Point of interest recommendation method and apparatus

Abstract

The application provides a point of interest recommendation method and device, the point of interest recommendation method comprises the following steps: obtaining user access information and a point of interest set; expanding a short sequence in the user access information to obtain an enhanced user access information sequence; inputting the enhanced user access information sequence and the point of interest set into a preset point of interest recommendation model to obtain a target point of interest recommendation result of a user. The method can sufficiently learn the preference information of the user visiting the point of interest, avoid the error caused by manual misjudgment, and improve the accuracy of the point of interest recommendation of the user.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to a method and apparatus for recommending points of interest. Background Technology

[0002] With the development of the "Internet + Tourism" concept, tourists have more options to choose from, making it increasingly important to recommend information that users are interested in in order to enhance their travel experience.

[0003] In related technologies, complex feature engineering is typically performed on users and items. This requires experts in the relevant fields to introduce a large amount of prior knowledge and integrate this prior knowledge into the feature engineering process to learn more user preference information. When a user has multiple points of interest, more expert prior knowledge needs to be introduced for guidance. The process of feature engineering users and items becomes more cumbersome, and the prior knowledge provided by experts is limited, which increases the error caused by human misjudgment, resulting in inaccurate information on tourist attractions recommended to users. Summary of the Invention

[0004] This invention provides a method and apparatus for recommending points of interest, which solves the problems of the cumbersome process of using expert prior knowledge to guide feature engineering in the prior art, the limited prior knowledge provided by experts, the increase in errors caused by human misjudgment, and the resulting in inaccurate information on points of interest recommended to users, thereby improving the accuracy of user point of interest recommendations.

[0005] This invention provides a method for recommending points of interest, comprising:

[0006] Obtain user access information and a set of points of interest, wherein the user access information includes the target points of interest accessed by the user, and the set of points of interest is associated with the target points of interest;

[0007] The short sequence in the user access information is expanded to obtain an enhanced user access information sequence;

[0008] The enhanced user access information sequence and the set of interest points are input into a preset interest point recommendation model to obtain the user's target interest point recommendation result. The interest point recommendation model is based on an attention mechanism.

[0009] According to the present invention, a method for recommending points of interest includes obtaining user access information and a set of points of interest, comprising:

[0010] Based on user access records, the user access information is obtained; based on the tourism event graph, the set of points of interest is obtained; the tourism event graph is obtained by data cleaning, event extraction, and event time sequence information extraction of tourism text data; and the user access records are obtained based on call detail records (CDRs).

[0011] According to the point-of-interest recommendation method provided by the present invention, the step of augmenting the user information with short sequences to obtain enhanced user information sequences includes:

[0012] The user access information is input into a preset pre-trained model to obtain an enhanced user access information sequence. The pre-trained model is obtained by training a Transformer network based on the reverse sequence of the sample user access information.

[0013] According to the present invention, an interest point recommendation method is provided, wherein the interest point recommendation model includes an embedding layer, an embedding propagation layer, and a prediction layer;

[0014] The step of inputting the enhanced user access information sequence and the set of interest points into a preset interest point recommendation model to obtain the target interest point recommendation result includes:

[0015] Based on the embedding layer, the user access information sequence is encoded to obtain a first vector representation, and each interest point in the interest point set is encoded to obtain a second vector representation;

[0016] Based on the embedded propagation layer and the first vector representation, a personalized profile of the user is obtained. Based on the personalized profile and the first vector representation, a user vector representation is obtained. The personalized profile is obtained based on the user's access to interest points, interest point types, and user access time corresponding to the first vector representation.

[0017] Based on the embedding propagation layer, the interest points corresponding to the second vector representation are aggregated to obtain the aggregation result, and the attention score of the aggregation result is calculated to obtain the neighbor representation of the second vector representation. Based on the second vector representation and the neighbor representation, the interest point vector representation is obtained.

[0018] Based on the prediction layer, the matching degree between the user vector representation and the interest point vector representation is scored to obtain the target interest point recommendation result.

[0019] According to a method for recommending points of interest provided by the present invention, the target point of interest recommendation result includes multiple points of interest, the user access information includes user identity information, and after obtaining the target point of interest recommendation result, the method further includes:

[0020] Obtain real-time pedestrian flow information, residential area distribution information, and crowd movement trajectory information for each point of interest;

[0021] Based on the user identity information and cluster centroids, a user's geographic location preference score is determined. The cluster centroids are obtained by clustering user type information based on user geographic location preference features and user side attribute features. The user geographic location preference features are obtained based on user access information, and the user side attribute features are obtained based on the user's side attribute information.

[0022] Based on the real-time pedestrian flow information of each point of interest, the population density score of each point of interest is obtained; based on the residential area distribution information of each point of interest, the residential area overlap score of each point of interest is obtained; based on the pedestrian movement trajectory information of each point of interest, the population density prediction score of each point of interest is obtained.

