User information integrated service management system based on multi-dimensional portrait

Abstract

The present application relates to user portrait analysis, intelligent business scheduling and information processing technical field, specifically to user information comprehensive business management system based on multidimensional portrait, including: obtaining cold baseline vector, initial hot disturbance time series tensor, business interaction log data and business server cluster load rate data; extracting time interval change rate, and combining the dynamic occupancy rate of core business resources to construct the attenuation factor, updating the initial hot disturbance time series tensor to obtain the target hot disturbance time series tensor, calculating the vector distance of the target hot disturbance time series tensor and the cold baseline vector, generating business merging recalculation instruction based on the evaluation result; generating business distribution instruction based on the target hot disturbance time series tensor, and modifying the retention time limit parameter of the personalized business content of the front-end business contact; the present application realizes the balance between portrait timeliness and system resource occupation.

Description

Technical Field

[0001] This invention relates to the fields of user profile analysis, intelligent business scheduling and information processing technology, specifically a comprehensive user information business management system based on multi-dimensional profiles. Background Technology

[0002] As the scale of internet platform business continues to expand, users generate a large amount of continuously changing business interaction data during browsing, clicking, adding to cart, comparing prices, and placing orders. The platform needs to use this data to identify short-term changes in user intent in a timely manner and adjust the recommendation flow, business routing, and caching strategies of front-end business touchpoints accordingly. How to balance the timeliness of user profiles, the accuracy of business distribution, and the control of system resource consumption in high-concurrency scenarios has also become a technical problem facing the field of comprehensive business management.

[0003] Traditional user information management methods currently mainly rely on the following approaches: building static user profiles based on historical behavior, distributing services based on fixed rules, and updating content according to a unified cache duration;

[0004] However, static user profiles, fixed rule distribution, and unified cache duration all have certain drawbacks. For example, static user profiles are not responsive enough to instantaneous changes in user interests within the current session, resulting in delayed distribution results; fixed rule methods lack the ability to adjust the load status of the business server cluster, which can easily lead to excessively frequent profile updates or underlying writes under high concurrency; and unified cache duration methods are difficult to dynamically adjust the content refresh rate according to the strength of the user's real-time intent, making it difficult to balance personalized display effects with the overall system load. Summary of the Invention

[0005] The purpose of this invention is to provide a comprehensive user information management system based on multi-dimensional profiles, and to solve the following technical problems:

[0006] It avoids high-frequency write impacts on the underlying image storage in a high-concurrency environment, and can achieve a balance between image timeliness and system resource consumption while maintaining the timeliness of business distribution.

[0007] The objective of this invention can be achieved through the following technical solutions:

[0008] A comprehensive user information management system based on multi-dimensional profiling, the system comprising:

[0009] The acquisition module is used to acquire a cold baseline vector representing the historical business characteristics of the target user, an initial hot perturbation time series tensor representing the current business characteristics of the target user, the business interaction log data of the target user, and the load rate data of the business server cluster. The business interaction log data includes timestamps and business identifiers, and the load rate data includes the dynamic occupancy rate of core business resources.

[0010] The processing module is used to extract the time interval change rate based on the business interaction log data, construct an attenuation factor in combination with the dynamic occupancy rate of the core business resources, update the initial thermal disturbance time series tensor using the attenuation factor to obtain the target thermal disturbance time series tensor, calculate the vector distance between the target thermal disturbance time series tensor and the cold baseline vector, compare the vector distance with the merging threshold dynamically generated based on the dynamic occupancy rate of the core business resources to obtain the evaluation result, and generate a business merging recalculation instruction based on the evaluation result.

[0011] The output module is used to generate and output service distribution instructions based on the target thermal perturbation time series tensor, and to modify the retention time parameters of personalized service content of the front-end service touchpoints based on the service distribution instructions.

[0012] Optional processing modules include:

[0013] The rate of change calculation unit is used to calculate the time difference between adjacent business interaction log data within a sliding time window of a preset length, so as to obtain the rate of change of the time interval.

[0014] The attenuation factor construction unit is used to map the dynamic occupancy rate of the core business resources to the available business computing resource balance, and to perform a scaling operation on the available business computing resource balance according to the time interval change rate to obtain the attenuation factor.

[0015] The tensor update unit is used to query a preset business feature mapping table, obtain the feature vector corresponding to the business identifier of the business interaction log data as the initial feature, multiply the initial feature by the decay factor to obtain the offset weight, and add the offset weight to the initial thermal perturbation time series tensor through matrix addition to obtain the target thermal perturbation time series tensor.

[0016] Optionally, the processing module also includes:

[0017] The distance calculation unit is used to perform a convergence operation on the target thermal perturbation time series tensor in the time step dimension to obtain the convergence result, and calculate the Euclidean distance between the convergence result and the cold baseline vector as the vector distance;

[0018] The threshold generation unit is used to dynamically generate a merged threshold by multiplying the dynamic occupancy rate of the core business resources by a preset base multiplier.

[0019] The recalculation judgment unit is used to trigger the recalculation operation of the target thermal perturbation time series tensor and the cold baseline vector to generate the service recalculation instruction when the Euclidean distance is greater than the merging threshold, and to block the recalculation operation when the Euclidean distance is not greater than the merging threshold.

[0020] Optionally, the system may also include:

[0021] The feedback module is used to statistically analyze the business execution conversion rate after the business distribution instruction is executed in real time, and to determine its trend by calculating the difference between the business execution conversion rates within adjacent sliding time windows.

[0022] The threshold adjustment module is used to lower the base multiplier when the difference is continuously negative, indicating a downward trend, and to keep the base multiplier unchanged when the difference is not continuously negative, indicating no downward trend.

[0023] Optionally, the system may also include:

[0024] The anomaly monitoring module is used to monitor whether the target user is in a cold start state by statistically analyzing the amount of historical business interaction log data, and to monitor whether the timestamp is in a disordered state by verifying the monotonicity of the timestamp.

[0025] The degradation processing module is used to disable the calculation logic of the decay factor and degenerate into a recommendation mechanism based on static rules to output the service distribution instruction when the target user is in the cold start state or the timestamp disorder state; and to execute the operation of the processing module when the target user is not in the cold start state and not in the timestamp disorder state.

[0026] Optionally, the system may also include:

[0027] The storage module is used to store the cold baseline vector in a long-cycle service feature archive and to store the initial thermal perturbation time series tensor and the target thermal perturbation time series tensor in a short-term service intent instant access archive. The capacity of the long-cycle service feature archive is greater than the capacity of the short-term service intent instant access archive, and the read speed of the short-term service intent instant access archive is greater than the read speed of the long-cycle service feature archive.

[0028] Optionally, the acquisition module is further configured to acquire a multi-dimensional user profile of the target user and extract combined profile features, wherein the initial thermal perturbation time series tensor is composed of the combined profile features and time series features, and the output module includes:

[0029] The instruction generation unit is used to extract the combined profile features from the target thermal perturbation time series tensor and generate the service distribution instruction based on the combined profile features;

[0030] The instruction issuing unit is used to issue the service distribution instruction to the corresponding service traffic intelligent scheduling node for route distribution;

[0031] The cache modification unit is used to locate the target user according to the business distribution instruction and modify the retention time parameter of the personalized business content of the target user at the front-end business touchpoint.

