User category identification methods, devices, equipment, and storage media

Abstract

This invention belongs to the field of behavior recognition technology and discloses a method, apparatus, device, and storage medium for user category identification. The method includes: dividing a walking signaling trajectory dataset into a labeled dataset and an unlabeled dataset, the labeled dataset containing user labels and trajectory labels; obtaining an attribute variable dataset from the labeled dataset, and performing supervised discretization on the continuous attribute variable data in the attribute variable dataset to obtain a new attribute variable dataset; obtaining probability-weighted category labels from the labeled and unlabeled datasets, and establishing a trajectory recognition model based on the probability-weighted category labels and the new attribute variable dataset; determining the trajectory recognition ratio based on the trajectory recognition model and the walking signaling trajectory of the target communication number; and determining the user category corresponding to the target communication number based on the trajectory recognition ratio and a ratio threshold. Through the above methods, the model accuracy is improved, thereby improving the recognition accuracy and enabling accurate identification of users with mobility impairments.

Description

Technical Field

[0001] This invention relates to the field of behavior recognition technology, and in particular to a user category recognition method, apparatus, device, and storage medium. Background Technology

[0002] Current methods for user category identification using signaling data largely employ population-based theories. These methods analyze signaling data reported by a population within a target area over a specific period to determine their residency characteristics. This data is then linked to operator user information to analyze and obtain data on permanent residents, migrant populations, and other age groups and genders. This allows medical institutions to identify individuals in scenarios such as breast and cervical cancer screening and chronic disease checkups for middle-aged women, the elderly, and migrant workers. However, this approach only addresses relatively broad groups. In practical applications, it typically only labels a small amount of walking trajectory data as either "trajectory of a patient with mobility impairment" or "trajectory of someone who is not with mobility impairment." This makes it difficult to apply mainstream supervised learning to such limited labeled datasets. Furthermore, the presence of systematic or random errors in mobile phone signaling reports, along with the randomness of individual walking behavior, hinders further segmentation and identification, making it difficult to target users in urgent need of healthcare. Consequently, the accuracy of identification is low. Summary of the Invention

[0003] The main objective of this invention is to provide a user category identification method, apparatus, device, and storage medium, aiming to solve the technical problems in the prior art where traditional methods for user category identification using signaling data are difficult to apply supervised learning, and the signaling reported by mobile phones has certain errors, resulting in low identification accuracy.

[0004] To achieve the above objectives, the present invention provides a user category identification method, the method comprising the following steps:

[0005] Obtain a pedestrian signaling trajectory dataset and divide the pedestrian signaling trajectory dataset into a labeled dataset and an unlabeled dataset, wherein the labeled dataset includes user labels and trajectory labels;

[0006] Obtain the attribute variable dataset of the labeled dataset, and perform supervised discretization on the continuous attribute variable data in the attribute variable dataset to obtain a new attribute variable dataset.

[0007] Obtain the probability-weighted category labels of the labeled dataset and the unlabeled dataset respectively, and establish a trajectory recognition model based on the probability-weighted category labels and the new attribute variable dataset;

[0008] Based on the trajectory recognition model and the walking signaling trajectory of the target communication number, the trajectory recognition ratio is determined;

[0009] Based on the trajectory recognition ratio and the ratio threshold, the user category corresponding to the target communication number is determined.

[0010] Optionally, the attribute variable dataset includes continuous attribute variable data and discrete attribute variable data. The supervised discretization of the continuous attribute variable data in the attribute variable dataset to obtain a new attribute variable dataset includes:

[0011] The number of intervals is determined based on the trajectory labels in the labeled dataset.

[0012] The continuous attribute variable data of the labeled dataset is transformed into an attribute data queue;

[0013] Based on the attribute data queue, an attribute data queue set is determined, wherein the attribute data queue set is composed of attribute data queue groups, and the attribute data queue groups are composed of the attribute data queues.

[0014] Sort the attribute data queues in the attribute data queue group in descending order, and calculate the expected frequency matrix between the attribute data queues;

[0015] Based on the expected frequency matrix, determine the chi-square value between the attribute data queues;

[0016] The attribute data queues are merged according to the chi-square value to obtain a new attribute data queue group;

[0017] When the number of attribute data queues in the new attribute data queue group is equal to the number of intervals, a new attribute data queue set is generated based on the new attribute data queue group.

[0018] Based on the new attribute data queue set, the continuous attribute variable data of the labeled dataset is updated to obtain new continuous attribute variable data;

[0019] New attribute variable data is obtained based on the new continuous attribute variable data and the new discrete attribute variable data.

[0020] Optionally, determining the trajectory recognition ratio based on the trajectory recognition model and the walking signaling trajectory of the target communication number includes:

[0021] The walking signaling trajectory is input into the trajectory recognition model to obtain the trajectory recognition result corresponding to the target communication number;

[0022] Based on the trajectory recognition results, determine the number of mobility impairment trajectories corresponding to the target communication number;

[0023] The trajectory recognition ratio corresponding to the target communication number is determined based on the number of mobility impairment trajectories of the target communication number and the total number of trajectories.

[0024] Optionally, the user category includes users with mobility impairments and users without mobility impairments. Determining the user category corresponding to the target communication number based on the trajectory recognition ratio and the ratio threshold includes:

[0025] Obtain the identification frequency data corresponding to the target communication number, determine the precision and recall based on the identification frequency data, and obtain the weighted harmonic average corresponding to the target communication number based on the precision and the recall.

[0026] Based on the weighted harmonic mean, a proportion threshold is determined, wherein the proportion threshold is the maximum weighted harmonic mean among the weighted harmonic means;

[0027] When the trajectory recognition ratio of the target communication number is greater than or equal to the ratio threshold, the user category corresponding to the target communication number is determined to be a user with mobility impairment;

[0028] When the trajectory recognition ratio of the target communication number is less than the ratio threshold, the user category corresponding to the target communication number is determined to be a non-mobile disabled user.

[0029] Optionally, obtaining the probability-weighted category labels of the labeled dataset and the unlabeled dataset respectively includes:

[0030] Obtain the first correspondence between the posterior probability data and the indicator vector of the true class;

[0031] Obtain the second correspondence between the indicator vector of the true category and the probability-weighted category label;

[0032] Based on the first correspondence and the second correspondence, determine the correspondence between the posterior probability data and the probability-weighted category labels;

[0033] Based on the posterior probability data of the labeled dataset and the correspondence between the posterior probability data and the probability-weighted class labels, the probability-weighted class labels of the labeled dataset are determined.

[0034] Based on the posterior probability data of the unlabeled dataset and the correspondence between the posterior probability data and the probability-weighted class labels, the probability-weighted class labels of the unlabeled dataset are determined.

[0035] Optionally, establishing a trajectory recognition model based on the probability-weighted category labels and the new attribute variable dataset includes:

[0036] Based on the trajectory labels and the labeled dataset, determine the trajectory set for people with mobility impairments and the trajectory set for people without mobility impairments;

[0037] Based on the probability-weighted category labels and the set of trajectory data for people with mobility impairments, determine the probability-weighted category labels of the set of trajectory data for people with mobility impairments; based on the probability-weighted category labels and the set of trajectory data for people without mobility impairments, determine the probability-weighted category labels of the set of trajectory data for people without mobility impairments.

[0038] The first parameter of the Naive Bayes classifier is determined based on the probability-weighted category labels of the set of trajectories with limited mobility, the probability-weighted category labels of the set of trajectories with limited mobility, and the third correspondence.

[0039] Based on the new attribute variable dataset, determine the indicator data;

[0040] Based on the indicated data and the probability-weighted category labels of the set of trajectories for people with mobility impairments, determine the sum of the trajectory products of the set of trajectories for people with mobility impairments; based on the indicated data and the probability-weighted category labels of the set of trajectories for people without mobility impairments, determine the sum of the trajectory products of the set of trajectories for people without mobility impairments.

[0041] The second parameter of the Naive Bayes classifier is determined based on the sum of the product of the trajectories of the set of trajectories with limited mobility, the sum of the product of the trajectories of the set of trajectories with limited mobility, and the fourth correspondence.

[0042] When the first parameter or the second parameter does not meet the preset convergence condition, the probability weighted category label is updated according to the first parameter and the second parameter, and the process returns to the step of determining the probability weighted category label of the set of disabled trajectories based on the probability weighted category label of the walking signaling trajectory and the set of disabled trajectories.

[0043] When the first parameter and the second parameter satisfy a preset convergence condition, the trajectory recognition model is established based on the first parameter and the second parameter.

[0044] Optionally, determining the indication data based on the new attribute variable dataset includes:

[0045] Based on the new attribute variable dataset, determine the correlation coefficients between the attribute variables, which include continuous attribute variables and discrete attribute variables;

[0046] Based on the correlation coefficient and correlation threshold, the continuous attribute variables and the discrete attribute variables are filtered to determine the feature attribute variables;

[0047] Based on the aforementioned feature attribute variables, determine the trajectory feature data corresponding to the labeled dataset;

[0048] Based on the trajectory feature data and trajectory feature value data corresponding to the labeled dataset, the indication data is determined.

