Recommendation strategy determination method and device based on multi-source heterogeneous data and electronic equipment

Abstract

The invention discloses a recommendation strategy determination method and device based on multi-source heterogeneous data and electronic equipment. The method comprises the steps that multi-source heterogeneous data of a user is obtained, and the multi-source heterogeneous data comprises data of multiple sources and types associated with the user; performing feature fusion on the multi-source heterogeneous data to obtain fusion features; performing prediction according to the fusion features through a deep learning network model to obtain a prediction result of the user, and generating an initial recommendation strategy according to the prediction result, the initial recommendation strategy being at least used for indicating products or services recommended to the user; feedback data is received, the initial recommendation strategy is adjusted according to the feedback data, a target recommendation strategy is obtained, and the feedback data is used for indicating the response of the user to the recommended content in the initial recommendation strategy. The technical problem that the accuracy of the generated recommendation strategy is low due to dependence on manual judgment and static rules is solved.

Description

Technical Field

[0001] This application relates to the field of data processing, and more specifically, to a method, apparatus, and electronic device for determining recommendation strategies based on multi-source heterogeneous data. Background Technology

[0002] Business expansion and customer management methods for related technologies face multiple challenges. With the advent of the big data era, massive amounts of heterogeneous user data from multiple sources have accumulated. However, limited by data processing technologies and algorithms, the value of this data has not been fully realized, making it difficult to meet personalized and dynamic recommendation needs. Specifically, existing technologies have the following main shortcomings:

[0003] Segmenting users based on simple clustering algorithms struggles to capture dynamic user changes and hinders the development of highly accurate recommendation strategies. It also has limitations in integrating external market information and industry dynamics, failing to achieve intelligent fusion of internal and external data. Customer opportunity identification in determining recommendation strategies relies heavily on the experience and intuition of account managers or on static and simple rule configurations, lacking flexibility and intelligence, and failing to meet personalized and dynamic recommendation needs. This reliance on human judgment is not only inefficient but also prone to bias, resulting in low accuracy of recommendation strategies, poor recommendation effects, low user conversion rates, and ultimately hindering the effective expansion of business and the increase of market share.

[0004] There is currently no effective solution to the above problems. Summary of the Invention

[0005] This application provides a method, apparatus, and electronic device for determining recommendation strategies based on multi-source heterogeneous data, so as to at least solve the technical problem that the accuracy of the generated recommendation strategies is low due to reliance on human judgment and static rules.

[0006] According to one aspect of the embodiments of this application, a method for determining a recommendation strategy based on multi-source heterogeneous data is provided, comprising: acquiring multi-source heterogeneous data of a user, wherein the multi-source heterogeneous data includes data from multiple sources and of various types associated with the user; performing feature fusion on the multi-source heterogeneous data to obtain fused features; performing prediction based on the fused features using a deep learning network model to obtain a prediction result for the user, and generating an initial recommendation strategy based on the prediction result, wherein the initial recommendation strategy is at least used to indicate products or services recommended to the user; receiving feedback data, and adjusting the initial recommendation strategy based on the feedback data to obtain a target recommendation strategy, wherein the feedback data is used to indicate the user's response to the content recommended in the initial recommendation strategy.

[0007] Optionally, feature fusion is performed on the multi-source heterogeneous data to obtain fused features, including: preprocessing the multi-source heterogeneous data to obtain preprocessed multi-source heterogeneous data; mapping the high-dimensional data in the preprocessed multi-source heterogeneous data to a low-dimensional space through linear transformation to obtain dimensionality-reduced multi-source heterogeneous data; and fusing the dimensionality-reduced multi-source heterogeneous data in a unified feature space to obtain fused features.

[0008] Optionally, a prediction result for the user is obtained by using a deep learning network model based on fused features, including: receiving fused features through the input layer of the deep learning network model; processing the fused features through multiple residual modules of the deep learning network model to obtain the output of the residual modules; passing the output of the residual modules to a global pooling layer for global pooling to obtain globally pooled features; passing the globally pooled features to multiple fully connected layers for processing, and determining the output of the last fully connected layer as the target feature; passing the target feature to the output layer for processing according to the task type to obtain the user's prediction result, wherein each residual module contains multiple convolutional layers, and the task type is used to indicate at least the output type of the prediction result.

[0009] Optionally, the initial recommendation strategy is adjusted based on feedback data to obtain the target recommendation strategy, including: constructing a Markov decision process, determining the state of the Markov decision process based on fusion features, determining the action of the Markov decision process based on the initial recommendation strategy, and determining the reward of the Markov decision process based on feedback data; and updating the initial recommendation strategy based on the state, action, and reward in the Markov decision process using a Q-learning algorithm to obtain the target recommendation strategy.

[0010] Optionally, the deep learning network model is trained in the following manner: The initial deep learning network model is trained using historical fusion features: using the historical fusion features as input, forward propagation is performed to obtain the output predicted by the initial deep learning network model; backpropagation is performed to calculate the gradient of the network parameters of the initial deep learning network based at least on the output, and the network parameters of the initial deep learning network are updated based on the gradient of the network parameters; the steps of forward propagation, backpropagation, and updating network parameters are repeated until the initial deep learning network model converges, thus obtaining the deep learning network model.

[0011] Optionally, the gradient of the network parameters of the initial deep learning network is calculated at least based on the output results through backpropagation, including: starting from the output layer of the initial deep learning network model, calculating the gradient of the loss function corresponding to the output layer based on the output results; using the chain rule, backpropagating the gradient of the loss function corresponding to the output layer along the initial deep learning network model, calculating layer by layer until the gradient of the loss function corresponding to the input layer is obtained, and determining the gradient of the loss function corresponding to each layer as the gradient of the network parameters of the initial deep learning network.

[0012] Optionally, the method further includes: continuously receiving real-time data streams from users and extracting features from the real-time data streams to obtain processed data, wherein the real-time data streams include new multi-source heterogeneous data generated after a first timestamp, the first timestamp being the timestamp corresponding to the acquisition of the user's multi-source heterogeneous data; and incrementally training the deep learning model based on the processed data.

[0013] According to another aspect of the embodiments of this application, a recommendation strategy determination apparatus based on multi-source heterogeneous data is also provided, comprising: an acquisition module for acquiring multi-source heterogeneous data of a user, wherein the multi-source heterogeneous data includes data from multiple sources and of various types associated with the user; a fusion module for performing feature fusion on the multi-source heterogeneous data to obtain fused features; a prediction module for making predictions based on the fused features using a deep learning network model to obtain prediction results for the user, and generating an initial recommendation strategy based on the prediction results, wherein the recommendation strategy is at least used to indicate the user's potential needs or behavioral tendencies, and the initial recommendation strategy is at least used to indicate products or services recommended to the user; and an adjustment module for receiving feedback data and adjusting the initial recommendation strategy based on the feedback data to obtain a target recommendation strategy, wherein the feedback data is used to indicate the user's response to the content recommended in the initial recommendation strategy.

