Private domain flow monitoring, analyzing and processing method and system

Abstract

The invention discloses a private domain flow monitoring, analyzing and processing method and system, particularly relates to the field of private domain flow management, is used for optimizing multi-marketing activity triggering time and solving the time conflict problem, and comprises the following steps of: constructing a multi-dimensional customer preference vector from customer interaction data through weighted time sliding window analysis and hierarchical clustering calculation; the acceptance of the customer to different activity types under each time slice is reflected; identifying and quantifying time period conflicts among multiple activities by utilizing a maximum matching optimization algorithm of a graph theory and a constraint satisfaction solver, and generating an activity priority weight list and a conflict mark of each customer; then, iteratively optimizing activity triggering time distribution based on the data structure by adopting a deep reinforcement learning algorithm, and realizing a personalized and conflict-free activity pushing strategy; therefore, the accuracy and conversion efficiency of marketing activities are remarkably improved, the customer experience is optimized, and the intelligence and competitiveness of private domain traffic management are enhanced.

Description

Technical Field

[0001] The present invention relates to the field of private domain traffic management, and more specifically, to a private domain traffic monitoring, analysis and processing method and system. Background Art

[0002] In the field of medical device sales, the use of private domain traffic has become the main means to promote customer conversion and enhance customer stickiness. Enterprises usually carry out various types of activities simultaneously in private domain platforms (such as WeChat official accounts, WeChat groups, and corporate WeChat), including product promotions, new product launches, customer care, and after-sales service activities. These activities often need to be pushed at precise time points to attract customer attention. However, when multiple activities are running in parallel, the triggering time points of different activities may conflict. For example, the same customer receives multiple different types of push messages in a short period of time. This conflict will not only cause information interference, making it difficult for customers to focus on specific activities, but may also cause customer dissatisfaction with private domain channels due to information overload, or even direct loss. This phenomenon is particularly evident in medical device sales, because the information categories and priorities that customers pay attention to are significantly affected by their specific needs (such as urgent need for consumables replenishment or equipment upgrades) and purchase cycles.

[0003] It is difficult for existing technologies to effectively identify and optimize the conflicts in time nodes between multiple activities, which is specifically manifested in the following aspects: First, the existing private domain traffic monitoring and analysis methods are mainly optimized for a single activity, and a global evaluation mechanism for push between multiple activities has not been established, resulting in different activities overlapping or too close in trigger time. Second, there are individual differences in customer response priorities for different activities. For example, some customers pay more attention to promotional information, while other customers need more functional upgrade suggestions. The existing system cannot dynamically adjust according to customer preferences. Third, the trigger logic of multiple activities is usually based on a fixed time period, and fails to fully consider the current behavioral activity status of customers and their possible decision-making paths, which makes the actual match between activity trigger time and customer needs low. The above problems not only affect the overall conversion effect of private domain activities, but also have an adverse impact on the long-term stickiness of private domain customers. It is urgent to solve them by building a set of monitoring, analysis and processing methods and systems for collaborative optimization of multiple activities.

[0004] In order to solve the above problems, a technical solution is now provided. Summary of the invention

[0005] In order to overcome the above-mentioned defects of the prior art, an embodiment of the present invention provides a private domain traffic monitoring, analysis and processing method and system, which constructs a multi-dimensional customer preference vector from customer interaction data through weighted time sliding window analysis and hierarchical clustering calculation to reflect the customer's acceptance of different activity types in each time segment; utilizes the maximum matching optimization algorithm and constraint satisfaction solver of graph theory to identify and quantify the time period conflicts between multiple activities, and generates an activity priority weight list and conflict markers for each customer; then, adopts a deep reinforcement learning algorithm to iteratively optimize the activity trigger time allocation based on the above data structure to realize a personalized and conflict-free activity push strategy; thereby significantly improving the accuracy and conversion efficiency of marketing activities, optimizing customer experience, and enhancing the intelligence and competitiveness of private domain traffic management to solve the problems raised in the above background technology.

[0006] To achieve the above object, the present invention provides the following technical solutions:

[0007] The private domain traffic monitoring, analysis and processing method comprises the following steps:

[0008] S1: Perform weighted time sliding window analysis and hierarchical clustering calculation on the customer interaction data obtained from private domain traffic, input the obtained time series feature parameters into the adaptive preference modeling module, and output a multi-dimensional customer preference vector containing the customer's acceptability scores for various activity information in different time segments and the corresponding time window data;

[0009] S2: The multi-dimensional customer preference vector and the preset parallel activity list are input into the maximum matching optimization algorithm and constraint satisfaction solver based on graph theory, and time period conflict detection and weighted calculation are performed on each parallel activity to generate a data structure containing a list of activity priority weights and a list of corresponding conflict markers for each customer;

[0010] S3: Input the step activity priority weight list, conflict marker list and multi-dimensional customer preference vector into the deep reinforcement learning algorithm for iterative optimization and solution, and output the trigger time allocation table for each activity in the corresponding customer time segment.