[0049] Furthermore, to achieve the above objectives, the present invention also proposes a user category identification device, the user category identification device comprising:

[0050] The data annotation module is used to acquire a pedestrian signaling trajectory dataset and divide the pedestrian signaling trajectory dataset into a labeled dataset and an unlabeled dataset, wherein the labeled dataset includes user labels and trajectory labels;

[0051] The model building module is used to obtain the attribute variable dataset of the labeled dataset, and to perform supervised discretization on the continuous attribute variable data in the attribute variable dataset to obtain a new attribute variable dataset.

[0052] The model building module is also used to obtain the probability-weighted category labels of the labeled dataset and the unlabeled dataset respectively, and to build a trajectory recognition model based on the probability-weighted category labels and the new attribute variable dataset;

[0053] The user identification module is used to determine the trajectory identification ratio based on the trajectory identification model and the walking signaling trajectory of the target communication number;

[0054] The user identification module is further configured to determine the user category corresponding to the target communication number based on the trajectory identification ratio and the ratio threshold.

[0055] Furthermore, to achieve the above objectives, the present invention also proposes a user category identification device, the user category identification device comprising: a memory, a processor, and a user category identification program stored in the memory and executable on the processor, the user category identification program being configured to implement the steps of the user category identification method as described above.

[0056] In addition, to achieve the above objectives, the present invention also proposes a storage medium storing a user category identification program, which, when executed by a processor, implements the steps of the user category identification method as described above.

[0057] In this invention, a pedestrian signaling trajectory dataset is acquired and divided into a labeled dataset and an unlabeled dataset. An attribute variable dataset is then acquired from the labeled dataset. The continuous attribute variable data in the attribute variable dataset undergoes supervised discretization to obtain a new attribute variable dataset. Probability-weighted class labels are acquired from both the labeled and unlabeled datasets. A trajectory recognition model is established based on the probability-weighted class labels and the new attribute variable dataset. Based on the trajectory recognition model and the pedestrian signaling trajectory of the target communication number, the trajectory recognition ratio is determined. Finally, based on the trajectory recognition ratio and a ratio threshold, the user category corresponding to the target communication number is determined. This invention effectively utilizes walking trajectories based on mobile phone signaling, applying chi-square optimal binning to identify scenarios involving users with mobility impairments. This enhances the interpretability of walking trajectory attribute variables. Compared to sample sets with few labeled samples, which are difficult to supervise with learning, this invention, based on the expectation-maximization method framework, uses probability-weighted class labels during model training to establish a Naive Bayes semi-supervised classifier, improving model accuracy. Since there are inherent errors in mobile phone signaling, this invention first identifies the walking trajectory category and then solves for the optimal proportion threshold to determine the user's category, reducing identification bias caused by signaling errors and improving accuracy. It achieves a semi-supervised learning hierarchical identification model from trajectory recognition to user identification, accurately identifying users with mobility impairments for screening medical service recipients. Attached Figure Description

[0058] Figure 1 This is a schematic diagram of the structure of a user category identification device in the hardware operating environment involved in the embodiments of the present invention;

[0059] Figure 2 This is a flowchart illustrating the first embodiment of the user category identification method of the present invention;

[0060] Figure 3 This is a flowchart illustrating the second embodiment of the user category identification method of the present invention;

[0061] Figure 4 This is a schematic diagram of the attribute data queue set of an embodiment of the user category identification method of the present invention;

[0062] Figure 5 This is a flowchart illustrating the third embodiment of the user category identification method of the present invention;

[0063] Figure 6 This is a flowchart illustrating the fourth embodiment of the user category identification method of the present invention;

[0064] Figure 7 This is a structural block diagram of the first embodiment of the user category identification device of the present invention.

[0065] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0066] It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the scope of the invention.

[0067] Reference Figure 1 , Figure 1 This is a schematic diagram of the user category identification device structure in the hardware operating environment involved in the embodiments of the present invention.

[0068] like Figure 1 As shown, the user category identification device may include: a processor 1001, such as a central processing unit (CPU), a communication bus 1002, a user interface 1003, a network interface 1004, and a memory 1005. The communication bus 1002 is used to enable communication between these components. The user interface 1003 may include a display screen or an input unit such as a keyboard; optionally, the user interface 1003 may also include a standard wired interface or a wireless interface. The network interface 1004 may optionally include a standard wired interface or a wireless interface (such as a Wireless-Fidelity (Wi-Fi) interface). The memory 1005 may be a high-speed random access memory (RAM) or a stable non-volatile memory (NVM), such as a disk drive. The memory 1005 may also optionally be a storage device independent of the aforementioned processor 1001.

[0069] Those skilled in the art will understand that Figure 1 The structure shown does not constitute a limitation on the user category identification device and may include more or fewer components than illustrated, or combine certain components, or have different component arrangements.

[0070] like Figure 1 As shown, the memory 1005, which serves as a storage medium, may include an operating system, a network communication module, a user interface module, and a user category identification program.

[0071] exist Figure 1In the user category identification device shown, the network interface 1004 is mainly used for data communication with the network server; the user interface 1003 is mainly used for data interaction with the user; the processor 1001 and the memory 1005 in the user category identification device of the present invention can be set in the user category identification device, and the user category identification device calls the user category identification program stored in the memory 1005 through the processor 1001 and executes the user category identification method provided in the embodiment of the present invention.

[0072] This invention provides a user category identification method, referring to... Figure 2 , Figure 2 This is a flowchart illustrating a first embodiment of a user category identification method according to the present invention.

[0073] In this embodiment, the user category identification method includes the following steps:

[0074] Step S10: Obtain the pedestrian signaling trajectory dataset and divide the pedestrian signaling trajectory dataset into a labeled dataset and an unlabeled dataset. The labeled dataset contains user labels and trajectory labels.

[0075] It should be noted that the execution subject in this embodiment is a smart terminal, such as a computer. The smart terminal is equipped with a user category recognition program, which can identify the user category and thus determine the user with mobility impairment.

[0076] As can be understood, a pedestrian signaling trajectory dataset is the entire collection of data samples that stores communication numbers and their corresponding pedestrian signaling trajectories. Communication numbers can be mobile phone numbers, and pedestrian signaling trajectories are the pedestrian walking trajectories recorded by mobile phone signaling. Each communication number can correspond to multiple pedestrian signaling trajectories. User tags refer to labels used to annotate communication numbers, and trajectory tags refer to labels used to annotate pedestrian signaling trajectories.

[0077] Further, step S10 includes: labeling communication numbers in the walking signaling trajectory dataset with user labels according to a preset labeling range; labeling walking signaling trajectories corresponding to communication numbers with user labels with trajectory labels; dividing the walking signaling trajectory dataset into a training dataset and a validation dataset according to a preset division ratio; and dividing the training dataset into a labeled dataset and an unlabeled dataset according to the trajectory labels.

[0078] It should be understood that this embodiment labels a portion of the communication numbers in the dataset. The set labeling range is the preset labeling range. This portion of communication numbers will be labeled with corresponding user tags. There are two types of user tags in this embodiment: tags for objects with mobility impairments and tags for objects without mobility impairments. The walking signaling trajectories of these communication numbers that have been labeled with user tags will also be labeled with trajectory tags. There are two types of trajectory tags in this embodiment: walking trajectory tags for objects with mobility impairments and walking trajectory tags for objects without mobility impairments. The walking signaling trajectories corresponding to communication numbers labeled with objects with mobility impairments will be labeled with walking trajectory tags for objects with mobility impairments. Walking signaling trajectories labeled with walking trajectory tags for objects with mobility impairments can be considered walking trajectories for objects with mobility impairments. The walking signaling trajectories corresponding to communication numbers labeled with tags for objects without mobility impairments will be labeled with walking trajectory tags for objects without mobility impairments. Walking signaling trajectories labeled with walking trajectory tags for objects without mobility impairments can be considered walking trajectories for objects without mobility impairments. Generally, the number of communication numbers that can be clearly labeled with user tags is usually small. Therefore, the number of walking signaling tracks labeled with track tags is usually small, and usually less than the number of walking signaling tracks without track tags.

[0079] It should be noted that the training dataset and validation dataset refer to the sets of training samples and validation samples required for subsequent model training, respectively. The walking signaling trajectories in these sets are individual samples. The preset split ratio is the ratio of the training dataset to the validation dataset, for example, 8:2. This can be set according to actual needs, and this embodiment does not impose any restrictions on it. Next, based on whether the trajectory is labeled, the training dataset is divided into a labeled dataset and an unlabeled dataset. The labeled dataset is the set of samples consisting of walking signaling trajectories in the training dataset that are labeled, and the unlabeled dataset is the set of samples consisting of walking signaling trajectories in the training dataset that are not labeled. At this point, the number of samples in the labeled dataset is usually smaller, and the number of samples in the unlabeled dataset is usually larger. The unlabeled dataset is denoted as T. U Let the labeled dataset be denoted as T. L If the entire training dataset is T, then the entire training dataset is T.

