Intelligent cloud data acquisition and dimension reduction customer acquisition method based on compressed sensing

Abstract

The invention belongs to the technical field of data processing, and particularly relates to an intelligent cloud data collection and dimension reduction customer obtaining method based on compressed sensing, and the method comprises a data collection step, a data analysis processing step, a data fusion processing step, and a target user customer obtaining processing step. The data acquisition step is used for acquiring user behavior data from multi-source data; the data analysis processing step is used for calculating the similarity between the users according to the time sequence characteristics, and constructing a behavior propagation equation based on the similarity between the users according to the similarity between the users; the data fusion processing step is used for performing feature fusion on a kernel tensor, a time sequence feature and a behavior propagation tensor to obtain a fusion feature; and the target user obtaining processing step is used for predicting the customer conversion rate of each user based on the fusion features. According to the method, the depth and the precision of user behavior analysis are remarkably improved, and an efficient and reliable solution is provided for accurate customer acquisition and user portrait construction.

Description

Technical Field

[0001] The present invention belongs to the field of data processing technology, and specifically relates to an intelligent cloud data collection and dimensionality reduction customer acquisition method based on compressed sensing. Background Art

[0002] Driven by the era of big data, user behavior analysis and accurate customer acquisition have become core concerns of many industries, especially in the fields of e-commerce, finance, social media, etc. Through in-depth mining of user behavior data, not only can the accurate construction of user portraits be achieved, but also personalized recommendations and advertising strategies can be optimized, thereby improving business conversion rates. However, the multi-source heterogeneity, dimensional complexity, and temporal dynamics of user behavior data have brought huge technical challenges to data analysis and target user screening. At present, a variety of technologies have been used to try to solve this problem, but there are still limitations, and more efficient and accurate analysis methods are urgently needed.

[0003] In the prior art, the core methods of user behavior analysis mainly include feature extraction methods based on matrix decomposition, behavior prediction models driven by deep learning, and traditional similarity calculation and recommendation system algorithms. Among them, matrix decomposition techniques (such as SVD and PCA) are often used for dimensionality reduction and main feature extraction of high-dimensional user behavior data. These methods have certain advantages in processing structured data and relatively low-dimensional user behavior data. However, with the diversification of data sources and the explosive growth of data dimensions, the single matrix decomposition method faces the problem of "dimensionality disaster", that is, when processing multi-dimensional heterogeneous data, the matrix model is difficult to capture the complex correlation between data. In addition, the traditional matrix method has poor adaptability to sparse data. When there are a large number of missing values ​​or high sparsity in the data, the performance of the model will significantly decrease, resulting in insufficient accuracy in behavior feature extraction. Deep learning models have been widely used in the field of user behavior prediction in recent years, such as time series analysis methods based on RNN and LSTM, feature extraction models based on CNN, and recommendation systems driven by multi-layer perceptron (MLP). These models can capture complex nonlinear relationships in user behavior data through deep feature learning. However, deep learning methods have extremely high requirements for data volume and computing resources, especially when processing large-scale, multimodal data. The training cost and real-time performance of model reasoning are difficult to meet the needs of practical applications. In addition, deep learning models are usually regarded as "black box" models, and their results are poorly interpretable. It is difficult to intuitively explain the relationship between user behavior characteristics and model outputs, which brings troubles to decision support in business applications. Summary of the invention

[0004] In view of this, the main purpose of the present invention is to provide an intelligent cloud data collection and dimensionality reduction customer acquisition method based on compressed sensing, so as to realize efficient collection, dimensionality reduction processing and accurate analysis of multi-source heterogeneous data. By constructing kernel tensors, time series features and behavior propagation equations, the present invention captures the global characteristics, time dynamics and group propagation laws of user behavior, generates highly expressive fusion features, and combines the dynamic weights of the gamma function to realize the sparsity control of user behavior and the conversion prediction of target customers. This method significantly improves the data processing efficiency, the depth and accuracy of user behavior analysis, and provides an efficient and reliable solution for accurate customer acquisition, personalized recommendations and user portrait construction.

[0005] The technical solution adopted by the present invention is as follows:

[0006] An intelligent cloud data collection and dimensionality reduction customer acquisition method based on compressed sensing, the method comprising: a data collection step, a data analysis and processing step, a data fusion processing step and a target user customer acquisition processing step; the data collection step is used to obtain user behavior data from multi-source data and represent the user behavior data as a multi-order tensor; the data analysis and processing step is used to perform tensor decomposition on the multi-order tensor to obtain a core tensor, construct a time series feature based on the decomposed core tensor, calculate the similarity between users according to the time series feature, and construct a behavior propagation equation based on the similarity between users according to the similarity between users to obtain a behavior propagation tensor; the data fusion processing step is used to perform feature fusion on the core tensor, the time series feature and the behavior propagation tensor to obtain a fusion feature; the target user customer acquisition processing step is used to predict the customer conversion rate of each user based on the fusion feature to obtain the conversion prediction rate that the user can be converted into a target customer; the conversion prediction rate of each user is compared with the set confidence level, and users with a conversion prediction rate higher than the confidence level are taken as target potential customers.

[0007] Furthermore, multi-order tensors It is expressed by the following formula:

[0008]

[0009] Among them, multi-order tensors N represents the behavior feature value of user n in data source s, dimension d, and time t; s represents the total number of sampling points in the data source s; Ω is the sampling space, which represents the area where the integral is calculated and is defined as the spatial range where the user and the sampling points are located; B d,s (t) is the behavior characteristic function of dimension d in data source s at time t, which is a quantity describing behavior characteristics, including: click rate, purchase volume or number of visits; t i is the sampling time point; ∥x n -x i∥2 indicates the user's location x n With the sampling point position x i The Euclidean distance of i (t) represents the feature vector of sampling point i at time t, reflecting the specific characteristics of the sampling point of user behavior data, including: purchase records or browsing behavior; σ t is the standard deviation of the Gaussian distribution.

[0010] Furthermore, in the data analysis and processing step, the multi-order tensor is decomposed using Tucker decomposition to obtain a core tensor, and the formula is as follows:

[0011]

[0012] Among them, the core tensor The following sparse constraints are satisfied:

[0013]

[0014] Among them, U N is the user modal basis matrix, U D is the dimension modal basis matrix, U T is the time modal basis matrix, U S is the modal basis matrix of the data source, which is a preset matrix; k is the sparsity parameter; ∥·∥0 represents the L0 norm; ∥·∥ F represents the Frobenius norm; rank(·) is the rank operation of the matrix; Represents the core tensor of user n. When it is user i, it is expressed as

[0015] Furthermore, the data analysis and processing step constructs the time series feature based on the decomposed core tensor through the following formula:

[0016]

[0017] Wherein, tra represents the transpose operation of the matrix; τ is the time integral variable; tr(·) represents the operation of the trace of the matrix; W(t-τ) represents the time weighting matrix when the time is t-τ; W(t) represents the time weighting matrix when the time is t, which is preset; represents the time series characteristics of user n. When it is user i, it is expressed as Represents the kernel tensor time gradient.

