Adaptive optimization method and system for push strategy based on personalized needs

Abstract

The present invention discloses a method and system for adaptive optimization of push strategies for personalized needs, which relates to the field of data processing technology. First, the information of the push target is acquired and processed, and then the product information is extracted and identified through natural language processing technology and a two-layer attention parsing model. A personalized tag set is then established to extract information from historical push interaction records. The subjective needs and objective tag sets are fused through a collaborative attention fusion network to generate standard personalized demand information. The plans in the push strategy library are then mapped into vectors, and the multi-dimensional weighted cosine-divergence joint similarity function is used to retrieve candidate plans. Finally, the strategy is evolved through a dynamic multi-objective optimization space, and the improved NSGA-III algorithm is used to generate a Pareto front solution set, and the optimal solution is selected as the target push strategy. This method can improve the accuracy and efficiency of the push strategy and increase user satisfaction.

Description

Technical Field

[0001] The present invention relates to the field of data processing technology, and in particular to a method and system for adaptively optimizing push strategies oriented to personalized needs. Background Art

[0002] In the era of big data, various online platforms generate massive amounts of data daily, including user behavior data, product information, service details, and push channel information. This information, when deeply mined and utilized, can provide crucial support for business decision-making. Push strategies leverage this data to deliver the information users need through push notifications, thereby improving user experience and satisfaction and generating greater value for businesses.

[0003] Currently, most push strategies rely on unified push rules to push information, which doesn't effectively meet users' personalized needs. As user needs become increasingly diverse and personalized, traditional push methods are no longer sufficient. Therefore, we need to develop a push strategy that addresses these personalized needs.

[0004] However, understanding users' true personalized needs and satisfying them through push strategies has long been a challenge for the industry. Many recent studies have attempted to address this issue, such as extracting product information through natural language processing and classifying users through clustering algorithms. However, these approaches all have drawbacks, such as requiring significant human and material resources for data mining, inaccurate product classification, and low accuracy.

[0005] The information push method for the intelligent marketing service platform in CN118246983A, a prior art technology, considers users' multiple marketing interest levels and advertising preferences, improving the accuracy of advertising push and audience response rates. However, this method primarily relies on a large-scale group feature label and advertising preference feature library, assigning weights and integrating features based on interest levels. While achieving a certain degree of personalized push, it still cannot accurately capture and meet the real-time and subjective needs of specific users. Summary of the Invention

[0006] The technical problem to be solved by the present invention is how to ensure the accuracy of the push strategy while combining the real-time needs and subjective needs of users to perform adaptive optimization of the personalized push strategy to improve user satisfaction and the effectiveness of the push.

[0007] In order to achieve the above object, the present invention is implemented by adopting the following technical solution: the method comprises:

[0008] Obtain a list of push targets and call intuitive feature tag sets of different push targets, where each push target corresponds to an intuitive feature tag group in the intuitive feature tag set;

[0009] Obtain the attributes and related requirements of the target product, extract product information through natural language processing technology, and use a two-layer attention parsing model to identify and weight product components;

[0010] Establish an objective personalized tag set for the target product, extract historical push interaction records from the product behavior log database, and use it to constrain the similarity between the generated distribution and the real distribution, ultimately generating standard demand information with subjective and objective consistency;

[0011] Integrate subjective and objective personalized needs, and fuse the subjective demand component set with the objective label set through the collaborative attention fusion network to generate standard personalized demand information;

[0012] Push strategy library matching: maps the plans in the push strategy library into structured vectors, and uses a multi-dimensional weighted cosine-divergence joint similarity function to retrieve candidate plans;

[0013] The push plan space is adaptively optimized. The original push plan set is strategically evolved through dynamic multi-objective optimization space. The improved NSGA-III algorithm is used to generate the Pareto front solution set, and the optimal solution is selected as the target push strategy.

[0014] In one solution, the push target list is extracted from a business database through a multi-source heterogeneous data interface. The database contains user behavior data, product information, service details, and push channel information. The extracted push targets are deduplicated and cleaned to ensure the uniqueness and validity of the push targets.

[0015] In one embodiment, the feature labels in the intuitive feature label set are defined by triplets, which respectively represent the feature name, data type and initial weight coefficient, where the data type includes discrete and continuous features, discrete features are generated into binary vectors using one-hot encoding, and continuous features are processed into scalar values through minimum-maximum normalization.

