CN118827771A - A next point of interest recommendation method based on adaptive graph and future preference - Google Patents

A next point of interest recommendation method based on adaptive graph and future preference Download PDF

Info

Publication number
CN118827771A
CN118827771A CN202410904235.7A CN202410904235A CN118827771A CN 118827771 A CN118827771 A CN 118827771A CN 202410904235 A CN202410904235 A CN 202410904235A CN 118827771 A CN118827771 A CN 118827771A
Authority
CN
China
Prior art keywords
user
preference
interest
future
term
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Granted
Application number
CN202410904235.7A
Other languages
Chinese (zh)
Other versions
CN118827771B (en
Inventor
李肖雪
李博涵
刘昕悦
殷崭祚
徐泽坤
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Nanjing University of Aeronautics and Astronautics
Original Assignee
Nanjing University of Aeronautics and Astronautics
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Nanjing University of Aeronautics and Astronautics filed Critical Nanjing University of Aeronautics and Astronautics
Priority to CN202410904235.7A priority Critical patent/CN118827771B/en
Publication of CN118827771A publication Critical patent/CN118827771A/en
Application granted granted Critical
Publication of CN118827771B publication Critical patent/CN118827771B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Classifications

    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04L—TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
    • H04L67/00—Network arrangements or protocols for supporting network services or applications
    • H04L67/50—Network services
    • H04L67/55—Push-based network services
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/042—Knowledge-based neural networks; Logical representations of neural networks
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/044—Recurrent networks, e.g. Hopfield networks
    • G06N3/0442—Recurrent networks, e.g. Hopfield networks characterised by memory or gating, e.g. long short-term memory [LSTM] or gated recurrent units [GRU]
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G06N3/0455—Auto-encoder networks; Encoder-decoder networks
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/0464—Convolutional networks [CNN, ConvNet]
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G06N3/082—Learning methods modifying the architecture, e.g. adding, deleting or silencing nodes or connections
    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04L—TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
    • H04L67/00—Network arrangements or protocols for supporting network services or applications
    • H04L67/50—Network services
    • H04L67/52—Network services specially adapted for the location of the user terminal

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • General Health & Medical Sciences (AREA)
  • Computing Systems (AREA)
  • Biomedical Technology (AREA)
  • Biophysics (AREA)
  • Computational Linguistics (AREA)
  • Data Mining & Analysis (AREA)
  • Evolutionary Computation (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Molecular Biology (AREA)
  • Artificial Intelligence (AREA)
  • General Engineering & Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Health & Medical Sciences (AREA)
  • Computer Networks & Wireless Communication (AREA)
  • Signal Processing (AREA)
  • Management, Administration, Business Operations System, And Electronic Commerce (AREA)

Abstract

本发明公开了一种基于自适应图与未来偏好的下一个兴趣点推荐方法，包括如下步骤：准备兴趣点历史签入数据集，数据集包括用户历史签入地点，签入时间，签入地点类别等信息，并对数据进行标准化处理；初始化POI节点嵌入信息，利用自适应图进行全局邻接矩阵的构建，利用相似度函数和卷积网络得到兴趣点最终嵌入表达；将时空节点编码信息与其他侧信息融合兴趣点嵌入表达，引入自注意机制，编码得到用户的长期偏好；之后利用长短期记忆递归神经网络，融合兴趣点的侧信息，引入注意力机制，得到用户的短期偏好；随后将长期偏好与短期偏好以自组装方式，加入注意力机制，以下一步时间时刻编码作为查询向量，捕捉下一个兴趣点的关联关系。本发明选择自适应图和图神经网络，可以对真实的兴趣点进行关联捕获，具有很好的推荐性能。

The present invention discloses a method for recommending the next point of interest based on an adaptive graph and future preferences, comprising the following steps: preparing a historical check-in data set of points of interest, the data set including information such as the user's historical check-in location, check-in time, and check-in location category, and performing standardization on the data; initializing POI node embedding information, using an adaptive graph to construct a global adjacency matrix, and using a similarity function and a convolutional network to obtain the final embedding expression of the point of interest; fusing the spatiotemporal node encoding information with other side information to obtain the embedded expression of the point of interest, introducing a self-attention mechanism, and encoding to obtain the user's long-term preference; then using a long short-term memory recursive neural network to fuse the side information of the point of interest, introducing an attention mechanism, and obtaining the user's short-term preference; then adding the attention mechanism to the long-term preference and the short-term preference in a self-assembly manner, using the next time moment encoding as a query vector, and capturing the association relationship of the next point of interest. The present invention selects an adaptive graph and a graph neural network, can capture the association of real points of interest, and has good recommendation performance.