[0018] Furthermore, the data analysis and processing step calculates the similarity between users according to the time series characteristics through the following formula, and calculates the similarity between users according to the similarity between users

[0019]

[0020] Among them, <·,·> represents the tensor inner product operation; U N (:,i) represents the basis vector of user i; U N (:,j) represents the basis vector of user j; σ s is the similarity adjustment parameter, which is the preset value; represents the temporal characteristics of user i; represents the temporal characteristics of user j; Represents the similarity between user i and user j; all similarities between users are taken as one element to form the user similarity matrix for The element in row i and column j.

[0021] Furthermore, the data analysis and processing step constructs a behavior propagation equation based on the similarity between users through the following formula:

[0022]

[0023] in, is the Laplace operator; represents the behavior propagation tensor of user n. When it is user i, it is expressed as div(·) is the divergence operator.

[0024] Furthermore, the data fusion processing step is used to fuse the core tensor, the time series feature and the behavior propagation tensor to obtain the fusion feature through the following formula:

[0025]

[0026] in, represents the fusion feature of user n. When it is user i, it is expressed as express The time gradient of for The corresponding form under the time-integrated variable; for The corresponding form under the time-integrated variable; for The corresponding form under the time-integrated variable; It is the tensor outer product operation.

[0027] Furthermore, the target user acquisition processing step uses the following formula to predict the customer conversion rate of each user based on the fusion features to obtain the conversion prediction rate r of the user being able to be converted into a target customer: i (t+Δt):

[0028]

[0029] Among them, r i (t+Δt) represents the conversion prediction rate of user i, which means the conversion prediction rate of the user who can be converted into the target customer after the window time Δt from time t; for The corresponding form under the time-integrated variable; for The corresponding form under the time-integrated variable; for The corresponding form under the time-integrated variable; is the dynamic weight based on the gamma function.

[0030] Furthermore, dynamic weights based on the gamma function Use the following formula to express it:

[0031]

[0032] By adopting the above technical scheme, the present invention produces the following beneficial effects: First, the present invention realizes efficient collection and dimensionality reduction processing of multi-source heterogeneous data through compressed sensing theory, and solves the "dimensionality disaster" problem of traditional technology in high-dimensional data processing. The core idea of ​​compressed sensing is to reconstruct high-dimensional data through a small number of sampling points based on the sparsity assumption, thereby significantly reducing data storage and computing costs. In the present invention, user behavior data is modeled as a high-order tensor. This multidimensional representation can capture the complex associations of data in multiple dimensions such as time, space, and behavior categories. However, high-dimensional tensor data processing faces the challenges of high computational complexity and large memory usage. The present invention reduces the dimensionality of high-dimensional data through tensor decomposition technology, combined with the sparsity assumption of compressed sensing, decomposes it into core tensors and modal basis matrices, retains the core information of the data and significantly reduces the data dimension. This processing method not only improves the integration efficiency of multi-source data, but also lays a high-quality data foundation for subsequent feature extraction and analysis. Secondly, the present invention demonstrates the ability to accurately capture the dynamic changes of user behavior in terms of time series modeling. User behavior has significant temporal dynamic characteristics, such as changes in purchase intentions and transfers of interest hotspots. The present invention extracts the changing rules of user behavior in the time dimension by constructing time series features. By combining the modeling of time weighted matrix and time gradient, the present invention can keenly capture the short-term fluctuations and long-term trends of user behavior, which is suitable for analyzing the short-term response behavior of users (such as purchase decisions in promotional activities) and revealing the long-term changes in users' interests. This deep modeling capability of time dynamics significantly improves the comprehensiveness and accuracy of user behavior analysis. Behavior propagation modeling is another major technical highlight of the present invention, which shows excellent ability in dynamic modeling of interactive relationships between users. Traditional user analysis methods usually only focus on individual behavior characteristics and are difficult to capture the mutual influence between users. The present invention simulates the diffusion and propagation process of user behavior in a group by introducing the similarity matrix and behavior propagation equation between users. The propagation model takes into account the synergy between user similarity, behavior dynamics and global features, and accurately describes the propagation path of user behavior in time and space through the combination of Laplace operator and divergence operator. This nonlinear propagation modeling method can not only identify the behavioral correlation between users, but also predict the potential direction of behavior diffusion, providing a scientific basis for the mining of target customers and the prediction of group behavior. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 A method flow chart of the method for intelligent cloud data collection and dimensionality reduction customer acquisition method based on compressed sensing provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0034] All features disclosed in this specification, or steps in all methods or processes disclosed, except mutually exclusive features and / or steps, can be combined in any manner.

[0035] Any feature disclosed in this specification (including any additional claims and abstract), unless otherwise stated, may be replaced by other equivalent or alternative features having similar purposes. That is, unless otherwise stated, each feature is only an example of a series of equivalent or similar features.

[0036] Example 1: Reference Figure 1 , an intelligent cloud data collection and dimensionality reduction customer acquisition method based on compressed sensing, the method comprising: a data collection step, a data analysis and processing step, a data fusion processing step and a target user customer acquisition processing step; the data collection step is used to obtain user behavior data from multi-source data and represent the user behavior data as a multi-order tensor; the data analysis and processing step is used to perform tensor decomposition on the multi-order tensor to obtain a core tensor, construct a time series feature based on the decomposed core tensor, calculate the similarity between users according to the time series feature, and construct a behavior propagation equation based on the similarity between users according to the similarity between users to obtain a behavior propagation tensor; the data fusion processing step is used to perform feature fusion on the core tensor, the time series feature and the behavior propagation tensor to obtain a fusion feature; the target user customer acquisition processing step is used to predict the customer conversion rate of each user based on the fusion feature to obtain the conversion prediction rate that the user can be converted into a target customer; the conversion prediction rate of each user is compared with the set confidence level, and users with a conversion prediction rate higher than the confidence level are taken as target potential customers.