[0011] In a preferred embodiment, step S1 includes the following contents:

[0012] a1. First collect customer interaction data;

[0013] a2. Use the weighted time sliding window analysis method to capture the temporal dynamic characteristics of customer behavior; set a time sliding window size and assign different weights to data points in each sliding window;

[0014] a3. Use wavelet transform to perform multi-scale decomposition on the weighted time series data and extract the behavioral characteristics under different frequency components;

[0015] a4. Take the extracted multi-scale behavior feature vector as input and use the hierarchical clustering algorithm to group customers; hierarchical clustering constructs a similarity matrix between customers. The similarity Derived from the dynamic time warping distance metric;

[0016] a5. Input the clustering results into the adaptive preference modeling module and construct a multidimensional customer preference vector for each customer through multidimensional regression analysis; the multidimensional customer preference vector P k Contains customer k's acceptability rating P for different activity types a at different time segments t k，a，t .

[0017] In a preferred embodiment, step S2 includes the following contents:

[0018] b1. First receive the multi-dimensional customer preference vector and obtain the preset parallel activity list Each of these activities Contains its type, scheduled push time window and related attributes, is the index of the activity, which uniquely identifies each activity in the parallel activity list;

[0019] b2. Map each activity into a multidimensional customer preference vector space to form an activity preference vector in represents customer k pairs of activities The acceptability score at time segment t; then, the similarity matrix between activities is calculated using cosine similarity

[0020] In a preferred embodiment, step S2 further includes the following contents:

[0021] b3. Construct an active conflict graph G = (V, E) where the vertex set The edge set E is generated based on the overlapping relationship of the activity time periods; if the activity and activities There is overlap in any time segment t, that is, Then add an edge to the active conflict graph Edge weight By combining similarity and time overlap.

[0022] In a preferred embodiment, step S2 further includes the following contents:

[0023] b4. Apply the maximum weight matching optimization algorithm on the constructed activity conflict graph to identify the maximum conflict pairs between activities; determine the key conflict pairs between activities by maximizing the sum of matching edge weights, and prioritize resolving conflicts with higher weights;

[0024] Subsequently, the priority of activities is dynamically adjusted using a constraint satisfaction solver;

[0025] The constraint satisfaction problem solver solves the activity priority weight list for each customer by defining variables, domains, and constraints. and a list of conflict markers in Indicates activity The priority weight of For activities Is there a time conflict flag? represents the total number of parallel marketing activities that customer k conducts simultaneously in the private domain traffic management system. (in ) represents an independent marketing campaign;

[0026] b5. Finally, the generated data structure includes a list of activity priority weights W for each customer. k and conflict marker list C k .

[0027] In a preferred embodiment, step S3 includes the following contents:

[0028] c1. First, integrate the multi-dimensional customer preference vector, activity priority weight list and conflict marker list as the input data of the deep reinforcement learning algorithm; define the current state S t Contains the following: t = {P k , W k , C k};

[0029] c2. The action space is defined as the set of all activity triggering time segments For each customer k and its corresponding activity action Indicates that the activity will Assigned to time segment T v , Construct a policy network π in deep reinforcement learning, whose input is the current state S t , the output is for each activity In each time segment T v The probability of action selection under

[0030] In a preferred embodiment, step S3 further includes the following contents:

[0031] c3. In order to guide the deep reinforcement learning algorithm to optimize the activity trigger time, the reward function R is designed to combine the conversion effect and customer satisfaction of the activity. The specific formula is as follows: in: Indicates activity In time segment T v ; α1 is the weight parameter, which controls the contribution of conversion rate and satisfaction to the reward respectively; γ1 is the penalty parameter, which is used to reduce the negative impact of time period conflict; is the indicator function, when the activity Assigned to time segment T v The value is 1 when , otherwise it is 0.

[0032] In a preferred embodiment, step S3 further includes the following contents:

[0033] c4. Use deep Q network as deep reinforcement learning algorithm;

[0034] c5. After multiple rounds of iterative training, the deep reinforcement learning algorithm learns the optimal activity trigger time allocation strategy; based on the trained strategy network, a specific trigger time segment is generated for each activity of each customer, forming a personalized activity trigger time allocation table Specifically, for each customer k and activity Trigger time (in ) is determined by the following formula: in, Indicates that in state S t Next, the activity Assigned to time segment T v Q value.

[0035] In a preferred embodiment, step S3 further includes the following contents:

[0036] c4. Use deep Q network as the deep reinforcement learning algorithm; the specific training process is as follows:

[0037] c4-1. Experience replay: Record the state, action, reward and next state of each interaction and store these experiences in the experience replay buffer

[0038] c4-2.Q value update: randomly sample a batch of experience from the experience replay buffer Update the Q network parameters θ using the loss function;

[0039] c4-3. Policy update: through ε - The greedy strategy selects action a t, where ε is gradually reduced to balance exploration and exploitation; the policy network selects the action with the highest Q value according to the Q value output by the current Q network Q(S, a; θ);

[0040] c4-4. Iterative optimization: Repeat c4-1 to c4-3, and gradually optimize the strategy network so that it can assign the optimal activity triggering time point to each customer under the multi-objective optimization framework.

