Point of interest representation and push method, device, electronic device and storage medium

By constructing user historical trajectory sequences and training position point embedding matrix, the problem of sparse interest point check-in data is solved, and the accuracy of interest point push is improved and geosemantic information enhancement is enhanced.

CN111782955BActive Publication Date: 2024-11-12ALIPAY (HANGZHOU) INFORMATION TECH CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202010624873.5
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2020-07-01
Publication Date
2024-11-12
Estimated Expiration
2040-07-01

AI Technical Summary

Technical Problem

The prior art is difficult to effectively deal with the sparse problem of point-of-interest check-in data, especially when there are too many points of interest, which leads to a decrease in the accuracy of point-of-interest recommendations.

Method used

By obtaining user's interest point check-in records and GPS acquisition point records, a historical trajectory sequence is constructed, and the position point embedding matrix is ​​trained using word representation model, and GPS acquisition point information is incorporated to give stronger geographical semantic information to the interest point, thereby realizing interest points vector representation and push.

Benefits of technology

Mapping points of interest from high-dimensional sparse space to low-dimensional dense space, improving the accuracy of point-of-interest push, enhancing the semantic information of points of interest in geographical location, and solving the problem of sparseness of point-of-interest check-in data.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN111782955B_ABST
    Figure CN111782955B_ABST
Patent Text Reader

Abstract

The embodiments of this specification disclose a method, device, electronic device and storage medium for representing and pushing points of interest, which construct a user's historical trajectory sequence based on the point of interest check-in records and GPS point sampling records of a user set; train a word representation model based on the user's historical trajectory sequence, and the obtained location point embedding matrix contains a location semantic vector representation of each point of interest in the point of interest check-in record, and a location semantic vector representation of each location point in the GPS point sampling record.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The embodiments of this specification relate to the field of data processing, and more particularly to a method, device, electronic device, and storage medium for representing and pushing points of interest. Background Art

[0002] POI (Point of interest) is a landmark or scenic spot on an electronic map, which is used to mark the commercial organizations, tourist attractions, historical sites, transportation facilities and other places represented by the place. There are millions of POIs in the real world, which makes the user's POI check-in data very sparse. With the rapid development of mobile Internet, LBSN (Location-based Social Network) is becoming more and more important in life. Through the check-in behavior and comments on the designated location, the experience and experience are shared with friends or netizens. The recommendation of POI has become the most important application in location services, guiding users to explore potential places of interest. It has a very large scene in tourism, catering and life services. Summary of the invention

[0003] The embodiments of this specification provide a method, device, electronic device and storage medium for representing and pushing points of interest to better push points of interest to users.

[0004] In a first aspect, an embodiment of the present specification provides a method for representing points of interest, including: obtaining check-in records of points of interest and GPS point sampling records of a user set, wherein several location points in the GPS point sampling records overlap with the points of interest in the check-in records of the points of interest; constructing a historical trajectory sequence of each user in the user set based on the check-in records of the points of interest and the GPS point sampling records; using the historical trajectory sequence of each user in the user set to train a word representation model to obtain a location point embedding matrix, wherein the location point embedding matrix contains a location semantic vector representation of each point of interest in the check-in records of the points of interest, and a location semantic vector representation of each location point in the GPS point sampling records.

[0005] In a second aspect, an embodiment of the present specification provides a method for pushing points of interest, including: determining a target user, wherein the target user is any user among all users who has a need for pushing points of interest;

[0006] The interest points are pushed to the target user according to a position point embedding matrix, wherein the position point embedding matrix is ​​obtained by the interest point representation method described in the first aspect.

[0007] In a third aspect, an embodiment of the present specification provides a device for representing points of interest, including: a record acquisition unit, used to acquire check-in records of points of interest and GPS point sampling records of a user set, wherein several location points in the GPS point sampling records coincide with the points of interest in the check-in records of the points of interest; a trajectory construction unit, used to construct a historical trajectory sequence of each user in the user set based on the check-in records of the points of interest and the GPS point sampling records; a model training unit, used to train a word representation model using the historical trajectory sequence of each user in the user set to obtain a location point embedding matrix, wherein the location point embedding matrix contains a location semantic vector representation of each point of interest in the check-in records of the points of interest, and a location semantic vector representation of each location point in the GPS point sampling records.

[0008] In a fourth aspect, an embodiment of the present specification provides an interest point pushing device, comprising: a determination unit, used to determine a target user, wherein the target user is any user among all users who has an interest point pushing demand; an interest point pushing module, used to push interest points to the target user according to a position point embedding matrix, wherein the position point embedding matrix is ​​obtained by the interest point representation method described in the first aspect.

[0009] In a fifth aspect, an embodiment of this specification provides an electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the steps of the method described in the first aspect or the second aspect when executing the program.

[0010] In a sixth aspect, an embodiment of the present specification provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the method described in the first aspect or the second aspect.