[0037] Specifically, compressed sensing is a breakthrough data processing theory, the core idea of ​​which is to reconstruct complete data through a small amount of random sampling under the premise that the data is sparse or compressible. In the present invention, this theory is widely used in the process of data collection and dimensionality reduction. Specifically, in the data collection step, the present invention obtains user behavior information by integrating multi-source data. These data may come from the user's multi-platform interactive behavior, including but not limited to e-commerce browsing records, social media operation logs, and sensor data of mobile devices. When processing such data, traditional methods often face the problems of high storage pressure, low transmission efficiency, and information loss. The present invention optimizes the data collection process through compressed sensing technology and realizes accurate reconstruction of the original data using a small number of sampling points. The core advantage of compressed sensing is that it is based on random linear measurement, maps the compression of high-dimensional signals to low-dimensional space, and restores the data through a sparsity reconstruction algorithm in the subsequent stage. This process not only reduces the storage and computing burden of the cloud computing platform, but also lays a foundation for data quality for subsequent analysis and processing. After obtaining user behavior data, the present invention uses tensor modeling to structure multidimensional data. Unlike traditional matrix or table representations, a tensor is a data structure in the form of a multidimensional array that can simultaneously express multidimensional information such as the temporal characteristics of user behavior, behavior categories, and user attributes. In the present invention, user behavior data is modeled as a multi-order tensor, and each dimension represents a different characteristic of the data. For example, a typical third-order tensor may contain a user dimension, a behavior type dimension, and a time series dimension. This modeling method can not only retain the complex relationship between multi-source data, but also achieve data dimensionality reduction operations through tensor decomposition technology. Tensor decomposition, as an efficient dimensionality reduction method, is an important technical support for the data analysis part of the present invention. By decomposing multi-order tensors, the present invention extracts a core tensor, and this core feature representation retains the main information of the original data and significantly reduces the dimension of the data. Compared with traditional dimensionality reduction methods such as principal component analysis (PCA), tensor decomposition has more advantages in processing multidimensional data, and it can more comprehensively capture the multidimensional structural characteristics of data.

[0038] On the basis of the core tensor, the present invention further analyzes the time series characteristics of users. The time series characteristics of user behavior are an important basis for evaluating user activity and interest changes, which reflects the operation frequency and behavior patterns of users in different time periods. For example, some users may be highly active in a specific time period, which may be related to their shopping habits, work arrangements or hot spots of interest. The present invention provides key inputs for user similarity calculation and behavior propagation modeling by extracting these time series characteristics. Based on the time series characteristics of users, the present invention calculates the similarity between users, which is not limited to the matching of static attributes, but also comprehensively considers the dynamic similarity of users in behavior patterns. The calculation of user similarity lays the foundation for behavior propagation analysis, and the behavior propagation equation is a key tool for simulating the user behavior propagation process in the present invention. The design of the behavior propagation equation is another innovation of the present invention. The behavior propagation between users is similar to the spread of information diffusion or influence in social networks. It describes how the behavior of a certain user is propagated to other users through social connections or interest similarity. In the present invention, the behavior propagation equation constructs a mathematical model that describes the propagation of user behavior by combining user similarity and behavior dynamics. This model not only takes into account the behavioral characteristics of individual users, but also integrates the propagation laws between groups. For example, when the behavioral characteristics of user A are highly similar to those of user B, user B may be significantly affected by the behavior of user A. This process is manifested in a multi-user environment as the construction of a behavior propagation tensor, which reflects the diffusion pattern and influence range of user behavior in the group.

[0039] After the data analysis and processing is completed, the present invention enters the data fusion stage. The core task of this stage is to fuse the core tensor, time series features and behavior propagation tensor to generate fusion features with more predictive value. The significance of feature fusion is to comprehensively utilize information from multiple sources to avoid information loss or analysis bias that may be caused by a single feature. In order to achieve this goal, the present invention adopts a deep learning embedding method to map various features into a unified feature space. The construction of this feature space not only retains the key information of various features, but also extracts the implicit relationship between features through the optimization training of the deep learning network. For example, the core tensor may reflect the overall characteristics of the user, while the time series features reflect the dynamic changes of user behavior, and the behavior propagation tensor describes the interactive relationship between users. The fusion of these features can generate a more comprehensive user portrait and provide high-quality input for subsequent user prediction. In the final target user prediction stage, the present invention predicts the user's customer conversion rate based on the fusion features. Customer conversion rate prediction is a key step in the customer acquisition process, which directly affects the accuracy of target customer screening and customer acquisition efficiency. The present invention adopts a prediction model based on deep learning, which generates the conversion prediction rate of each user through deep learning of fusion features. This prediction process not only takes into account the characteristics of individual users, but also combines the propagation law of group behavior to generate more reliable prediction results. In order to further improve the stability and credibility of the prediction, the present invention introduces a confidence screening mechanism, that is, the conversion prediction rate of each user is compared with a preset confidence threshold, and only users above the threshold are selected as target potential customers. The confidence screening mechanism effectively avoids the problem of error accumulation in the prediction and ensures the accuracy of target customer screening.

[0040] Example 2: Multi-order tensors It is expressed by the following formula:

[0041]

[0042] Among them, multi-order tensors N represents the behavior feature value of user n in data source s, dimension d, and time t; s represents the total number of sampling points in the data source s; Ω is the sampling space, which represents the area where the integral is calculated and is defined as the spatial range where the user and the sampling points are located; B d,s (t) is the behavior characteristic function of dimension d in data source s at time t, which is a quantity describing behavior characteristics, including: click rate, purchase volume or number of visits; t i is the sampling time point; ∥x n -x i ∥2 indicates the user's location x n With the sampling point position x i The Euclidean distance ofi (t) represents the feature vector of sampling point i at time t, reflecting the specific characteristics of the sampling point of user behavior data, including: purchase records or browsing behavior; σ t is the standard deviation of the Gaussian distribution.

[0043] Specifically, in Embodiment 2 of the present invention, the formula is expressed by a multi-order tensor X n,d,t,s The construction of , describes in detail the dynamic characteristics of user behavior features in space, time and behavior dimensions in multi-source data scenarios. This formula embodies the core idea of ​​compressed sensing theory, that is, under the assumption of sparsity, high-dimensional behavior feature representation is reconstructed through limited data sampling, while combining the spatiotemporal correlation of user behavior and the complexity of multi-dimensional feature interaction, thus providing strong mathematical support for the accurate prediction of target users. In the formula, the tensor X n,d,t,s The behavioral feature values ​​of user n in data source s, dimension d and time t are defined. Its essence is a structured data representation in the form of a multidimensional array. By comprehensively modeling user behavior, it fully considers the complex relationship of multi-source data, especially the dynamic characteristics in time and space, so that details that are often ignored in traditional data analysis methods can be quantified and expressed. The construction of the formula links user behavior with its spatial environment in the form of integration. The integration area Ω represents the spatial range of the user and the sampling point. Through this integration operation, the formula can integrate the characteristics of the sampling point in a global scope, thereby capturing the relationship between the user and the surrounding environment. For example, a user's shopping behavior may be affected not only by its own characteristics, but also by the potential transmission effect of the behavior of other nearby users. By weighted summarizing the behavioral characteristics of the sampling points in the integration area, the formula can construct a global user behavior feature representation, and the construction of this feature depends on multiple important variables in the formula.