[0041] The private domain traffic monitoring, analysis and processing system includes: a customer clustering unit, a conflict detection unit and a scheduling optimization unit;

[0042] Customer clustering unit: Performs weighted time sliding window analysis and hierarchical clustering calculation on customer interaction data obtained from private domain traffic, inputs the obtained time series feature parameters into the adaptive preference modeling module, and outputs a multi-dimensional customer preference vector containing customer acceptability scores for various activity information in different time segments and corresponding time window data;

[0043] Conflict detection unit: inputs the multi-dimensional customer preference vector output by the customer clustering unit and the preset parallel activity list into the maximum matching optimization algorithm based on graph theory and the constraint satisfaction solver, performs time period conflict detection and weighted calculation on each parallel activity, and generates a data structure containing a list of activity priority weights and a list of corresponding conflict markers for each customer;

[0044] Scheduling optimization unit: The activity priority weight list and conflict mark list generated in the conflict detection unit and the multi-dimensional customer preference vector obtained in the customer clustering unit are input into the deep reinforcement learning algorithm for iterative optimization and solution, and the trigger time allocation table for each activity in the corresponding customer time segment is output.

[0045] The technical effects and advantages of the private domain traffic monitoring, analysis and processing method and system of the present invention are as follows:

[0046] The present invention significantly improves the trigger time management efficiency of multiple parallel marketing activities in private domain traffic by introducing the construction of multi-dimensional customer preference vectors, maximum matching optimization and constraint satisfaction solving technology based on graph theory, and iterative optimization of deep reinforcement learning algorithms. First, weighted time sliding window analysis and hierarchical clustering calculation are used to accurately capture customers' detailed preferences for various activities in different time segments, forming a high-dimensional, dynamically updated customer preference vector. This multi-dimensional representation not only reflects the customer's historical behavior patterns, but also adapts to changes in their current interests and needs in real time. Secondly, through the maximum matching optimization algorithm of graph theory combined with the constraint satisfaction solver, the time period conflicts between multiple activities are systematically identified and quantified, ensuring that high-priority activities are pushed first within the customer's optimal acceptance period, while reasonably allocating low-priority activities to avoid information overload and customer interference. Finally, under the multi-objective optimization framework, the deep reinforcement learning algorithm iteratively learns and optimizes the activity trigger time allocation strategy based on the complex data structure generated in the first two steps to achieve personalized and conflict-free activity push. Overall, it has significantly improved the accuracy and conversion efficiency of marketing activities, optimized customer experience, and ensured that the private domain traffic management system has greater adaptability and competitiveness in a complex and changing market environment. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 It is a flowchart of the private domain traffic monitoring, analysis and processing method of the present invention;

[0048] Figure 2 It is a structural schematic diagram of the private domain traffic monitoring, analysis and processing system of the present invention. DETAILED DESCRIPTION

[0049] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0050] Embodiment 1: Figure 1 The present invention provides a private domain traffic monitoring and analysis processing method, including:

[0051] S1. Perform weighted time sliding window analysis and hierarchical clustering calculation on the customer interaction data obtained from private domain traffic, input the obtained time series feature parameters into the adaptive preference modeling module, and output a multi-dimensional customer preference vector containing customer acceptability scores for various activity information in different time segments and corresponding time window data;

[0052] S2. Input the multi-dimensional customer preference vector outputted from step S1 and the preset parallel activity list into the maximum matching optimization algorithm and constraint satisfaction solver based on graph theory, perform time period conflict detection and weighted calculation on each parallel activity, and generate a data structure containing a list of activity priority weights and a list of corresponding conflict markers for each customer;

[0053] S3. Input the activity priority weight list and conflict marker list generated in step S2 and the multi-dimensional customer preference vector obtained in step S1 into the deep reinforcement learning algorithm for iterative optimization and solution, and output the trigger time allocation table for each activity in the corresponding customer time segment.

[0054] In the field of medical device sales, private traffic management systems improve customer stickiness and sales conversion rates by directly interacting with customers. However, with the parallel development of multiple marketing activities, how to accurately identify and understand customers' acceptance of various activities in different time periods has become the key to improving marketing efficiency. Existing technologies rely on simple statistical analysis and static rules in customer preference modeling, and it is difficult to dynamically capture real-time changes in customer behavior and multi-dimensional preferences. Therefore, developing a method that can deeply analyze customer interaction data and construct a detailed multi-dimensional customer preference vector is of great significance for optimizing private activity triggering strategies.

[0055] Step S1 includes the following contents:

[0056] This step aims to extract key behavioral features from customer interaction data obtained from private domain traffic through data analysis technology, and construct a multi-dimensional customer preference vector to reflect the customer's acceptability of different types of activities in different time segments.

[0057] a1. Data collection and preprocessing: First, collect customer interaction data through the private domain traffic monitoring module, including access timestamp, stay time, click behavior, activity response records and interaction frequency. The collected data needs to be cleaned to remove duplicate, missing and abnormal data points to ensure data quality. Subsequently, the timestamp data is standardized and converted into a unified time base (such as UTC time) for subsequent analysis.