[0032] Optional, the acquisition module includes:

[0033] The log consumption unit is used to consume the business interaction log data forwarded by the business gateway in real time through a message queue;

[0034] The monitoring and reading unit is used to read the load rate data of the business server cluster in real time through the monitoring agent.

[0035] The beneficial effects of this invention are:

[0036] 1. This invention introduces a hot and cold dual-state profiling mechanism, which constructs a decay factor by calculating the rate of change of behavior time interval and the dynamic occupancy rate of core business resources, and dynamically updates the thermal perturbation time series tensor. This method can quickly and accurately capture the user's short-term intention changes, and can also adjust the update amplitude according to the server load to prevent system overload under high concurrency.

[0037] 2. This invention establishes a dynamic merging logic constrained by both load and conversion rate; based on the vector distance between the hot tensor and the cold baseline, a dynamic merging threshold is generated in combination with the server load, and the threshold is adjusted in a closed loop according to the business conversion rate trend, which effectively avoids the high-frequency write impact caused by fixed rules and realizes the adaptive write-back frequency of the underlying layer.

[0038] 3. This invention enables fine-grained control of edge node caching strategies; the system generates business distribution instructions based on the latest hot profiles, and accordingly modifies the retention time parameters of personalized business content at front-end business touchpoints, adjusting the refresh rhythm as needed based on the strength of the user's real-time intent, breaking the rigid limitations brought about by uniform caching duration. Attached Figure Description

[0039] Other features, objects, and advantages of the invention will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings:

[0040] Figure 1 This is a schematic diagram of the modules of the user information integrated business management system based on multi-dimensional profile provided in the embodiments of this application. Detailed Implementation

[0041] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0042] Please see Figure 1 The user information integrated business management system based on multi-dimensional profiles includes: an acquisition module, which is used to acquire the cold baseline vector representing the historical business characteristics of the target user, the initial hot perturbation time series tensor representing the current business characteristics of the target user, the business interaction log data of the target user, and the load rate data of the business server cluster. The business interaction log data includes timestamps and business identifiers, and the load rate data includes the dynamic occupancy rate of core business resources.

[0043] The processing module is used to extract the time interval change rate based on business interaction log data, construct an attenuation factor in combination with the dynamic occupancy rate of core business resources, update the initial thermal disturbance time series tensor using the attenuation factor to obtain the target thermal disturbance time series tensor, calculate the vector distance between the target thermal disturbance time series tensor and the cold baseline vector, compare the vector distance with the merging threshold dynamically generated based on the dynamic occupancy rate of core business resources to obtain the evaluation result, and generate a business merging recalculation instruction based on the evaluation result.

[0044] The output module is used to generate and output service distribution instructions based on the target thermal perturbation time series tensor, as well as to modify the retention time parameters of personalized service content of the front-end service touchpoints based on the service distribution instructions.

[0045] This embodiment provides a specific implementation of a comprehensive user information business management system based on multi-dimensional profiles. Specifically, the system is deployed in a high-concurrency business interaction scenario. The platform needs to identify the user's instantaneous purchase intention in a timely manner with system resource consumption below a preset threshold during the user's continuous browsing, adding to cart, comparing prices and placing orders, and determine the homepage recommendation stream, event route entry and edge caching strategy accordingly.

[0046] Specifically, the system acquires the target user's cold baseline vector, initial thermal perturbation time series tensor, business interaction log data, and load rate data of the business server cluster;

[0047] The cold baseline vector is used to characterize the relatively stable business preferences over a long period, which can be understood as the steady-state profile of the user formed within a preset long-term time window; the initial hot perturbation time series tensor is used to characterize the short-term fluctuation characteristics within the current session, which can be understood as the dynamic offset profile of the user within the current preset short-term time window.

[0048] The data structure of the initial thermal perturbation time series tensor includes a time step dimension and a business feature dimension. It is composed of a two-dimensional matrix or a three-dimensional tensor formed by concatenating the feature vectors corresponding to multiple time slices in the current session in chronological order.

[0049] Business interaction log data should include at least a timestamp and a business identifier, which can be the entry number for businesses such as maternal and infant product fairs, cross-border e-commerce fairs, and digital limited-time sales. Load rate data should include at least the dynamic utilization rate of core business resources, used to describe the current load pressure on the business server cluster. During processing, the time interval change rate should be extracted from the business interaction log data first.

[0050] For ease of explanation, suppose the same user generates three log entries consecutively in a certain session, with timestamps of 10:00:00, 10:00:08 and 10:00:20, and corresponding time differences of 8 seconds and 12 seconds, respectively. The rate of change of the time interval during this period can be simply understood as the degree of change of the later time difference relative to the earlier time difference.

[0051] In a simplified calculation model, it can be directly used as a variable with clear positive and negative directions. For example, a value of 0.5 indicates that the user's operation pace has slowed down compared to the previous stage. If the dynamic utilization rate of the core business resources of the business cluster is 60% at the same time, the system combines the utilization rate with the aforementioned variable to generate a decay factor.

[0052] The specific combination method can be achieved by mapping the dynamic occupancy rate of core business resources to the system's idle capacity, and scaling the capacity proportionally according to the rate of change of time interval; the attenuation factor can be understood as: under the combined effect of changes in user rhythm and server load, how much offset should be applied to the current dynamic profile for this new behavior;

[0053] The system updates the initial thermal perturbation time series tensor using a decay factor. For ease of deduction, assuming the feature vector corresponding to a certain business identifier, such as the maternal and infant venue, is [0.8, 0.1, 0.1] and the decay factor is 0.5, the offset weights [0.4, 0.05, 0.05] can be obtained. If the original value of the slice corresponding to the current time in the initial thermal perturbation time series tensor is [0.2, 0.3, 0.1], then it is updated to [0.6, 0.35, 0.15]. The system continuously performs this type of superposition on each time slice to form the target thermal perturbation time series tensor.

[0054] The system calculates the vector distance between the target thermal disturbance time series tensor and the cold baseline vector, and determines whether to generate a service merging and recalculation instruction based on the distance evaluation result;

[0055] To illustrate with simplified data, if the cold baseline vector is [0.3, 0.3, 0.2], and the current hot aggregation result is [0.6, 0.35, 0.15], then there is a deviation between the two that is greater than the preset distance threshold.

[0056] If the deviation exceeds the preset threshold, it means that the user's current intent can no longer be explained by the historical steady-state profile alone, and the merging and recalculation of hot and cold profiles should be triggered; if it does not exceed the threshold, it means that short-term deviations can still be processed in the cache state, and there is no need to immediately penetrate to the underlying long-term profile storage.

[0057] The output module generates and outputs business distribution instructions based on the target thermal perturbation time series tensor. The business distribution instructions may include, but are not limited to: routing the target user to the maternal and infant special sale recommendation stream, improving the card ranking of a certain venue, and switching to a specific activity service cluster.