Description

Next interest point recommendation method based on self-adaptive graph and future preference
Technical Field
The invention belongs to the field of recommendation systems, and discloses a next interest point recommendation method realized by applying adaptive graph learning and future preference.
Background
Today, the rapid development of modernization, the demand for personalized services is increasing, various leisure, entertainment and play sites are layered endlessly, and the rapid development of location-based social networks (LBSN) has spawned a recommendation system in combination with itself, resulting in POI recommendations. The explosive growth of network geographic resources has led to a dramatic increase in demand for POI recommendations, promoting their explosive growth, and providing valuable geographic information to service providers and users. In view of the increasing demand for real-time and personalization from business scenarios, the next POI recommendation is attracting wide attention in academia and industry. Along with the maturation of space-time big data, the internet of things, deep learning technology and the like, point of interest (POI) recommendation is rapidly developed. In POI recommendation, there is a strong temporal and spatial correlation between sign-on data, and the conventional recommendation method makes capturing user preferences a great challenge due to the individuality and complexity of the user preferences. Moreover, how to consider the geographical dependency relationship between global POI nodes is also a direction of urgent research.
POI (Point of Interest) recommends refer to recommending surrounding or interesting places to the user based on the user's location information and historical behavior. Such a recommendation system may help a user find new places, plan travel routes, find surrounding service facilities, etc. In the process of making recommendations, there are mainly two problems: firstly, a global POI map is constructed by adopting a GNN-based method. However, GNN-based approaches are limited by the inherent limitations of fixed graphs: namely, the construction of the input graph is limited by priori knowledge, and the selection of the statistical indexes leads to incomplete graph structure and noise to a certain extent, so that the performance of characterization learning is restricted; secondly, it is: existing future preference methods ignore interactions of attribute features, particularly spatio-temporal information. Personal access is affected not only by the user's travel plan, but also by the spatiotemporal relationship between the different POIs. Ignoring the spatiotemporal weighted information would consider all historic check-in behavior as a unified contribution, thereby weakening the ability to construct other side information features.
In recent years, deep learning has been attracting attention because of its ability to model data in a high-dimensional nonlinear manner, and has achieved good results in recommendation systems. But mostly rely on a predefined adjacency matrix based on priori knowledge as the basis for global graph construction, and the graph structure is not well processed by only relying on the graph neural network and the predefined adjacency matrix. How to handle efficient modeling and related space-time coding of long-time-ordered data is also a problem to be solved in dealing with user future preferences. The existing models do not adequately consider how to learn and learn the user's future preferences using the adaptive graph to model the user's final preferences, resulting in recommendations.
Disclosure of Invention
Existing POI recommendation methods mostly rely on GNN graph convolution networks and static construction of POI global graphs, but GNN-based methods are limited by inherent limitations of fixed graphs: namely, the construction of the input graph is limited by priori knowledge, the selection of statistical indexes leads to a certain degree of graph structural incompleteness and noise, and the performance of characterization learning is restricted, so that we propose to solve the problem by using the adaptive graph, learn the adjacency matrix of the adaptive graph through similarity measurement, and automatically deduce the potential geographic relation among POIs. Meanwhile, the existing future preference recommendation method mainly does not consider space-time coding, and in order to solve the problems, the invention aims to provide a next POI recommendation method based on adaptive graph learning and future preference.
In order to achieve the above purpose, the invention adopts the following technical scheme:
A next interest point recommending method based on an adaptive graph and future preference comprises the following steps:
Step 1, preparing a historical sign-on dataset of a user, wherein the historical dataset comprises data values such as longitude and latitude of a sign-on place, category of the sign-on place, time of the sign-on place and the like, establishing a preprocessing method, constructing a historical sign-on track of the user and an initial interest point adjacent matrix, and obtaining initial input data.
In the step 1, we use three real data sets from new york city, phoenix city and singapore, each data set includes sign-on time, sign-on location category, sign-on user id, etc., taking singapore city as an example, the data set includes 8648 users, 33172 interest points 355337 sign-on data, sparsity of the data set is 0.12%, time information in the data is selected, including hours and whether it is workday, category information, longitude and latitude information, and user information as main characteristic values, and recommendation sequence of next interest point is predicted.
Step 2, constructing a final embedding expression of the interest points, inputting the final embedding expression as an initial embedding vector of the interest points, utilizing the initial embedding expression of the POI, taking cosine similarity as a measurement function, learning an adjacent matrix of the self-adaptive graph through similarity measurement, automatically deducing potential geographic relation between the interest points, introducing a threshold value to carry out sparsification processing on the graph, and carrying out convolution aggregation on neighborhood information on the basis to obtain the final embedding expression of the interest pointsAnd as a subsequent input.
In the step2, the overall structure of the adaptive graph learning is as follows:
wherein we initialize an initial embedded representation of a global POI D id denotes the feature dimension, and then the dependency between each pair of POIs is inferred by computing the embedded similarity between the two POIs. Using cosine similarity as a metric function, each entry a i,j in the adjacency matrix a, which is a fully connected graph, a i,j e (-1, 1), is finally obtained, and the entries based on the adaptive graph in the matrix reflect an accurate estimate of the geographic dependency between points of interest. But this result can lead to excessive trivial edges, creating a noisy structure. Therefore, a threshold is introduced to perform graphics sparsification, so as to perfect dynamic characteristics related to spatial dependence, to obtain a non-negative adjacency matrix A i,j E (0, 1), and to enhance expressive force of interest points, neighborhood information is aggregated, and initial characteristics of nodes in the graph are convolved, wherein the expression is as follows:
Acquiring a high-order representation of a point of interest UsingAs the final embedded representation of the point of interest.
Step 3, constructing long-term preference of usersThe input is the historical track data of the user within n-1 days, which is composed of the input data x obtained by connecting the final embedded expression of the interest point and the initial embedded expression of other side informationWherein the method comprises the steps ofRepresenting a single-day user track formed by a plurality of embedded data, then constructing a space interval matrix and a time interval matrix of the user embedded track, and encoding the space-time interval matrix to obtain space-time weight omega T,ωD. Secondly, by using a self-attention mechanism, track data in a specific number of days of a user is used as input to obtain association weights among nodes, and long-term preference of the user is obtained