[0080] In the specific implementation, the walking signaling trajectory dataset is labeled with trajectory and number. First, the communication numbers are labeled with user tags, and then the mobile phone numbers that have been labeled with user tags are labeled with trajectory tags. Thus, a small number of labeled walking signaling trajectories and a large number of unlabeled walking signaling trajectory samples constitute a complete data sample set. Then, the data sample set is divided into a training dataset and a validation dataset. The training dataset and the validation dataset are then divided into a small set of labeled samples and a large set of unlabeled samples, respectively.

[0081] It is understood that this embodiment will perform quality checks on the walking signaling trajectories uploaded by the mobile terminal. If missing values ​​or outliers are found, the quality check is considered to be unsatisfactory, and relevant preprocessing will be performed to ensure the quality of the walking signaling trajectories in the walking signaling trajectory dataset, thereby ensuring the quality of the samples used for model training and improving the accuracy of the model.

[0082] Step S20: Obtain the attribute variable dataset from the labeled dataset, and perform supervised discretization on the continuous attribute variable data in the attribute variable dataset to obtain a new attribute variable dataset.

[0083] It should be understood that an attribute variable dataset refers to a data set consisting of data corresponding to the attribute variables of the walking signaling trajectory. Attribute variables typically include continuous attribute variables and discrete attribute variables, corresponding to continuous attribute variable data and discrete attribute variable data, respectively.

[0084] It should be noted that, based on the physical kinematic characteristics of the walking trajectory and the natural and social attributes of the user to whom the walking trajectory belongs, this embodiment initially selected multiple attribute variables. The attribute variables for physical kinematic characteristics are Kine1 and Kine2, and the attribute variables for natural and social attributes are Peo1, Peo2, and Peo3. Among them, Kine1, Kine2, Peo1, Peo2, and Peo3 represent the continuous movement duration, continuous movement distance, user age, user gender, and whether the user is a regular user of the walking signaling trajectory, respectively. Kine1, Kine2, Peo1, and Peo2 are continuous attribute variables, and Peo3 is a discrete attribute variable. Kine1 and Kine2 can form the attribute vector of physical kinematic characteristics Kine = [Kine1, Kine2], and Peo1, Peo2, and Peo3 can form the attribute vector of natural and social attributes Peo = [Peo1, Peo2, Peo3].

[0085] It is understood that supervised discretization can use the chi-square optimal binning algorithm or other suitable algorithms; this embodiment does not impose any restrictions on this. After supervised discretization, new continuous attribute variable data will be obtained, resulting in a new attribute variable dataset, i.e., the new attribute variable dataset.

[0086] In the specific implementation, the attribute variables related to the walking trajectory are selected. Based on the chi-square optimal binning algorithm, the continuous attribute variables are discretized in a supervised manner, which can improve the interpretability of the attribute variables and reduce the risk of overfitting of the trajectory recognition model.

[0087] Step S30: Obtain the probability-weighted category labels of the labeled dataset and the unlabeled dataset respectively, and establish a trajectory recognition model based on the probability-weighted category labels and the new attribute variable dataset.

[0088] It should be understood that if there is a labeled dataset with the i-th walking signaling trajectory t... i According to the trajectory labels, the walking signaling trajectory t i There are two categories: walking signaling trajectories labeled with characteristics of mobility impairment (walking trajectory with mobility impairment characteristics), and walking signaling trajectories labeled with non-mobility impairment characteristics (walking trajectory without mobility impairment characteristics), denoted by c1 and c0 respectively. Here, the basic event "any walking signaling trajectory is a walking trajectory with mobility impairment characteristics" is represented as y = c1, and the basic event "any walking signaling trajectory is not a walking trajectory with mobility impairment characteristics" is represented as y = c0. Then P(y = c1|t) i The event "the i-th walking signaling trajectory is a walking trajectory with mobility impairment characteristics" can be represented by A. i1 The probability of P(y=c1|t) occurring can also be called the "posterior probability," representing the "posterior probability" that "the i-th walking trajectory" belongs to the class "walking trajectory with mobility impairment characteristics." i When ) = 1, the i-th walking signaling trajectory belongs to the "walking trajectory with mobility impairment characteristics" category, and is therefore considered to have a "positive label". Similarly, P(y = c0|t) i This can be represented as "the i-th walking signaling trajectory is not a walking trajectory with mobility impairment characteristics" (this event is denoted as A). i0 When P(y=c0|t) i When ) = 1, the walking signaling trajectory is considered to have a "negative label". Since for the i-th walking signaling trajectory, event A... i1 And Event A i0 If the events are complementary, then the sum of the probabilities of the two events is 1, meaning that the two posterior probabilities satisfy the relationship: P(y=c0|t) i )+P(y=c1|t i ) = 1.

[0089] It should be noted that, assuming there is a way to know the true category of the i-th walking signaling trajectory, then for the i-th walking signaling trajectory, in event A, which is a mutually exclusive event... i1 And Event A i0 In this context, there is one and only one event with a posterior probability of 1. If A is used... ij Unified representation of event A i1 And Event A i0 Then we can obtain the following expression:

[0090]

[0091] Therefore, the true category of the i-th walking signaling trajectory can be represented as a vector z of binary indicator variables. i , z i =[z i0 , z i1 ], where z i0 and z i1 Each is a binary indicator variable, z i0 With A i0 Correspondingly, is event A. i0 The binary indicator variable, z i1 With A i1 Correspondingly, is event A. i1 The two binary indicator variables are used to determine the i-th walking signaling trajectory, while the i-th walking signaling trajectory is composed of a vector z formed by its two binary indicator variables. i This refers to the "vector of binary indicator variables for the true categories," or the "indicator vector of the true categories." When A... i1 At the time of its establishment, A i0 This is definitely not true; at this point, z i1 =1,z i0 =0, which indicates that the i-th walking signaling trajectory is a walking trajectory with mobility impairment characteristics, so we can obtain P(y=c1|t i ) = 1, P(y = c0|t i )=0; when A i0 At the time of its establishment, A i1 This is necessarily not true, therefore z is not true at this time. i1 =0, z i0 =1, which means that the i-th walking signaling trajectory is not a walking trajectory with mobility impairment characteristics, so we can obtain P(y=c1|t i ) = 0, P(y = c0|t i ) = 1. z i0 and z i1 The value of z ranges from 0 to 1, i.e. ij ∈{0,1}, where j=0,1, and z i0 +z i1 =1, that is, z i0 and z i1 The sum is always 1. Furthermore, for the i-th walking signaling trajectory t... i The true category indicator vector, P(y=c0|t) i )=z i0 P(y=c0|t) i )=z i1 That is, at this time for t i Its posterior probability P(y=c j |t i The indicator vector z of its true classij It has been determined.

[0092] Understandably, due to the unlabeled dataset T U The true category of the walking signaling trajectory is unknown. It is typically necessary to define an estimate variable for the indicator vector of the true category, and continuously update it through iterative calculations to make it approach the true value, ultimately obtaining the unlabeled dataset T. U The true category of each walking trajectory sample.

[0093] It should be understood that, assuming the true class indicator vector z i z in i0 and z i1 During the iterative calculation process, P(y=c1|t) will be used. i ) and P(y=c0|t i Approximate substitution, as z i0 and z i1 The estimated value in the intermediate iteration process and at this time Where the superscript g denotes the g-th iteration, then It is z i0 Similarly, the estimated value in the g-th iteration... It is z i1 The estimated value in the g-th iteration is denoted as... and The vector formed is because and As an estimate of the "indicator vector of the true category" in the iteration, this embodiment will and The "probability-weighted category label" is called the i-th walking signaling trajectory. and The value range of is [0, 1], and it should satisfy . That is, the sum of all probability-weighted category labels for the i-th walking signaling trajectory is 1. Furthermore, if the i-th walking signaling trajectory t... i If the true category indicator vector is unknown, then t is called i If the sample is unlabeled, then t i The posterior probability P(y=c j |t i It is necessary to obtain its initial value through certain assumptions.

[0094] Further, obtaining the probability-weighted class labels for the labeled dataset and the unlabeled dataset respectively includes: obtaining a first correspondence between posterior probability data and the indicator vector of the true class; obtaining a second correspondence between the indicator vector of the true class and the probability-weighted class label; determining the correspondence between posterior probability data and the probability-weighted class label based on the first and second correspondences; determining the probability-weighted class label of the labeled dataset based on the posterior probability data of the labeled dataset and the correspondence between the posterior probability data and the probability-weighted class label; and determining the probability-weighted class label of the unlabeled dataset based on the posterior probability data of the unlabeled dataset and the correspondence between the posterior probability data and the probability-weighted class label.