[0058] a2. Weighted time sliding window analysis: In order to capture the temporal dynamic characteristics of customer behavior, a weighted time sliding window analysis method is used. A time sliding window size T is set (for example, one hour in 24 hours is a time segment), and different weights are assigned to data points in each sliding window. The weight function w(t) is designed as a function based on exponential decay to emphasize the importance of recent behavior: w(t) = e -λ(T-t); where λ is the decay rate parameter, which is used to control the decay speed of the weight over time, and t is the position of the time point in the sliding window (for example, the number of hours from the current time). In this way, recent interactive behaviors have a more significant impact on customer preferences.

[0059] a3. Behavioral feature extraction and time series modeling: Use wavelet transform to perform multi-scale decomposition on weighted time series data and extract behavioral features under different frequency components. Wavelet transform can effectively separate short-term and long-term behavioral patterns and capture customer behavior changes at different time scales. The specific steps are as follows: Among them, X(t) represents the original time series data, c i is the wavelet coefficient, ψ i is the wavelet basis function, and N is the number of decomposition layers. Through wavelet transform, multi-scale behavior feature vectors are obtained, covering customer behavior patterns at different time scales.

[0060] a4. Hierarchical clustering calculation: The extracted multi-scale behavior feature vector is used as input, and the hierarchical clustering algorithm is used to group customers. Hierarchical clustering constructs a similarity matrix between customers. The similarity It is derived from the dynamic time warping distance metric: Among them, X i and X j Represent the time series behavior data of customer i and customer j respectively.

[0061] a5. Adaptive Preference Modeling: Input the clustering results into the adaptive preference modeling module and construct a multidimensional customer preference vector P for each customer through multidimensional regression analysis. k . Multidimensional customer preference vector P k Contains customer k's acceptability rating P for different activity types a at different time segments t k，a，t , the calculation formula is as follows:

[0062]

[0063] Among them, β 0，a，t is a constant term, β m，a，t is the regression coefficient, X k，m，t is the mth behavior feature of customer k for activity type a in time segment t, X k，n，t represents the value of the nth behavior feature of customer k in time segment t, β m，n，a，t is the regression coefficient of the behavioral feature interaction term, and ∈ is the error term. Through multidimensional regression analysis, the preference vector not only considers the independent impact of each behavioral feature on preference, but also captures the synergistic effect between different features through interaction terms, thereby generating a detailed and dynamic multidimensional customer preference vector, providing basic data for subsequent activity conflict detection and priority sorting.

[0064] Step S1 constructs a detailed and dynamic multi-dimensional customer preference vector through weighted time sliding window analysis, wavelet transform, multi-scale behavioral feature extraction, hierarchical clustering, and multi-dimensional regression analysis. This process not only captures the customer's acceptance of various activities in different time periods, but also improves the accuracy and adaptability of the preference model through weight functions and multi-dimensional interaction terms. The constructed multi-dimensional customer preference vector provides a solid data foundation for activity conflict detection and priority sorting in subsequent steps, ensuring the personalization and optimization of private domain activity triggering strategies.

[0065] In step S1, a multi-dimensional customer preference vector for each customer is successfully constructed through weighted time sliding window analysis and hierarchical clustering calculations. The vector records in detail the customer's acceptability scores for different types of activities in each time segment. In order to further optimize the push strategy of private domain activities, it is necessary to identify and solve the time period conflicts between multiple parallel activities on this basis. When dealing with multi-activity time conflicts, the existing technology often ignores the complex interactive relationship between customers' individual preferences and activities, resulting in poor activity push effects. Therefore, step S2 introduces a maximum matching optimization algorithm based on graph theory and a constraint satisfaction solver. By accurately detecting time period conflicts and calculating activity priorities, a list of activity priority weights and a list of conflict markers are generated for each customer to ensure the rationality and efficiency of subsequent activity scheduling.

[0066] Step S2 includes the following contents:

[0067] Step S2 aims to identify the time period conflicts between multiple parallel activities based on the multi-dimensional customer preference vector generated in step S1, and prioritize the activities, and generate the activity priority weight list Wk and conflict marker list C for each customer through graph theory methods and constraint satisfaction solving technology. k .

[0068] b1. Input data preparation: First receive the multi-dimensional customer preference vector P output in step S1 k = {P k，a，t}, where P k，a，t represents the acceptability rating of customer k for activity type a in time segment t. At the same time, obtain the preset parallel activity list Each of these activities Contains its type, scheduled push time window and related attributes, Is the index of the activity, which uniquely identifies each activity in the parallel activity list.

[0069] b2. Activity mapping and similarity calculation: Map each activity to a multi-dimensional customer preference vector space to form an activity preference vector in represents customer k pairs of activities The acceptability score at time segment t. Next, calculate the similarity matrix between activities Measuring activity using the cosine similarity formula and Activities Similarity in customer k preference space: in, For activities and Activities The dot product of and Respectively for activities and Activities The Euclidean norm of the preference vector .

