An online and offline deeply integrated campus one-card self-service management method

Abstract

The present application relates to the field of wisdom campus and artificial intelligence technology, and more particularly to a kind of online and offline deep integration's campus one-card self-service management method. Including the following steps: S1: preset time monitoring window, according to time monitoring window obtains the multi-source data information and reference data of campus one-card, constructs the behavior trajectory tensor and thought trajectory tensor of student user according to multi-source data information;S2: according to the behavior trajectory tensor and thought trajectory tensor of student user, calculate behavior motif and thought motif;According to thought motif, construct virtual time sequence space, and calculate the thought matching degree of virtual time sequence space;S3: according to the behavior motif of student user, construct the dynamic behavior portrait vector of student user;According to dynamic behavior portrait vector and similarity, obtain three kinds of group portrait vector.The present application is matched with group through thought and behavior aspects and obtains the self-service of campus one-card.

Description

Technical Field

[0001] This invention relates to the fields of smart campus and artificial intelligence technology, and in particular to a method for managing campus card self-service that deeply integrates online and offline methods. Background Technology

[0002] As a core infrastructure of smart campuses, the campus card system has accumulated massive amounts of student behavior data, including consumption records, access control information, and library borrowing records. Currently, when using the campus card system, it often fails to adaptively recommend services to student users. Existing profiling methods are mainly divided into two categories. The first category is the statistical feature method, which describes users by calculating aggregate indicators such as total consumption amount and card swipe frequency. This method is simple and efficient, but it completely loses the sequential and rhythmic information of behavior over time, cannot distinguish between users with regular routines and those with random fluctuations, and is also slow to react to short-term behavioral changes. The second category is the sequence modeling method, which directly inputs the raw behavioral data of continuous time slots into a recurrent neural network and extracts the hidden states as user representations. This method preserves the temporal structure, but has the following drawbacks: First, it is too sensitive to noise in the raw data and is prone to overfitting; second, the model output lacks interpretability and is difficult to correspond to the behavior pattern of leaving early and returning late; third, when the monitoring window is long, the sequence length can reach thousands, resulting in high computational overhead and making it difficult to deploy on a large scale in real time on campus. Summary of the Invention

[0003] To overcome the shortcomings of campus cards in adapting to service pushes, this invention provides a campus card self-service management method that deeply integrates online and offline services.

[0004] The technical solution of this invention is: a campus card self-service management method that deeply integrates online and offline methods, comprising the following steps: S1: Preset time monitoring window, obtain multi-source data information and reference data of campus card according to the time monitoring window, and construct student user behavior trajectory tensor and thought trajectory tensor according to the multi-source data information; S2: Calculate the behavioral and thought sequences based on the student user's behavioral trajectory tensor and thought sequence tensor; construct a virtual temporal space based on the thought sequences, and calculate the thought matching degree of the virtual temporal space; S3: Construct a dynamic behavior profile vector of student users based on the behavioral sequence of the student users; calculate the similarity of multiple student users based on reference data; and obtain a first group profile vector, a second group profile vector, and a third group profile vector based on the dynamic behavior profile vector and the similarity. S4: Based on the first group portrait vector, the second group portrait vector, the third group portrait vector, and the thinking matching degree, the probability of the current student user's service demand and the recommendation confidence are obtained by calculating the comprehensive matching degree; S5: Adaptively push campus card information to current student users based on the probability of their service needs and the recommendation confidence.

[0005] Preferably, the preset time monitoring window acquires multi-source data information and reference data of the campus card based on the time monitoring window, and constructs a student user's behavior trajectory tensor and thought trajectory tensor based on the multi-source data information, including: presetting a time monitoring window based on the campus card data system refresh time, acquiring multi-source data information and reference data from the campus card with the time monitoring window as the period, the multi-source data information including consumption data information, access control data information, borrowing data information, and click data information, the reference data including multi-source data information of other student users as reference data for the current student user, constructing a behavior trajectory tensor of the current student user at different times based on the current student user's consumption data information, access control data information, and borrowing data information, the behavior trajectory tensor including behavior duration, behavior category, and behavior characteristics, and constructing a thought trajectory tensor of the current student user at different times based on the current student user's click data information, the thought trajectory tensor including click browsing duration, browsing category, and browsing characteristics.

[0006] Preferably, the step of calculating the behavioral and cognitive order based on the student user's behavioral trajectory tensor and thought trajectory tensor includes: expanding the current student user's behavioral trajectory tensor according to behavioral categories and behavioral characteristics to obtain behavioral feature vectors at different time intervals within a duration; expanding the current student user's thought trajectory tensor according to browsing categories and browsing characteristics to obtain thought feature vectors at different time intervals within a duration; pre-setting an order length parameter; constructing behavioral and cognitive order vectors based on the behavioral feature vectors, thought feature vectors, and the order length parameter; obtaining the behavioral and cognitive order frequencies of all student users' behavioral and cognitive order vectors at different time intervals based on reference data; obtaining a behavioral order index and a behavioral order frequency histogram based on the behavioral order frequency; and obtaining a cognitive order index and a cognitive order frequency histogram based on the cognitive order frequency.

[0007] Preferably, the step of obtaining the behavior sequence index and behavior sequence frequency histogram based on the behavior sequence frequency, and obtaining the thought sequence index and thought sequence frequency histogram based on the thought sequence frequency, includes: constructing a behavior sequence codebook based on the behavior sequence frequency, constructing a thought sequence codebook based on the thought sequence frequency, calculating the spatial distance between the behavior sequence vector and the behavior sequence codebook using the nearest neighbor criterion for the current element, using the codeword with the smaller spatial distance as the index of the element in the current behavior sequence vector, and merging behavior sequence vectors with the same behavior sequence index into one behavior sequence index class; and calculating the spatial distance between the thought sequence vector and the thought sequence codebook using the nearest neighbor criterion for the current element. The spatial distance between the left and right codebooks of the current element is calculated. The codeword with the smaller spatial distance is used as the index of the element in the current thought sequence vector. Thought sequence vectors with the same thought sequence index are merged into one thought sequence index class. The sequence frequency of the thought sequence index class within different time gaps in the current time monitoring window is calculated based on the number of time gaps. The sequence frequency of the thought sequence index is a thought sequence frequency histogram. The sequence frequency of the behavior sequence index class within different time gaps in the current time monitoring window is calculated based on the number of time gaps. The sequence frequency of the behavior sequence index is a behavior sequence frequency histogram. The formula for calculating the sequence frequency is: ; in, The fundamental sequence frequency, This represents the total number of base sequences for student users within the time monitoring window. This is the sequence number of the current radix. For student users in the first A basal index This is an indicator function; it takes the value 1 if the condition within the parentheses is true, and 0 otherwise. This is the index class number of the base order.

