Activity footprint multi-aspect preference pattern mining method based on holographic semantic trajectory big data

Abstract

The application discloses a kind of based on holographic semantic track big data activity footprint multi-aspect preference mode mining method, first, the GPS track of user original activity footprint is associated with multi-dimensional attribute semantic data according to space-time position and converges, respectively generates group and individual holographic semantic track data set;Second, the GPS track is grid encoded, and the multi-dimensional attribute semantic data is discretized coding;Respectively construct group and individual activity tensor;Finally, establish double-layer Tucker tensor decomposition method, decompose group activity tensor and extract group user activity base mode;Again based on group user activity base mode, individual activity tensor is decomposed, and individual user activity multi-aspect preference mode is extracted.The significant effect of the application is that multi-aspect preference mode from different attribute dimension combination can be mined from holographic semantic track big data, the association interaction between mode is quantitatively described, and the fine portrait of user activity footprint preference mode is realized at group and individual level.

Description

Technical Field

[0001] This invention relates to the Internet field, and more specifically to a method for mining preference patterns. Background Technology

[0002] Against the backdrop of the rapid development and widespread adoption of mobile internet and smart devices, mobile apps, wearable devices (such as fitness trackers and smartwatches), and social networking platforms have recorded massive amounts of user-shared activity footprint data (such as outdoor physical activities like jogging, cycling, and walking). Simultaneously, ubiquitous sensing devices also record relevant contextual data (such as weather, temperature, and individual user attributes) generated during these activity footprints. Compared to traditional survey data, this associated footprint and contextual data offers advantages such as low cost, large scale, and high spatiotemporal resolution. By linking and integrating this footprint and contextual data (multi-attribute semantic data), a digital twin of the user's activity footprint can be reconstructed, i.e., a holographic semantic trajectory.

[0003] Holographic semantic trajectories not only contain the spatiotemporal patterns of user activities, but also include multi-faceted preference patterns, such as time + location, time + weather, and time + weather + location. These multi-faceted preference patterns are composed of base patterns from different attribute dimensions, which more finely and comprehensively represent the user's preference characteristics and interest habits, and can effectively support typical applications such as personalized recommendations, activity behavior prediction, and urban planning and design.

[0004] However, most existing trajectory big data mining is based only on footprint data and focuses on spatiotemporal pattern mining, without further extracting multi-faceted preference pattern knowledge based on contextual data; this undoubtedly reduces the diversity and semantics of analyzing activity preference patterns.

[0005] Multifaceted preference patterns are composed of base patterns derived from multiple attribute dimensions of data, representing users' activity preference characteristics under different contextual attribute combinations, and can better characterize user preferences. However, mining these preference patterns from large datasets of high-dimensional heterogeneous holographic semantic trajectories using existing technologies still faces many challenges. The main reason is that holographic semantic trajectory data contains multi-source, heterogeneous attribute data, making it difficult for existing methods to model efficiently. In addition, multifaceted preference patterns come from attribute data of different dimensions, and there are complex implicit dependencies between base patterns. Existing methods, such as nonnegative matrix factorization (NMF), singular value decomposition (SVD), and principal component analysis (PCA), are difficult to quantitatively characterize the correlations and interactions between base patterns of different dimensions. Although tensor decomposition methods can effectively model high-dimensional semantic trajectory data and capture the interaction relationships between patterns of different attribute dimensions, it is difficult to mine multifaceted preference patterns at both the individual and group levels. In particular, the imbalance and sparsity of trajectory data distribution remain difficult to model and handle. Summary of the Invention

[0006] To address the above technical problems, this invention establishes a two-layer tensor decomposition method to extract multi-faceted preference patterns from holographic semantic trajectory big data. The main technical solutions adopted are as follows:

[0007] A method for mining multi-faceted preference patterns in activity footprints based on holographic semantic trajectory big data includes the following steps:

[0008] Step 1: Associate and aggregate the GPS trajectory of the user's original activity footprint with multi-dimensional attribute semantic data according to spatiotemporal location to generate a group holographic semantic trajectory dataset and an individual holographic semantic trajectory dataset respectively;

[0009] Step 2: Grid-encode the GPS trajectory and discretize the multi-dimensional attribute semantic data; construct the group activity tensor and the individual activity tensor respectively.

[0010] Step 3: Establish a two-layer Tucker tensor decomposition method to decompose the group activity tensor to extract the group user activity base pattern; then decompose the individual activity tensor based on the group user activity base pattern to extract the individual user activity multi-faceted preference pattern.

