Lightweight next position prediction method and system based on zero-sample two-stage dynamic weighted sorting

Abstract

The invention discloses a light-weight next position prediction method and system based on zero-sample two-stage dynamic weighted sorting. The method comprises the following steps: firstly, constructing a three-dimensional evaluation matrix, and then respectively adopting an improved entropy weight-TOPSIS algorithm and a CRITIC-TOPSIS algorithm to obtain a first-stage position identification candidate set and a second-stage position identification candidate set according to the three-dimensional evaluation matrix, and finally, acquiring a next position candidate set of the current sign-in position of the user according to the first-stage position identifier candidate set and the second-stage position identifier candidate set. The method does not need model training, is suitable for data scarce and cross-scene environments, has the characteristics of high real-time performance and lightweight calculation, is suitable for online prediction application, has good interpretability and expandability, and is superior to a traditional model in the aspects of accuracy and stability.

Description

Technical Field

[0001] This invention belongs to the field of human mobility behavior modeling and location intelligent recommendation technology, specifically involving a lightweight next-position prediction method and system with zero-sample two-stage dynamic weighted ranking. Background Technology

[0002] With the development of mobile internet and location technology, location-based services (LBS) have been widely applied in fields such as travel recommendations, urban transportation, business layout, and social behavior research. Among these, next location prediction is one of the core tasks of LBS, aiming to predict a user's next likely location based on their historical check-in data and contextual characteristics. High-precision next location prediction results can not only improve the intelligence level of travel recommendations but also provide important support for infectious disease transmission prediction, urban traffic management, energy scheduling, and business decision-making.

[0003] Currently, mainstream research in academia and industry focuses on sequence modeling methods based on deep learning, such as recurrent neural networks (RNN) and graph neural networks (GNN). These methods typically capture potential behavioral patterns by learning users' spatiotemporal trajectory patterns, possessing strong nonlinear fitting capabilities and prediction accuracy. However, with the expansion of data scale and the diversification of application scenarios, existing methods still have the following prominent problems: (1) High dependence on model training, making it difficult to transfer and generalize. Deep learning models require a large amount of labeled data and computing resources during training. Data differences between different cities or platforms can lead to a significant decrease in model transfer performance, making it difficult to deploy quickly in new scenarios. (2) Insufficient interpretability and controllability. Due to the complex structure and high parameter dimension of deep neural networks, they are essentially black box models, and their prediction results are difficult to fundamentally explain, which is not conducive to scenarios such as urban travel analysis that require interpretable decision-making basis. (3) Individual differences and behavioral sparsity are not fully modeled. User check-in data often has characteristics such as strong temporal fluctuations and sparse spatial distribution. Existing models usually start from the overall pattern and have weak identification capabilities for users with few samples or unpopular locations. (4) Insufficient real-time performance and high resource consumption. Large neural networks have long inference times, making them unsuitable for real-time online prediction or edge computing environments; simultaneously, model updates are costly and difficult to support dynamically changing travel behavior patterns. In summary, most existing next-location prediction technologies rely on deep learning models, making it difficult to balance real-time performance, interpretability, and cross-scenario adaptability. Against this backdrop, there is an urgent need for a lightweight method that requires no training, can adaptively evaluate the importance of candidate locations based on statistical features, and achieve dynamic ranking, thereby ensuring computational efficiency while enhancing prediction accuracy and the algorithm's sensitivity to time fluctuations and individual user differences. Summary of the Invention

[0004] To address the problems existing in the background technology, this invention provides a lightweight next-position prediction method and system based on zero-sample, two-stage dynamic weighted ranking. This solves the technical problem that most existing next-position prediction technologies rely on deep learning models, making it difficult to balance real-time performance, interpretability, and cross-scenario adaptability. Addressing the issues of strong training dependence, insufficient modeling of individual differences, static feature weighting, and poor real-time performance in existing methods, this invention proposes a lightweight prediction framework based on the fusion of statistical features and dynamic weights. This framework enables adaptive inference of user travel behavior without requiring deep learning model training. This invention improves prediction accuracy, interpretability, and cross-scenario adaptability while maintaining computational efficiency.

[0005] The technical solution adopted in this invention is: I. A lightweight next position prediction method based on zero-sample two-stage dynamic weighted ranking: S1. Obtain check-in samples of user movement behavior, serialize all check-in samples in chronological order to obtain historical check-in sequences, obtain historical check-in location subsequences, historical check-in activity subsequences and next location candidate sets based on historical check-in sequence, construct location access frequency features, activity subnetwork access frequency features and location subnetwork access frequency features based on historical check-in location subsequences, historical check-in activity subsequences and next location candidate sets respectively, and construct a three-dimensional evaluation matrix based on the three types of features.

[0006] S2. Based on the three-dimensional evaluation matrix, the improved entropy weight-TOPSIS algorithm is used to sort the candidate set of the next position in the first stage, resulting in the first stage candidate set of position identifiers; based on the three-dimensional evaluation matrix, the CRITIC-TOPSIS algorithm is used to sort the candidate set of the next position in the second stage, resulting in the second stage candidate set of position identifiers.

[0007] S3. Select the first few location identifiers from the first stage location identifier candidate set, remove the selected location identifiers from the second stage location identifier candidate set to obtain the second stage location identifier candidate set after removal, and then select the first few location identifiers from the second stage location identifier candidate set to supplement it. Thus, all the location identifiers selected in succession constitute the next location candidate set of the user's current check-in location.

[0008] Step S1 specifically involves: S11. Obtain the current check-in sample of the user's movement behavior and several consecutive historical check-in samples preceding the current check-in sample. Serialize all check-in samples in chronological order to obtain the historical check-in sequence. Each check-in sample in the historical check-in sequence includes a location identifier, an activity type identifier, and a timestamp.

[0009] S12. The historical check-in position subsequence is formed by combining the position identifiers of the current check-in sample and the h check-in samples before the current check-in sample in the historical check-in sequence.

[0010] S13. The current check-in sample and all activity identifiers of the h check-in samples before the current check-in sample in the historical check-in sequence constitute the historical check-in activity subsequence.

[0011] S14. The set of position identifiers obtained by deduplicating the same position identifiers in the historical check-in position subsequence is used as the next position candidate set of the current check-in sample; each position identifier in the next position candidate set is a candidate position.

[0012] S15. Construct location access frequency features, location subnetwork access frequency features, and activity subnetwork access frequency features based on the historical check-in location subsequence, historical check-in activity subsequence, and next location candidate set, respectively. Construct a three-dimensional evaluation matrix based on the three types of features.

[0013] Step S15 specifically involves: S151. The sequence formed by the frequency of each location identifier in the next candidate location set appearing in the historical check-in location subsequence is used as the location access frequency feature.

[0014] S152. The sequence consisting of the total frequency of each candidate position appearing in the preset window after all marked positions are counted in the historical check-in position subsequence is used as the access frequency feature of the position sub-network.