[0044] Euclidean distance in the formula ∥x n -x i∥2 is used to characterize the spatial relationship between the user location and the sampling point, reflecting the spatial dependence of user behavior. This distance factor introduces a natural weighting mechanism, that is, the sampling points closer to the user contribute more to their behavioral characteristics, while the influence of the sampling points farther away gradually weakens. This weight distribution conforms to the propagation law of user behavior in reality. For example, a user may be more easily affected by promotional activities at nearby shopping points, but rarely participate in activities that are farther away. Through the weighted calculation of the Euclidean distance, the formula reflects the spatial sparsity in behavior modeling, avoids excessive reliance on the behavioral characteristics of non-related sampling points, and reduces the computational complexity, providing computational efficiency guarantees for subsequent data processing. The dynamic changes in time are reflected by the Gaussian decay function in the formula. This function describes the correlation of user behavior in the time dimension, that is, the influence of sampling point behavior gradually weakens over time. This Gaussian decay mechanism is achieved through the parameter σ t Controlling the diffusion range of time can accurately capture the temporal patterns of user behavior. For example, in an e-commerce scenario, a user's purchasing behavior may be strongly affected by short-term promotional activities, and this influence gradually fades after the activity ends. By introducing Gaussian distribution, the formula takes into account the temporal dynamics and sparsity characteristics of user behavior in temporal modeling, so that behavior propagation can more realistically reflect the user's actual behavior patterns. This temporal dynamic modeling not only improves the expressive power of multi-order tensors, but also lays the foundation for the subsequent calculation of user similarity and behavior propagation modeling.

[0045] In addition, the formula also passes through the eigenvector v i (t) and characteristic function B d,s (t) captures the characteristics of the sampling points and data sources, further enriching the semantic information represented by the tensor. i (t) describes the behavioral characteristics of the sampling point at a specific time, such as purchase records or browsing behavior, and its role is to give each sampling point a unique behavioral identifier. d,s(t) describes the global characteristics of the behavior from the perspective of dimension and data source, such as click-through rate, visit volume, etc. The introduction of this function realizes the unified modeling of different data sources and behavior dimensions, and facilitates the fusion of multi-source data. In the formula, these two features are integrated with the user behavior characteristics through the interaction of time and space, thereby constructing a more comprehensive user behavior expression. Compared with traditional behavior analysis technology, this user behavior modeling method based on multi-order tensors has significant innovation and advantages. Traditional methods usually only consider user behavior in a single dimension or time point, ignoring the dynamic characteristics of user behavior in multi-dimensional space, while the present invention realizes the comprehensive quantification of user behavior in time, space and data source dimensions through tensor modeling and formula construction. In particular, under the guidance of compressed sensing theory, the present invention utilizes the sparsity characteristics of data to reconstruct a high-dimensional tensor representation through the behavior of limited sampling points, greatly reducing the complexity of data collection and storage. In addition, through the distance factor and Gaussian attenuation function in the formula, the present invention effectively combines the spatial and temporal dependence of user behavior, making the behavior modeling more in line with the actual scene, and providing more accurate and reliable input data for subsequent user behavior prediction.

[0046] Embodiment 3: In the data analysis and processing step, the multi-order tensor is decomposed by using Tucker decomposition to obtain a core tensor, and the formula is as follows:

[0047]

[0048] Among them, the core tensor The following sparse constraints are satisfied:

[0049]

[0050] Among them, U N is the user modal basis matrix, U D is the dimension modal basis matrix, U T is the time modal basis matrix, U S is the modal basis matrix of the data source, which is a preset matrix; k is the sparsity parameter; ∥·∥0 represents the L0 norm; ∥·∥ F represents the Frobenius norm; rank(·) is the rank operation of the matrix; Represents the core tensor of user n. When it is user i, it is expressed as

[0051] Specifically, Tucker decomposition is a classic tensor decomposition method. By decomposing high-dimensional data into a low-dimensional core tensor and the product of several modal basis matrices, it effectively extracts the core feature information of the data and significantly reduces the dimension of the data, thereby providing a more concise and efficient data representation for subsequent user behavior analysis and target user screening. The core idea of ​​this formula is to represent the original multi-order tensor as a linear mapping of each modal basis matrix on its corresponding modality, thereby constructing a sparse low-rank core tensor. By mapping the basis matrices representing the user modality, dimensional modality, time modality, and data source modality, the noise, redundant information, and low-correlation data in the original tensor are filtered out, and the core features of the data are retained. In actual operation, the core tensor is designed to meet the sparse constraint conditions, and its sparsity is controlled by the parameter k. The design of this sparsity constraint is based on the core assumption of compressed sensing theory, that is, user behavior data is sparse in multidimensional space, and only a few features have a significant impact on user behavior. Through this constraint, the core tensor can not only compress high-dimensional data, but also retain important information in the process of dimensionality reduction, eliminate irrelevant or noisy data, and ensure the accuracy and effectiveness of the data. Especially when user behavior involves multi-source heterogeneous data, this constraint plays an important role in reducing the dimensionality curse and improving data processing efficiency. Furthermore, the ∥x in the sparse constraint formula n,d,t,s ∥ and rank(x n,d,t,s ) respectively measure the data intensity and rank of the tensor. Data intensity describes the activity level of user behavior in a specific scenario, such as the number of purchases, browsing depth, etc.; while the rank of the tensor indicates the degree of linear independence of the data. By combining these two factors, the sparse constraint dynamically adjusts the sparsity of the core tensor, so that it can adapt to the complexity of the data and effectively control the dimensionality reduction effect of the data. This dynamic adjustment mechanism enables the present invention to flexibly adapt to different data scenarios, whether it is a cold start scenario with strong behavioral sparsity or a large-scale recommendation system environment with complex data dimensions, it can achieve excellent performance. The intelligent cloud data collection and dimensionality reduction customer acquisition method based on compressed sensing here realizes efficient compression and sparse representation of data through Tucker decomposition, thereby greatly reducing the computational complexity of multi-source data processing. At the same time, the core tensor, as a low-dimensional core representation of behavioral features, provides more reliable and accurate data support for subsequent behavioral propagation modeling and target user conversion prediction. Compared with the traditional tensor decomposition method, the present invention introduces sparsity constraints, combines the sparsity advantages of compressed sensing theory with the dimensionality reduction capability of Tucker decomposition, solves the noise interference and redundancy problems of high-dimensional data in the analysis process, and fundamentally improves the utilization efficiency and feature expression capability of multi-source heterogeneous data.

[0052] Embodiment 4: The data analysis and processing step constructs a time series feature based on the decomposed core tensor through the following formula:

[0053]

[0054] Wherein, tra represents the transpose operation of the matrix; τ is the time integral variable; tr(·) represents the operation of the trace of the matrix; W(t-τ) represents the time weighting matrix when the time is t-τ; W(t) represents the time weighting matrix when the time is t, which is preset; represents the time series characteristics of user n. When it is user i, it is expressed as Represents the core tensor time gradient.