[0008] Preferably, the step of constructing a virtual temporal space based on the thought sequence and calculating the thought matching degree of the virtual temporal space includes: obtaining a standard thought sequence through a standard campus card recommendation sequence, wherein the standard thought sequence includes the standard thought sequence obtained after clicking on all recommended information on the campus card line; constructing a virtual temporal space of the current student user's operations on the campus card line based on the current student user's thought sequence; using the thought sequence frequency corresponding to the thought sequence index in the thought sequence frequency histogram as a weight; using the current student user's thought sequence as the compared time series; using the standard thought sequence as the comparison time series; and calculating the thought matching degree between the comparison time series and the compared time series, wherein the thought matching degree calculation formula is: ; in, To compare time series and the time series being compared The degree of matching of thinking, To compare time series, equivalent to the current student user's thought process sequence, For the time series being compared, it is equivalent to the standard thought order. To compare time series Virtual spatial points in time intervals For the time series being compared in Virtual spatial points in time intervals For the time series being compared The The index class number is in The fundamental sequence frequency of the time slot, The total number of time intervals. This is the sequence number of the current time slot. It is an exponential function. Let be the square of the Euclidean norm. This is the Gaussian kernel bandwidth parameter.

[0009] Preferably, the step of constructing a dynamic behavior profile vector for the student user based on the student user's behavior sequence includes: extracting the behavior sequence frequency histogram and behavior sequence index obtained from the current student user's behavior sequence at different time intervals; obtaining local sequence distribution feature vectors for the current student user at different time intervals; obtaining memory state vectors for each time interval using memory neural networks from the local sequence distribution feature vectors; calculating attention weights for each time interval based on the memory state vectors and the corresponding frequencies of the behavior sequence frequency histogram using an attention mechanism; and calculating the dynamic behavior profile vector for the current student user based on the memory state vectors and the attention weights for each time interval, wherein the calculation formulas for attention weights and dynamic behavior profile vectors are as follows: ; ; in, For attention weights, For the attention mechanism parameter vector, This is the weight matrix for the attention mechanism. for The memory state vector at each moment, This represents the average value of the order frequency in the order frequency histogram. This is the attention mechanism bias vector. It is a hyperbolic tangent activation function, calculated element-wise. For vector transpose, For regularization operations, To correct the activation function of the linear unit, specifically to find its maximum value, This is a dynamic behavioral profile vector for current student users. This is the weight matrix of the fully connected layer. This is the bias vector of the fully connected layer. The total number of time intervals. This is the sequence number of the current time slot. This is the normalized representation vector.

[0010] Preferably, the step of calculating the dynamic behavior profile vector of the current student user based on the memory state vector and the attention weight of each time interval includes: calculating the representation vector of the current student user in the entire time monitoring window using a weighted summation algorithm based on the memory state vector of the current student user in each time interval and the attention weight of each time interval; normalizing the representation vector to obtain a normalized representation vector; and mapping the normalized representation vector to obtain the dynamic behavior profile vector of the current student user. The dynamic behavior profile vector includes the dynamic behavior evolution of the current student user within the time monitoring window and obtains the contribution level of different time intervals through an attention mechanism.

[0011] Preferably, the step of calculating the similarity of multiple student users based on reference data, and obtaining the first group portrait vector, the second group portrait vector, and the third group portrait vector based on the dynamic behavior portrait vector and the similarity, includes: pre-setting a sample number, wherein the sample number includes a pre-set number of student users used for sampling reference within the current time monitoring window; extracting reference data of the pre-set sample number; obtaining reference dynamic behavior portrait vectors of the pre-set sample number within the current time monitoring window; calculating the pairwise similarity between the reference dynamic behavior portrait vectors of the pre-set sample number in the reference data using a cosine similarity algorithm; pre-setting a similarity fitting range; and matching the similarity of two sample dynamic behavior portrait vectors with the similarity fitting range. By comparing and filtering samples with high similarity, three different sample datasets are obtained, including a first sample dataset, a second sample dataset, and a third sample dataset. The dynamic behavior profile vectors in the three different sample datasets are fitted to obtain a first group profile vector, a second group profile vector, and a third group profile vector with a current preset number of samples. The first group profile vector includes the average center of the evolution of all student behaviors in the first sample dataset, the second group profile vector includes the average center of the evolution of all student behaviors in the second sample dataset, and the third group profile vector includes the average center of the evolution of all student behaviors in the third sample dataset.

[0012] Preferably, the step of obtaining the service demand probability and recommendation confidence of the current student user by calculating the comprehensive matching degree based on the first group portrait vector, the second group portrait vector, the third group portrait vector, and the thinking matching degree includes: calculating the first matching degree, the second matching degree, and the third matching degree by using the dynamic behavior portrait vector of the current student user and the first group portrait vector, the second group portrait vector, and the third group portrait vector through exponentially decaying Euclidean distance; calculating the comprehensive matching degree of the current student user based on the first matching degree, the second matching degree, the third matching degree, and the thinking matching degree; calculating the service demand probability and recommendation confidence degree based on the comprehensive matching degree, wherein the service demand probability includes the probability that the current student needs to push messages, and the recommendation confidence degree includes the message category that the current student needs to push messages to; and performing self-service and service recommendations for the current student's campus card based on the service demand probability and recommendation confidence degree, wherein the calculation formulas for the service demand probability and recommendation confidence degree are: ; ; ; in, The first match degree, For the second degree of matching, The third degree of matching, For mindset matching degree, To serve the probability of demand, To recommend the confidence level, For steepness parameters, This is the offset threshold. For overall matching accuracy, Scaling factor This is the midpoint of the probability.