[0058] Meanwhile, the system modifies the retention time parameters of personalized business content at the front-end business touchpoints based on the distribution instruction; for example, for target users with strong purchase intent and continuous behavior, the cache lifespan of the activity page is reduced from 300 seconds to 60 seconds in order to refresh personalized content faster; for users with stable and unchanging behavior, the cache time is maintained at a value greater than the preset time threshold to reduce the pressure on edge nodes to return to the origin.

[0059] As an anomaly handling mechanism, if the cold baseline vector is found to be missing during the acquisition phase, the platform category average vector can be used as a temporary substitute, and then backfilled after the historical data is supplemented; if the initial thermal perturbation time series tensor does not exist, a zero tensor is created as the starting point of the current session.

[0060] If duplicate messages appear in the business interaction log, the combination of timestamp and business identifier is used as the idempotent key for deduplication. If the dynamic utilization rate of core business resources fails to be collected, the most recent valid sample value can be used. If multiple consecutive sampling periods fail, the merging and recalculation are paused, and only the hot distribution output is retained to avoid making incorrect writing decisions based on distorted load.

[0061] After the peak business period began, a user first browsed the maternal and infant diaper section, then entered the milk powder activity page 8 seconds later, and returned to the diaper section 12 seconds later and added it to the shopping cart; the platform read the user's long-term profile and found that he mainly preferred daily necessities in the past three months, but the frequency of maternal and infant related operations in the current session was greater than the preset frequency threshold.

[0062] The system thus generates strong thermal disturbances in the maternal and infant direction. Without immediately rewriting the long-term profile, the second screen of the homepage is switched to the maternal and infant promotion flow. At the same time, the caching time of the target user's activity page on the edge node is shortened, so that the relevant business combination information can appear faster in subsequent refreshes.

[0063] When this offset continues to accumulate and exceeds the merging threshold, a merging and recalculation instruction is generated to write the new purchasing tendency back into the long-term profile.

[0064] Specifically, the merge and recalculation operation involves extracting the converged result of the target thermal perturbation time series tensor as the thermal feature vector. And adopt the update step size coefficient set by the system. Among them, the step size coefficient is updated. The value is inversely proportional to the amount of historical business interaction log data of the target user and directly proportional to the number of time segments contained in the target hot perturbation time series tensor. The hot feature vector is weighted and fused into the original cold baseline vector. In this context, the specific update operators are summarized in the following formula:

[0065]

[0066] The system uses this formula to obtain the updated cold baseline vector. And complete the writing back of the underlying data for long-term profiling; among them, This represents the multiplication operation between a scalar and a vector. This represents vector addition.

[0067] The purpose of this step is to process long-term stable features and short-term session features in a layered manner, so that the system can maintain the timeliness of business distribution in a high-concurrency environment, while suppressing the impact of high-frequency writes to the underlying profile storage, thereby achieving a balance between profile timeliness and system resource consumption.

[0068] In a preferred embodiment of the present invention, the processing module includes: a rate of change calculation unit, used to calculate the time difference between adjacent business interaction log data within a sliding time window of a preset length, so as to obtain the time interval change rate; and an attenuation factor construction unit, used to map the dynamic occupancy rate of core business resources to the available business computing resource reserve, and to perform a proportional scaling operation on the available business computing resource reserve according to the time interval change rate to obtain the attenuation factor.

[0069] The tensor update unit is used to query a preset business feature mapping table, obtain the feature vector corresponding to the business identifier of the business interaction log data as the initial feature, multiply the initial feature by the decay factor to obtain the offset weight, and then add the offset weight to the initial thermal perturbation time series tensor through matrix addition to obtain the target thermal perturbation time series tensor.

[0070] This embodiment provides a mechanism for breaking down the processing process in detail; specifically, in the aforementioned overall solution, if only a general description is given of updating the hot profile based on logs, two defects are likely to occur in the engineering: first, the difference in the rhythm of behavior at different time periods cannot be quantified stably;

[0071] Secondly, the impact of server load on the intensity of image updates lacks a clear entry point, which may lead to the indiscriminate amplification of hot state offsets under high concurrency.

[0072] Therefore, this embodiment further introduces a rate of change calculation unit, a decay factor construction unit, and a tensor update unit to form an executable data stream;

[0073] Specifically, the rate of change calculation unit processes adjacent logs within a sliding time window of a preset length. Assuming the window length is 30 seconds, four logs enter the window in chronological order, with timestamps t1, t2, t3, and t4, corresponding to intervals of 4 seconds, 6 seconds, and 3 seconds. The unit first obtains the adjacent time difference sequence [4, 6, 3], and then further calculates the rate of change of the time interval.

[0074] To facilitate the explanation of the algorithm logic, a simplified approach can be adopted: divide the difference between adjacent time differences by the previous time difference to obtain the change rate sequence [0.5, -0.5]; where 0.5 indicates that the behavior rhythm slows down, and -0.5 indicates that the behavior rhythm speeds up again; in this way, the original timestamp is converted into a change quantity that can be directly used in the calculation.

[0075] If only the above changes are present, there is still a drawback: the system cannot perceive the current cluster pressure. To address this, the attenuation factor construction unit performs numerical calculations on the dynamic occupancy rate of core business resources and the rate of change of time intervals to obtain the attenuation factor.

[0076] Specifically, to avoid a lack of interpretability in the numerical calculation process, the system presets an adjustment coefficient. adjustment coefficient The scaling factor used to map the rate of change of time intervals to the [0,1] interval is determined by the historical highest load extreme value of the business server cluster, minus the dynamic occupancy rate of core business resources from 1. Characterize the available business computing resource reserves, combined with the time interval change rate. Based on the formula:

[0077]

[0078] in, This represents scalar subtraction. This represents a scalar multiplication operation; the attenuation factor is calculated. Assuming a preset adjustment coefficient The value is 0.66, representing the dynamic utilization rate of core business resources. When the system has 40% available margin, it can generate a responsive attenuation value greater than the preset first attenuation threshold. When the dynamic occupancy rate of core business resources rises to 85%, even if the rate of change of behavior is the same, an attenuation value less than the preset second attenuation threshold should be generated.

[0079] Using micro-data as an example, if the rate of change is -0.5, it indicates that the user's operation pace has accelerated, while the dynamic utilization rate of core business resources is 40%. Substituting these values ​​into the formula, the decay factor can be obtained. For ease of subsequent calculations, it can be approximated as 0.8;

[0080] If the dynamic utilization rate of core business resources is 85% at another moment, the decay factor can be calculated similarly. It can be approximated as 0.2; thus, the same user pace will produce different update intensities under low and high loads.

[0081] The system pre-builds or configures a business feature mapping table, which stores the mapping relationship between each business identifier and a feature vector of fixed dimensions. The system obtains the feature vector corresponding to the current business identifier in the business interaction log data by querying the business feature mapping table. Furthermore, the tensor update unit takes the feature vector corresponding to the business identifier as the initial feature and multiplies it with the decay factor to obtain the offset weight. Assuming that the feature vector corresponding to the milk powder activity page is [0.7, 0.2, 0.1], when the decay factor is 0.8, the offset weight is [0.56, 0.16, 0.08].