[0011] One or more technical solutions provided by the embodiments of this specification achieve at least the following technical effects or advantages:

[0012] The embodiments of the present specification construct a user's historical trajectory sequence based on the user's set of interest point check-in records and GPS point sampling records; a word representation model is trained based on the user's historical trajectory sequence, and the obtained location point embedding matrix contains the location semantic vector representation of each location point in the GPS point sampling record, thereby realizing the mapping of interest points from high-dimensional sparse space to low-dimensional dense space to obtain the vector representation of interest points, solving the problem of sparse interest point check-in data caused by too many interest points, and the GPS point sampling information integrated into the vector representation of the interest point gives the interest point stronger semantic information in terms of the geographical location. Therefore, the use of the location point embedding matrix for interest point push is more accurate. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] In order to more clearly illustrate the embodiments of this specification or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of this specification. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying creative work.

[0014] Figure 1 A schematic diagram of a scenario of a method for pushing points of interest in an embodiment of this specification;

[0015] Figure 2 This is a flowchart of a method for representing points of interest in an embodiment of this specification;

[0016] Figure 3 Schematic diagram of a training word representation model in an embodiment of this specification;

[0017] Figure 4 This is a flow chart of the method for pushing points of interest in an embodiment of this specification;

[0018] Figure 5 This is a functional module diagram of the interest point representation device in the embodiment of this specification;

[0019] Figure 6 This is a functional module diagram of the device for pushing points of interest in the embodiments of this specification;

[0020] Figure 7 It is a schematic diagram of the structure of an electronic device in an embodiment of this specification. DETAILED DESCRIPTION

[0021] In order to make the purpose, technical solutions and advantages of the embodiments of this specification more clear, the technical solutions in the embodiments of this specification will be clearly and completely described below in conjunction with the drawings in the embodiments of this specification. Obviously, the described embodiments are part of the embodiments of this specification, not all of the embodiments. Based on the embodiments in this specification, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this specification.

[0022] The present specification provides a method for pushing points of interest, which is applied to Figure 1 In the scenario shown. Figure 1As shown, the scenario includes: a server 20 and a client 10 for which the server 20 provides business services. The client 10 can be deployed on a mobile device, and the client 10 can be any target application based on LBSN. The client 10 reports the check-in records of points of interest and GPS point collection records to the server 20. The server 20 collects the check-in records of points of interest and GPS point collection records reported by each client 10, thereby obtaining the check-in records of points of interest and GPS point collection records of the user set. Several location points in the GPS point collection records coincide with the points of interest in the check-in records of points of interest; according to the check-in records of points of interest and GPS point collection records of the user set, a historical trajectory sequence of each user in the user set is constructed, and a word representation model is trained using the historical trajectory sequence of each user in the user set to obtain a location point embedding matrix, wherein the location point embedding matrix contains the location semantic vector representation of each point of interest in the check-in records of points of interest, and the location semantic vector representation of each point of interest in the GPS point collection records. The server 20 pushes the points of interest to the target user according to the location point embedding matrix. Alternatively, the server 20 sends the obtained position point embedding matrix to other servers, and the other server devices push the points of interest to the target user according to the position point embedding matrix.

[0023] By mapping from high-dimensional sparse space to low-dimensional dense space to obtain the vector representation of points of interest, the problem of sparse sign-in data of points of interest caused by too many points of interest is solved. The GPS sampling information integrated into the vector representation of points of interest gives the points of interest stronger semantic information in terms of geographical location, making up for the defect that the points of interest themselves lack sufficient text description. Therefore, it is more accurate to push points of interest using the location point embedding matrix.

[0024] In the first aspect, the embodiment of this specification provides a method for representing points of interest, which is applied to a server, referring to Figure 2 As shown, the interest point representation method includes the following steps:

[0025] S200: Obtaining the interest point check-in records and GPS point sampling records of the user set, wherein a number of location points in the GPS point sampling records overlap with the interest points in the interest point check-in records.

[0026] If the purpose is to push only local POIs in a certain geographical area, the POI sign-in records and GPS point-collecting records of users are collected for the geographical area, and the user set only includes users who sign in to POIs based on LBSN in the geographical area. If the purpose is to push POIs in different places, the POI sign-in records and GPS point-collecting records of users are collected for different geographical areas. The user set can include users who sign in to POIs based on LBSN under the target application, and even users who sign in to POIs based on LBSN under multiple target applications.

[0027] Specifically, the interest point check-in record of the user set includes: the interest point check-in record of each user in the user set, wherein each user's interest point check-in record includes more than one interest point check-in data. Specifically, any user checks in to the current interest point based on LBSN, and a corresponding interest point check-in data will be generated. Each interest point check-in data generated can be represented by a triple (u, v, t), which means that user u visited interest point v at time t.

[0028] In order to give points of interest stronger semantic information in terms of geographical location, the GPS point collection records of each user in the user set are integrated into the embodiments of this specification. Specifically, the GPS point collection records of each user are collected by the mobile device where the target application is located. The GPS point collection records of each user are a series of location points with geographical identification collected by the mobile device through GPS. Each location point can be represented by a triple (u, l, t), indicating that user u is at location point l at time t. The location points collected by GPS can be points of interest or ordinary location points.

[0029] For example, if U = {u1,u2,…,u n} represents the user set, then L={l1,l2,…,l m} represents the set of location points in the GPS point collection records corresponding to the user set U, V = {v1,v2,…,v p} represents the set of interest points in the check-in records of the interest points corresponding to the user set U. The location point set L contains the interest point set V, and m≥p.