In the step 3, the overall structure of the long-term preference encoder is as follows:
Wherein the user history track The sign-in point data of the user on the same day is obtained, and the sign-in record of each user in the sequence has three characteristics: user id u, POI representationCategory id c. Furthermore, since POIs are closely related to time information, we add two new time information T to capture user activities at a specific time, e.g. people like to go to parks on holidays, go to bars at night. Thus, the two pieces of time information are a weekly id feature t d containing 2 categories and an hourly id feature t h containing 24 categories, respectively. The embedding layer converts each id feature into a d id -dimensional dense vector, which is represented as a feature dimension, and then the pooling layer embeds and merges all five into check-in data x in the generated sequence, the formula is as follows:
The dimension of x is d h＝5did and the trace is expressed as Where m is the total number of checked-ins in the whole track.
Then, considering that the user long-term preference modeling is affected by the spatio-temporal environment, we consider the effect of the spatio-temporal weights when designing the long-term preference encoder. We note the space-time interval between two registration points p i,pj as Δt i,j,ΔDi,j, respectively. The time impact weight and the distance impact weight are denoted as W T(ΔTi,j),WDΔDi,j, respectively, specifically we use Havercosine periodic function to capture user time period data and set the time decay weight as follows:
Wperiod(ΔTi,j)＝hvc(2πΔTi,j)
Where α is the time decay rate, the rate at which the control weight decreases over the time interval ΔT i,j.
From a spatial perspective, we consider that the closer the POI is to the current location, the more helpful the location prediction. The more helpful the position prediction. The distance exponential decay weight thus represents:
where β is the time decay rate, controlling the rate at which the weight decreases with distance interval Δd i,j.
Finally, we use self-attention mechanism to encode long-term preference, the long-term preference is obtained by fusing multiple side information and time-space information, the formula is as follows:
Wherein the method comprises the steps of All belonging to the transformation weight matrix.
Step 4, constructing a user long-term preference H short which is input as a history track of the user on the same dayShort-term preferences H short for the user are constructed using a long-term memory recurrent neural network LSTM.
In the step 4, the overall structure of the short-term preference H short is as follows:
The input is the historical track of the user on the same day, and the short-term preference encoder captures the time perception sequence dependency of the user from the current sign-in behavior of the user to obtain the short-term preference of the user as H short.
The calculation of the short-term preference encoder may be defined as follows:
we use the current trajectory S short＝(x1,x1,...xm) as input data to feed into LSTM network as follows:
wherein, AndThe hidden states of time stamps t i and t i-1, respectively.
Step 5, constructing user future preference, considering self-activity plan of the user, introducing attention mechanism, embedding expression as query vector through future time, taking hidden state matrix of long-term preference as content vector, extracting potential future behavior from daily periodic behavior of the user, obtaining future time preference weight, which implicitly simulates activity schedule of the userCombining the obtained time preference weight with short-term preference H short, embedding the user as a query vector to obtain user future preference H f, and learning probability distribution of the user preference to obtain a final prediction result.
In the step 5, the overall structure of the future preference H f is as follows:
wherein input X is a future time preference weight obtained by fusing future behavior with long-term encoding Because of the consideration that the user may have a self-activity plan when going out, time-aware attention may be used to identify the behavior in the past long-term trajectory that is most relevant to the future time context, in particular, the user long-term trajectoryIs encoded asThe state matrix is hidden and future time preference weights are obtained by extracting potential future behaviors from the user's daily periodic behaviors. Using the future time embedded t k+1 as a query vector, the formula is as follows:
the weights are then combined with short-term preferences by means of a connection to generate hidden states characterizing the user preferences The final future preference of the user is determined byIs summarized, so the final future preference of the user is:
where u represents the user embedded representation. We then learn the probability distribution of the user preferences, decoding the points of interest:
Wherein the method comprises the steps of Is a transformation matrix.
Drawings
FIG. 1 is an overall block diagram of adaptive graph learning in the present invention;
FIG. 2 is an overall block diagram of a medium and long term preference encoder of the present invention;
FIG. 3 is an overall block diagram of constructing a short-term preference encoder in accordance with the present invention;
FIG. 4 is an overall block diagram of a user's future preferences in the present invention;
FIG. 5 is an overall block diagram of a model in the present invention;
Detailed Description
The present invention will be further explained below with reference to the drawings in the embodiments of the present invention.
The invention discloses a next interest point recommending method based on an adaptive graph and future preference, which comprises the following steps:
Step 1, preparing a historical sign-on dataset of a user, wherein the historical dataset comprises data values such as longitude and latitude of a sign-on place, category of the sign-on place, time of the sign-on place and the like, establishing a preprocessing method, constructing a historical sign-on track of the user and an initial interest point adjacent matrix, and obtaining initial input data.
Step 2, constructing a final embedding expression of the interest points, inputting the final embedding expression as an initial embedding vector of the interest points, utilizing the initial embedding expression of the POI, taking cosine similarity as a measurement function, learning an adjacent matrix of the self-adaptive graph through similarity measurement, automatically deducing potential geographic relation between the interest points, introducing a threshold value to carry out sparsification processing on the graph, and carrying out convolution aggregation on neighborhood information on the basis to obtain the final embedding expression of the interest pointsAnd as a subsequent input.
Step 3, constructing long-term preference of usersThe input is the historical track data of the user within n-1 days, which is composed of the input data x obtained by connecting the final embedded expression of the interest point and the initial embedded expression of other side informationWherein the method comprises the steps ofRepresenting a single-day user track formed by a plurality of embedded data, then constructing a space interval matrix and a time interval matrix of the user embedded track, and encoding the space-time interval matrix to obtain space-time weight omega T,ωD. Secondly, by using a self-attention mechanism, track data in a specific number of days of a user is used as input to obtain association weights among nodes, and long-term preference of the user is obtained