[0015] S153. The sequence consisting of the total frequency of the activity identifier corresponding to each candidate position in the preset window after all marked activities in the historical check-in activity subsequence is used as the access frequency feature of the activity sub-network.

[0016] S154. The location access frequency features, activity subnetwork access frequency features and location subnetwork access frequency features are sequentially concatenated to form an initial three-dimensional evaluation matrix. The range normalization of each column of the initial three-dimensional evaluation matrix is ​​performed to obtain the three-dimensional evaluation matrix.

[0017] The improved entropy weight-TOPSIS algorithm is specifically as follows: D1. Calculate the positive ideal solution and the negative ideal solution based on the three-dimensional evaluation matrix.

[0018] D2. Construct a time stability correction factor based on the set of timestamps of each location identifier accessed in the next location candidate set; construct a user sparsity correction factor based on the set of users accessed for each location identifier in the next location candidate set.

[0019] D3. Calculate the weight of each feature in the first stage of the three-dimensional evaluation matrix based on the time stability correction factor, the user sparsity correction factor, and the information entropy of each feature in the three-dimensional evaluation matrix.

[0020] D4. Based on the positive ideal solution, negative ideal solution, and the weight of each column of features in the three-dimensional evaluation matrix in the first stage, the TOPSIS method is used to sort the candidate set of the next position in the first stage, thus obtaining the candidate set of position identifiers after the first stage sorting, i.e., the candidate set of position identifiers in the first stage.

[0021] The time stability correction factor and the user sparsity correction factor are set according to the following formulas: CV time =(1 / K)∑ i=1 K ((1 / p)∑ z=1 p (t i,z -t i ') 2 ) 1 / 2 CV user =(1 / K)∑ i=1 K (1 / Q i ) Among them, CV time and CV user These represent the time stability correction factor and the user sparsity correction factor, respectively; i is the index; K represents the number of location identifiers in the next location candidate set; p represents the total number of timestamps in the timestamp set of the i-th location identifier in the next location candidate set that have been accessed; z is the index; t i,z This represents the z-th timestamp of the i-th location identifier in the next location candidate set being visited; t i ' represents the average of all timestamps in the set of timestamps for the i-th location identifier in the next location candidate set that was visited; Q i This represents the number of users in the user set whose i-th position identifier in the next position candidate set is visited.

[0022] The weights of each feature column in the three-dimensional evaluation matrix for the first stage are set according to the following formula: w e,t,h,j u =(1-e t,h,j u )CV j ) / (∑ y=1 3 ((1-e t,h,y u )CV y j=1,2,3 CV1=CVtime ;CV2=1;CV3=CV user e t,h,j u =(1 / lnK)∑ i=1 K (z t,h,ij u ·lnz t,h,ij u ) z t,h,ij u =(m t,h,ij u ) / (∑ i=1 K m t,h,ij u ) Where j and y are both indices; w e,t,h,j u e represents the weight of the first stage of the j-th column feature; t,h,j u and e t,h,y u These represent the information entropy of the features in the j-th and y-th columns of the three-dimensional evaluation matrix, respectively; CV j and CV y CV1, CV2, and CV3 represent the correction values ​​corresponding to the features in columns j and y, respectively; CV1, CV2, and CV3 represent the correction values ​​in columns 1, 2, and 3 of the three-dimensional evaluation matrix, respectively; CV time and CV user These represent the time stability correction factor and the user sparsity correction factor, respectively; i is the index; K represents the number of location identifiers in the next location candidate set; z t,h,ij u Let m be the weight value of the feature in column j for the i-th position identifier; t,h,ij u This represents the eigenvalue of the i-th position in the j-th column of the three-dimensional evaluation matrix.

[0023] The CRITIC-TOPSIS algorithm is specifically as follows: F1. Calculate the positive and negative ideal solutions based on the three-dimensional evaluation matrix.

[0024] F2. Obtain the contrast intensity and conflict intensity of each feature column based on all eigenvalues ​​in the three-dimensional evaluation matrix.

[0025] F3. Obtain the weight of each feature in the second stage of the three-dimensional evaluation matrix based on the contrast intensity and conflict intensity of each feature.

[0026] F4. Based on the weights of the features in the second stage of each column of the three-dimensional evaluation matrix, the next position candidate set is sorted in the second stage using the TOPSIS method to obtain the position identifier candidate set after the second stage sorting, which is the second stage position identifier candidate set.

[0027] The contrast intensity and conflict intensity of each column of features are set according to the following formula: σ t,h,j u =((1 / K)∑ i=1 K (m t,h,ij u -m t,h,·j u’ ) 2 ) 1 / 2 I t,h,j u =∑ j’=1 3 (1-|r jj’ |), j≠j' r jj’ =(∑ i=1 K ((m t,h,ij u -m t,h,·j u’ (m) t,h,ij’ u -m t,h,·j’ u’ ))) / ((∑ i=1 K (m t,ij u -m t,h,·j u’ ) 2 ) 1 / 2 ·(∑ i=1 K (m t,ij’ u -m t,h,·j’ u’ ) 2 ) 1 / 2 ) Where i, j, and j' all represent indices; σ t,h,j u Indicates the contrast intensity of the feature in column j; I t,h,j u The value r represents the conflict intensity of the feature in column j; jj’ The feature correlation value between the features in column j and column j' represents the value of the feature correlation between them; K represents the number of location identifiers in the next location candidate set; m t,h,ij u and m t,h,ij’u These represent the feature values ​​of the i-th position identifier in the j-th and j'-th columns, respectively; m t,h,·j u’ and m t,h,·j’ u’ Let represent the average eigenvalues ​​obtained by summing and averaging all eigenvalues ​​corresponding to the j-th column and the j'-th column, respectively.

[0028] The weights of each feature column in the second stage of the three-dimensional evaluation matrix are set according to the following formula: w c,t,h,j u =(σ t,h,j u ·I t,h,j u ) / (∑ j=1 3 (σ t,h,j u ·I t,h,j u j=1,2,3 Where j represents the index; w c,t,h,j u σ represents the weight of the second stage of the j-th column feature; t,h,j u Indicates the contrast intensity of the feature in column j; I t,h,j u This represents the conflict intensity of the feature in column j.

[0029] II. A lightweight next-position prediction system with zero-sample two-stage dynamic weighted ranking: The three-dimensional evaluation matrix construction module obtains check-in samples of user movement behavior, serializes all check-in samples in chronological order to obtain historical check-in sequences, obtains historical check-in location subsequences, historical check-in activity subsequences, and next location candidate sets based on the historical check-in sequence, and constructs location access frequency features, location subnetwork access frequency features, and activity subnetwork access frequency features based on the historical check-in location subsequences, historical check-in activity subsequences, and next location candidate sets, respectively, and constructs a three-dimensional evaluation matrix based on the three types of features.

[0030] The first-stage location identifier candidate set acquisition module uses an improved entropy weight-TOPSIS algorithm based on the three-dimensional evaluation matrix to sort the next location candidate set in the first stage, resulting in the first-stage sorted location identifier candidate set, which is the first-stage location identifier candidate set.