[0014] According to another aspect of the embodiments of this application, a non-volatile storage medium is also provided, wherein a program is stored in the non-volatile storage medium, and the program controls the device where the non-volatile storage medium is located to execute the above-mentioned recommendation strategy determination method based on multi-source heterogeneous data when it runs.

[0015] According to another aspect of the embodiments of this application, an electronic device is also provided, including: a memory and a processor, the processor being configured to run a program stored in the memory, wherein the program executes the above-described method for determining recommendation strategies based on multi-source heterogeneous data during runtime.

[0016] According to another aspect of the embodiments of this application, a computer program product is also provided, including computer instructions, which, when executed by a processor, implement the above-described method for determining recommendation strategies based on multi-source heterogeneous data.

[0017] In this embodiment, the method involves acquiring multi-source heterogeneous data of the user, including data from various sources and types associated with the user; fusing features of the multi-source heterogeneous data to obtain fused features; using a deep learning network model to make predictions based on the fused features to obtain the user's prediction result; and generating an initial recommendation strategy based on the prediction result. The recommendation strategy at least indicates the user's potential needs or behavioral tendencies, and the initial recommendation strategy at least indicates the products or services recommended to the user. Feedback data is received, and the initial recommendation strategy is adjusted based on the feedback data to obtain a target recommendation strategy. The feedback data indicates how the user responds to the content recommended in the initial recommendation strategy. By acquiring multi-source heterogeneous data and fusing features to obtain fused features, then using a deep learning network model to make predictions based on the fused features to obtain the user's prediction result, generating an initial recommendation strategy based on the prediction result, and then adjusting the initial recommendation strategy based on the feedback data to obtain the target recommendation strategy, the method achieves the goal of dynamically adjusting the recommendation rules and improving the accuracy of the recommendation strategy. This solves the technical problem of low accuracy in generated recommendation strategies due to reliance on human judgment and static rules. Attached Figure Description

[0018] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments of this application and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings:

[0019] Figure 1 This is a hardware structure block diagram of a computer terminal for implementing a recommendation strategy determination method for multi-source heterogeneous data, according to an embodiment of this application.

[0020] Figure 2 This is a flowchart of a method for determining a recommendation strategy for multi-source heterogeneous data according to an embodiment of this application;

[0021] Figure 3 This is a schematic diagram of a device for determining a recommendation strategy for multi-source heterogeneous data according to an embodiment of this application. Detailed Implementation

[0022] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present application, and not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative effort should fall within the scope of protection of the present application.

[0023] The information collected in this application embodiment is information and data authorized by the user or fully authorized by all parties. The collection, storage, use, processing, transmission, provision, disclosure and application of the relevant data all comply with the relevant laws, regulations and standards of the relevant regions, and necessary confidentiality measures have been taken. It does not violate public order and good morals, and provides corresponding operation entry points for users to choose to authorize or reject the automated decision results. If the user chooses to reject, the process will proceed to the expert decision-making process.

[0024] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0025] To better understand the embodiments of this application, the technical terms involved in the embodiments of this application are explained below:

[0026] Reinforcement Learning (RL) is a machine learning method that learns optimal decision-making strategies through interaction with the environment.

[0027] Incremental learning: a machine learning technique that allows a model to be updated with new data without being completely retrained.

[0028] Principal Component Analysis (PCA): A data dimensionality reduction technique used to transform high-dimensional data into low-dimensional data while retaining its main feature information.

[0029] In related technologies, customer opportunity identification when determining recommendation strategies mainly relies on the experience and intuition of account managers, or on static and simple rule configurations, lacking flexibility and intelligence, and making it difficult to meet personalized and dynamic recommendation needs. This method, which relies on human judgment, is not only inefficient but also prone to bias, resulting in low accuracy of recommendation strategies, poor recommendation effects, and low user conversion rates, ultimately hindering the effective expansion of enterprise business and the increase of market share. Therefore, there is a technical problem of low accuracy in the generated recommendation strategies due to reliance on human judgment and static rules. To solve this problem, this application provides relevant solutions, which are described in detail below.

[0030] According to an embodiment of this application, an embodiment of a recommendation strategy determination method based on multi-source heterogeneous data is provided. It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions. Furthermore, although a logical order is shown in the flowchart, in some cases, the steps shown or described may be executed in a different order than that shown here.

[0031] The methods and embodiments provided in this application can be executed on a computer terminal or similar computing device. Figure 1 A hardware block diagram of a computer terminal for implementing a recommendation strategy determination method based on multi-source heterogeneous data is shown. Figure 1 As shown, the computer terminal 10 may include one or more processors 102 (shown as 102a, 102b, ..., 102n in the figure) 102 (processor 102 may include, but is not limited to, a microprocessor MCU or a programmable logic device FPGA, etc.), a memory 104 for storing data, and a transmission device 106 for communication functions. In addition, it may also include: a display, an input / output interface (I / O interface), a universal serial bus (USB) port (which may be included as one of the ports of a BUS bus), a network interface, a power supply, and / or a camera. Those skilled in the art will understand that... Figure 1 The structure shown is for illustrative purposes only and does not limit the structure of the aforementioned electronic device. For example, computer terminal 10 may also include... Figure 1 The more or fewer components shown, or having the same Figure 1 The different configurations shown.

[0032] It should be noted that the aforementioned one or more processors 102 and / or other data processing circuits are generally referred to herein as "data processing circuits". These data processing circuits may be embodied, in whole or in part, in software, hardware, firmware, or any other combination thereof. Furthermore, the data processing circuits may be a single, independent processing module, or may be integrated, in whole or in part, into any other element within the computer terminal 10. As involved in the embodiments of this application, the data processing circuits serve as a form of processor control (e.g., selection of a variable resistor termination path connected to an interface).

[0033] The memory 104 can be used to store software programs and modules of application software, such as the program instructions / data storage device corresponding to the recommendation strategy determination method based on multi-source heterogeneous data in this embodiment of the application. The processor 102 executes various functional applications and data processing by running the software programs and modules stored in the memory 104, thereby realizing the above-mentioned recommendation strategy determination method based on multi-source heterogeneous data. The memory 104 may include high-speed random access memory, and may also include non-volatile memory, such as one or more magnetic storage devices, flash memory, or other non-volatile solid-state memory. In some instances, the memory 104 may further include memory remotely located relative to the processor 102, and these remote memories can be connected to the computer terminal 10 via a network. Examples of such networks include, but are not limited to, the Internet, corporate intranets, local area networks, mobile communication networks, and combinations thereof.

[0034] The transmission device 106 is used to receive or send data via a network. Specific examples of the network described above may include a wireless network provided by the communication provider of the computer terminal 10. In one example, the transmission device 106 includes a Network Interface Controller (NIC), which can connect to other network devices via a base station to communicate with the Internet. In another example, the transmission device 106 may be a Radio Frequency (RF) module, used for wireless communication with the Internet.

[0035] The display may be, for example, a touchscreen liquid crystal display (LCD) that allows the user to interact with the user interface of the computer terminal 10.