[0055] Specifically, the core of the formula is to use the core tensor to describe the main features of user behavior in multidimensional space, and at the same time adjust the influence of behavior in time through the time weighting matrix W(t-τ). The time weighting matrix is ​​essentially a preset function used to assign weights to behaviors at different time points, and this weight changes dynamically with time differences. For example, when a user's behavior occurs at a time point closer to the current time t, the weighting matrix will assign it a greater weight, reflecting its stronger influence on the current moment; while for time points farther away, its influence gradually decays, which conforms to the time-dependent characteristics of actual user behavior. Such a time-weighted design not only makes the formula flexible in time adjustment, but also reduces the interference of long-term behavior noise on feature modeling to a certain extent, thereby enhancing the ability to capture short-term behavior trends. This is of great significance in intelligent customer acquisition scenarios. For example, by analyzing the user's recent visit frequency and behavior intensity, their purchase intention can be more accurately predicted. In order to further refine the analysis of user behavior changes, the formula introduces the time gradient of the core tensor and measures the intensity of changes in the time dimension through its Frobenius norm. The time gradient is a key indicator that describes the rate of change of the kernel tensor in the time dimension, and can keenly reflect the mutation or stability of user behavior in time. For example, when a user suddenly increases the frequency of a certain behavior (such as browsing a certain type of product or participating in a certain type of activity) over a period of time, the time gradient will show a larger value, indicating a significant change in behavior; on the contrary, when the behavior remains stable, the value of the time gradient will be smaller. By introducing this gradient into the formula, the time series feature can capture the changing trend of user behavior and dynamically adjust it according to the impact of the change on the behavior. This design effectively solves the problem that traditional time series analysis methods cannot take into account both the stability and mutation of behavior, so that the modeling results can not only describe the long-term stable trend of users, but also focus on significant short-term behavioral changes.

[0056] On this basis, the formula further controls the time gradient through an exponential decay function to ensure the robustness and sparsity of the time series features. The design of the exponential decay function makes full use of the sparsity assumption in the theory of compressed sensing, that is, significant changes in user behavior are often concentrated in a few key time points, while the behavior in most time periods is relatively stable. Specifically, when the time gradient is large, the exponential decay factor will be significantly reduced, weakening the impact of mutation behavior on the overall time series features to avoid the interference of abnormal fluctuations; when the time gradient is small, the exponential decay factor is close to 1, allowing stable behavior to make a greater contribution to the time series features. Through this mechanism, the formula can capture behavioral changes while suppressing excessive behavioral fluctuations, making the constructed time series features more stable and reliable. In addition, the trace operation of the time modal basis matrix is ​​introduced in the denominator of the exponential decay to measure the feature complexity in the time dimension, thereby further optimizing the adjustment range of the time gradient. As an important attribute of the matrix, the trace operation reflects the overall characteristics of the kernel tensor in the time dimension, providing a mathematical basis for the optimization of the formula in time dynamic control.

[0057] By decomposing and controlling the sparsity of the core tensor, the construction of time series features achieves data compression and dimensionality reduction, providing lightweight input for subsequent behavior propagation modeling and conversion rate prediction. At the same time, the combination of time weighted matrix and time gradient enables time series features to more finely characterize the evolution of user behavior over time, providing support for the dynamic update of user portraits. This dynamic modeling method is particularly suitable for real-time analysis needs in large-scale data environments, such as capturing users' purchase intentions in real time on e-commerce platforms or monitoring users' interaction trends on social media platforms. In addition, through the sparsity constraints of exponential decay and trace operations, the formula effectively reduces the interference of noise on the modeling results, making the calculation of time series features not only more efficient, but also more robust, and able to adapt to complex analysis needs in multi-source heterogeneous data scenarios. Compared with traditional static feature analysis methods, time series features can fully reflect the dynamic changes of user behavior over time, providing richer information for the accurate prediction of target users. For example, in the customer acquisition scenario, by analyzing the behavior change trend of potential customers in the short term, combined with the modeling results of time series features, customers' purchase intentions and conversion possibilities can be more accurately judged. Especially in an environment with huge data volumes and complex behavioral patterns, this formula significantly improves data processing efficiency and analysis accuracy through the sparsity assumption of compressed sensing, providing theoretical support and technical guarantee for realizing intelligent user analysis and precise customer acquisition.

[0058] Embodiment 5: The data analysis and processing step calculates the similarity between users according to the time series characteristics through the following formula, and calculates the similarity between users according to the similarity between users

[0059]

[0060] Among them, <·,·> represents the tensor inner product operation; U N (:,i) represents the basis vector of user i; U N (:,j) represents the basis vector of user j; σ s is the similarity adjustment parameter, which is the preset value; represents the temporal characteristics of user i; represents the temporal characteristics of user j; Represents the similarity between user i and user j; all similarities between users are taken as one element to form the user similarity matrix for The element in row i and column j.

[0061] Specifically, in the formula, the similarity of user time series features is calculated by normalizing the inner product of the tensor and the Frobenius norm. and They describe the behavior dynamics of user i and user j at time t respectively. Their inner product reflects the consistency of the direction of behavior change of the two users, while the norm normalization The influence of feature magnitude is eliminated, so that the calculation results focus on the similarity of behavior patterns. This similarity calculation method reveals the degree of consistency of the behavior trends of two users in a certain period of time by comparing the dynamic characteristics of user behavior. For example, when the purchase behavior, browsing frequency or interaction pattern of users i and j at time t have the same change trend, the similarity of time series features will tend to a higher value, indicating that they may have similar interests or needs. On the other hand, the formula further considers the relative position relationship of users in the basis vector space by introducing the Euclidean distance of user basis vectors. The basis vector describes the global characteristics of the user, such as long-term behavioral preferences or basic attribute differences, while the square value of the distance quantifies the feature differences between the two users. By combining this distance information in the exponential decay function, the formula can dynamically adjust the impact of the basis vector difference on the similarity. Specifically, when the basis vectors of two users are close, the similarity of their behavior characteristics will be amplified; when the user basis vectors are very different, the similarity will decrease rapidly with the increase of distance. This design theoretically integrates the dual information of temporal dynamic features and global static features, providing a more comprehensive perspective for multi-dimensional similarity modeling of user behaviors.

[0062] σ in the formula S As a similarity adjustment parameter, it is used to control the sensitivity of basis vector differences to similarity. S The smaller the value, the weaker the impact of the basis vector difference, that is, more emphasis is placed on the temporal dynamic similarity of users; while the smaller σ SA value of σ will emphasize the global difference of the basis vectors, thus tending to screen out users with high similarity in long-term behavior patterns. This parameter setting is highly flexible and can be optimized according to specific application scenarios. For example, in short-term marketing activities, you may pay more attention to the user's recent behavior characteristics. In this case, you need to set a larger σ S In the analysis of long-term user loyalty, a smaller σ may be required. S value to highlight the global features. Finally, the similarity between all users is integrated into a user similarity matrix Each element represents the behavioral similarity between user i and user j at time t. This matrix not only intuitively shows the relationship between all users, but also provides key input for behavior propagation modeling. For example, a high-similarity user group in the matrix may represent a potential customer group with similar interests or needs, while a low-similarity user may reflect personalized preferences and behavioral characteristics. This matrix representation based on user similarity can well adapt to subsequent behavior propagation modeling and user grouping analysis, and provides an efficient data structure for accurate customer acquisition. Through the calculation formula of user similarity, the dynamic features, static features and time change trends of multidimensional data are effectively integrated, breaking through the limitation of traditional similarity analysis that only focuses on a single dimension. This formula not only realizes the combination of time series features and global features in theory, but also significantly improves the accuracy and robustness of similarity analysis in practice. For example, compared with simple cosine similarity or Euclidean distance, this formula captures users' short-term behavior trends through time series features, and combines basis vector differences to reflect users' long-term characteristics, making the calculation results of user similarity closer to real needs. Especially in large-scale data environments, kernel tensor dimensionality reduction based on compressed sensing theory greatly reduces the computational complexity, making the generation of user similarity matrix both efficient and accurate.