[0013] Preferably, the adaptive push of campus card information to the current student user based on the current student user's service demand probability and recommendation confidence further includes: determining self-service recommendations based on the current student user's service demand probability through the campus card; and determining the self-service recommendation category based on the current student user's recommendation confidence through the campus card. The self-service recommendation categories include balance reporting, course borrowing, and daily life categories. The self-service recommendation judgment condition is: if the current student user's service demand probability is greater than a preset demand threshold and the recommendation confidence is greater than half of the maximum recommendation confidence, then self-service is provided to the campus card, and the recommendation category is the daily life category. If the probability of a current student user's service demand is greater than the preset demand threshold and the recommendation confidence is less than half of the maximum recommendation confidence, then the campus card will be provided with self-service, and the recommended categories will be balance reporting and course borrowing. If the probability of a current student user's service demand is less than the preset demand threshold and the recommendation confidence is less than half of the maximum recommendation confidence, then the campus card will be silenced.

[0014] Compared with the prior art, the present invention has the following beneficial effects: 1. This invention clusters and summarizes behavioral patterns over consecutive short periods of time, smoothing the random fluctuations in the original data into stable patterns and effectively filtering out the interference of occasional noise. 2. This method, on the one hand, statistically analyzes the total proportion of various typical short-term patterns throughout the entire observation period to characterize user behavior habits; on the other hand, it performs time-series coding on the pattern distribution within local time periods to capture the migration of behavior habits over time. After adaptive fusion of the two types of information, it can reflect both the user's lifestyle and seasonal changes, providing a more comprehensive basis for subsequent service recommendations. 3. The user features extracted by this method can reflect behavioral dynamics and have stable noise resistance. When used to predict service demand for loss reporting and balance reminders, the accuracy is significantly better than existing statistical methods and simple sequence methods. At the same time, it is directly related to specific short-term behavioral patterns, making the reasons for push decisions clear and traceable, and meeting the campus management requirements for interpretability. Attached Figure Description

[0015] Figure 1 This is a flowchart illustrating the online-offline integrated campus card self-service management method of the present invention. Figure 2 This is a schematic diagram of the structure of the campus card self-service management method that deeply integrates online and offline services according to the present invention. Detailed Implementation

[0016] The present invention will be described in detail below with reference to the accompanying drawings and specific embodiments, but this is not intended to limit the present invention.

[0017] A campus card self-service management method that deeply integrates online and offline methods, such as... Figure 1 and 2 As shown, it includes the following steps: S1: Preset time monitoring window, obtain multi-source data information and reference data of campus card based on the time monitoring window, and construct student user behavior trajectory tensor and thought trajectory tensor based on multi-source data information; S2: Calculate the behavioral and thought sequences based on the student user's behavioral trajectory tensor and thought sequence tensor; construct a virtual temporal space based on the thought sequence and calculate the thought matching degree of the virtual temporal space; S3: Construct dynamic behavior profile vectors for student users based on their behavioral patterns; calculate the similarity of multiple student users based on reference data; and obtain the first group profile vector, the second group profile vector, and the third group profile vector based on the dynamic behavior profile vector and the similarity. S4: Based on the first group portrait vector, the second group portrait vector, the third group portrait vector, and the thinking matching degree, the probability of the current student user's service demand and the recommendation confidence are obtained by calculating the comprehensive matching degree; S5: Adaptively push campus card information to current student users based on the probability of their service needs and the recommendation confidence.

[0018] Based on the campus card data system refresh time preset time monitoring window, multi-source data information and reference data from the campus card are acquired periodically within the time monitoring window. Multi-source data information includes consumption data information, access control data information, borrowing data information, and click data information. Reference data includes multi-source data information from other student users as reference data for the current student user. Based on the current student user's consumption data information, access control data information, and borrowing data information, a behavioral trajectory tensor of the current student user at different times is constructed. The behavioral trajectory tensor includes the duration of behavior, behavior category, and behavior characteristics. Based on the current student user's click data information, a thought trajectory tensor of the current student user at different times is constructed. The thought trajectory tensor includes the duration of click browsing, browsing category, and browsing characteristics.

[0019] It needs to be explained that the specific method for obtaining the time monitoring window is as follows: by analyzing the data logs of the campus card, the data refresh cycle is obtained, and the time monitoring window is based on the multiple of the data refresh cycle; the consumption data information records every transaction record of students using campus cards to pay in canteens, supermarkets, and vending machines, including transaction timestamps, merchant numbers, and consumption amounts; the access control data information records the card swiping records of students entering and exiting dormitory buildings, teaching buildings, laboratories, and libraries, including access timestamps, building numbers, and directions; the borrowing data information records the operation records of students borrowing and returning books in the library, including operation timestamps, book classification numbers, and the number of books borrowed; the click data information records the click behavior records of students on various push messages on the campus card mobile application and self-service terminals, including click timestamps, push message types, and click action types; the specific method for obtaining the behavior trajectory tensor is: discretizing continuous time into time slots of equal length. For example, in this embodiment, each time monitoring window is 10 minutes long, and is further divided into 5 time slots, each time slot being 2 minutes long; the behavior trajectory tensor is a three-dimensional number. The group consists of three dimensions: time slot number, behavior category, and behavior feature. Behavior categories include consumption, access control, and borrowing, coded by the numbers 1, 2, and 3 respectively. The behavior feature is a fixed-length vector containing two components: behavior duration and behavior characteristics. Behavior duration refers to the time span from the first occurrence to the last occurrence of the same behavior category within the current time slot, measured in seconds. For example, a student might have three consumption behaviors within a 2-minute time slot: the first consumption occurs 20 seconds after the time slot begins, and the last consumption occurs 11 seconds after the time slot begins. If the duration is 0 seconds, the duration of the behavior is 90 seconds. If no such behavior occurs within the time slot, the duration of the behavior is recorded as 0 seconds. Behavioral characteristics are defined according to different categories: For consumption behavior, behavioral characteristics include total consumption amount in yuan, number of transactions, average consumption amount, and merchant diversity count; for access control behavior, behavioral characteristics include number of passes, number of different buildings, and number of changes in entry and exit directions; for borrowing behavior, behavioral characteristics include number of books borrowed, entropy value of book category used to measure the dispersion of borrowed books and book return status. Borrowing behavior characteristics are represented by 0 and 1.