[0036] In the above operating environment, this application provides an embodiment of a recommendation strategy determination method based on multi-source heterogeneous data. It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions. Also, although a logical order is shown in the flowchart, in some cases, the steps shown or described can be executed in a different order than that shown here.

[0037] like Figure 2 The diagram shown is a flowchart of a method for determining a recommendation strategy based on multi-source heterogeneous data according to an embodiment of this application, including:

[0038] Step S202: Obtain the user's multi-source heterogeneous data.

[0039] In the technical solution provided in step S202, multi-source heterogeneous data includes data from multiple sources and of various types associated with the user.

[0040] In some embodiments of this application, a recommendation system based on multi-source heterogeneous data fusion can be used to execute the method of steps S202 to S208 of this application. The recommendation system based on multi-source heterogeneous data fusion includes a data acquisition module for executing step S202; a data preprocessing module for preprocessing the multi-source heterogeneous data to obtain preprocessed multi-source heterogeneous data; a multi-source data fusion module for executing step S204; a deep learning analysis module for executing step S206; and a recommendation engine module for executing step S208. By organizing the massive amounts of user-related data of various types accumulated in the current big data pool, and then from a business perspective, the data used to construct this system (i.e., the aforementioned multi-source heterogeneous data) is determined.

[0041] For example, in the network operator business sector, multi-source heterogeneous data can include four parts: user consumption behavior data, network infrastructure data, social attribute data, and social media data. The aforementioned consumption behavior data can be the consumption behavior data of target users (e.g., government and enterprise users, also known as government and enterprise customers), used to describe the communication behavior and usage patterns of target users (e.g., government and enterprise users). Typical data extracted from the consumption behavior data that conforms to user characteristics and aggregated into the big data pool are shown in the following table:

[0042]

[0043]

[0044] Network infrastructure data is based on base stations. It involves querying logs from various devices and network monitoring systems within the base station's coverage area, such as base stations, routers, and switches. This includes device traffic data, and data synchronized with a Geographic Information System (GIS) system, such as grid data, coordinates, building surfaces, and regional data. Typical data extracted from network infrastructure is shown in the table below.

[0045]

[0046] Social attribute data is primarily collected through publicly available external information or commercial data sources, covering external social attributes such as company size, industry, and financial status. The data includes the company's number of employees, number of branches, years of establishment, profit, revenue, investment amount, industry classification and ranking, and bidding project information. Typical data for extracting social attributes is shown in the table below:

[0047]

[0048]

[0049] Social media data is collected by analyzing a company's interactions across various social media platforms, including company posts, user comments and feedback, and social media activity. Data collection methods include API access, purchasing from third-party data service providers, and web scraping. Typical examples of social media data extraction are shown in the table below:

[0050] category content Social interaction The number of posts, user comments, likes, and shares published by the company. User Feedback User evaluations and feedback on a company's products and services. Activity The time and frequency of a company's activity on social media platforms.

[0051] After acquiring the aforementioned multi-source heterogeneous data, it needs to be stored. For example, the following storage methods can be used: For structured data in the multi-source heterogeneous data: business data is stored in a relational database (e.g., PostgreSQL; PostgreSQL is a widely used open-source relational database management system), including user behavior data, base station traffic, etc.; For unstructured data in the multi-source heterogeneous data: large amounts of logs and unstructured data (such as network device logs, application logs, social media data, etc.) are processed using a big data platform (e.g., Spark; Spark is an open-source cluster computing framework for large-scale data processing). In actual storage, distributed storage services provided by the operator's cloud service platform are used for storage and backup. When the multi-source heterogeneous data includes four parts: telecommunications user consumption behavior data, network infrastructure data, social attribute data, and social media data, the data table design for each part is as follows: Consumption behavior data table: stores the behavior records of each user, such as traffic, call duration, terminal device type, etc., with the enterprise user ID as the primary key. Network infrastructure data table: stores information such as traffic and device access volume for each base station, associated with the base station ID and geographical information. The Social Attributes data table stores company information, including financial data and industry classifications, with the company ID as the primary key. The Social Media data table stores the company's interactions on social media, including post content and user feedback, with the company ID and timestamp as the primary key.

[0052] After obtaining the above multi-source heterogeneous data, step S204 is executed to perform feature fusion on the multi-source heterogeneous data to obtain fused features.

[0053] In the technical solution provided in step S204, there are several ways to perform feature fusion on multi-source heterogeneous data to obtain fused features: preprocessing the multi-source heterogeneous data to obtain preprocessed multi-source heterogeneous data; mapping high-dimensional data in the preprocessed multi-source heterogeneous data to a low-dimensional space through linear transformation to obtain dimensionality-reduced multi-source heterogeneous data; and fusing the dimensionality-reduced multi-source heterogeneous data in a unified feature space to obtain fused features. The following is one specific implementation method:

[0054] First, the multi-source heterogeneous data undergoes preprocessing, namely data cleaning and standardization, ultimately yielding preprocessed multi-source heterogeneous data. Data cleaning includes missing value handling, duplicate value handling, and outlier handling. Missing value handling includes record deletion: records with extremely low missing data ratios are directly deleted. Missing value imputation: important fields are imputed according to business rules. Important fields are pre-defined fields in the multi-source heterogeneous data, and business rules are pre-defined imputation rules for each important field. For example, if the field "average call duration" for a user is missing, it is imputed with the average value of the user's group. Duplicate value handling includes deduplication: using fields such as user ID and device ID to check for duplicate records; merging duplicates: for multiple records with the same ID, they are merged based on the latest record to avoid errors caused by data duplication. Outlier handling includes extreme value detection: detecting extreme values ​​that clearly do not conform to business logic, such as abnormally high user data usage, which is considered erroneous data and removed. The handling strategy is to handle outliers through deletion, correction, etc., and to filter out abnormal records using rules during modeling. Data standardization, through techniques such as normalization, standardization, and categorical data encoding, transforms data of different magnitudes from multi-source heterogeneous data to the same scale, avoiding misleading models due to different feature value ranges. Data normalization scales data of different magnitudes (e.g., consumption data, traffic data) from multi-source heterogeneous data to the range [0,1]. Standardization transforms some data from multi-source heterogeneous data (e.g., base station traffic fluctuations, user behavior patterns) into a standard normal distribution with a mean of 0 and a standard deviation of 1. Category data encoding converts some data from multi-source heterogeneous data (e.g., variables such as terminal device type, package type, and terminal device brand) into binary vectors using one-hot encoding.

[0055] Next, a linear transformation is used to map the high-dimensional data in the preprocessed multi-source heterogeneous data to a low-dimensional space, resulting in dimensionality-reduced multi-source heterogeneous data. For example, this can be achieved by using Principal Component Analysis (PCA) to project the high-dimensional data from the multi-source heterogeneous data to a low-dimensional space using a linear transformation, preserving the main variation information of the data. The formula used is:

[0056]

[0057] Where X is the original data matrix (i.e., the data matrix corresponding to the preprocessed multi-source heterogeneous data), X c Principal components are components along a single dimension. For example, in call data related to consumer behavior, μ is the mean vector, C is the covariance matrix (describing the correlation between features in the original data, with the shape of "feature number × feature number"), v is the eigenvector, λ is the eigenvalue, PC is the principal component, V is the matrix composed of eigenvectors, and n is the number of samples.

