Personalized comprehensive old-age care service recommendation method and system based on artificial intelligence big data

Abstract

The invention provides a personalized comprehensive old-age care service recommendation method and system based on artificial intelligence big data, and relates to the technical field of computer data processing and artificial intelligence. Travel data of a plurality of target users and browsing data of a community service APP are collected, the travel data comprise travel timestamps and in-out directions, and the travel timestamps and the in-out directions of the target users are acquired; the browsing data comprises a plurality of browsing records, browsing frequency and staying duration; constructing a behavior feature vector according to the record quantity of the travel data in the in-out direction in the preset time period, and constructing an interest feature vector according to the browsing data; respectively inputting the behavior feature vector and the interest feature vector into a neural network based on deep canonical correlation analysis to generate a fusion feature vector of each target user; dividing all the target users into a plurality of social groups based on the fused feature vectors; and based on the low-dimensional manifold space, the community activity information is sent to the edge node objects in the same social group, so that precise social recommendation for the solitary old people is realized.

Description

Technical Field

[0001] This application relates to the fields of computer data processing and artificial intelligence technology, and in particular to a method and system for recommending personalized comprehensive elderly care services based on artificial intelligence big data. Background Technology

[0002] As the aging population continues to grow, smart community-based elderly care has become a key area for improving the quality of life for seniors. Personalized service recommendation technology based on big data has broad application prospects in promoting the social participation of the elderly. By using intelligent algorithms to analyze the living habits and interests of seniors, it has significant social value in alleviating the growing social isolation and loneliness among this population.

[0003] Existing community-based elderly care service recommendation methods typically rely on basic demographic tags of the elderly or single-dimensional historical browsing records, primarily employing traditional collaborative filtering or rule-based matching algorithms for information delivery. These existing technologies largely treat users' online digital reading behavior and offline physical travel patterns as isolated data silos, mechanically matching based solely on simple click popularity or static tags, lacking in-depth correlation analysis of multi-source heterogeneous behavioral data.

[0004] However, this approach, relying solely on data from a single perspective, struggles to capture the deep-seated consistency between the thoughts and actions of the elderly. This can lead to recommended activities exceeding their actual capabilities or failing to align with their genuine, underlying social desires. Furthermore, existing methods lack effective identification of the potential social network topology within communities, failing to leverage the shared influence of active core members to engage isolated seniors. Consequently, recommendation services struggle to translate into genuine offline social interactions. Therefore, current technologies suffer from low recommendation accuracy and poor social integration due to the lack of effective integration of multi-dimensional behavioral features and social structure analysis. Summary of the Invention

[0005] The purpose of this application is to provide a personalized comprehensive elderly care service recommendation method and system based on artificial intelligence and big data, in order to solve the technical problems of low recommendation accuracy and poor social integration effect caused by the lack of effective integration of multi-dimensional behavioral features and social structure analysis in the existing technology.

[0006] Firstly, this application provides a method for recommending personalized comprehensive elderly care services based on artificial intelligence and big data, including:

[0007] We collect travel data and browsing data from community service apps from multiple target users. Travel data includes travel timestamps and directions of entry and exit, while browsing data includes multiple browsing records, browsing frequency, and dwell time.

[0008] Based on the number of records of travel directions within a preset time period, construct a behavioral feature vector, and based on browsing data, construct an interest feature vector.

[0009] The behavioral feature vector and interest feature vector are respectively input into a neural network based on deep canonical correlation analysis. By jointly optimizing the objective function of the neural network, a fused feature vector for each target user is generated.

[0010] Based on the fused feature vectors, Gaussian similarity between target users is calculated to construct a Laplacian matrix. By mapping the Laplacian matrix to a low-dimensional manifold space, a clustering algorithm is used to divide all target users into multiple social groups.

[0011] Based on low-dimensional manifold space, we identify edge node objects and central node objects in social groups, extract community activity information that central node objects frequently participate in, and send the community activity information to edge node objects within the same social group.

[0012] Optionally, the method further includes:

[0013] Obtain the generation time of each browsing record in the browsing data, and calculate the time weight coefficient of each browsing record based on the difference between the generation time and the current time;

[0014] Based on the browsing data, construct an interest feature vector, including:

[0015] Based on the browsing data, an interest feature vector is constructed using the time weighting coefficient of each browsing record.

[0016] Optionally, based on the browsing data, an interest feature vector is constructed using the time weight coefficient of each browsing record, including:

[0017] The content tags of browsing records in the browsing data of each target user are converted into semantic embedding vectors, and the basic attention of each browsing record is calculated based on the browsing frequency and dwell time in the browsing data.

[0018] The basic attention level is multiplied by the corresponding time weight coefficient to obtain the comprehensive weight value of each browsing record;

[0019] The semantic embedding vector of each browsing record is multiplied by the corresponding comprehensive weight value to obtain a weighted semantic vector. All weighted semantic vectors of the same target user are then summed to obtain the interest feature vector.

[0020] Optionally, the behavioral feature vector and interest feature vector are respectively input into a neural network based on deep canonical correlation analysis. By jointly optimizing the objective function of the neural network, a fused feature vector for each target user is generated, including:

[0021] A neural network based on deep canonical correlation analysis is constructed. The neural network includes a first mapping channel with a first network weight parameter and a second mapping channel with a second network weight parameter.

[0022] Input the behavioral feature vector into the first mapping channel to obtain the first feature vector, and input the interest feature vector into the second mapping channel to obtain the second feature vector;

[0023] Calculate the covariance of the first eigenvector and the second eigenvector, and calculate the Pearson correlation coefficient based on the covariance;

[0024] Construct an objective function with the negative of the Pearson correlation coefficient as the loss value, and use the gradient descent algorithm to jointly optimize the objective function until the improvement of the Pearson correlation coefficient is within a preset threshold. Then determine the convergence of the first network weight parameters and the second network weight parameters.

[0025] Using the converged weight parameters of the first network and the converged weight parameters of the second network, the behavioral feature vector and the interest feature vector are remapped to obtain the first target feature vector and the second target feature vector.

[0026] The first feature vector and the second feature vector of the target are concatenated to obtain the fused feature vector of each target user.

[0027] Optionally, an objective function is constructed using the negative of the Pearson correlation coefficient as the loss value. The objective function is then jointly optimized using a gradient descent algorithm until the increase in the Pearson correlation coefficient is within a preset threshold. The convergence of the first and second network weight parameters is then determined, including:

[0028] A preset number of training users are randomly selected, and the behavioral feature vector and interest feature vector of each training user are used as a training sample.

[0029] According to the preset arrangement, the first feature vector and the second feature vector corresponding to all training samples are stacked to obtain the first feature matrix and the second feature matrix.

[0030] The first and second feature matrices are decentered and standardized respectively to generate canonical correlation matrices.

[0031] Add all the diagonal elements of the canonical correlation matrix to obtain the sum of the Pearson correlation coefficients;

[0032] The negative of the sum of Pearson correlation coefficients is used as the loss value to construct the objective function;

[0033] Using the gradient descent algorithm, calculate the first partial derivative of the objective function with respect to the weight parameters of the first network and the second partial derivative with respect to the weight parameters of the second network.

[0034] Based on the first and second partial derivatives, the first and second network weight parameters are updated by subtraction, and the sum of the updated Pearson correlation coefficients is calculated until the increase in the Pearson correlation coefficient before and after the update is less than a preset threshold, at which point the convergence of the first and second network weight parameters is determined.