Step 4, constructing a user long-term preference H short which is input as a history track of the user on the same dayShort-term preferences H short for the user are constructed using a long-term memory recurrent neural network LSTM.
Step 5, constructing user future preference, considering self-activity plan of the user, introducing attention mechanism, embedding expression as query vector through future time, taking hidden state matrix of long-term preference as content vector, extracting potential future behavior from daily periodic behavior of the user, obtaining future time preference weight, which implicitly simulates activity schedule of the userCombining the obtained time preference weight with short-term preference H short, embedding the user as a query vector to obtain user future preference H f, and learning probability distribution of the user preference to obtain a final prediction result.
In the step 1, we use three real data sets from new york city, phoenix city and singapore, each data set includes sign-on time, sign-on location category, sign-on user id, etc., taking singapore city as an example, the data set includes 8648 users, 33172 interest points 355337 sign-on data, sparsity of the data set is 0.12%, time information in the data is selected, including hours and whether it is workday, category information, longitude and latitude information, and user information as main characteristic values, and recommendation sequence of next interest point is predicted.
In the step 2, please refer to fig. 1 for the overall structure of the adaptive graph learning, the overall structure of the adaptive graph learning is as follows:
wherein we initialize an initial embedded representation of a global POI D id denotes the feature dimension, and then the dependency between each pair of POIs is inferred by computing the embedded similarity between the two POIs. Using cosine similarity as a metric function, each entry a i,j in the adjacency matrix a, which is a fully connected graph, a i,j e (-1, 1), is finally obtained, and the entries based on the adaptive graph in the matrix reflect an accurate estimate of the geographic dependency between points of interest. But this result can lead to excessive trivial edges, creating a noisy structure. Therefore, a threshold is introduced to perform graphics sparsification, so as to perfect dynamic characteristics related to spatial dependence, to obtain a non-negative adjacency matrix A i,j E (0, 1), and to enhance expressive force of interest points, neighborhood information is aggregated, and initial characteristics of nodes in the graph are convolved, wherein the expression is as follows:
Acquiring a high-order representation of a point of interest UsingAs the final embedded representation of the point of interest.
In the step 3, please refer to fig. 2 for the overall structure of the long-term preference encoder, and the long-term preference encoder is integrated
The body structure is as follows:
Wherein the user history track The sign-in point data of the user on the same day is obtained, and the sign-in record of each user in the sequence has three characteristics: user id u, POI representationCategory id c. Furthermore, since POIs are closely related to time information, we add two new time information T to capture user activities at a specific time, e.g. people like to go to parks on holidays, go to bars at night. Thus, the two pieces of time information are a weekly id feature t d containing 2 categories and an hourly id feature t h containing 24 categories, respectively. The embedding layer converts each id feature into a d id -dimensional dense vector, which is represented as a feature dimension, and then the pooling layer embeds and merges all five into check-in data x in the generated sequence, the formula is as follows:
The dimension of x is d h＝5did and the trace is expressed as Where m is the total number of checked-ins in the whole track.
Then, considering that the user long-term preference modeling is affected by the spatio-temporal environment, we consider the effect of the spatio-temporal weights when designing the long-term preference encoder. We note the space-time interval between two registration points p i,pj as Δt i,j,ΔDi,j, respectively. The time impact weight and the distance impact weight are denoted as W T(ΔTi,j),WDΔDi,j, respectively, specifically we use Havercosine periodic function to capture user time period data and set the time decay weight as follows:
Wperiod(ΔTi,j)＝hvc(2πΔTi,j)
Where α is the time decay rate, the rate at which the control weight decreases over the time interval ΔT i,j.
Spatially, we consider that the closer a POI is to the current location, the more helpful it is in location prediction
Large. The more helpful the position prediction. The distance exponential decay weight thus represents:
where β is the time decay rate, controlling the rate at which the weight decreases with distance interval Δd i,j.
Finally, we use self-attention mechanism to encode long-term preference, which fuses multiple side messages
The information and the air information are obtained by the following formula:
Wherein the method comprises the steps of All belonging to the transformation weight matrix.
In the step 4, referring to fig. 3, the short-term preference encoder overall structure is shown in the short-term preference H short overall structure
The structure is as follows:
The input is the historical track of the user on the same day, and the short-term preference encoder captures the time perception sequence dependency of the user from the current sign-in behavior of the user to obtain the short-term preference of the user as H short.
The calculation of the short-term preference encoder may be defined as follows:
we use the current trajectory S short＝(x1,x1,...xm) as input data to feed into LSTM network as follows:
wherein, AndThe hidden states of time stamps t i and t i-1, respectively.
In the step 5, referring to fig. 4, the overall structure of the future preference H f is as follows:
wherein input X is a future time preference weight obtained by fusing future behavior with long-term encoding Because of the consideration that the user may have a self-activity plan when going out, time-aware attention may be used to identify the behavior in the past long-term trajectory that is most relevant to the future time context, in particular, the user long-term trajectoryIs encoded asThe state matrix is hidden and future time preference weights are obtained by extracting potential future behaviors from the user's daily periodic behaviors. Using the future time embedded t k+1 as a query vector, the formula is as follows:
the weights are then combined with short-term preferences by means of a connection to generate hidden states characterizing the user preferences The final future preference of the user is determined byIs summarized, so the final future preference of the user is:
where u represents the user embedded representation. We then learn the probability distribution of the user preferences, decoding the points of interest:
Wherein the method comprises the steps of Is a transformation matrix.
In the invention, in the aspect of data set processing, unlike a general flow prediction algorithm, the real traffic flow data set is adopted as training, so that the real availability of the algorithm is enhanced; in the aspect of network structure, the invention uses a multi-head attention spread graph convolution network to dynamically simulate traffic flow mode, thereby enhancing the accuracy of traffic flow prediction; in the aspect of predicting the target, the method selects the traffic flow prediction in the short-term time in the future as the target, so that the prediction content of the method is more accurate and real-time. Experiments show that the model provided by the method can basically realize effective prediction of traffic flow, and the finally obtained effect has authenticity.
The foregoing is merely a preferred embodiment of the present invention and it should be noted that modifications and adaptations to those skilled in the art may be made without departing from the principles of the present invention, which are intended to be comprehended within the scope of the present invention.