[0020] The current student user's behavior trajectory tensor is expanded according to behavior category and behavior characteristics to obtain behavior feature vectors at different time intervals within the duration. The current student user's thought trajectory tensor is expanded according to browsing category and browsing characteristics to obtain thought feature vectors at different time intervals within the duration. A pre-defined sequence length parameter is used to construct behavior sequence vectors and thought sequence vectors based on the behavior feature vectors, thought feature vectors, and sequence length parameter. Based on reference data, the behavior sequence frequency and thought sequence frequency of all student users' behavior sequence vectors and thought sequence vectors at different time intervals are obtained. Based on the behavior sequence frequency, behavior sequence indexes and behavior sequence frequency histograms are obtained. Based on the thought sequence frequency, thought sequence indexes and thought sequence frequency histograms are obtained.

[0021] It needs to be explained that the specific method for obtaining the behavioral feature vector is as follows: In each time interval, all behavioral feature values ​​of the three behavioral categories are concatenated into a one-dimensional vector. The specific method for obtaining the thought feature vector is as follows: There are three browsing categories, and all browsing feature values ​​of the three browsing categories are concatenated into a one-dimensional vector, which is the thought feature vector for that time interval. The specific method for presetting the sequence length parameter is: it is adjusted proportionally according to the time monitoring window. For example, if the total duration of the time monitoring window is 10 minutes, with a total of 5 time intervals, then the sequence length parameter is set to 2. The method for constructing the behavioral sequence vector is as follows: For each time interval... and If the sequence length parameter is greater than or equal to the current time interval, the behavior feature vector of the current time interval is sequentially concatenated with the behavior feature vector of the preceding consecutive time intervals minus the sequence length parameter. In this embodiment, the sequence length parameter is 2. Therefore, for the 2nd to 5th time intervals, each behavior sequence vector is formed by concatenating the behavior feature vector of the previous time interval and the behavior feature vector of the current time interval. For example, the behavior sequence vector of the 2nd time interval is equal to the behavior feature vector of the 1st time interval directly concatenated with the behavior feature vector of the 2nd time interval, and the behavior sequence vector of the 3rd time interval is equal to the behavior feature vector of the 2nd time interval concatenated with the behavior feature vector of the 3rd time interval.

[0022] A behavior-based codebook is constructed based on behavior-based codebook frequency, and a thought-based codebook is constructed based on thought-based codebook frequency. The spatial distance between the behavior-based codebook and its left and right codebooks is calculated using the nearest neighbor criterion. Codewords with smaller spatial distances are used as indices of elements in the current behavior-based codebook. Behavior-based codebooks with the same behavior-based codebook index are grouped into one behavior-based codebook index class. Similarly, the spatial distance between the thought-based codebook and its left and right codebooks is calculated using the nearest neighbor criterion. Codewords with smaller spatial distances are used as indices of elements in the current thought-based codebook. Thought-based codebooks with the same thought-based codebook index are grouped into one thought-based codebook index class. The codebook frequency of the thought-based codebook index class is calculated within different time gaps in the current time monitoring window based on the number of time gaps. The codebook frequency of the thought-based codebook index is represented by a thought-based codebook frequency histogram. The codebook frequency of the behavior-based codebook index is also calculated within different time gaps in the current time monitoring window based on the number of time gaps. The codebook frequency of the behavior-based codebook index is represented by a behavior-based codebook frequency histogram. The formula for calculating the codebook frequency is as follows: ; in, The fundamental sequence frequency, This represents the total number of base sequences for student users within the time monitoring window. This is the sequence number of the current radix. For student users in the first A basal index This is an indicator function; it takes the value 1 if the condition within the parentheses is true, and 0 otherwise. This is the index class number of the base order.

[0023] It needs to be explained that the specific meaning of the formula is to obtain the frequency of the motif in the motif index by calculating the total number of motifs and the characteristic function when the condition is met. The specific construction method of the behavioral motif codebook and the thought motif codebook is as follows: count the occurrence frequency of the behavioral motif vector and the thought motif vector of all users in different time slots, and construct the behavioral motif codebook and the thought motif codebook. The Euclidean distance is calculated as follows: regard the current behavioral motif vector and the codeword as the coordinates of two points in a high-dimensional space, calculate the sum of the squares of the differences in each dimension between the two points, and then take the square root. The specific calculation method of the spatial distance is as follows: sort the current motif vector and the current motif codebook synchronously, use the nearest neighbor criterion to map the motif vector to the current motif codebook, the mapping position is the most similar codeword in the motif codebook, and the index of this codeword is used as the index of the current motif vector.

[0024] The standard thinking sequence is obtained through the standard campus card recommendation sequence. This standard thinking sequence includes the standard thinking sequence obtained after clicking on all recommended information on the campus card website. A virtual time-series space of the current student user's operations on the campus card website is constructed based on the current student user's thinking sequence. The thinking sequence frequency corresponding to the thinking sequence index in the thinking sequence frequency histogram is used as the weight. The current student user's thinking sequence is used as the compared time series, and the standard thinking sequence is used as the comparison time series. The thinking matching degree is calculated between the comparison time series and the compared time series, where the formula for calculating the thinking matching degree is: ; in, To compare time series and the time series being compared The degree of matching of thinking, To compare time series, equivalent to the current student user's thought process sequence, For the time series being compared, it is equivalent to the standard thought order. To compare time series Virtual spatial points in time intervals For the time series being compared in Virtual spatial points in time intervals For the time series being compared The The index class number is in The fundamental sequence frequency of the time slot, The total number of time intervals. This is the sequence number of the current time slot. It is an exponential function. Let be the square of the Euclidean norm. This is the Gaussian kernel bandwidth parameter.