[0070] b3. Conflict graph construction and edge weight assignment: Construct an active conflict graph G = (V, E) where the vertex set The edge set E is generated based on the overlapping relationship of the activity time periods. and activities There is overlap in any time segment t, that is, Then add an edge to the active conflict graph Edge weight By combining similarity And the time overlap calculation yields: in, is the indicator function, when the activity and When overlapping in time segment t, otherwise is the total number of time segments. This weight formula emphasizes the severity of multiple time segment overlaps through square processing, and combines similarity to ensure that conflicts of highly similar activities are given higher weights, thereby improving the accuracy of conflict detection.

[0071] b4. Maximum matching optimization and constraint satisfaction solution: On the constructed activity conflict graph, the maximum weight matching optimization algorithm is applied to identify the maximum conflicting pairs between activities. The algorithm determines the key conflicting pairs between activities by maximizing the sum of matching edge weights and prioritizes solving conflicts with higher weights.

[0072] Then, the priority of activities is adjusted dynamically using the constraint satisfaction solver. The following constraints are defined:

[0073] Time non-overlap: High-priority activities cannot overlap with low-priority activities in the same time segment.

[0074] Example: Assume that the multi-dimensional customer preference vector of customer D shows that the acceptance score for device upgrade is higher than that for promotional offer. If the system plans to push the device upgrade activity from 10:00 to 11:00 a.m. and the promotional offer activity from 10:00 to 10:30 a.m. at the same time, according to the constraint that the time does not overlap, the system should adjust the push time of the promotional offer activity. For example, adjust the push time of the promotional offer activity to 11:00 to 11:30 a.m. to avoid overlapping with the time period of the device upgrade activity.

[0075] Priority dependency: Based on the multidimensional customer preference vector, high preference activities have higher priority than low preference activities.

[0076] Example: The multi-dimensional customer preference vector of customer E shows that after-sales service is the most acceptable, followed by new product launches, and promotional offers. During activity scheduling, the system gives after-sales service activities the highest priority and arranges them to be pushed in the time segment with high acceptance of customer E; new product launch activities are the second highest priority and are arranged to be pushed in the time segment with the second highest acceptance; and promotional offers are arranged to be pushed in the time segment with relatively low acceptance according to the remaining resources, ensuring that high-priority activities are pushed first.

[0077] Resource limitations: Push resources (such as number of messages, push frequency) must be within an acceptable range and not exceed the limits set by the system.

[0078] Example: The system sets a maximum of 5 marketing messages per customer per day, and a maximum of 1 message per hour. For customer F, if the system plans to push 7 activities (including new product launches, promotional offers, after-sales service, etc.), the system will select the 5 activities with the highest priority for push based on the activity priority weight list and conflict marker list, and ensure that the push time interval between these activities is at least one hour. For example, the five highest priority activities of after-sales service, new product launches, equipment upgrades, health lectures, and technical support are selected for push, while the two low-priority activities of promotional offers and limited-time discounts are abandoned or postponed to the next day.

[0079] CSPSolver solves the activity priority weight list for each customer by defining variables (activity trigger time), domains (optional time segments) and constraints. and a list of conflict markers in Indicates activity The priority weight of For activities A flag indicating whether there is a time conflict (1 indicates a conflict, 0 indicates no conflict), represents the total number of parallel marketing activities that customer k conducts simultaneously in the private domain traffic management system. (in ) represents an independent marketing activity, such as new product launch, promotional offer, customer care or after-sales service.

[0080] CSPSolver, i.e. constraint satisfaction problem solver, is an algorithm or software tool specifically used to solve constraint satisfaction problems. Constraint satisfaction problems involve finding a combination of all variable values ​​based on a set of variables, the range of values ​​of each variable, and a series of constraints, so that all constraints are satisfied. CSPSolver effectively narrows the search space and improves the efficiency of solution through systematic search strategies such as backtracking search, forward check, minimum residual value heuristic, and constraint propagation. In addition, CSPSolver can also use heuristic algorithms and optimization techniques, such as local search and genetic algorithms, to further enhance its ability to solve complex problems. In the present invention, CSPSolver is applied to time period conflict detection and priority sorting of multiple parallel activities. By defining variables (activity trigger time), domains (optional time segments) and constraints (such as time non-overlap, priority dependency, and resource constraints), the time allocation of activities is systematically optimized to ensure that various marketing activities can be pushed efficiently and conflict-free within the best time window preferred by customers, thereby improving the overall performance and marketing effect of the private domain traffic management system.

[0081] b5. Output data structure generation: Finally, the data structure generated in step S2 includes a list of activity priority weights for each customer W k and conflict marker list C k These data structures will serve as input to the deep reinforcement learning algorithm in step S3 to ensure the optimal scheduling of activity triggering times based on accurate conflict identification and priority sorting.

