Intention degree hierarchical management system based on online and offline fusion behavior track

Abstract

The present application relates to the technical fields of data processing and intelligent analysis, in particular to an intention degree grading management system based on online and offline fusion behavior track, comprising: collecting online interaction event data, terminal motion data, offline behavior event data and visual step frequency data of the same user to form online behavior track and offline behavior track; determining candidate target object based on the mapping relationship between offline area identification and online target object identification; generating terminal anchor point sequence and visual anchor point sequence according to terminal motion data and visual step frequency data, and determining and updating time difference to calibrate the time of online behavior track; calculating intention degree score based on calibrated online behavior track, offline behavior track and candidate target object, and determining intention level according to the score change of current evaluation period and previous evaluation period, and outputting corresponding service action instruction; the present application improves the reliability of intention recognition.

Description

Technical Field

[0001] This invention relates to the field of data processing and intelligent analysis technology, specifically to an intention-based hierarchical management system based on online and offline integrated behavioral trajectories. Background Technology

[0002] Currently, online-offline integrated behavior analysis solutions typically process online interaction records and offline behavior records independently, or only roughly bind the two types of events based on a wide time window. When users interact with objects offline while simultaneously viewing, saving, or comparing them online, existing solutions struggle to accurately match the online and offline behaviors of the same user towards the same target object due to factors such as heterogeneous clock deviations, unclear object associations, and the lack of a unified time benchmark for multi-source behavioral evidence. This can easily lead to candidate object drift, inaccurate identification of synchronous behaviors, interference with scoring by outdated behaviors, and invalid dwell times being misjudged as valid intentions, thus reducing the accuracy of intention level management and service action output. Summary of the Invention

[0003] The purpose of this invention is to provide an intention-level hierarchical management system based on online and offline integrated behavioral trajectories, avoiding misjudgments of intention caused by heterogeneous clock deviations and unclear object associations, and more accurately corresponding online and offline behaviors of the same target, thereby improving the accuracy of intention-level hierarchical management. Specifically, the technical solution of this invention includes:

[0004] Collect online interaction event data, terminal motion data, offline behavior event data and visual cadence data of the same user. Among them, multiple online interaction event data are sorted by recording time to form an online behavior trajectory, and multiple offline behavior event data are sorted by event time to form an offline behavior trajectory.

[0005] Based on the area identifiers in offline behavioral event data, and according to the preset mapping relationship between offline area identifiers and online target object identifiers, candidate target objects are determined; terminal anchor point sequences and visual anchor point sequences are generated based on terminal motion data and visual cadence data;

[0006] The update time difference is determined based on the terminal anchor sequence and the visual anchor sequence, and the online behavior trajectory is time-calibrated based on the update time difference to obtain the calibrated online behavior trajectory.

[0007] Based on the calibrated online and offline behavioral trajectories and candidate target objects, determine the user's intention score for the candidate target objects within the preset current evaluation period;

[0008] The intention level is determined based on the intention score for the current assessment period and the change in the intention score compared to the previous assessment period, and corresponding service action instructions are output.

[0009] Preferably, online interactive event data includes user identifier, target object identifier, action type, recording time, and action weight; offline behavioral event data includes user identifier, area identifier, offline action type, event time, and action confidence; terminal motion data includes sampling time and vertical acceleration component; and visual gait frequency data includes visual sampling time and human vertical displacement change value.

[0010] Preferably, the mapping relationship between offline area identifiers and online target object identifiers is queried to determine candidate target objects;

[0011] The target object mapping relationship includes the mapping relationship between product code, electronic price tag code, or shelf area code;

[0012] When no corresponding target object mapping relationship is found, perform one of the following operations: assign a default location based on the region identifier;

[0013] Perform downgraded matching based on the higher-level region; generate mapping failure records containing region identifier, timestamp, and miss status marker.

[0014] Preferably, the anchor point generation module is used to determine the terminal anchor point sequence and the visual anchor point sequence, including: dividing the terminal motion data and visual cadence data into multiple analysis windows:

[0015] According to the preset gait frequency band, the vertical acceleration component in the terminal motion data is bandpass processed, and the terminal peak extraction threshold is determined based on the statistical characteristics of the mean, standard deviation and peak distribution within the current analysis window to generate the terminal anchor point sequence;

[0016] Based on the preset gait frequency band, the visual gait frequency data is smoothed and bandpass processed, and the visual peak extraction threshold is determined based on the statistical characteristics of the mean, standard deviation and peak distribution within the current analysis window to generate a visual anchor point sequence.