[0025] It needs to be explained that the specific meaning of the thinking matching degree formula is to calculate the virtual distance between the current student's actual thinking order and the standard thinking order using the Euclidean norm and Gaussian kernel, and use the virtual distance as the thinking matching degree. The comparison time series and the compared time series are aligned point by point through the same time interval, and multiple index classes will not appear in the same time interval, so the comparison time series and the compared time series do not need to be re-aligned and matched. The standard thinking order is specifically obtained by pushing a series of push messages to the campus card according to the student's standard daily routine and diet schedule, and generating a time series table of push messages according to the preset best recommendation strategy. For example, in this embodiment, 100 trained volunteer students are selected and required to click on every push message that appears when using the campus card online application. Regardless of the push content, the volunteers must click and view the complete content, and the performance of these 100 volunteers in the standard All click response behaviors under the recommended sequence, including browsing features such as timestamps of each click, click delay, browsing duration, and number of repeated clicks, are then used to construct a thought trajectory tensor from the click data of these volunteers in the same way as ordinary students. Standard thought sequences are then obtained through sequence construction and clustering steps. The virtual temporal space is constructed by embedding discrete thought sequence index sequences sequentially into a measurable geometric space. The Gaussian kernel bandwidth parameter is obtained through cross-validation. A fixed number of student records are randomly selected from historical data. Each record contains its thought sequence sequence and a label indicating an abnormal event. For different initial values ​​of the Gaussian kernel bandwidth parameter, such as 0.5, 1.0, 1.5, and 2.0, the thought matching degree is calculated, and a simple classifier is trained using the current matching degree as a feature. The Gaussian kernel bandwidth parameter that maximizes classification accuracy is selected. In this embodiment, the optimal value for the Gaussian kernel bandwidth parameter is 1.2.

[0026] The behavior sequence frequency histogram and behavior sequence index of the current student user are extracted at different time intervals to obtain local sequence distribution feature vectors for the current student user at different time intervals. These local sequence distribution feature vectors are then used to obtain memory state vectors for each time interval through a memory neural network. Based on the memory state vectors and the corresponding frequencies in the behavior sequence frequency histogram, an attention weight for each time interval is calculated using an attention mechanism. Finally, the dynamic behavior profile vector of the current student user is calculated based on the memory state vectors and the attention weights for each time interval. The formulas for calculating the attention weights and the dynamic behavior profile vector are as follows: ; ; in, For attention weights, For the attention mechanism parameter vector, This is the weight matrix for the attention mechanism. Let be the memory state vector at time t. This represents the average value of all behavioral sequence frequencies in the behavioral sequence frequency histogram. This is the attention mechanism bias vector. It is a hyperbolic tangent activation function, calculated element-wise. For vector transpose, For regularization operations, To correct the activation function of the linear unit, specifically to find its maximum value, This is a dynamic behavioral profile vector for current student users. This is the weight matrix of the fully connected layer. This is the bias vector of the fully connected layer. The total number of time intervals. This is the sequence number of the current time slot. This is the normalized representation vector.

[0027] It needs to be explained that the attention weight formula specifically means that the attention weight in the attention mechanism network is calculated using the attention mechanism parameter vector, the attention mechanism weight matrix, and the memory state vector. The attention weight includes the weight corresponding to each time interval in the behavior radix frequency histogram. The dynamic behavior profile vector formula specifically means that the dynamic behavior profile vector is calculated using the attention weight and the normalized representation vector; the attention mechanism parameter vector... The dimension is 32, and the weight matrix of the fully connected layer is... The dimension is 32*64, and the bias vector of the fully connected layer is... The dimension is 32, and the formula for calculating the normalized representation vector is: ; in, The normalized representation vector is the memory state vector at time t. The attention weights, attention mechanism parameter vector, attention mechanism weight matrix, attention mechanism bias vector, fully connected layer weight matrix, and fully connected layer bias vector are automatically trained from labeled historical data through supervised learning. Specifically, all parameters are uniformly initialized using the PyTorch framework with Xavier initialization. The initial values ​​for the attention mechanism bias vector and the fully connected layer bias vector are zero vectors. The attention mechanism parameter vector is obtained by sampling small random numbers using a uniform distribution. Training data is then acquired; in this embodiment, the training data includes anonymized campus card records and push logs from universities over the past 12 months. The Adam optimizer is used for training the data, with an initial learning rate set to... The batch size was set to 256, and the training lasted for 50 epochs. The attention mechanism parameter vector, attention mechanism weight matrix, attention mechanism bias vector, fully connected layer weight matrix, and fully connected layer bias vector were obtained through iterative optimization. The memory state vector was obtained by sequentially inputting the local motif distribution feature vector into a memory neural network in chronological order. In this embodiment, the memory neural network used is a Long Short-Term Memory (LSTM) network. The hidden state dimension of the LSTM network is preset to 64. The network parameters include input gate weights, forget gate weights, output gate weights, and cell state update weights. These parameters are learned from historical data through backpropagation during the training phase.

[0028] Based on the memory state vector and attention weight of the current student user in each time interval, a weighted summation algorithm is used to calculate the representation vector of the current student user in the entire time monitoring window. The representation vector is then normalized to obtain a normalized representation vector. The normalized representation vector is then mapped to obtain the dynamic behavior profile vector of the current student user. The dynamic behavior profile vector includes the evolution of the current student user's dynamic behavior within the time monitoring window, and the contribution level of different time intervals is obtained through an attention mechanism.