[0023] Based on the user's geographic location preference score, the population density score of each point of interest, the residential area overlap score of each point of interest, and the population density prediction score of each point of interest, a comprehensive score is obtained for each point of interest corresponding to the user.

[0024] Based on the comprehensive score of each interest point, the optimized target interest point recommendation result is obtained.

[0025] The present invention also provides an interest point recommendation device, comprising:

[0026] The information acquisition module is used to acquire user access information and a set of points of interest, wherein the user access information includes the target points of interest accessed by the user, and the set of points of interest is associated with the target points of interest.

[0027] A sequence augmentation module is used to augment the short sequences in the user access information to obtain an enhanced user access information sequence;

[0028] The point of interest recommendation module inputs the enhanced user access information sequence and the point of interest set into a preset point of interest recommendation model to obtain the user's target point of interest recommendation result. The point of interest recommendation model is based on an attention mechanism.

[0029] According to the present invention, an interest point recommendation device is provided, wherein the target interest point recommendation result includes multiple interest points, the user access information includes user identity information, and the device further includes:

[0030] The recommendation results optimization module is used to obtain real-time pedestrian flow information, residential area distribution information, and crowd movement trajectory information for each point of interest;

[0031] Based on the user identity information and cluster centroids, a user's geographic location preference score is determined. The cluster centroids are obtained by clustering user type information based on user geographic location preference features and user side attribute features. The user geographic location preference features are obtained based on user access information, and the user side attribute features are obtained based on the user's side attribute information.

[0032] Based on the real-time pedestrian flow information of each point of interest, the population density score of each point of interest is obtained; based on the residential area distribution information of each point of interest, the residential area overlap score of each point of interest is obtained; based on the pedestrian movement trajectory information of each point of interest, the population density prediction score of each point of interest is obtained.

[0033] Based on the user's geographic location preference score, the population density score of each point of interest, the residential area overlap score of each point of interest, and the population density prediction score of each point of interest, a comprehensive score is obtained for each point of interest corresponding to the user.

[0034] Based on the comprehensive score of each interest point, the optimized target interest point recommendation result is obtained.

[0035] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the point of interest recommendation method as described above.

[0036] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the point of interest recommendation method as described above.

[0037] The present invention also provides a computer program product, including a computer program that, when executed by a processor, implements the point of interest recommendation method as described above.

[0038] The method and apparatus for recommending points of interest provided by this invention expands user access information with short sequences to ensure sufficient data content for the input model. The expanded and enhanced user access sequences and associated points of interest are then input into the point of interest recommendation model for prediction. The target point of interest recommendation result can fully learn the user's preference information for accessing points of interest, reduce errors caused by human misjudgment, and improve the accuracy of user point of interest recommendation. Attached Figure Description

[0039] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0040] Figure 1 This is one of the flowcharts of the point of interest recommendation method provided by the present invention;

[0041] Figure 2 This is a schematic diagram of the short sequence augmentation algorithm based on a pre-trained model provided by the present invention;

[0042] Figure 3 This is the second flowchart of the point-of-interest recommendation method provided by the present invention;

[0043] Figure 4 This is a flowchart illustrating the method for obtaining the vector representation of interest points provided by this invention;

[0044] Figure 5 This is a flowchart illustrating the method for optimizing scenic area recommendation results provided by the present invention;

[0045] Figure 6 This is a schematic diagram of the structure of the point-of-interest recommendation device provided by the present invention;

[0046] Figure 7 This is a schematic diagram of the physical structure of an electronic device provided by the present invention. Detailed Implementation

[0047] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0048] The following is combined Figures 1-6 The present invention describes a method and apparatus for recommending points of interest.

[0049] Figure 1 This is one of the flowcharts illustrating the point-of-interest recommendation method provided by this invention, such as... Figure 1 As shown, this point-of-interest recommendation method includes the following steps:

[0050] Step 110: Obtain user access information and a set of points of interest. The user access information includes the target points of interest visited by the user, and the set of points of interest is associated with the target points of interest. In this step, the user access information includes user identity information, the type and location information of the points of interest visited by the user, and the time information of the user's stay at the points of interest. For example, user A arrives at attraction a at 8:00 AM and leaves attraction a at 8:10 AM.

[0051] In this step, the user access information also includes the user's trajectory information at multiple points of interest. For example, user A visited attractions a, b, and c in sequence in the morning.

[0052] In this embodiment, user access information can be derived from the call detail record (CDR) of the communication device. This information includes the user's historical locations visited and the start and end times of those visits. For example, a user's historical activity information can be represented as:

[0053] H={v1,t 1s ,t 1e ,v2,t 2s ,t 2e ,...,v n ,t ns ,t ne},

[0054] Among them, v1 to v n t represents the n points of interest that the user browses. 1s This indicates the start time of the user browsing the v1st point of interest, t. 1e This indicates the end time for the user to browse the v1st point of interest, and so on.