[0030] S202: construct a historical trajectory sequence of each user in the user set based on the check-in records of the points of interest and the GPS point sampling records of the user set.

[0031] Specifically, for each user in the user set, the user's historical trajectory sequence is constructed based on the user's interest point check-in record and the user's GPS point collection record. Specifically, the user's location points within a certain historical time range T can be sorted according to the time sequence, or the user's location points of a certain number of wanderings M can be sorted according to the time sequence to construct the user's historical trajectory sequence, for example: represents the historical trajectory sequence of user u, where the historical trajectory sequence of user u includes several points of interest and other location points that user u has visited in the past. Thus, the historical trajectory sequence of each user in the user set can be obtained. Figure 3 In the historical trajectory sequence, black blocks represent locations of interest points, and white blocks represent other locations.

[0032] S204. Use the historical trajectory sequence of each user in the user set to train a word representation model to obtain a location point embedding matrix, wherein the location point embedding matrix includes a location semantic vector representation of each point of interest in the point of interest check-in record and a location semantic vector representation of each point of interest in the GPS point sampling record.

[0033] Thus, the position semantic vector representation of each point of interest in the point of interest check-in record and the position semantic vector representation of other positions (position points other than the point of interest) are obtained.

[0034] Specifically, step S204 includes the following detailed steps: generating training sample data according to the historical trajectory sequence of each user in the user set, wherein the training sample data includes training examples for each point of interest in the point of interest check-in record; training a word representation model according to the training sample data to obtain a trained word representation model; and obtaining a position point embedding matrix according to the hidden layer weight vector of the trained word representation model.

[0035] In the embodiments of this specification, the word representation model can be but is not limited to using the Skip-Gram model. It should be noted that the Skip-Gram model is a three-layer neural network including an input layer, a hidden layer, and an output layer. The hidden layer does not use an activation function, but the output layer uses a sotfmax function. In the embodiments of this specification, the historical trajectory sequence of each user in the user set is used to generate training examples for each point of interest to train the Skip-Gram model. The weight matrix of the hidden layer in the trained Skip-Gram model is the position point embedding matrix. Each row of the hidden layer weight vector in the position point embedding matrix corresponds to the position semantic vector representation of one of the position points.

[0036] More specifically, the implementation process of generating training sample data based on the historical trajectory sequence of each user in the user set is as follows:

[0037] Select an i-th interest point from the historical trajectory sequence of an n-th user in a user set, where n is 1 to N, i is 1 to Z, N is the number of users in the user set, and Z is the number of interest points in the historical trajectory sequence of the n-th user; determine the context position point of the i-th interest point from the historical trajectory sequence of the n-th user according to a pre-set context window, and construct a positive sample based on the context position point of the i-th interest point and the i-th interest point; select C interest points different from the i-th interest point, and construct corresponding negative samples based on the context position point of the i-th interest point and the C interest points, and construct a positive sample based on the context position point of the i-th interest point and the i-th interest point; and form a group of training samples for the i-th interest point based on the positive samples and the negative samples.

[0038] In the specific implementation process, the form of the training sample for the i-th interest point can be expressed as follows:

[0039] (l i ,context(l i ), train a three-layer neural network through the training examples of each i-th interest point, predict the context position point of the i-th interest point, and maximize the output probability of the context position point.

[0040] The pre-set context window is d and the number of neurons in the hidden layer is k. k determines the dimension of the vector used to represent the interest point. Taking the i-th interest point selected as interest point l0 as an example, the process of constructing a set of training examples for interest point l0 is described as follows:

[0041] Select the 2d positions before and after the interest point l0 as the context position points of the interest point l0, expressed as: context(l0), and construct a positive sample based on the interest point l0 and the context position point context(l0);

[0042] Since there are usually too many location point data, negative sampling methods are needed to reduce the training pressure. Therefore, C points of interest that are different from the point of interest l0 are randomly selected, and the C points of interest are respectively combined with the 2d location points before and after to form negative samples for the point of interest l0. Since different users have the same points of interest in their historical trajectory sequences, for example, user A and user B have visited and checked in at the same point of interest, or user A has visited the same point of interest one after another. Therefore, for each point of interest in the point of interest check-in record, one or more groups of training samples can be obtained. Predict the 2d location points before and after the point of interest l0, so that the output probability of the 2d locations before and after is maximized. Based on this, the output probability P(context(l0),l0)=1, the output probability P(context(l0),l i )=0,i=1,2,…,c, that is, the expectation is to maximize the following probability:

[0043]

[0044] In the above formula, σ represents the sigmoid function, context(l0) represents the 2d position points before and after the interest point l0, is the linear objective function of the neural network, w represents the weight, and θ is the parameter of the linear objective function.

[0045] The Skip-Gram model is trained based on one or more training examples for each point of interest. The Skip-Gram model is trained to obtain the hidden layer weight vector of the Skip-Gram model, and the hidden layer weight vector of the Skip-Gram model after training is used as the location point embedding matrix (Location Embedding), which is expressed as follows:

[0046]

[0047] Where W = (w1, w2, ..., w m ) T is the location point embedding matrix, which contains the hidden layer weight vector of each location point in the GPS point collection record, refer to Figure 3 , w i =(w i1 ,w i2 ,…,w ik ) T is the hidden layer weight vector of position point i, that is, the position semantic vector representation of position point i, m is the number of position points, k represents the vector dimension of the position points, and corresponds to the number of neurons k in the hidden layer.