[0082] If the attenuation factor is 0.2, the offset weight becomes [0.14, 0.04, 0.02]; this offset weight is superimposed on the corresponding temporal position of the initial thermal perturbation temporal tensor by matrix addition or tensor slice addition;

[0083] Assuming the original slice is the latest row in [[0.1, 0.0, 0.0], [0.2, 0.1, 0.1]], the superimposed slice can be updated to [[0.1, 0.0, 0.0], [0.76, 0.26, 0.18]] or [[0.1, 0.0, 0.0], [0.34, 0.14, 0.12]], thus verifying the modulation effect of the load rate on the thermal offset amplitude;

[0084] As an exception handling mechanism, if there are fewer than two log entries within the sliding time window, the difference between adjacent time periods cannot be calculated. In this case, the change rate of the time interval can be set to zero, and only the preset basic weights of the initial features can be injected.

[0085] If the time difference between adjacent logs is zero, it indicates that there is replay logs or millisecond-level concurrent log backlog in a short period of time. In this case, deduplication can be performed first or the minimum protection amount can be added to the zero interval to prevent subsequent operation instability. If the dynamic utilization rate of core business resources is abnormally higher than 100% or lower than 0%, it is considered as an abnormal monitoring sampling. The sampling is discarded and rolled back to the most recent valid value.

[0086] If the business identifier does not find a corresponding vector in the feature mapping table, it can be mapped to other business backup vectors to avoid the entire update failing; 5 minutes before the peak of the promotion, a user entered three pages in a row: diapers, milk powder, and baby wipes. The time difference within the window was shortened from 7 seconds to 3 seconds, and then to 2 seconds.

[0087] Based on this, the system determines that the user's operation pace in the maternal and infant category has significantly accelerated. If the cluster load is still within a controllable range at this time, a higher decay factor is constructed to quickly amplify the hot offset in the maternal and infant direction, so that the subsequent recommendation stream can focus on relevant products more quickly. If the cluster has entered a high-pressure state at this time, a lower decay factor is constructed to retain the trend but reduce the write impulse, thereby prioritizing the overall service stability.

[0088] The purpose of this mechanism is to couple the rhythm of user behavior with the server's stress state in the same processing link and implement it through executable sliding windows, numerical calculations, and matrix overlay, thereby achieving fine control over the intensity of hot profile updates.

[0089] In a preferred embodiment of the present invention, the processing module further includes: a distance calculation unit, used to perform a convergence operation on the target thermal perturbation time series tensor in the time step dimension to obtain a convergence result, and calculate the Euclidean distance between the convergence result and the cold baseline vector as the vector distance; and a threshold generation unit, used to dynamically generate a merging threshold by multiplying the dynamic occupancy rate of core business resources with a preset base multiplier.

[0090] The recalculation judgment unit is used to trigger the recalculation operation of the target thermal perturbation time series tensor and the cold baseline vector to generate a service recalculation instruction when the Euclidean distance is greater than the merging threshold, and to block the recalculation operation when the Euclidean distance is not greater than the merging threshold.

[0091] This embodiment provides a mechanism for determining when hot and cold images should be merged. Specifically, while the continuous accumulation of thermal perturbation time-series tensors can improve instantaneous response speed, the lack of clear merging trigger rules will result in two opposite defects: first, merging is too frequent, causing the underlying storage to be rewritten at high frequency; second, merging does not occur for a long time, causing long-term images to gradually become distorted.

[0092] Therefore, this embodiment introduces a distance calculation unit, a threshold generation unit, and a recalculation judgment unit to form a dynamic merging logic subject to load constraints;

[0093] Specifically, the distance calculation unit calculates the Euclidean distance between the converged result of the target thermal perturbation time series tensor and the cold baseline vector;

[0094] Specifically, the convergence operation refers to performing mean pooling or weighted summation on the target thermal perturbation time series tensor along the time step dimension, reducing its dimension to one dimension, the same as the cold baseline vector, thus obtaining the dimension-aligned converged result;

[0095] For ease of explanation, assume the cold baseline vector is [0.3, 0.4, 0.3], while the current hot aggregation result is [0.7, 0.2, 0.1]. The differences between the two in the three feature dimensions are 0.4, -0.2, and -0.2, respectively. After squaring, summing, and taking the square root, a distance value can be obtained. There is no need to go into complex calculations here. It is only necessary to explain that the larger the distance value, the more difficult it is for the current session offset to be explained by the original steady-state profile, and the more necessary it is to trigger cold and hot merging.

[0096] However, if merging is performed immediately for any distance greater than the preset distance threshold, write jitter may still occur under high load. To address this, the threshold generation unit multiplies the dynamic utilization rate of core business resources with a preset base multiplier to generate a dynamic merging threshold.

[0097] When the dynamic utilization rate of core business resources is 30% and the preset base multiplier is 1.0, a dynamic merging threshold of 0.3, which is lower than the preset utilization threshold, can be obtained; when the dynamic utilization rate of core business resources rises to 90% and the base multiplier remains unchanged at 1.0, a dynamic merging threshold of 0.9, which is greater than the preset first attenuation threshold, can be obtained.

[0098] In this way, the system updates long-term profiles more actively when under low load, and retains hot results in memory or cache more restrainedly when under high load.

[0099] The recalculation judgment unit outputs an action based on the comparison result between the aforementioned distance value and the dynamic threshold. For example, if the Euclidean distance is 0.55 and the dynamic occupancy rate of core business resources is 30%, and the dynamic merging threshold is 0.3, then it is judged that the distance exceeds the threshold, that is, 0.55 is greater than 0.3, and a business merging recalculation instruction should be generated.

[0100] If the dynamic occupancy rate of core business resources is 90% and the dynamic merging threshold is 0.9, then the same distance value, i.e., 0.55, which is no greater than 0.9, will not trigger merging, but will block the current recalculation. The blocking here does not discard the user's current intent, but keeps it in the thermal perturbation time series tensor for real-time distribution, and merges it after the load falls back or the offset continues to accumulate.

[0101] As an anomaly handling mechanism, if the dimensions of the cold baseline vector and the hot aggregation result are inconsistent, dimension alignment is performed first, such as padding missing dimensions with zeros or rearranging them according to a unified mapping table before calculating the distance.

[0102] If the dynamic threshold is less than zero due to an incorrect configuration of the base multiplier, it should be immediately rolled back to the preset safety threshold to prevent the system from triggering a merge for all requests. If the same user triggers the merge judgment multiple times within a short period, a minimum recalculation interval can be set. Before the interval is reached, only the highest offset snapshot is recorded and the recalculation command is not issued repeatedly to avoid high-concurrency recalculation conflicts.

[0103] After the peak business period officially started, a user who originally preferred household cleaning products suddenly started searching for baby bottles, diapers, and baby food within the first 10 minutes, and quickly switched between multiple maternal and infant activity pages; at this time, the distance between the hot aggregation results and the original steady-state profile increased significantly.

[0104] If the business cluster load is low at that moment, the system will quickly issue a merge recalculation command to write the user's maternal and infant inclination into the long-term profile.