[0055] In this step, the set of points of interest can be a collection of multiple points of interest that are a certain distance away from the current point of interest accessed by the user. For example, an event graph can be constructed using multiple historical access data of the user, and the set of k-hop neighbors of the current point of interest can be determined from the event graph.

[0056] In some embodiments, the activities that users participate in during their trip can be the focus, and most of the information during the user's travel process can be collected. The spatiotemporal relationship between the user's activities and events during their trip can be modeled, a travel event graph can be constructed, and the required user access information and set of points of interest can be obtained from the travel time graph.

[0057] Step 120: Expand the short sequence in the user access information to obtain an enhanced user access information sequence.

[0058] In this step, if a new user or a user has too little historical access information, it is necessary to expand the user's historical access information; otherwise, it will lead to a cold start problem.

[0059] In this embodiment, a short sequence refers to a sequence whose length is less than a sequence length threshold, which can be set according to user requirements.

[0060] In this embodiment, it is necessary to set a sequence length threshold and an enhanced candidate location number threshold to ensure that the data dimensions of the enhanced user access information sequence are consistent.

[0061] In this embodiment, a self-attention training model can be used to perform short sequence augmentation on user access information. For example, a pre-trained model can be trained using a Transformer network, and the user access information can be augmented using this pre-trained model to obtain an enhanced user access information sequence.

[0062] Step 130: Input the enhanced user access information sequence and interest point set into the preset interest point recommendation model to obtain the user's target interest point recommendation result. The interest point recommendation model is based on the attention mechanism.

[0063] In this step, the interest point recommendation model can be trained by a graph attention network.

[0064] In this embodiment, the point of interest recommendation model may include an embedding layer, an embedding propagation layer, and a prediction layer. The embedding layer is used to receive user access information sequences and point of interest sets, and encode the received information to obtain parameter forms that the model can read and learn. The embedding propagation layer is used to aggregate the parameter parts learned by the model with the original input parts to obtain new parameter representations, thereby improving the accuracy of input data prediction. The prediction layer matches the point of interest that the user has visited and the set of point of interest associated with the point of interest based on user information to obtain the matching results between the user and multiple point of interest.

[0065] In this step, the target interest point recommendation result can be the matching degree between the user and multiple interest points predicted by the interest point recommendation model. For example, the target interest point recommendation result includes matching scores of user A with 5 interest points of 0.7, 0.2, 0.6, 0.9 and 0.85 respectively.

[0066] In this step, the enhanced user access information sequence includes a pseudo-prior user access information sequence.

[0067] In some embodiments, when a user matches too many points of interest, in order to reduce calculations, a matching score threshold can be set to remove points of interest with a low matching score from among multiple points of interest; for example, the matching score threshold can be set to 0.3, then among the above 5 points of interest, the points of interest with a matching score of 0.2 will be removed.

[0068] The interest point recommendation method provided in this invention expands user access information with short sequences to ensure sufficient data content for the input model. The expanded and enhanced user access sequences and associated interest points are then input into the interest point recommendation model for prediction. The target interest point recommendation result can fully learn the user's preference information for accessing interest points, avoid the error rate of human misjudgment, and improve the accuracy of user interest point recommendation.

[0069] In some embodiments, obtaining user access information and a set of points of interest includes: obtaining user access information and a set of points of interest based on a tourism event graph and user access records. The tourism event graph is obtained by data cleaning, event extraction, and event time sequence information extraction of tourism text data, and the user access records are obtained based on call detail records (CDRs).

[0070] In this embodiment, the structured event graph data stored in the tourism event graph is obtained by preprocessing (e.g., data cleaning) unstructured tourism travelogue text data, wherein the tourism text data includes unstructured text data crawled from the Internet.

[0071] In this embodiment, the activities that users participate in during their trip are the central focus, covering most of the activity information during the user's travel process, and the spatiotemporal relationships during the trip are modeled to obtain a travel event map.

[0072] In this embodiment, the tourism event graph can support the W3C standard (SPARQL, RDF). Compared to the event logic graph, the completed graph and the traditional knowledge graph form a system, laying the foundation for subsequent graph fusion.

[0073] In this embodiment, event extraction refers to extracting valuable information from a large amount of unstructured text and transforming it into structured information. For example, event extraction has two basic sub-tasks: (1) event trigger word extraction, which first identifies the keywords that trigger the event from the sentence and classifies them into event types; (2) event argument extraction, which classifies event elements, first identifies specific arguments, and then assigns corresponding roles.

[0074] In this embodiment, an event-triggered word extraction algorithm based on an attention mechanism and an argument extraction algorithm based on machine reading comprehension can be used to extract events in the tourism field.