[0048] For example, if the obtained location point embedding matrix is ​​10,000 rows * 300 columns, and each row corresponds to a location semantic vector representation of a location point, then the 10,000 rows * 300 columns location point embedding matrix contains location semantic vector representations for 10,000 location points (including several points of interest), and the location semantic vector representation for each location point is a 300-dimensional vector.

[0049] In a specific application scenario, corresponding to the POI check-in record and GPS point sampling record obtained in step S200: the obtained location point embedding matrix contains the location semantic vector representation of each location point in the target geographical area, for example, the target geographical area may be a city, a province, or an urban area; if in addition to pushing local POIs to users, it is also necessary to push remote POIs to users, then the location point embedding matrix contains the location semantic vector representation of each location point in multiple different geographical areas, for example, it contains the location semantic vector representation of each location point in the national geographical area.

[0050] Therefore, the location embedding matrix can solve the sparse POI matrix problem caused by too many POIs and sparse check-in data of POIs on the one hand, and give the POIs semantic information in terms of their geographical location on the other hand.

[0051] In a second aspect, the embodiments of this specification provide a method for pushing points of interest, which is applied to a server and pushes points of interest to a user according to the position point embedding matrix obtained in the first aspect. Figure 4 As shown, the method for pushing points of interest provided in the embodiment of this specification includes the following steps:

[0052] S400: Determine a target user, wherein the target user is any user among all users who has an interest in push requirements.

[0053] Specifically, the target user may be any user within or outside the user set who has a need to push points of interest. For example, if the point of interest push method is applied to a target application, all users may be every registered user of the target application.

[0054] S402: Pushing points of interest to the target user according to the position point embedding matrix, wherein the position point embedding matrix is ​​obtained by the method for representing points of interest according to the first aspect.

[0055] Specifically, step S402 includes the following specific steps:

[0056] S4021. Calculate push basis information for the target user according to the location point embedding matrix, where the push basis information includes any one of the following for the target user:

[0057] ① Interest point similarity information. Specifically, the interest point similarity information includes the similarity between each candidate interest point and the interest point where the target user is currently located. Specifically, the interest point similarity between each candidate interest point and the interest point where the target user is currently located can be obtained through the following steps A1 to A4:

[0058] Step A1: Detect the current point of interest visited by the target user; specifically, the current point of interest visited by the target user may be the point of interest that the target user currently signs in based on the LBSN.

[0059] Step A2: Recall multiple candidate interest points based on the current interest point. Specifically, each interest point in the position point embedding matrix except the current interest point can be used as a candidate interest point, or some interest points can be recalled as candidate interest points according to a preset rule to reduce the amount of calculation, such as random recall or recall of interest points of the same type.

[0060] Step A3: Obtain the location semantic vector representation of the current point of interest and the location semantic vector representation of each candidate point of interest from the location point embedding matrix. According to the sequential identification of the current point of interest, find the corresponding row hidden layer weight vector from the location point embedding matrix as the location semantic vector representation of the current point of interest. For example, if the current point of interest is the 100th point of interest in the set of points of interest, then the location semantic vector of the 100th point of interest obtained from the location point embedding matrix is ​​the location semantic vector representation of the current point of interest. The method of obtaining the location semantic vector representation of the candidate point of interest is similar and will not be repeated here for the sake of brevity of the specification.

[0061] Step A4: For each candidate interest point, the interest point similarity between the current interest point and the candidate interest point is obtained based on the location semantic vector representation of the current interest point and the location semantic vector representation of the candidate interest point. The interest point similarity may refer to: the cosine distance between the location semantic vector representation of the candidate interest point and the location semantic vector representation of the current interest point.

[0062] ② User similarity information. The user similarity information includes the user similarity between each candidate user and the target user. Specifically, the user similarity information can be obtained through the following steps B1 to B3:

[0063] Step B1: Obtain the location semantic vector representation of each location point in the historical trajectory sequence of the target user from the location point embedding matrix, and generate the trajectory vector representation of the target user according to the location semantic vector representation of each location point in the historical trajectory sequence of the target user.

[0064] Specifically, the trajectory vector representation of the target user is obtained by averaging the position semantic vector representation of each position point in the historical trajectory sequence of the target user.

[0065] Step B2: determining multiple candidate users for the target user from the user set;

[0066] Specifically, each user in the user set can be used as a candidate user. In order to reduce the amount of calculation, multiple users associated with the target user are selected from the user set as candidate users. For example, the friends of the target user in the target application are selected as candidate users, and another example is: multiple users who belong to the same age group as the target user are selected as candidate users. However, the selection method for candidate users is not limited to the above examples.

[0067] Step B3, the user similarity between each candidate user and the target user is obtained in the following manner: obtaining the position semantic vector representation of each position point in the historical trajectory sequence of the candidate user from the position point embedding matrix; and generating the trajectory vector representation of the candidate user according to the position semantic vector representation of each position point in the historical trajectory sequence of the candidate user; and obtaining the user similarity between the candidate user and the target user according to the trajectory vector representation of the candidate user and the trajectory vector representation of the target user.