[0095] It should be noted that the posterior probability data refers to the walking signaling trajectory t. i The posterior probability P(y=c1|t) relative to the category "is a walking trajectory with mobility impairment characteristics" i ) and the posterior probability P(y=c0|t) relative to the category "walking trajectory not a feature of mobility impairment". i For labeled datasets, the true class indicator vector is usually calculated based on the posterior probability data. The first correspondence refers to the formula for calculating the true class indicator vector, i.e., z. i0 =P(y=c0|t) i ), z i1 =P(y=c0|t) i ), where z i0 and z i1 These are the binary indicator variables that constitute the indicator vector. The probability-weighted class label is usually determined based on the indicator vector of the true class. The second correspondence refers to the calculation formula for the probability-weighted class label. The walking signaling trajectory t i The probability-weighted category label is z i0 and z i1 For unlabeled datasets, the first correspondence refers to the formula for calculating the estimated values ​​of the indicator vectors of the true classes, i.e. and in, and These are the estimated values ​​of the binary indicator variables that constitute the indicator vector, respectively. The probability-weighted category label is determined based on the estimated value of the indicator vector of the true category. The second correspondence refers to the calculation formula for the probability-weighted category label. The walking signaling trajectory t... i The probability-weighted category label is and

[0096] Understandably, before training the model, it's necessary to initialize the probability-weighted class labels for each walking signaling trajectory in the training dataset. This is because there is a labeled dataset T. L All samples in the dataset are labeled and annotated. Let T be... L The set consisting of positively labeled samples is T. P And T L The set consisting of negatively labeled samples is T. RN For T P For each walking signaling trajectory in the set, the posterior probability of belonging to the category "walking trajectory with mobility impairment characteristics" is 1, that is, for the i-th walking trajectory sample t i i = 1, 2, ..., |T P |, where |T P | represents set|T P The number of elements in |, i.e., t i ∈T P .

[0097] In the specific implementation, for set T RN Each walking signaling trajectory t in n ∈T RN All are assigned initial probability-weighted class labels. and in, For set T P Each sample t in n ∈T RN They are all assigned initial probability-weighted category labels. and in, Generally, during model iteration, T P and T RN The probability-weighted categories do not change during the iteration process. Unlabeled dataset T U Each walking signaling trajectory in the dataset has an unknown true category, i.e., it is an unlabeled sample. Therefore, a negative assumption needs to be made initially: assuming the unlabeled dataset T... U Each walking signaling trajectory in the dataset belongs to the category of "not a correct walking trajectory sample," meaning it is initialized as a negative-label sample based on the assumption. Therefore, for the unlabeled dataset T... U For each walking signaling trajectory in the dataset, the posterior probability of belonging to the "not a correct walking trajectory sample" category is 1, that is, for the u-th walking signaling trajectory t... u P(y=c0|t) u ) = 1, P(y = c1|t u ) = 0, where u = 1, 2, ..., |T U |,|TU | represents the unlabeled dataset T U The number of elements, then the u-th walking signaling trajectory t u The probability-weighted category labels are sum and .

[0098] It should be understood that a trajectory recognition model is established by training a Naive Bayes classifier based on probability-weighted class labels and a dataset of new attribute variables, and updating the parameters of the Naive Bayes classifier using the expectation-maximization (EM) strategy. This embodiment selects the Naive Bayes classifier model as the initial basic classification model. First, an initial Naive Bayes classification model is established using a labeled dataset. Then, based on the Naive Bayes classifier model, the EM iterative method framework is applied to establish the iterative update formula for the classification model, entering the iterative training process, and finally obtaining the trajectory recognition model. This embodiment establishes a semi-supervised Naive Bayes classifier (trajectory recognition model) by using probability-weighted class labels, avoiding the "deviation" that can easily occur in self-training algorithms and improving model accuracy.

[0099] Step S40: Determine the trajectory recognition ratio based on the trajectory recognition model and the walking signaling trajectory of the target communication number.

[0100] It should be noted that the trajectory recognition model is the final recognition model that can identify whether each walking signaling trajectory belongs to the category of "walking trajectory with mobility impairment characteristics" or the category of "walking trajectory without mobility impairment characteristics". The target communication number is the communication number of the user to be judged / identified, and the trajectory recognition ratio refers to the proportion of walking signaling trajectories belonging to the "walking trajectory with mobility impairment characteristics" category under the target communication number out of all walking signaling trajectories.

[0101] Further, step S40 includes: inputting the walking signaling trajectory into the trajectory recognition model to obtain the trajectory recognition result corresponding to the target communication number; determining the number of mobility impairment trajectories corresponding to the target communication number based on the trajectory recognition result; and determining the trajectory recognition ratio corresponding to the target communication number based on the number of mobility impairment trajectories of the target communication number and the total number of trajectories.

[0102] It is understandable that the trajectory recognition result is the recognition result obtained by the trajectory recognition model for each walking trajectory signaling input under the target communication number. The number of mobility impairment trajectories refers to the number of walking signaling trajectories under the target communication number that belong to the category of "walking trajectory with mobility impairment characteristics" (Num). p The total number of trajectories refers to the number of walking signaling trajectories under the target communication number, Num.p / Num represents the trajectory recognition ratio.

[0103] In the specific implementation, the walking signaling trajectory under the target communication number is input into the trajectory recognition model. Based on the recognition results, the proportion of walking signaling trajectories under the target communication number that belong to the category of "walking trajectory with mobility impairment characteristics" is calculated to determine the category of the user corresponding to the target communication number.

[0104] Step S50: Determine the user category corresponding to the target communication number based on the trajectory recognition ratio and the ratio threshold.

[0105] It should be understood that the proportion threshold refers to the percentage of walking signaling trajectories belonging to the "walking trajectory with mobility impairment characteristics" category out of all walking signaling trajectories. It is typically the optimal threshold for a classification problem. In this embodiment, the proportion threshold is calculated based on a weighted harmonic average. There are two user categories: users with mobility impairments and non-users with mobility impairments. When the trajectory recognition proportion is greater than or equal to the proportion threshold, the user belonging to the target communication number is considered a "user with mobility impairment." When the trajectory recognition proportion is less than the proportion threshold, the user belonging to the target communication number is considered a "non-user with mobility impairment."

[0106] In this embodiment, a walking signaling trajectory dataset is acquired and divided into labeled and unlabeled datasets. The attribute variable dataset from the labeled dataset is then obtained. Continuous attribute variable data in the attribute variable dataset undergoes supervised discretization to obtain a new attribute variable dataset. Probability-weighted class labels are obtained from both the labeled and unlabeled datasets. A trajectory recognition model is established based on the probability-weighted class labels and the new attribute variable dataset. Based on the trajectory recognition model and the walking signaling trajectory of the target communication number, the trajectory recognition ratio is determined. Based on the trajectory recognition ratio and a ratio threshold, the user category corresponding to the target communication number is determined. This embodiment effectively applies chi-square optimal binning to the scenario of identifying users with mobility impairments based on walking trajectories from mobile phone signaling, making the interpretability of walking trajectory attribute variables stronger. Furthermore, this embodiment uses probability-weighted class labels during model training, improving model accuracy. In addition, this embodiment first identifies the walking trajectory category and then solves for the optimal ratio threshold to determine the user category, reducing identification bias caused by signaling errors and improving identification accuracy. This realizes a semi-supervised learning hierarchical identification mode from trajectory identification to user identification, accurately identifying users with mobility impairments.

[0107] Reference Figure 3 , Figure 3 This is a flowchart illustrating a second embodiment of a user category identification method according to the present invention.

[0108] Based on the above embodiments, step S20 includes:

[0109] Step S201: Determine the number of intervals based on the trajectory labels of the labeled dataset.

[0110] It should be noted that the number of intervals refers to the total number of intervals ultimately defined. In the labeled dataset T... L In this process, based on the trajectory label of each walking signal trajectory, the total number of intervals KFnum is set using whether it is a "walking trajectory with mobility impairment characteristics" as the dividing identifier.

[0111] It is understandable that the labeled dataset T is extracted. L The attribute variable data corresponding to the attribute variables of each walking signaling trajectory constitute the attribute variable dataset D. L Then, extract the continuous attribute variable data to form the continuous attribute variable dataset D. L,c The remaining discrete attribute variable data then constitute the discrete attribute variable dataset D. L,d D L,d =D L -D L ,c Similarly, for the unlabeled dataset T... U The corresponding attribute variable dataset D can also be obtained. U And the corresponding continuous attribute variable dataset D U,c and discrete attribute variable dataset D U,d D U,d =D U -D U,c .

[0112] Step S202: Transform the continuous attribute variable data of the labeled dataset into an attribute data queue, and determine the attribute data queue set based on the attribute data queue.

[0113] It should be understood that an attribute data queue is a data queue composed of attribute variable data, an attribute data queue group is a data queue group composed of attribute data queues, and an attribute data queue set is a data set composed of attribute data queue groups. The attribute data queue set is composed of attribute data queue groups, and the attribute data queue groups are composed of attribute data queues. The relationship between attribute data queues, attribute data queue groups, and attribute data queue sets is as follows: Figure 4 As shown, the attribute data queue set consists of several attribute data queue groups, and each attribute data queue group consists of several attribute data queues, as shown in Figure I1 to I2. n S1 to S2 represent n attribute data queues respectively. m Let S[] represent m attribute data queue groups, and S[] represent the attribute data queue set.

[0114] In the specific implementation, there will be a labeled dataset T. L Dataset D of continuous attribute variables L，c Each attribute is transformed into a maximum and minimum value, which is used as its own attribute data queue I. I contains the data value and partitioning identifier of each data, and is stored in the attribute data queue group S. Each attribute data queue group S contains all the data in a single attribute. All attribute data queue groups S constitute an attribute data queue set S[]. S[] is the data source that is finally provided to the chi-square optimal binning algorithm, containing all the data in all attributes.