[0017] Based on the comparison between the deviation of the adjacent anchor point interval from the average interval and the preset stability threshold, the step frequency stability of the terminal anchor point sequence and the visual anchor point sequence within the current analysis window is determined respectively.

[0018] Preferably, the process of determining the update time difference includes: acquiring the terminal anchor sequence and visual anchor sequence within each analysis window; and discretizing the terminal anchor sequence and visual anchor sequence into corresponding terminal pulse sequence and visual pulse sequence.

[0019] The relative time difference of each analysis window is determined based on the cross-correlation results of the terminal pulse sequence and the visual pulse sequence; the window weight of each analysis window is determined based on the cross-correlation peak intensity and anchor point continuity; and the relative time difference of each analysis window is fused based on the window weight of each analysis window to obtain the global relative time difference.

[0020] The update time difference is determined based on the current global relative time difference and the effective time difference that meets the stability condition in the previous analysis period.

[0021] Preferably, when determining a user's intention rating for a candidate target object, the intention calculation module's calculation process includes determining an online interest score; determining the online interest score includes:

[0022] Extract online interaction event data associated with the candidate target object;

[0023] Online interest scores are generated based on the action weights corresponding to online interactive event data and the time decay characteristics based on the recording time.

[0024] Preferably, the process of determining the offline contact score includes:

[0025] Extract offline behavioral event data associated with the candidate target object;

[0026] The offline baseline value is determined based on the action weight and action confidence corresponding to the offline action type.

[0027] The offline decay value is determined based on the time difference between the event time of the offline behavioral event data and the end time of the current evaluation period;

[0028] The offline contact score is generated based on the offline baseline value and the offline attenuation value.

[0029] Preferably, the process of determining the synchronization score includes: screening candidate event pairs from the calibrated online and offline behavior trajectories according to preset synchronization judgment rules, wherein the candidate event pairs meet the following conditions:

[0030] The following conditions must be met: belonging to the same candidate target object; the time difference between the recording time of online interaction event data and the event time of offline behavior event data is less than the synchronization threshold; and the online action type and the offline action type satisfy the preset correspondence.

[0031] The synchronous decay value of each candidate event pair is determined based on the time difference and the preset decay function;

[0032] By combining the pre-configured synchronization weights based on the event type combinations, the synchronization decay values ​​of each candidate event pair are accumulated to obtain a synchronization score.

[0033] Preferably, the process of determining the deduction value includes:

[0034] Based on the preset invalid stay determination rules, the offline stay events corresponding to the candidate target objects are extracted. The invalid stay determination rules include at least the stay duration threshold, target area association conditions, and abnormal behavior conditions.

[0035] The deduction value is determined based on the characteristics of the abnormal behavior and the duration of the stay.

[0036] Preferably, the process of outputting the corresponding service action instruction includes:

[0037] The first, second, third, or fourth level is determined based on the intention score and three preset stratification thresholds. The three preset stratification thresholds satisfy the following conditions in sequence: the first stratification threshold is less than the second stratification threshold, and the second stratification threshold is less than the third stratification threshold.

[0038] The score changes are determined based on the intention scores of the current assessment cycle and the previous assessment cycle.

[0039] Based on the judgment results of whether the user's intention rating of the candidate target object reaches the preset level and whether the change of the rating meets the preset conditions, determine whether to output the key attention mark;

[0040] Based on the final determined level of intent, output corresponding service action instructions in the form of sales guide intervention, promotional push, experience reminder, continued observation, or routine record.

[0041] Compared with the prior art, the present invention has the following beneficial effects:

[0042] This invention collects online interaction event data, terminal motion data, offline behavioral event data, and visual cadence data of the same user. It then uses the mapping relationship between offline area identifiers and online target object identifiers to determine candidate target objects. After determining the update time difference using terminal anchor point sequences and visual anchor point sequences, it performs time calibration on the online behavioral trajectory. Furthermore, based on the calibrated online and offline behavioral trajectories and candidate target objects, it calculates an intention score and outputs corresponding service action instructions. This provides an intention-level hierarchical management scheme that can jointly judge the online and offline behaviors of the same user and the same target object under a unified time reference. Compared to existing methods that process online and offline records independently or only use a wide time window for rough binding, this effectively reduces the impact of candidate object drift, inaccurate synchronous behavior recognition, interference from outdated behaviors in the score, and false positives from invalid dwell times, thereby improving the accuracy of intention recognition and service action output. Attached Figure Description

[0043] To more clearly illustrate the technical solutions in the embodiments of this application and the prior art, the accompanying drawings used in the description of the embodiments and the prior art will be briefly introduced below:

[0044] Figure 1 This is a schematic diagram of the module of the intention level management system based on online and offline integrated behavioral trajectories in an embodiment of the present invention. Detailed Implementation

[0045] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments.