[0011] The multi-dimensional attribute semantic data includes weather data and temperature data. Attached Figure Description

[0012] Figure 1 This is a flowchart of the present invention;

[0013] Figure 2 A schematic diagram illustrating one way to represent holographic semantic trajectory data;

[0014] Figure 3 This is a schematic diagram illustrating the construction of the activity tensor in Example 2;

[0015] Figure 4 This is a schematic diagram illustrating the principle of the two-layer Tucker decomposition method in Example 2 for mining multi-faceted preference patterns in activities.

[0016] Figure 5 This is an example diagram of the base pattern results of group user activities obtained in Example 2;

[0017] Figure 6 This is an example diagram showing the results of the multi-faceted preference patterns of group user activities obtained in Example 2;

[0018] Figure 7 This is an experimental diagram illustrating the determination of the decomposition modulus in Embodiment 2 of the present invention;

[0019] Figure 8This is an example diagram showing the results of the multi-faceted preference pattern of individual user activities obtained in Embodiment 2 of the present invention. Detailed Implementation

[0020] The present invention will be further described below with reference to the embodiments and accompanying drawings.

[0021] Example 1:

[0022] A method for mining multi-faceted preference patterns in activity footprints based on holographic semantic trajectory big data includes the following steps:

[0023] Step 1: Associate and aggregate the GPS trajectory of the user's original activity footprint with multi-dimensional attribute semantic data according to spatiotemporal location to generate a group holographic semantic trajectory dataset and an individual holographic semantic trajectory dataset; the multi-dimensional attribute semantic data includes weather data and temperature data;

[0024] Specifically, step 1 further includes the following sub-steps:

[0025] Step 1.1: Referring to the GPS trajectory and multi-dimensional attribute semantic data, select and determine the attribute dimension Aspects = {A1, A2, ..., A...} of the holographic semantic trajectory dataset. i ,…,A N} and preprocess the data for each attribute dimension;

[0026] The attribute dimensions include time dimension, space dimension, weather dimension, and temperature dimension;

[0027] A i Let i be the i-th attribute dimension;

[0028] i = 1, 2, 3, ..., N; N is the number of selected attribute dimensions;

[0029] Step 1.2: For any of the GPS trajectories, the trajectory line T = {p1, p2, ..., p...} τ ,…,p Γ Based on the spatiotemporal correlation of the GPS trajectory, multi-dimensional attribute semantic data are integrated to generate a holographic semantic trajectory dataset p`. τ ;

[0030] p τ Let τ be the τth trajectory point on trajectory line T;

[0031] T represents the number of trajectory points on trajectory line T;

[0032]

[0033] Let be the j-th feature value of the i-th attribute dimension;

[0034] j = 1, 2, 3, ..., I i ;I i The number of feature values ​​for the i-th attribute dimension;

[0035] Step 1.3: Divide the trajectory dataset; the holographic semantic trajectory dataset corresponding to all users is the group holographic semantic trajectory dataset. The holographic semantic trajectory dataset corresponding to an individual user is the individual holographic semantic trajectory dataset.

[0036] For the m-th trajectory of all users;

[0037] m = 1, 2, 3, ..., K; K is the total number of trajectories for all users;

[0038] This refers to the nth trajectory of individual user u.

[0039] n = 1, 2, 3, ..., U, where U is the total number of trajectories of individual user u.

[0040] Step 2: Grid-encode the GPS trajectory and discretize the multi-dimensional attribute semantic data;

[0041] Construct the group activity tensor and the individual activity tensor respectively;

[0042] Step 2 includes the following sub-steps:

[0043] Step 2.1: Process the holographic semantic trajectory dataset p` τ The spatiotemporal location attributes in the dataset are gridded for encoding the holographic semantic trajectory dataset p`. τ The multi-dimensional attribute semantic data is discretized and encoded to obtain the encoded i-th dimension attribute feature set.

[0044] This represents the j-th feature value of the i-th attribute dimension after encoding.

[0045] I i The number of feature values ​​for the i-th attribute dimension;

[0046] Step 2.2: Aggregate and statistically analyze the trajectory activity traffic into a spatial grid, and construct a group activity tensor based on the encoding results of multi-dimensional attribute data associated with spatiotemporal attributes. The element value represents the number of group trajectories under all combinations of specific attribute values ​​across all dimensions;

[0047] The elements of a tensor all belong to the real number field;

[0048] Step 2.3: For the individual holographic semantic trajectory dataset Traj u Based on spatiotemporal attributes and grid coding, attribute information is aggregated to construct individual activity tensors. The element value represents the number of individual trajectories under all combinations of specific attribute values ​​across all dimensions;

[0049] All elements representing a tensor belong to the real number field.

