Financial marketing SAAS platform based on DeepSeek

Abstract

The present invention relates to the field of user behavior analysis technology, specifically a SAAS platform for financial marketing based on DeepSeek, wherein the platform is equipped with a control system, and the system includes a behavior feature acquisition module, a marketing strategy priority evaluation module, a window behavior analysis module, a strategy deployment adaptation module, and a resource dynamic mapping module. In the present invention, by collecting the time features of user behavior events and transaction type identifiers, combining financial product status labels with user portrait attributes, cross-platform timing alignment is completed to form a cross-system behavior timing chain, and the distribution features of high-frequency trading behaviors are extracted based on the behavior timing chain. The urgency of strategy requirements is dynamically corrected in combination with product status labels, and the time window is divided according to the strategy priority sequence. The trigger density features and time interval fluctuation features are extracted, and strategy optimization is performed by comparing the trigger density with the carrying threshold, dynamically allocating computing and storage resources, and improving resource allocation efficiency and system stability.

Description

Technical Field

[0001] The present invention relates to the technical field of user behavior analysis, and in particular to a SAAS platform for financial marketing based on DeepSeek. Background Art

[0002] The field of user behavior analytics encompasses technologies that collect, process, and model user behavioral data from financial activities to support marketing decisions. The core of this area is to identify potential needs and behavioral patterns by integrating multi-dimensional data such as user transaction records, browsing preferences, and interaction feedback, thereby generating targeted marketing strategies. This systematic technical process encompasses everything from raw data cleaning and feature extraction to behavioral pattern mining and dynamic prediction, ultimately forming actionable marketing plans. This ensures that financial institutions can optimize resource allocation and improve service precision based on objective behavioral data.

[0003] The DeepSeek-based financial marketing SAAS platform refers to a technical solution that utilizes a specific data processing architecture to integrate cross-platform user behavior data and generate financial product recommendation strategies through preset analysis models. The technical matters targeted by this patent subject include standardized access processes for multi-source heterogeneous behavior data, behavior patterns based on time series features and association rules, analysis methods, and strategy generation mechanisms for financial scenarios. By building a distributed data storage layer to achieve real-time collection of high-concurrency behavior logs, a rule engine is used to automatically map user behavior tags with dynamic weight calculations, and a strategy is combined with an execution interface to output analysis results directly to the marketing system of financial institutions to complete strategy deployment.

[0004] Existing technologies for collecting and processing user behavior data suffer from incomplete integration of multi-source, heterogeneous data. Standardized access mechanisms are limited to standardizing basic fields and fail to address the alignment requirements for complex time series features. This leads to timestamp drift and broken behavior chains in cross-platform behavioral data, weakening the accuracy of behavioral pattern recognition. Regarding strategy generation, existing technologies focus on rule-based or static model-based deductions and lack dynamic risk correction mechanisms. This distorts strategy prioritization and prevents timely adjustments to rapidly changing user needs, impacting the accuracy of marketing responses. During the behavioral analysis phase, coarse window division and low granularity in behavioral feature extraction result in a lack of detailed behavioral support during strategy execution, impacting strategy adaptation and execution effectiveness. Regarding resource scheduling, static or estimated resource allocation methods are often used, ignoring actual task volume fluctuations and resource usage dynamics. This can easily lead to resource waste or bottlenecks, resulting in system instability and delayed response. For example, during high-concurrency promotional periods, uneven resource allocation often leads to the failure or delayed delivery of some tasks, severely impacting overall marketing effectiveness and customer experience. Summary of the Invention

[0005] The purpose of the present invention is to solve the shortcomings of the prior art and to propose a SAAS platform for financial marketing based on DeepSeek.

[0006] To achieve the above objectives, the present invention adopts the following technical solutions: DeepSeek-based financial marketing SAAS platform includes:

[0007] The behavior feature collection module calls the platform interface to obtain the time features of user behavior events and transaction type identifiers, simultaneously collects financial product status tags and user profile attributes, aligns the time features of behavior events across platforms, and generates a cross-system behavior time series chain;

[0008] The marketing strategy priority assessment module extracts the distribution characteristics of high-frequency trading behaviors based on the cross-system behavior time series chain, calculates the urgency of strategy requirements in combination with the financial product status tags, dynamically modifies the urgency through the user's risk preference tags, and generates a strategy priority sequence;

[0009] The window behavior parsing module divides the cross-system behavior time series chain into time windows according to the policy classification in the policy priority sequence, extracts the trigger density characteristics and time interval fluctuation characteristics of the policy-related behaviors in the window, and generates a policy behavior feature set;

[0010] The policy deployment adaptation module compares the trigger density characteristics in the policy behavior feature set with the preset load threshold, executes policy splitting processing when the trigger density exceeds the load threshold, and executes policy optimization processing when the resource occupancy characteristics exceed the limit threshold, and generates a policy deployment plan and optimization instruction set.

[0011] As a further solution of the present invention, the cross-system behavior timing chain includes a behavior event timestamp, a transaction type identifier, a financial product status label, and a user portrait attribute; the strategy priority sequence includes high-frequency trading behavior distribution characteristics, strategy demand urgency, and a dynamic correction factor; the strategy behavior feature set includes trigger density characteristics, time interval fluctuation characteristics, and strategy-related behaviors; the strategy deployment plan and optimization instruction set include strategy splitting instructions, strategy optimization instructions, and resource occupancy limit parameters.

[0012] As a further solution of the present invention, the behavior feature collection module includes:

[0013] The behavior time extraction submodule obtains the time characteristics of user behavior events and transaction type identifiers returned by the platform interface, extracts the occurrence time, duration, and frequency parameters based on the event time field, matches the time characteristics with the transaction type identifier according to the chronological order, organizes the time intervals between adjacent events, and generates event time interval quantities;

[0014] The product status synchronization submodule synchronously collects financial product status tags and user profile attributes based on the event time interval, filters data records that are consistent with the behavioral event time, classifies and processes them according to the transaction type identifier and product status identifier, organizes the product status information corresponding to the transaction type, and generates transaction product status information;

[0015] The timing chain generation submodule organizes the time characteristics of behavioral events under the transaction type according to the status of the transaction product, arranges the event records in chronological order, filters the data segments with time intervals less than the synchronization reference value, groups them into continuous behavior chains, and generates a cross-system behavior timing chain.