[0063] Embodiment 6: The data analysis and processing step constructs a behavior propagation equation based on the similarity between users through the following formula:

[0064]

[0065] in, is the Laplace operator; represents the behavior propagation tensor of user n. When it is user i, it is expressed as div(·) is the divergence operator.

[0066] Specifically, the first part of the formula is expressed by the Laplace operator Describes the diffusion characteristics of user behavior. The diffusion process here is composed of the similarity matrix between users and timing characteristics Determined jointly. The user similarity matrix reflects the association of behavioral patterns between users, while the timing characteristics describe the dynamic characteristics of user behavior over time. The role of the Laplace operator here is to measure the local rate of change of behavior between users and simulate the propagation of behavior between similar users. For example, when users i and j have a high similarity in the similarity matrix, the diffusion rate of behavior propagation will increase significantly, thereby accelerating the diffusion of the behavior in high-similarity user groups. In contrast, for user groups with lower similarity, the diffusion rate will be naturally suppressed. This diffusion mechanism can well simulate the law of user behavior propagation in social networks, such as the rapid diffusion of hot product recommendations or promotional information among similar user groups. In addition, the diffusion term also adjusts the intensity of the diffusion through the Frobenius norm of the modal basis matrix, so that it can adapt to the propagation needs of different time scales and behavioral characteristics. The second part of the formula is obtained through the kernel tensor and the temporal gradient of the temporal features Establish the interaction term of behavior propagation. This interaction term describes how the global characteristics of user behavior affect its local temporal dynamics in a physical sense. It is the core representation of multimodal features. It extracts the main features of user behavior through compressed sensing and tensor decomposition, representing the global characteristics of users in multiple dimensions. It reflects the rate of change of user behavior over time. For example, when a user's recent purchasing behavior suddenly increases or decreases, its time gradient will change significantly. The interaction term combines these two to describe how global features regulate local time changes. For example, a user's long-term interest preferences may determine the direction of change in his short-term behavior. This global and local association is mathematically expressed in the interaction term. At the same time, this term adjusts the interaction intensity through the ratio of the basis matrix to ensure the adaptability of behavior propagation under different user feature dimensions. The third part of the formula uses the divergence operator Describes the change in the distribution of the kernel tensor gradient under similarity regulation. The role of the divergence operator here is to characterize the fluidity of behavior propagation in the spatial dimension, that is, how the similarity between users affects the path and range of behavior propagation. For example, in a high-similarity user group, behavioral characteristics will propagate between users at a higher frequency, while in a low-similarity user group, this propagation will be significantly suppressed. Through this divergence term, the formula can simulate the flow direction and intensity of user behavior in space, so that the modeling of behavior propagation is not only limited to the time dimension, but also can achieve a more accurate description in space. This term also regulates the divergence process through the ratio of the basis matrix to adapt to the propagation needs in different data source scenarios.

[0067] Embodiment 7: The data fusion processing step is used to fuse the core tensor, the time series feature and the behavior propagation tensor to obtain the fusion feature by the following formula:

[0068]

[0069] in, represents the fusion feature of user n. When it is user i, it is expressed as express The time gradient of for The corresponding form under the time-integrated variable; for The corresponding form under the time-integrated variable; for The corresponding form under the time-integrated variable; It is the tensor outer product operation.

[0070] Specifically, the core idea of ​​the formula is to use the tensor outer product Combine multiple features to construct a high-dimensional feature tensor The introduction of the outer product operation makes the core tensor The multimodal global characteristics and temporal features represented Dynamic changes and propagation tensors The behavior diffusion information of can be integrated simultaneously. Through this joint representation, the formula can capture the interactive relationship of user behavior in multiple dimensions. For example, the kernel tensor Can reflect users’ long-term preferences and temporal characteristics describes the behavior changes of users in a specific time period, and the propagation tensor It reveals the diffusion law of user behavior in the group. The outer product combination of the three unifies this information in a high-dimensional feature space, so that the fused feature not only contains the independent characteristics of each mode, but also reflects the complex interaction between the modes. The time integration operation in the formula further enhances the dynamics of the fused feature. By integrating the time variable τ, the formula can globally summarize the changing trend of user behavior in the time dimension, thereby generating a fusion representation that comprehensively reflects the full cycle characteristics of user behavior. For example, for a user who has undergone significant behavioral changes in a certain period of time, the time integration can accurately capture the cumulative impact of these changes, while for users with relatively stable behavior, it can highlight the stability of their long-term characteristics. In addition, the integration process adjusts the influence of behavior in time by introducing dynamic weights, so that recent behavior contributes more to the fused feature, while the influence of behavior at more distant time points gradually decays. This dynamic weighting mechanism ensures the timeliness and pertinence of the fused feature.

[0071] Time gradient plays the role of sparsity constraint in the formula. By propagating tensors The formula can measure the rate of change of user behavior propagation in the time dimension by calculating the time gradient of the time gradient. When the time gradient is large, it indicates that there are rapid changes in the behavior propagation process. At this time, the exponential decay factor will be significantly reduced, suppressing the interference of drastic fluctuations on the fusion features; when the time gradient is small, the exponential decay factor tends to 1, allowing the smooth propagation process to make a greater contribution to the fusion features. This design is based on the sparsity assumption of compressed sensing theory, that is, significant changes in user behavior are usually concentrated in a few key time points, while the propagation process in most time periods is relatively smooth. By controlling the sparsity of the time gradient, the formula effectively avoids the negative impact of abnormal data or drastic fluctuations on the fusion features, making the generated features more robust. In addition, the formula normalizes the factor The amplitude of exponential decay is adjusted to balance the weight relationship between the behavior propagation gradient and the dynamics of the temporal features. The normalization factor reflects the overall change intensity of the temporal features, making the adjustment of exponential decay more adaptive. For example, when the user's temporal features change greatly, the normalization factor can amplify the inhibitory effect of the temporal gradient, thereby highlighting the contribution of important behavioral changes to the fusion features; when the temporal features change less, the normalization factor weakens the influence of the temporal gradient, thereby maintaining the stability of the fusion features.