[0035] Optionally, based on the fused feature vectors, Gaussian similarity among target users is calculated to construct a Laplacian matrix. This Laplacian matrix is ​​then mapped to a low-dimensional manifold space, and a clustering algorithm is used to divide all target users into multiple social groups, including:

[0036] Calculate the Euclidean distance between any two fused feature vectors, and substitute each Euclidean distance into the Gaussian kernel function to obtain the Gaussian similarity between target users, so as to construct a Gaussian similarity matrix;

[0037] Based on the row sum of the Gaussian similarity matrix, construct a diagonal matrix, and perform matrix subtraction between the diagonal matrix and the Gaussian similarity matrix to obtain the Laplace matrix;

[0038] Eigenvalue decomposition is performed on the Laplacian matrix to obtain multiple eigenvalues. All eigenvalues ​​are arranged in ascending order to obtain an eigenvalue sequence. The first number of non-zero eigenvalues ​​of a predetermined dimension are selected from the eigenvalue sequence. Based on the eigenvectors corresponding to the non-zero eigenvalues, an eigenvector matrix is ​​constructed.

[0039] Each row of data in the feature vector matrix is ​​used as a coordinate point in a low-dimensional manifold space, and clustering is performed on all coordinate points in the low-dimensional manifold space to obtain multiple social groups, where the target users in each social group have the same cluster label.

[0040] Optionally, based on a low-dimensional manifold space, edge node objects and central node objects in the social group are identified, and information on frequently participated community activities of the central node objects is extracted. This community activity information is then sent to edge node objects within the same social group, including:

[0041] Based on the low-dimensional manifold space, the arithmetic mean of the coordinates of the points within each social group is calculated to obtain the center coordinates of each social group.

[0042] Calculate the spatial Euclidean distance between the coordinates of each target user in the same social group and the center coordinates. Target users whose spatial Euclidean distance is less than or equal to a preset distance threshold are designated as center node objects, and target users whose spatial Euclidean distance is greater than the preset distance threshold are designated as edge node objects.

[0043] The number of times the central node object participates in each community activity within a preset time period is counted, and the participation counts of all community activities are sorted in descending order to obtain the participation popularity sequence of each social group.

[0044] The community activity information is selected from the top preset number of community activities in the popularity sequence, and the terminal device identifier of the edge node object in each social group is obtained. The community activity information is then sent to the edge node object in the same social group through the terminal device identifier.

[0045] Secondly, this application provides a personalized comprehensive elderly care service recommendation system based on artificial intelligence and big data, including:

[0046] The data collection module is used to collect travel data and browsing data of community service apps from multiple target users. Travel data includes travel timestamps and directions of entry and exit, while browsing data includes multiple browsing records, browsing frequency, and dwell time.

[0047] The module is used to construct behavioral feature vectors based on the number of records of inbound and outbound directions within a preset time period, and to construct interest feature vectors based on browsing data.

[0048] The optimization module is used to input the behavioral feature vector and interest feature vector into the neural network based on deep canonical correlation analysis, and generate a fused feature vector for each target user by jointly optimizing the objective function of the neural network.

[0049] The building module is also used to calculate the Gaussian similarity between target users based on the fused feature vectors to construct the Laplacian matrix. By mapping the Laplacian matrix to a low-dimensional manifold space, a clustering algorithm is used to divide all target users into multiple social groups.

[0050] The determination module is used to identify edge node objects and central node objects in a social group based on a low-dimensional manifold space, extract information on community activities that the central node objects frequently participate in, and send the community activity information to the edge node objects within the same social group.

[0051] Thirdly, this application provides an electronic device, comprising:

[0052] Memory, used to store computer programs;

[0053] A processor is used to execute computer programs to implement the steps of the personalized comprehensive elderly care service recommendation method based on artificial intelligence big data as described in the first aspect above.

[0054] Fourthly, this application provides a computer-readable storage medium storing a computer program, which, when executed by a processor, can implement the steps of the personalized comprehensive elderly care service recommendation method based on artificial intelligence big data as described in the first aspect above.

[0055] The personalized comprehensive elderly care service recommendation method based on artificial intelligence and big data provided in this application achieves a digital representation of users' physical activity and digital interests by simultaneously collecting travel pattern data in the physical dimension and browsing interest data in the digital dimension, thus solving the problem that unstructured data is difficult to use for model calculation. This improves the accuracy of user profiles. It overcomes the deficiency of traditional clustering algorithms in discovering non-convex community structures in the original high-dimensional space, ensuring that the segmented groups have a high degree of homogeneity in their lifestyles and interests. It also solves the problem that existing recommendation services are difficult to translate into actual offline social behavior, improving the social integration effect of peripheral nodes.

[0056] Furthermore, this application constructs a deep canonical correlation analysis neural network with a dual-channel mapping structure to map user behavior feature vectors and interest feature vectors to a nonlinear space, and establishes an objective optimization mechanism with the inverse of the Pearson correlation coefficient as the loss function. Gradient descent is used to jointly optimize the dual-channel parameters until the network converges. This maximizes the correlation between the two types of features, and then concatenates the mapped features to generate a deep fusion feature vector that combines physical behavior patterns with digital interest preferences. This solves the problem of mismatch between recommendation activities and users' actual intentions or abilities caused by one-sided feature representation. Attached Figure Description

[0057] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0058] Figure 1 A flowchart illustrating the personalized comprehensive elderly care service recommendation method based on artificial intelligence big data provided in this application embodiment;

[0059] Figure 2 A flowchart illustrating a method for generating fused feature vectors provided in an embodiment of this application;

[0060] Figure 3 A flowchart illustrating a method for dividing all target users into multiple social groups, provided in an embodiment of this application;

[0061] Figure 4 A schematic diagram of the structure of a personalized comprehensive elderly care service recommendation system based on artificial intelligence big data provided in this application embodiment;

[0062] Figure 5 This is a schematic diagram of the hardware structure of the electronic device provided in the embodiments of this application. Detailed Implementation

[0063] In response to the problems of inaccurate recommendations and difficulties in social integration caused by the disconnect between online digital interests and offline physical capabilities in community-based elderly care scenarios, this application collects travel data and APP browsing data simultaneously to obtain multi-source heterogeneous information reflecting users' actions and thoughts. This maximizes the intrinsic consistency between behavior and interests in a non-linear space, thereby solving the problem of activity recommendations that do not match the actual wishes or abilities of the elderly due to the one-sidedness of traditional profiling.

[0064] Meanwhile, this application constructs a Laplace matrix and maps it to a low-dimensional manifold space to deeply explore the implicit social topology within the community, accurately identify core and peripheral nodes within like-minded social groups, and replaces mechanical matching with a strategy of core-driven peripheral integration, thereby achieving precise guidance for social integration and solving the technical bottleneck of existing recommendation services being difficult to translate into actual offline social behavior.

[0065] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0066] The core of this application is to provide a personalized comprehensive elderly care service recommendation method based on artificial intelligence and big data. A flowchart illustrating one specific implementation method is shown below. Figure 1 As shown, the method includes:

[0067] Step 101: Collect travel data and browsing data of community service apps from multiple target users. Travel data includes travel timestamps and directions of entry and exit, while browsing data includes multiple browsing records, browsing frequency, and dwell time.