Claims (6)

1.一种基于自适应图与未来偏好的下一个兴趣点推荐方法，包括如下步骤：1. A method for recommending next points of interest based on adaptive graph and future preference, comprising the following steps: 步骤1，准备用户历史签入数据集，历史数据集包含签入地点经纬度，签入地点的类别，签入地点的时间等数据值，建立预处理的方法，构建用户的历史签入轨迹，以及初始的兴趣点邻接矩阵，得到初始输入数据。Step 1: Prepare the user's historical check-in dataset. The historical dataset contains data values such as the latitude and longitude of the check-in location, the category of the check-in location, and the time of the check-in location. Establish a preprocessing method to construct the user's historical check-in trajectory and the initial point of interest adjacency matrix to obtain the initial input data. 步骤2，构建兴趣点的最终嵌入表达，其输入为兴趣点的初始嵌入向量，利用POI的初始嵌入表示，将余弦相似度作为度量函数，通过相似性度量学习自适应图的邻接矩阵，从而自动推断兴趣点之间的潜在地理关系并引入了一个阈值来对图进行稀疏化处理，并且在此基础上，对邻域信息进行了卷积聚合，得到兴趣点的最终嵌入表示并作为后续的输入。Step 2: Construct the final embedding representation of the POI. The input is the initial embedding vector of the POI. Using the initial embedding representation of the POI, the cosine similarity is used as the metric function. The adjacency matrix of the adaptive graph is learned through the similarity metric, so as to automatically infer the potential geographical relationship between the POIs and introduce a threshold to sparse the graph. On this basis, the neighborhood information is convolutionally aggregated to obtain the final embedding representation of the POI. And serve as subsequent input. 步骤3，构建用户长期偏好其输入为兴趣点最终嵌入表达与其他侧信息的初始嵌入进行连接得到的输入数据x所构成的用户n-1天内的历史轨迹数据其中表示由多个嵌入数据形成的用户单天用户轨迹，然后构建用户嵌入轨迹的空间间隔矩阵与时间间隔矩阵，将时空间隔矩阵编码得到时空权重ωT,ωD。其次，利用自注意力机制，以用户特定天数内的轨迹数据为输入，得到节点之间的关联权重，得到用户的长期偏好 Step 3: Build long-term user preferences Its input is the historical trajectory data of the user within n-1 days, which is formed by connecting the final embedding expression of the interest point with the initial embedding of other side information. in Represents the user's single-day user trajectory formed by multiple embedded data, and then constructs the spatial interval matrix and time interval matrix of the user's embedded trajectory, and encodes the time-space interval matrix to obtain the time-space weights ω T , ω D . Secondly, using the self-attention mechanism, taking the user's trajectory data within a specific number of days as input, the association weights between nodes are obtained to obtain the user's long-term preference 步骤4，构建用户长期偏好Hshort，其输入为当天用户的历史轨迹利用长短期记忆递归神经网络LSTM，构建用户的短期偏好Hshort。Step 4: Construct the user's long-term preference H short , whose input is the user's historical trajectory on that day The long short-term memory recurrent neural network LSTM is used to construct the user's short-term preference H short . 步骤5，构建用户未来偏好，考虑用户的自我活动计划，引入注意力机制，通过未来时刻嵌入表达作为查询向量，以长期偏好的隐藏状态矩阵作为内容向量，从用户的日常周期性行为中提取潜在的未来行为，获得未来时间偏好权重，这隐含模拟了用户的活动时间表将得到的时间偏好权重与短期偏好Hshort结合，并以用户嵌入作为查询向量，得到用户未来偏好Hf，学习用户偏好的概率分布，得到最终的预测结果。Step 5: Construct user future preferences. Consider the user's self-activity plan and introduce the attention mechanism. Use the future moment embedding expression as the query vector and the long-term preference hidden state matrix as the content vector to extract potential future behaviors from the user's daily periodic behavior and obtain the future time preference weight, which implicitly simulates the user's activity schedule. The obtained time preference weight is combined with the short-term preference H short , and the user embedding is used as the query vector to obtain the user's future preference H f , learn the probability distribution of user preferences, and obtain the final prediction result. 2.根据权利要求1所述的一种基于自适应图与未来偏好的下一个兴趣点推荐方法，其特征在于：所述步骤1中，我们使用了来自纽约市，凤凰城市以及新加坡三个真实数据集，每个数据集都包含了签入的时间，签入地点类别，签入用户id等信息，以新加坡市为例，该数据集包含了8648个用户，33172个兴趣点355337条签入数据，该数据集的稀疏性为0.12％，选取该数据中的时间信息，包含小时与是否为工作日，类别信息，经纬度信息，以及用户信息作为主要的特征值，预测下一个兴趣点的推荐序列。2. According to claim 1, a method for recommending the next point of interest based on adaptive graph and future preference is characterized in that: in the step 1, we used three real data sets from New York City, Phoenix City and Singapore, each data set contains information such as check-in time, check-in location category, check-in user ID, etc. Taking Singapore as an example, the data set contains 8648 users, 33172 points of interest and 355337 check-in data. The sparsity of the data set is 0.12%. The time information in the data, including hours and whether it is a weekday, category information, longitude and latitude information, and user information are selected as the main feature values to predict the recommendation sequence of the next point of interest. 