[0115] Step S203: Sort the attribute data queues in the attribute data queue group in descending order, calculate the expected frequency matrix between the attribute data queues, and determine the chi-square value between the attribute data queues based on the expected frequency matrix.

[0116] In the specific implementation, the attribute data queues I in each attribute data queue group S are sorted from largest to smallest. Each attribute data queue I is iterated over, and the expected frequency matrix E between adjacent attribute data queues I is calculated. ij The calculation formula is shown below:

[0117]

[0118] In the formula, C j This represents the number of all values ​​in the j-th class, i.e. N represents the total number of data points. R i This represents the number of values ​​falling within the i-th interval. Next, calculate the chi-square value X between attribute data queues I. 2 The calculation formula is shown below:

[0119]

[0120] In the formula, E ij Let A represent the desired frequency matrix. ij Indicates an event.

[0121] Step S204: Merge the attribute data queues according to the chi-square value to obtain a new attribute data queue group. When the number of attribute data queues in the new attribute data queue group is equal to the number of intervals, generate a new attribute data queue set according to the new attribute data queue group.

[0122] In the specific implementation, the attribute data queue I with the smallest chi-square value is found and merged, thereby updating the attribute data queue group S. Steps S203 to S204 are repeated until the number of attribute data queues I is equal to the number of intervals KFnum. After each attribute data queue group S is updated, a new attribute data queue set S is finally obtained.

[0123] Step S205: Update the continuous attribute variable data of the labeled dataset according to the new attribute data queue set to obtain new continuous attribute variable data, and obtain new attribute variable data based on the new continuous attribute variable data and the discrete attribute variable data.

[0124] In the specific implementation, based on the intervals divided by the data in the attribute data queue set S[], the labeled dataset T is... L Dataset D of continuous attribute variables L,c and unlabeled dataset T U Dataset D of continuous attribute variables U,c The continuous variable is discretized by updating (discretizing it to the interval it lies in). Since D... L,c and D U,c Updated; attribute variable dataset D needs to be recalculated. L and D U D L =D L,c ∪D L,d D U =D U,c ∪D U,d Similarly, for the validation dataset containing continuous attribute variables and the dataset containing discrete attribute variables D... L,d The process involves updating (continuous values ​​are transformed into their corresponding intervals), thus completing the discretization of continuous variables.

[0125] In this embodiment, based on the trajectory labels of the labeled dataset, the number of intervals is determined. The continuous attribute variable data of the labeled dataset is transformed into an attribute data queue set. The attribute data queues in the attribute data queue set are sorted in descending order, and the chi-square value between the attribute data queues is calculated. The attribute data queues are merged according to the chi-square value to obtain a new attribute data queue group. When the number of attribute data queues in the new attribute data queue group is equal to the number of intervals, a new attribute data queue set is generated, and the continuous attribute variable data of the labeled dataset is updated to obtain new continuous attribute variable data. This embodiment effectively applies chi-square optimal binning to the scenario of identifying users with mobility impairments based on walking trajectories from mobile phone signaling, making the interpretability of walking trajectory attribute variables stronger and improving the accuracy of subsequent model building.

[0126] Reference Figure 5, Figure 5 This is a flowchart illustrating a third embodiment of a user category identification method according to the present invention.

[0127] Based on the first embodiment described above, step S30 includes:

[0128] Step S301: Obtain the probability-weighted category labels of the labeled dataset and the unlabeled dataset respectively. Based on the trajectory labels and the labeled dataset, determine the trajectory set for people with mobility impairments and the trajectory set for people without mobility impairments.

[0129] It should be noted that the set of tracks for people with mobility impairments is the collection of walking signaling tracks in the labeled dataset that belong to the category of "walking tracks with mobility impairment characteristics", while the set of tracks for people without mobility impairments is the collection of walking signaling tracks in the labeled dataset that belong to the category of "walking tracks that do not belong to mobility impairment characteristics".

[0130] Step S302: Based on the probability-weighted category labels and the set of trajectory sets for people with mobility impairments, determine the probability-weighted category labels of the trajectory sets for people with mobility impairments. Based on the probability-weighted category labels and the set of trajectory sets for people without mobility impairments, determine the probability-weighted category labels of the trajectory sets for people without mobility impairments.

[0131] It is understandable that the probability-weighted category label sum refers to the sum of the probability-weighted category labels corresponding to each walking signaling trajectory in the set. First, the probability-weighted category label sum of the walking signaling trajectories in the set of mobility impairment trajectories is calculated, and the calculation formula is as follows:

[0132]

[0133] In the formula, the iteration number g = 0, and T represents the training dataset. This represents the i-th walking signaling trajectory t in the trajectory set for people with mobility impairments. i The probability-weighted category labels are then calculated. Next, the sum of the probability-weighted category labels for walking signaling trajectories in the non-mobility disability trajectory set is calculated, as shown in the following formula:

[0134]

[0135] In the formula, the iteration number g = 0, and T represents the training dataset. This represents the i-th walking signaling trajectory t in the non-mobility disability trajectory set. i The probability-weighted category labels.

[0136] Step S303: Determine the first parameter of the Naive Bayes classifier based on the probability-weighted class label sum of the trajectory set of people with limited mobility, the probability-weighted class label sum of the trajectory set of people without limited mobility, and the third correspondence.

[0137] It should be understood that there are two important parameters that need to be calculated in the Naive Bayes classifier, namely the first parameter. With the second parameter Typically, these are estimated values; the first and second parameters can form a vector. The third correspondence is the formula for calculating the first parameter in the Naive Bayes classifier:

[0138]

[0139] In the formula, T represents the training dataset. Let C = {c0, c1} represent the probability-weighted category labels, where C = {c0, c1} represents the set of categories. In this embodiment, |C| = 2, representing the number of categories. The constant terms 1 and |C| added to the numerator and denominator are to avoid the numerator and denominator being 0. The first parameter represents P(c) j ), that is, c j Prior probability of a class. (i.e., when j=0) represents the initial estimate of the first parameter corresponding to class c0. (i.e., when j=1) represents the initial estimate of the first parameter corresponding to class c1. Substitute the relevant data into the third correspondence mentioned above to calculate the initial estimate of the first parameter.

[0140] Step S304: Determine the indicator data based on the new attribute variable dataset.

[0141] Further, step S304 includes: determining the correlation coefficient between attribute variables based on the new attribute variable dataset; filtering the continuous attribute variables and the discrete attribute variables based on the correlation coefficient and the correlation threshold to determine the feature attribute variables; determining the trajectory feature data corresponding to the labeled dataset based on the feature attribute variables; and determining the indication data based on the trajectory feature data and trajectory feature value data corresponding to the labeled dataset.

[0142] It should be noted that the correlation coefficient is used to measure the correlation between attribute variable data obtained after chi-square optimal binning. In this embodiment, it is necessary to select relatively independent attribute variables as features and remove attribute variables with high correlation. This can be done by setting a correlation threshold. The attribute variables with low correlation that are finally selected are the feature attribute variables. For example, among the five attribute variables Kine1, Kine2, Peo1, Peo2, and Peo3, if Peo2 has the highest correlation, then Kine1, Kine2, Peo1, and Peo3 are selected as feature attribute variables, which can also be considered as trajectory features. These four feature attribute variables can form a feature attribute vector F = (f {1} f {2} , ..., f{|F|} ) = (Kine1, Kine2, Peo1, Peo3), at this time |F| = 4.

[0143] It can be understood that trajectory feature data is the feature attribute vector F of each walking signaling trajectory, F = (f {1} f {2} , ..., f {|F|} Let f be the k-th trajectory feature. {k} We can determine the possible values ​​and the number of possible values ​​for each trajectory feature to obtain the set L of possible values ​​for the k-th feature component. k That is, trajectory feature value data, f {k，l} Representing trajectory features f {k} Let f be a possible value (the l-th value), where k = 1, 2, ..., |F|, l = 1, 2, ..., |L|. k |

[0144] It should be understood that the k-th trajectory feature of the i-th labeled walking signaling trajectory is denoted as... Based on t for each walking signaling trajectory i ∈T L Trajectory features, judging one by one from all possible values ​​of the k-th feature. Whether the values ​​of |T are equal, that is, for i = 1, 2, ..., |T L | and k = 1, 2, ..., |F| and l = 1, 2, ..., |L| k | Calculate the indicator function, i.e., the indicator data:

[0145]

[0146] Where i = 1, 2, ..., |T L |,k=1,2,...,|F|,l=1,2,...|L k |

[0147] Step S305: Determine the sum of trajectory products of the disabled trajectory set based on the indication data and the probability weighted category labels of the disabled trajectory set; determine the sum of trajectory products of the non-disabled trajectory set based on the indication data and the probability weighted category labels of the non-disabled trajectory set.

[0148] It should be noted that the trajectory product refers to the product of the exponential data and the probability-weighted category label of the same category. The sum of trajectory products is the sum of the trajectory products N of the walking signaling trajectories. k，l，j The calculation formula is shown below:

[0149]

[0150]

[0151] Where, N k,l,0 Corresponding to class c0, N k,l,1 Corresponding to class c1.