[0031] The second-stage location identifier candidate set acquisition module uses the CRITIC-TOPSIS algorithm to sort the next location candidate set in the second stage according to the three-dimensional evaluation matrix, and obtains the second-stage sorted location identifier candidate set, which is the second-stage location identifier candidate set.

[0032] The next position candidate set acquisition module selects the first few position identifiers from the first stage position identifier candidate set, removes the selected position identifiers from the second stage position identifier candidate set, and obtains the second stage position identifier candidate set after removal. Then, it selects the first few position identifiers from the second stage position identifier candidate set to supplement it. Thus, all the position identifiers selected in succession constitute the final next position candidate set.

[0033] The beneficial effects of this invention are: 1. No model training required, suitable for data-scarce and cross-scenario environments. This invention constructs a statistical feature matrix based on historical check-in sequences and uses a two-stage dynamic weighted sorting method to replace the traditional neural network training process, achieving a completely training-free prediction process. Therefore, this method can be directly applied to new user or new region data, avoiding the high cost of model training and parameter migration, and significantly improving the versatility and deployment efficiency of the method.

[0034] 2. A time stability correction factor and a user sparsity correction factor are introduced to enhance the ability to characterize individual differences. These factors reflect the temporal regularity of user access behavior and the location access sparsity, respectively. This correction mechanism can dynamically adjust the contribution weights of different features, making the prediction results more reflective of user behavior rhythms and individual preferences, thereby effectively improving the problem of inaccurate predictions for unpopular locations and small sample sizes using traditional methods.

[0035] 3. An improved two-stage fusion mechanism of entropy-weighted TOPSIS and CRITIC-TOPSIS is proposed to enhance ranking accuracy and robustness. The first stage, the improved entropy-weighted TOPSIS model, is used to characterize the importance distribution of features, and adaptive weight allocation is achieved by combining a time stability correction factor and a user sparsity correction factor. The second stage, the CRITIC-TOPSIS model, compensates for the redundant influence between features by calculating the feature contrast strength and conflict. After fusing the results of the two stages, both feature discriminative power and information independence can be taken into account, making the comprehensive evaluation of candidate positions more comprehensive and stable.

[0036] 4. It possesses high real-time performance and lightweight computational characteristics, making it suitable for online prediction applications. The overall complexity of the algorithm in this invention is linear, and the time taken for a single prediction can be controlled within milliseconds, far lower than algorithms based on deep learning models. It can run directly on mobile devices, edge devices, and real-time recommendation systems. Compared to deep learning models, this method can quickly output results without requiring dedicated GPU memory or a parallel acceleration environment, significantly reducing the computational burden on the system.

[0037] 5. Excellent interpretability and scalability. Each candidate location output by this invention comes with a feature weight index, which can intuitively explain the formation logic of the prediction results and facilitate subsequent user profile analysis and model auditing. In addition, this method can be further extended to introduce a geographic distance decay factor and a point of interest (POI) semantic similarity index, achieving multi-dimensional feature fusion without changing the algorithm structure, and providing a scalable statistical reasoning framework for the next location prediction.

[0038] 6. Experiments show that the proposed method outperforms traditional models in both prediction accuracy and stability. On a typical urban user historical check-in dataset, the method of this invention demonstrates greater advantages over advanced deep learning models in multiple metrics, while reducing computation time by more than 88%. These results indicate that the proposed method has significant advantages in balancing performance and efficiency, possessing good engineering applicability and potential for widespread adoption. Attached Figure Description

[0039] Figure 1 This is a flowchart of the method of the present invention.

[0040] Figure 2 This is a block diagram of the system of the present invention.

[0041] Figure 3 This is a bar chart comparing the time required for multiple algorithms to predict a single sample in an embodiment of the present invention.

[0042] Figure 4 This is a schematic diagram illustrating the scenario in which the algorithm is implemented in an embodiment of the present invention.

[0043] Figure 5 This is a bar chart comparing the effects of multiple algorithms in the embodiments of the present invention. Detailed Implementation

[0044] The present invention will now be described in more detail with reference to the accompanying drawings and embodiments. However, the present invention is not limited thereto. For those skilled in the art, several improvements and modifications can be made without departing from the principles of the present invention, and these improvements and modifications are also considered to be within the scope of protection of the present invention. Contents not described in detail in this specification are prior art known to those skilled in the art.

[0045] like Figure 1As shown, the lightweight next position prediction method in this embodiment includes the following steps: S11. Obtain the current check-in sample of the current user's current movement behavior and obtain several consecutive historical check-in samples before the current check-in sample from the database. Serialize all check-in samples in chronological order to obtain the historical check-in sequence.

[0046] Each check-in sample in the historical check-in sequence includes a location identifier, an activity type identifier, and a timestamp.

[0047] A day is divided into 24 time slices as discrete time labels; timestamps are mapped to discrete time labels; missing or abnormal records are removed or repaired; and the reliability of subsequent statistics is ensured.

[0048] The location identifier refers to a unique serial number label for each location; the activity type refers to the activity a user performs at this location, such as working or drinking coffee; the activity type identifier refers to a unique serial number label for each activity; the timestamp refers to a serial number representing the time of day, an integer ranging from 1 to 24, with each integer representing a slice of time in one hour.

[0049] Since subsequent statistical calculations require structured data, sequence preparation is necessary.

[0050] Serializing all user check-in samples in chronological order yields the following historical check-in sequence: S u ={(g i u ,a i u ,t i u )} i=1 n Where u is the user index, representing the u-th user; S u This represents the historical check-in sequence of the u-th user; i is the index, representing the ith check-in sample; n represents the length of the historical check-in sequence and also the total number of check-in samples; g i u This represents the location identifier of the i-th check-in sample of the u-th user; all location identifiers belong to the location identifier set G; a i u The activity type identifier for the i-th check-in sample of the u-th user; t i u Let represent the timestamp of the i-th check-in sample of the u-th user; therefore, each element (check-in sample) in the user's historical check-in sequence can be represented as a structured form of a triple consisting of a location identifier, an activity type identifier, and a timestamp.

[0051] S12. The position identifiers of the current check-in sample and the h pre-set check-in samples in the historical check-in sequence are used to form the historical check-in position subsequence.

[0052] The historical check-in location subsequence is set according to the following formula: S g,t-h:t u ={g t-h u ,g t-h+1 u ,…,g t u} Where t represents the index of the current check-in sample; h represents the number of check-in samples selected from the historical check-in sequence before the current check-in sample; S g,t-h:t u g represents the historical check-in position subsequence of the current check-in sample of the u-th user (the subsequence consisting of the current check-in sample and all position identifiers of the previous h check-in samples in the historical check-in sequence for the u-th user); t-h u and g t-h+1 u These represent the position identifiers of the first h and first h-1 check-in samples of the current check-in sample for the u-th user, respectively; g t u This indicates the location identifier of the current sign-in sample.

[0053] S13. The activity identifiers of the current check-in sample and the h pre-set check-in samples in the historical check-in sequence are used to form a historical check-in activity subsequence.