[0068] In this step, travel data refers to the physical passage log information of users recorded by IoT sensing devices at community entrances and exits, used to represent users' activity patterns and travel rhythms in the real physical space. Browsing data refers to the interaction behavior logs generated by users on the community service mobile terminal application, used to represent users' reading interests and focus in the virtual digital space.

[0069] In this embodiment, the system connects to the property management database of Community A via an IoT data interface to read access control gate passage time logs within a specified time window in real time or at regular intervals to obtain travel data. This travel data specifically records user identification, card swipe timestamps, and entry / exit direction indicators. Simultaneously, the system connects to the backend server of the community service software via an application programming interface to synchronously extract the electronic bulletin board click and browsing records of the same group of users to obtain browsing data.

[0070] For example, for target user U, the collected travel data can be represented as one or more data items in the form of [10-01 07:30:00, Exit], indicating that user U swiped their card to leave the community at that time. The collected browsing data can be represented as one or more data items in the form of [Health Knowledge Section, 5, 300], indicating that user U clicked on the Health Knowledge Section 5 times within the statistical period, with a cumulative reading time of 300 seconds.

[0071] Step 102: Based on the number of records of inbound and outbound directions in the travel data within a preset time period, construct a behavioral feature vector, and based on the browsing data, construct an interest feature vector.

[0072] In this step, the behavioral feature vector refers to a numerical sequence representing the distribution of a user's travel activity across different time periods within a preset time frame. The interest feature vector refers to a numerical sequence representing a user's preference for and focus on different service content in the virtual space.

[0073] In this embodiment, a behavioral feature vector is first constructed based on travel data. A preset time period sequence is defined, assuming a day is divided into 24 time slots, with each hour as the unit. The collected travel data is traversed, and records belonging to the current statistical period are filtered out. For each time slot, the total number of records of the target user's entry and exit directions within that time slot is counted. The number of records obtained from all time slots is arranged in chronological order to form a one-dimensional vector, thus obtaining the behavioral feature vector.

[0074] Simultaneously, interest feature vectors are constructed based on browsing data. Multiple browsing records are analyzed, and the content tags, browsing frequency, and dwell time for each record are extracted. The content tags are converted into computer-readable numerical vectors, and the user's level of attention to the content is calculated based on browsing frequency and dwell time. By combining the numerical vectors with the level of attention, an interest feature vector reflecting the user's overall interests and preferences is generated.

[0075] Step 10201: Obtain the generation time of each browsing record in the browsing data, and calculate the time weight coefficient of each browsing record based on the difference between the generation time and the current time.

[0076] In this embodiment, the browsing data is first parsed, and the timestamp field of each browsing record is extracted to obtain the generation time. The current system time at the time of data processing is obtained. The time difference is calculated by subtracting the generation time from the current time. This time difference can be expressed in days or hours.

[0077] For example, if the current time is October 5th, and user U has two browsing records, the first record R1 was created on October 5th, meaning it was created on that day. The time difference is... The value is 0. The second record, R2, was generated on October 1st, and the time difference is 0. The time difference is 4 days. This time difference is substituted into the preset exponential decay function for calculation, as shown in formula (1):

[0078] (1)

[0079] in, This represents the time weighting coefficient. It is the base of the natural logarithm. This is a preset decay constant used to control the rate at which the weight decreases over time; its value is typically between 0 and 1. This represents the difference between the current time and the time when the browsing history was generated. A decay constant can be set. The time weighting coefficient is 0.1 for record R1. The time weighting coefficient is 1 for record R2. The weight is 0.67. It can be seen that the record of the day has the highest weight, and the weight of the record R2 decreases as time goes by.

[0080] Based on the browsing data, construct an interest feature vector, including:

[0081] Based on the browsing data, an interest feature vector is constructed using the time weighting coefficient of each browsing record.

[0082] Step 201: Convert the content tags of browsing records in the browsing data of each target user into semantic embedding vectors, and calculate the basic attention of each browsing record based on the browsing frequency and dwell time in the browsing data.

[0083] In this embodiment, a pre-built natural language processing (NLP) algorithm library is invoked. This library stores word vector dictionaries or context encoders generated based on massive corpora. Content tag text from browsing history is passed as input to the NLP algorithm library. Through matching retrieval or encoding calculations within the library, the corresponding semantic embedding vector is output. For example, for the content tag L1 "community free clinic," a 100-dimensional semantic embedding vector is obtained after transformation by the algorithm library. =[0.1,0.5,...]. Simultaneously calculate the basic attention level. Extract the browsing frequency and dwell time from this browsing record.

[0084] To eliminate the influence of dimensions, the browsing frequency and dwell time are normalized by dividing by the preset maximum browsing frequency and dwell time, respectively. The normalized frequency ratio is then multiplied by the dwell time ratio to calculate the basic attention level. For example, consider target user U, who has two browsing records R1 and R2. Record R1's content tag is "community free clinic," and the browsing frequency is... Five times, duration of stay The interval is 300 seconds. Set the maximum frequency. For 10 times, maximum duration The duration is 600 seconds. Therefore, the base attention level is calculated. It is 0.25.

[0085] Step 202: Multiply the basic attention score by the corresponding time weight coefficient to obtain the comprehensive weight value of each browsing record.

[0086] In practice, the time weight coefficient corresponding to each browsing step and the corresponding basic attention level are obtained. These two are multiplied together to obtain the comprehensive weight value. For example, for record R1, the basic attention level is... The time weighting coefficient is 0.25. If the value is 1, then the overall weight value is... This is the product of the two. For recording R², the base attention... Both are 0.25, but the time weighting coefficient is different. The overall weight value is 0.67. This is the product of the two. The final calculated result is the overall weight value for each browsing record.

[0087] Step 203: Multiply the semantic embedding vector of each browsing record by the corresponding comprehensive weight value to obtain a weighted semantic vector, and sum all the weighted semantic vectors of the same target user to obtain the interest feature vector.

[0088] In this embodiment, for each browsing record, the value of each dimension of its corresponding semantic embedding vector is multiplied by the record's overall weight value to obtain a weighted semantic vector. For example, for record R1, the semantic vector... Multiply by the overall weight This yields a weighted semantic vector. For record R2, the semantic vector Multiply by the overall weight This yields a weighted semantic vector. Next, iterate through all the browsing history of the target user and sum all the calculated weighted semantic vectors along their corresponding dimensions.

[0089] For example, and Adding them together, the final composite vector is generated. This is the interest feature vector of the target user. This vector comprehensively reflects the user's semantic content preferences and the dynamic strength of these preferences over time. Through this process, although the user visits the two sections with the same frequency and duration, the community free clinic is the most recently focused topic, so it has a larger proportion in the final feature vector, thus more accurately reflecting the user's current focus of interest.

[0090] Step 103: Input the behavioral feature vector and interest feature vector into the neural network based on deep canonical correlation analysis respectively, and generate the fused feature vector for each target user by jointly optimizing the objective function of the neural network.

[0091] In this step, the neural network for deep canonical correlation analysis refers to a neural network architecture for deep canonical correlation analysis (DCCA), which includes two deep nonlinear mapping channels used to learn the maximum correlation representation between two sets of heterogeneous view data. The objective function is the mathematical expression used to guide parameter optimization during network training.

[0092] like Figure 2 As shown, Figure 2 A flowchart illustrating a method for generating fused feature vectors provided in an embodiment of this application is shown.

[0093] Step 301: Construct a neural network based on deep canonical correlation analysis. The neural network includes a first mapping channel with a first network weight parameter and a second mapping channel with a second network weight parameter.