[0075] In this embodiment, the event relationships in the tourism scenario are defined as four types: Before, After, Vague, and Overlap. An event temporal relationship extraction algorithm based on the BERT pre-trained language model is used to distinguish the event relationships extracted in the previous step, which can enrich the types of event relationships in tourism texts and more accurately construct the spatiotemporal transfer information of users in travelogue texts.

[0076] In this embodiment, the user's access information during the trip includes structured data and unstructured data. Both of these data contain key information for constructing a tourism event map. Therefore, different schemes need to be adopted to extract information from the structured data and unstructured data, and then organize them into a suitable data format according to the defined schema.

[0077] In this embodiment, a general pipeline for extracting valid information from a user's access information during their journey can be included, encompassing everything from web crawling to information extraction, information organization, and information storage.

[0078] In this embodiment, the tourism event graph knowledge graph can be stored in a Resource Description Framework (RDF), a Neo4j database, or a MySQL database. For example, the extracted structured data can be stored in a Neo4j database and queried and used through Cypher query statements.

[0079] In this embodiment, the tourism event graph stores information such as the type of points of interest accessed and the user's historical access trajectory. Based on the points of interest accessed by the user at the current moment, the k-hop neighbor algorithm can be used to obtain the k associated points of interest accessed at the current moment, thus obtaining the set of points of interest.

[0080] The point of interest recommendation method provided in this invention abstracts the feature information in user access information into a graph structure by constructing a tourism event graph. At the same time, it can store the contextual knowledge of user access information, thereby obtaining user access information and point of interest set with rich feature information, which is beneficial to improving the prediction accuracy of subsequent models.

[0081] In this embodiment, the user information is augmented with a short sequence to obtain an enhanced user information sequence, including: inputting user access information into a preset pre-trained model to obtain an enhanced user access information sequence, and the pre-trained model is obtained by training a Transformer network based on the reverse sequence of sample user access information.

[0082] It is understandable that when the user access information contains few feature details, it can easily lead to a cold start problem in the model's feature learning.

[0083] In this embodiment, user access information includes a historical sequence of user access to points of interest.

[0084] Figure 2 This is a schematic diagram of the short sequence augmentation algorithm based on a pre-trained model provided by the present invention. Figure 2 In the illustrated embodiment, a pre-trained model is trained using the reverse sequence of user access point history information (corresponding to sample user access information) and a Transformer network. For the user access point history sequence S U =[v1,v2,…,v n The probability p of the next point of interest in a reverse sequence can be predicted using a pre-trained model, and this probability p can be expressed as:

[0085] p(v0=v|S),

[0086] Where v0 represents the next point of interest in the reverse sequence, and v represents the predicted point of interest.

[0087] It should be noted that although the pre-trained model is trained in reverse, the self-attention module can still reflect the correlation between historical access sequences.

[0088] In some embodiments, it is also necessary to design the threshold M for the short sequence to be expanded and the number k of enhanced candidate locations to unify the data dimensions of user access information and the set of points of interest; for short sequences with a sequence length less than M, the output of the expanded enhanced user sequence is:

[0089] S U =[v -k+1 ,v -k+2 ,…,v0],

[0090] Where, v1, v2, ..., v n Corresponding to POI1, POI2, ..., POI n .

[0091] The interest point recommendation method provided in this invention trains a Transformer network to obtain a pre-trained model by using the reverse sequence of sample user access information, and inputs the user access information into the pre-trained model to obtain an enhanced user access sequence, thereby reducing the cold start phenomenon of the model. At the same time, it unifies the data dimensions of user access information and interest point set, which facilitates the matching and prediction of user access information and interest point set in subsequent processes.

[0092] In some embodiments, the point of interest (POI) recommendation model includes an embedding layer, an embedding propagation layer, and a prediction layer. The enhanced user access information sequence and the set of POIs are input into a preset POI recommendation model to obtain a target POI recommendation result. This includes: based on the embedding layer, encoding the user access information sequence to obtain a first vector representation, and encoding the set of POIs to obtain a second vector representation; based on the embedding propagation layer and the first vector representation, obtaining a personalized user profile; based on the personalized profile and the first vector representation, obtaining a user vector representation, wherein the personalized profile is obtained based on the user's accessed POIs, POI types, and user access time corresponding to the first vector representation; based on the embedding propagation layer, aggregating the POIs corresponding to the second vector representation to obtain an aggregation result, calculating the attention score of the aggregation result to obtain a neighbor representation of the second vector representation, and obtaining a POI vector representation based on the second vector representation and the neighbor representation; and based on the prediction layer, scoring the matching degree between the user vector representation and the POI vector representation to obtain the target POI recommendation result.

[0093] In this embodiment, the user access information sequence includes the user's historical trajectory.