[0082] Step S2 combines the multi-dimensional customer preference vector generated in step S1 with the preset parallel activity list, and uses the maximum weight matching optimization algorithm and constraint satisfaction solving technology of graph theory to systematically identify and quantify the time period conflicts between multiple activities. The specific process includes activity mapping and similarity calculation, conflict graph construction and edge weight assignment, maximum matching optimization and constraint satisfaction solving, and finally generates a list of activity priority weights and a list of conflict markers for each customer. This process ensures the accurate identification and reasonable priority sorting of conflicts between multiple activities, provides a solid data foundation for subsequent personalized time point optimization scheduling, and improves the intelligence and optimization level of private domain activity push strategy.

[0083] In step S1, a multidimensional customer preference vector for each customer is successfully constructed through weighted time sliding window analysis and hierarchical clustering calculations. The vector records in detail the customer's acceptability ratings for different types of activities in each time segment. Then, in step S2, based on these multidimensional customer preference vectors, the system uses the maximum matching optimization algorithm and constraint satisfaction solver of graph theory to identify and quantify the time period conflicts between multiple parallel activities, and generates a list of activity priority weights and a list of conflict markers for each customer. In order to further optimize the trigger time allocation of private domain activities, step S3 introduces a deep reinforcement learning algorithm. Through iterative optimization, based on the output data of steps S1 and S2, a personalized and conflict-free activity trigger time allocation table is generated, thereby improving the intelligence and marketing efficiency of the private domain traffic management system.

[0084] Step S3 includes the following contents:

[0085] Step S3 aims to achieve personalized optimization scheduling of multi-activity triggering times through deep reinforcement learning algorithms based on multi-dimensional customer preference vectors, activity priority weight lists, and conflict marker lists, to ensure the efficiency and conflict-free nature of activity push.

[0086] c1. Input data integration and state definition: First, the multi-dimensional customer preference vector output from step S1, the activity priority weight list and the conflict marker list output from step S2 are integrated as the input data of the deep reinforcement learning algorithm. Define the current state S t Contains the following: t = {P k , W k , C k};

[0087] in:

[0088] P k is the multidimensional preference vector of customer k, which represents the acceptability score of different activity types in each time segment.

[0089] W k It is a list of activity priority weights, indicating the priority of each activity.

[0090] C k It is a list of conflict markers, indicating whether there is a time period conflict among the activities.

[0091] c2. Action definition and policy network construction: The action space is defined as the set of all possible activity triggering time segments For each customer k and its corresponding activity action Indicates that the activity will Assigned to time segment T v , Construct a policy network π in deep reinforcement learning, whose input is the current state S t , the output is for each activity In each time segment T v The probability of action selection under

[0092] c3. Reward function design and optimization goal: In order to guide the deep reinforcement learning algorithm to optimize the activity trigger time, the reward function R is designed to combine the conversion effect and customer satisfaction of the activity. The specific formula is as follows:

[0093]

[0094] in:

[0095] Indicates activity In time segment T v Expected conversion rate increase.

[0096] α1 is a weight parameter that controls the contribution of conversion rate and satisfaction to the reward.

[0097] γ1 is a penalty parameter used to reduce the negative impact of time period conflicts.

[0098] is the indicator function, when the activity Assigned to time segment T v The reward function takes the value of 1 when the trigger time is reached, otherwise it takes the value of 0. The reward function comprehensively considers the conversion rate improvement and customer satisfaction, encourages the deep reinforcement learning algorithm to choose an efficient trigger time, and optimizes the activity time allocation by punishing conflicting events to avoid customer information overload.

[0099] c4. Use deep Q network as the deep reinforcement learning algorithm. The specific training process is as follows:

[0100] c4-1. Experience replay: Record the state, action, reward and next state of each interaction and store these experiences in the experience replay buffer

[0101] c4-2.Q value update: randomly sample a batch of experience from the experience replay buffer Update the Q network parameters θ using the following loss function:

[0102]

[0103] Among them, γ2 is the discount factor, θ - are the target network parameters, and Q(S,a;θ) represents the value estimate of the Q network for the state-action pair (S,a).

[0104] c4-3. Policy update: through ε - The greedy strategy selects action a t , where ε is gradually reduced to balance exploration and exploitation. The policy network selects the action with the highest Q value according to the Q value output by the current Q network Q(S,a;θ).

[0105] c4-4. Iterative optimization: Repeat c4-1 to c4-3, and gradually optimize the strategy network so that it can assign the optimal activity trigger time point to each customer under the multi-objective optimization framework (such as maximizing conversion rate and minimizing customer interference).

[0106] c5. Output trigger time allocation table generation: After multiple rounds of iterative training, the deep reinforcement learning algorithm learns the optimal activity trigger time allocation strategy. Based on the trained strategy network, the system generates specific trigger time segments for each activity of each customer, forming a personalized activity trigger time allocation table.

[0107] Specifically, for each customer k and activity Trigger time (in ) is determined by the following formula: in, Indicates that in state S t Next, the activity Assigned to time segment T v Q value.