[0068] Specifically, the trajectory vector representation of each candidate user is obtained by averaging the position semantic vector representation of each position point in the historical trajectory sequence of the candidate user.

[0069] By calculating the cosine distance between the trajectory vector representation of the candidate user and the trajectory vector representation of the target user, the historical trajectory similarity between the candidate user and the target user is obtained. The calculated historical trajectory similarity is the user similarity between the candidate user and the target user:

[0070]

[0071] cosθ ij represents the historical trajectory similarity between the candidate user and the target user. The trajectory vector representation of the historical trajectory candidate user and the trajectory vector representation of the target user are avg(l i )、avg(l j ).

[0072] ③ Potential behavior trajectory. Specifically, the potential behavior trajectory is the result of predicting the target user's future behavior trajectory based on the target user's historical trajectory sequence. Specifically, the target user's potential behavior trajectory can be predicted through the following steps C1 to C3:

[0073] Step C1: Obtain the location semantic vector representation of each location point in the historical trajectory sequence of the target user from the location point embedding matrix.

[0074] Specifically, it can be triggered to obtain the location semantic vector representation of each location point in the historical trajectory sequence of the target user within a certain historical time range according to a preset period, or when it is detected that the target user is currently at a certain point of interest

[0075] Step C2: Obtain a trajectory vector matrix of the target user according to the position semantic vector representation of each position point in the historical trajectory sequence of the target user.

[0076] Step C3: predicting the target user's potential behavior trajectory according to the trajectory vector matrix, where the potential behavior trajectory contains more than one point of interest.

[0077] For example, a trajectory prediction model is obtained by training a CNN (Convolutional Neural Networks) model or an RNN (Recurrent Neural Network) model. Specifically, the trajectory vector matrix of the target user is used as the input of the trajectory prediction model, and the trajectory vector matrix of the target user is processed by the trajectory prediction model to obtain the potential behavior trajectory of the target user.

[0078] S4022: Pushing points of interest to target users based on the push basis information.

[0079] Specifically, the points of interest pushed to the target user include one or more. Depending on the push basis information, the points of interest pushed to the target user will also be different, which are described below:

[0080] ① If the push basis information is the interest point similarity between each candidate interest point and the current interest point, correspondingly, step S4022 specifically includes: selecting interest points similar to the current interest point from multiple candidate interest points according to the interest point similarity between each candidate interest point and the current interest point; and pushing the selected similar interest points to the target user.

[0081] Specifically, one or more interest points are selected as similar interest points according to the similarity between each candidate interest point and the current interest point from high to low. For example, if the current interest point of the target user is shopping mall A, the similar interest point pushed to the target user may be shopping mall B. In this way, other interest points similar to the current interest point can be selected and pushed to the target user.

[0082] When the push basis information is implementation mode ①, only the semantic vector representation of the location point of interest in the location point embedding matrix needs to be used, and the semantic vector representation of other location points can be ignored.

[0083] Through the vector correlation of different points of interest, we can obtain more similar points of interest, get rid of geographical restrictions, and realize not only the push of local points of interest, but also the push of remote points of interest.

[0084] ② If the push basis information is the user similarity between each candidate user and the target user, correspondingly, step S4022 specifically includes: selecting a reference user from multiple candidate users according to the user similarity between each candidate user and the target user, wherein one or more candidate users can be selected as reference users according to the user similarity from high to low; and pushing the current point of interest visited by the reference user to the target user.

[0085] By pushing the current points of interest visited by the reference user to the target user, points of interest that the target user may be interested in can be mined, and remote recommendation of points of interest can be achieved.

[0086] ③ If the push is based on the potential behavior trajectory of the target user of the information, the potential behavior trajectory contains one or more points of interest. Therefore, step S4022 specifically includes: determining the potential points of interest for the target user from the potential behavior trajectory; and pushing the potential points of interest to the target user. Combined with the deep learning algorithm, the potential points of interest are predicted, so as to push points of interest that are consistent with the user's historical behavior habits and interest preferences.

[0087] In a third aspect, based on the same inventive concept as the aforementioned method for representing points of interest, an embodiment of the present specification provides a device for representing points of interest, referring to Figure 5 As shown, including:

[0088] A record acquisition unit 501 is used to acquire the interest point check-in records and GPS point sampling records of the user set, where some location points in the GPS point sampling records overlap with the interest points in the interest point check-in records;

[0089] The trajectory construction unit 502 is used to construct a historical trajectory sequence of each user in the user set according to the check-in records of the points of interest and the GPS point collection records;

[0090] The model training unit 503 is used to train the word representation model using the historical trajectory sequence of each user in the user set to obtain a location point embedding matrix, wherein the location point embedding matrix includes the location semantic vector representation of each point of interest in the point of interest check-in record and the location semantic vector representation of each point of interest in the GPS point sampling record.

[0091] In an optional implementation, the model training unit 503 includes:

[0092] A training sample generation subunit, used to generate training sample data according to the historical trajectory sequence of each user in the user set, wherein the training sample data includes a training sample for each point of interest in the point of interest check-in record;

[0093] A training subunit, used for training the word representation model according to the training sample data to obtain a trained word representation model;

[0094] The weight acquisition subunit is used to obtain the position point embedding matrix according to the hidden layer weight vector of the trained word representation model.