[0072] Fusion Features The construction provides a comprehensive and efficient user behavior representation for accurate customer acquisition. By dynamically fusing the core tensor, time series features and behavior propagation tensor in the time dimension, the formula can not only capture the behavioral characteristics of users in different modes, but also reflect the evolution of behavior in time and space. This multi-dimensional, multi-modal feature representation provides an important basis for the construction of user portraits and the prediction of customer conversion rates. For example, in e-commerce scenarios, fused features can be used to analyze users' long-term preferences and recent behavioral changes, and help predict their response to specific promotional activities; on social media platforms, fused features can be used to identify user groups with similar propagation laws, thereby optimizing information push and advertising strategies. Compared with traditional feature fusion methods, the formula of the present invention realizes the organic unity of global features, time series features and behavior propagation information through the combination of tensor outer product and time dynamics. In particular, through the sparsity constraints of the time gradient and the dynamic weighting mechanism, the formula captures significant changes in user behavior while avoiding the interference of noise data, making the fused features more stable and accurate in complex data environments. This design has significant technical advantages in intelligent cloud data collection and analysis. It can not only adapt to the processing needs of multi-source heterogeneous data, but also provide strong technical support for target user screening and behavior prediction in precise customer acquisition.

[0073] Embodiment 8: The target user acquisition processing step uses the following formula to predict the customer conversion rate of each user based on the fusion features to obtain the conversion prediction rate r of the user to the target customer: i (t+Δt):

[0074]

[0075] Among them, r i (t+Δt) represents the conversion prediction rate of user i, which means the conversion prediction rate of the user who can be converted into the target customer after the window time Δt from time t; for The corresponding form under the time-integrated variable; for The corresponding form under the time-integrated variable; for The corresponding form under the time-integrated variable; is the dynamic weight based on the gamma function.

[0076] Specifically, the formula It is the core probability control part, used to calculate the fusion feature F i (τ) and the behavior propagation tensor P i (τ) and maps it to the probability space of [0,1]. Fusion feature F i (τ) integrates the user's global characteristics, temporal dynamics, and propagation laws, while the behavior propagation tensor P i (τ) describes the spread and intensity of user behavior in the group. The ratio of the two reflects the balance between the independence and spread of user behavior characteristics. For example, when the spread of user behavior is large (high P i (τ)) but the fusion feature is small (low F i (τ)), the conversion rate may be low, indicating that user behavior may be more driven by group effects; conversely, a higher fusion feature value indicates that individual user behavior is more independent and the conversion rate may be higher. By nonlinearly mapping the sigmoid function to the contrast value, the formula can smoothly handle the impact of different feature combinations on the conversion rate. Secondly, the gamma function weight The introduction of gives the formula dynamic control capabilities. The form of the gamma function combines the temporal feature gradient The proportional relationship with the fusion feature reflects the impact of the user behavior change rate on the conversion rate. ), the gamma function weights will be adjusted accordingly to highlight the importance of these changes in conversion rate prediction; when the behavior changes are relatively stable, the weights tend to be smooth, reducing the interference of unnecessary fluctuations on the prediction results. In addition, the gamma function parameter k, as a control factor, can be flexibly adjusted according to the application scenario. For example, in a highly dynamic environment, increase the k value to enhance the influence of behavioral changes, or in a stable scenario, reduce the k value to highlight long-term behavioral characteristics. The exponential decay part of the formula further smoothes the impact of the time gradient on the conversion rate. In the design, when the gradient of the time series feature is large, the attenuation factor is reduced, thereby limiting the amplification effect of abnormal behavior changes on the prediction results; when the time series feature changes are small, the attenuation factor is close to 1, making the stability of the behavior contribute more to the prediction results. This mechanism is achieved through the matrix trace The balance between behavioral dynamics and data source characteristics is adjusted. For example, for user behavior in complex data sources, the matrix trace can dynamically adjust the amplitude of the attenuation factor to ensure the adaptability of the formula to different data sources. The design of this conversion rate prediction formula comprehensively integrates the individual characteristics of users and the laws of group behavior, providing a scientific basis for accurate customer acquisition. By capturing the dynamic changes of user behavior within the window period through time integration, the formula can predict the probability of users converting into target customers in the future. For example, on e-commerce platforms, users' recent browsing and purchasing behaviors and their dissemination effects can be analyzed to accurately assess their likelihood of participating in promotional activities; in social networks, users' response to information dissemination can be predicted through behavioral dissemination and group effects. Compared with traditional static prediction methods, this formula has significant advantages in dynamic behavior modeling and multimodal feature fusion.

[0077] Example 9: Dynamic weight based on gamma function Use the following formula to express it:

[0078]

[0079] Specifically, the formula of the dynamic weight of the gamma function depends on the integral lower limit, which is determined by the ratio of the temporal gradient of the user behavior to the strength of the fused feature, that is, This quantitative index comprehensively reflects the significance of the change in user behavior and the stability of the overall characteristics. For users with large behavioral changes, their time gradient values ​​are high, which makes the decay rate of the gamma function weights relatively fast, thereby highlighting the contribution of these behavioral changes in the weight distribution; while for users with relatively stable behavior, the decay of the gamma function weights is slow, so that the weight distribution more evenly reflects the long-term stability of their behavior. The core idea of ​​this dynamic adjustment of weight distribution is highly consistent with the sparsity assumption of the compressed sensing theory in the present invention, that is, by highlighting a few significant behavioral change points, reducing the influence of unimportant features, and ultimately achieving accurate modeling of the target behavior. The role of the sparsity parameter k in the gamma function further enhances the flexibility and adaptability of sparsity regulation. When k>1, the gamma function pays more attention to the sparsity of the weight distribution, that is, by emphasizing the contribution of significant behavioral change points, suppressing the influence of minor behavioral characteristics. This design is particularly applicable in fast-response business scenarios, such as user conversion prediction during marketing activities. In this scenario, it is necessary to quickly identify the target users who respond most actively to the activity, and a larger k value can ensure that the model preferentially captures these user characteristics with significant behavior. When k≤1, the weight distribution of the gamma function tends to be smoother, focusing more on the overall trend of behavior, which is more advantageous when analyzing users' long-term behavior patterns or loyalty. For example, in the member analysis of an e-commerce platform, a smaller k value can highlight the user's continuous behavioral characteristics and provide a more reliable basis for user stratification and long-term value prediction. The integral form of the gamma function further reflects the important role of time gradient in dynamic weight distribution. Time gradient It is a key indicator for measuring the rate of change of user behavior. Its value directly affects the setting of the lower limit of the gamma function integral, which has a significant impact on the dynamic weight distribution. When the time gradient of user behavior is large, it indicates that their behavior has changed rapidly in a short period of time. At this time, the gamma function weight will automatically concentrate on these time points, highlighting the importance of these rapid changes in conversion prediction; for time periods with slower behavior changes, the gamma function will assign lower weights to avoid the interference of stable behavior on the overall prediction. This nonlinear weight distribution mechanism enables the model to more accurately capture key changes in user behavior, thereby showing higher accuracy and flexibility in target user screening.