[0016] As a further solution of the present invention, the marketing strategy priority evaluation module includes:

[0017] The transaction behavior extraction submodule obtains the cross-system behavior time series chain transaction data, detects transaction timestamps, types and amounts, collects continuous transaction sequences, classifies transaction features, filters high-frequency behaviors, and generates high-frequency transaction behavior distribution values;

[0018] The product status assessment submodule uses the financial product status tag based on the high-frequency trading behavior distribution value to detect the current status and changes of the product, extract the activity range and change range, and obtain the urgency of the strategy demand;

[0019] The priority sequence generation submodule calls the user risk preference label according to the urgency of the policy requirements, detects the risk level and tolerance range, sets the correction factor, adjusts the urgency measure, and sorts and generates the policy priority sequence.

[0020] As a further solution of the present invention, the window behavior parsing module includes:

[0021] The policy classification submodule obtains the cross-system behavior time series chain and policy priority sequence, assigns the behavior to the corresponding policy category according to the priority, arranges the policy categories based on the time sequence and labels the behavior nodes, counts the number of occurrences of the policy categories in the overall time series chain, and generates a policy distribution value;

[0022] The time window division submodule calls the strategy distribution value, determines the time distribution characteristics of the behavior node according to the density change of the strategy behavior in the time series chain, calibrates the time window boundary according to the change of the time interval of the behavior node, and generates the time window boundary value based on the accumulation of the strategy category;

[0023] The behavior feature extraction submodule calls the time window demarcation value, collects the strategy category behavior within the time window, extracts the time fluctuation characteristics according to the behavior triggering time series, and generates a strategy behavior feature set based on the local behavior quantity change characteristics.

[0024] As a further solution of the present invention, the specific calculation formula for counting the number of occurrences of strategy categories in the overall time series chain is:

[0025] ;

[0026] in, represents the weighted occurrence number of strategy category j, Represents the global priority correction coefficient of the behavior node in the timing chain, represents the original priority of the k-th behavior node belonging to policy category j, represents the number of time units between the kth behavior node and the end point of the current timing chain, d represents the total number of behavior nodes belonging to strategy category j, To prevent the smoothing constant from being zero in the denominator, represents the collaborative weight factor of cross-system behavior nodes, represents the number of behavior nodes of the mth associated subsystem in strategy category j, and q represents the total number of associated subsystems.

[0027] As a further solution of the present invention, the policy deployment adaptation module includes:

[0028] The density comparison submodule obtains the trigger density feature and the preset load threshold in the policy behavior feature set, calls the trigger density feature to compare with the preset load threshold, and filters the policy behavior features that exceed the load threshold interval according to the comparison result to generate the trigger density offset;

[0029] The resource judgment submodule obtains the resource occupancy characteristics and the limit threshold in the policy behavior feature set based on the trigger density deviation, calls the resource occupancy characteristics and the limit threshold to perform difference judgment, filters the resource occupancy characteristics that exceed the limit threshold, and generates the resource load breakthrough rate;

[0030] The solution generation submodule calls the policy behavior characteristics that exceed the load threshold according to the resource load breakthrough rate, divides the policy set according to the resource load breakthrough rate, calls the divided policy set to formulate deployment parameters and perform instruction mapping processing, and generates a policy deployment plan and an optimized instruction set.

[0031] As a further solution of the present invention, the specific calculation formula for comparing the call trigger density feature with the preset load threshold is:

[0032] ;

[0033] Among them, H represents the trigger density deviation trend parameter, ρ represents the average trigger density value of all samples in the strategy behavior feature set, and ρ irepresents the density statistics of the i-th feature dimension in the strategy behavior feature set in the sample set, T represents the benchmark value of the preset load threshold, represents the load threshold correction corresponding to the i-th characteristic dimension, α represents the dynamic adjustment coefficient, μ represents the harmonic mean of the historical density mean and the current density peak, γ represents the density fluctuation tolerance factor, and β i represents the nonlinear offset gain coefficient of the i-th feature dimension, η represents the threshold stability correction parameter, and n represents the total number of dimensions of the strategy behavior feature.

[0034] As a further embodiment of the present invention, the system further comprises:

[0035] The resource dynamic mapping module dynamically allocates computing resource capacity characteristics and storage resource quota characteristics based on the task volume characteristics and optimization amplitude characteristics of the policy deployment scheme and the optimization instruction set, and generates a cross-system resource configuration table;

[0036] The cross-system resource configuration table includes task volume characteristics, optimization range characteristics, computing resource capacity characteristics, and storage resource quota characteristics.

[0037] As a further solution of the present invention, the resource dynamic mapping module includes:

[0038] The task feature extraction submodule obtains the task volume data and task category parameters in the strategy deployment plan and the optimization instruction set, calls the task volume data to determine the capacity demand value, calls the task category parameters to identify the attribute classification, selects high-demand task items based on the capacity demand value and the attribute classification, and generates the task capacity feature value;

[0039] The optimization amplitude parsing submodule obtains the optimization target parameter and optimization priority parameter in the optimization instruction set based on the task capacity characteristic value, calls the optimization target parameter to determine the adjustment range value, calls the optimization priority parameter to arrange the application order, selects high-priority task items according to the adjustment range value and the application order, and generates the task adjustment amplitude characteristic value;

[0040] The resource relationship generation submodule obtains the initial capacity parameters of computing resources and the initial quota parameters of storage resources according to the task adjustment amplitude characteristic value, calls the initial capacity parameters of computing resources to determine the allocation ratio, calls the initial quota parameters of storage resources to determine the adjustment ratio, establishes a mapping parameter set based on the allocation ratio and the adjustment ratio, and generates a cross-system resource configuration table.

[0041] Compared with the prior art, the advantages and positive effects of the present invention are:

[0042] In the present invention, by collecting the time characteristics of user behavior events and transaction type identifiers, combining financial product status tags and user portrait attributes, cross-platform timing alignment is completed to form a cross-system behavior timing chain. Based on the behavior timing chain, the distribution characteristics of high-frequency trading behaviors are extracted, and the urgency of strategy requirements is dynamically corrected in combination with product status tags. The time window is divided according to the strategy priority sequence, and the trigger density characteristics and time interval fluctuation characteristics are extracted. By comparing the trigger density with the carrying threshold, strategy optimization is performed, and computing and storage resources are dynamically allocated to improve resource allocation efficiency and system stability. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 is a system flow chart of the present invention;

[0044] Figure 2 This is a system block diagram of the present invention. DETAILED DESCRIPTION

[0045] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0046] In the description of the present invention, it should be understood that the terms "length," "width," "up," "down," "front," "back," "left," "right," "vertical," "horizontal," "top," "bottom," "inside," "outside," and the like, indicating positions or relationships, are based on the positions or relationships shown in the accompanying drawings and are intended only to facilitate the description of the present invention and simplify the description. They do not indicate or imply that the devices or elements referred to must have a specific orientation, be constructed, or operate in a specific orientation. Therefore, they should not be construed as limiting the present invention. Furthermore, in the description of the present invention, "plurality" means two or more, unless otherwise expressly and specifically defined.