[0094] In this embodiment, the first vector is a low-dimensional vector representation of the entities and relationships involved in the user's historical trajectory, and the second vector is a low-dimensional vector representation of the entities and relationships involved in the tourism event graph.

[0095] Figure 3 This is the second flowchart illustrating the point-of-interest recommendation method provided by this invention. Figure 3 In the illustrated embodiment, the user access sequence is first converted into a spatiotemporal sequence, that is, t s -t e The input includes information such as the POI entity at a given time, POI type, and the start and end times of the POI's stay. This information is then hybridized and encoded before being fed into n LSTM basic units. The outputs of these n LSTM basic units are used as the user vector representation; where user access information can be represented as H = {v1, t}. 1s ,t 1e ,v2,t 2s ,t 2e ,...,v n ,t ns ,t ne}, where v1-v n Let t represent the n attractions that the user is interested in. is Indicates the start time of the user's browsing of the i-th attraction, t ie This represents the end time of the user's browsing of the i-th attraction. Through the embedding layer, the n attractions that the user is interested in are represented as vector units, and the start and end times of the user's browsing of the attractions are encoded to obtain the corresponding vector representations. For example... and By using an embedding propagation layer, the sequential relationship between the corresponding POI entities, POI types, and the user's dwell time on the POI entities in the user access information is combined as the model input. An LSTM layer is applied to capture the user's historical information to obtain a personalized user profile. Finally, the user's own embedding vector is aggregated to obtain the user's final representation vector, i.e., the user vector representation.

[0096] In this embodiment, the input information of the point-of-interest (POI) recommendation model can be enriched by adding type background knowledge to the location POI (Point-of-interest) information involved in the user access information. For example, based on the characteristics of tourist cities, the POI types can be simplified into six types: attractions, bays, shopping, hotels, islands, and performances. The output of the model is the probability value of user u accessing candidate point-of-interest recommendation v. Finally, k POIs are obtained for personalized recommendations for the current user based on the probability ranking. The probability values ​​can be expressed as:

[0097] y uv =f(u,v|G,H,θ),

[0098] Among them, y uv This represents the probability that user u will access candidate point of interest recommendation v given the inputs are user history access information H and tourism event map G. The learning parameter θ is the learning parameter during model training.

[0099] exist Figure 3 In the illustrated embodiment, based on the target POI in the tourism event graph, the k-hop neighbor set of the target POI can be obtained. Then, based on the specificity of the relation path corresponding to each i-hop neighbor set, the attention score of each is calculated to obtain the neighbor representation of the current POI entity. Finally, the neighbor representation is aggregated with the POI entity's own embedding vector to obtain the final vector representation of the project, i.e., the point of interest vector representation.

[0100] In this embodiment, the matching probability between the interest point vector representation and the user vector representation is obtained through the prediction layer, and a top-k recommendation list is provided to the next module based on the predicted probability.

[0101] Specifically, given the final representation u of user u h The final representation of project v p The preference prediction of u for v can be defined as follows:

[0102]

[0103] Where σ is the activation function Predict the probability that user u will access a POI project. Between 0 and 1, select the k candidate POIs with the highest predicted probability from the candidate set in order.

[0104] Figure 4 This is a flowchart illustrating the method for obtaining the vector representation of interest points provided by this invention. Figure 4 In the illustrated embodiment, for a location point of interest (POI) entity corresponding to a single item v, multiple other entities in the tourism event graph are associated with it through different relational paths. These entities can supplement the representation of the current POI entity. It should be noted that utilizing higher-order connectivity is an important means of achieving high-quality recommendations. Through multi-hop relationships, the location entity e corresponding to the current point of interest recommendation v in the tourism event graph can be obtained. v The set of k-jump neighbors.

[0105] In this embodiment, the graph attention network can supplement the item representation with neighbor information. The k-hop neighbor entities and their corresponding k-hop relationship paths are input into the Attention score calculation module, and the calculated normalized score is returned to the corresponding neighbor representation. The user then aggregates the k-hop neighbor representation.

[0106] The point-of-interest (POI) recommendation method provided in this invention, by setting an embedding layer, an embedding propagation layer, and a prediction layer, predicts the matching degree between the enhanced user access information sequence and the POI set, and obtains the target POI recommendation result. It can make full use of the contextual knowledge of the trajectory information of user access POIs in the tourism event graph and the entity and relationship information between the k POI sets associated with the POI, thereby improving the accuracy of the matching degree prediction between user access information and POI sets.