[0058] For example, when multi-source heterogeneous data includes four parts: telecommunications user consumption behavior data, network infrastructure data, social attribute data, and social media data, PCA dimensionality reduction yields four principal components (P...). 消费 ,P 基础 ,P 社会 ,P 社交 This refers to the multi-source heterogeneous data after dimensionality reduction, where q represents q consumer instance objects.

[0059] P 消费 =[PC 消费1 PC 消费2 ,...,PC 消费q ]

[0060] P 基础 =[PC 基础1 PC 基础2 ,...,PC 基础q ]

[0061] P 社会 =[PC 社会1 PC 社会2 ,...,PC 社会q ]

[0062] P 社交 =[PC 社交1 PC 社交2 ,...,PC 社交q ],

[0063] Then, data fusion is performed, fusing the dimensionality-reduced multi-source heterogeneous data in a unified feature space to obtain fused features. For example, for the four principal components mentioned above (P... 消费 ,P 基础 ,P 社会 ,P 社交 The fused feature vector (i.e., the fused feature P mentioned above) 融合 )for:

[0064] P 融合 =[P 消费 ,P 基础 ,P社会 ,P 社交 ],

[0065] Step S206: The deep learning network model makes predictions based on the fusion features to obtain the prediction results for the user, and generates an initial recommendation strategy based on the prediction results.

[0066] In the technical solution provided in step S206, the initial recommendation strategy is used at least to indicate the products or services recommended to the user, and may also indicate the user's potential needs or behavioral tendencies.

[0067] In the technical solution provided in step S206, there are multiple ways to obtain the user's prediction result by making predictions based on fused features using a deep learning network model. For example: receiving fused features through the input layer of the deep learning network model; processing the fused features through multiple residual modules of the deep learning network model to obtain the output of the residual modules; passing the output of the residual modules to a global pooling layer for global pooling to obtain globally pooled features; passing the globally pooled features to multiple fully connected layers for processing, and determining the output of the last fully connected layer as the target feature; passing the target feature to the output layer for processing according to the task type to obtain the user's prediction result. Each residual module contains multiple convolutional layers, and the task type is used to indicate at least the output type of the prediction result.

[0068] The deep learning network model described above is trained as follows: The initial deep learning network model is trained using historical fusion features. Using these features as input, forward propagation yields the predicted output of the initial deep learning network model. Backpropagation is then used to calculate the gradients of the network parameters of the initial deep learning network, based at least on the output results, and the network parameters are updated accordingly. This process of forward propagation, backpropagation, and updating network parameters is repeated until the initial deep learning network model converges, resulting in the final deep learning network model. Calculating the gradients of the network parameters of the initial deep learning network, based at least on the output results, through backpropagation can be achieved as follows: Starting from the output layer of the initial deep learning network model, the gradient of the loss function corresponding to the output layer is calculated. Using a chain rule, the gradient of the loss function corresponding to the output layer is backpropagated along the initial deep learning network model, layer by layer, until the gradient of the loss function corresponding to the input layer is obtained. The gradients of the loss functions corresponding to each layer are then determined as the gradients of the network parameters of the initial deep learning network.

[0069] Deep neural networks (DNNs) are used to automatically extract complex nonlinear features, thereby improving the accuracy of user profiling and the performance of recommendation systems. A multi-layered, nested, concatenated N (N is an integer greater than or equal to 1) residual convolutional module is constructed to extend the network depth, while also supporting on-demand adjustment of the parameters of each module. For the basic residual convolutional module, each module uses a 1D convolution as its core, combined with batch normalization, activation functions, and residual connections to ensure consistent input and output dimensions, allowing for direct concatenation to extend the network depth. Assume the input to the residual convolutional module is a static feature vector x∈R. 1×d (1×d is the input dimension, determined by the PCA dimensionality reduction result, such as 8 dimensions). For example, the specific steps for constructing the modules of the initial deep learning network model are as follows:

[0070] The initial deep learning network model architecture includes an input layer that receives the fused feature vectors (PCA principal components, i.e., the fused features mentioned above), an initial convolutional layer (1D convolution), N residual modules (N can be expanded as needed to increase network depth), a global pooling layer (1D global average pooling, aimed at compressing spatial dimensions and preserving core features), and an output layer (depending on the task, it can be a classification layer (for user segmentation) or a regression layer (for behavior prediction). The residual modules are constructed as follows:

[0071] (1) The model input is: X = P 融合 Assume the module input is a static feature vector x∈R 1×d , where d is the input dimension, which is determined by the PCA dimensionality reduction result.

[0072] (2) Convolutional layer 1 (1D):

[0073]

[0074] Where x is the module input vector, with dimension 1×d, and is the principal component of PCA dimensionality reduction (i.e., the aforementioned fusion feature). i:i+k-1 W1 is a slice representation of the input vector in a 1D convolution operation, used to extract continuous local elements from a static feature vector. i b1 is the weight of the i-th convolutional kernel, with a dimension of k×C, where C is the number of convolutional kernels, which is configurable. b1 is the bias term, with a dimension of 1×C. k is the kernel size, for example, it can be fixed at 2 to capture the correlation between two adjacent principal components, such as enterprise size and base station coverage. z1 is the convolutional output, with a dimension of 1×(d-k+1)×C. Since the stride s=1, the output length remains d-k+1=d-1, which is approximately the same as the input length.

[0075] (3) Batch Normalization (BN):

[0076]

[0077] Where B represents the batch size (the number of samples during training). μ B , The mean and variance within the batch are given. ε is used to prevent small values ​​with a denominator of 0. γ and β are learnable parameters (dimension 1×C), representing the scaling factor and offset, respectively. z′1 is the batch normalized output, with the same dimension as z1, ensuring a stable numerical range. 1,i This represents the convolution output of the i-th sample. This represents the intermediate variable for batch normalization.

[0078] (4) The ReLU activation function is mainly used to introduce nonlinearity and increase feature representation. The output result after the activation function is a1:

[0079] a1 = ReLU(z′1).

[0080] (5) Convolutional layer 2 (1D convolution):

[0081]

[0082] Among them, a 1,i:i+k-1 It is a slice representation of the input vector, used to extract continuous local elements from the static feature vector. W2 i b1 represents the kernel weights, with a dimension of k×C, where k is the kernel size, the same as the kernel size of convolutional layer 1. b2 is the bias term, with a dimension of 1×C. z2 is the convolution output of convolutional layer 2, with a dimension of 1×d×C. Since the input length is d-1, it is restored to d after convolution, which is the same as the length of the module input vector x of the input module.

[0083] (6) Residual connection, which performs element-wise addition of data to alleviate gradient vanishing in deep networks and supports multi-level nesting.

