic spot v is defined as: Among them, N v is the attribute set of attraction v; Step 7.6, considering the historical rating user set R of the scenic spot v , the interactive behavior characteristics of scenic spot v are defined as: Among them, e s is the initialization embedding vector representation of the sth user in the historical rating user set of the scenic spot; In step 7.7, considering the co-occurrence feature representation of scenic spot attributes and the interactive behavior feature representation, they are spliced ​​and fused to obtain the feature representation of the scenic spot: v=[v Att ,v Int ] (20)。 9. The tourist attraction recommendation method integrating attribute co-occurrence and interactive behavior features according to claim 8, characterized in that: In step S8, the user's rating of the scenic spot is calculated by inner product based on the user's feature representation and the feature representation of the scenic spot. The specific steps are: For the user's feature representation u and the scenic spot's feature representation v, the user's rating of the scenic spot is calculated based on the inner product, that is, the rating prediction function is defined as: y uv =u T v (21)。 10. The tourist attraction recommendation method integrating attribute co-occurrence and interactive behavior features according to claim 9, characterized in that: In step S9, the candidate attractions are sorted according to the user's ratings to generate a list of recommended attractions. The specific steps are as follows: For a new attraction v' in the candidate set of attractions, the user's rating of the attraction is calculated using the feature representation of user u and attraction v' through the rating prediction function. The attractions are sorted according to the ratings and the top k attractions with the highest ratings are selected for recommendation. In step S10, the recommended attractions are marked with recommendation reasons based on the co-occurrence attributes of the user and the attractions. The specific steps are as follows: For the recommendation results, according to the attention weight calculation method of step S3, the co-occurrence weight of the matching of user attributes and scenic spot attributes is obtained, and the corresponding recommendation reason explanation is generated for the recommendation result based on the user attributes and scenic spot attributes with the highest co-occurrence weight.

Images