[0016] In one embodiment, the feature fusion function is a weighted sum function, in which the feature weight coefficient is dynamically calculated by the entropy weight method, and the weighted sum function selects a vector or scalar output form according to the feature type to generate an intuitive feature label group for each push target.

[0017] In one solution, the user's subjective personalized needs are obtained, and the demand text is extracted through natural language processing technology. The demand components are identified and weighted using a two-layer attention parsing model to form a set of subjective personalized demand components to facilitate the optimization of subsequent push strategies.

[0018] In one scheme, the user's historical push interaction records are extracted, the historical behaviors are weighted by the time decay function, a user-target association matrix is constructed, and the data is reduced in dimension using sparse principal component analysis to generate the user's objective personalized label set.

[0019] In one solution, the end-to-end adaptive evolution of the push strategy performs strategy evolution on the original push plan set through a dynamic multi-objective optimization space, adopts the improved NSGA-III algorithm for multi-objective optimization to generate a Pareto front solution set, and selects the optimal solution as the target push strategy.

[0020] In another aspect, a system for adaptively optimizing a push strategy for personalized needs is provided, wherein the system is configured to execute the method for adaptively optimizing a push strategy for personalized needs, and the system comprises:

[0021] A push target acquisition module is used to obtain a list of push targets and call intuitive feature tag sets of different push targets, wherein each push target corresponds to an intuitive feature tag group in the intuitive feature tag set;

[0022] A user demand interaction module is used to interact with the target user, obtain the target user's subjective personalized needs, and analyze the subjective personalized needs to obtain a subjective personalized needs component set;

[0023] A historical data parsing module is used to extract the historical push data of the target user and parse the historical push data to establish an objective personalized tag set for the target user;

[0024] a personalized demand fusion module, configured to fuse the subjective personalized demand component set with the objective personalized label set to obtain standard personalized demand information, wherein the standard personalized demand information includes demand labels and corresponding demand weights;

[0025] A push strategy matching module is used to input the standard personalized demand information into the push strategy library;

[0026] The push plan optimization module is used to construct a push plan space based on the original push plan set to perform adaptive optimization and obtain a target push strategy.

[0027] Beneficial effects of the present invention:

[0028] By acquiring and processing information about push targets, combined with natural language processing technology and a dual-layer attention parsing model, product information can be extracted and understood more accurately, making push notifications more targeted.

[0029] By establishing a personalized tag set and extracting historical push interaction records, we can better understand users' historical behaviors and needs, and thus generate push strategies that better meet user needs.

[0030] By fusing subjective needs and objective label sets through a collaborative attention fusion network, we can accurately capture the user's real needs and improve the accuracy of push notifications.

[0031] By dynamically optimizing the original push plan set with multiple objectives and using the improved NSGA-III algorithm to generate the Pareto frontier solution set, the plan strategy can be dynamically adjusted to improve the push effect.

[0032] Through the above steps, we can achieve adaptive optimization of push strategies for personalized needs, improve the accuracy and efficiency of push strategies, enhance user experience, and help improve the company's economic benefits and user satisfaction.

[0033] Therefore, the technical solution of the present invention can not only realize the adaptive optimization of the push strategy and improve the effectiveness of the push, but also improve user satisfaction, promote user stickiness, and help improve business results. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Figure 1 Schematic diagram of the process of the adaptive optimization method of push strategy for personalized needs of the present invention;

[0035] Figure 2 This is a structural diagram of the push strategy adaptive optimization system for personalized needs of the present invention.

[0036] Explanation of the accompanying reference numerals: push target acquisition module 11, user demand interaction module 12, historical data analysis module 13, personalized demand fusion module 14, push strategy matching module 15, push plan optimization module 16. DETAILED DESCRIPTION

[0037] The above technical solution will be described in detail below in conjunction with the accompanying drawings and specific implementation methods of the specification to better understand the above technical solution. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments of the present invention. It should be understood that the present invention is not limited to the example embodiments used only to explain the present invention. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention. In addition, it should be noted that, for the convenience of description, only the parts related to the present invention, rather than all, are shown in the drawings. Example

[0038] Figure 1As shown in FIG, the push strategy adaptive optimization method for personalized needs of the present invention includes:

[0039] S1. Obtain a list of push targets, and call intuitive feature tag sets of different push targets, wherein each push target corresponds to an intuitive feature tag group in the intuitive feature tag set.