[0084] y = z² + P(x),

[0085] Where y represents the feature vector output after convolution and residual connection, and P(x) is the input projection. If the number of input and output channels is the same, there are no additional parameters. The dimensions of input x and output z2 are perfectly matched, that is:

[0086] y = z² + x

[0087] If the number of channels in the input and output are different, then 1D convolution is needed to adjust the number of channels in the input x to match the dimension of z2, that is:

[0088] P(x)=Conv1D(x,W P ,b P ),

[0089] Among them, W P For 1D projective convolution kernel, b P This is the bias term for projective convolution. For example, if the input x is 1×8×32 (32 channels) and the output z2 is 1×8×64 (64 channels), then W... P The x value is 1×32×64. x is converted to 1×8×64 through convolution, and then added to z2 to obtain y. Conv1D refers to the convolution operation.

[0090] The loss function for the initial deep learning network model during training is set as follows:

[0091] The classification task uses the cross-entropy loss function:

[0092]

[0093] Where, parameter L is the loss value, representing the difference between the model's (deep learning network model's) prediction (i.e., the output of the initial deep learning network model's prediction) and the true result. c is the class index, representing the class number of the currently calculated sample. C is the total number of classes, representing the total number of all classes that need to be classified. c Let y be the true class label (0 or 1) of the sample (i.e., the historical multi-source heterogeneous data corresponding to the historical fusion features). If the sample belongs to class c, then y c =1, if the sample does not belong to class c, then y c =0. The model predicts the probability, which represents the probability that the model predicts the sample belongs to class c, and the value ranges from 0 to 1.

[0094] The regression task uses the mean squared error (MSE) loss function:

[0095]

[0096] Where L MSE y is the mean squared error, which measures the average deviation between the model's (deep learning network model's) predictions of continuous values ​​and the actual values. n represents the total number of samples involved in training (i.e., the total number of historical multi-source heterogeneous data corresponding to historical fusion features). i is the sample index, representing the i-th sample (e.g., the i-th enterprise user), ranging from 1 to n. i Let be the true value of the i-th sample, and represent the actual observed value of that sample. Let be the predicted value of the i-th sample, representing the model's prediction result for that sample (i.e., the output result of the initial deep learning network model).

[0097] The training process of the initial deep learning model constructed in the above steps is as follows: Forward propagation: Calculate the network output. Backward propagation: Calculate the gradient of the loss function with respect to the parameters of each layer. Parameter update: Update the network weights using an optimization algorithm (such as Adaptive Moment Estimation (Adam). The Adam algorithm can adaptively adjust the learning rate and, combined with first-order and second-order moment estimation, effectively update the network weights to minimize the loss function). Specifically: Forward propagation calculates the network output: Using historical fusion features as input, the output predicted by the initial deep learning network model is obtained through forward propagation; Backward propagation calculates the gradient of the loss function with respect to the parameters of each layer: Starting from the output layer of the initial deep learning network model, the gradient of the loss function corresponding to the output layer is calculated based on the output result (the loss function is determined according to different task types, for example, the cross-entropy loss function and the mean squared error (MSE) loss function mentioned above); The gradient of the loss function corresponding to the output layer is backpropagated along the initial deep learning network model using the chain rule, calculated layer by layer until the gradient of the loss function corresponding to the input layer is obtained, and the gradient of the loss function corresponding to each layer is determined as the gradient of the network parameters of the initial deep learning network. Update network weights using an optimization algorithm (such as Adam): Update the network parameters (e.g., network weights) of the initial deep learning network based on the gradient of the network parameters. Repeat the forward propagation, backpropagation, and network parameter update steps until the initial deep learning network model converges (the value of the loss function no longer decreases significantly over multiple iterations, or the preset number of training epochs is reached), thus obtaining the deep learning network model.

[0098] For example, historical multidimensional asynchronous data of enterprise users can be randomly obtained from a database for training. First, the historical multidimensional asynchronous data is preprocessed to obtain preprocessed historical multidimensional asynchronous data. Then, a linear transformation is used to map the high-dimensional data in the preprocessed historical multidimensional asynchronous data to a low-dimensional space, resulting in dimensionality-reduced historical multi-source heterogeneous data. The dimensionality-reduced historical multi-source heterogeneous data is then fused in a unified feature space to obtain historical fused features. Finally, the historical fused features are used to train the initial deep learning network model.

[0099] The following is a specific implementation example. Historical multidimensional asynchronous data includes: Consumer behavior data (X, where X refers to the original high-dimensional feature matrix of this type of data): call duration 600 minutes / month, data usage 100GB / month, consumption 2000 / month. Infrastructure data (Y): base station coverage of 100 grids, peak data traffic of 80GB, device access type is high-end equipment. Social attribute data (Z): company size 1000+ employees, industry is fintech, profit 100 million. Social media data (W): the company has high activity on social media, and user feedback is positive. First, the historical multidimensional asynchronous data is preprocessed (data cleaning and data standardization): For the above four types of data, the standardization formula is as follows, taking consumer behavior data as an example:

[0100]

[0101] Where μ is the mean and σ is the standard deviation. T′ 消费 These represent the preprocessed data corresponding to the above data. The preprocessed historical multidimensional asynchronous data at this point includes T′. 消费 、T′ 基础 、T′ 社会 、T′ 社交 Then, Principal Component Analysis (PCA) is used to map the high-dimensional data in the preprocessed historical multidimensional asynchronous data to a low-dimensional space through linear transformation, resulting in dimensionality-reduced historical multi-source heterogeneous data.

[0102] Let X be the original high-dimensional feature matrix of a certain type of data (such as consumer behavior data). Then the matrix is ​​the number of samples n × the number of features m, where m features use attributes from the corresponding behavioral features, such as call duration, data usage, and spending amount. The following steps are performed on the original high-dimensional feature matrix (original data matrix) X of this type of data:

[0103] 1) Data centralization: Calculate the mean vector μ of X to obtain the centered data (the principal components are the components in a certain dimension) X. c :

[0104] X c =X-μ,

[0105] 2) Calculate the covariance matrix C to describe X. c The correlation between features, where n is the number of samples:

[0106]

[0107] 3) Eigenvalue decomposition: For the covariance matrix, if there exists a non-zero eigenvector v and an eigenvalue λ, satisfying:

[0108] C·ν=λ·ν,

[0109] Then, QR decomposition (a matrix factorization technique) is used to decompose the covariance matrix to obtain all eigenvalues ​​λ. 1～m (m is the total number of eigenvalues) and eigenvector v, find the direction with the most information in the data.

[0110] 4) Extract principal components, sort them by eigenvalue λ from largest to smallest, select the k largest corresponding eigenvectors (k is the number of extracted principal components, the low-dimensional feature dimension after dimensionality reduction of the high-dimensional original data), form a matrix V(m×k), and retain the k comprehensive feature directions with the most information. The principal components can then be represented as:

[0111] PC = X c ·V,

[0112] Among them, PC is the principal component, an n×k matrix that contains the core features of a certain type of behavioral data.

[0113] After obtaining the core features of each type of data in the historical multidimensional asynchronous data through the above steps, they are combined into historical fusion features.