[0108] Step S3 uses a deep reinforcement learning algorithm, combined with the multi-dimensional customer preference vector generated by step S1 and the activity priority weight list and conflict marker list generated by step S2, to achieve personalized optimization scheduling of multi-activity trigger times. The specific process includes input data integration and state definition, action definition and strategy network construction, reward function design and optimization goal setting, deep reinforcement learning algorithm training and strategy optimization, and the final generation of the trigger time allocation table. In a complex multi-activity environment, it is possible to dynamically adjust and optimize the activity trigger time points to ensure efficient push and conflict-free execution of marketing activities, thereby significantly improving the intelligence level of the private domain traffic management system and the marketing conversion efficiency.

[0109] Embodiment 2: Figure 2 The present invention provides a private domain traffic monitoring, analysis and processing system, including: a customer clustering unit, a conflict detection unit and a scheduling optimization unit;

[0110] Customer clustering unit: Performs weighted time sliding window analysis and hierarchical clustering calculation on customer interaction data obtained from private domain traffic, inputs the obtained time series feature parameters into the adaptive preference modeling module, and outputs a multi-dimensional customer preference vector containing customer acceptability scores for various activity information in different time segments and corresponding time window data;

[0111] Conflict detection unit: inputs the multi-dimensional customer preference vector output by the customer clustering unit and the preset parallel activity list into the maximum matching optimization algorithm based on graph theory and the constraint satisfaction solver, performs time period conflict detection and weighted calculation on each parallel activity, and generates a data structure containing a list of activity priority weights and a list of corresponding conflict markers for each customer;

[0112] Scheduling optimization unit: The activity priority weight list and conflict mark list generated in the conflict detection unit and the multi-dimensional customer preference vector obtained in the customer clustering unit are input into the deep reinforcement learning algorithm for iterative optimization and solution, and the trigger time allocation table for each activity in the corresponding customer time segment is output.

[0113] The above formulas are all dimensionless and numerical calculations. The formula is a formula for the most recent real situation obtained by collecting a large amount of data and performing software simulation. The preset parameters in the formula are set by technicians in this field according to actual conditions.

[0114] The above description is only by way of illustration of certain exemplary embodiments of the present invention. It is undoubted that those skilled in the art can modify the described embodiments in various ways without departing from the spirit and scope of the present invention. Therefore, the above drawings and descriptions are illustrative in nature and should not be construed as limiting the scope of protection of the claims of the present invention.

[0115] It should be noted that, in this article, if there are relational terms such as first and second, etc., they are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the term includes, contains or any other variant thereof is intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, an element defined by the sentence includes a... does not exclude the existence of other identical elements in the process, method, article or device including the element.

[0116] The above is only a specific implementation of the present application, but the protection scope of the present application is not limited thereto. Any person skilled in the art who is familiar with the present technical field can easily think of changes or substitutions within the technical scope disclosed in the present application, which should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.

Claims

1. A private domain traffic monitoring, analysis and processing method, characterized in that: Includes steps: S1: Perform weighted time sliding window analysis and hierarchical clustering calculation on the customer interaction data obtained from private domain traffic, input the obtained time series feature parameters into the adaptive preference modeling module, and output a multi-dimensional customer preference vector containing the customer's acceptability scores for various activity information in different time segments and the corresponding time window data; S2: The multi-dimensional customer preference vector and the preset parallel activity list are input into the maximum matching optimization algorithm and constraint satisfaction solver based on graph theory, and time period conflict detection and weighted calculation are performed on each parallel activity to generate a data structure containing a list of activity priority weights and a list of corresponding conflict markers for each customer; S3: Input the step activity priority weight list, conflict marker list and multi-dimensional customer preference vector into the deep reinforcement learning algorithm for iterative optimization and solution, and output the trigger time allocation table for each activity in the corresponding customer time segment.

2. The private domain traffic monitoring and analysis processing method according to claim 1 is characterized in that: Step S1 includes the following contents: a1. First collect customer interaction data; a2. Use weighted time sliding window analysis method to capture the temporal dynamic characteristics of customer behavior; Set a time window size and assign different weights to data points in each window; a3. Use wavelet transform to perform multi-scale decomposition on the weighted time series data and extract the behavioral characteristics under different frequency components; a4. Take the extracted multi-scale behavior feature vector as input and use the hierarchical clustering algorithm to group customers; hierarchical clustering constructs a similarity matrix between customers. The similarity Derived from the dynamic time warping distance metric; a5. Input the clustering results into the adaptive preference modeling module and construct a multidimensional customer preference vector for each customer through multidimensional regression analysis; Multidimensional customer preference vector P k Contains customer k's acceptability rating P for different activity types a at different time segments t k,a,t .

3. The private domain traffic monitoring and analysis processing method according to claim 2 is characterized in that: Step S2 includes the following contents: b1. First receive the multi-dimensional customer preference vector and obtain the preset parallel activity list Each of these activities Contains its type, scheduled push time window and related attributes, is the index of the activity, which uniquely identifies each activity in the parallel activity list; b2. Map each activity into a multidimensional customer preference vector space to form an activity preference vector in represents customer k pairs of activities The acceptability score at time segment t; then, the similarity matrix between activities is calculated using cosine similarity 4. The private domain traffic monitoring and analysis processing method according to claim 3 is characterized in that: Step S2 also includes the following contents: b3. Construct an active conflict graph G = (V, E) where the vertex set The edge set E is generated based on the overlapping relationship of the activity time periods; if the activity and activities There is overlap in any time segment t, that is, Then add an edge to the active conflict graph Edge weight By combining similarity and time overlap.