[0094] In this embodiment, a deep learning framework is used to build a deep canonical correlation analysis model. For example, a deep canonical correlation analysis network is constructed, with the first channel processing a 24-dimensional behavioral feature vector and the second channel processing a 100-dimensional interest feature vector. The output dimension of both channels is set to 10 dimensions.

[0095] The first mapping channel is designed as a fully connected structure comprising an input layer, several hidden layers, and an output layer. For example, the number of nodes in the input layer corresponds to the dimension of the behavioral feature vector, the hidden layers use a linear rectified function as the activation function, and the number of nodes in the output layer is set to the preset feature mapping dimension. Similarly, a structurally symmetrical second mapping channel is constructed, with the number of nodes in its input layer corresponding to the dimension of the interest feature vector, and the number of nodes in its output layer consistent with the first channel. The weight parameters of the first and second networks are initialized, typically using random initialization or identical initialization methods. Although these two channels receive input independently, they share the same optimization objective during training, thus forming a complete deep canonical correlation analysis network architecture.

[0096] The training process of a neural network can specifically include:

[0097] Network Model Construction: First, a deep neural network model is constructed, comprising a first mapping channel and a second mapping channel. The first mapping channel receives and processes behavioral feature vectors, and the second mapping channel receives and processes interest feature vectors. Both channels employ a multi-layer fully connected network structure, with each layer connected by a weight matrix and configured with a non-linear activation function to enable the network to handle complex non-linear relationships. The network weight parameters for both channels are initialized, and the number of nodes in the output layers of both channels is set to be the same; this number represents the dimension of the final mapped features.

[0098] Forward Propagation and Feature Mapping: During training, a batch of training samples is randomly selected from the user dataset. The behavioral feature vectors of these samples are input into the first mapping channel, and after multiple nonlinear transformations, the first feature vector matrix is ​​obtained. Simultaneously, the corresponding interest feature vectors are input into the second mapping channel, and after the same transformation process, the second feature vector matrix is ​​obtained. The output obtained at this point represents an abstract representation of the original data in the deep feature space.

[0099] Objective Function Construction: To measure the consistency of output features between two channels, a loss function based on the Pearson correlation coefficient needs to be constructed. First, the first and second eigenvector matrices are centered. Next, the autocovariance matrices of each matrix and the cross-covariance matrix between them are calculated. A canonical correlation matrix is ​​constructed using these covariance matrices, and the sum of the main diagonal elements of this matrix is ​​calculated. This sum represents the overall correlation of the current batch of samples across the two views. To accommodate the gradient descent optimization mechanism of neural networks, the negative of this sum of correlation coefficients is defined as the loss function value.

[0100] Backpropagation and Joint Optimization: Based on the calculated loss function value, the backpropagation algorithm is used to calculate the gradient of the loss function with respect to the network weight parameters in the first and second mapping channels. These gradients indicate the direction of parameter adjustment that increases the correlation coefficient. Subsequently, the gradient descent optimization algorithm is used to synchronously update the weight parameters of the two channels by subtraction based on the calculated gradient values. The above process of forward propagation, correlation calculation, and parameter update is repeated until the loss function value tends to stabilize. At this point, the network training is considered complete, and the converged parameters can be used to extract high-quality fusion features.

[0101] Step 302: Input the behavioral feature vector into the first mapping channel to obtain the first feature vector, and input the interest feature vector into the second mapping channel to obtain the second feature vector.

[0102] In the embodiments of this application, the forward propagation process of the neural network is performed. For example, in the forward propagation, the first channel maps the 24-dimensional input to a 10-dimensional first feature vector. The second channel maps the 100-dimensional input to a 10-dimensional second feature vector. The standardized behavioral feature vector is input into the input layer of the first mapping channel. The data is then processed by matrix multiplication and nonlinear activation functions in each hidden layer, and finally the first feature vector is output by the output layer.

[0103] Simultaneously, the standardized interest feature vector is input into the second mapping channel, and after hierarchical operations similar to those of the first mapping channel, its output layer outputs the second feature vector. Since the number of output layer nodes of the two channels is preset to the same value, the generated first and second feature vectors have the same dimension and both reside in the same nonlinear mapping space.

[0104] Step 303: Calculate the covariance of the first eigenvector and the second eigenvector, and calculate the Pearson correlation coefficient based on the covariance.

[0105] In this step, covariance refers to the expected value of the overall error between the first and second eigenvectors across all dimensions, representing the degree of synchronization between the changes of the two vectors.

[0106] In the embodiments of this application, calculations are typically performed based on batch samples. For a batch of samples, a first feature matrix and a second feature matrix are obtained by stacking them. After decentering these two matrices, their cross-covariance matrix and their respective autocovariance matrix are calculated using matrix multiplication. Using singular value decomposition or other matrix decomposition techniques, the canonical correlation matrix is ​​derived based on these covariance matrices. The sum of the main diagonal elements of this matrix is ​​the sum of the Pearson correlation coefficients.

[0107] Step 304: Construct an objective function with the negative of the Pearson correlation coefficient as the loss value, and use the gradient descent algorithm to jointly optimize the objective function until the improvement of the Pearson correlation coefficient is within a preset threshold. Then determine the convergence of the first network weight parameters and the second network weight parameters.

[0108] In this step, the loss value refers to the scalar that the optimization algorithm needs to minimize.

[0109] Step 311: Randomly select a preset number of training users, and use the behavioral feature vector and interest feature vector of each training user as a training sample.

[0110] In this step, the preset quantity refers to the batch size of samples drawn from the total sample set in each training iteration.

[0111] In this embodiment, a certain number of user data are extracted from all target user data stored in the database using a random sampling algorithm. For example, a preset number of 32 is set, meaning 32 users are extracted each time. For each selected user, their corresponding behavioral feature vector and interest feature vector are extracted, and these two vectors are combined into a training sample.

[0112] Step 312: According to the preset arrangement, stack the first feature vector and the second feature vector corresponding to all training samples to obtain the first feature matrix and the second feature matrix.

[0113] In this step, the first feature matrix refers to a two-dimensional numerical matrix formed by combining the first feature vectors of all samples in a batch according to a specific rule. The second feature matrix refers to a two-dimensional numerical matrix formed by combining the second feature vectors of all samples in a batch according to the same rule.

[0114] In this embodiment, the training samples are first input into the neural network for forward propagation. The first mapping channel outputs a batch of first feature vectors, and the second mapping channel outputs a batch of second feature vectors. These vectors are then stacked row by row according to the order in which the samples are selected. For example, if there are 32 samples, and each feature vector has a dimension of 10, then the first and second feature matrices formed after stacking will both have a dimension of 32 rows and 10 columns.

[0115] Step 313: Decentralize and standardize the first feature matrix and the second feature matrix respectively to generate a canonical correlation matrix.

[0116] In this step, the typical correlation matrix refers to the intermediate matrix calculated based on two decentralized matrices, which contains the correlation structure information between the two view data.

[0117] In this embodiment, the arithmetic mean of each column of the first feature matrix is ​​calculated to obtain a first mean vector. This first mean vector is then subtracted from each row of the first feature matrix to obtain a first decentralized matrix. Similarly, a second decentralized matrix is ​​obtained. Based on these two decentralized matrices, the autocovariance matrix is ​​calculated. , and cross-covariance matrix Finally, the canonical correlation matrix is ​​calculated using the formula. As shown in formula (2):

[0118] (2)

[0119] Step 314: Add all the diagonal elements of the canonical correlation matrix to obtain the sum of the Pearson correlation coefficients.