[0050] Step 3: Establish a two-layer Tucker tensor decomposition method to decompose the group activity tensor to extract the group user activity base pattern;

[0051] The individual activity tensor is then decomposed based on the group user activity base pattern to extract multi-faceted preference patterns of individual user activities;

[0052] Step 3 includes the following sub-steps:

[0053] Step 3.1: Establish a two-level Tucker tensor decomposition method, using the first-level tensor decomposition to decompose the group activity tensor. It is decomposed into a modal product of a population core tensor and factor matrices of each dimension;

[0054] Among them, the factor matrix records the principal components of the data distribution in each dimension, namely the basic pattern of group user activities (or: activity preference basic pattern); the group core tensor captures the correlation and interaction between different basic patterns in each dimension, namely the multi-faceted preference pattern of group user activities (or: multi-faceted activity preference pattern).

[0055] The Tucker decomposition formula is as follows:

[0056]

[0057] For the core tensor of the group,

[0058] M i Represents the decomposition modulus of the i-th attribute dimension;

[0059] × i Let be the modal product of the tensor and the i-th dimension of the matrix;

[0060] A (i) For the factor matrix,

[0061] To decompose and reconstruct the tensor;

[0062] The group core tensor And factor matrix A (i)The objective function L1 can be solved by optimizing the following minimization method using alternating least squares or gradient descent.

[0063]

[0064] ‖·‖ is the Frobenius norm;

[0065] λ1 is the population core tensor The solution process includes penalty parameters for L2 norm regularization;

[0066] Step 3.2: Use the root mean square error (RMSE) to measure the original tensor. and decomposition and reconstruction tensor The difference is used as an evaluation index for the decomposition quality in order to obtain the optimal decomposition modulus;

[0067]

[0068] Step 3.3: Decompose the base patterns obtained from the first-level tensor decomposition (i.e., factor matrix A) (i) As the factor matrix for individual tensor decomposition, the individual tensors are decomposed using a second-level tensor decomposition model.

[0069] Solving for the individual core tensor This is used as a multifaceted preference pattern for individual user activities.

[0070] Individual core tensor The objective function L2 can be solved by optimizing the following minimization:

[0071]

[0072] λ2 is the individual core tensor The solution process includes penalty parameters for L2 norm regularization.

[0073] Example 2:

[0074] like Figure 1 As shown, a method for mining multi-faceted preference patterns in jogging activity footprints based on holographic semantic trajectory big data includes the following steps:

[0075] Step 1, as follows Figure 2 As shown, the original GPS trajectory of a jogging activity is linked and integrated with multi-dimensional attribute semantic data to construct a holographic semantic trajectory dataset. Specifically:

[0076] Step 1.1: Select six dimensions as the attribute semantic data of the holographic semantic trajectory: activity time H (hours), days D, space (grid) S, jogging duration L, temperature T, and weather W. Then Aspects = {H, D, S, L, T, W}; correspondingly, A1 = H, A2 = D, A3 = S, A4 = L, A5 = T, A6 = W, N = 6.

[0077] Step 1.2: For any jogging trajectory T = {p1, p2, ..., p...} τ ,…,p Γ}, the original jogging GPS track points p τ =(x τ ,y τ ,t τ Based on spatiotemporal information, multi-dimensional attribute semantic data are correlated and integrated to generate holographic semantic trajectory data p`. τ ;

[0078] x τ Let x and y coordinates be the coordinates of the τth trajectory point. τ Let y be the y-coordinate of the τth trajectory point, and t be the t-coordinate of the trajectory τ Let τ be the time corresponding to the τth trajectory point;

[0079] Let be the j-th feature value of the i-th attribute dimension;

[0080] j = 1, 2, 3, ..., I i ;I i Let be the number of feature values ​​for the i-th attribute dimension. This value is usually chosen manually. For example, when i and j take certain values:

[0081] p` τ =(18:30, Saturday, grid 1440, 35 minutes, 25℃, sunny);

[0082] Step 1.3: Divide the holographic semantic trajectory dataset into a group trajectory set Traj. c and individual trajectory set Traj u ;

[0083]

[0084] For the m-th trajectory of all users;

[0085] m = 1, 2, 3, ..., K; K is the total number of trajectories for all users;

[0086] This refers to the nth trajectory of individual user u.

[0087] n = 1, 2, 3, ..., U, where U is the total number of trajectories of individual user u.