[0054] The historical check-in activity sub-sequence is set according to the following formula: S a,t-h:t u ={a t-h u ,a t-h+1 u ,…,a t u} Where t represents the index of the current check-in sample; h represents the number of check-in samples selected from the historical check-in sequence before the current check-in sample; S a,t-h:t u This represents the historical check-in activity subsequence of the current check-in sample for the u-th user (the subsequence consisting of the current check-in sample t and all activity identifiers of the previous h check-in samples in the historical check-in sequence for the u-th user); a t-h u and a t-h+1 uThese represent the activity identifiers of the first h and first h-1 check-in samples of the current check-in sample for the u-th user, respectively; a t u This indicates the activity identifier of the current sign-in sample.

[0055] S14. The set of location identifiers obtained by deduplicating the identical location identifiers in the historical check-in location subsequence is used as the next location candidate set for the current check-in sample. Each location identifier in the next location candidate set is a candidate location. The activity identifier corresponding to the candidate location is a candidate activity.

[0056] The resulting candidate set for the next position is set according to the following formula: C t,h u ={c t,h u(1) ,c t,h u(2) ,…,c t,h u(K)} Among them, C t,h u Let K be the candidate set of the next location for the current check-in sample of the u-th user; K represents the total number of location identifiers (number of elements) in the candidate set of the next location; c t,h u(1) c t,h u(2) and c t,h u(K) These represent the 1st, 2nd, and Kth position identifiers (elements) in the candidate set of the next position for the current check-in sample, respectively. Since there may be duplicate position identifiers in the historical check-in position subsequence, K... <h。

[0057] S15. Construct location access frequency features, location subnetwork access frequency features, and activity subnetwork access frequency features based on the historical check-in location subsequence, historical check-in activity subsequence, and next location candidate set, respectively. Construct a three-dimensional evaluation matrix based on the three types of features.

[0058] S151. Construct a location access frequency feature based on the historical check-in location subsequence and the next location candidate set; the location access frequency feature refers to the frequency of each location identifier in the next location candidate set appearing in the historical check-in location subsequence, taking into account that locations frequently visited by users are more likely to be the next locations visited by users, representing the user's access preference for that location, and is set according to the following formula: F t,h u ={f t,h u(1) ,f t,h u(2) ,…,f t,hu(K)} Among them, F t,h u This represents the location access frequency characteristic calculated based on the historical check-in location subsequence; f t,h u(1) f t,h u(2) and f t,h u(K) These represent the position identifiers c in the candidate set for the next position. t,h u(1) c t,h u(2) and c t,h u(K) The frequency of occurrence of the corresponding location in the historical check-in position subsequence.

[0059] S152. Using the current sign-in sample's location identifier as the marked location, calculate the total frequency of each candidate location appearing within a preset window after all marked locations are in the historical sign-in location subsequence.

[0060] Considering the higher-order characteristics of user access habits, this is used to characterize short-term travel associations; unlike the location access frequency feature F. t,h u The location subnetwork access frequency feature only considers the location identifier g of the current check-in sample. t u Subsequence S of historical check-in locations g,t-h:t u The total frequency of each candidate location appearing within a neighboring window of length q when the current location appears; these locations are highly correlated with the current location because the user's journey is likely to form subnetworks (e.g., home-restaurant-office). The next location candidate set C is calculated. t,h u The location subnetwork access frequency characteristic can be obtained by calculating the total frequency of each location identifier in the historical check-in location subsequence within a neighboring window of length q following the location identifier that is the same as the user's current location identifier. This frequency can be set according to the following formula: D t,h,gs u ={d t,h,gs u(1) ,d t,h,gs u(2) ,…,d t,h,gs u(K)} Among them, D t,h,gs u Indicates the location subnetwork access frequency characteristics; d t,h,gs u(1) d t,h,gs u(2) and d t,h,gsu(K) These represent the 1st, 2nd, and Kth elements in the location subnetwork access frequency feature, respectively; d t,h,gs u(1) d t,h,gs u(2) and d t,h,gs u(K) They also represent the candidate set C for the next position. t,h u The first, second, and Kth positions in the array represent the total frequency of each occurrence; gs represents the position.

[0061] S153. Using the activity identifier of the current check-in sample as the marked activity, calculate the total frequency of the activity identifier corresponding to each candidate position in the preset window after all marked activities in the historical check-in activity subsequence.

[0062] Considering the higher-order characteristics of user access habits, this is used to characterize short-term travel associations; unlike the location access frequency feature F. t,h u The activity subnetwork access frequency characteristics only consider the location identifier g of the current check-in sample. t u Corresponding activity identifier a t u Subsequence S of historical check-in locations g,t-h:t u The total frequency of activity identifiers corresponding to each candidate location within a neighboring window of length q when a certain activity occurs is calculated. These activities are highly correlated with the activity identifiers of the user's current check-in sample, as the user's itinerary is likely to form subnetworks (e.g., home-restaurant-office). The next location candidate set C is calculated. t,h u The activity sub-network access frequency characteristic can be obtained by calculating the total frequency of the activity identifier corresponding to each location identifier in the historical check-in activity sub-sequence within a neighboring window of length q following the activity identifier that is the same as the user's current activity identifier. This frequency can be set according to the following formula: D t,h,as u ={d t,h,as u(1) ,d t,h,as u(2) ,…,d t,h,as u(K)} Among them, D t,h,as u Indicates the frequency characteristics of active subnetwork access; d t,h,as u(1) d t,h,as u(2) and d t,h,as u(K)These represent the 1st, 2nd, and Kth elements in the active subnetwork access frequency characteristic, respectively; d t,h,as u(1) d t,h,as u(2) and d t,h,as u(K) They also represent the candidate set C for the next position. t,h u The total frequency of the activity identifier corresponding to the 1st, 2nd, and 3rd position identifiers in the table; as represents the activity.

[0063] It should be noted that, due to the candidate set C for the next position t,h u The activity identifier corresponding to each location identifier in the network may be duplicated, therefore the activity subnetwork access frequency characteristic D t,h,as u The elements in the index may be repeated, but this does not affect the representativeness of the index.

[0064] The feature lengths of the obtained location access frequency features, location subnetwork access frequency features, and activity subnetwork access frequency features are all k'.

[0065] S154. The location access frequency feature, the activity subnetwork access frequency feature and the location subnetwork access frequency feature are sequentially concatenated to form an initial three-dimensional evaluation matrix. The range normalization is performed on the feature of each column dimension (a total of three columns) of the initial three-dimensional evaluation matrix to make the features of different dimensions have a consistent measurement standard, thus obtaining the three-dimensional evaluation matrix.