[0047] See also Figure 1-2 The financial marketing SAAS platform based on DeepSeek includes:

[0048] The behavior feature collection module calls the platform interface to obtain the time features of user behavior events and transaction type identifiers, simultaneously collects financial product status tags and user profile attributes, aligns the time features of behavior events across platforms, and generates a cross-system behavior time series chain;

[0049] The marketing strategy priority assessment module extracts the distribution characteristics of high-frequency trading behaviors based on the cross-system behavior time series chain, calculates the urgency of strategy requirements based on the financial product status tags, and dynamically adjusts the urgency based on the user's risk preference tags to generate a strategy priority sequence.

[0050] The window behavior parsing module divides the cross-system behavior time series chain into time windows based on the policy classification in the policy priority sequence, extracts the trigger density characteristics and time interval fluctuation characteristics of the policy-related behaviors within the window, and generates a policy behavior feature set;

[0051] The policy deployment adaptation module compares the trigger density characteristics in the policy behavior feature set with the preset load threshold. When the trigger density exceeds the load threshold, the policy splitting process is executed. When the resource occupancy characteristics exceed the limit threshold, the policy optimization process is executed to generate a policy deployment plan and optimization instruction set.

[0052] The resource dynamic mapping module dynamically allocates computing resource capacity characteristics and storage resource quota characteristics based on the task volume characteristics and optimization amplitude characteristics in the policy deployment plan and optimization instruction set, and generates a cross-system resource configuration table.

[0053] The cross-system behavior timing chain includes the behavior event timestamp, transaction type identifier, financial product status label, and user portrait attributes. The strategy priority sequence includes the distribution characteristics of high-frequency trading behavior, the urgency of strategy requirements, and the dynamic correction factor. The strategy behavior feature set includes the trigger density characteristics, the time interval fluctuation characteristics, and the strategy association behavior. The strategy deployment plan and optimization instruction set include the strategy splitting instructions, the strategy optimization instructions, and the resource occupation limit parameters. The cross-system resource configuration table includes the task volume characteristics, the optimization amplitude characteristics, the computing resource capacity characteristics, and the storage resource quota characteristics.

[0054] See also Figure 2 , the behavior feature collection module includes:

[0055] The behavior time extraction submodule obtains the time characteristics of user behavior events and transaction type identifiers returned by the platform interface, extracts the occurrence time, duration, and frequency parameters based on the event time field, matches the time characteristics with the transaction type identifier according to the chronological order, organizes the time intervals between adjacent events, and generates event time interval quantities;

[0056] The behavior time extraction submodule obtains the user behavior event time characteristics and transaction type identification returned by the platform interface. First, it reads the data records returned by the platform interface, including the event time field and the transaction type identification field. It reads each record, extracts the event time as the occurrence time, and directly records the time field as the specific time point of the event, such as 10:00 on April 1, 2025. If the event duration field exists, it is directly used. If not, the default duration of each event is 5 seconds. The frequency parameter is obtained by counting the number of times the same transaction type behavior occurs every day. The statistical method is to filter out the number of records of the same transaction type for a certain user in one day, for example, If a user makes three payments in one day, the payment frequency is recorded as three times per day. After sorting all events, they are arranged in ascending time order. After the events of the same user are arranged in sequence, the time interval between two adjacent events is calculated. The interval time is obtained by directly reading the difference between the time fields of the latter and the previous event. For example, if the first time is 10:00 and the second time is 10:05, the time interval is 5 minutes. After sorting all time intervals, the interval range is defined. A time interval less than 5 minutes is defined as a short interval, and a time interval greater than or equal to 5 minutes is defined as a long interval. By sorting out the number of time intervals in different intervals, an overall feature description and classification record of the event time interval quantity is generated.

[0057] The product status synchronization submodule collects financial product status tags and user profile attributes based on the event time interval, filters data records that are consistent with the behavioral event time, classifies and processes them according to the transaction type identifier and product status identifier, organizes the product status information corresponding to the transaction type, and generates transaction product status data;

[0058] The product status synchronization submodule synchronously collects financial product status tags and user profile attributes based on the time interval between events. During data collection, it extracts the product status field and user profile field from the financial product system records and selects data records that coincide with the behavioral event or have a time deviation of no more than one minute. For example, if a behavioral event occurs at 10:00 and the product status is recorded at 10:01, the two are considered a match. During the screening process, a direct comparison is made to determine whether the difference in seconds between the two time fields is less than or equal to 60 seconds. After the screening is completed, the behavioral event transaction type and the corresponding product status are classified. For example, if the transaction type is purchase and the product status is on sale, this event is classified into the purchase and sale product collection. Then, all classified records are sorted and the number of different product statuses under each transaction type is counted. For example, if the on sale status appears ten times and the unsale status appears twice in a purchase transaction, the on sale product ratio is 83.33%. A baseline value is set at 95%. If the ratio falls below this baseline, the data type is marked as requiring synchronization and correction, completing the classification process and generating transaction product status data.

[0059] The time series chain generation submodule organizes the time characteristics of behavioral events under the transaction type based on the transaction product status, arranges the event records in chronological order, filters out data segments with time intervals less than the synchronization benchmark value, groups them into continuous behavioral chains, and generates a cross-system behavioral time series chain;

[0060] The time series chain generation submodule organizes the time characteristics of the corresponding behavioral events under each transaction type based on the transaction product status. First, all behavioral event records under each transaction type are extracted, such as the event record list of all payment transactions. All records are sorted in ascending order according to the time field. After sorting, the time interval of each pair of adjacent events is calculated in turn. If the time interval between an event and the next event is less than five minutes, the two events are classified as part of the same continuous behavioral chain. For example, the first event time is 10 o'clock, and the second event time is 10:02. The time interval is two minutes, which is less than the benchmark value of five minutes. Therefore, the two payment transactions are grouped together. In order to be included in the same chain, if the time interval exceeds five minutes, it will be divided into a new chain and grouped in turn to form a continuous behavior chain. During the chain sorting process, the start time, end time, transaction type and number of behaviors contained in each chain are recorded at the same time. For example, a chain starts at 10 o'clock, ends at 10:02, the transaction type is payment, and contains two behavior events. For the chain length division standard, a single behavior chain is defined as a low-frequency chain, a two-behavior chain is a medium-frequency chain, and a three-or-more behavior chain is a high-frequency chain. The generation and sorting of cross-system behavior timing chains are completed in the above manner.