[0088] Step 2, as follows Figure 3 As shown, the six-dimensional attribute semantic data is encoded, specifically by gridding the GPS trajectory and discretizing the multi-dimensional attribute semantic data, and then encoding the data according to the divided group trajectory set Traj. c and individual trajectory set Traj u Constructing the group activity tensor and individual activity tensor Specifically:

[0089] Step 2.1: As shown in Table 1, for the holographic semantic trajectory dataset p` τ The attribute data of the six dimensions are encoded to construct a tensor, resulting in the attribute feature set of the i-th dimension, denoted as Attr. i ;

[0090] in For the j-th feature value of the i-th attribute dimension after encoding, I i The number of feature values ​​for the i-th attribute dimension;

[0091] Table 1. Attribute Coding Table

[0092]

[0093] Step 2.2: Statistically analyze the group trajectory set Traj c The number of jogging trajectories under different attribute feature values ​​in all dimensions is used to construct a group jogging tensor. The element values ​​represent the group jogging flow under specific time, space, activity duration, temperature, and weather conditions;

[0094] Step 2.3: Statistically analyze the trajectory sets Traj for each body. u The number of trajectories under different attribute feature values ​​in all dimensions is used to construct the jogging tensor for each individual. Individual jogging tensor Tensor shape and group jogging tensor Consistent with the shape of the tensor The element value represents the individual jogging flow under specific time, space, activity duration, temperature, and weather conditions;

[0095] All elements representing a tensor belong to the real number field.

[0096] Step 3: Establish a two-level Tucker tensor decomposition method. Use this method to decompose the group jogging tensor to extract the group user activity base patterns. Then, based on the group user activity base patterns, decompose the individual jogging tensor to extract individual user activity preference patterns. Specifically, such as... Figure 4 As shown:

[0097] Step 3.1: Use the first-level Tucker tensor decomposition method to decompose the group jogging tensor. It is decomposed into the modal product of a population core tensor and the factor matrices (basis modes) of each dimension. The Tucker decomposition formula is as follows:

[0098]

[0099] In the formula:

[0100] For the core tensor of the group;

[0101] × i Let be the modal product of the tensor and the i-th dimension of the matrix;

[0102] Let i be a factor matrix, where i ∈ Aspects;

[0103] To decompose and reconstruct the tensor;

[0104] Group core tensor And factor matrix A (i) The objective function L1 can be solved by minimizing the following objective function using methods such as alternating least squares or gradient descent. L1 is calculated as follows:

[0105]

[0106] In the formula:

[0107] ‖·‖ is the Frobenius norm;

[0108] λ1 is the penalty parameter for L2 norm regularization in this process;

[0109] Factor matrix A (i) (Example of a user activity base pattern or activity preference base pattern) Figure 5 As shown, an example of a group core tensor (a pattern of multiple user activity preferences or multiple activity preferences) is as follows: Figure 6 As shown;

[0110] Step 3.2: Define an approximate range for the decomposition modulus (number of modes in each dimension) for each dimension, and perform a series of tensor decompositions as described in Step 3.1; then use the root mean square error (RMSE) to measure the original tensor. and decomposition and reconstruction tensor The difference in RMSE is used as an evaluation index for decomposition quality (the smaller the RMSE, the higher the decomposition quality). RMSE is calculated as follows:

[0111]

[0112] like Figure 7 As shown, the decomposition modulus for each dimension is determined as M. H =3,M D =2,M S =5,M L =4,M T =4 and M W =3;

[0113] Step 3.3: Decompose the individual jogging tensor using the second-level Tucker tensor decomposition method. Specifically, the factor matrix of the decomposition result is inherited from the decomposition result (base mode) A of the factor matrix of the group jogging tensor in step 3.1. (i) Only the individual core tensor needs to be solved. The decomposition process is as follows:

[0114]

[0115] In the formula For individual core tensors, The objective function L2 can be solved by minimizing it as follows, and L2 is calculated as follows:

[0116]

[0117] In the formula, λ2 is the penalty parameter for L2 norm regularization of the process;

[0118] Individual core tensor As a multi-faceted preference pattern for individual user activities, for example... Figure 8 As shown.

[0119] Beneficial Effects: This invention designs a method for mining multi-faceted preference patterns in activity footprints by modeling holographic semantic trajectory big data. This method can mine multi-faceted preference patterns from different attribute dimensions in holographic semantic trajectory big data, quantitatively characterize the correlation and interaction between patterns, and achieve a refined profile of user activity footprint preference patterns at both the group and individual levels.

[0120] Finally, it should be noted that the above description is merely a preferred embodiment of the present invention. Those skilled in the art, under the guidance of the present invention, can make various similar representations without departing from the spirit and claims of the present invention, and such modifications all fall within the protection scope of the present invention.