[0120] In this embodiment, the computational method needs to be selected based on actual computing resources and numerical stability. When dealing with scenarios requiring high numerical stability and strict accuracy, singular value decomposition (SVD) is used. SVD is performed on the canonical correlation matrix T, directly obtaining the sequence of singular values ​​on its diagonal. All values ​​in this sequence are summed to obtain the total Pearson correlation coefficient. When reducing computational complexity and the matrix satisfies the positive definite symmetry condition, eigenvalue calculation is used. The eigenvalues ​​of matrix T are calculated, and the sum of the square roots of the eigenvalues ​​is taken as the total Pearson correlation coefficient.

[0121] For example, using the singular value decomposition method, assuming the sum of the singular values ​​of T is 8.5, then the sum of the Pearson correlation coefficients is also 8.5. In the context of neural network optimization, maximizing this sum means that the data of the two views have achieved optimal alignment across all feature dimensions.

[0122] Step 315: Use the negative of the sum of Pearson correlation coefficients as the loss value to construct the objective function.

[0123] In this embodiment, a mathematical expression is established. The input variable of this expression is the output feature of the neural network, and the output value is a scalar indicator used to measure the performance of the current model. In this scheme, to enable the gradient descent algorithm to work, the mathematical expression is directly defined as: the objective function value is equal to the negative of the sum of the Pearson correlation coefficients. For example, if the calculated sum of the Pearson correlation coefficients is 8.5, then substituting it into this expression yields a loss value of -8.5.

[0124] Step 316: Using the gradient descent algorithm, calculate the first partial derivative of the objective function with respect to the weight parameters of the first network and the second partial derivative with respect to the weight parameters of the second network.

[0125] In this embodiment, automatic differentiation is used to calculate the gradient layer by layer in reverse, starting from the objective function, according to the chain rule. First, the gradient of the objective function with respect to the first and second feature matrices is calculated. Then, it is backpropagated through a fully connected layer to finally obtain the first partial derivative with respect to the first network weight parameters and the second partial derivative with respect to the second network weight parameters.

[0126] For example, using the backpropagation algorithm, the direction of parameter adjustment that would further reduce the loss value of -8.5, i.e. further increase the correlation coefficient of 8.5, is calculated, and the first and second partial derivatives are obtained.

[0127] Step 317: Based on the first and second partial derivatives, update the first and second network weight parameters by subtraction, and calculate the sum of the updated Pearson correlation coefficients until the increase in the Pearson correlation coefficients before and after the update is less than a preset threshold, thus determining that the first and second network weight parameters have converged.

[0128] In this step, the improvement refers to the increase in the sum of the target optimization index, namely the Pearson correlation coefficient, during two consecutive iterations, and is used to determine whether the model training has entered a stable state.

[0129] In this embodiment, a preset learning rate is set to control the step size of each parameter update. The first partial derivative is multiplied by the learning rate to obtain a first correction amount. The current first network weight parameters are subtracted from the first correction amount to obtain the updated first network weight parameters. Similarly, the second network weight parameters are updated using the second partial derivative. For example, assuming the current total Pearson correlation coefficient is 8.5 and the learning rate is set to 0.01, the first partial derivative is calculated to be 0.5. The first network weight parameters are then updated. The second network weight parameters are updated similarly.

[0130] After each update, forward propagation and correlation calculation are performed again using the updated parameters to obtain the updated sum of Pearson correlation coefficients. Next, the difference between the updated sum of Pearson correlation coefficients and the original sum of Pearson correlation coefficients is calculated to obtain the improvement. This improvement is compared to a preset threshold. If the improvement is greater than or equal to the preset threshold, the model is still in the fast learning phase, and the gradient calculation and parameter update steps are repeated. If the improvement is less than the preset threshold, the model is considered to have converged, and the training process stops.

[0131] The first and second network weight parameters obtained at this point are the final convergence parameters. For example, using the new parameters, the sum of the new correlation coefficients is calculated to be 8.505. The calculated improvement is 0.005. Assume the preset threshold is 0.001. Therefore, training continues. After several iterations, the sum before a certain update is 9.2000, and the sum after the update is 9.20005. The improvement is 0.00005. The convergence condition is met, and training stops.

[0132] Step 305: Using the converged first network weight parameters and the converged second network weight parameters, remap the behavior feature vector and interest feature vector to obtain the target first feature vector and the target second feature vector.

[0133] In this embodiment, the state of the neural network is first fixed, parameter updates are stopped, and the first and second network weight parameters after training are loaded. The behavioral feature vector of the target user to be processed is input into the first mapping channel. The data undergoes fixed weight matrix multiplication and nonlinear activation operations within the network, and the output layer outputs the first target feature vector. Simultaneously, the interest feature vector of the target user is input into the second mapping channel, undergoes the same fixed-parameter network operation, and the output layer outputs the second target feature vector.

[0134] For example, Input the pre-trained first channel, output a 10-dimensional target first feature vector. =[0.1,0.2,...]. This represents the interest feature vector. Input the pre-trained second channel, and output a 10-dimensional target second feature vector. =[0.8,0.9,...]. These two feature vectors reside in the same highly correlated feature subspace.

[0135] Step 306: Concatenate the first feature vector and the second feature vector of the target to obtain the fused feature vector of each target user.

[0136] In this embodiment, elements of the second feature vector of the target are appended to the end of the first feature vector of the target to form a new long vector. For example, the first feature vector of the target... and the second feature vector of the target The two vectors are concatenated to generate a fused feature vector. =[0.1, 0.2, ..., 0.8, 0.9, ], if and Both are 10-dimensional, then The total dimensions are 20.

[0137] Step 104: Based on the fused feature vectors, calculate the Gaussian similarity between target users to construct the Laplacian matrix. By mapping the Laplacian matrix to a low-dimensional manifold space, and using a clustering algorithm, all target users are divided into multiple social groups.

[0138] In this step, the Laplacian matrix is ​​an algebraic representation combining the degree matrix and adjacency matrix of a graph, used to describe the connection strength and topological structure between user nodes in a social network. A low-dimensional manifold space is a low-dimensional coordinate system with an inherent geometric structure onto which high-dimensional data is projected through eigenvalue decomposition, used to reveal the nonlinear community structure hidden within complex high-dimensional data. Clustering algorithms are unsupervised learning methods used to group similar objects into the same group. A social group refers to a set of users in a low-dimensional manifold space who are spatially close and share highly similar behavioral and interest characteristics.

[0139] like Figure 3 As shown, Figure 3 The illustration shows a flowchart of a method for dividing all target users into multiple social groups according to an embodiment of this application.

[0140] Step 401: Calculate the Euclidean distance between any two fused feature vectors, and substitute each Euclidean distance into the Gaussian kernel function to obtain the Gaussian similarity between target users, so as to construct a Gaussian similarity matrix.

[0141] In this step, Gaussian similarity refers to mapping the Euclidean distance between two vectors to a value between 0 and 1 using a Gaussian kernel function. Gaussian similarity is used to represent the degree of closeness between two users in the fused feature space; the closer the value is to 1, the more similar they are. The Gaussian similarity matrix is ​​a symmetric square matrix where the element in the i-th row and j-th column stores the Gaussian similarity value between the i-th user and the j-th user.