3.根据权利要求1所述的一种基于自适应图与未来偏好的下一个兴趣点推荐方法，其特征在于：所述步骤2中，自适应图学习的整体结构如下：3. The next point of interest recommendation method based on adaptive graph and future preference according to claim 1, characterized in that: in step 2, the overall structure of adaptive graph learning is as follows: input初始兴趣点嵌入→自适应图学习→兴趣点的最终嵌入表达 Input initial interest point embedding → adaptive graph learning → final embedding expression of interest points 其中我们初始化全局POI的初始嵌入表示did表示特征维度，然后通过计算两POI之间的嵌入相似度来推断每对POI之间的依赖关系。使用余弦相似度作为度量函数，最后得到邻接矩阵A中的每个条目Ai,j,这是一个全连接图，Ai,j∈(-1,1)，该矩阵中基于自适应图的项反映了对兴趣点之间地理依赖关系的准确估计。但这一结果会导致过多的琐碎边，从而产生噪声结构。因此引入一个阈值来进行图形稀疏化，目的是完善与空间依赖性相关的动态特征，以获得非负邻接矩阵Ai,j∈(0,1)，并且为了增强兴趣点的表现力，对邻域信息进行了聚合，对图中节点的初始特征进行卷积，表达如下：We initialize the initial embedding representation of the global POI d id represents the feature dimension, and then the dependency between each pair of POIs is inferred by calculating the embedding similarity between the two POIs. Using cosine similarity as the metric function, each entry A i,j in the adjacency matrix A is finally obtained. This is a fully connected graph, A i,j ∈(-1,1). The items based on the adaptive graph in this matrix reflect the accurate estimation of the geographic dependency between points of interest. However, this result will lead to too many trivial edges, resulting in a noisy structure. Therefore, a threshold is introduced to perform graph sparsification, with the aim of improving the dynamic features related to spatial dependencies to obtain a non-negative adjacency matrix A i,j ∈(0,1), and in order to enhance the expressiveness of the points of interest, the neighborhood information is aggregated and the initial features of the nodes in the graph are convolved, as expressed as follows: 获取兴趣点的高阶表示使用作为兴趣点的最终嵌入表达。Get a high-level representation of interest points use As the final embedding expression of the interest point. 4.根据权利要求1所述的一种基于自适应图与未来偏好的下一个兴趣点推荐方法，其特征在于：所述步骤3中，长期偏好编码器的整体结构如下：4. The next point of interest recommendation method based on adaptive graph and future preference according to claim 1, characterized in that: in step 3, the overall structure of the long-term preference encoder is as follows: 用户历史轨迹→长期偏好编码器→ User history → Long-term preference encoder → 其中用户历史轨迹由用户当天签入点数据得到，序列中每个用户的签到记录都有三个特征：用户id u、POI表示类别id c。此外，由于POI与时间信息密切相关，我们增加了两个新的时间信息T来捕捉特定时间的用户活动，例如，人们喜欢在休息日去公园，晚上去酒吧。因此，这两个时间信息分别是包含2个类别的每周id特征td和包含24个类别的每小时id特征th。嵌入层将每个id特征转换成did维稠密向量，表示其为特征维数，然后池化层将所有五个嵌入合并生成序列中的签到数据x，公式如下所示：The user's historical trajectory The check-in point data of the user on the day is obtained. Each user's check-in record in the sequence has three features: user id u, POI representation Category id c. In addition, since POI is closely related to time information, we add two new time information T to capture user activities at specific times, for example, people like to go to parks on weekends and go to bars at night. Therefore, the two time information are weekly id features td containing 2 categories and hourly id features th containing 24 categories. The embedding layer converts each id feature into a d id- dimensional dense vector, indicating that it is a feature dimension, and then the pooling layer merges all five embeddings to generate the check-in data x in the sequence. The formula is as follows: x的维度为dh＝5did，轨迹表示为其中m为整条轨迹中的总签入个数。The dimension of x is d h =5d id , and the trajectory is expressed as Where m is the total number of check-ins in the entire trajectory. 之后，考虑到用户长期偏好建模受到时空环境的影响，我们在设计长期偏好编码器时考虑了时空权重的影响。我们将两个注册点pi,pj之间的时空间隔分别记为ΔTi,j,ΔDi,j。时间影响权重和距离影响权重分别记为WT(ΔTi,j),WDΔDi,j，具体来说，我们使用Havercosine周期性函数捕捉用户时间周期数据，并且设置了时间衰减权重，具体公式如下：Afterwards, considering that the user's long-term preference modeling is affected by the spatiotemporal environment, we considered the influence of spatiotemporal weights when designing the long-term preference encoder. We denote the spatiotemporal interval between two registration points p i ,p j as ΔT i,j ,ΔD i,j . The time influence weight and distance influence weight are denoted as W T (ΔT i,j ) and W D ΔD i,j , respectively. Specifically, we use the Havercosine periodic function to capture the user's time period data and set the time decay weight. The specific formula is as follows: Wperiod(ΔTi,j)＝hvc(2πΔTi,j)W period (ΔT i,j )=hvc (2πΔT i,j ) 其中α是时间衰减率，控制权重随时间间隔ΔTi,j下降的速度。