[0107] In some embodiments, the target point of interest recommendation result includes multiple points of interest, and the user access information includes user identity information. After obtaining the target point of interest recommendation result, the method further includes: acquiring real-time pedestrian flow information, residential area distribution information, and crowd movement trajectory information for each point of interest; determining the user's geographic location preference score based on the user identity information and cluster centroid, wherein the cluster centroid is obtained by clustering user type information based on user geographic location preference features and user side attribute features, and the user geographic location preference features are obtained based on user access information, and the user side attribute features are obtained based on user side attribute information; obtaining the crowd density score for each point of interest based on the real-time pedestrian flow information, obtaining the residential area overlap score for each point of interest based on the residential area distribution information, and obtaining the crowd density prediction score for each point of interest based on the crowd movement trajectory information; obtaining the comprehensive score for each point of interest corresponding to the user based on the user's geographic location preference score, the crowd density score for each point of interest, the residential area overlap score for each point of interest, and the crowd density prediction score for each point of interest; and obtaining the optimized target point of interest recommendation result based on the comprehensive score for each point of interest.

[0108] In this embodiment, after obtaining the user's k points of interest recommendations, each point of interest recommendation is comprehensively evaluated from aspects such as user's geographical location preference, the density of the population at the point of interest, the degree of overlap of residents, and the estimated population density of the memory, so as to obtain a comprehensive score for each point of interest.

[0109] In this embodiment, the cluster centroid can be obtained according to the following steps: (1) design a user type clustering algorithm based on the user's historical access records, and design a clustering distance function according to the type of the user's historical access interest points; (2) after completing the clustering, save the cluster centroid data and obtain the user type according to the distance between the user and the centroid of each cluster.

[0110] In this embodiment, the user type clustering algorithm can be the K-means algorithm, or other types of clustering algorithms.

[0111] In this embodiment, a scoring strategy is designed for different interest types corresponding to different user types to obtain user geolocation preference scores.

[0112] Figure 5 This is a flowchart illustrating the scenic area recommendation result optimization method provided by the present invention. Figure 5In the illustrated embodiment, the distance between the requesting user data and the cluster centroid is first obtained to determine the user's geographic location preference score. Then, the movement flow of people in the scenic area is obtained using a mobile simulation system. For each candidate point of interest, predictions are made to obtain the future flow of people in the scenic area and scores are assigned to obtain a population density score. The candidate points of interest are also scored based on changes in the flow of people in the scenic area to obtain a population density score. Additionally, the overlap score between the scenic area and residential areas is calculated based on the overlap between the scenic area and residential areas. Finally, the above four types of scores are weighted and summed to obtain a comprehensive score for each point of interest. The scores are then sorted and used as the personalized recommendation result for the user.

[0113] In this embodiment, the user's geographic location preference score S1 is used. i Crowd density score S2 i Residential area overlap score S3 i Crowd density prediction score S4 i The final score for the candidate scenic spots list is obtained by weighting the four parts together. i .

[0114] Score i =w1*S1 i +w2*S2 i +w3*S3 i +w4*S4 i ;

[0115] Where w1 is S1 i The weights, w2 is S2 i The weights, w3 is S3 i The weights, w4 is S4 i The weights w1, w2, w3, and w4 can all be set according to the user's actual needs.

[0116] In this embodiment, the flow of visitors to scenic spots is statistically analyzed, and different types of scenic spots are classified into different levels. Different crowd flow standards are set for different types of points of interest, with scenic spots with higher crowd density receiving lower scores. The impact of recommendations on local residents is reduced by introducing the overlap of residential areas. The determination of whether candidate points of interest are located in residential areas is made by calculating whether they are located in residential areas, and different scores are set for different areas. The predicted score of the crowd density of scenic spots is mainly based on the scoring of recommended candidate points of interest by crowd movement simulation, which predicts whether it will lead to an increase in the crowd density of scenic spots. The movement intention preferences of large-scale urban populations at different times are introduced by crowd movement simulation, thereby judging the future changes in the flow of people in scenic spots and predicting the score based on the synchronous changes caused by the flow of people in scenic spots.

[0117] The point-of-interest (POI) recommendation method provided in this invention comprehensively evaluates each POI recommendation from aspects such as user geographic location preference, POI population density, resident overlap, and memory population density prediction. It also optimizes the POI recommendation results by combining real-time scenic area traffic information, trajectory features, and techniques such as population movement simulation.

[0118] The interest point recommendation device provided by the present invention is described below. The interest point recommendation device described below can be referred to in correspondence with the interest point recommendation method described above.

[0119] Figure 6 This is a schematic diagram of the structure of the point-of-interest recommendation device provided by the present invention, as shown below. Figure 6 As shown, the point of interest recommendation device includes: an information acquisition module 610, a sequence expansion module 620, and a point of interest recommendation module 630.

[0120] The information acquisition module 610 is used to acquire user access information and a set of points of interest. The user access information includes the target points of interest accessed by the user, and the set of points of interest is associated with the target points of interest.