[0095] In an optional implementation manner, the training sample generation subunit is specifically used to:

[0096] Select the i-th point of interest from the historical trajectory sequence of the n-th user in the user set, where n is 1 to N, i is 1 to Z, N is the number of users in the user set, and Z is the number of points of interest in the historical trajectory sequence of the n-th user;

[0097] According to a pre-set context window, determine the context location point of the i-th interest point from the historical trajectory sequence of the n-th user;

[0098] Select C interest points different from the i-th interest point, construct corresponding negative samples based on the context position point of the i-th interest point and the C interest points, and construct positive samples based on the context position point of the i-th interest point and the i-th interest point;

[0099] A set of training examples for the i-th interest point is formed based on the positive examples and the negative examples.

[0100] In a fourth aspect, the present specification provides a device for pushing points of interest, referring to Figure 6 As shown, including:

[0101] A determination unit 601 is used to determine a target user, wherein the target user is any user among all users who has an interest in point push requirements;

[0102] The interest point push module 602 is used to push interest points to the target user according to the position point embedding matrix. The position point embedding matrix is ​​obtained in the aforementioned interest point representation method embodiment, and will not be repeated here for the sake of brevity.

[0103] In an optional implementation manner, the point of interest push module 602 includes:

[0104] The information calculation unit 6021 is used to obtain push basis information for the target user according to the location point embedding matrix, wherein the push basis information includes the following information for the target user: interest point similarity information, user similarity information, and potential behavior trajectory;

[0105] The push unit 6022 is used to push the points of interest to the target user according to the push basis information.

[0106] In an optional implementation manner, the information computing unit 6021 includes:

[0107] A detection subunit, used to detect a current point of interest visited by a target user;

[0108] A recall subunit, used to recall multiple candidate interest points according to the current interest point;

[0109] A first vector acquisition subunit is used to acquire a position semantic vector representation of the current interest point and a position semantic vector representation of each candidate interest point among multiple candidate interest points from the position point embedding matrix;

[0110] The calculation subunit is used to obtain the interest point similarity between the current interest point and the candidate interest point for each candidate interest point among the multiple candidate interest points according to the position semantic vector representation of the current interest point and the position semantic vector representation of the candidate interest point.

[0111] In an optional implementation manner, the push unit 6022 includes:

[0112] An interest point selection subunit, for selecting an interest point similar to the current interest point from a plurality of candidate interest points according to the interest point similarity between each candidate interest point and the current interest point;

[0113] The first push subunit is used to push similar points of interest to target users.

[0114] In an optional implementation manner, the information computing unit 6021 includes:

[0115] A trajectory representation subunit is used to obtain the location semantic vector representation of each location point in the historical trajectory sequence of the target user from the location point embedding matrix, and generate the trajectory vector representation of the target user according to the location semantic vector representation of each location point in the historical trajectory sequence of the target user;

[0116] A user determination subunit, used to determine multiple candidate users for a target user from a user set;

[0117] The similarity calculation subunit is used to obtain, for each candidate user, a location semantic vector representation of each location point in the candidate user's historical trajectory sequence from the location point embedding matrix; and generate a trajectory vector representation of the candidate user based on the location semantic vector representation of each location point in the candidate user's historical trajectory sequence; and obtain the user similarity between the candidate user and the target user based on the trajectory vector representation of the candidate user and the trajectory vector representation of the target user.

[0118] In an optional implementation, the push unit 6022 includes:

[0119] A user selection subunit, used to select a reference user from multiple candidate users according to the user similarity between each candidate user and the target user;

[0120] The second push subunit is used to push the current point of interest visited by the reference user to the target user.

[0121] In an optional implementation manner, the information computing unit 6021 includes:

[0122] The second vector acquisition subunit is used to acquire the location semantic vector representation of each location point in the historical trajectory sequence of the target user from the location point embedding matrix;

[0123] A trajectory obtaining subunit is used to obtain a trajectory vector matrix for the target user according to the position semantic vector representation of each position point in the historical trajectory sequence of the target user;

[0124] The trajectory prediction subunit is used to predict the potential behavior trajectory of the target user according to the trajectory vector matrix, and the potential behavior trajectory contains more than one point of interest.

[0125] In an optional implementation manner, the push unit 6022 includes:

[0126] An interest point determination subunit, used to determine potential interest points for a target user from a potential behavior trajectory;

[0127] The third push subunit is used to push potential points of interest to target users.

[0128] In a fifth aspect, based on the same inventive concept as the aforementioned method for representing points of interest and method for pushing points of interest, the embodiments of this specification further provide an electronic device, such as Figure 7 As shown, it includes a memory 704, a processor 702, and a computer program stored in the memory 704 and executable on the processor 702. When the processor 702 executes the program, the steps of the aforementioned interest point representation method embodiment are implemented, or the steps of the aforementioned interest point push method embodiment are implemented.