[0040] Specifically, the push target refers to the execution object or specific goal of the push activity, such as specific products, services or content, the push channels that can be used (such as online advertising, social media, offline activities, etc.), and the push methods that can be implemented (such as discount promotions, giveaways, etc.); the push target list is a form that includes all push targets in the target scenario, and each entry in the form contains basic information about the target.

[0041] In the implementation process of step S1, the system first extracts the set of all push targets in the target scenario from the business database through the multi-source heterogeneous data interface and builds a push target list. Each push target is represented as a tuple ,in is the unique identifier of the target, Represents a set of push channel parameters. A set of push method parameters. It is a set of target attributes (such as product category, price range, applicable population, etc.). In order to establish a computable target feature model, the system uses the feature engineering module to perform structured transformation on the original parameters and generate an intuitive feature label set with unified dimensions. , where each feature label By triple Definition, which represents the feature name, data type (discrete / continuous) and initial weight coefficient. For discrete features (such as channel type), one-hot encoding is used to generate a binary vector , where k represents the total number of categories of the feature, and each dimension corresponds to a Boolean value mapping of the existence of a specific category. Continuous features (such as discount strength) are normalized to scalar values through min-max normalization. , where X represents the value set of the feature in the historical data.

[0042] Intuitive feature label set for each push target Through the feature fusion function Generate, where It is a feature conversion operator that selects vector or scalar output form according to the feature type. is the feature weight coefficient, which is dynamically calculated by the entropy weight method. Specifically, the feature weight calculation process is defined as: for a data set containing N targets, the feature Information entropy ,in Represents the normalized eigenvalue distribution probability, the final weight , Represents the original eigenvalue of the i-th feature on the k-th push target before normalization. This mathematical processing gives higher weights to features with high discrimination (low entropy), effectively improving the representation ability of the feature set. The feature matrix of the push target finally generated is It will be used as the input space for subsequent strategy optimization. is a set of real numbers, where Indicates the total number of dimensions after all features are transformed. is a feature conversion function, where each row of the matrix corresponds to a fused feature vector for a target. This process is implemented through an automated feature pipeline, supporting an incremental update mechanism. When new target information is pushed, the system only needs to partially recalculate the weights and distribution parameters of the affected features, ensuring the real-time and consistency of the feature set.

[0043] S2. Obtain the attributes and related requirements of the target product, extract product information through natural language processing technology, and use a two-layer attention parsing model to identify and weight product components.

[0044] Specifically, target products refer to objects that require relevant push promotion activities, such as supplies in the fields of services, equipment, resources, etc.; the attributes and related needs of products refer to the specific characteristics and potential needs of the products themselves, which vary according to the categories, functions, quality, market positioning, etc. of different products. For example, they involve the functional goals, market strategies, applicable scenarios and expected effects of the products.

[0045] In the implementation process of step S2, the system collects the subjective demand text expressed in natural language by the target user through the multimodal interactive interface. , where each statement Contains demand semantic units. To extract structured demand components, the system uses a two-layer attention parsing model to semantically deconstruct the original text: First, the BiLSTM-CRF model constrained by the domain dictionary is used to perform basic entity recognition and define the entity extraction function ,in is the entity term,

[0046] is the entity type label; then the demand component association matrix is constructed based on the self-attention mechanism , whose elements Representing an entity and The semantic association strength of is a trainable parameter matrix, and d is the word vector dimension. The demand component weights are generated through the gated fusion layer. in is the sigmoid function, is the gate parameter matrix, which ultimately forms the subjective personalized demand component set In order to eliminate the ambiguity of user expressions, the system introduces a demand confidence calibration factor ,in represents the KL divergence between the weight distribution of current demand component j and the historical demand distribution, is the attenuation coefficient, and the calibrated demand weight is updated to Finally, the standard demand component weights are obtained through normalization Construct a complete set of subjective personalized demand components This process continuously optimizes the entity recognition and weight calculation modules through fine-tuning of the pre-trained language model and online learning mechanisms, ensuring adaptability to emerging user demand expression patterns.