[0066] The three-dimensional evaluation matrix is ​​set according to the following formula: M t,h u =[F t,h u ',D t,h,gs u ',D t,h,as u '] Among them, M t,h u This represents a three-dimensional evaluation matrix considering the h most recent location identifiers of user u, with a dimension of 3k'; F t,h u ' is the location access frequency feature after range normalization; D t,h,gs u ' is the location subnetwork access frequency characteristic after range normalization; D t,h,as u 'This represents the frequency characteristics of active subnetwork access after range normalization; S21. Based on the three-dimensional evaluation matrix, the improved entropy weight-TOPSIS algorithm is used to sort the next position candidate set in the first stage, and the position identifier candidate set after the first stage sorting is obtained, that is, the first stage position identifier candidate set.

[0067] D1. The positive ideal solution of the three-dimensional evaluation matrix calculated from the three-dimensional evaluation matrix is ​​P. t,h u =[p t,h u(1) ,p t,h u(2) ,p t,h u(3) The negative ideal solution is N. t,h u =[n t,h u(1) ,n t,h u(2) ,n t,h u(3) ],in: p t,h u(j) =max 1≤r≤K (m t,h,rj u ), j=1,2,3 n t,h u(j) =min 1≤r≤K (m t,h,rj u ), j=1,2,3 Where j is the index, representing the feature of the j-th column in the three-dimensional evaluation matrix, j=1,2,3; p t,h u(j) The positive ideal solution P is represented by t,h u The j-th column feature in n; t,h u(j) N represents the negative ideal solution t,h u The j-th column of the feature; r is the index, representing the r-th location identifier in the next location candidate set; K represents the number of location identifiers in the next location candidate set; m t,h,rj u This represents the feature value of the r-th position identifier in the j-th column (the feature value in the r-th row and j-th column).

[0068] D2. Construct a time stability correction factor based on the set of timestamps of each location identifier visited in the next location candidate set: Considering that in actual human travel behavior, the access behavior to different locations has significant differences in time stability, if a user has visited a location at many different time periods, it is more like a random behavior, and the probability of its recurrence in the future is difficult to predict. Conversely, if it only appears at fixed times, it indicates that it has periodicity, high regularity, and stronger predictability. If the i-th location identifier c in the next location candidate set... t,h u(i) The set of timestamps accessed (searched from 1-24) is T. i ={t i,1 ,t i,2 ,…,t i,p A time stability correction factor is introduced to the location access frequency feature: CV time =(1 / K)∑ i=1 K (1 / p∑ z=1 p (t i,z -t i ') 2 ) 1 / 2 Among them, CV time Represents the time stability correction factor; i and z are both indices; K represents the number of position identifiers in the next position candidate set; p represents the i-th position identifier c in the next position candidate set. t,h u(i) The total number of timestamps in the accessed timestamp set; t i,z c represents the identifier of the i-th position in the candidate set for the next position. t,h u(i) The z-th timestamp accessed; t i 'Represents the identifier c of the i-th position in the next position candidate set. t,h u(i) The average of all timestamps in the set of accessed timestamps.

[0069] Construct a user sparsity correction factor based on the set of users visited by each location identifier in the next location candidate set: Obtain the historical check-in sequence of all users. Considering that in actual human travel behavior, different users have different access habits for different types of locations, if the current user's access habits are relatively rare, frequently visiting locations with low access frequency, then their travel characteristics may not conform to common subnetwork rules. If the i-th location identifier c in the next location candidate set... t,h u(i) The set of users accessed is U i ={u i,1 ,ui,2 ,…,u i,Qi Introduce a user sparsity correction factor to the location subnetwork access frequency characteristics: CV user =(1 / K)∑ i=1 K (1 / Q i ) Among them, CV user K represents the user sparsity correction factor; K represents the number of location identifiers in the next location candidate set; i is the index; Q represents the number of location identifiers in the next location candidate set. i c represents the identifier of the i-th position in the candidate set for the next position. t,h u(i) The number of users in the user set that was visited.

[0070] D3. Calculate the weight of each feature in the first stage of the three-dimensional evaluation matrix based on the time stability correction factor, the user sparsity correction factor, and the information entropy of each feature in the three-dimensional evaluation matrix.

[0071] The weights of each feature column in the three-dimensional evaluation matrix for the first stage are set according to the following formula: w e,t,h,j u =(1-e t,h,j u )CV j ) / (∑ y=1 3 ((1-e t,h,y u )CV y j=1,2,3 CV1=CV time ;CV2=1;CV3=CV user e t,h,j u =(1 / lnK)∑ i=1 K (z t,h,ij u ·lnz t,h,ij u ), j=1,2,3 z t,h,ij u =(m t,h,ij u ) / (∑ i=1 K m t,h,ij u ), j=1,2,3 Where j and y are both indices; w e,t,h,j u e represents the weight of the first stage of the j-th column feature;t,h,j u and e t,h,y u These represent the information entropy of the features in the j-th and y-th columns of the three-dimensional evaluation matrix, respectively; CV j and CV y These represent the correction values ​​corresponding to the features in column j and column y, respectively; CV time and CV user These represent the time stability correction factor and the user sparsity correction factor, respectively; i is the index; K represents the number of location identifiers in the next location candidate set; z t,h,ij u Let be the weight value of the feature in the j-th column for the i-th position; ln represents the logarithm to the base e; m t,h,ij u Represents the three-dimensional evaluation matrix M t,h u The eigenvalue at the i-th position in the j-th column (the eigenvalue in the i-th row and j-th column); CV1, CV2, and CV3 represent the correction values ​​in the 1st, 2nd, and 3rd columns of the three-dimensional evaluation matrix, respectively.

[0072] D4. Based on the three-dimensional evaluation matrix, the positive ideal solution, the negative ideal solution, and the weight of each column of features in the three-dimensional evaluation matrix in the first stage, the TOPSIS method is used to sort the candidate set of the next position in the first stage, and the candidate set of position identifiers after the first stage sorting is obtained, namely the first stage candidate set of position identifiers Rank1.

[0073] The TOPSIS method is implemented in the following steps: H1. Calculate the distance from each position identifier in the next position candidate set to the positive and negative ideal solutions based on the three-dimensional evaluation matrix, the weight of each feature in the first stage of the three-dimensional evaluation matrix, the positive ideal solution, and the negative ideal solution.

[0074] s t,h,p,i u =(∑ j=1 3 w e,t,h,j u (m t,h,ij u -p t,h u(j) ) 2 ) 1 / 2 i=1,,2,…,K s t,h,n,i u =(∑ j=1 3 w e,t,h,j u (m t,h,ij u -n t,hu(j) ) 2 ) 1 / 2 i=1,,2,…,K Among them, s t,h,p,i u s represents the distance from the i-th position marker to the positive ideal solution; t,h,n,i u w represents the distance from the i-th position marker to the negative ideal solution; e,t,h,j u The m represents the weight of the j-th column feature in the first stage; t,h,ij u p represents the feature value of the i-th position identifier in the j-th column (the feature value in the r-th row and j-th column); t,h u(j) The positive ideal solution P is represented by t,h u The j-th column feature in n; t,h u(j) N represents the negative ideal solution t,h u The first column of features.