[0046] The intention-based hierarchical management system, based on integrated online and offline behavioral trajectories, includes:

[0047] The data collection module is used to collect online interaction event data, terminal motion data, offline behavior event data and visual cadence data of the same user. Among them, multiple online interaction event data are sorted according to the recording time to form an online behavior trajectory, and multiple offline behavior event data are sorted according to the event time to form an offline behavior trajectory.

[0048] The object association module is used to determine candidate target objects based on the area identifiers in offline behavioral event data and according to the preset mapping relationship between offline area identifiers and online target object identifiers;

[0049] Anchor point generation module, used to generate terminal anchor point sequence and visual anchor point sequence based on terminal motion data and visual cadence data;

[0050] The time calibration module is used to determine the update time difference based on the terminal anchor sequence and the visual anchor sequence, and to perform time calibration on the online behavior trajectory based on the update time difference to obtain the calibrated online behavior trajectory.

[0051] The intention calculation module is used to determine the user's intention score for the candidate target object within the preset current evaluation period based on the calibrated online behavior trajectory, offline behavior trajectory and candidate target object;

[0052] The hierarchical management module is used to determine the intention level based on the intention score in the current assessment period and the change in the intention score relative to the previous assessment period, and output the corresponding service action instructions.

[0053] Online interaction event data includes user identifier, target object identifier, action type, recording time, and action weight;

[0054] Offline behavioral event data includes user identifier, area identifier, offline action type, event time, and action confidence.

[0055] Terminal motion data includes sampling time and vertical acceleration components;

[0056] Visual cadence data includes visual sampling time and changes in the vertical displacement of the human body;

[0057] The object association module is used to query the target object mapping relationship based on the offline region identifier and the online target object identifier in order to determine the candidate target object;

[0058] The target object mapping relationship includes the mapping relationship between product code, electronic price tag code, or shelf area code;

[0059] When no corresponding target object mapping relationship is found, perform one of the following operations: assign a default location based on the region identifier;

[0060] Perform downgrade matching based on the higher-level region;

[0061] Generate a mapping failure record containing a region identifier, timestamp, and a miss status flag.

[0062] This implementation provides an intention-based hierarchical management system based on integrated online and offline behavioral trajectories, such as... Figure 1 As shown; the system collects behavioral data of the same user both online and offline, and determines candidate target objects based on the mapping relationship between offline areas and online target objects;

[0063] By using terminal motion data and visual cadence data, anchor point sequences that can be used for alignment are extracted, the update time difference between online behavior trajectory and offline behavior trajectory is calculated, and the online behavior trajectory is time-calibrated.

[0064] Based on a unified time frame, the calibrated online behavior trajectory, offline behavior trajectory, and candidate target objects are jointly analyzed to obtain the user's intention score for the candidate target objects. Combined with the score changes between the current evaluation period and the previous evaluation period, corresponding service action instructions are output. Through this processing link, online events and offline events that were originally recorded independently can be judged under the same time reference, thereby improving the reliability of intention recognition.

[0065] Based on the above implementation method, online interaction event data, terminal motion data, offline behavior event data and visual cadence data of the same user are obtained. According to the user identifier and the evaluation period, the data are collected and organized to generate online behavior trajectory and offline behavior trajectory.

[0066] The evaluation period is the time window limit for the system to collect user behavior data, determine intent, and update levels. It can be configured as a fixed time span set by the system, such as 5 minutes or 10 minutes, or as a single continuous session period dynamically triggered by events such as users entering and leaving the monitoring area.

[0067] Specifically, online interaction event data refers to records of object-related interactions generated by users on the terminal side, such as browsing object details, collecting objects, and adding objects to the shopping cart. It includes at least user identifier, target object identifier, action type, recording time, and action weight. The action weight is a pre-configured numerical parameter used to quantify the degree of contribution of different action types to the user's final intention conversion.

[0068] Offline behavioral event data consists of user behavior records identified in physical scenarios, such as entering an area, staying, touching a sample, and leaving an area. It includes at least user identifier, area identifier, offline action type, event time, and action confidence. Terminal motion data consists of vertical acceleration sampling sequences output by terminal sensors. Visual gait data consists of sequences of vertical displacement changes of the human body extracted from the visual side, such as displacement fluctuations formed by head and shoulder undulations.