5. The private domain traffic monitoring and analysis processing method according to claim 4 is characterized in that: Step S2 also includes the following contents: b4. Apply the maximum weight matching optimization algorithm on the constructed activity conflict graph to identify the maximum conflict pairs between activities; determine the key conflict pairs between activities by maximizing the sum of matching edge weights, and prioritize resolving conflicts with higher weights; Subsequently, the priority of activities is dynamically adjusted using a constraint satisfaction solver; The constraint satisfaction problem solver solves the activity priority weight list for each customer by defining variables, domains, and constraints. and a list of conflict markers in Indicates activity The priority weight of For activities Is there a time conflict flag? represents the total number of parallel marketing activities that customer k conducts simultaneously in the private domain traffic management system. (in ) represents an independent marketing campaign; b5. Finally, the generated data structure includes a list of activity priority weights W for each customer. k and conflict marker list C k .

6. The private domain traffic monitoring and analysis processing method according to claim 5 is characterized in that: Step S3 includes the following contents: c1. First, integrate the multi-dimensional customer preference vector, activity priority weight list and conflict marker list as the input data of the deep reinforcement learning algorithm; define the current state S t Contains the following: t = {P k ,W k ,C k }; c2. The action space is defined as the set of all activity triggering time segments For each customer k and its corresponding activity action Indicates that the activity will Assigned to time segment T v , Construct a policy network π in deep reinforcement learning, whose input is the current state S t , the output is for each activity In each time segment T v The probability of action selection under 7. The private domain traffic monitoring and analysis processing method according to claim 6 is characterized in that: Step S3 also Includes the following: c3. In order to guide the deep reinforcement learning algorithm to optimize the activity trigger time, the reward function R is designed to combine the conversion effect and customer satisfaction of the activity. The specific formula is as follows: in: Indicates activity In time segment T v ; α1 is the weight parameter, which controls the contribution of conversion rate and satisfaction to the reward respectively; γ1 is the penalty parameter, which is used to reduce the negative impact of time period conflict; is the indicator function, when the activity Assigned to time segment T v The value is 1 when , otherwise it is 0.

8. The private domain traffic monitoring and analysis processing method according to claim 7 is characterized in that: Step S3 also includes the following: c4. Use deep Q network as deep reinforcement learning algorithm; c5. After multiple rounds of iterative training, the deep reinforcement learning algorithm learns the optimal activity trigger time allocation strategy; based on the trained strategy network, a specific trigger time segment is generated for each activity of each customer, forming a personalized activity trigger time allocation table Specifically, for each customer k and activity Trigger time (in ) is determined by the following formula: in, Indicates that in state S t Next, the activity Assigned to time segment T v Q value.

9. The private domain traffic monitoring and analysis processing method according to claim 8 is characterized in that: Step S3 also includes the following: c4. Use deep Q network as the deep reinforcement learning algorithm; the specific training process is as follows: c4-1. Experience replay: Record the state, action, reward and next state of each interaction and store these experiences in the experience replay buffer c4-2.Q value update: randomly sample a batch of experience from the experience replay buffer Update the Q network parameters θ using the loss function; c4-3. Policy update: through ε - The greedy strategy selects action a t , where ε is gradually reduced to balance exploration and exploitation; the policy network selects the action with the highest Q value according to the Q value output by the current Q network Q(S,a;θ); c4-4. Iterative optimization: Repeat c4-1 to c4-3, and gradually optimize the strategy network so that it can assign the optimal activity triggering time point to each customer under the multi-objective optimization framework.

10. A private domain traffic monitoring, analysis and processing system, used to implement the private domain traffic monitoring, analysis and processing method according to any one of claims 1 to 9, characterized in that: include: Customer clustering unit, conflict detection unit and scheduling optimization unit; Customer clustering unit: Performs weighted time sliding window analysis and hierarchical clustering calculation on customer interaction data obtained from private domain traffic, inputs the obtained time series feature parameters into the adaptive preference modeling module, and outputs a multi-dimensional customer preference vector containing customer acceptability scores for various activity information in different time segments and corresponding time window data; Conflict detection unit: inputs the multi-dimensional customer preference vector output by the customer clustering unit and the preset parallel activity list into the maximum matching optimization algorithm based on graph theory and the constraint satisfaction solver, performs time period conflict detection and weighted calculation on each parallel activity, and generates a data structure containing a list of activity priority weights and a list of corresponding conflict markers for each customer; Scheduling optimization unit: The activity priority weight list and conflict mark list generated in the conflict detection unit and the multi-dimensional customer preference vector obtained in the customer clustering unit are input into the deep reinforcement learning algorithm for iterative optimization and solution, and the trigger time allocation table for each activity in the corresponding customer time segment is output.

Images