[0047] S3. Establish an objective personalized tag set for the target product, extract historical push interaction records from the product behavior log database, and use it to constrain the similarity between the generated distribution and the real distribution, and finally generate standard demand information with subjective and objective consistency.

[0048] Specifically, the target user's description of their needs may deviate from their actual needs, that is, the user's ability to express subjective personalized needs may be limited or fluctuating. Therefore, we cannot rely solely on the user's subjective description to understand the needs. Therefore, information is extracted from the target user's historical behavior data, that is, it can be considered that the performance and feedback of the push plan in the historical records are a faithful reflection of personalized needs. By analyzing the user's historical push data, we can gain an in-depth understanding of user behavior and needs, and reduce the deviation caused by the subjective description of subjective personalized needs.

[0049] In the implementation process of step S3, the system extracts the target user’s behavior log database in the time window. Historical push interaction record set , where each record Contains the subject identifier , interaction timestamp , interaction type ( {click, conversion, favorite, share}) and interaction depth , click duration, purchase amount, etc. To construct an objective personalized label set, the system first uses the time decay function Weight the historical behavior, where is the decay coefficient, which characterizes the time decay rate of the behavior utility. Based on the characteristic matrix of the push target (From S1), establish the user-target association matrix , whose elements represents the user’s cumulative preference strength for feature dimension j, where is the set of labels containing feature j, is the interaction type weight function, which is preset by the expert system\ Quantify the value differences between different interaction behaviors.

[0050] In order to eliminate the influence of feature collinearity, the system uses sparse principal component analysis (SPCA) to Perform dimensionality reduction to solve optimization problems in is the sparse loading matrix, is a low-dimensional representation, is the sparsity control parameter. Extract the feature combination corresponding to the first d principal components , contribution rate is the lth singular value) is used to calculate the objective label weight in Represents the non-zero load feature index set corresponding to the lth principal component. The final generated objective personalized label set Converted into business-interpretable label descriptions through the semantic mapping module, such as the load factor exceeding the threshold Feature combination Mapped to semantic labels such as "high-frequency, high-value category preference". This process is dynamically updated through a sliding time window mechanism. When new interaction data arrives, the system incrementally recalculates the attenuation weights and principal component loads to ensure the timeliness of the label set and uses KL divergence to Monitor the degree of user preference drift and trigger the tag set reconstruction process when it exceeds the threshold.

[0051] in Represents the non-zero load feature index set corresponding to the lth principal component. The final generated objective personalized label set Converted into business-interpretable label descriptions through the semantic mapping module, such as the load factor exceeding the threshold Feature combination Mapped to semantic labels such as "high-frequency, high-value category preference". This process is dynamically updated through a sliding time window mechanism. When new interaction data arrives, the system incrementally recalculates the attenuation weights and principal component loads to ensure the timeliness of the label set and uses KL divergence to Monitor the degree of user preference drift and trigger the tag set reconstruction process when it exceeds the threshold.

[0052] In the implementation process of step S4, the system constructs a collaborative attention fusion network to integrate the subjective demand components. With objective label set Perform semantic alignment and weight fusion. First, establish a cross-modal correlation matrix , whose elements Indicates subjective demand component With objective labels The semantic similarity of and They are respectively a subjective semantic encoder based on BERT fine-tuning and an objective label encoder based on feature embedding. The collaborative weights are generated through a two-way attention mechanism: subjective side attention weight Quantify the support strength of objective labels for subjective needs and the objective side attention weight Characterizes the extent to which subjective demands modify objective labels.

[0053] Perform bidirectional projection fusion based on attention weights and define the fusion weight calculation function:

[0054] ;

[0055] in ; is the dynamic balance factor, through KL divergence and JS divergence Quantify the difference between subjective and objective distributions, and is a trainable parameter matrix. For labels with semantic conflicts , satisfy (Threshold ), start the conflict resolution mechanism ;

[0056] in is the penalty coefficient, For A set of low-correlation objective labels. Finally, the standard demand weight is generated through normalization. And merge subjective and objective tags to generate standard personalized demand information ,middle is a semantically merged label that satisfies and is the subset of subjective and objective labels after semantic clustering. The fusion process optimizes the encoder and attention parameters through the adversarial training strategy, and the discriminator network With the generator network The loss functions are:

[0057] ;

[0058] in is the maximum mean difference, which is used to constrain the similarity between the generated distribution and the real distribution, and finally generate standard demand information with subjective and objective consistency.