[0142] In this embodiment, the fused feature vector of each target user is obtained, and the squared Euclidean distance between the fused feature vectors of any two users is calculated. The bandwidth parameter of the Gaussian kernel function is set. This parameter controls the rate at which similarity decreases with increasing distance. The Euclidean distance is substituted into the Gaussian kernel function formula for calculation, as shown in formula (3):

[0143] (3)

[0144] in, This represents the Gaussian similarity between user i and user j. and These are their fused feature vectors, The square of the Euclidean distance is used. All the calculated similarity values ​​are then filled into the corresponding positions in the matrix to construct the Gaussian similarity matrix. To construct a Laplacian matrix suitable for spectral clustering, the elements on the main diagonal of the Gaussian similarity matrix need to be forced to 0. For example, suppose there are three target users A, B, and C, and the fused feature vectors obtained after processing are as follows: , , First, calculate the distance between them and construct a Gaussian similarity matrix. Assuming as follows:

[0145]

[0146] Step 402: Construct a diagonal matrix based on the row sum of the Gaussian similarity matrix, and perform matrix subtraction between the diagonal matrix and the Gaussian similarity matrix to obtain the Laplace matrix.

[0147] In this step, a diagonal matrix is ​​a square matrix in which all elements except the main diagonal are zero.

[0148] In this embodiment, each row of the Gaussian similarity matrix is ​​traversed, and all elements in that row are summed to obtain the row sum. These row sums are then sequentially filled into the main diagonal of a zero matrix to obtain a diagonal matrix. .in Next, perform matrix subtraction using a diagonal matrix. Subtract the Gaussian similarity matrix The Laplace matrix is ​​obtained. For example, calculating the degree matrix. The degree of A is 0.9, the degree of B is 0.9, and the degree of C is 0.2.

[0149]

[0150] Calculate the Laplace matrix :

[0151]

[0152] Step 403: Perform eigenvalue decomposition on the Laplacian matrix to obtain multiple eigenvalues. Sort all eigenvalues ​​in ascending order to obtain an eigenvalue sequence. Select the first preset number of non-zero eigenvalues ​​from the eigenvalue sequence. Construct an eigenvector matrix based on the eigenvectors corresponding to the non-zero eigenvalues.

[0153] In this step, the eigenvector matrix refers to a matrix composed of selected eigenvectors as column vectors, used to map the original samples to a low-dimensional space.

[0154] In this embodiment of the application, a linear algebra library is invoked to process the real symmetric Laplace matrix. Perform eigenvalue decomposition and solve the equations To obtain all eigenvalues and the corresponding feature vector Sort all eigenvalues ​​in ascending order. Typically, the smallest eigenvalue of the Laplacian matrix is ​​0, and its corresponding eigenvector is an all-1 vector, which does not contain discriminative information and is therefore discarded. From the remaining non-zero eigenvalues, select the smallest eigenvalues. One, of which The preset number of dimensionality reduction dimensions is typically used to determine the desired number of social groups. Extract this... The eigenvectors corresponding to each eigenvalue are arranged column-wise to form a... eigenvector matrix ,in This represents the total number of users.

[0155] For example, for Perform eigenvalue decomposition to obtain and sort the eigenvalues. Assume a predetermined number of dimensionality reduction dimensions. The value is 2. Select the eigenvectors corresponding to the two smallest non-zero eigenvalues. and Construct the eigenvector matrix Size is .

[0156] Step 404: Treat each row of data in the feature vector matrix as a coordinate point in the low-dimensional manifold space, and perform clustering operations on all coordinate points in the low-dimensional manifold space to obtain multiple social groups, where the target users in each social group have the same clustering label.

[0157] In this embodiment of the application, the feature vector matrix is... Each row is considered a data sample, and these samples are distributed across a... In a 1 / 2-dimensional Euclidean space, this space is the low-dimensional manifold space. Clustering algorithms are performed within this space. Random initialization is then performed. For each cluster center, calculate the distance from each coordinate point to all cluster centers and assign the coordinate point to the nearest cluster. Update the cluster centers and repeat the iteration until convergence. Finally, each coordinate point obtains a cluster label. Target users with the same cluster label are grouped into the same social group.

[0158] For example, matrix The first line This represents the coordinates of user A in the low-dimensional manifold space. (Matrix) The second line Let these be the coordinates of user B. (Matrix) The third line Let A and B be the coordinates of user C. Cluster these three points. Since A and B are closely connected in the similarity matrix, their coordinates in low-dimensional space will be very close, while C is farther away. The final clustering result might be: A and B belong to group 1, C might belong to group 2, or C might be classified into group 1 but be far from the center.

[0159] Step 105: Based on the low-dimensional manifold space, identify the edge node objects and the central node objects in the social group, extract the community activity information that the central node objects participate in frequently, and send the community activity information to the edge node objects in the same social group.

[0160] In this step, the central node object refers to the user located at the geometric center or closest to other members within the social group in the low-dimensional manifold space. It represents the core active figures and influential members within the social group. The peripheral node object refers to the user located at the geometric edge or furthest from the central node within the social group in the low-dimensional manifold space. It represents elderly people at risk of social isolation or in a state of isolation who require assistance. Community activity information refers to the names, times, and locations of specific activities frequently participated in by the selected central node objects.

[0161] Step 501: Based on the low-dimensional manifold space, perform an arithmetic mean operation on the coordinates of the points within each social group to obtain the center coordinates of each social group.

[0162] In this embodiment, each social group is traversed. The coordinates of all target users belonging to the current social group are obtained. For each dimension, the coordinate values ​​of all users are added together, and then divided by the total number of users to calculate the average value for that dimension. The average values ​​of all dimensions are combined to obtain the center coordinates of the social group.

[0163] For example, suppose we obtain a social group G1 after clustering, and suppose this group contains users A, B, and C for distance calculation purposes. User A's coordinates... User B coordinates User C coordinates The coordinates of the center are calculated as follows: .

[0164] Step 502: Calculate the spatial Euclidean distance between the coordinates of each target user in the same social group and the center coordinates. Target users whose spatial Euclidean distance is less than or equal to a preset distance threshold are designated as center node objects, and target users whose spatial Euclidean distance is greater than the preset distance threshold are designated as edge node objects.

[0165] In this embodiment of the application, a distance threshold is set. For each target user, their coordinates are used. and center coordinates Calculate the Euclidean distance The calculated distance With threshold Compare. If If so, then the user is determined to be the central node object. If so, the user is determined to be an edge node object.

[0166] For example, A to the center is B to the center is C to the center is Set a threshold The determination result is: A is the center node, B is the center node, and C is the edge node.

[0167] Step 503: Count the number of times the central node object participates in each community activity within the preset time period, sort the number of participations in all community activities in descending order, and obtain the participation popularity sequence of each social group.

[0168] In this embodiment, the system connects to a community activity database and retrieves the activity check-in records of the user(s) marked as central node objects within a preset time period. The activity names are extracted from the records. The total number of times each activity was participated in by all central node objects is counted. Finally, all activities are sorted according to the total number of participations to generate an ordered list.

[0169] For example, consider the activities of A and B: A recently participated in chess 3 times and calligraphy 1 time. B participated in chess 2 times and Tai Chi 1 time. The total activity count is: chess 5 times, calligraphy 1 time, Tai Chi 1 time. Sort these activities to obtain a popularity sequence.