Claims

1. A method for mining multi-faceted preference patterns in activity footprints based on holographic semantic trajectory big data, characterized in that, Includes the following steps: Step 1: Associate and aggregate the GPS trajectory of the user's original activity footprint with multi-dimensional attribute semantic data according to spatiotemporal location to generate a group holographic semantic trajectory dataset and an individual holographic semantic trajectory dataset respectively; Step 2: Grid-encode the GPS trajectory and discretize the multi-dimensional attribute semantic data; Construct the group activity tensor and the individual activity tensor respectively; Step 3: Establish a two-layer Tucker tensor decomposition method to decompose the group activity tensor to extract the group user activity base pattern; The individual activity tensor is decomposed based on the base pattern of group user activities to extract multi-faceted preference patterns of individual user activities; Step 1 includes the following sub-steps: Step 1.1: Referring to the GPS trajectory and multi-dimensional attribute semantic data, select and determine the attribute dimensions of the holographic semantic trajectory dataset. And preprocess the data for each attribute dimension; The attribute dimensions include time dimension, space dimension, weather dimension, and temperature dimension; For the first i Each attribute dimension; ； The number of selected attribute dimensions; Step 1.2: For any of the GPS trajectory lines... Based on the spatiotemporal correlation of the GPS trajectory, multi-dimensional attribute semantic data are integrated to generate a holographic semantic trajectory dataset. ; trajectory line The first A trajectory point; trajectory line The number of trajectory points on the track; ； For the first i The first attribute dimension j One eigenvalue; j =1,2,3,…, I i ; I i For the first i The number of feature values ​​in each attribute dimension; Step 1.3: Divide the trajectory dataset; the holographic semantic trajectory dataset corresponding to all users is the group holographic semantic trajectory dataset. The holographic semantic trajectory dataset corresponding to an individual user is the individual holographic semantic trajectory dataset. ; For all users m A trajectory; ; The total number of trajectories for all users; For individual users u The n A trajectory; , For individual users u The total number of trajectories; Step 2 includes the following sub-steps: Step 2.1: Process the holographic semantic trajectory dataset The spatiotemporal location attributes in the dataset are gridded for encoding the holographic semantic trajectory dataset. The multi-dimensional attribute semantic data in the data is discretized and encoded; the encoded first... i Dimensional attribute feature set }; For the encoded first The first attribute dimension One eigenvalue; I i For the first i The number of feature values ​​in each attribute dimension; Step 2.2: Aggregate and statistically analyze the trajectory activity traffic into a spatial grid, and construct a group activity tensor based on the encoding results of multi-dimensional attribute data associated with spatiotemporal attributes. ; The elements of a tensor all belong to the real number field; Step 2.3: For the individual holographic semantic trajectory dataset Based on spatiotemporal attributes and grid coding, attribute information is aggregated to construct individual activity tensors. ; The elements of a tensor all belong to the real number field; Step 3 includes the following sub-steps: Step 3.1: Establish a two-level Tucker tensor decomposition method, using the first-level tensor decomposition to decompose the group activity tensor. It can be decomposed into the modal product of a population core tensor and factor matrices of each dimension. The Tucker decomposition formula is as follows: ； For the core tensor of the group, ; M i Indicates the first i Decomposition modulus of each attribute dimension; For tensors and matrices Modal product of dimensions; For the factor matrix, , ; To decompose and reconstruct the tensor; Step 3.2: Measure the original tensor using the root mean square error. and decomposition and reconstruction tensor The difference is used as an evaluation index for the decomposition quality in order to obtain the optimal decomposition modulus; Step 3.3: Decompose the factor matrix obtained from the first-level tensor decomposition. As the factor matrix for individual tensor decomposition, the individual tensor is decomposed using a second-level tensor decomposition model. ; ； Solving for the individual core tensor This is used as a multi-faceted preference pattern for individual user activities; In step 3.1, the objective function is minimized using either alternating least squares or gradient descent. Solving the population core tensor sum factor matrix ; ； It is the Frobenius norm; It is the core tensor of the group The solution process The penalty parameter for norm regularization; In step 3.2, the root mean square error RMSE Calculate using the following formula: ； In step 3.3, the objective function is minimized by optimization. Solving for the individual core tensor ; ； It is the individual core tensor The solution process The penalty parameter for norm regularization.

2. The method for mining multi-faceted preference patterns of activity footprints based on holographic semantic trajectory big data according to claim 1, characterized in that: The multi-dimensional attribute semantic data includes weather data and temperature data.

Images