[0121] The sequence expansion module 620 is used to expand the short sequence in the user access information to obtain an enhanced user access information sequence;

[0122] The point of interest recommendation module 630 inputs the enhanced user access information sequence and the set of points of interest into the preset point of interest recommendation model to obtain the user's target point of interest recommendation result. The point of interest recommendation model is based on the attention mechanism.

[0123] The interest point recommendation device provided in this embodiment of the invention expands user access information with short sequences to ensure sufficient data content for the input model. The expanded and enhanced user access sequence and associated interest points are then input together into the interest point recommendation model for prediction. The target interest point recommendation result can fully learn the user's preference information for accessing interest points, reduce errors caused by human misjudgment, and improve the accuracy of user interest point recommendation.

[0124] According to the present invention, the target interest point recommendation result includes multiple interest points, the user access information includes user identity information, and the device further includes: a recommendation result optimization module 640.

[0125] The recommendation result optimization module 640 is used to obtain real-time pedestrian flow information, residential area distribution information, and crowd movement trajectory information for each point of interest;

[0126] Based on user identity information and cluster centroids, user geolocation preference scores are determined. Cluster centroids are obtained by clustering user type information based on user geolocation preference features and user side attribute features. User geolocation preference features are obtained based on user access information, and user side attribute features are obtained based on user side attribute information.

[0127] Based on the real-time pedestrian flow information of each point of interest, the population density score of each point of interest is obtained; based on the residential area distribution information of each point of interest, the residential area overlap score of each point of interest is obtained; based on the pedestrian movement trajectory information of each point of interest, the population density prediction score of each point of interest is obtained.

[0128] Based on the user's geographic location preference score, the population density score of each point of interest, the residential area overlap score of each point of interest, and the population density prediction score of each point of interest, a comprehensive score is obtained for each point of interest corresponding to the user.

[0129] Based on the comprehensive score of each point of interest, the optimized target point of interest recommendation results are obtained.

[0130] The point of interest recommendation device provided in this invention comprehensively evaluates each point of interest recommendation from aspects such as user geographic location preference, point of interest population density, resident overlap, and memory population density prediction. It also optimizes the point of interest recommendation results by combining real-time traffic information of scenic spots, trajectory features, and techniques such as population movement simulation.

[0131] Figure 7 An example is a schematic diagram of the physical structure of an electronic device, such as... Figure 7 As shown, the electronic device may include a processor 710, a communications interface 720, a memory 730, and a communication bus 740. The processor 710, communications interface 720, and memory 730 communicate with each other via the communication bus 740. The processor 710 can call logical instructions in the memory 730 to execute an interest point recommendation method. This method includes: acquiring user access information and an interest point set; the user access information includes the user's target interest points, and the interest point set is associated with the target interest points; expanding the short sequences in the user access information to obtain an enhanced user access information sequence; and inputting the enhanced user access information sequence and the interest point set into a preset interest point recommendation model to obtain the user's target interest point recommendation result. The interest point recommendation model is based on an attention mechanism.

[0132] Furthermore, the logical instructions in the aforementioned memory 730 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, essentially, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0133] On the other hand, the present invention also provides a computer program product, which includes a computer program that can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer can execute the point of interest recommendation method provided by the above methods. The method includes: acquiring user access information and a set of points of interest, wherein the user access information includes the user's target points of interest, and the set of points of interest is associated with the target points of interest; expanding the short sequence in the user access information to obtain an enhanced user access information sequence; and inputting the enhanced user access information sequence and the set of points of interest into a preset point of interest recommendation model to obtain the user's target point of interest recommendation result, wherein the point of interest recommendation model is based on an attention mechanism.

[0134] In another aspect, the present invention also provides a non-transitory computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements an interest point recommendation method provided by the above methods. The method includes: acquiring user access information and an interest point set, wherein the user access information includes the user's target interest point, and the interest point set is associated with the target interest point; expanding a short sequence in the user access information to obtain an enhanced user access information sequence; and inputting the enhanced user access information sequence and the interest point set into a preset interest point recommendation model to obtain the user's target interest point recommendation result, wherein the interest point recommendation model is based on an attention mechanism.