[0059] S5. Push strategy library matching: mapping the plans in the push strategy library into structured vectors, and using the multi-dimensional weighted cosine-divergence joint similarity function to retrieve candidate plans.

[0060] Specifically, each plan contains plan component information, such as target labels, label weights, etc., which are used to describe the applicable scenarios and adapted personalized needs of the plan.

[0061] In the implementation process of step S5, the system pushes M in the strategy library to the plan by constructing a multi-dimensional feature space. Mapping to structured vectors ,in is the label vector of the plan component (element Indicates the plan For the kth standard requirement label matching strength), are meta-feature vectors (representing the timeliness of the plan, the execution cost, and the historical conversion rate), Generate timestamp for the plan. Based on standard personalized demand information , define the multidimensional weighted cosine-divergence joint similarity function:

[0062] ;

[0063] in Normalize the matching strength for the Sigmoid activation function (parameters (a, b) are fitted by historical feedback data). is the label discrimination weight, is the meta-feature adaptive weight coefficient, is the feature importance parameter determined by the entropy weight method. For time-sensitive features, a time decay correction factor is introduced. , is the preset maximum effective period), and the original similarity is adjusted to:

[0064] ;

[0065] The system uses an improved ball tree index structure to accelerate neighbor search and define a distance metric function Construct multi-dimensional space partition. Retrieve the satisfying , is a dynamic threshold, determined by the Pareto quantile of the historical matching distribution) and based on the comprehensive score Sort in descending order, where is the time-sensitive coefficient. The final output is the original push plan set ,in Indicates pressing The sorted rth plan is verified by the orthogonal projection mechanism to ensure ( The process adopts an incremental update strategy. When a new plan is added to the strategy library, the index structure is dynamically optimized through the R* tree node splitting algorithm to maintain Retrieval efficiency of time complexity.

[0066] S6: Adaptive optimization of the push plan space, which performs strategic evolution on the original push plan set through dynamic multi-objective optimization space.

[0067] In the implementation process of step S6, the system constructs a dynamic multi-objective optimization space to optimize the original push plan set. Perform strategy evolution and define the three-dimensional optimization objective function:

[0068] ;

[0069] in is the time attenuation coefficient, Represents the user budget, time limit and device preference constraint vector. The improved NSGA-III algorithm is used for multi-objective optimization to generate the Pareto front solution set. , its dominant relationship is defined as: Dominant solution If and only if And there exists at least one j, .

[0070] According to the constraints, an adaptive penalty function is introduced: ;

[0071] in is the Sigmoid constraint relaxation function, is the tolerance threshold of each dimension constraint. The solution space exploration is realized through the quantum genetic crossover operator, and the chromosome crossover probability is defined as:

[0072] ;

[0073] in is the population diversity coefficient, and are the average fitness and optimal fitness of the \( t \)th generation population, respectively. Perform fuzzy membership decision on the generated Pareto solution set and calculate the comprehensive utility value of each solution:

[0074] ;

[0075] in represents the preference weight extracted from the user's implicit feedback data through the LSTM network, Controls the sharpness of the weight distribution for the temperature coefficient. Final selection The largest solution is used as the target push strategy , and encode it into lightweight execution rules through policy distillation technology: ;

[0076] in To select weights, is the indicator function. The optimization process is continuously updated through the online incremental learning mechanism. When new user feedback data arrives, according to the policy gradient Adjust the reference point distribution of NSGA-III to achieve end-to-end adaptive evolution of push strategies.

[0077] Example 2

[0078] Figure 2 This is a schematic diagram of the structure of the push strategy adaptive optimization system for personalized needs of the present invention. For example, Figure 1 The flowchart of the push strategy adaptive optimization method for personalized demand in the present invention can be shown as follows: Figure 2 The structure shown is implemented.

[0079] Based on the same concept as the personalized demand-oriented push strategy adaptive optimization method in the above embodiment, the present invention also provides a personalized demand-oriented push strategy adaptive optimization system including:

[0080] The push target acquisition module 11 is used to acquire a push target list and call intuitive feature tag sets of different push targets, wherein each push target corresponds to an intuitive feature tag group in the intuitive feature tag set.