[0080] The introduction of dynamic weights of the gamma function not only enhances the modeling capability of sparse features, but also plays a core role in the temporal evolution analysis of dynamic features. Compared with the traditional static weight method, the gamma function realizes the dynamic adaptation capability of feature weight distribution through the joint regulation of the sparsity parameter k and the time gradient. This capability is particularly suitable for the analysis scenario of multi-source heterogeneous data, in which user behavior often has high temporal dynamics and feature complexity. For example, in the analysis of e-commerce promotions, the gamma function can dynamically adjust the weights according to the degree of response of user behavior to promotional activities, thereby providing more accurate support for the identification of high-value customers and the formulation of conversion strategies. The dynamic weights of the gamma function also show important robustness in dealing with abnormal behavior and noisy data. Through the joint action of the integral lower limit and the sparsity parameter, the gamma function can effectively suppress the negative impact of abnormal behavior on the model prediction results. For example, when a user has abnormal short-term high-frequency behavior in a certain period of time, the nonlinear attenuation mechanism of the gamma function weight can quickly reduce the weights of these abnormal points, thereby avoiding the distortion of the model prediction results. This mechanism is particularly important in multi-source heterogeneous data scenarios, because in such scenarios, data noise and abnormal behavior are inevitable. Through the dynamic weight distribution of the gamma function, the model can enhance its ability to handle abnormal data without affecting the overall prediction accuracy.

[0081] Although the specific embodiments of the present invention are described above, it should be understood by those skilled in the art that these specific embodiments are only illustrative, and those skilled in the art may omit, replace, and change the details of the above methods and systems in various ways without departing from the principles and essence of the present invention. For example, merging the above method steps so as to perform substantially the same functions in substantially the same manner to achieve substantially the same results is within the scope of the present invention. Therefore, the scope of the present invention is limited only by the appended claims.

Claims

1. Intelligent cloud data collection and dimensionality reduction customer acquisition method based on compressed sensing, characterized in that: The method comprises: a data collection step, a data analysis and processing step, a data fusion processing step and a target user acquisition processing step; the data collection step is used to obtain user behavior data from multi-source data and represent the user behavior data as a multi-order tensor; the data analysis and processing step is used to perform tensor decomposition on the multi-order tensor to obtain a core tensor, construct a time series feature based on the decomposed core tensor, calculate the similarity between users according to the time series feature, and construct a behavior propagation equation based on the similarity between users according to the similarity between users to obtain a behavior propagation tensor; the data fusion processing step is used to perform feature fusion on the core tensor, the time series feature and the behavior propagation tensor to obtain a fusion feature; the target user acquisition processing step is used to predict the customer conversion rate of each user based on the fusion feature to obtain the conversion prediction rate that the user can be converted into a target customer; the conversion prediction rate of each user is compared with a set confidence level, and users with a conversion prediction rate higher than the confidence level are taken as target potential customers.

2. The method for intelligent cloud data collection and dimensionality reduction based on compressed sensing according to claim 1, characterized in that: Multi-rank tensors It is expressed by the following formula: Among them, multi-order tensors N represents the behavior feature value of user n in data source s, dimension d, and time t; s represents the total number of sampling points in the data source s; Ω is the sampling space, which represents the area where the integral is calculated and is defined as the spatial range where the user and the sampling points are located; B d,s (t) is the behavior characteristic function of dimension d in data source s at time t, which is a quantity describing behavior characteristics, including: click rate, purchase volume or number of visits; t i is the sampling time point; ∥x n -x i ∥2 indicates the user's location x n With the sampling point position x i The Euclidean distance of i (t) represents the feature vector of sampling point i at time t, reflecting the specific characteristics of the sampling point of user behavior data, including: purchase records or browsing behavior; σ t is the standard deviation of the Gaussian distribution.

3. The method for intelligent cloud data collection and dimensionality reduction based on compressed sensing according to claim 2, characterized in that: In the data analysis and processing step, the multi-order tensor is decomposed by using Tucker decomposition to obtain a core tensor, and the formula is as follows: Among them, the core tensor The following sparse constraints are satisfied: Among them, U N is the user modal basis matrix, U D is the dimension modal basis matrix, U T is the time modal basis matrix, U S is the modal basis matrix of the data source, which is a preset matrix; k is the sparsity parameter; ∥·∥0 represents the L0 norm; ∥·∥ F represents the Frobenius norm; rank(·) is the rank operation of the matrix; Represents the core tensor of user n. When it is user i, it is expressed as 4. The method for intelligent cloud data collection and dimensionality reduction based on compressed sensing according to claim 3, characterized in that: The data analysis and processing step constructs the time series feature based on the decomposed core tensor through the following formula: Wherein, tra represents the transpose operation of the matrix; τ is the time integral variable; tr(·) represents the operation of the trace of the matrix; W(t-τ) represents the time weighting matrix when the time is t-τ; W(t) represents the time weighting matrix when the time is t, which is preset; represents the time series characteristics of user n. When it is user i, it is expressed as Represents the core tensor time gradient.

5. The method for intelligent cloud data collection and dimensionality reduction based on compressed sensing according to claim 4, characterized in that: The data analysis and processing step calculates the similarity between users according to the time series characteristics through the following formula, and calculates the similarity between users according to the similarity between users Among them, <·,·> represents the tensor inner product operation; U N (:,i) represents the basis vector of user i; U N (:,j) represents the basis vector of user j; σ s is the similarity adjustment parameter, which is the preset value; represents the temporal characteristics of user i; represents the temporal characteristics of user j; Represents the similarity between user i and user j; all similarities between users are taken as one element to form the user similarity matrix for The element in row i and column j.

6. The method for intelligent cloud data collection and dimensionality reduction based on compressed sensing according to claim 5, characterized in that: The data analysis and processing step constructs a behavior propagation equation based on the similarity between users through the following formula: in, is the Laplace operator; represents the behavior propagation tensor of user n. When it is user i, it is expressed as div(·) is the divergence operator.

7. The method for intelligent cloud data collection and dimensionality reduction based on compressed sensing according to claim 6, characterized in that: The data fusion processing step is used to fuse the core tensor, the time series feature and the behavior propagation tensor to obtain the fusion feature through the following formula: in, represents the fusion feature of user n. When it is user i, it is expressed as express The time gradient of for The corresponding form under the time-integrated variable; for The corresponding form under the time-integrated variable; for The corresponding form under the time-integrated variable; It is the tensor outer product operation.

8. The method for intelligent cloud data collection and dimensionality reduction based on compressed sensing according to claim 7, characterized in that: The target user acquisition processing step uses the following formula to predict the customer conversion rate of each user based on the fusion features to obtain the conversion prediction rate r of the user to the target customer. i (t+Δt): Among them, r i (t+Δt) represents the conversion prediction rate of user i, which means the conversion prediction rate of the user who can be converted into the target customer after the window time Δt from time t; for The corresponding form under the time-integrated variable; for The corresponding form under the time-integrated variable; for The corresponding form under the time-integrated variable; is the dynamic weight based on the gamma function.

9. The method for intelligent cloud data collection and dimensionality reduction based on compressed sensing according to claim 8, characterized in that: Dynamic weights based on gamma function Use the following formula to express it:

Images