[0061] See also Figure 2 , the marketing strategy priority assessment module includes:

[0062] The transaction behavior extraction submodule obtains cross-system behavior time series chain transaction data, detects transaction timestamps, types, and amounts, collects continuous transaction sequences, classifies transaction features, filters high-frequency behaviors, and generates high-frequency transaction behavior distribution values;

[0063] The transaction behavior extraction submodule must first be connected to the data interfaces of multiple independent transaction systems, and continuously call the data acquisition service to complete the cross-platform transaction record collection. Each record must include user identification, transaction time, transaction type and transaction amount. The transaction time detection is based on whether the time interval between adjacent transactions is less than 5 minutes. If the time difference between two adjacent transactions is less than or equal to 5 minutes, it is considered a continuous transaction. For example, the transfer of 500 yuan at 08:00 and the payment of 120 yuan at 08:03 are only 3 minutes apart, which meets the continuity standard. When classifying transaction types, they are matched according to the pre-established coding table. For example, transfers are classified as code T01, consumption is classified as code T02, and transaction amounts are classified according to the set intervals. For example, amounts from 0 to 100 yuan are classified as the first tier, 100 to 500 yuan are classified as the second tier, and 500 to 1000 yuan are classified as the third tier. Each zone The number of transactions between periods is accumulated. When screening for high-frequency behaviors, the proportion of the number of transactions in each period to the total number of transactions needs to be calculated. When the proportion of transactions in a certain amount period exceeds 50%, it is determined to be a high-frequency trading period. For example, in a certain day's records, there are 23 total transactions, of which 15 are in the amount range of 0 to 100 yuan. The calculated proportion is 15 divided by 23, which is approximately 65%. If it exceeds the 50% threshold, it is marked as a high-frequency trading period, and the distribution value of high-frequency trading behavior is set to 65%. In the process of detecting continuous transactions, the traversal comparison method can be used to calculate the difference between the current transaction time and the previous transaction time in sequence. The transaction feature classification process adopts the type dictionary direct matching method and is divided into grades according to the amount. The screening process is judged based on the set frequency threshold. The distribution value generation process is the ratio conversion of the number of high-frequency occurrences in each amount period to the total number of times, and finally the distribution value extraction and archiving of high-frequency trading behavior are completed.

[0064] The product status assessment submodule uses the distribution value of high-frequency trading behavior to call the financial product status label, detect the current status and changes of the product, extract the activity range and change range, and obtain the urgency of strategic requirements;

[0065] After receiving the high-frequency trading behavior distribution value, the product status assessment submodule needs to synchronously call the latest status label of the corresponding financial product. Status labels such as active and stagnant are set by the unified standard of the financial system. When dividing the activity interval, 0 to 0.3 is low activity, 0.3 to 0.7 is medium activity, and 0.7 to 1.0 is high activity. The obtained high-frequency trading behavior distribution value is mapped to the activity interval. For example, the distribution value of a certain product is 0.652, which falls within the medium activity interval. Then, to detect the change in product status, the activity value of the previous period and the current activity value must be compared. For example, the activity value of the previous period was 0.45, and the current value is 0.652. The difference between the two is 0.202. This change exceeds the preset change threshold of 0.1, so it is classified as a significant change in activity. When determining the urgency of strategy requirements, the grading can be set according to the change amplitude. For example, a change of less than or equal to 0.1 is classified as low urgency, greater than 0.1 and less than or equal to 0.3 is classified as medium urgency, and greater than 0.3 is classified as high urgency. In this example, the change is 0.202, which falls into the medium urgency range. The detection process requires selecting the activity flag field from the time series transaction record, and calculating the current product change status through field comparison and change range calculation. The activity range extraction process is completed by directly comparing the activity value with the upper and lower bounds of the set interval. The change range is calculated by subtracting the historical value from the current value. For example, the activity increases from 0.45 to 0.652, with a change range of 0.202. The change range determination range is set according to business statistical standards. For example, based on the change range of general financial product activity in the market, most changes are concentrated between 0.05 and 0.25. Therefore, it is reasonable to set the medium urgency range to 0.1 to 0.3.

[0066] The priority sequence generation submodule calls the user's risk preference label according to the urgency of the strategy requirements, detects the risk level and tolerance range, sets the correction factor, adjusts the urgency measurement, and sorts and generates the strategy priority sequence value;

[0067] After receiving the urgency level of the strategy requirement, the priority sequence generation submodule needs to synchronously read the user's risk preference label. Risk preferences are usually divided into low risk, medium risk, and high risk. The risk tolerance range is directly associated with the risk preference label. For example, medium and high risk preferences correspond to the high risk range of 0.6 to 1.0. When setting the urgency correction factor, the corresponding factor value is determined by the risk preference type. For example, the low risk preference setting factor is 0.8, the medium risk setting factor is 1.0, and the high risk setting factor is 1.2. If the urgency level is medium urgency, the corresponding basic urgency measurement setting value is 2. The corrected urgency measurement is obtained by multiplying the basic urgency measurement with the risk preference correction factor. For example, the medium urgency base value 2 is multiplied by the high risk correction factor. Factor 1.2, the corrected urgency measure is 2.4, and then the strategies are prioritized according to the corrected urgency measure values, and all strategies are arranged in descending order. For example, strategy A has a corrected urgency measure of 3.0, strategy B has a corrected urgency measure of 2.4, and strategy C has a corrected urgency measure of 1.8. Finally, the strategy execution order is formed. The user risk preference field needs to be directly extracted from the customer portrait system. The risk level matching is completed by comparing the risk preference and the risk level interval. The setting of the correction factor is formulated with reference to the risk-bearing capacity standard document. The adjustment of the urgency measure is obtained by simply multiplying the basic urgency measure value and the correction factor. The sorting process adopts the descending order method. The sorted priority list is used for subsequent strategy module calls.