Where α is the time decay rate, which controls how fast the weight decreases over the time interval ΔT i,j . 从空间角度来看，我们认为POI离当前位置越近，对位置预测的帮助就越大。就越有助于位置预测。因此距离指数衰减权重表示：From a spatial perspective, we believe that the closer the POI is to the current location, the more helpful it is for location prediction. Therefore, the distance exponential decay weight is expressed as: 其中β是时间衰减率，控制着权重随距离间隔ΔDi,j下降的速度。Where β is the time decay rate, which controls how fast the weight decreases with the distance interval ΔD i,j . 最后，我们采用自注意力机制进行长期偏好的编码，长期偏好融合多个侧信息及时空信息得到，公式如下：Finally, we use the self-attention mechanism to encode long-term preferences. Long-term preferences are obtained by integrating multiple side information and spatiotemporal information. The formula is as follows: 其中都属于变换权重矩阵。in All belong to the transformation weight matrix. 5.根据权利要求1所述的一种基于自适应图与未来偏好的下一个兴趣点推荐方法，其特征在于，所述步骤4中，短期偏好Hshort整体结构如下：5. The next point of interest recommendation method based on adaptive graph and future preference according to claim 1, characterized in that in step 4, the overall structure of the short-term preference H short is as follows: 当天历史轨迹→短期偏好编码器→Hshort History of the day →Short-term preference encoder→H short 其中，输入为当天用户的历史轨迹，短期偏好编码器从用户当前的签到行为中捕获到用户的时间感知顺序依赖性，得到用户短期偏好为Hshort。The input is the historical trajectory of the user on that day. The short-term preference encoder captures the user's time-aware sequential dependency from the user's current check-in behavior, and obtains the user's short-term preference as H short . 短期偏好编码器的计算可以定义如下：The computation of the short-term preference encoder can be defined as follows: 我们利用当前轨迹Sshort＝(x1,x1,...xm)作为输入数据，馈入LSTM网络中，具体公式如下：We use the current trajectory S short = (x 1 , x 1 , ... x m ) as input data and feed it into the LSTM network. The specific formula is as follows: 其中，和分别为时间戳ti与ti-1的隐藏状态。in, and are the hidden states of timestamps ti and ti-1 respectively. 6.根据权利要求1所述的一种基于自适应图与未来偏好的下一个兴趣点推荐方法，其特征在于，所述步骤5中，未来偏好Hf整体结构如下：6. The next point of interest recommendation method based on adaptive graph and future preference according to claim 1, characterized in that in step 5, the overall structure of the future preference H f is as follows: input X→长期偏好编码→ input X → long-term preference encoding → 其中input X为通过融合融合未来行为与长期编码得到的未来时间偏好权重因为考虑到用户外出时可能会有自我活动计划，所以时间感知注意力可用于识别过去长期轨迹中与未来时间上下文最相关的行为，具体来说，用户长期轨迹被编码为隐藏状态矩阵，通过从用户的日常周期性行为中提取潜在的未来行为来获得未来时间偏好权重。使用未来时间嵌入tk+1作为查询向量，公式如下：Where input X is the future time preference weight obtained by integrating future behavior and long-term coding Because users may have self-planned activities when they are out and about, time-aware attention can be used to identify behaviors in past long-term trajectories that are most relevant to future temporal contexts. Specifically, user long-term trajectories Encoded as Hidden state matrix, obtains future time preference weights by extracting potential future behaviors from the user's daily periodic behaviors. Use the future time embedding t k+1 as the query vector, the formula is as follows: 之后通过连接方式将权重与短期偏好结合生成表征用户偏好的隐藏状态用户最终的未来偏好是由中的每个隐藏状态汇总而成的，因此用户最终未来偏好为：The weights are then combined with short-term preferences through connections to generate a hidden state that represents the user's preference. The user's final future preferences are determined by Each hidden state in is aggregated, so the user's final future preference is: 其中u表示用户嵌入表示。之后我们学习用户偏好的概率分布，解码兴趣点：Where u represents the user embedding representation. Then we learn the probability distribution of user preferences and decode the points of interest: 其中是一个变换矩阵。in is a transformation matrix.
CN202410904235.7A 2024-07-08 2024-07-08 A Next Interest Recommendation Method Based on Adaptive Graph and Future Preferences Active CN118827771B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202410904235.7A CN118827771B (en) 2024-07-08 2024-07-08 A Next Interest Recommendation Method Based on Adaptive Graph and Future Preferences