[0152] In the specific implementation, and for each value f {k,l} First, calculate the product of the indication data of each walking signaling trajectory with the probability-weighted category label of the same category. Then, sum the products of all walking trajectory samples to obtain the sum of the trajectory products of the disabled trajectory set and the non-disabled trajectory set.

[0153] Step S306: Determine the second parameter of the Naive Bayes classifier based on the sum of trajectory products of the set of disabled trajectories, the sum of trajectory products of the set of non-disabled trajectories, and the fourth correspondence.

[0154] It is understandable that the fourth correspondence is the same as the formula for calculating the second parameter in the Naive Bayes classifier:

[0155]

[0156] In the formula, T represents the training dataset. L represents the probability-weighted category label sum. k This represents the trajectory feature value data, with constant terms 1 and |L added to the numerator and denominator. k This is to avoid the numerator and denominator being 0. The second parameter represents P(f) {k，l} |c j ), that is, c j Trajectory features f under the premise of class occurrence {k} Calculate the conditional probability of the l-th possible value. Substitute the relevant data into the fourth correspondence mentioned above to calculate the initial estimate of the second parameter.

[0157] Step S307: When the first parameter and the second parameter satisfy the preset convergence condition, establish a trajectory recognition model based on the first parameter and the second parameter.

[0158] It should be understood that when either the first or second parameter does not meet the preset convergence condition, the probability-weighted category labels are updated based on the first and second parameters, and the process returns to the step of determining the probability-weighted category labels of the disabled trajectory set based on the probability-weighted category labels of the walking signaling trajectory and the disabled trajectory set. The preset convergence condition refers to... That is, the difference between the parameter value calculated in the current iteration and the parameter value obtained in the previous iteration is small enough (ε can be set to an extremely small value). When the first parameter and the second parameter simultaneously satisfy the preset convergence condition, the algorithm converges, the iteration stops, and the final trajectory recognition model is obtained. If at least one of the first parameter and the second parameter does not satisfy the preset convergence condition, based on the EM method framework, the next iteration is continued, the probability-weighted category label is updated, and the process returns to step S302. The first parameter and the second parameter are updated according to the updated probability-weighted category label until the first parameter and the second parameter simultaneously satisfy the preset convergence condition.

[0159] Understandably, the update of probability-weighted class labels falls into two cases. The first case involves the first update, i.e., the first iteration (g+1 = 1st iteration, g = 0), where the unlabeled dataset T... U Each walking signaling trajectory t u If no label is assigned, then for T respectively U Each unlabeled sample t in u Match the second parameter corresponding to each attribute variable data. Recorded as And for each unlabeled sample t u Find the corresponding data of all its attribute variables. The product of the products is recorded as follows: Right now:

[0160]

[0161] in, Corresponding to class c0, Corresponding to class c1, j = 0, 1. Then, for each unlabeled sample t... u ,Will and Substituting into the following formula, we can calculate the "probability-weighted class label" updated in the (g+1)th iteration. The calculation expression is shown below:

[0162]

[0163] In the formula, The first case involves updating the probability-weighted label in the first iteration of the walking trajectory sample, where g = 0. The second case occurs if the update is not the first time, but in the (g+1)th iteration of EM (i.e., g+1 ≥ 2, g ≥ 1). In this case, for each unlabeled sample t... u Match the second parameter corresponding to each attribute variable data. (The value of the second parameter in the g-th iteration), denoted as And for each unlabeled sample tu Find the corresponding data of all its attribute variables. The product of the products, and for each unlabeled sample t u Calculate the product of the corresponding matching results for all walking trajectory feature values, and denote the product as . Right now:

[0164]

[0165] In the formula, Corresponding to class c0, Corresponding to class c1, j = 0, 1. For each unlabeled sample t u ,Will and Substituting into the following formula, we can calculate the "probability-weighted class label" updated in the (g+1)th iteration. The calculation expression is shown below:

[0166]

[0167] In the formula, This refers to the probability-weighted label updated in the (g+1)th iteration of the walking trajectory samples, where g ≥ 1. Furthermore, That is, the probability-weighted label updated in the (g+1)th iteration. It is calculated using the parameters of the Naive Bayes classifier updated at the g-th time.

[0168] In this embodiment, probability-weighted class labels are obtained from both labeled and unlabeled datasets. Based on these probability-weighted class labels and the new attribute variable dataset, a Naive Bayes classifier is trained. The parameters of the Naive Bayes classifier are then updated using an expectation-maximization (EM) strategy to establish a trajectory recognition model. This embodiment, based on the EM method framework, uses probability-weighted class labels during model training to establish a semi-supervised Naive Bayes classifier, thereby improving model accuracy.

[0169] Reference Figure 6 , Figure 6 This is a flowchart illustrating the fourth embodiment of a user category identification method according to the present invention.

[0170] Based on the first embodiment described above, step S50 includes:

[0171] Step S501: Obtain the identification frequency data corresponding to the target communication number, determine the precision and recall based on the identification frequency data, and obtain the weighted harmonic average corresponding to the target communication number based on the precision and the recall.

[0172] It should be noted that after the semi-supervised model identifies the walking trajectory with mobility impairment characteristics, due to the certain randomness error in mobile phone signaling, the ratio of the number of "walking trajectories with mobility impairment characteristics" identified by the classification algorithm for each user to which the communication number belongs to may be less than or equal to 1. The closer it is to 1, the more accurate the model is. The deviation that does not reach 1 is mainly caused by the random error of the signaling data. Therefore, this embodiment uses a validation dataset to verify and evaluate the model.

[0173] Understandably, based on the recognition results of the validation set, the frequency of the following information regarding walking trajectory data is first counted: the number of cases where "is a walking trajectory with mobility impairment characteristics" is predicted as "is a walking trajectory with mobility impairment characteristics" (TP). tr The number of "walking tracks that are not characteristic of mobility impairment" predicted as "walking tracks that are not characteristic of mobility impairment" (TN) tr The number of "walking tracks that are not characteristic of mobility impairment" predicted as "walking tracks that are characteristic of mobility impairment" (FP) tr The number of "walking tracks that are characteristic of mobility impairment" predicted as "walking tracks that are not characteristic of mobility impairment" (FN) tr Next, the accuracy P of the classification of "walking trajectories with mobility impairment characteristics" was approximated using statistical frequency. tr With recall rate R tr The calculation formula is shown below:

[0174]

[0175] Then, based on the calculated accuracy P tr With recall rate R tr Calculate the weighted harmonic mean F1 tr The calculation formula is shown below:

[0176]

[0177] Verification has shown that the weighted harmonic mean F1 tr The higher the value, the more effective the model. Therefore, it is necessary to find the optimal scale threshold based on the weighted harmonic mean.

[0178] In the specific implementation, the total number of walking signaling trajectories Num under the i-th target communication number is calculated based on the identification results of the walking signaling trajectories corresponding to the target communication number, and the number Num that is identified as "walking trajectories with mobility impairment characteristics" is counted. p And the number of walking tracks identified as "not characteristic of mobility impairment" (Num) n Set k = 0.1, 0.2, 0.3, 0.4, ..., 0.9. For each value of k, determine the trajectory recognition ratio Num for each target communication number.p The relationship between / Num and k, when Num p When / Num≥k, the user belonging to the target communication number is considered to be a user with mobility impairment. p When / Num < k, the user belonging to the target communication number is considered to be a non-mobility-impaired user. Next, the following correlation frequencies (identification frequency data) are calculated regarding the communication number; the number TP predicted as "is a mobility-impaired user". num The number of "non-mobility-impaired users" predicted as "non-mobility-impaired users" (TN) num The number of "non-mobility users" predicted as "mobility users" (FP) num The number of users who are "users with mobility impairments" predicted as "users without mobility impairments" (FN) num Based on these identification frequency data, the accuracy P is calculated. num and recall rate R num And based on the accuracy P num and recall rate R num Calculate the corresponding weighted harmonic mean F1 num .

[0179] Step S502: Determine the proportion threshold based on the weighted harmonic mean, wherein the proportion threshold is the maximum weighted harmonic mean among the weighted harmonic means.

[0180] In the specific implementation, the highest value is taken as the weighted harmonic average F1. num The k corresponding to the (maximum weighted harmonic mean) is used as the final proportional threshold.

[0181] Step S503: When the trajectory recognition ratio of the target communication number is greater than or equal to the ratio threshold, the user category corresponding to the target communication number is determined to be a user with mobility impairment; when the trajectory recognition ratio of the target communication number is less than the ratio threshold, the user category corresponding to the target communication number is determined to be a user without mobility impairment.

[0182] In practice, when the trajectory recognition ratio is greater than or equal to the ratio threshold, the user to which the target communication number belongs is considered to be a "user with mobility impairment". When the trajectory recognition ratio is less than the ratio threshold, the user to which the target communication number belongs is considered to be a "non-mobility impairment user".