[0068] See also Figure 2 , the window behavior parsing module includes:

[0069] The policy classification submodule obtains the cross-system behavior time series chain and policy priority sequence, assigns behaviors to corresponding policy categories based on priority, arranges policy categories based on chronological order and labels behavior nodes, counts the number of occurrences of policy categories in the overall time series chain, and generates a policy distribution value;

[0070] The specific calculation formula for counting the number of occurrences of strategy categories in the overall time series chain is:

[0071] ;

[0072] in, represents the weighted occurrence number of strategy category j, Represents the global priority correction coefficient of the behavior node in the timing chain, represents the original priority of the k-th behavior node belonging to policy category j, represents the number of time units between the kth behavior node and the end point of the current timing chain, d represents the total number of behavior nodes belonging to strategy category j, To prevent the smoothing constant from being zero in the denominator, represents the collaborative weight factor of cross-system behavior nodes, represents the number of behavior nodes of the mth associated subsystem in strategy category j, and q represents the total number of associated subsystems;

[0073] Global priority correction factor The standard deviation of the priority distribution of nodes in the cross-system behavior timing chain is calculated. The standard deviation is 0.45 and the mean priority is 7.2. =1.1 by It is concluded that the correction coefficient increases with the increase of standard deviation;

[0074] The total number of behavior nodes belonging to policy category j is d=3, which is obtained by counting the actual number of behavior nodes belonging to the temporal chain;

[0075] The original priority of the first behavior node , the second , the third ,The priority is quantified by the task execution urgency score in the system log, ,which ranges from 1 to 10;

[0076] Interval time =2, =5, =0, the unit is hour, calculated by the time stamp difference of the timing chain node;

[0077] Smoothing constant =0.1, set according to the minimum time unit of the timing chain 0.1 hour to prevent When the denominator is zero;

[0078] Synergy Weight Factor =0.9, calculated based on the cross-system interaction frequency accounting for 45%, , the weight increases linearly with the interaction frequency;

[0079] Number of subsystem behavior nodes , ,According to the statistics of subsystem activity log, the activity threshold ≥ 3 times / hour is converted into the number of nodes, and the total number of associated subsystems q=2.

[0080] Calculation process:

[0081] Compute the absolute value term:

[0082] Node 1: ;

[0083] Node 2: ;

[0084] Node 3: ;

[0085] Sum of absolute values: 6.07+2.92+31.33=40.32;

[0086] Calculate synergy terms:

[0087] Subsystem node mean: ;

[0088] Synergy value: 0.9×4.5=4.05;

[0089] Total weighted statistic value:

[0090] ;

[0091] The results show that the weighted statistical value of 44.37 is the comprehensive behavior density index of strategy category j in the time chain. The numerical results quantify the dynamic weight characteristics of the strategy distribution value by superimposing the priority correction term and the subsystem coordination term, and directly serve as the input data for generating the strategy distribution value.

[0092] The time window division submodule calls the strategy distribution value, determines the time distribution characteristics of the behavior node according to the density change of the strategy behavior in the time series chain, calibrates the time window boundary according to the change of the time interval of the behavior node, and generates the time window boundary value based on the accumulation of the strategy category;

[0093] After calling the strategy distribution value, the time window division submodule first sets the initial time window width, for example, every 5 minutes is a basic time window, and then analyzes the time interval between the behavior nodes, and calculates the difference in their occurrence time for two adjacent behavior nodes. For example, behavior a1 occurs at 10:00 and b1 occurs at 10:02. The time interval between the two is 2 minutes, which is less than the set 5-minute window width. Therefore, it is determined that the two belong to the same time window. When it is found that the time interval between two behavior nodes is significantly greater than the set time window width, for example, the interval is 15 minutes, which far exceeds the 5-minute threshold, a new time window is divided here. For the method of judging density changes, all behavior nodes are traversed, and the time difference of each pair of adjacent nodes is continuously calculated. When multiple large time intervals appear continuously, it means that the behavior density is high. The degree of decrease requires further refinement of the time window division, and the accumulation of policy categories is also used as an auxiliary basis. The number of policy category occurrences in a window is accumulated. For example, in a window, policy category 1 appears 3 times, reaching the set threshold of 3 times, and the current window boundary is confirmed. When the number of behaviors is less than the set threshold, the window is extended until the requirement is met. To determine whether to create a new time window, the method of real-time comparison of the current time interval and the threshold is adopted. If the time interval is greater than 5 minutes, a new window is immediately delineated. For the definition of high density and low density, a time interval of less than 2 minutes is set as high density, and greater than 10 minutes is set as low density. For example, a time interval of 1 minute is determined to be high density, and a time interval of 12 minutes is determined to be low density. Finally, through continuous comparison and the accumulation of the number of policy categories, the time window boundary value, that is, the start and end boundaries of each window, is generated.

[0094] The behavior feature extraction submodule calls the time window demarcation value to collect the strategy category behavior within the time window, extracts the time fluctuation characteristics based on the behavior trigger time series, and combines the local behavior quantity change characteristics to generate the strategy behavior feature set;

[0095] After calling the time window demarcation value, the behavior feature extraction submodule collects all policy category behaviors recorded in each time window in turn, and then extracts the time fluctuation characteristics based on the triggering time point of each behavior. For any two adjacent behavior nodes, their time difference is recorded, and the overall time change trend is obtained by summarizing all time differences. For example, if the time intervals are mostly concentrated in 1 to 3 minutes, it means that the behavior occurs frequently and the fluctuation is small. If the time interval changes in a large range, such as from 2 minutes to 20 minutes, it means that the fluctuation is large. Further combined with the local behavior quantity change characteristics, the number of behaviors is counted within each time window, such as every 5-minute window. For example, if 6 behaviors occur in a 5-minute window, the behavior is dense. The degree is 6 times divided by 5 minutes, and the density is 1.2 times per minute. Then compare the behavior density changes in adjacent time windows one by one. For example, the density of the first window is 1.2 times / minute, and the density of the second window is 2 times / minute, and the density increases by 0.8 times / minute. If the change threshold is set to 0.5 times / minute, it is considered that the behavior density has increased significantly. As for how to set the standard for significant fluctuations, it is assumed that a change value greater than 1 time / minute is judged to be a significant increase, less than -1 time / minute is judged to be a significant decrease, and changes between the two are judged to be general. Finally, based on the time mean, time fluctuation amplitude, local behavior number and change situation in each window, a complete set of strategic behavior feature sets is extracted to provide basic data support for subsequent processing.

[0096] See also Figure 2 , the policy deployment adaptation module includes:

[0097] The density comparison submodule obtains the trigger density feature and the preset load threshold in the policy behavior feature set, calls the trigger density feature to compare with the preset load threshold, and filters the policy behavior features that exceed the load threshold range based on the comparison results to generate the trigger density offset;

[0098] The specific calculation formula for comparing the trigger density feature with the preset load threshold is:

[0099] ;

[0100] Among them, H represents the trigger density deviation trend parameter, ρ represents the average trigger density value of all samples in the strategy behavior feature set, and ρ i represents the density statistics of the i-th feature dimension in the strategy behavior feature set in the sample set, T represents the benchmark value of the preset load threshold, represents the load threshold correction corresponding to the i-th characteristic dimension, α represents the dynamic adjustment coefficient (calculated by the geometric mean of the system load factor and the resource utilization rate), μ represents the harmonic mean of the historical density mean and the current density peak, γ represents the density fluctuation tolerance factor (calculated by the product of the historical maximum density difference and the current environment coefficient), βi represents the nonlinear offset gain coefficient of the i-th feature dimension, η represents the threshold stability correction parameter (calculated by the ratio of the variance of the preset threshold to the sample size), and n represents the total number of dimensions of the strategic behavior feature;

[0101] Parameter assignment and acquisition methods:

[0102] ρ i The data is collected through real-time monitoring system, and the value range is 0.1 to 2.0. The measured sample data is ρ i =0.85, ρ2=1.20, ρ3=1.50.