[0075] H2. Obtain the proximity value of each location identifier to the positive ideal solution based on the distance of each location identifier to the positive ideal solution and the negative ideal solution. Sort the location identifiers in descending order based on the proximity value of each location identifier to the positive ideal solution to obtain the first-stage sorted candidate set of location identifiers, namely the first-stage candidate set of location identifiers Rank1.

[0076] The proximity value of each location identifier to the positive ideal solution is set according to the following formula: v t,h,i u =s t,h,n,i u / (s t,h,p,i u +s t,h,n,i u ), i=1,2,…,K Among them, v t,h,i u v represents the proximity value from the i-th position identifier to the positive ideal solution. t,h,i u ∈[0,1];s t,h,p,i u s represents the distance from the i-th position marker to the positive ideal solution; t,h,n,i u This represents the distance from the i-th position identifier to the negative ideal solution.

[0077] The first-stage candidate set of location identifiers is set according to the following formula: Rank1={c rank1,t,h u(1) ,crank1,t,h u(2) ,…,c rank1,t,h u(K)} Wherein, Rank1 represents the candidate set of location identifiers in the first stage; c rank1,t,h u(1) c rank1,t,h u(2) and c rank1,t,h u(K) These represent the 1st, 2nd, and Kth location identifiers in the first-stage location identifier candidate set, respectively.

[0078] S22. Based on the three-dimensional evaluation matrix, the CRITIC-TOPSIS algorithm is used to sort the next position candidate set in the second stage, resulting in the position identifier candidate set after the second stage sorting, i.e., the second stage position identifier candidate set.

[0079] F1, using the positive ideal solution P obtained in step D1 t,h u and negative ideal solution N t,h u .

[0080] F2. Obtain the contrast intensity and conflict intensity of each feature column based on all eigenvalues ​​in the three-dimensional evaluation matrix.

[0081] F3. Obtain the weight of each feature in the second stage of the three-dimensional evaluation matrix based on the contrast intensity and conflict intensity of each feature.

[0082] The weights of each feature column in the second stage of the three-dimensional evaluation matrix are set according to the following formula: w c,t,h,j u =(σ t,h,j u ·I t,h,j u ) / (∑ j=1 3 (σ t,h,j u ·I t,h,j u j=1,2,3 σ t,h,j u =((1 / K)(∑ i=1 K (m t,h,ij u -m t,h,·j u ˉ) 2 )) 1 / 2 I t,h,j u =∑j’=1 3 (1-|r jj’ |), j≠j' r jj’ =(∑ i=1 K ((m t,h,ij u -m t,h,·j u’ (m) t,h,ij’ u -m t,h,·j’ u’ ))) / ((∑ i=1 K (m t,ij u -m t,h,·j u’ ) 2 ) 1 / 2 ·(∑ i=1 K (m t,ij’ u -m t,h,·j’ u’ ) 2 ) 1 / 2 ) Where i, j, and j' all represent indices; w c,t,h,j u σ represents the weight of the second stage of the j-th column feature; t,h,j u Indicates the contrast intensity of the feature in column j; I t,h,j u The value r represents the conflict intensity of the feature in column j; jj’ Represents the feature correlation value between the features in column j and column j'; |r jj’ | indicates taking the feature correlation value r jj’ The absolute value; K represents the number of location identifiers in the next location candidate set; m t,h,ij u and m t,h,ij’ u These represent the feature values ​​corresponding to the i-th position identifier in the j-th and j'-th columns (the feature value in the i-th row and j-th column and the feature value in the i-th row and j'-th column, respectively); m t,h,·j u’ m represents the average eigenvalue obtained by summing and averaging all eigenvalues ​​in column j; t,h,·j’ u’ This represents the average eigenvalue obtained by summing and averaging all eigenvalues ​​in column j'.

[0083] F4. Based on the weights of the features in the second stage of each column of the three-dimensional evaluation matrix, the same TOPSIS algorithm as in step 21 is used to sort the next position candidate set in the second stage, resulting in the position identifier candidate set after the second stage sorting, i.e., the second stage position identifier candidate set Rank2.

[0084] The resulting second-stage location identifier candidate set is set according to the following formula: Rank2={c rank2,t,h u(1) ,c rank2,t,h u(2) ,…,c rank2,t,h u(K)} Wherein, Rank2 represents the candidate set of location identifiers for the second stage; c rank2,t,h u(1) c rank2,t,h u(2) and c rank2,t,h u(K) These represent the 1st, 2nd, and Kth location identifiers in the candidate set of location identifiers for the second stage, respectively.

[0085] Step S3: Select the top K1 (TOP K1) location identifiers from the first-stage location identifier candidate set Rank1, remove the selected location identifiers from the second-stage location identifier candidate set Rank2, and obtain the second-stage location identifier candidate set Rank2' after removal. Then select the top K2 (TOP K2) location identifiers from the second-stage location identifier candidate set Rank2' to supplement it. Thus, the combination of the K1 and K2 location identifiers selected successively constitutes the final next location candidate set Rank.

[0086] The final set of candidate next positions is set according to the following steps: Rank={c rank,t u(1) ,c rank,t u(2) ,…,c rank,t u(K1) ,…,c rank,t u(K1+K2)} Where Rank represents the final set of candidate positions for the next position; c rank,t u(1) c rank,t u(2) c rank,t u(K1) and c rank,t u(K1+K2) These represent the 1st, 2nd, K1th, and (K1+K2)th position identifiers in the final candidate set Rank for the next position, respectively.

[0087] Rank1 reflects a ranking based on the dominant trend of behavioral preferences, while Rank2 provides a differentiated ranking perspective based on feature conflict. By taking Rank1 and Rank2 into account, the final ranked candidate set of positions is obtained.

[0088] The resulting Rank set represents the set of locations where the current user is most likely to move next, improving the intelligence level of user travel recommendations and providing important support for infectious disease transmission prediction, urban traffic management, energy scheduling, and business decision-making.

[0089] like Figure 2 As shown, this embodiment also provides a next location prediction system, including: The three-dimensional evaluation matrix construction module obtains check-in samples of user movement behavior, serializes all check-in samples in chronological order to obtain historical check-in sequences, obtains historical check-in location subsequences, historical check-in activity subsequences, and next location candidate sets based on the historical check-in sequence, and constructs location access frequency features, location subnetwork access frequency features, and activity subnetwork access frequency features based on the historical check-in location subsequences, historical check-in activity subsequences, and next location candidate sets, respectively, and constructs a three-dimensional evaluation matrix based on the three types of features.

[0090] The first-stage location identifier candidate set acquisition module uses an improved entropy weight-TOPSIS algorithm based on the three-dimensional evaluation matrix to sort the next location candidate set in the first stage, resulting in the first-stage sorted location identifier candidate set, which is the first-stage location identifier candidate set.

[0091] The second-stage location identifier candidate set acquisition module uses the CRITIC-TOPSIS algorithm to sort the next location candidate set in the second stage according to the three-dimensional evaluation matrix, and obtains the second-stage sorted location identifier candidate set, which is the second-stage location identifier candidate set.