[0114] Then, the initial deep learning model is trained: Model input X = P 融合 P here 融合 This represents the historical fusion features. Then, forward propagation: based on the residual modules of the initial deep learning model, the parameter settings are simplified. For example, the residual module parameters are: number of kernels C = 4, kernel size k = 2, kernel weights W1 and W2 are randomly initialized, and biases b1 and b2 are 0. The initial convolution result x0 is:

[0115] x0=Conv 1D(P 融合 ,W0,b0),

[0116] Where W0 is 1×6×4 (converting 6-dimensional features into 4 channels), and b0 is the initial bias value, using randomly initialized data following a normal distribution. The residual module calculation results are obtained through the convolutional layer 1, batch normalization, activation function, convolutional layer 2, and residual connections constructed above. The process is listed below:

[0117] Convolutional layer 1: z1 = Conv1D(x0, W1, b1), W1 is 2×4×4 (using two 4×4 convolutional kernels for feature extraction), output z1 ∈ R 1×5×4 .

[0118] Batch normalization: z′1 = BN(z1), using the BN procedure of the residual module to process the data.

[0119] Activation function: a1 = ReLU(z′1), non-negative results are preserved.

[0120] Convolutional layer 2: z2 = Conv1D(a1, W2, b2), output z2 ∈ R 1×6×4 Keep it consistent with x0.

[0121] Residual connection: y0 = z2 + x0 (the initial convolution result x0 and the output z2 have the same dimensions).

[0122] Global pooling is used to process the data after the residual module has processed it. This data then passes through a fully connected layer, and finally, softmax is used to transform the data. The pooling layer formula is:

[0123]

[0124] Among them, H c This represents the output value of the c-th channel after pooling. This represents the i-th value of the c-th channel in the residual module output. L is the feature length of each channel.

[0125] After local pooling, the output is obtained through a fully connected layer:

[0126] z out =H′1·W out +b out ,

[0127] Among them, z out W represents the original score for user group c, and is the value of the k-th feature after H′1 pooling, where k is the index of the feature vector after pooling, corresponding to the output of each channel after global pooling. out The weight matrix of the fully connected layer is a mapping weight from the number of pooled features to the number of user categories, with dimensions of [number of categories × number of pooled features], and follows a normal distribution with a mean of 0 and a small standard deviation, randomly initialized. out It is the base offset value for adjusting the output results, with a dimension of [number of categories × 1], and is initialized to 0.

[0128] For the loss calculation in the above process, for example, if the cross-entropy loss function is used for classification tasks, and then a softmax transformation is performed to obtain the final output value:

[0129] When performing backpropagation and parameter updates, the following steps are executed:

[0130] Calculate the gradient:

[0131]

[0132] Among them, z out The raw score for a certain user group. The error between the predicted probability and the true label of the model (deep learning network model). L is the cross-entropy loss function, and y is the true class label of the sample (i.e., historical multi-source heterogeneous data corresponding to historical fused features). This represents the predicted probability of the sample.

[0133] Update parameters (Adam optimizer):

[0134]

[0135] Where θ is the parameter to be updated, for example, W out Or, the convolution weights in the residual module. θ 旧 These are the parameter values ​​before the update, and η is the learning rate, which is dynamically updated based on the training and prediction results, initially set to 0.001. The gradient of the loss with respect to the parameter θ is the result of backpropagation calculation.

[0136] In practical applications, a trained deep learning network model is used to make predictions based on fused features. The fused features are received through the input layer of the deep learning network model; multiple residual modules of the deep learning network model process the fused features to obtain the output of the residual modules; the output of the residual modules is passed to a global pooling layer for global pooling to obtain globally pooled features; the globally pooled features are then passed to multiple fully connected layers for processing, and the output of the last fully connected layer is determined as the target feature. The target feature is then passed to the output layer for processing according to the task type to obtain the user's prediction result. Based on the prediction result, an initial recommendation strategy is obtained. The initial recommendation strategy can take various forms: for example, selecting packages, services, or products with the highest predicted probabilities (e.g., the top 10) as recommendations; or generating multiple possible recommendation strategies based on user characteristics and prediction results, such as recommending packages with high data allowance and network optimization services to users with high network dependence; or generating customized recommendation strategies based on the user's social attributes and behavioral characteristics. Finally, the generated initial recommendation strategy is output through a recommendation engine, such as a network operator's CRM system.

[0137] Furthermore, to enable deep learning models to dynamically update and promptly reflect user behavior and market changes, real-time data acquisition, processing (e.g., using Apache Flink for real-time data processing, including cleaning, standardization, and feature extraction) and analysis are achieved. This can be accomplished through methods such as: continuously receiving real-time data streams from users (e.g., using Apache Kafka as a message queue to receive real-time data streams), extracting features from the real-time data streams to obtain processed data. The real-time data streams include new multi-source heterogeneous data generated after a first timestamp, where the first timestamp corresponds to the timestamp corresponding to the acquisition of the user's multi-source heterogeneous data; and incrementally training the deep learning model based on the processed data (i.e., inputting the processed new multi-source heterogeneous data into the deep learning model in real-time for incremental training or fine-tuning to maintain the model's up-to-date state). Incremental training, in the field of machine learning and deep learning, refers to a training method where, when new data becomes available, the model does not need to be trained from scratch, but can be updated and fine-tuned based solely on the new data, building upon existing training. Incremental learning formula and derivation:

[0138] Set the parameters of the deep learning model at the current time t to θ. t The new multi-source heterogeneous data is for X t+1 Then, the parameters are updated using gradient descent to obtain the updated deep learning model parameters:

[0139] θ t+1 =θ t -η▽ θ L(X t+1 ,θ t ),

[0140] Where η is the learning rate, ▽ θ L is the gradient of the loss function with respect to the parameters.

[0141] Step S208: Receive feedback data and adjust the initial recommendation strategy based on the feedback data to obtain the target recommendation strategy. Finally, based on the generated recommendation result (i.e., the target recommendation strategy), push personalized package suggestions to the user.

[0142] In the technical solution provided in step S208, the feedback data is used to indicate the user's response to the content recommended in the initial recommendation strategy. There are several ways to adjust the initial recommendation strategy based on the feedback data to obtain the target recommendation strategy. For example: constructing a Markov decision process, determining the state of the Markov decision process based on fusion features, determining the action of the Markov decision process based on the initial recommendation strategy, and determining the reward of the Markov decision process based on the feedback data; or updating the initial recommendation strategy in the Markov decision process based on the state, action, and reward using a Q-learning algorithm to obtain the target recommendation strategy.