[0170] Step 504: Select the top preset number of community activities from the participation popularity sequence as community activity information, and obtain the terminal device identifier of the edge node object within each social group. Send the community activity information to the edge node object within the same social group through the terminal device identifier.

[0171] In this embodiment, the top N activities in the popularity sequence are extracted and packaged into recommended content. Users marked as edge nodes within the same social group are located, and their terminal device identifiers are retrieved from the user profile database. Recommended cards containing these N popular activities are sent to the edge node objects' terminals via a push notification interface. For example, selecting the top activity: chess. The terminal device identifier of user C is obtained, and a message is pushed to C's terminal app: "Your neighbors are all participating in chess activities recently; welcome to try it out!"

[0172] Figure 4This is a schematic diagram illustrating a specific implementation of the personalized comprehensive elderly care service recommendation system based on artificial intelligence and big data provided in this application. (Refer to...) Figure 4 The system may include:

[0173] The data collection module 21 is used to collect travel data and browsing data of community service apps from multiple target users. Travel data includes travel timestamps and directions of entry and exit, while browsing data includes multiple browsing records, browsing frequency, and dwell time.

[0174] Module 22 is used to construct a behavioral feature vector based on the number of records of entry and exit directions within a preset time period of travel data, and to construct an interest feature vector based on browsing data.

[0175] Optimization module 23 is used to input behavioral feature vectors and interest feature vectors into a neural network based on deep canonical correlation analysis, and generate a fused feature vector for each target user by jointly optimizing the objective function of the neural network.

[0176] Module 22 is also used to calculate the Gaussian similarity between target users based on the fused feature vectors to construct the Laplacian matrix. By mapping the Laplacian matrix to a low-dimensional manifold space, a clustering algorithm is used to divide all target users into multiple social groups.

[0177] The determination module 24 is used to determine the edge node objects and the central node objects in the social group based on the low-dimensional manifold space, extract the community activity information of the central node objects that participate in frequently, and send the community activity information to the edge node objects in the same social group.

[0178] The personalized comprehensive elderly care service recommendation system based on artificial intelligence big data in this application embodiment is used to implement the aforementioned personalized comprehensive elderly care service recommendation method based on artificial intelligence big data. Therefore, the specific implementation of the personalized comprehensive elderly care service recommendation system based on artificial intelligence big data can be found in the embodiment section of the personalized comprehensive elderly care service recommendation method based on artificial intelligence big data above. The specific implementation can be referred to the description of the corresponding embodiments, and will not be repeated here.

[0179] Figure 5 A schematic diagram of the hardware structure of the electronic device provided in an embodiment of this application is shown.

[0180] This application also provides an electronic device, including: a memory for storing a computer program; and a processor for executing the computer program to implement the steps of the above-described personalized comprehensive elderly care service recommendation method based on artificial intelligence big data.

[0181] The electronic device may include a processor 510 and a memory 520 storing computer program instructions.

[0182] Specifically, the processor 510 may include a central processing unit (CPU), an application-specific integrated circuit (ASIC), or one or more integrated circuits that can be configured to implement the embodiments of this application.

[0183] Memory 520 may include mass storage for data or instructions. For example, and not limitingly, memory 520 may include a hard disk drive (HDD), floppy disk drive, flash memory, optical disk, magneto-optical disk, magnetic tape, or Universal Serial Bus (USB) drive, or a combination of two or more of these. Where appropriate, memory 520 may include removable or non-removable (or fixed) media. Where appropriate, memory 520 may be internal or external to the integrated gateway disaster recovery device. In a particular embodiment, memory 520 is non-volatile solid-state memory.

[0184] Memory may include read-only memory (ROM), random access memory (RAM), disk storage media devices, optical storage media devices, flash memory devices, and electrical, optical, or other physical / tangible memory storage devices. Therefore, typically, memory includes one or more tangible (non-transitory) computer-readable storage media (e.g., memory devices) encoded with software including computer-executable instructions, and when the software is executed (e.g., by one or more processors), it is operable to perform the operations described with reference to the method according to the first aspect of this disclosure.

[0185] The processor 510 reads and executes computer program instructions stored in the memory 520 to implement any of the personalized comprehensive elderly care service recommendation methods based on artificial intelligence big data in the above embodiments.

[0186] In one example, the electronic device may also include a communication interface 530 and a bus 540. Wherein, such as Figure 5 As shown, the processor 510, memory 520, and communication interface 530 are connected through bus 540 and complete communication with each other.

[0187] The communication interface 530 is mainly used to realize communication between various modules, devices, units and / or equipment in the embodiments of this application.

[0188] Bus 540 includes hardware, software, or both, that couples components of an online data traffic metering device together. For example, and not limitingly, the bus may include an Accelerated Graphics Port (AGP) or other graphics bus, an Enhanced Industry Standard Architecture (EISA) bus, a Front Side Bus (FSB), HyperTransport (HT) interconnect, an Industry Standard Architecture (ISA) bus, an Infinite Bandwidth Interconnect, a Low Pin Count (LPC) bus, a memory bus, a Microchannel Architecture (MCA) bus, a Peripheral Component Interconnect (PCI) bus, a PCI-Express (PCI-X) bus, a Serial Advanced Technology Attachment (SATA) bus, a Video Electronics Standards Association Local (VLB) bus, or other suitable buses, or combinations of two or more of these. Where appropriate, bus 540 may include one or more buses. Although specific buses are described and illustrated in embodiments of this application, any suitable bus or interconnect is contemplated herein.

[0189] This application also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of any of the above-described methods for recommending personalized comprehensive elderly care services based on artificial intelligence big data.

[0190] In one exemplary embodiment, the aforementioned computer-readable storage medium may include, but is not limited to, various media capable of storing computer programs, such as USB flash drives, read-only memory, random access memory, portable hard drives, magnetic disks, or optical disks.

[0191] Embodiments of the present invention also provide a computer program product, which includes a computer program that, when executed by a processor, implements the steps in any of the embodiments of the personalized comprehensive elderly care service recommendation method based on artificial intelligence big data.

[0192] Those skilled in the art will further recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0193] The above provides a detailed description of the personalized comprehensive elderly care service recommendation method and system based on artificial intelligence and big data provided in this application. Specific examples have been used to illustrate the principles and implementation methods of this application. The descriptions of the embodiments above are only for the purpose of helping to understand the method and its core ideas. It should be noted that those skilled in the art can make several improvements and modifications to this application without departing from the principles of this application, and these improvements and modifications also fall within the protection scope of this application.

Claims

1. A personalized comprehensive elderly care service recommendation method based on artificial intelligence and big data, characterized in that, include: The system collects travel data and browsing data from community service apps from multiple target users. The travel data includes travel timestamps and directions of entry and exit, while the browsing data includes multiple browsing records, browsing frequency, and dwell time. Based on the number of records of the travel data in the direction of entry and exit within a preset time period, a behavioral feature vector is constructed, and based on the browsing data, an interest feature vector is constructed. The behavioral feature vector and the interest feature vector are respectively input into a neural network based on deep canonical correlation analysis. By jointly optimizing the objective function of the neural network, a fused feature vector for each target user is generated. Based on the fused feature vector, Gaussian similarity between the target users is calculated to construct a Laplacian matrix. The Laplacian matrix is ​​then mapped to a low-dimensional manifold space, and a clustering algorithm is used to divide all target users into multiple social groups. Based on the low-dimensional manifold space, edge node objects and central node objects in the social group are determined, and information on community activities that the central node objects frequently participate in is extracted. The community activity information is then sent to the edge node objects within the same social group.