[0183] In this embodiment, identification frequency data corresponding to the target communication number is acquired. Based on the identification frequency data, precision and recall are determined. Based on precision and recall, a weighted harmonic mean corresponding to the target communication number is obtained. The proportion threshold is the maximum weighted harmonic mean among the weighted harmonic means. When the trajectory identification proportion of the target communication number is greater than or equal to the proportion threshold, the user category corresponding to the target communication number is determined to be a user with mobility impairment. When the trajectory identification proportion of the target communication number is less than the proportion threshold, the user category corresponding to the target communication number is determined to be a user without mobility impairment. This embodiment first identifies the walking trajectory category and then solves for the optimal proportion threshold to determine the user category, reducing identification bias caused by signaling errors and improving identification accuracy. It realizes a semi-supervised learning hierarchical identification mode from trajectory identification to user identification.

[0184] Furthermore, embodiments of the present invention also propose a storage medium storing a user category identification program, wherein the user category identification program, when executed by a processor, implements the steps of the user category identification method as described above.

[0185] Reference Figure 7 , Figure 7 This is a structural block diagram of the first embodiment of the user category identification device of the present invention.

[0186] like Figure 7 As shown, the user category identification device proposed in this embodiment of the invention includes:

[0187] The data labeling module 10 is used to acquire a walking signaling trajectory dataset and divide the walking signaling trajectory dataset into a labeled dataset and an unlabeled dataset. The labeled dataset includes user labels and trajectory labels.

[0188] The model building module 20 is used to obtain the attribute variable dataset of the labeled dataset, and to perform supervised discretization on the continuous attribute variable data in the attribute variable dataset to obtain a new attribute variable dataset.

[0189] The model building module 20 is also used to obtain the probability-weighted category labels of the labeled dataset and the unlabeled dataset respectively, and to establish a trajectory recognition model based on the probability-weighted category labels and the new attribute variable dataset.

[0190] User identification module 30 is used to determine the trajectory identification ratio based on the trajectory identification model and the walking signaling trajectory of the target communication number.

[0191] The user identification module 30 is further configured to determine the user category corresponding to the target communication number based on the trajectory identification ratio and the ratio threshold.

[0192] In this embodiment, the chi-square optimal binning can be effectively applied to the scenario of identifying users with mobility impairments based on walking trajectories obtained from mobile phone signaling. This makes the interpretability of walking trajectory attribute variables stronger. Furthermore, the use of probability-weighted category labels during model training improves model accuracy. In addition, this embodiment first identifies the walking trajectory category and then solves for the optimal proportion threshold to determine the user category, reducing identification bias caused by signaling errors and improving identification accuracy. This realizes a semi-supervised learning hierarchical identification mode from trajectory identification to user identification.

[0193] In one embodiment, the model building module 20 is further configured to determine the number of intervals based on the trajectory labels of the labeled dataset;

[0194] The continuous attribute variable data of the labeled dataset is transformed into an attribute data queue;

[0195] Based on the attribute data queue, an attribute data queue set is determined, wherein the attribute data queue set is composed of attribute data queue groups, and the attribute data queue groups are composed of the attribute data queues.

[0196] Sort the attribute data queues in the attribute data queue group in descending order, and calculate the expected frequency matrix between the attribute data queues;

[0197] Based on the expected frequency matrix, determine the chi-square value between the attribute data queues;

[0198] The attribute data queues are merged according to the chi-square value to obtain a new attribute data queue group;

[0199] When the number of attribute data queues in the new attribute data queue group is equal to the number of intervals, a new attribute data queue set is generated based on the new attribute data queue group.

[0200] Based on the new attribute data queue set, the continuous attribute variable data of the labeled dataset is updated to obtain new continuous attribute variable data;

[0201] New attribute variable data is obtained based on the new continuous attribute variable data and the new discrete attribute variable data.

[0202] In one embodiment, the user identification module 30 is further configured to input the walking signaling trajectory into the trajectory identification model to obtain the trajectory identification result corresponding to the target communication number;

[0203] Based on the trajectory recognition results, determine the number of mobility impairment trajectories corresponding to the target communication number;

[0204] The trajectory recognition ratio corresponding to the target communication number is determined based on the number of mobility impairment trajectories of the target communication number and the total number of trajectories.

[0205] In one embodiment, the user identification module 30 is further configured to acquire identification frequency data corresponding to the target communication number, determine precision and recall based on the identification frequency data, and obtain the weighted harmonic average corresponding to the target communication number based on the precision and recall.

[0206] Based on the weighted harmonic mean, a proportion threshold is determined, wherein the proportion threshold is the maximum weighted harmonic mean among the weighted harmonic means;

[0207] When the trajectory recognition ratio of the target communication number is greater than or equal to the ratio threshold, the user category corresponding to the target communication number is determined to be a user with mobility impairment;

[0208] When the trajectory recognition ratio of the target communication number is less than the ratio threshold, the user category corresponding to the target communication number is determined to be a non-mobile disabled user.

[0209] In one embodiment, the model building module 20 is further configured to obtain a first correspondence between the posterior probability data and the indicator vector of the true class;

[0210] Obtain the second correspondence between the indicator vector of the true category and the probability-weighted category label;

[0211] Based on the first correspondence and the second correspondence, determine the correspondence between the posterior probability data and the probability-weighted category labels;

[0212] Based on the posterior probability data of the labeled dataset and the correspondence between the posterior probability data and the probability-weighted class labels, the probability-weighted class labels of the labeled dataset are determined.

[0213] Based on the posterior probability data of the unlabeled dataset and the correspondence between the posterior probability data and the probability-weighted class labels, the probability-weighted class labels of the unlabeled dataset are determined.

[0214] In one embodiment, the model building module 20 is further configured to determine a set of trajectories for people with mobility impairments and a set of trajectories for people without mobility impairments based on the trajectory labels and the labeled dataset;

[0215] Based on the probability-weighted category labels and the set of trajectory data for people with mobility impairments, determine the probability-weighted category labels of the set of trajectory data for people with mobility impairments; based on the probability-weighted category labels and the set of trajectory data for people without mobility impairments, determine the probability-weighted category labels of the set of trajectory data for people without mobility impairments.

[0216] The first parameter of the Naive Bayes classifier is determined based on the probability-weighted category labels of the set of trajectories with limited mobility, the probability-weighted category labels of the set of trajectories with limited mobility, and the third correspondence.

[0217] Based on the new attribute variable dataset, determine the indicator data;

[0218] Based on the indicated data and the probability-weighted category labels of the set of trajectories for people with mobility impairments, determine the sum of the trajectory products of the set of trajectories for people with mobility impairments; based on the indicated data and the probability-weighted category labels of the set of trajectories for people without mobility impairments, determine the sum of the trajectory products of the set of trajectories for people without mobility impairments.

[0219] The second parameter of the Naive Bayes classifier is determined based on the sum of the product of the trajectories of the set of trajectories with limited mobility, the sum of the product of the trajectories of the set of trajectories with limited mobility, and the fourth correspondence.

[0220] When the first parameter or the second parameter does not meet the preset convergence condition, the probability weighted category label is updated according to the first parameter and the second parameter, and the process returns to the step of determining the probability weighted category label of the set of disabled trajectories based on the probability weighted category label of the walking signaling trajectory and the set of disabled trajectories.

[0221] When the first parameter and the second parameter satisfy a preset convergence condition, the trajectory recognition model is established based on the first parameter and the second parameter.

[0222] In one embodiment, the model building module 20 is further configured to determine the correlation coefficient between attribute variables based on the new attribute variable dataset, wherein the attribute variables include continuous attribute variables and discrete attribute variables;

[0223] Based on the correlation coefficient and correlation threshold, the continuous attribute variables and the discrete attribute variables are filtered to determine the feature attribute variables;

[0224] Based on the aforementioned feature attribute variables, determine the trajectory feature data corresponding to the labeled dataset;

[0225] Based on the trajectory feature data and trajectory feature value data corresponding to the labeled dataset, the indication data is determined.

[0226] It should be understood that the above are merely illustrative examples and do not constitute any limitation on the technical solutions of the present invention. In specific applications, those skilled in the art can make settings as needed, and the present invention does not impose any restrictions on this.

[0227] It should be noted that the workflow described above is merely illustrative and does not limit the scope of protection of this invention. In practical applications, those skilled in the art can select some or all of the workflow to achieve the purpose of this embodiment according to actual needs, and no restrictions are imposed here.

[0228] The sequence numbers of the above embodiments of the present invention are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0229] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as read-only memory (ROM) / RAM, magnetic disk, optical disk) and includes several instructions to cause a terminal device (which may be a mobile phone, computer, server, or network device, etc.) to execute the methods described in the various embodiments of the present invention.

[0230] The above are merely preferred embodiments of the present invention and do not limit the scope of the patent. Any equivalent structural or procedural transformations made based on the description and drawings of the present invention, or direct or indirect applications in other related technical fields, are similarly included within the scope of patent protection of the present invention.