[0143] In some embodiments of this application, a novel adaptive recommendation algorithm is constructed. Through deep learning and real-time data stream processing technology, it achieves intelligent fusion and dynamic analysis of multi-source heterogeneous data, thereby providing highly personalized and accurate recommendation strategies. This adaptive recommendation algorithm can automatically optimize and adjust recommendation strategies based on real-time user feedback and behavioral data, thus providing more accurate and personalized recommendations and improving user conversion rates and satisfaction. First, steps S202 to S204 are executed to fuse multi-source heterogeneous data. By fusing this data, a comprehensive user profile is constructed, capturing the user's multi-dimensional characteristics and behavioral patterns. Then, step S206 is executed. Deep learning is a machine learning method that automatically extracts data features through multi-layer neural networks. Deep learning network models can be used for feature extraction, user behavior prediction, and the construction of recommendation systems. Afterwards, real-time data stream processing is performed. Real-time data stream processing refers to the immediate processing and analysis of continuously generated data to ensure that the system can quickly respond to user behavior and market changes, achieving incremental updates of the deep learning model and dynamically updating the model and recommendation strategy. Finally, the adaptive recommendation algorithm automatically optimizes and adjusts the recommendation strategy based on the user's real-time feedback and behavioral data (i.e., the feedback data mentioned above) (corresponding to step S208), providing more accurate and personalized recommendations and improving user conversion rates and satisfaction. The steps of automatically optimizing and adjusting the recommendation strategy based on the user's real-time feedback and behavioral data (i.e., the feedback data mentioned above) using the adaptive recommendation algorithm are as follows:

[0144] An initial recommendation strategy is obtained using a deep learning model (e.g., the initial recommendation strategy is a prediction of packages or services that the user might be interested in). Feedback data is then received (the user's response to the recommendations in the initial strategy, such as clicks, purchases, or ignores). Based on the feedback data, the recommendation model is optimized through online learning or reinforcement learning to adjust the initial recommendation strategy and obtain the target recommendation strategy. The following is the reinforcement learning formula and derivation: The recommendation problem is modeled as a Markov Decision Process (MDP), defining states, actions, and rewards. For example, the state of the Markov Decision Process is determined based on fused features, and... The initial recommendation strategy determines the actions in the Markov decision process, and the feedback data is used as the reward for the Markov decision process. Specifically: State (S): The user's current behavior and characteristics (i.e., the fused features mentioned above). Action (A): The content recommended by the initial recommendation strategy (e.g., a package or service, product, etc.). Reward (R): The user's response to the recommendation (i.e., determined based on the feedback data; for example, purchasing in the feedback data is a positive reward, ignoring it is a negative reward). Then, the initial recommendation strategy is updated in the Markov decision process based on the state, action, and reward using the Q-learning algorithm (i.e., continuously updating the Q-value) to obtain the target recommendation strategy.

[0145]

[0146] Where Q(S,A) is the long-term value estimate of action A in state S. α is the learning rate, representing the degree of influence of new feedback on the Q value; γ is the discount factor, representing the importance attached to the future; and R is the real-time reward, representing the user's direct feedback to the current recommendation. Let S' be the future value of the optimal action in the next state. S' is the next state of S, i.e., the state after S is updated. Through continuous iteration, the optimization goal of the recommendation strategy is to ensure that the recommended package (action A) receives a higher user response (reward R) in the long run under the user's current state (S), that is, the Q value will gradually approach the expected reward under the optimal strategy, thereby optimizing the recommendation strategy.

[0147] For a state S and action A in a Markov Decision Process (MDP), the Q-value (Q(SA)) is defined as the expected cumulative reward that the agent can obtain in the future after taking action A, starting from state S. The Q-value is continuously updated using a Q-learning algorithm to ultimately determine the target action A (i.e., the target recommendation strategy). Selection is made based on the updated Q-value to ensure that the recommended content can best meet user needs and improve user conversion rates.

[0148] The loss function in the above process is to minimize the recommendation error and improve the recommendation accuracy.

[0149] Among them, y i These are real tags (encoding of users' genuine feedback to recommendations, such as purchasing or ignoring). is the predicted value, and n is the total number of recommendations.

[0150] In some embodiments of this application, testing and verification can also be performed, for example, through the following method: 1. Cross-validation: Method: K-fold cross-validation, the data is divided into K parts, K-1 parts are used for training in sequence, and the remaining parts are used for verification. The calculation formula is as follows:

[0151]

[0152] Evaluation metrics: Silhouette Score: Measures the tightness and separation of clusters. Log Likelihood: Evaluates how well the model fits the data. Accuracy: The correctness of the prediction results.

[0153] 2. Split Test: Experimental Design: Experimental Group: Using the recommendation strategy of the system in this application embodiment. Control Group: Using the traditional recommendation strategy. Results: The recommendation success rate of the experimental group was improved by approximately 40%, demonstrating the significant advantage of the system.

[0154] The system implemented in this application can accurately segment different types of users, with each user group exhibiting unique characteristics in terms of behavior, network, and social attributes. For different user groups, such as a network operator's company, corresponding pricing standards and recommendation strategies can be formulated to meet the needs of enterprise users, and customer satisfaction and corporate revenue can be improved through value-added services and customized packages. Using actual production data, the system generates characteristics of different user groups and provides standards for these groups, as follows:

[0155]

[0156]

[0157] After actual classification, the customer profile data is as follows:

[0158]

[0159]

[0160] Based on the network operator's existing customer classification criteria and customer data, one or more recommended plans are suggested to enterprise customers seeking to switch. The plan recommendation list is sorted from highest to lowest rating, displaying the plan that best meets the customer's needs. See the table below:

[0161]

[0162]

[0163] We utilize multiple methods, including SMS, email, and app push notifications, to deliver package recommendations to customers. The push notifications are concise and clear, offering a quick "one-click order" mechanism. Recommendation campaigns are designed to reach customers through multiple channels, and strategies are adjusted in real-time based on customer feedback. If a customer is interested in a package but hasn't purchased it, further discounts or package upgrades will be offered. The effectiveness of recommendation campaigns is tracked and evaluated, analyzing key metrics such as customer conversion rates and purchase rates. Based on the evaluation results, we optimize the recommendation algorithm and strategies to improve customer satisfaction.

[0164] This application also provides a schematic diagram of a device for determining recommendation strategies based on multi-source heterogeneous data, such as... Figure 3 As shown, it includes:

[0165] The acquisition module 302 is used to acquire multi-source heterogeneous data of users, wherein the multi-source heterogeneous data includes data from multiple sources and types associated with users.

[0166] The fusion module 304 is used to perform feature fusion on multi-source heterogeneous data to obtain fused features.

[0167] The prediction module 306 is used to make predictions based on fused features through a deep learning network model to obtain the prediction results for the user, and to generate an initial recommendation strategy based on the prediction results. The initial recommendation strategy is used to indicate at least the products or services recommended to the user.

[0168] The adjustment module 308 is used to receive feedback data and adjust the initial recommendation strategy based on the feedback data to obtain the target recommendation strategy. The feedback data is used to indicate the user's response to the content recommended in the initial recommendation strategy.

[0169] It should be noted that, Figure 3 The recommended strategy determination device based on multi-source heterogeneous data shown is used to perform... Figure 2 The method for determining recommendation strategies based on multi-source heterogeneous data shown herein, therefore Figure 2 The relevant explanations in the method for determining recommendation strategies based on multi-source heterogeneous data also apply to the device for determining recommendation strategies based on multi-source heterogeneous data, and will not be repeated here.