[0029] The system sets a preset sample size, which includes the number of student users to be used for sampling reference within the current time monitoring window. It extracts reference data for this preset sample size, obtains reference dynamic behavior profile vectors for this preset sample size within the current time monitoring window, calculates the pairwise similarity between these reference dynamic behavior profile vectors using a cosine similarity algorithm, sets a preset similarity fitting range, compares the similarity between two preset sample dynamic behavior profile vectors with the similarity fitting range, and filters out samples with high similarity. It then obtains three different sample datasets: a first sample dataset, a second sample dataset, and a third sample dataset. The system fits the dynamic behavior profile vectors from these three different sample datasets to obtain the first group profile vector, the second group profile vector, and the third group profile vector for the current preset sample size. The first group profile vector includes the average center of all student behavior evolution in the first sample dataset, the second group profile vector includes the average center of all student behavior evolution in the second sample dataset, and the third group profile vector includes the average center of all student behavior evolution in the third sample dataset.

[0030] It needs to be explained that, based on the three different sample datasets obtained, the specific construction methods for the first, second, and third group portrait vectors are as follows: The mean value for each dimension is calculated for each of the three different sample datasets. For example, if the first dataset contains 85 dynamic behavior portrait vectors, the first component of each of the 85 vectors is added together and divided by 85 to obtain the first component of the first group portrait vector. The second to 128th components are processed in the same way, ultimately resulting in a 128-dimensional mean vector. This mean vector is the first group portrait vector. The sample size... The specific preset method is as follows: log analysis is performed on the campus card of the current student user, and the group portrait vector of the campus card is filtered and fitted to find the optimal number of samples when the accuracy of the group portrait vector is the highest. The optimal number of samples is used as the preset number of samples. The specific calculation method of cosine similarity is: calculate the dot product and divide it by the norm length of the divisor and the dividend. The value range of cosine similarity is [0, 1]. The specific method of obtaining the similarity fitting range is to manually adjust the actual clustering effect. In this embodiment, the similarity fitting range is [-0.2, +0.2]. The first, second, and third matching degrees are calculated using exponentially decaying Euclidean distance from the current student user's dynamic behavior profile vector and the first, second, and third group profile vectors. The comprehensive matching degree of the current student user is then calculated based on the first, second, and third matching degrees and the mindset matching degree. The service demand probability and recommendation confidence are then calculated based on the comprehensive matching degree. The service demand probability includes the probability that the current student needs a push notification, and the recommendation confidence includes the message category that the current student needs to receive. Self-service and service recommendations are then provided to the current student's campus card based on the service demand probability and recommendation confidence. The formulas for calculating the service demand probability and recommendation confidence are as follows: ; ; ; in, The first match degree, For the second degree of matching, The third degree of matching, For mindset matching degree, To serve the probability of demand, To recommend the confidence level, For steepness parameters, This is the offset threshold. For overall matching accuracy, Scaling factor This is the midpoint of the probability.

[0031] It needs to be explained that the specific meaning of the service demand probability formula includes calculating the service probability using a linear weighted method of natural exponential for the first matching degree, second matching degree, third matching degree, and mindset matching degree. The specific meaning of the recommendation confidence formula is to obtain the sensitivity of the current student to the message category under the current service demand probability based on the performance of the service demand probability in different ranges of historical data, and to divide different message category intervals according to the sensitivity. The specific method of obtaining the steepness parameter is as follows: the preset steepness parameter is between [5, 10], the service demand probability of historical samples is calculated with different steepness parameters, and the error of the actual service demand and the service demand probability is compared. The service demand probability with the smallest error is obtained, and the steepness parameter with the smallest service demand probability is used as the current steepness parameter. The specific method of obtaining the offset threshold is as follows: calculate the comprehensive matching degree in historical samples, and obtain the offset threshold through the median. The shift threshold is set to 0.5 in this embodiment. The probability midpoint is obtained as follows: the initial value is set to 0.65. Based on the frequency distribution of service demand probability, the probability midpoint is adjusted according to the actual needs of current student users and campus card users. In this embodiment, the probability midpoint is set to 0.7. The scaling factor is obtained as follows: based on the probability midpoint of 0.7, the scaling factor is set to a value range of [5, 10]. The recommendation confidence is calculated under different scaling factors. The accuracy of the recommendation confidence is calculated by controlling variables. In this embodiment, the accuracy is the success rate of student users obtaining the desired recommended service when using the campus card self-service 10 times. The recommended service obtained by student users is obtained through the current student user's click records. The record with the highest accuracy is selected, and the scaling factor of the record with the highest accuracy is used as the scaling factor parameter.

[0032] The system determines self-service recommendations based on the probability of current student users' service needs using the campus card. It also determines the self-service recommendation category for current student users based on the recommendation confidence level using the campus card. The self-service recommendation categories include balance reporting, course borrowing, and daily life. The self-service recommendation criteria are as follows: if the probability of current student users' service needs is greater than the preset demand threshold and the recommendation confidence level is greater than half of the maximum recommendation confidence level, then self-service is provided through the campus card, and the recommended category is daily life. If the probability of a current student user's service demand is greater than the preset demand threshold and the recommendation confidence is less than half of the maximum recommendation confidence, then the campus card will be provided with self-service, and the recommended categories will be balance reporting and course borrowing. If the probability of a current student user's service demand is less than the preset demand threshold and the recommendation confidence is less than half of the maximum recommendation confidence, then the campus card will be silenced.

[0033] It needs to be explained that the maximum recommendation confidence score includes the maximum recommendation score of the message category when a student user pushes a message to a student based on the probability of service demand. The category of the pushed message is determined by comparing the recommendation confidence score and the maximum recommendation confidence score. The maximum recommendation confidence score is obtained by calculating the recommendation confidence score in the historical logs and obtaining the maximum recommendation confidence score in the historical logs through the maximum-minimum value method. The preset demand threshold includes classifying the service demands of different student users and obtaining the average service demand probability of different student users when demanding campus card self-service. When the current student user's service demand probability is less than the preset demand threshold and the recommendation confidence score is less than half of the maximum recommendation confidence score, no proactive service recommendation is made for campus card, i.e., a silent treatment is performed. For example, if the student's service demand probability is 0.55 and the recommendation confidence score is 0.20, both conditions do not meet the threshold requirements, so the current student's status is determined to be normal, and no reminders or service access are required.