[0092] The next position candidate set acquisition module selects the first few position identifiers from the first stage position identifier candidate set, removes the selected position identifiers from the second stage position identifier candidate set, and obtains the second stage position identifier candidate set after removal. Then, it selects the first few position identifiers from the second stage position identifier candidate set to supplement it. Thus, all the position identifiers selected in succession constitute the final next position candidate set.

[0093] Furthermore, to demonstrate the beneficial effects of the present invention, this embodiment obtains the user's historical check-in records (continuous historical check-in samples) based on the user's check-in activities on the check-in website, and then obtains the user's historical check-in sequence. Based on the user's historical check-in sequence, the existing MC, GRU, Time-LSTM, DeepMove, Flashback, MSSRM, and HGARN algorithms are directly used for next position prediction. The comparison results of directly using the MC, GRU, Time-LSTM, DeepMove, Flashback, MSSRM, and HGARN algorithms and the method of the present invention under the same hardware and software conditions are shown in the following table: Model Recall rate @1 Recall rate @5 Recall rate @10 Normalized loss cumulative gain @5 Normalized loss cumulative gain @10 Time (in seconds) required to predict a single sample GRU 0.210 0.503 0.626 0.363 0.403 0.009 Time-LSTM 0.225 0.526 0.653 0.382 0.424 0.016 DeepMove 0.206 0.519 0.649 0.369 0.412 0.023 Flashback 0.220 0.541 0.677 0.386 0.430 0.014 MSSRM 0.239 0.554 0.668 0.404 0.441 0.012 HGARN 0.277 0.574 0.657 0.438 0.464 0.014 This invention 0.288 0.579 0.701 0.443 0.482 0.001 As shown in the table above and Figure 3 , Figure 4 , Figure 5 It is evident that the method proposed in this invention outperforms existing models in multiple evaluation metrics. As shown in the table, this invention achieves recall rates of 0.288, 0.579, and 0.701 for @1, @5, and @10, respectively, representing improvements of approximately 1.1%, 0.5%, and 4.4% compared to the best-performing deep learning model HGARN. Furthermore, it achieves 0.443 and 0.482 for @5 and @10 of the normalized loss cumulative gain, respectively, indicating that this method excels in the rationality of candidate position ranking and top relevance. In addition, this invention demonstrates exceptional inference efficiency, with a single-sample prediction time of only 0.001 seconds, an improvement of approximately one order of magnitude compared to the average deep learning model. Since this method requires no training process and only performs explicit calculations based on historical check-in data, it significantly reduces computational costs and storage requirements, making it suitable for deployment in resource-constrained or real-time prediction scenarios. In summary, this invention combines high accuracy, high interpretability, and high efficiency. By introducing time stability correction factors and user sparsity correction factors, the algorithm can maintain stable prediction performance under conditions of few samples. The two-stage weighting strategy based on information entropy and feature conflict degree enables the model to achieve a good balance between interpretability and generalization, and has significant application potential and promotion value.

[0094] The above-described embodiments are merely preferred embodiments provided to fully illustrate the present invention, and the scope of protection of the present invention is not limited thereto. Equivalent substitutions or modifications made by those skilled in the art based on the present invention are all within the scope of protection of the present invention. The scope of protection of the present invention is defined by the claims.

Claims

1. A lightweight next location prediction method of zero-shot two-stage dynamic weighted ranking, characterized in that, Comprise the following steps: S1, obtaining the check-in sample of user mobile behavior, sequencing all check-in samples in time sequence to obtain the historical check-in sequence, obtaining the historical check-in position subsequence, the historical check-in activity subsequence and the next position candidate set according to the historical check-in sequence respectively, constructing the position access frequency feature, the activity subnetwork access frequency feature and the position subnetwork access frequency feature according to the historical check-in position subsequence, the historical check-in activity subsequence and the next position candidate set respectively, and constructing the three-dimensional evaluation matrix according to the three types of features; S2, the improved entropy weight-TOPSIS algorithm is used to sort the next position candidate set according to the three-dimensional evaluation matrix to obtain the first stage position identification candidate set;The CRITIC-TOPSIS algorithm is used to sort the next position candidate set according to the three-dimensional evaluation matrix to obtain the second stage position identification candidate set; S3, select the first several position identifications from the first stage position identification candidate set, remove the selected position identifications from the second stage position identification candidate set, obtain the second stage position identification candidate set after removal, and select the first several position identifications from the second stage position identification candidate set for supplementation, so that all the position identifications selected in succession form the next position candidate set of the current check-in position of the user.

2. The lightweight next position prediction method of claim 1, wherein, The step S1 is specifically: S11, obtaining the current check-in sample of user mobile behavior and the previous h continuous historical check-in samples of the current check-in sample, and sequencing all check-in samples in time sequence to obtain the historical check-in sequence; Each check-in sample in the historical check-in sequence comprises a position identification, an activity type identification and a timestamp; S12, all position identifications of the current check-in sample and the previous h check-in samples of the current check-in sample in the historical check-in sequence form a historical check-in position subsequence; S13, all activity identifications of the current check-in sample and the previous h check-in samples of the current check-in sample in the historical check-in sequence form a historical check-in activity subsequence; S14, the set of position identifications obtained after removing the same position identifications in the historical check-in position subsequence is taken as the next position candidate set of the current check-in sample;Each position identification in the obtained next position candidate set is a candidate position; S15, constructing the position access frequency feature, the position subnetwork access frequency feature and the activity subnetwork access frequency feature according to the historical check-in position subsequence, the historical check-in activity subsequence and the next position candidate set respectively, and constructing the three-dimensional evaluation matrix according to the three types of features.

3. The lightweight next position prediction method of claim 1, wherein, The step S15 is specifically: S151, the sequence formed by the frequency of each position identification in the historical check-in position subsequence in the next position candidate set is taken as the position access frequency feature; S152, the sequence formed by the total frequency of each candidate position appearing in the preset window after marking all positions in the historical check-in position subsequence is taken as the position subnetwork access frequency feature; S153, a sequence constituted by total frequencies of appearance of activity identifiers corresponding to each candidate position in a preset window after all marked activities counted in the historical check-in activity subsequence as activity subnetwork access frequency features; S154, concatenating the position access frequency features, the activity subnetwork access frequency features and the position subnetwork access frequency features in turn to form an initial three-dimensional evaluation matrix, and performing range normalization on each column dimension feature of the initial three-dimensional evaluation matrix to obtain a three-dimensional evaluation matrix.

4. The lightweight next position prediction method of claim 1, wherein, The improved entropy weight-TOPSIS algorithm is specifically: D1, calculating a positive ideal solution and a negative ideal solution according to the three-dimensional evaluation matrix; D2, constructing a time stability correction factor according to a timestamp set of each position identifier accessed in the next position candidate set; constructing a user sparsity correction factor according to a user set of each position identifier accessed in the next position candidate set; D3, calculating a first-stage weight of each column feature in the three-dimensional evaluation matrix according to the time stability correction factor, the user sparsity correction factor and the information entropy of each column feature in the three-dimensional evaluation matrix; D4, performing a first-stage sorting on the next position candidate set by the TOPSIS method according to the positive ideal solution, the negative ideal solution and the first-stage weight of each column feature in the three-dimensional evaluation matrix, to obtain a first-stage sorted position identifier candidate set, that is, a first-stage position identifier candidate set.