Claims

1. A user category identification method, characterized in that, The user category identification method includes: Obtain a pedestrian signaling trajectory dataset and divide the pedestrian signaling trajectory dataset into a labeled dataset and an unlabeled dataset, wherein the labeled dataset includes user labels and trajectory labels; Obtain the attribute variable dataset of the labeled dataset, and perform supervised discretization on the continuous attribute variable data in the attribute variable dataset based on the chi-square optimal binning algorithm to obtain a new attribute variable dataset; Obtain the probability-weighted category labels of the labeled dataset and the unlabeled dataset respectively, and establish a trajectory recognition model based on the probability-weighted category labels and the new attribute variable dataset; Based on the trajectory recognition model and the walking signaling trajectory of the target communication number, the trajectory recognition ratio is determined; Based on the trajectory recognition ratio and the ratio threshold, the user category corresponding to the target communication number is determined; The step of obtaining the probability-weighted class labels for the labeled dataset and the unlabeled dataset includes: obtaining a first correspondence between posterior probability data and indicator vectors of the true classes; obtaining a second correspondence between indicator vectors of the true classes and probability-weighted class labels; determining a correspondence between posterior probability data and probability-weighted class labels based on the first and second correspondences; determining the probability-weighted class labels for the labeled dataset based on the posterior probability data of the labeled dataset and the correspondence between the posterior probability data and probability-weighted class labels; and determining the probability-weighted class labels for the unlabeled dataset based on the posterior probability data of the unlabeled dataset and the correspondence between the posterior probability data and probability-weighted class labels. The method for establishing a trajectory recognition model based on the probability-weighted category labels and the new attribute variable dataset includes: determining a set of trajectories for people with mobility impairments and a set of trajectories for people without mobility impairments based on the trajectory labels and the labeled dataset; determining the sum of probability-weighted category labels for the set of trajectories for people with mobility impairments based on the probability-weighted category labels and the set of trajectories for people with mobility impairments; determining the sum of probability-weighted category labels for the set of trajectories for people without mobility impairments based on the probability-weighted category labels and the set of trajectories for people without mobility impairments; and determining the first probability-weighted category label of the Naive Bayes classifier based on the sum of probability-weighted category labels for the set of trajectories for people with mobility impairments, the sum of probability-weighted category labels for the set of trajectories for people without mobility impairments, and a third correspondence. The system takes one parameter; determines indication data based on the new attribute variable dataset; determines the sum of trajectory products of the disabled trajectory set based on the indication data and the probability-weighted category labels of the disabled trajectory set; determines the sum of trajectory products of the non-disabled trajectory set based on the indication data and the probability-weighted category labels of the non-disabled trajectory set; determines the second parameter of the Naive Bayes classifier based on the sum of trajectory products of the disabled trajectory set, the sum of trajectory products of the non-disabled trajectory set, and a fourth correspondence; and establishes the trajectory recognition model based on the first parameter and the second parameter when the first parameter and the second parameter satisfy a preset convergence condition.

2. The method as described in claim 1, characterized in that, The attribute variable dataset includes continuous attribute variable data and discrete attribute variable data. The supervised discretization of the continuous attribute variable data in the attribute variable dataset to obtain a new attribute variable dataset includes: The number of intervals is determined based on the trajectory labels in the labeled dataset. The continuous attribute variable data of the labeled dataset is transformed into an attribute data queue; Based on the attribute data queue, an attribute data queue set is determined, wherein the attribute data queue set is composed of attribute data queue groups, and the attribute data queue groups are composed of the attribute data queues. Sort the attribute data queues in the attribute data queue group in descending order, and calculate the expected frequency matrix between the attribute data queues; Based on the expected frequency matrix, determine the chi-square value between the attribute data queues; The attribute data queues are merged according to the chi-square value to obtain a new attribute data queue group; When the number of attribute data queues in the new attribute data queue group is equal to the number of intervals, a new attribute data queue set is generated based on the new attribute data queue group. Based on the new attribute data queue set, the continuous attribute variable data of the labeled dataset is updated to obtain new continuous attribute variable data; New attribute variable data is obtained based on the new continuous attribute variable data and the new discrete attribute variable data.

3. The method as described in claim 1, characterized in that, The determination of the trajectory recognition ratio based on the trajectory recognition model and the walking signaling trajectory of the target communication number includes: The walking signaling trajectory is input into the trajectory recognition model to obtain the trajectory recognition result corresponding to the target communication number; Based on the trajectory recognition results, determine the number of mobility impairment trajectories corresponding to the target communication number; The trajectory recognition ratio corresponding to the target communication number is determined based on the number of mobility impairment trajectories of the target communication number and the total number of trajectories.

4. The method as described in claim 3, characterized in that, The user categories include users with mobility impairments and users without mobility impairments. Determining the user category corresponding to the target communication number based on the trajectory recognition ratio and ratio threshold includes: Obtain the identification frequency data corresponding to the target communication number, determine the precision and recall based on the identification frequency data, and obtain the weighted harmonic average corresponding to the target communication number based on the precision and the recall. Based on the weighted harmonic mean, a proportion threshold is determined, wherein the proportion threshold is the maximum weighted harmonic mean among the weighted harmonic means; When the trajectory recognition ratio of the target communication number is greater than or equal to the ratio threshold, the user category corresponding to the target communication number is determined to be a user with mobility impairment; When the trajectory recognition ratio of the target communication number is less than the ratio threshold, the user category corresponding to the target communication number is determined to be a non-mobile disabled user.

5. The method as described in claim 1, characterized in that, The method further includes: If the first parameter or the second parameter does not meet the preset convergence condition, the probability weighted category label is updated according to the first parameter and the second parameter, and the process returns to the step of determining the probability weighted category label of the disabled trajectory set based on the probability weighted category label of the walking signaling trajectory and the disabled trajectory set.

6. The method as described in claim 1, characterized in that, The step of determining the indicator data based on the new attribute variable dataset includes: Based on the new attribute variable dataset, determine the correlation coefficients between the attribute variables, which include continuous attribute variables and discrete attribute variables; Based on the correlation coefficient and correlation threshold, the continuous attribute variables and the discrete attribute variables are filtered to determine the feature attribute variables; Based on the aforementioned feature attribute variables, determine the trajectory feature data corresponding to the labeled dataset; Based on the trajectory feature data and trajectory feature value data corresponding to the labeled dataset, the indication data is determined.

7. A user category identification device, characterized in that, The user category identification device includes: The data annotation module is used to acquire a pedestrian signaling trajectory dataset and divide the pedestrian signaling trajectory dataset into a labeled dataset and an unlabeled dataset, wherein the labeled dataset includes user labels and trajectory labels; The model building module is used to obtain the attribute variable dataset of the labeled dataset, and based on the chi-square optimal binning algorithm, to perform supervised discretization on the continuous attribute variable data in the attribute variable dataset to obtain a new attribute variable dataset. The model building module is also used to obtain the probability-weighted category labels of the labeled dataset and the unlabeled dataset respectively, and to build a trajectory recognition model based on the probability-weighted category labels and the new attribute variable dataset; The user identification module is used to determine the trajectory identification ratio based on the trajectory identification model and the walking signaling trajectory of the target communication number; The user identification module is further configured to determine the user category corresponding to the target communication number based on the trajectory identification ratio and the ratio threshold. The model building module is also used to obtain a first correspondence between the posterior probability data and the indicator vector of the true class; Obtain the second correspondence between the indicator vector of the true category and the probability-weighted category label; Based on the first correspondence and the second correspondence, determine the correspondence between the posterior probability data and the probability-weighted category labels; Based on the posterior probability data of the labeled dataset and the correspondence between the posterior probability data and the probability-weighted class labels, the probability-weighted class labels of the labeled dataset are determined. Based on the posterior probability data of the unlabeled dataset and the correspondence between the posterior probability data and the probability-weighted class labels, the probability-weighted class labels of the unlabeled dataset are determined. The model building module is also used to determine the trajectory set for people with mobility impairments and the trajectory set for people without mobility impairments based on the trajectory labels and the labeled dataset; Based on the probability-weighted category labels and the set of trajectory data for people with mobility impairments, determine the probability-weighted category labels of the set of trajectory data for people with mobility impairments; based on the probability-weighted category labels and the set of trajectory data for people without mobility impairments, determine the probability-weighted category labels of the set of trajectory data for people without mobility impairments. The first parameter of the Naive Bayes classifier is determined based on the probability-weighted category labels of the set of trajectories with limited mobility, the probability-weighted category labels of the set of trajectories with limited mobility, and the third correspondence. Based on the new attribute variable dataset, determine the indicator data; Based on the indicated data and the probability-weighted category labels of the set of trajectories for people with mobility impairments, determine the sum of the trajectory products of the set of trajectories for people with mobility impairments; based on the indicated data and the probability-weighted category labels of the set of trajectories for people without mobility impairments, determine the sum of the trajectory products of the set of trajectories for people without mobility impairments. The second parameter of the Naive Bayes classifier is determined based on the sum of the product of the trajectories of the set of trajectories with limited mobility, the sum of the product of the trajectories of the set of trajectories with limited mobility, and the fourth correspondence. When the first parameter and the second parameter satisfy a preset convergence condition, the trajectory recognition model is established based on the first parameter and the second parameter.

8. A user category identification device, characterized in that, The device includes: a memory, a processor, and a user category identification program stored in the memory and executable on the processor, the user category identification program being configured to implement the steps of the user category identification method as described in any one of claims 1 to 6.

9. A storage medium, characterized in that, The storage medium stores a user category identification program, which, when executed by a processor, implements the steps of the user category identification method as described in any one of claims 1 to 6.

Images