2. The method according to claim 1, characterized in that, The method further includes: Obtain the generation time of each browsing record in the browsing data, and calculate the time weight coefficient of each browsing record based on the difference between the generation time and the current time; The step of constructing an interest feature vector based on the browsing data includes: Based on the browsing data, an interest feature vector is constructed using the time weighting coefficient of each browsing record.

3. The method according to claim 2, characterized in that, Based on the browsing data, an interest feature vector is constructed using the time weighting coefficient of each browsing record, including: The content tags of browsing records in the browsing data of each target user are converted into semantic embedding vectors, and the basic attention of each browsing record is calculated based on the browsing frequency and dwell time in the browsing data. The basic attention level is multiplied by the corresponding time weight coefficient to obtain the comprehensive weight value of each browsing record; The semantic embedding vector of each browsing record is multiplied by the corresponding comprehensive weight value to obtain a weighted semantic vector. All weighted semantic vectors of the same target user are then summed to obtain the interest feature vector.

4. The method according to claim 1, characterized in that, The behavioral feature vector and the interest feature vector are respectively input into a neural network based on deep canonical correlation analysis. By jointly optimizing the objective function of the neural network, a fused feature vector for each target user is generated, including: A neural network based on deep canonical correlation analysis is constructed, the neural network including a first mapping channel with a first network weight parameter and a second mapping channel with a second network weight parameter; The behavioral feature vector is input into the first mapping channel to obtain the first feature vector, and the interest feature vector is input into the second mapping channel to obtain the second feature vector; Calculate the covariance between the first feature vector and the second feature vector, and calculate the Pearson correlation coefficient based on the covariance; Construct an objective function with the negative of the Pearson correlation coefficient as the loss value, and use the gradient descent algorithm to jointly optimize the objective function until the improvement of the Pearson correlation coefficient is within a preset threshold, and determine that the first network weight parameters and the second network weight parameters have converged. Using the converged first network weight parameters and the converged second network weight parameters, the behavior feature vector and the interest feature vector are remapped to obtain the target first feature vector and the target second feature vector. The first feature vector and the second feature vector of the target are concatenated to obtain the fused feature vector of each target user.

5. The method according to claim 4, characterized in that, Construct an objective function with the negative of the Pearson correlation coefficient as the loss value, and jointly optimize the objective function using a gradient descent algorithm until the improvement of the Pearson correlation coefficient is within a preset threshold. Determine that the first network weight parameters and the second network weight parameters have converged, including: A preset number of training users are randomly selected, and the behavioral feature vector and interest feature vector of each training user are used as a training sample. According to the preset arrangement, the first feature vector and the second feature vector corresponding to all training samples are stacked to obtain the first feature matrix and the second feature matrix. The first feature matrix and the second feature matrix are respectively decentered and standardized to generate a canonical correlation matrix; The sum of the Pearson correlation coefficients is obtained by adding all the diagonal elements of the canonical correlation matrix. The negative of the sum of the Pearson correlation coefficients is used as the loss value to construct the objective function; Using the gradient descent algorithm, the first partial derivative of the objective function with respect to the weight parameters of the first network and the second partial derivative with respect to the weight parameters of the second network are calculated respectively. Based on the first partial derivative and the second partial derivative, the first network weight parameters and the second network weight parameters are updated by subtraction, and the sum of the updated Pearson correlation coefficients is calculated until the increase in the Pearson correlation coefficient before and after the update is less than a preset threshold, at which point the convergence of the first network weight parameters and the second network weight parameters is determined.

6. The method according to claim 1, characterized in that, Based on the fused feature vectors, Gaussian similarity among the target users is calculated to construct a Laplacian matrix. This Laplacian matrix is ​​then mapped to a low-dimensional manifold space, and a clustering algorithm is used to divide all target users into multiple social groups, including: Calculate the Euclidean distance between any two fused feature vectors, and substitute each Euclidean distance into the Gaussian kernel function to obtain the Gaussian similarity between the target users, so as to construct a Gaussian similarity matrix; Based on the row sum of the Gaussian similarity matrix, a diagonal matrix is ​​constructed. The diagonal matrix and the Gaussian similarity matrix are then subtracted to obtain the Laplace matrix. The Laplacian matrix is ​​decomposed into eigenvalues ​​to obtain multiple eigenvalues. All eigenvalues ​​are arranged in ascending order to obtain an eigenvalue sequence. The first preset number of non-zero eigenvalues ​​are selected from the eigenvalue sequence. Based on the eigenvectors corresponding to the non-zero eigenvalues, an eigenvector matrix is ​​constructed. Each row of data in the feature vector matrix is ​​used as a coordinate point in a low-dimensional manifold space, and clustering is performed on all coordinate points in the low-dimensional manifold space to obtain multiple social groups, wherein the target users in each social group have the same cluster label.

7. The method according to claim 1, characterized in that, Based on the low-dimensional manifold space, edge node objects and central node objects in the social group are determined, and information on frequently participated community activities of the central node objects is extracted. This community activity information is then sent to edge node objects within the same social group, including: Based on the low-dimensional manifold space, the arithmetic mean of the coordinate points within each social group is calculated to obtain the center coordinates of each social group. Calculate the spatial Euclidean distance between the coordinates of each target user in the same social group and the center coordinates. Target users whose spatial Euclidean distance is less than or equal to a preset distance threshold are designated as center node objects, and target users whose spatial Euclidean distance is greater than the preset distance threshold are designated as edge node objects. The number of times the central node object participated in each community activity within a preset time period was counted, and the number of participations in all community activities was sorted in descending order to obtain the participation popularity sequence of each social group. Select a preset number of community activities from the participation popularity sequence as community activity information, obtain the terminal device identifier of the edge node object within each social group, and send the community activity information to the edge node object within the same social group through the terminal device identifier.

8. A personalized comprehensive elderly care service recommendation system based on artificial intelligence and big data, characterized in that: include: The data collection module is used to collect travel data and browsing data of community service apps from multiple target users. The travel data includes travel timestamps and directions of entry and exit, and the browsing data includes multiple browsing records, browsing frequency, and dwell time. The construction module is used to construct a behavioral feature vector based on the number of records of the direction of entry and exit in the travel data within a preset time period, and to construct an interest feature vector based on the browsing data. An optimization module is used to input the behavioral feature vector and the interest feature vector into a neural network based on deep canonical correlation analysis, and generate a fusion feature vector for each target user by jointly optimizing the objective function of the neural network. The construction module is also used to calculate the Gaussian similarity between the target users based on the fused feature vector to construct a Laplacian matrix, and to divide all target users into multiple social groups by mapping the Laplacian matrix to a low-dimensional manifold space and using a clustering algorithm. The determination module is used to determine the edge node objects and the central node objects in the social group based on the low-dimensional manifold space, extract the community activity information of the central node objects that participate in frequently, and send the community activity information to the edge node objects in the same social group.

9. An electronic device, characterized in that, include: Memory, used to store computer programs; A processor, configured to execute the computer program to implement the steps of the personalized comprehensive elderly care service recommendation method based on artificial intelligence big data as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, enables the implementation of the personalized comprehensive elderly care service recommendation method based on artificial intelligence big data as described in any one of claims 1 to 7.

Images