[0170] It should be noted that each module in the above-mentioned recommendation strategy determination device based on multi-source heterogeneous data can be a program module (e.g., a set of program instructions to implement a certain function) or a hardware module. For the latter, it can be manifested in the following forms, but is not limited to them: each of the above modules is manifested as a processor, or the functions of each of the above modules are implemented by a processor.

[0171] This application also provides a non-volatile storage medium, which includes a stored program, wherein, during program execution, the device where the non-volatile storage medium is located executes the recommendation strategy determination method based on multi-source heterogeneous data from any of the above embodiments.

[0172] This application also provides an electronic device, which includes a processor for running a program, wherein the recommendation strategy determination method based on multi-source heterogeneous data of any of the above embodiments is executed during program execution.

[0173] According to another aspect of the embodiments of this application, a computer program product is also provided, including a computer program that, when executed by a processor, implements the recommendation strategy determination method based on multi-source heterogeneous data of any of the above embodiments.

[0174] In the above embodiments of this application, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0175] In the several embodiments provided in this application, it should be understood that the disclosed technical content can be implemented in other ways. The device embodiments described above are merely illustrative; for example, the division of units can be a logical functional division, and in actual implementation, there may be other division methods. For instance, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the displayed or discussed mutual coupling, direct coupling, or communication connection may be through some interfaces; the indirect coupling or communication connection between units or modules may be electrical or other forms.

[0176] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0177] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0178] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to related technologies, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, read-only memory (ROM), random access memory (RAM), portable hard drives, magnetic disks, or optical disks.

[0179] The above description is only a preferred embodiment of this application. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of this application, and these improvements and modifications should also be considered within the scope of protection of this application.

Claims

1. A method for determining a recommendation strategy based on multi-source heterogeneous data, characterized in that, The method comprises: acquiring multi-source heterogeneous data of a user, wherein the multi-source heterogeneous data comprises a plurality of sources and types of data associated with the user; performing feature fusion on the multi-source heterogeneous data to obtain fused features; performing prediction on the fused features by using a deep learning network model to obtain a prediction result of the user, and generating an initial recommendation strategy according to the prediction result, wherein the initial recommendation strategy is used at least to indicate a product or service recommended to the user; receiving feedback data, and adjusting the initial recommendation strategy according to the feedback data to obtain a target recommendation strategy, wherein the feedback data is used to indicate a response of the user to the content recommended in the initial recommendation strategy.

2. The method of claim 1, wherein, The feature fusion on the multi-source heterogeneous data to obtain the fused features comprises: performing data preprocessing on the multi-source heterogeneous data to obtain preprocessed multi-source heterogeneous data; mapping high-dimensional data in the preprocessed multi-source heterogeneous data to a low-dimensional space by using linear transformation to obtain reduced multi-source heterogeneous data; fusing the reduced multi-source heterogeneous data in a unified feature space to obtain the fused features.

3. The method of claim 1, wherein, The prediction on the fused features by using the deep learning network model to obtain the prediction result of the user comprises: receiving the fused features by using an input layer of the deep learning network model; processing the fused features by using a plurality of residual modules of the deep learning network model to obtain outputs of the residual modules; performing global pooling operation on the outputs of the residual modules by using a global pooling layer to obtain features after global pooling; processing the features after global pooling by using a plurality of fully connected layers, determining an output of a last fully connected layer as a target feature, and processing the target feature by using an output layer according to a task type to obtain the prediction result of the user, wherein each residual module comprises a plurality of convolutional layers, and the task type is used at least to indicate an output type of the prediction result.

4. The method of claim 1, wherein, The adjustment of the initial recommendation strategy according to the feedback data to obtain the target recommendation strategy comprises: constructing a Markov decision process, determining a state of the Markov decision process according to the fused features, determining an action of the Markov decision process according to the initial recommendation strategy, and determining a reward of the Markov decision process according to the feedback data; updating the initial recommendation strategy according to the state, the action, and the reward in the Markov decision process by using a Q-learning algorithm to obtain the target recommendation strategy.

5. The method of claim 1, wherein, The deep learning network model is trained in the following manner: training an initial deep learning network model by using historical fused features: taking the historical fused features as input, and obtaining an output result predicted by the initial deep learning network model by using forward propagation; calculating gradients of network parameters of the initial deep learning network according to at least the output result by using back propagation, and updating the network parameters of the initial deep learning network according to the gradients of the network parameters. The steps of forward propagation, back propagation and updating network parameters are repeatedly performed until the initial deep learning network model converges, and the deep learning network model is obtained.

6. The method of claim 5, wherein, The calculating the gradients of the network parameters of the initial deep learning network according to the output result by back propagation comprises: Starting from an output layer of the initial deep learning network model, the gradient of a loss function corresponding to the output layer is calculated according to the output result; The gradient of the loss function corresponding to the output layer is back propagated along the initial deep learning network model by the chain rule, and the gradient of the loss function corresponding to the input layer is calculated layer by layer until the gradient of the loss function corresponding to the input layer is calculated, and the gradient of the loss function corresponding to each layer is determined as the gradient of the network parameter of the initial deep learning network.

7. The method of claim 1, wherein, The method further comprises: Continuously receiving a real-time data stream of the user, and performing feature extraction on the real-time data stream to obtain processed data, wherein the real-time data stream comprises new multi-source heterogeneous data generated after a first timestamp, and the first timestamp is a timestamp corresponding to the acquisition of the multi-source heterogeneous data of the user; Incrementally training the deep learning model according to the processed data. 8.A device for determining a recommendation strategy based on multi-source heterogeneous data, characterized in that, Comprise: An acquisition module configured to acquire multi-source heterogeneous data of a user, wherein the multi-source heterogeneous data comprises data of multiple sources and types associated with the user; A fusion module configured to perform feature fusion on the multi-source heterogeneous data to obtain fused features; A prediction module configured to perform prediction according to the fused features by a deep learning network model to obtain a prediction result of the user, and generate an initial recommendation strategy according to the prediction result, wherein the initial recommendation strategy is used at least to indicate a product or service recommended to the user; An adjustment module configured to receive feedback data, and adjust the initial recommendation strategy according to the feedback data to obtain a target recommendation strategy, wherein the feedback data is used to indicate a response of the user to the content recommended in the initial recommendation strategy.

9. A non-volatile storage medium, comprising: The non-volatile storage medium stores a program, wherein the program controls the device in which the non-volatile storage medium is located to execute the recommendation strategy determination method based on multi-source heterogeneous data according to any one of claims 1 to 7 when the program is running.

10. An electronic device, comprising: Comprise: A memory and a processor, wherein the processor is configured to run a program stored in the memory, and the program performs the recommendation strategy determination method based on multi-source heterogeneous data according to any one of claims 1 to 7 when the program is running.

11. A computer program product comprising computer instructions, characterized in that, The computer instructions are executed by the processor to implement the recommendation strategy determination method based on multi-source heterogeneous data according to any one of claims 1 to 7.

Images