[0105] If it is during a period of high concurrency peak, the system will not directly write the aggregation results into the long-term database. Instead, it will only use the current hot results to drive the recommendation flow and routing strategy, and then complete the merging when the peak has passed.

[0106] The purpose of this mechanism is to make the merging of hot and cold profiles no longer dependent on a fixed period, but determined by the degree of user intent deviation and the system stress state, thereby achieving dynamic convergence of underlying write operations.

[0107] In a preferred embodiment of the present invention, the system further includes: a feedback module, used to statistically analyze the business execution conversion rate after the execution of the business distribution instruction in real time, and to determine its changing trend by calculating the difference between the business execution conversion rates within adjacent sliding time windows; and a threshold adjustment module, used to lower the base multiplier when the difference is continuously negative, indicating a downward trend, and to maintain the base multiplier unchanged when the difference is not continuously negative, indicating no downward trend.

[0108] This embodiment provides a mechanism for further closed-loop correction of dynamic thresholds. Specifically, although the aforementioned dynamic thresholds have taken into account server load, there is still a drawback: if the base multiplier is fixed for a long time, the system may be too conservative in certain business stages, causing hot offsets to be delayed in being written back to the long-term profile, resulting in a continuous decline in subsequent distribution effects.

[0109] Therefore, this embodiment introduces a feedback module and a threshold adjustment module, using the business execution conversion rate as a feedback signal to adaptively correct the base multiplier;

[0110] Specifically, the feedback module provides real-time statistics on the business execution conversion rate after the business distribution instructions are executed; this conversion rate can be defined as click-through rate, add-to-cart rate, order rate, or task completion rate, depending on the specific business.

[0111] In this e-commerce mainline, the combined conversion rate of clicking to enter the activity page and generating add-to-cart should be prioritized; assuming the system uses a 5-minute sliding time window, the conversion rates obtained from three consecutive windows are 4.2%, 3.9%, and 3.5% respectively, then the difference between adjacent windows is -0.3% and -0.4% respectively, showing a continuous negative trend, indicating that the current distribution effect is showing a continuous downward trend;

[0112] If we only observe the continuous downward trend without adjusting the threshold, the system may still maintain its original conservative strategy and fail to correct the deviation in time. Therefore, the threshold adjustment module lowers the base multiplier when the difference is continuously negative.

[0113] After the base multiplier is lowered, the dynamic merging threshold, obtained by multiplying the dynamic utilization rate of core business resources by the multiplier, will also decrease accordingly, thereby increasing the probability of triggering the recalculation of hot and cold profile merging.

[0114] For example, if the original base multiplier is 1.0 and the dynamic utilization rate of core business resources is 50%, the dynamic threshold is 0.5. After the conversion rate drops for three consecutive windows, the base multiplier is lowered to 0.8, and the dynamic threshold becomes 0.4. Under the same user offset, more users will be included in the merged recalculation scope, and the long-term profile can catch up with the actual interest changes more quickly.

[0115] Conversely, if the difference between adjacent windows is not consistently negative, for example, if the conversion rates of the three windows are 4.1%, 4.2%, and 4.0%, then although there is a single fluctuation, it does not constitute a continuous downward trend. In this case, the base multiplier is kept unchanged to avoid the system frequently changing the global threshold due to short-term noise, which would cause oscillations in the recalculation frequency.

[0116] As an anomaly handling mechanism, if the sample size of a window is too low, for example, if there are only a few exposures or clicks in the window, the statistical confidence of the conversion rate of the window is insufficient. In this case, the trend judgment can be skipped and the base multiplier of the previous period can be used. If there is no effective conversion data in multiple consecutive windows, the base multiplier can be frozen and only the observation log can be recorded.

[0117] If a change in metrics occurs due to a change in activity, such as switching from order rate to click-through rate, the trend sequence should be restarted at the point of the metric change to avoid miscalculations caused by mixing different indicators. In the first 15 minutes after the opening of a specific business peak event, the system distributed a batch of target users to the diaper special event, but the feedback module found that the conversion rate of these users adding to their cart dropped from 4.2% to 3.5%.

[0118] Analysis revealed that some users' immediate intent had shifted from low-priced consumables to bundled milk powder purchases, while their long-term profiles had not yet been synchronized. As a result, the system automatically lowered the base multiplier, causing more users with such hot-state shifts to trigger a merge and recalculation. After the next round of correction, the recommendation cards seen by the relevant users on the homepage were switched to specific business recommendation streams, and the business execution conversion rate rebounded.

[0119] The purpose of this mechanism is to reverse the distribution effect on the threshold generation process, forming a closed-loop adjustment of output result - threshold correction - re-output, thereby achieving adaptive matching between the portrait merging frequency and business effect.

[0120] In a preferred embodiment of the present invention, the system further includes: an anomaly monitoring module, used to monitor whether the target user is in a cold start state by statistically analyzing the amount of historical business interaction log data, and to monitor whether the target user is in a timestamp disorder state by verifying the monotonicity of the timestamp; and a degradation processing module, used to disable the calculation logic of the decay factor and degenerate into a recommendation mechanism based on static rules to output business distribution instructions when the target user is in a cold start state or a timestamp disorder state, and to execute the operation of the processing module when the target user is not in a cold start state and is not in a timestamp disorder state.

[0121] This embodiment provides a degradation processing mechanism for abnormal scenarios. Specifically, the aforementioned scheme relies on the collaborative calculation of decay factors based on historical logs, current log rhythm, and load information, but it will expose obvious defects in two types of scenarios: one is that newly registered users have almost no historical behavior, making it impossible to form a reliable cold baseline.

[0122] Another type is when the terminal disconnects and reconnects or multiple terminals report concurrently, causing the timestamps to become out of order, which distorts the rate of change of the time interval. If hot updates and merging judgments are still forcibly performed, the errors will be amplified. Therefore, this embodiment introduces an anomaly monitoring module and a degradation processing module.

[0123] Specifically, the anomaly monitoring module determines whether the target user is in a cold start state by statistically analyzing the amount of historical business interaction log data.

[0124] To facilitate execution, a minimum sample threshold can be set, for example, if there are fewer than 5 valid logs in the last 7 days, it is considered a cold start; here, valid logs can be limited to logs that have been deduplicated, have valid timestamps, and are recognizable by business identifiers; if a user only visits the page 1 to 2 times, the cold baseline vector is easily affected by accidental behavior and should not be used as a stable reference.

[0125] The anomaly monitoring module verifies the monotonicity of timestamps to identify timestamp disorder. If three logs received in the same session have timestamps of 10:00:05, 10:00:02, and 10:00:09 respectively, then the second log is reversed and the normal monotonically increasing relationship has been disrupted.

[0126] At this point, the rate of change calculated based on the difference between adjacent time intervals will lose its physical meaning and may produce negative intervals or abnormal sudden increases in value. Therefore, it should be regarded as a timestamp disorder.

[0127] When the degradation processing module detects a cold start state or a timestamp disorder state, it disables the calculation logic of the decay factor and degenerates into a recommendation mechanism based on static rules to output business distribution instructions.