5. The lightweight next position prediction method of zero sample two-stage dynamic weighted sorting according to claim 4, characterized in that: The time stability correction factor and the user sparsity correction factor are set according to the following formula: CV time = (1 / K)∑ i=1 K ((1 / p)∑ z=1 p (t i,z -t i ’) 2 ) 1 / 2 CV user = (1 / K)∑ i=1 K (1 / Q i ) wherein CV time and CV user respectively represent a temporal stability correction factor and a user sparsity correction factor; i is an index; K represents the number of position identifiers in the next position candidate set; p represents the total number of time stamps of the time stamp set visited by the i-th position identifier in the next position candidate set; z is an index; t i,z represents the z-th time stamp visited by the i-th position identifier in the next position candidate set; t i ’ represents the mean of all time stamps in the time stamp set visited by the i-th position identifier in the next position candidate set; Q i represents the number of users of the user set visited by the i-th position identifier in the next position candidate set.

6. The lightweight next position prediction method of zero sample two-stage dynamic weighted sorting according to claim 4, characterized in that: The first-stage weight of each column feature in the three-dimensional evaluation matrix is set according to the following formula: w e,t,h,j u =((1-e t,h,j u )CV j ) / (∑ y=1 3 ((1-e t,h,y u )CV y )), j = 1, 2, 3 CV1 = CV time ; CV2 = 1 ; CV3 = CV user e t,h,j u =(1 / lnK)∑ i=1 K (z t,h,ij u ·lnz t,h,ij u ) z t,h,ij u =(m t,h,ij u ) / (∑ i=1 K m t,h,ij u ) wherein j and y are indexes; w e,t,h,j u is the weight of the first stage for the jth column feature; e t,h,j u and e t,h,y u respectively represent the information entropy of the jth column and yth column features in the three-dimensional evaluation matrix; CV j and CV y respectively represent the correction values corresponding to the jth column and yth column features; CV1, CV2 and CV3 respectively represent the correction values of the first column, second column and third column in the three-dimensional evaluation matrix; CV time and CV user respectively represent the time stability correction factor and the user sparsity correction factor; i is an index; K represents the number of position identifiers in the next position candidate set; z t,h,ij u is the proportion value of the ith position identifier in the jth column feature; m t,h,ij u represents the feature value of the ith position identifier in the jth column in the three-dimensional evaluation matrix.

7. The lightweight next position prediction method of claim 1, wherein, The CRITIC-TOPSIS algorithm is specifically: F1, calculating a positive ideal solution and a negative ideal solution according to the three-dimensional evaluation matrix; F2, obtaining a contrast intensity and a conflict intensity of each column feature according to all feature values in the three-dimensional evaluation matrix; F3, obtaining a second-stage weight of each column feature in the three-dimensional evaluation matrix according to the contrast intensity and the conflict intensity of each column feature; F4, performing a second-stage sorting on the next position candidate set by the TOPSIS method according to the second-stage weight of each column feature in the three-dimensional evaluation matrix, to obtain a second-stage sorted position identifier candidate set, that is, a second-stage position identifier candidate set.

8. The lightweight next position prediction method of zero sample two-stage dynamic weighted sorting according to claim 7, characterized in that: The contrast intensity and the conflict intensity of each column feature are set according to the following formula: σ t,h,j u = ((1 / K)∑ i=1 K (m t,h,ij u -m t,h,·j u’ ) 2 ) 1 / 2 I t,h,j u =∑ j’=1 3 (1-|r jj’ |),j≠j’ r jj’ = (∑ i=1 K ((m t,h,ij u -m t,h,·j u’ )(m t,h,ij’ u -m t,h,·j’ u’ ))) / ((∑ i=1 K (m t,ij u -m t,h,·j u’ ) 2 ) 1 / 2 ·(∑ i=1 K (m t,ij’ u -m t,h,·j’ u’ ) 2 ) 1 / 2 ) where i, j and j' represent indices; σ t,h,j u represents the contrast intensity of the jth column feature; I t,h,j u represents the conflict intensity of the jth column feature; r jj’ represents the feature correlation value between the jth column and j'th column features; K represents the number of position identifications in the next position candidate set; m t,h,ij u and m t,h,ij’ u respectively represent the feature values of the ith position identification in the jth column and j'th column; m t,h,·j u’ and m t,h,·j’ u’ respectively represent the average feature values obtained by summing and averaging all the feature values corresponding to the jth column and j'th column.

9. The lightweight next position prediction method of zero sample two-stage dynamic weighted sorting according to claim 7, characterized in that: The second-stage weight of each column feature in the three-dimensional evaluation matrix is set according to the following formula: w c,t,h,j u =(σ t,h,j u ·I t,h,j u ) / (∑ j=1 3 (σ t,h,j u ·I t,h,j u )),j=1,2,3 where j represents an index; w c,t,h,j u represents the weight of the second stage of the jth column feature; σ t,h,j u represents the contrast intensity of the jth column feature; I t,h,j u represents the conflict intensity of the jth column feature.

10. A next location prediction system employing a lightweight next location prediction method with zero-shot two-stage dynamic weighted ranking as claimed in any one of claims 1-9, characterized in that, including: The three-dimensional evaluation matrix construction module obtains check-in samples of user mobile behaviors, serializes all check-in samples in time sequence to obtain a historical check-in sequence, obtains a historical check-in position sub-sequence, a historical check-in activity sub-sequence and a next position candidate set respectively according to the historical check-in sequence, respectively constructs a position access frequency feature, a position sub-network access frequency feature and an activity sub-network access frequency feature according to the historical check-in position sub-sequence, the historical check-in activity sub-sequence and the next position candidate set, and constructs a three-dimensional evaluation matrix according to the three types of features; The first-stage position identifier candidate set acquisition module adopts an improved entropy weight-TOPSIS algorithm to perform first-stage sorting on the next position candidate set according to the three-dimensional evaluation matrix, obtains a first-stage sorted position identifier candidate set, and the first-stage sorted position identifier candidate set is the first-stage position identifier candidate set; The second-stage position identifier candidate set acquisition module adopts a CRITIC-TOPSIS algorithm to perform second-stage sorting on the next position candidate set according to the three-dimensional evaluation matrix, obtains a second-stage sorted position identifier candidate set, and the second-stage sorted position identifier candidate set is the second-stage position identifier candidate set; The next position candidate set acquisition module selects a plurality of position identifiers from the first-stage position identifier candidate set, removes the selected position identifiers from the second-stage position identifier candidate set, obtains a second-stage position identifier candidate set after removal, and selects a plurality of position identifiers from the second-stage position identifier candidate set after removal for supplementation, so that all the position identifiers selected in succession form the final next position candidate set.

Images