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202410904235.7A CN118827771B (en) 2024-07-08 2024-07-08 A Next Interest Recommendation Method Based on Adaptive Graph and Future Preferences

Publications (2)

Publication Number Publication Date
CN118827771A true CN118827771A (en) 2024-10-22
CN118827771B CN118827771B (en) 2025-11-25

Family

ID=93076037

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202410904235.7A Active CN118827771B (en) 2024-07-08 2024-07-08 A Next Interest Recommendation Method Based on Adaptive Graph and Future Preferences

Country Status (1)

Country Link
CN (1) CN118827771B (en)

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN119807553A (en) * 2025-03-17 2025-04-11 武汉东信同邦信息技术有限公司 Next point of interest recommendation method, device, equipment and storage medium
CN120821908A (en) * 2025-06-30 2025-10-21 中国测绘科学研究院 Method and device for recommending regions of interest by embedding user dependency into adaptive meta-paths

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112380426A (en) * 2020-10-23 2021-02-19 南京邮电大学 Interest point recommendation method and system based on graph embedding and user long-term and short-term interest fusion
CN115982468A (en) * 2023-01-04 2023-04-18 重庆邮电大学 A Point-of-Interest Recommendation Method Combining User Periodic Preferences and Short-term Interests
CN117370670A (en) * 2023-10-31 2024-01-09 南京航空航天大学 A sequence recommendation method that integrates timing awareness and user preferences
CN117828197A (en) * 2023-12-19 2024-04-05 北京工业大学 A point of interest recommendation method combining discontinuity and multi-level periodicity

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112380426A (en) * 2020-10-23 2021-02-19 南京邮电大学 Interest point recommendation method and system based on graph embedding and user long-term and short-term interest fusion
CN115982468A (en) * 2023-01-04 2023-04-18 重庆邮电大学 A Point-of-Interest Recommendation Method Combining User Periodic Preferences and Short-term Interests
CN117370670A (en) * 2023-10-31 2024-01-09 南京航空航天大学 A sequence recommendation method that integrates timing awareness and user preferences
CN117828197A (en) * 2023-12-19 2024-04-05 北京工业大学 A point of interest recommendation method combining discontinuity and multi-level periodicity

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
沈冬东;汪海涛;姜瑛;陈星;: "一种融合知识图谱与长短期偏好的下一项推荐算法", 小型微型计算机系统, no. 04, 9 April 2020 (2020-04-09) *

Cited By (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN119807553A (en) * 2025-03-17 2025-04-11 武汉东信同邦信息技术有限公司 Next point of interest recommendation method, device, equipment and storage medium
CN119807553B (en) * 2025-03-17 2025-07-08 武汉东信同邦信息技术有限公司 Method, device, equipment and storage medium for recommending next interest point
CN120821908A (en) * 2025-06-30 2025-10-21 中国测绘科学研究院 Method and device for recommending regions of interest by embedding user dependency into adaptive meta-paths

Also Published As

Publication number Publication date
CN118827771B (en) 2025-11-25

Similar Documents

Publication Publication Date Title
Wu et al. Location prediction on trajectory data: A review
CN113570859B (en) A Traffic Flow Prediction Method Based on Asynchronous Space-Time Dilated Graph Convolutional Networks
CN108875007B (en) Method and device for determining point of interest, storage medium, electronic device
CN118827771B (en) A Next Interest Recommendation Method Based on Adaptive Graph and Future Preferences
CN115240425A (en) Traffic prediction method based on multi-scale space-time fusion graph network
CN105389332B (en) It is a kind of geography social networks under user's similarity calculation method
Jiao et al. A novel next new point-of-interest recommendation system based on simulated user travel decision-making process
CN112231592B (en) Graph-based network community discovery method, device, equipment, and storage medium
CN113068131B (en) Method, device, equipment and storage medium for predicting user movement mode and track
CN114780866B (en) Personalized intelligent recommendation method based on spatio-temporal context interest learning model
CN114039871B (en) Method, system, device and medium for cellular traffic prediction
CN112632404B (en) Next point of interest recommendation method based on multi-granularity self-attention
CN110334293A (en) A kind of facing position social networks has Time Perception position recommended method based on fuzzy clustering
CN118606543A (en) A point of interest recommendation method based on spatiotemporal weighted dual-graph convolutional neural network
CN119066712A (en) Privacy protection method and device for sparse traffic flow prediction based on cloud-edge-end collaboration
CN115827898A (en) User intention-oriented spatio-temporal knowledge graph construction method
Sun et al. TCSA-Net: A temporal-context-based self-attention network for next location prediction
Xu et al. Simulating continuous-time human mobility trajectories
CN117633371A (en) Recommended methods, devices and readable storage media based on multi-attention mechanisms
CN120429503A (en) A point of interest recommendation method based on region-aware dynamic hypergraph and two-layer modeling
CN121051506A (en) A cross-scale self-supervised training method for spatiotemporal trajectory fundamental models
CN111738447B (en) Mobile social network user relationship inference method based on spatio-temporal relationship learning
CN114911996B (en) Missing POI Identification Method
CN114943407A (en) Area planning method, device, equipment, readable storage medium and program product
CN114638308A (en) Method and device for acquiring object relationship, electronic equipment and storage medium

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
GR01 Patent grant
GR01 Patent grant