[0128] The static rules here can be category hot lists, regional best-selling lists, demographic rules, event main venue priority rules, etc.; for example, for newly registered users from East China, priority is given to distributing to the new customer zone with high conversion rates in the current site; for active users with disordered timestamps, their hot profiles are temporarily frozen, and distribution is carried out according to the most recent stable cold baseline plus the platform's best-selling weight.

[0129] If the abnormal monitoring results indicate that the user is neither in a cold start state nor has a timestamp error, then the normal processing chain is resumed, including extraction of time interval change rate, construction of decay factor, update of thermal perturbation tensor and subsequent merging judgment;

[0130] As an exception handling mechanism, if a user has just exceeded the cold start threshold, for example, from 4 valid logs to 5 valid logs, a transition window can be set to refer to both static rules and light hot signals in the first few minutes to avoid jitter due to insufficient samples when switching to normal mode.

[0131] If timestamp disorder occurs only once and the disorder magnitude is less than the preset disorder tolerance threshold, it is possible to try rearranging by time before resuming use, rather than immediately downgrading the entire system; if cold start and timestamp disorder occur simultaneously, the more stringent static rule distribution will be executed first, and any merging and recalculation within this period will be prohibited.

[0132] During the promotional period, a newly registered user entered the platform for the first time and only clicked on the maternal and infant advertisement on the homepage once, with insufficient historical data. The system recognized that the user was in a cold start state, so it did not calculate the thermal disturbance attenuation, but instead prioritized the user to the initial default business cluster and the global high-priority business zone.

[0133] Another long-time user repeatedly opened the activity page in the weak network environment of the subway, and the logs were reported in reverse order. The system detected that the timestamp was no longer monotonically increasing, so it suspended the current hot update for this user and switched to using the most recent stable profile to output recommendations. The user will be reconnected to the normal link after the logs return to normal.

[0134] The purpose of this mechanism is to provide a clear exit point for the profile update process that relies on time series and historical samples, so as to prevent the spread of abnormal data, prevent the writing of erroneous profiles, and ensure that business distribution is not interrupted.

[0135] In a preferred embodiment of the present invention, the system further includes: a storage module, used to store the cold baseline vector in a long-cycle service feature archive and to store the initial thermal perturbation time series tensor and the target thermal perturbation time series tensor in a short-term service intent instant access archive, wherein the capacity of the long-cycle service feature archive is greater than the capacity of the short-term service intent instant access archive, and the read speed of the short-term service intent instant access archive is greater than the read speed of the long-cycle service feature archive.

[0136] This embodiment provides a cold and hot tiered storage mechanism. Specifically, in the aforementioned scheme, if the cold baseline vector and the hot perturbation time series tensor are stored in the same type of database, an engineering bottleneck will occur: long-term profiles require large storage capacity and persistent preservation, while hot time series data requires high-frequency read / write and fast response, and the storage requirements of the two conflict with each other. Therefore, this embodiment decouples the two through a storage module.

[0137] Specifically, the cold baseline vector is stored in a long-term business feature archive. The capacity of the long-term business feature archive is larger than that of the short-term business intent instant access archive, making it suitable for carrying long-term profiles of massive users. It can be a wide-table database, a distributed columnar database, or other large-capacity databases with a focus on persistence.

[0138] Thermal disturbance-related data, including the initial thermal disturbance time series tensor and the updated target thermal disturbance time series tensor, are stored in the short-term business intent instant access library. The read speed of the short-term business intent instant access library is higher than that of the long-term business feature archive library, making it more suitable as a data carrier for session state and near real-time state, such as high-frequency instant business response support components.

[0139] For ease of explanation, assume a platform has 100 million users; the cold baseline vector is stored in a long-term business feature archive according to user ID, and a single read may take tens of milliseconds, but it can support the target level of storage capacity; the hot perturbation time series tensor only retains data of the most recent windows of the current active users, and is stored in a short-term business intent real-time access library, and a single read can be compressed to within a few milliseconds.

[0140] In this way, during real-time distribution, the system prioritizes reading the hot results in the short-term business intent immediate access library to quickly determine the current recommendation flow; and only accesses the cold baseline vector in the long-term business feature archive library when initialization or hot-cold merging is required.

[0141] This hierarchical scheme is also reflected in the update path; when the target thermal perturbation time series tensor arrives for each session behavior, it is first written to the short-term business intent instant access library to handle high-frequency fluctuations; only after the recalculation judgment is passed will the merged long-term result be written back to the long-term business feature archive library.

[0142] To give a micro example, if the same user makes 20 browsing actions within 10 minutes, the short-term business intent instant access database may be updated 20 times, while the long-term business feature archive database may only be updated once or even 0 times in these 20 actions; thus, write penetration to large-capacity databases can be significantly reduced.

[0143] As an exception handling mechanism, if the hot perturbation time series tensor of a user in the short-term business intent immediate access library does not exist due to expiration, the system can read its cold baseline vector from the long-term business feature archive library and reconstruct the initial hot perturbation time series tensor with all zeros or the default template.

[0144] If the long-cycle business feature archive is temporarily unavailable, the system can continue to perform short-term distribution based on the existing hot results in the short-term business intent instant access library, but it is prohibited to initiate new cold and hot merge write-back; if the memory pressure of the short-term business intent instant access library is too high, it can be tiered and eliminated according to the user's most recent active time or tensor update time, and the elimination event is recorded in the compensation queue for subsequent idle period recovery.

[0145] After the peak of high-concurrency business begins, a massive number of users refresh the activity page simultaneously. The platform's recommendation decision for active users is mainly based on quickly retrieving their current hot perturbation time series tensor from the short-term business intent instant storage library, and completing the routing and content switching in milliseconds. Meanwhile, their long-term preferences, historical consumption capacity and other steady-state information are still stored in the long-term business feature archive library, and are only supplemented or recalculated and written back when necessary.

[0146] In this way, even if a surge in read and write requests occurs during peak sales periods, the system will not experience access congestion because all profiles are concentrated on reading and writing to the same database.

[0147] The purpose of this mechanism is to ensure that long-term storage and real-time access are properly positioned, thereby achieving a structural balance between capacity, speed, and write pressure.

[0148] In a preferred embodiment of the present invention, the acquisition module is further configured to acquire a multi-dimensional user profile of the target user and extract combined profile features. The initial thermal perturbation time series tensor is composed of a combination of combined profile features and time series features. The output module includes: an instruction generation unit, configured to extract the combined profile features from the target thermal perturbation time series tensor and generate a service distribution instruction based on the combined profile features.

[0149] The instruction issuing unit is used to issue business distribution instructions to the corresponding business traffic intelligent scheduling nodes for routing distribution; the cache modification unit is used to locate the target user according to the business distribution instructions and modify the retention time parameters of the target user's personalized business content at the front-end business touchpoint.

[0150] This embodiment provides a profile feature construction and instruction execution mechanism for the output link; specifically, in the aforementioned scheme, if the thermal perturbation time series tensor consists of only a single behavior sequence, although it can reflect changes in behavior strength, it still has a drawback:

[0151] The system has difficulty directly explaining why a certain route is assigned to a user, and it is also not conducive to the use of unified features by downstream business traffic intelligent scheduling nodes and caching systems; therefore, this embodiment further introduces multi-dimensional user profiles and combined profile features to make output instructions easier for business systems to consume.