[0103] T: The base value of the preset load threshold, defined by the system configuration file, is T=0.75.

[0104] α is calculated as the geometric mean of the system load factor (monitored value 0.6) and the resource utilization (monitored value 0.7). The formula is: .

[0105] The historical density average of μ is 0.5 (the average value of the past 30 days is queried through the historical database), and the current density peak is 0.8 (the maximum value of real-time monitoring). The formula is: .

[0106] γ is calculated as the product of the historical maximum density difference (the maximum fluctuation value of historical data is 0.3) and the current environmental coefficient (the environmental monitoring value is 0.9). The formula is: ×0.9=0.27.

[0107] β i Output from the feature importance analysis model, the value range is 0.1 to 0.6, and β1=0.55, β2=0.30, and β3=0.48 are taken.

[0108] η is calculated as the ratio of the preset threshold variance (calculated value 0.1) to the sample size (current sample size 100), and the formula is .

[0109] The measured number of dimensions n=3.

[0110] Formula calculation derivation process:

[0111] The first term of the calculation formula:

[0112] ;

[0113] The results show that the dynamic deviation between the current trigger density feature and the threshold is 0.370 after normalization, reflecting that the density fluctuation is within the tolerance range.

[0114] The second term of the calculation formula:

[0115] Summing over three dimensions:

[0116] Dimension 1:

[0117] ;

[0118] Dimension 2:

[0119] ;

[0120] Dimension 3:

[0121] ;

[0122] Sum result: 11.704+8.049+14.914=34.667;

[0123] Calculate H:

[0124] H=0.370+34.667=35.037;

[0125] The result shows that the trigger density offset trend parameter H is 35.037. When H exceeds the preset threshold (such as H ≥ 1.0), it is determined that the corresponding strategy behavior characteristics exceed the load threshold range, and the trigger density offset degree needs to be generated.

[0126] The resource judgment submodule obtains the resource occupancy characteristics and limit thresholds in the policy behavior feature set based on the trigger density deviation, calls the resource occupancy characteristics and limit thresholds to determine the difference, filters the resource occupancy characteristics that exceed the limit threshold, and generates the resource load breakthrough rate;

[0127] The resource judgment submodule extracts resource usage characteristics, such as CPU usage and memory usage, from overloaded policy behavior features. It then sets reasonable limit thresholds for each resource, such as 80% for CPU usage and 75% for memory usage. It then sequentially reads the current resource usage for each policy behavior feature and compares the extracted usage with the corresponding limit thresholds. If the CPU usage for a policy behavior feature reaches 85% while the set limit is only 80%, the feature is considered to have exceeded the CPU resource limit. If the memory usage is 70%, it is considered within the reasonable range. In a real-world scenario, for example, if the CPU resource usage monitoring data for a cloud server instance shows 85%, exceeding the standard value of 80%, it is directly classified as a resource load violation. To quantify the severity of the resource load violation, the difference between the resource usage and the limit threshold is calculated and converted into a percentage of the limit threshold to obtain the resource load violation rate. In this example, the CPU limit violation is 6.25%, which is recorded in the resource load violation rate data set for subsequent processing.

[0128] The solution generation submodule calls the policy behavior characteristics that exceed the load threshold based on the resource load breakthrough rate, divides the policy set according to the resource load breakthrough rate, calls the divided policy set to formulate deployment parameters and perform instruction mapping processing, and generates a policy deployment plan and optimized instruction set;

[0129] The solution generation submodule further categorizes policy behavior characteristics based on resource load breakthrough rates, setting resource breakthrough rate classification criteria. For example, a breakthrough rate of 0% to 5% is classified as light overload, 5% to 10% as moderate overload, and 10% or more as severe overload. The breakthrough rate values ​​corresponding to each policy behavior characteristic are extracted and categorized according to the breakthrough rate range. For example, if a policy behavior characteristic has a breakthrough rate of 6.25%, it falls into the moderate overload category, while another policy behavior characteristic with a breakthrough rate of 11% falls into the severe overload category. Targeted deployment parameters are formulated based on the classification results. Moderate overload policy behavior characteristics are handled using rate limiting and caching mechanisms, while severe overload policy behavior characteristics employ more stringent isolation and circuit breaker measures. Specifically, the moderate overload category can be capped at three accesses per second and have a 30-second cache expiration. The severe overload category can have a circuit breaker mechanism activated, with three consecutive failures triggering isolation. Taking a microservice interface as an example, if the breach rate of interface A is 6.25%, a rate limit of 3 calls per second is set for it, and caching is added to ensure a stable load on system resources. Finally, based on the breach rate and classification of each policy's behavioral characteristics, a comprehensive policy deployment plan and corresponding optimization instruction set are generated for the system.

[0130] See also Figure 2 , the resource dynamic mapping module includes:

[0131] The task feature extraction submodule obtains the task volume data and task category parameters from the strategy deployment plan and optimization instruction set, uses the task volume data to determine the capacity demand value, uses the task category parameters to identify the attribute classification, selects high-demand task items based on the capacity demand value and attribute classification, and generates the task capacity feature value;

[0132] The task feature extraction submodule extracts task volume data and task category parameters from the strategy deployment plan and optimization instruction set. First, according to the time nodes and task distribution records in the deployment plan, the number of tasks in each time period is queried. For example, the task volume in April was recorded as 1,200, and the average task inflow rate per hour was 50. In the stage of extracting task category parameters, the task type labels are divided into categories A, B, and C, and high priority, normal, and low priority weights are set respectively. The weight of category A is set to 0.8, category B is set to 0.5, and category C is set to 0.2. In the link of determining the capacity demand value, the extracted task volume is multiplied by the average inflow rate for calculation. For example, 1,200 tasks require 60,000 units of processing capacity at an average inflow rate of 50 per hour. Then, based on the task category parameters, according to the set The tasks are classified according to the specified attribute classification threshold. If the capacity requirement of category A tasks is greater than or equal to 20,000 units, they are classified as high-demand tasks. The capacity requirement of category B tasks is between 10,000 and 19,999 units. The capacity requirement of category C tasks is less than 10,000 units. According to the above set standards, tasks are screened and high-demand tasks that meet the conditions are selected. For example, 500 tasks are screened out of category A, and the average capacity requirement of each task is 25,000 units. Finally, the task capacity characteristic value is formed by multiplying the number of tasks by the category weight. For example, the capacity characteristic value of category A tasks is 500 multiplied by 0.8, which is 400. Combined with the specific system, for example, in a logistics sorting system, if the average daily package processing volume is 1,200, high-priority sorting tasks need to be pre-configured with 400 units of capacity resources to achieve targeted and efficient resource allocation.