[0129] Among them, Figure 7 In the embodiment of the present invention, a bus architecture (represented by bus 700) is shown, which may include any number of interconnected buses and bridges, and bus 700 links various circuits including one or more processors represented by processor 702 and memory represented by memory 704. Bus 700 may also link various other circuits such as peripherals, voltage regulators, and power management circuits, which are well known in the art and are therefore not further described herein. Bus interface 706 provides an interface between bus 700 and receiver 701 and transmitter 703. Receiver 701 and transmitter 703 may be the same element, namely a transceiver, providing a unit for communicating with various other devices over a transmission medium. Processor 702 is responsible for managing bus 700 and general processing, while memory 704 may be used to store data used by processor 702 when performing operations.

[0130] In the sixth aspect, based on the same inventive concept as the aforementioned embodiments of the method for representing points of interest and the method for pushing points of interest, the embodiments of this specification also provide a computer-readable storage medium on which a computer program is stored, which, when executed by a processor, implements the steps of the aforementioned embodiment of the method for representing points of interest, or implements the steps of the aforementioned embodiment of the method for pushing points of interest.

[0131] This specification is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of this specification. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that has the functions specified in one or more boxes.

[0132] These computer program instructions may also be stored in a computer-readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture including an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0133] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0134] Those skilled in the art will readily appreciate other embodiments of the invention after considering the specification and practicing the invention disclosed herein. The present invention is intended to cover any variations, uses or adaptations of the present invention that follow the general principles of the present invention and include common knowledge or customary techniques in the art that are not disclosed in this disclosure. The description and examples are to be considered exemplary only, and the true scope and spirit of the present invention are indicated by the following claims.

[0135] It should be understood that the present invention is not limited to the precise structure described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from the scope thereof. The scope of the present invention is limited only by the appended claims. The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modification, equivalent substitution, improvement, etc. made within the spirit and principle of the present invention shall be included in the scope of protection of the present invention.

Claims

1. A method for pushing points of interest, comprising: Determine a target user, wherein the target user is any user among all users who is interested in pushing a demand; Pushing points of interest to the target user according to the position point embedding matrix includes: obtaining push basis information for the target user according to the position point embedding matrix, wherein, in the case where the push basis information includes a potential behavior trajectory, obtaining a position semantic vector representation of each position point in the historical trajectory sequence of the target user from the position point embedding matrix; obtaining a trajectory vector matrix for the target user according to the position semantic vector representation of each position point in the historical trajectory sequence of the target user; predicting a potential behavior trajectory of the target user according to the trajectory vector matrix, wherein the potential behavior trajectory includes more than one point of interest, and pushing the point of interest to the target user according to the push basis information; The location point embedding matrix is ​​obtained by the following steps: obtaining the interest point check-in records and GPS point collection records of the user set, wherein several location points in the GPS point collection records overlap with the interest points in the interest point check-in records, and the GPS point collection records are location points with geographical identification collected by a mobile device through GPS; constructing a historical trajectory sequence of each user in the user set according to the interest point check-in records and the GPS point collection records; using the historical trajectory sequence of each user in the user set to train a word representation model to obtain a location point embedding matrix, wherein the location point embedding matrix contains a location semantic vector representation of each interest point in the interest point check-in record, and a location semantic vector representation of each location point in the GPS point collection record.

2. According to the method as claimed in claim 1, the push basis information can also be the following information for the target user: interest point similarity information, user similarity information.

3. The method according to claim 2, wherein when the push basis information includes the interest point similarity information, obtaining the push basis information for the target user according to the location point embedding matrix comprises: Detecting a current point of interest visited by the target user; Recalling a plurality of candidate interest points according to the current interest point; Acquire a position semantic vector representation of the current point of interest and a position semantic vector representation of each candidate point of interest among the multiple candidate points of interest from the position point embedding matrix; For each candidate interest point among the multiple candidate interest points, the interest point similarity between the current interest point and the candidate interest point is obtained according to the position semantic vector representation of the current interest point and the position semantic vector representation of the candidate interest point.

4. The method according to claim 3, wherein the step of pushing the point of interest to the target user according to the push basis information comprises: Selecting similar interest points to the current interest point from the plurality of candidate interest points according to the interest point similarity between each candidate interest point and the current interest point; The similar points of interest are pushed to the target user.

5. The method according to claim 2, wherein when the push basis information includes the user similarity information, obtaining the push basis information for the target user according to the location point embedding matrix comprises: Acquire a location semantic vector representation of each location point in the historical trajectory sequence of the target user from the location point embedding matrix, and generate a trajectory vector representation of the target user according to the location semantic vector representation of each location point in the historical trajectory sequence of the target user; Determine a plurality of candidate users for the target user from the user set; For each of the candidate users, a location semantic vector representation of each location point in the historical trajectory sequence of the candidate user is obtained from the location point embedding matrix; and a trajectory vector representation of the candidate user is generated based on the location semantic vector representation of each location point in the historical trajectory sequence of the candidate user; and a user similarity between the candidate user and the target user is obtained based on the trajectory vector representation of the candidate user and the trajectory vector representation of the target user.

6. The method according to claim 5, wherein the step of pushing the point of interest to the target user according to the push basis information comprises: Selecting a reference user from the plurality of candidate users according to a user similarity between each of the candidate users and the target user; The current points of interest visited by the reference user are pushed to the target user.

7. The method according to claim 1, wherein the step of pushing the point of interest to the target user according to the push basis information comprises: Determining potential points of interest for the target user from the potential behavior trajectory; The potential points of interest are pushed to the target user.