[0152] Specifically, in addition to the original cold baseline vector and log data, the acquisition phase also acquires multi-dimensional user profiles of the target users and extracts combined profile features from them; the multi-dimensional user profiles may include user region, membership level, price sensitivity, historical preference categories, active time period, terminal type, etc.

[0153] The so-called combined profile features are not simply a combination of elements, but rather business-executable tags formed by combining multiple dimensions; for example, East China region + high frequency of maternal and infant purchases + high frequency of specific business interactions + high confidence preference + nighttime activity, etc.

[0154] The initial thermal perturbation time series tensor is composed of these combined profile features and time series features. Therefore, the tensor contains information about who the user is, as well as information about what the user was just doing and how the rhythm changed.

[0155] During the output phase, the instruction generation unit extracts combined profile features from the target thermal perturbation time series tensor and generates service distribution instructions accordingly.

[0156] For ease of deduction, suppose a user's combined profile features are currently encoded as [strong maternal and infant focus, price sensitive, active at noon], and their time-series features show that they have continuously visited the pages for milk powder, diapers, and complementary foods within the last 2 minutes. Then, the generated business distribution instruction can contain two parts: one is the routing direction, such as entering the maternal and infant discount stream; the other is the content strategy, such as prioritizing the return of discount coupons and bundled purchase product cards.

[0157] If the generated instruction is not explicitly distributed through a distribution link, downstream systems may still deviate in execution due to inconsistent interpretation. Therefore, the instruction distribution unit sends the service distribution instruction to the corresponding intelligent service traffic scheduling node for routing distribution. Here, the intelligent service traffic scheduling node can be a traffic scheduling gateway, a recommendation flow gateway, or a service orchestration node.

[0158] When distributing traffic, the target user identifier, routing target, priority, and validity period can be included. For example, for high-intent maternal and infant users, their request traffic can be routed to the maternal and infant promotion service cluster; for general browsing users, it can still be routed to the general recommendation cluster.

[0159] Furthermore, the cache modification unit locates the target user based on the business distribution instruction and modifies the retention time parameters of the personalized business content on the front-end business touchpoint; the modification here is not a unified site-wide adjustment, but rather a fine-grained control of edge cache items for the target user.

[0160] For example, if a user is identified as a high-conversion-intent user, the cache time for their activity page or recommendation stream is reduced from 180 seconds to 30 seconds to reflect real-time offers and inventory changes more quickly; if user behavior does not fluctuate much, the cache time is maintained at a value greater than the preset time threshold to reduce edge-to-origin requests.

[0161] As an anomaly handling mechanism, if some dimensions are missing in the multidimensional profile, such as new users not having membership level or price sensitivity data, unknown placeholder dimensions can be used to participate in the combination to avoid interruption of the combined feature construction; if the combined profile features conflict with the time-series features, such as long-term digital preference but current short-term maternal and infant behavior is abnormally strong, the hot window weight can be increased to reflect the current session intent first.

[0162] If the intelligent scheduling node for business traffic is temporarily unreachable, the business distribution instruction can be temporarily stored in the message queue for retry, and the valid route from the previous round can be used during the retry. If the edge node does not support user-level cache modification, it can degenerate into adjusting the cache lifetime at the meeting venue level or user group level.

[0163] During the peak business hours at noon, a high-level member from East China continuously browsed the mother and baby section for milk powder, diapers, and baby wipes, while their historical profile showed that they were sensitive to coupons.

[0164] The system constructs a profile of the user with combined characteristics of East China region, strong preference for maternal and infant products, price sensitivity, and midday activity, and superimposes the current access rhythm to form a hot perturbation time series tensor; after the instruction is generated, the platform routes the user to the maternal and infant promotion recommendation cluster, and at the same time shortens the edge caching time so that the user can see the changes in discount coupons, bundled purchases and limited-time inventory in a timely manner on the next refresh.

[0165] The purpose of this mechanism is to transform abstract profiling results into routable, deployable, and cacheable business instructions, thereby enabling direct linkage between profiling decisions and network-side execution actions.

[0166] In a preferred embodiment of the present invention, the acquisition module includes: a log consumption unit, used to consume business interaction log data forwarded by the business gateway in real time through a message queue; and a monitoring and reading unit, used to read the load rate data of the business server cluster in real time through a monitoring agent.

[0167] This embodiment provides a real-time data acquisition mechanism oriented towards data entry points; specifically, in the aforementioned scheme, if only the acquisition of logs and load data is abstractly described, a problem will still be exposed in actual deployment:

[0168] Behavior logs and cluster loads typically come from different systems, with inconsistent sampling timings and significant differences in throughput. Without a dedicated collection unit, situations can easily arise where behavior arrives but load is missing, or load updates are timely but behavior is delayed. Therefore, this embodiment further breaks down the acquisition process into a log consumption unit and a monitoring and reading unit.

[0169] Specifically, the log consumption unit consumes business interaction log data forwarded by the business gateway in real time through a message queue. The business gateway is located at the user request entry point and can write logs containing user identifiers, timestamps, business identifiers, etc., to the message queue when a user performs actions such as visiting the homepage, entering a venue, clicking on a product, adding it to the cart, or submitting an order. The log consumption unit reads these messages in partition order and can perform intra-stream aggregation based on the user identifier.

[0170] For ease of sandboxing explanation, assume the message queue receives three messages within a certain second: the first message is at 10:00:01, from user A, in the maternal and infant section; the second message is at 10:00:05, from user A, in the formula details page; and the third message is at 10:00:09, from user A, in the diaper details page.

[0171] The log consumption unit sends the logs to the subsequent sliding window for processing in sequence; if there are only behavior logs, the system still cannot adjust the merging strategy under load fluctuations; therefore, the monitoring and reading unit reads the load rate data of the business server cluster in real time through the monitoring agent.

[0172] The load rate data here includes not only the dynamic utilization rate of core business resources, but also the load rate of business state cache. The dynamic utilization rate of core business resources reflects the instantaneous computing pressure, while the load rate of business state cache is more directly related to the risk level of state-based stream processing.

[0173] For example, if the dynamic utilization rate of core business resources is 55% and the business state cache load rate is 48%, it means that the cluster still has good margin; if the dynamic utilization rate of core business resources is only 60%, but the business state cache load rate has risen to 88%, it indicates that there is obvious state accumulation. At this time, even if the core computing nodes have not yet reached the maximum full load threshold, it is advisable to be cautious about triggering new merge recalculation or large-scale tensor expansion.

[0174] In one specific implementation, the log consumption unit and the monitoring reading unit can use different sampling rhythms and then perform time alignment at the processing node; for example, business logs are entered one by one in an event-driven manner, while load data is pulled at a period of 1 second or 5 seconds; the processing node can attach the most recent effective load snapshot to each behavior log, so that each hot update has a load context adjacent to its time.