[0081] The user demand interaction module 12 is used to interact with the target user, obtain the target user's subjective personalized demand, and analyze the subjective personalized demand to obtain a subjective personalized demand component set.

[0082] The historical data parsing module 13 is used to extract the historical push data of the target user and parse the historical push data to establish an objective personalized tag set for the target user.

[0083] The personalized demand fusion module 14 is configured to fuse the subjective personalized demand component set with the objective personalized label set to obtain standard personalized demand information, wherein the standard personalized demand information includes demand labels and corresponding demand weights.

[0084] The push strategy matching module 15 is used to input the standard personalized demand information into the push strategy library, and perform proximity matching with the standard personalized demand information based on the plan component information of the push plan in the push strategy library, obtain the first N closest push plans, and obtain the original push plan set, where N is a positive integer greater than or equal to 2.

[0085] The push plan optimization module 16 is used to construct a push plan space based on the original push plan set to perform adaptive optimization and obtain a target push strategy.

[0086] In some embodiments, the user demand interaction module 12 performs the following steps:

[0087] Interactively extract the target user's multi-dimensional demand data and perform data preprocessing, including data cleaning, text normalization, and format conversion, to generate multi-dimensional demand information.

[0088] The multidimensional demand information is parsed based on natural semantic recognition and matched with the intuitive feature label set to obtain the target user's subjective personalized demand label set, and a principal component analysis is performed on the subjective personalized demand label set to determine the principal component contribution corresponding to the subjective personalized demand label as the demand weight, and output the demand weight set.

[0089] The demand weight set is associated with the subjective personalized demand label set to generate the subjective personalized demand component set.

[0090] In some embodiments, the historical data parsing module 13 performs the following steps:

[0091] Based on the historical push data, associated historical push strategies and historical push feedback are obtained.

[0092] A historical policy tag set of the historical push policy is obtained by parsing, and a correlation between the historical policy tag set and the historical push feedback is analyzed to obtain a correlation coefficient set.

[0093] The correlation coefficient set is used as the label weight of the historical strategy label set, and the associated output is the objective personalized label set.

[0094] In some embodiments, the execution steps of the personalized demand fusion module 14 include:

[0095] Analyze and evaluate the data quality of the historical push data to obtain a data quality coefficient.

[0096] The deviation between the data quality coefficient and the standard data quality is calculated to obtain a quality deviation degree, wherein the quality deviation degree is a vector.

[0097] According to the quality deviation, the fusion weight ratio of the subjective personalized demand component set and the objective personalized label set is adjusted.

[0098] Obtain the union of the subjective personalized demand label set and the objective personalized label set as the demand label set of the standard personalized demand information, and calculate the weighted mean of the demand weight and the label weight of each pair of matches based on the fusion weight ratio, and associate and store them as standard personalized demand information.

[0099] In some embodiments, the push plan optimization module 16 includes a quick optimization unit, and the execution steps include:

[0100] Randomly mutate the N original push plans in the original push plan set to generate a new push plan set.

[0101] Re-evaluate the plan component information of the newly added push plan set and the original push plan set, and match them to obtain the optimal plan.

[0102] In some embodiments, the push plan optimization module 16 includes a precision optimization unit, and the execution steps include:

[0103] Based on the original push plan set, a push plan space is defined and constructed, and the N original push plans in the original push plan set are mapped to N sample points in the push plan space.

[0104] Based on the historical usage records of the N original push plans, optimization parameters of N sample points are initialized, wherein the optimization parameters include an optimization step size and an optimization direction.

[0105] It should be understood that the embodiments mentioned in this specification focus on their differences from other embodiments. The specific embodiments in the aforementioned embodiment one are also applicable to the push strategy adaptive optimization system for personalized needs described in embodiment two. For the sake of brevity of the specification, no further elaboration will be given here.

[0106] It should be understood that the embodiments disclosed in the present invention and the above description can enable those skilled in the art to use the present invention to implement the present invention. At the same time, the present invention is not limited to the embodiments mentioned above. It should be understood that those skilled in the art can still modify the technical solutions described in the above embodiments or replace some of the technical features therein with equivalents; and such modifications or replacements do not deviate from the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention and are all included in the scope of protection of the present invention.