[0135] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0136] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0137] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. An interest point recommendation method, characterized in that, include: Obtain user access information and a set of points of interest, wherein the user access information includes the target points of interest accessed by the user, and the set of points of interest is associated with the target points of interest; The short sequence in the user access information is expanded to obtain an enhanced user access information sequence; The enhanced user access information sequence and the set of interest points are input into a preset interest point recommendation model to obtain the user's target interest point recommendation result. The interest point recommendation model is obtained based on an attention mechanism. The point of interest recommendation model includes an embedding layer, an embedding propagation layer, and a prediction layer; The step of inputting the enhanced user access information sequence and the set of interest points into a preset interest point recommendation model to obtain the target interest point recommendation result includes: Based on the embedding layer, the user access information sequence is encoded to obtain a first vector representation, and each interest point in the interest point set is encoded to obtain a second vector representation; Based on the embedded propagation layer and the first vector representation, a personalized profile of the user is obtained. Based on the personalized profile and the first vector representation, a user vector representation is obtained. The personalized profile is obtained based on the user's access to interest points, interest point types, and user access time corresponding to the first vector representation. Based on the embedding propagation layer, the interest points corresponding to the second vector representation are aggregated to obtain the aggregation result, and the attention score of the aggregation result is calculated to obtain the neighbor representation of the second vector representation. Based on the second vector representation and the neighbor representation, the interest point vector representation is obtained. Based on the prediction layer, the matching degree between the user vector representation and the interest point vector representation is scored to obtain the target interest point recommendation result; The user access information sequence includes the user's historical trajectory; the first vector is a low-dimensional vector representation of the entities and relationships involved in the user's historical trajectory, and the second vector is a low-dimensional vector representation of the entities and relationships involved in the tourism event graph.

2. The point-of-interest recommendation method according to claim 1, characterized in that, The acquisition of user access information and point of interest set includes: Based on user access records, the user access information is obtained; based on the tourism event graph, the set of points of interest is obtained; the tourism event graph is obtained by data cleaning, event extraction, and event time sequence information extraction of tourism text data; and the user access records are obtained based on call detail records (CDRs).

3. The point-of-interest recommendation method according to claim 1, characterized in that, The step of augmenting the user information with a short sequence to obtain an enhanced user information sequence includes: The user access information is input into a preset pre-trained model to obtain an enhanced user access information sequence. The pre-trained model is obtained by training a Transformer network based on the reverse sequence of the sample user access information.

4. The point-of-interest recommendation method according to claim 1, characterized in that, The target interest point recommendation result includes multiple interest points, the user access information includes user identity information, and after obtaining the target interest point recommendation result, the method further includes: Obtain real-time pedestrian flow information, residential area distribution information, and crowd movement trajectory information for each point of interest; Based on the user identity information and cluster centroids, a user's geographic location preference score is determined. The cluster centroids are obtained by clustering user type information based on user geographic location preference features and user side attribute features. The user geographic location preference features are obtained based on user access information, and the user side attribute features are obtained based on the user's side attribute information. Based on the real-time pedestrian flow information of each point of interest, the population density score of each point of interest is obtained; based on the residential area distribution information of each point of interest, the residential area overlap score of each point of interest is obtained; based on the pedestrian movement trajectory information of each point of interest, the population density prediction score of each point of interest is obtained. Based on the user's geographic location preference score, the population density score of each point of interest, the residential area overlap score of each point of interest, and the population density prediction score of each point of interest, a comprehensive score is obtained for each point of interest corresponding to the user. Based on the comprehensive score of each interest point, the optimized target interest point recommendation result is obtained.

5. An interest point recommendation device corresponding to the interest point recommendation method as described in claim 1, characterized in that, include: The information acquisition module is used to acquire user access information and a set of points of interest, wherein the user access information includes the target points of interest accessed by the user, and the set of points of interest is associated with the target points of interest. A sequence augmentation module is used to augment the short sequences in the user access information to obtain an enhanced user access information sequence; The point of interest recommendation module inputs the enhanced user access information sequence and the point of interest set into a preset point of interest recommendation model to obtain the user's target point of interest recommendation result. The point of interest recommendation model is based on an attention mechanism.

6. The point-of-interest recommendation device according to claim 5, characterized in that, The target interest point recommendation result includes multiple interest points, the user access information includes user identity information, and the device further includes: The recommendation result optimization module is used to obtain real-time pedestrian flow information, residential area distribution information and crowd movement trajectory information for each interest point after obtaining the recommendation results of the target interest point; Based on the user identity information and cluster centroids, a user's geographic location preference score is determined. The cluster centroids are obtained by clustering user type information based on user geographic location preference features and user side attribute features. The user geographic location preference features are obtained based on user access information, and the user side attribute features are obtained based on the user's side attribute information. Based on the real-time pedestrian flow information of each point of interest, the population density score of each point of interest is obtained; based on the residential area distribution information of each point of interest, the residential area overlap score of each point of interest is obtained; based on the pedestrian movement trajectory information of each point of interest, the population density prediction score of each point of interest is obtained. Based on the user's geographic location preference score, the population density score of each point of interest, the residential area overlap score of each point of interest, and the population density prediction score of each point of interest, a comprehensive score is obtained for each point of interest corresponding to the user. Based on the comprehensive score of each interest point, the optimized target interest point recommendation result is obtained.

7. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the point of interest recommendation method as described in any one of claims 1 to 4.

8. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the point of interest recommendation method as described in any one of claims 1 to 4.

9. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the point of interest recommendation method as described in any one of claims 1 to 4.

Images