[0175] As an exception handling mechanism, if there is a backlog in the message queue, the log consumption unit can prioritize consuming messages corresponding to high-priority business identifiers, or perform tiered processing based on the user's recent activity level; if the same log is delivered repeatedly, idempotent filtering is performed based on the message's unique number and timestamp.

[0176] If the monitoring agent loses connection for a short time, resulting in the loss of load data, the most recent valid snapshot will be used and the sampling aging degree will be marked. After the preset aging time is exceeded, the new round of merging judgment will be stopped. If there is a significant contradiction between the dynamic utilization rate of core business resources and the business status cache load rate, such as the core computing node being lower than the preset utilization rate threshold but the memory being higher than the preset utilization rate threshold, the recalculation frequency can be constrained by the business status cache load rate to prevent state failures.

[0177] Within the first 30 seconds after the core business summit venue opens, the business gateway continuously writes massive amounts of user clicks and venue switching logs to the message queue. The log consumption unit reads in real time the continuous actions of a target user from the homepage to the maternal and infant venue, then to the milk powder details page, and then to the coupon entry.

[0178] At the same time, the monitoring agent sends back the dynamic utilization rate of the core business resources and the business state cache load rate of the current service cluster every second. After aligning these two data sources, the processing node finds that although the core computing node is only at 58%, the business state cache load rate is close to 85%. Therefore, the subsequent updates for the user's hot disturbance can continue to be executed, but the merge recalculation trigger is controlled more conservatively to avoid putting pressure on the state system and the underlying storage at the same time during peak periods.

[0179] The purpose of this mechanism is to provide a stable, real-time and alignable data entry point for subsequent profile calculations, thereby enabling synchronous perception of behavioral and system-side signals.

[0180] The foregoing has provided a detailed description of one embodiment of the present invention, but this description is merely a preferred embodiment and should not be construed as limiting the scope of the invention. All equivalent variations and modifications made within the scope of the claims of this invention should still fall within the patent coverage of this invention.

Claims

1. A comprehensive user information business management system based on multi-dimensional profiles, characterized in that: The system includes: The acquisition module is used to acquire a cold baseline vector representing the historical business characteristics of the target user, an initial hot perturbation time series tensor representing the current business characteristics of the target user, the business interaction log data of the target user, and the load rate data of the business server cluster. The business interaction log data includes timestamps and business identifiers, and the load rate data includes the dynamic occupancy rate of core business resources. The processing module is used to extract the time interval change rate based on the business interaction log data, construct an attenuation factor in combination with the dynamic occupancy rate of the core business resources, update the initial thermal disturbance time series tensor using the attenuation factor to obtain the target thermal disturbance time series tensor, calculate the vector distance between the target thermal disturbance time series tensor and the cold baseline vector, compare the vector distance with the merging threshold dynamically generated based on the dynamic occupancy rate of the core business resources to obtain the evaluation result, and generate a business merging recalculation instruction based on the evaluation result. The output module is used to generate and output service distribution instructions based on the target thermal perturbation time series tensor, and to modify the retention time parameters of personalized service content of the front-end service touchpoints based on the service distribution instructions.

2. The user information integrated business management system based on multi-dimensional profiling according to claim 1, characterized in that, The processing module includes: The rate of change calculation unit is used to calculate the time difference between adjacent business interaction log data within a sliding time window of a preset length, so as to obtain the rate of change of the time interval. The attenuation factor construction unit is used to map the dynamic occupancy rate of the core business resources to the available business computing resource balance, and to perform a scaling operation on the available business computing resource balance according to the time interval change rate to obtain the attenuation factor. The tensor update unit is used to query a preset business feature mapping table, obtain the feature vector corresponding to the business identifier of the business interaction log data as the initial feature, multiply the initial feature by the decay factor to obtain the offset weight, and add the offset weight to the initial thermal perturbation time series tensor through matrix addition to obtain the target thermal perturbation time series tensor.

3. The user information integrated business management system based on multi-dimensional profiling according to claim 2, characterized in that, The processing module further includes: The distance calculation unit is used to perform a convergence operation on the target thermal perturbation time series tensor in the time step dimension to obtain the convergence result, and calculate the Euclidean distance between the convergence result and the cold baseline vector as the vector distance; The threshold generation unit is used to dynamically generate a merged threshold by multiplying the dynamic occupancy rate of the core business resources by a preset base multiplier. The recalculation judgment unit is used to trigger the recalculation operation of the target thermal perturbation time series tensor and the cold baseline vector to generate the service recalculation instruction when the Euclidean distance is greater than the merging threshold, and to block the recalculation operation when the Euclidean distance is not greater than the merging threshold.

4. The user information integrated business management system based on multi-dimensional profiling according to claim 3, characterized in that, The system also includes: The feedback module is used to statistically analyze the business execution conversion rate after the business distribution instruction is executed in real time, and to determine its trend by calculating the difference between the business execution conversion rates within adjacent sliding time windows. The threshold adjustment module is used to lower the base multiplier when the difference is continuously negative, indicating a downward trend, and to keep the base multiplier unchanged when the difference is not continuously negative, indicating no downward trend.

5. The user information integrated business management system based on multi-dimensional profiling according to claim 1, characterized in that, The system also includes: The anomaly monitoring module is used to monitor whether the target user is in a cold start state by statistically analyzing the amount of historical business interaction log data, and to monitor whether the timestamp is in a disordered state by verifying the monotonicity of the timestamp. The degradation processing module is used to disable the calculation logic of the decay factor and degenerate into a recommendation mechanism based on static rules to output the service distribution instruction when the target user is in the cold start state or the timestamp disorder state; and to execute the operation of the processing module when the target user is not in the cold start state and not in the timestamp disorder state.

6. The user information integrated business management system based on multi-dimensional profiling according to claim 1, characterized in that, The system also includes: The storage module is used to store the cold baseline vector in a long-cycle service feature archive and to store the initial thermal perturbation time series tensor and the target thermal perturbation time series tensor in a short-term service intent instant access archive. The capacity of the long-cycle service feature archive is greater than the capacity of the short-term service intent instant access archive, and the read speed of the short-term service intent instant access archive is greater than the read speed of the long-cycle service feature archive.

7. The user information integrated business management system based on multi-dimensional profiling according to claim 1, characterized in that, The acquisition module is further configured to acquire a multi-dimensional user profile of the target user and extract combined profile features. The initial thermal perturbation time series tensor is composed of the combined profile features and the time series features. The output module includes: The instruction generation unit is used to extract the combined profile features from the target thermal perturbation time series tensor and generate the service distribution instruction based on the combined profile features; The instruction issuing unit is used to issue the service distribution instruction to the corresponding service traffic intelligent scheduling node for route distribution; The cache modification unit is used to locate the target user according to the business distribution instruction and modify the retention time parameter of the personalized business content of the target user at the front-end business touchpoint.

8. The user information integrated business management system based on multi-dimensional profiling according to claim 1, characterized in that, The acquisition module includes: The log consumption unit is used to consume the business interaction log data forwarded by the business gateway in real time through a message queue; The monitoring and reading unit is used to read the load rate data of the business server cluster in real time through the monitoring agent.

Images