8. The method according to claim 1, wherein the step of training a word representation model using the historical trajectory sequence of each user in the user set to obtain a position point embedding matrix comprises: Generating training sample data according to the historical trajectory sequence of each user in the user set, wherein the training sample data includes a training example for each point of interest in the check-in record of the point of interest; Training a word representation model according to the training sample data to obtain a trained word representation model; The position point embedding matrix is ​​obtained according to the hidden layer weight vector of the trained word representation model.

9. The method according to claim 8, wherein generating training sample data according to the historical trajectory sequence of each user in the user set comprises: Selecting the i-th point of interest from the historical trajectory sequence of the n-th user in the user set, where n is 1 to N, i is 1 to Z, N is the number of users in the user set, and Z is the number of points of interest in the historical trajectory sequence of the n-th user; According to a preset context window, determining a context location point of the i-th point of interest from the historical trajectory sequence of the n-th user; Select C points of interest different from the i-th point of interest, construct corresponding negative samples based on the context position points of the i-th point of interest and the C points of interest, and construct positive samples based on the context position points of the i-th point of interest and the i-th point of interest; A group of training examples for the i-th interest point is formed based on the positive examples and the negative examples.

10. A device for pushing points of interest, comprising: A determination unit, used to determine a target user, wherein the target user is any user among all users who has an interest in pushing a demand; An interest point push module, used for pushing interest points to the target user according to a position point embedding matrix, comprising: an information calculation unit, used for obtaining push basis information for the target user according to the position point embedding matrix, wherein, in the case where the push basis information includes a potential behavior trajectory, obtaining a position semantic vector representation of each position point in the historical trajectory sequence of the target user from the position point embedding matrix; obtaining a trajectory vector matrix for the target user according to the position semantic vector representation of each position point in the historical trajectory sequence of the target user; predicting a potential behavior trajectory of the target user according to the trajectory vector matrix, wherein the potential behavior trajectory includes more than one interest point, and a push unit, used for pushing interest points to the target user according to the push basis information; The location point embedding matrix is ​​obtained by the following steps: obtaining the interest point check-in records and GPS point collection records of the user set, wherein several location points in the GPS point collection records overlap with the interest points in the interest point check-in records, and the GPS point collection records are location points with geographical identification collected by a mobile device through GPS; constructing a historical trajectory sequence of each user in the user set according to the interest point check-in records and the GPS point collection records; using the historical trajectory sequence of each user in the user set to train a word representation model to obtain a location point embedding matrix, wherein the location point embedding matrix contains a location semantic vector representation of each interest point in the interest point check-in record, and a location semantic vector representation of each location point in the GPS point collection record.

11. The device according to claim 10, wherein the push basis information can also be the following information for the target user: interest point similarity information, user similarity information.

12. The device according to claim 11, wherein when the push basis information includes the interest point similarity information, the information calculation unit comprises: A detection subunit, configured to detect a current point of interest visited by the target user; A recall subunit, configured to recall a plurality of candidate interest points according to the current interest point; A first vector acquisition subunit, configured to acquire a position semantic vector representation of the current point of interest and a position semantic vector representation of each candidate point of interest among the multiple candidate points of interest from the position point embedding matrix; The calculation subunit is used to obtain, for each candidate interest point among the multiple candidate interest points, the interest point similarity between the current interest point and the candidate interest point according to the position semantic vector representation of the current interest point and the position semantic vector representation of the candidate interest point.

13. The device according to claim 12, wherein the pushing unit comprises: an interest point selection subunit, configured to select an interest point similar to the current interest point from the plurality of candidate interest points according to the interest point similarity between each candidate interest point and the current interest point; The first push subunit is used to push the similar points of interest to the target user.

14. The device according to claim 11, wherein when the push basis information includes the user similarity information, the information calculation unit comprises: a trajectory representation subunit, configured to obtain a position semantic vector representation of each position point in the historical trajectory sequence of the target user from the position point embedding matrix, and generate a trajectory vector representation of the target user according to the position semantic vector representation of each position point in the historical trajectory sequence of the target user; A user determination subunit, configured to determine a plurality of candidate users for the target user from the user set; The similarity calculation subunit is used to obtain, for each candidate user, a position semantic vector representation of each position point in the historical trajectory sequence of the candidate user from the position point embedding matrix; and generate a trajectory vector representation of the candidate user based on the position semantic vector representation of each position point in the historical trajectory sequence of the candidate user; and obtain the user similarity between the candidate user and the target user based on the trajectory vector representation of the candidate user and the trajectory vector representation of the target user.

15. The device according to claim 14, wherein the pushing unit comprises: a user selection subunit, configured to select a reference user from the plurality of candidate users according to a user similarity between each of the candidate users and the target user; The second push subunit is used to push the current point of interest visited by the reference user to the target user.

16. The device according to claim 10, wherein the pushing unit comprises: An interest point determination subunit, configured to determine potential interest points for the target user from the potential behavior trajectory; The third push subunit is used to push the potential points of interest to the target user.

17. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the steps of any one of the methods of claims 1-9 when executing the program.

18. A computer-readable storage medium having a computer program stored thereon, wherein when the program is executed by a processor, the steps of the method according to any one of claims 1 to 9 are implemented.