[0034] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for managing campus card self-service that deeply integrates online and offline methods, characterized in that, Includes the following steps: S1: Preset time monitoring window, obtain multi-source data information and reference data of campus card according to the time monitoring window, and construct student user behavior trajectory tensor and thought trajectory tensor according to the multi-source data information; S2: Calculate the behavioral and thought sequences based on the student user's behavioral trajectory tensor and thought sequence tensor; construct a virtual temporal space based on the thought sequences, and calculate the thought matching degree of the virtual temporal space; S3: Construct a dynamic behavior profile vector of student users based on the behavioral sequence of the student users; calculate the similarity of multiple student users based on reference data; and obtain a first group profile vector, a second group profile vector, and a third group profile vector based on the dynamic behavior profile vector and the similarity. S4: Based on the first group portrait vector, the second group portrait vector, the third group portrait vector, and the thinking matching degree, the probability of the current student user's service demand and the recommendation confidence are obtained by calculating the comprehensive matching degree; S5: Adaptively push campus card information to current student users based on the probability of their service needs and the recommendation confidence.

2. The method for managing campus card self-service that deeply integrates online and offline methods according to claim 1, characterized in that, The preset time monitoring window acquires multi-source data and reference data from the campus card system. Based on this multi-source data, it constructs a student user's behavioral trajectory tensor and thought trajectory tensor. This includes: setting a preset time monitoring window based on the campus card system refresh time; acquiring multi-source data and reference data from the campus card system periodically within this window; the multi-source data includes consumption data, access control data, borrowing data, and click data; the reference data includes multi-source data from other student users as reference data for the current student user; constructing a behavioral trajectory tensor for the current student user at different times based on the current student user's consumption data, access control data, and borrowing data; and constructing a thought trajectory tensor for the current student user at different times based on the current student user's click data; the thought trajectory tensor includes the duration of clicks / browses, the browsing category, and browsing characteristics.

3. The method for managing campus card self-service that deeply integrates online and offline methods according to claim 1, characterized in that, The step of calculating behavioral and cognitive orderings based on student users' behavioral trajectory tensors and thought trajectory tensors includes: expanding the current student user's behavioral trajectory tensor according to behavioral categories and characteristics to obtain behavioral feature vectors at different time intervals within a given duration; expanding the current student user's thought trajectory tensor according to browsing categories and browsing characteristics to obtain thought feature vectors at different time intervals within a given duration; pre-setting an ordering length parameter; constructing behavioral and cognitive ordering vectors based on the behavioral feature vectors, thought feature vectors, and the ordering length parameter; obtaining behavioral and cognitive ordering frequencies of all student users' behavioral and cognitive ordering vectors at different time intervals based on reference data; obtaining behavioral ordering indices and behavioral ordering frequency histograms based on the behavioral ordering frequencies; and obtaining cognitive ordering indices and cognitive ordering frequency histograms based on the cognitive ordering frequencies.

4. The method for managing campus card self-service that deeply integrates online and offline methods according to claim 3, characterized in that, The step of obtaining the behavior motif index and behavior motif frequency histogram based on the behavior motif frequency, and obtaining the thought motif index and thought motif frequency histogram based on the thought motif frequency, includes: constructing a behavior motif codebook based on the behavior motif frequency, constructing a thought motif codebook based on the thought motif frequency, calculating the spatial distance between the behavior motif vector and the behavior motif codebook using the nearest neighbor criterion for the current element, using the codeword with the smaller spatial distance as the index of the element in the current behavior motif vector, and merging behavior motif vectors with the same behavior motif index into one behavior motif index class; calculating the spatial distance between the thought motif vector and the thought motif codebook using the nearest neighbor criterion. Calculate the spatial distance between the left and right codebooks of the current element, and use the codeword with the smaller spatial distance as the index of the element in the current thought sequence vector. Thought sequence vectors with the same thought sequence index are merged into one thought sequence index class. Calculate the sequence frequency of the thought sequence index class within different time gaps in the current time monitoring window based on the number of time gaps. The sequence frequency of the thought sequence index is a thought sequence frequency histogram. Calculate the sequence frequency of the behavior sequence index class within different time gaps in the current time monitoring window based on the number of time gaps. The sequence frequency of the behavior sequence index is a behavior sequence frequency histogram. The formula for calculating the sequence frequency is: ； in, The fundamental sequence frequency, This represents the total number of base sequences for student users within the time monitoring window. This is the sequence number of the current radix. For student users in the first A basal index This is an indicator function; it takes the value 1 if the condition within the parentheses is true, and 0 otherwise. This is the index class number of the base order.

5. The method for managing campus card self-service that deeply integrates online and offline methods according to claim 1, characterized in that, The step of constructing a virtual time-series space based on the thought sequence and calculating the thought matching degree of the virtual time-series space includes: obtaining a standard thought sequence through a standard campus card recommendation sequence, wherein the standard thought sequence includes the standard thought sequence obtained after clicking on all recommended information on the campus card line; constructing a virtual time-series space of the current student user's operations on the campus card line based on the current student user's thought sequence; using the thought sequence frequency corresponding to the thought sequence index in the thought sequence frequency histogram as a weight; using the current student user's thought sequence as the compared time series; using the standard thought sequence as the comparison time series; and calculating the thought matching degree between the comparison time series and the compared time series, wherein the thought matching degree calculation formula is: ； in, To compare time series and the time series being compared The degree of matching of thinking, To compare time series, equivalent to the current student user's thought process sequence, For the time series being compared, it is equivalent to the standard thought order. To compare time series Virtual spatial points in time intervals For the time series being compared in Virtual spatial points in time intervals Let $\mathbf$ be the order frequency of the $i$-th index class number of the time series $V_b$ being compared at time interval $t$. The total number of time intervals. This is the sequence number of the current time slot. It is an exponential function. Let be the square of the Euclidean norm. This is the Gaussian kernel bandwidth parameter.