[0133] The optimization amplitude analysis submodule obtains the optimization target parameters and optimization priority parameters in the optimization instruction set based on the task capacity characteristic value, calls the optimization target parameters to determine the adjustment range value, calls the optimization priority parameters to arrange the application order, selects high-priority task items based on the adjustment range value and application order, and generates the task adjustment amplitude characteristic value;

[0134] The optimization amplitude parsing submodule extracts the optimization target parameters and optimization priority parameters from the optimization instruction set based on the generated task capacity characteristic value. In the stage of extracting the optimization target parameters, the minimization of processing time and the optimization of resource occupancy are set as the main goals, among which the minimization of processing time is given a weight of 0.7 and the optimization of resources is given a weight of 0.3. In the stage of calling the optimization target parameters to determine the adjustment range value, the task capacity characteristic value is combined with the optimization target coefficient. If the task capacity characteristic value is 400 and the processing time minimization coefficient is 0.9, the adjustment range value is about 360. In the stage of calling the optimization priority parameters to arrange the application order, the processing time is minimized according to the weight from high to low. Put it first, and resource optimization comes second. According to the adjustment range value and priority application order, high-priority task items are screened. The screening criteria are tasks with an adjustment range value higher than 300 units and a corresponding optimization target weight greater than 0.5. For example, 300 priority optimization tasks are screened, and the adjustment amplitude characteristic value of each task is obtained by multiplying the adjustment range value by the priority weight, such as 360 multiplied by 0.7 to get 252. Combined with actual scenarios, such as in data center task scheduling optimization, for video encoding tasks with high processing time requirements, high-priority processing tasks are screened out according to the above method and optimization strategies are assigned to form a specific adjustment task list, which provides a clear basis for subsequent resource reallocation.

[0135] The resource relationship generation submodule obtains the initial capacity parameters of computing resources and the initial quota parameters of storage resources based on the task adjustment amplitude characteristic value, calls the initial capacity parameters of computing resources to determine the allocation ratio, calls the initial quota parameters of storage resources to determine the adjustment ratio, establishes a mapping parameter set based on the allocation ratio and the adjustment ratio, and generates a cross-system resource configuration table;

[0136] The resource relationship generation submodule obtains the initial capacity parameters of computing resources and the initial quota parameters of storage resources based on the task adjustment amplitude characteristic value. When extracting the initial capacity parameters of computing resources, the system baseline configuration data is read. For example, the initial computing resources are 80,000 core hours. In the stage of extracting the initial quota parameters of storage resources, the total amount of storage space of the initial configuration is read, such as 100TB. In the stage of calling the initial capacity parameters of computing resources to determine the allocation ratio, the task adjustment amplitude characteristic value is compared with the total computing capacity. For example, the task adjustment amplitude characteristic value is 252 units. Compared with the total computing capacity of 80,000 core hours, the allocation ratio is about 0.00315. In the stage of calling the initial quota parameters of storage resources to determine the adjustment ratio, the task adjustment amplitude characteristic value is compared with the total storage capacity. For comparison, if 100TB is converted to 102400MB, the adjustment ratio is approximately 0.00246. Based on the above allocation ratio and adjustment ratio, a mapping parameter set is established to allocate a clear CPU resource ratio and storage resource ratio for each task. For example, a mapping table is established to mark the resource allocation ratio corresponding to the task ID. Task001 is allocated 0.00315 computing resources and 0.00246 storage resources. In actual system configuration, resource scheduling platforms such as Kubernetes can be used to issue resource allocation instructions based on the above mapping data to ensure that tasks can be reasonably allocated in proportion among different subsystems. For example, in a hybrid cloud environment, data backup tasks are preferentially allocated to high-capacity nodes based on this configuration to reduce overall resource conflicts.

[0137] The above are merely preferred embodiments of the present invention and do not limit the present invention in any other form. Any technician familiar with the profession may use the technical content disclosed above to change or modify it into an equivalent embodiment with equivalent changes and apply it to other fields. However, any simple modification, equivalent change and modification made to the above embodiment based on the technical essence of the present invention without departing from the content of the technical solution of the present invention shall still fall within the scope of protection of the technical solution of the present invention.

Claims

1. A financial marketing SAAS platform based on DeepSeek, equipped with a control system, characterized by: The system comprises: The behavior feature collection module calls the platform interface to obtain the time features of user behavior events and transaction type identifiers, simultaneously collects financial product status tags and user profile attributes, aligns the time features of behavior events across platforms, and generates a cross-system behavior time series chain; The marketing strategy priority assessment module extracts the distribution characteristics of high-frequency trading behaviors based on the cross-system behavior time series chain, calculates the urgency of strategy requirements in combination with the financial product status tags, dynamically modifies the urgency through the user's risk preference tags, and generates a strategy priority sequence; The window behavior parsing module divides the cross-system behavior time series chain into time windows according to the policy classification in the policy priority sequence, extracts the trigger density characteristics and time interval fluctuation characteristics of the policy-related behaviors in the window, and generates a policy behavior feature set; The policy deployment adaptation module compares the trigger density characteristics in the policy behavior feature set with the preset load threshold, executes policy splitting processing when the trigger density exceeds the load threshold, and executes policy optimization processing when the resource occupancy characteristics exceed the limit threshold, and generates a policy deployment plan and optimization instruction set.

2. The DeepSeek-based financial marketing SAAS platform according to claim 1, characterized in that: The cross-system behavior timing chain includes the behavior event timestamp, transaction type identifier, financial product status label, and user portrait attributes; the strategy priority sequence includes the high-frequency trading behavior distribution characteristics, strategy demand urgency, and dynamic correction factors; the strategy behavior feature set includes trigger density characteristics, time interval fluctuation characteristics, and strategy-related behaviors; the strategy deployment plan and optimization instruction set include strategy splitting instructions, strategy optimization instructions, and resource occupancy limit parameters.