Claims

1. A push strategy adaptive optimization method for personalized needs, characterized by: The method comprises: Obtain a list of push targets and call intuitive feature tag sets of different push targets, where each push target corresponds to an intuitive feature tag group in the intuitive feature tag set; Obtain the attributes and related requirements of the target product, extract product information through natural language processing technology, and use a two-layer attention parsing model to identify and weight product components; Establish an objective and personalized tag set for the target product, extract historical push interaction records from the product behavior log database, and use a time window to constrain the similarity between the generated distribution and the real distribution, ultimately generating standard demand information with subjective and objective consistency; Integrate subjective and objective personalized needs, and fuse the subjective demand component set with the objective label set through the collaborative attention fusion network to generate standard personalized demand information; Push strategy library matching: maps the plans in the push strategy library into structured vectors, and uses a multi-dimensional weighted cosine-divergence joint similarity function to retrieve candidate plans; Adaptive optimization of the push plan space: the original push plan set is strategically evolved through dynamic multi-objective optimization space, the improved NSGA-III algorithm is used to generate the Pareto front solution set, and the optimal solution is selected as the target push strategy; The end-to-end adaptive evolution of the push strategy is achieved by evolving the original push plan set through a dynamic multi-objective optimization space. The improved NSGA-III algorithm is used for multi-objective optimization to generate a Pareto front solution set and select the optimal solution as the target push strategy. The optimization process is continuously updated through an online incremental learning mechanism. When new user feedback data arrives, the reference point distribution of NSGA-III is adjusted according to the policy gradient to achieve end-to-end adaptive evolution of the push strategy. The feature fusion function is a weighted sum function, in which the feature weight coefficient is dynamically calculated by the entropy weight method. The weighted sum function selects vector or scalar output form according to the feature type to generate an intuitive feature label group for each push target; The system extracts the user's historical push interaction records, weights historical behaviors using a time decay function, constructs a user-target association matrix, and uses sparse principal component analysis to reduce the data dimension to generate the user's objective personalized tag set. This is dynamically updated through a sliding time window mechanism. When new interaction data arrives, the system incrementally recalculates the decay weights and principal component loadings to ensure the timeliness of the tag set. The system also monitors the degree of user preference drift using KL divergence, triggering the tag set reconstruction process when the threshold is exceeded. The feature labels in the intuitive feature label set are defined by triplets, which respectively represent the feature name, data type and initial weight coefficient, where the data type includes discrete and continuous features. Discrete features are generated into binary vectors using one-hot encoding, and continuous features are processed into scalar values through minimum-maximum normalization.

2. The push strategy adaptive optimization method for personalized needs according to claim 1, characterized in that: The push target list is extracted from the business database through a multi-source heterogeneous data interface. The database contains user behavior data, product information, service details and push channel information, and the extracted push targets are deduplicated and cleaned to ensure the uniqueness and validity of the push targets.

3. The push strategy adaptive optimization method for personalized needs according to claim 1, characterized in that: Obtain the user's subjective personalized needs, extract the demand text through natural language processing technology, use the double-layer attention parsing model to identify and weight the demand components, and form a set of subjective personalized demand components to facilitate the optimization of subsequent push strategies.

4. A push strategy adaptive optimization system for personalized needs, characterized by: The system is used to execute the push strategy adaptive optimization method for personalized needs according to any one of claims 1 to 3, and the system includes: A push target acquisition module is used to obtain a list of push targets and call intuitive feature tag sets of different push targets, wherein each push target corresponds to an intuitive feature tag group in the intuitive feature tag set; A user demand interaction module is used to interact with the target user, obtain the target user's subjective personalized needs, and analyze the subjective personalized needs to obtain a subjective personalized needs component set; A historical data parsing module is used to extract the historical push data of the target user and parse the historical push data to establish an objective personalized tag set for the target user; a personalized demand fusion module, configured to fuse the subjective personalized demand component set with the objective personalized label set to obtain standard personalized demand information, wherein the standard personalized demand information includes demand labels and corresponding demand weights; A push strategy matching module is used to input the standard personalized demand information into the push strategy library; The push plan optimization module is used to construct a push plan space based on the original push plan set to perform adaptive optimization and obtain a target push strategy.

Images