6. The method for managing campus card self-service that deeply integrates online and offline methods according to claim 1, characterized in that, The step of constructing a dynamic behavior profile vector for student users based on the student user's behavior sequence includes: extracting the behavior sequence frequency histogram and behavior sequence index obtained from the current student user's behavior sequence at different time intervals; obtaining local sequence distribution feature vectors for the current student user at different time intervals; obtaining memory state vectors for each time interval using a memory neural network; calculating attention weights for each time interval based on the memory state vectors and the corresponding frequencies of the behavior sequence frequency histogram using an attention mechanism; and calculating the current student user's dynamic behavior profile vector based on the memory state vectors and the attention weights for each time interval. The calculation formulas for the attention weights and the dynamic behavior profile vector are as follows: ； ； in, For attention weights, For the attention mechanism parameter vector, This is the weight matrix for the attention mechanism. Let be the memory state vector at time t. This represents the average value of the order frequency in the order frequency histogram. This is the attention mechanism bias vector. It is a hyperbolic tangent activation function, calculated element-wise. For vector transpose, For regularization operations, To correct the activation function of the linear unit, specifically to find its maximum value, This is a dynamic behavioral profile vector for current student users. This is the weight matrix of the fully connected layer. This is the bias vector of the fully connected layer. The total number of time intervals. This is the sequence number of the current time slot. This is the normalized representation vector.

7. The method for managing campus card self-service that deeply integrates online and offline methods according to claim 6, characterized in that, The step of calculating the dynamic behavior profile vector of the current student user based on the memory state vector and the attention weight of each time interval includes: calculating the representation vector of the current student user in the entire time monitoring window using a weighted summation algorithm based on the memory state vector of the current student user in each time interval and the attention weight of each time interval; normalizing the representation vector to obtain a normalized representation vector; and mapping the normalized representation vector to obtain the dynamic behavior profile vector of the current student user. The dynamic behavior profile vector includes the dynamic behavior evolution of the current student user within the time monitoring window, and the contribution degree of different time intervals is obtained through an attention mechanism.

8. The method for managing campus card self-service that deeply integrates online and offline methods according to claim 1, characterized in that, The step of calculating the similarity of multiple student users based on reference data, and obtaining a first group portrait vector, a second group portrait vector, and a third group portrait vector based on the dynamic behavior profile vector and the similarity, includes: setting a preset sample size, which includes a preset number of student users for sampling reference within the current time monitoring window; extracting reference data of the preset sample size; obtaining reference dynamic behavior profile vectors of the preset sample size within the current time monitoring window; calculating the pairwise similarity between the reference dynamic behavior profile vectors of the preset sample size in the reference data using a cosine similarity algorithm; setting a preset similarity fitting range; and performing a similarity fitting within the preset similarity range on the dynamic behavior profile vectors of two samples. By comparison and selection of samples with high similarity, three different sample datasets are obtained, namely, a first sample dataset, a second sample dataset, and a third sample dataset. The dynamic behavior profile vectors in the three different sample datasets are fitted to obtain the first group profile vector, the second group profile vector, and the third group profile vector with the current preset number of samples. The first group profile vector includes the average center of the evolution of all student behaviors in the first sample dataset, the second group profile vector includes the average center of the evolution of all student behaviors in the second sample dataset, and the third group profile vector includes the average center of the evolution of all student behaviors in the third sample dataset.

9. The method for managing campus card self-service that deeply integrates online and offline methods according to claim 1, characterized in that, The step of obtaining the service demand probability and recommendation confidence of the current student user by calculating the comprehensive matching degree based on the first group portrait vector, the second group portrait vector, the third group portrait vector, and the thinking matching degree includes: calculating the first matching degree, the second matching degree, and the third matching degree by using the exponentially decaying Euclidean distance to calculate the dynamic behavior portrait vector of the current student user and the first group portrait vector, the second group portrait vector, and the third group portrait vector; calculating the comprehensive matching degree of the current student user based on the first matching degree, the second matching degree, the third matching degree, and the thinking matching degree; calculating the service demand probability and recommendation confidence degree based on the comprehensive matching degree; wherein the service demand probability includes the probability that the current student needs to push messages; and the recommendation confidence degree includes the message category that the current student needs to push messages to; and performing self-service and service recommendations for the current student's campus card based on the service demand probability and recommendation confidence degree, wherein the calculation formulas for the service demand probability and recommendation confidence degree are as follows: ； ； ； in, The first match degree, For the second degree of matching, The third degree of matching, For mindset matching degree, To serve the probability of demand, To recommend the confidence level, For steepness parameters, This is the offset threshold. For overall matching accuracy, Scaling factor This is the midpoint of the probability.

10. The method for managing campus card self-service that deeply integrates online and offline methods according to claim 1, characterized in that, The adaptive push of campus card information to the current student user based on the current student user's service demand probability and recommendation confidence also includes: determining self-service recommendations based on the current student user's service demand probability through the campus card, and determining the self-service recommendation category based on the current student user's recommendation confidence through the campus card. The self-service recommendation categories include balance reporting, course borrowing, and daily life categories. The self-service recommendation judgment condition is: if the current student user's service demand probability is greater than a preset demand threshold and the recommendation confidence is greater than half of the maximum recommendation confidence, then self-service is provided to the campus card, and the recommendation category is the daily life category. If the probability of a current student user's service demand is greater than the preset demand threshold and the recommendation confidence is less than half of the maximum recommendation confidence, then the campus card will be provided with self-service, and the recommended categories will be balance reporting and course borrowing. If the probability of a current student user's service demand is less than the preset demand threshold and the recommendation confidence is less than half of the maximum recommendation confidence, then the campus card will be silenced.

Images