3. The DeepSeek-based financial marketing SAAS platform according to claim 1, characterized in that: The behavior feature collection module includes: The behavior time extraction submodule obtains the time characteristics of user behavior events and transaction type identifiers returned by the platform interface, extracts the occurrence time, duration, and frequency parameters based on the event time field, matches the time characteristics with the transaction type identifier according to the chronological order, organizes the time intervals between adjacent events, and generates event time interval quantities; The product status synchronization submodule synchronously collects financial product status tags and user profile attributes based on the event time interval, filters data records that are consistent with the behavioral event time, classifies and processes them according to the transaction type identifier and product status identifier, organizes the product status information corresponding to the transaction type, and generates transaction product status information; The timing chain generation submodule organizes the time characteristics of behavioral events under the transaction type according to the status of the transaction product, arranges the event records in chronological order, filters the data segments with time intervals less than the synchronization reference value, groups them into continuous behavior chains, and generates a cross-system behavior timing chain.

4. The DeepSeek-based financial marketing SAAS platform according to claim 3, characterized in that: The marketing strategy priority evaluation module includes: The transaction behavior extraction submodule obtains the cross-system behavior time series chain transaction data, detects transaction timestamps, types and amounts, collects continuous transaction sequences, classifies transaction features, filters high-frequency behaviors, and generates high-frequency transaction behavior distribution values; The product status assessment submodule uses the financial product status tag based on the high-frequency trading behavior distribution value to detect the current status and changes of the product, extract the activity range and change range, and obtain the urgency of the strategy demand; The priority sequence generation submodule calls the user risk preference label according to the urgency of the policy requirements, detects the risk level and tolerance range, sets the correction factor, adjusts the urgency measure, and sorts and generates the policy priority sequence.

5. The DeepSeek-based financial marketing SAAS platform according to claim 4, characterized in that: The window behavior analysis module includes: The policy classification submodule obtains the cross-system behavior time series chain and policy priority sequence, assigns the behavior to the corresponding policy category according to the priority, arranges the policy categories based on the time sequence and labels the behavior nodes, counts the number of occurrences of the policy categories in the overall time series chain, and generates a policy distribution value; The time window division submodule calls the strategy distribution value, determines the time distribution characteristics of the behavior node according to the density change of the strategy behavior in the time series chain, calibrates the time window boundary according to the change of the time interval of the behavior node, and generates the time window boundary value based on the accumulation of the strategy category; The behavior feature extraction submodule calls the time window demarcation value, collects the strategy category behavior within the time window, extracts the time fluctuation characteristics according to the behavior triggering time series, and generates a strategy behavior feature set based on the local behavior quantity change characteristics.

6. The DeepSeek-based financial marketing SAAS platform according to claim 5, characterized in that: The specific calculation formula for counting the number of occurrences of the strategy category in the overall time series chain is: ； in, represents the weighted occurrence number of strategy category j, Represents the global priority correction coefficient of the behavior node in the timing chain, represents the original priority of the k-th behavior node belonging to policy category j, represents the number of time units between the kth behavior node and the end point of the current timing chain, d represents the total number of behavior nodes belonging to strategy category j, To prevent the smoothing constant from being zero in the denominator, represents the collaborative weight factor of cross-system behavior nodes, represents the number of behavior nodes of the mth associated subsystem in strategy category j, and q represents the total number of associated subsystems.

7. The DeepSeek-based financial marketing SAAS platform according to claim 5, characterized in that: The policy deployment adaptation module includes: The density comparison submodule obtains the trigger density feature and the preset load threshold in the policy behavior feature set, calls the trigger density feature to compare with the preset load threshold, and filters the policy behavior features that exceed the load threshold interval according to the comparison result to generate the trigger density offset; The resource judgment submodule obtains the resource occupancy characteristics and the limit threshold in the policy behavior feature set based on the trigger density deviation, calls the resource occupancy characteristics and the limit threshold to perform difference judgment, filters the resource occupancy characteristics that exceed the limit threshold, and generates the resource load breakthrough rate; The solution generation submodule calls the policy behavior characteristics that exceed the load threshold according to the resource load breakthrough rate, divides the policy set according to the resource load breakthrough rate, calls the divided policy set to formulate deployment parameters and perform instruction mapping processing, and generates a policy deployment plan and an optimized instruction set.

8. The DeepSeek-based financial marketing SAAS platform according to claim 7, characterized in that: The specific calculation formula for comparing the call trigger density feature with the preset load threshold is: ； Among them, H represents the trigger density deviation trend parameter, ρ represents the average trigger density value of all samples in the strategy behavior feature set, and ρ i represents the density statistics of the i-th feature dimension in the strategy behavior feature set in the sample set, T represents the benchmark value of the preset load threshold, represents the load threshold correction corresponding to the i-th characteristic dimension, α represents the dynamic adjustment coefficient, μ represents the harmonic mean of the historical density mean and the current density peak, γ represents the density fluctuation tolerance factor, β i represents the nonlinear offset gain coefficient of the i-th feature dimension, η represents the threshold stability correction parameter, and n represents the total number of dimensions of the strategy behavior feature.

9. The DeepSeek-based financial marketing SAAS platform according to claim 1, characterized in that: The system further comprises: The resource dynamic mapping module dynamically allocates computing resource capacity characteristics and storage resource quota characteristics based on the task volume characteristics and optimization amplitude characteristics of the policy deployment scheme and the optimization instruction set, and generates a cross-system resource configuration table; The cross-system resource configuration table includes task volume characteristics, optimization range characteristics, computing resource capacity characteristics, and storage resource quota characteristics.

10. The DeepSeek-based financial marketing SAAS platform according to claim 9, characterized in that: The resource dynamic mapping module includes: The task feature extraction submodule obtains the task volume data and task category parameters in the strategy deployment plan and the optimization instruction set, calls the task volume data to determine the capacity demand value, calls the task category parameters to identify the attribute classification, selects high-demand task items based on the capacity demand value and the attribute classification, and generates the task capacity feature value; The optimization amplitude parsing submodule obtains the optimization target parameter and optimization priority parameter in the optimization instruction set based on the task capacity characteristic value, calls the optimization target parameter to determine the adjustment range value, calls the optimization priority parameter to arrange the application order, selects high-priority task items according to the adjustment range value and the application order, and generates the task adjustment amplitude characteristic value; The resource relationship generation submodule obtains the initial capacity parameters of computing resources and the initial quota parameters of storage resources according to the task adjustment amplitude characteristic value, calls the initial capacity parameters of computing resources to determine the allocation ratio, calls the initial quota parameters of storage resources to determine the adjustment ratio, establishes a mapping parameter set based on the allocation ratio and the adjustment ratio, and generates a cross-system resource configuration table.

Images