[0069] In practical implementation, online interaction event data can be written to the local cache queue by the terminal tracking component, while offline behavior event data and visual cadence data can be written to the store-side cache queue to be integrated by the edge processing process. The fusion computing process reads and deserializes the data into event objects according to the same user identifier. For data segments that have been read, the corresponding cache space can be released after the current evaluation period ends or the current analysis window is closed.

[0070] This embodiment establishes a processing data set with a unified input basis based on the same user identifier; it constructs online and offline behavioral trajectories through user data collection and trajectory organization, thus solving the problem of multi-source heterogeneous data fusion processing.

[0071] Based on the above implementation method, the area identifier in the offline behavior event data is obtained, and the area identifier is queried according to the preset mapping relationship between the offline area identifier and the online target object identifier to determine the candidate target object.

[0072] The area identifier is used to represent the object carrying area corresponding to the offline event, such as the shelf area, display area, and trial area; the target object mapping relationship is used to project the offline area onto the online object, which can be composed of product code, electronic price tag code, or shelf area code; the candidate target object represents the set of objects that are associated with the user's offline contact area within the current evaluation period, such as a product or a group of products in the same shelf area;

[0073] In practical implementation, the mapping relationship can be pre-stored in the mapping table synchronously issued by the store product platform. The mapping table resides in memory in the form of an object dictionary or array index during the fusion computing process. After the object association module reads the area identifier in the offline behavior event, it can directly look up the corresponding target object identifier in the current process without having to access external storage again, thereby reducing query latency.

[0074] When no corresponding target object mapping relationship is found, default allocation can be performed by region identifier, or downgraded matching can be performed by the upper level region. Specifically, in the pre-built regional spatial tree topology structure of the system, starting from the sub-region node that does not currently match the mapping, the parent region node is searched upwards. If the parent region node is bound to a default target object group of a broad category in the mapping table, then the default target object group is used as the current candidate target object.

[0075] If the mapping still fails to hit, the process continues to traverse to the next higher level until the root node is reached, or a mapping failure record containing the region identifier, system cycle marker, and miss status marker is generated for subsequent supplementary maintenance of the missed regions. The system cycle marker is used to record the timestamp of the specific running batch or corresponding evaluation period in which the mapping failure event occurred, so that the background system can trace back the timeline to investigate the cause of the missing mapping.

[0076] This embodiment uses a preset mapping relationship to perform precise matching and establish a precise correspondence between offline regions and online objects to avoid data association errors; it can generate candidate target objects based on a deterministic mapping between region identifiers and object identifiers to prevent offline events from failing to be assigned to specific objects;

[0077] Based on the above implementation method, the calibrated online behavior trajectory, offline behavior trajectory and candidate target object are obtained. According to the preset scoring logic, the online behavior, offline behavior and their time correspondence of the same candidate target object are jointly calculated to determine the user's intention score for the candidate target object. The intention level is determined according to the score change between the current evaluation period and the previous evaluation period, and the corresponding service action instruction is output.

[0078] Intention score is used to characterize the overall result of a user’s attention to a candidate target. The change in score reflects the relationship between the current evaluation period score and the previous evaluation period score, whether it increases, decreases or remains the same. Service action instructions are structured outputs that can be directly executed by the business system, such as sales guide intervention instructions, discount push instructions, experience reminder instructions, continue observation instructions or routine recording instructions.

[0079] In practical implementation, the hierarchical management module can encapsulate the determined intention level and service action into a structured message and write it into the store-side business message queue. The sales guide terminal or marketing system can obtain the corresponding instructions by continuously listening to the message queue and rewrite the message status bit to the consumed status after reading.

[0080] This implementation method further transforms the scoring results into triggerable business actions, so that the identification results can be directly used for on-site services and subsequent recording; it can realize intention scoring and level output based on the fusion trajectory under a unified time, avoiding the scoring results remaining at the analysis level.

[0081] By identifying candidate target objects and scoring them on a unified timeline, the online interaction features and offline behavioral features of the same target object can be accurately matched and calculated. In scenarios where multiple target objects interact in parallel, the data range of candidate target objects can be limited by region mapping, which can effectively filter out the interference of non-related object features on the calculation process.

[0082] In the application embodiment, the system can be deployed in a processing environment that includes user terminals, store edge nodes, and fusion computing nodes; user terminals continuously generate online interaction event data and terminal motion data, while store edge nodes continuously generate offline behavior event data and visual cadence data; fusion computing nodes read the corresponding data from the queue to be fused according to the user identifier, and complete the determination of candidate target objects, time calibration, intention scoring, and level output;

[0083] After receiving the service action instruction, the business execution system triggers reminders on the shopping guide terminal, pushes promotional information, or updates observation records. Subsequent new user events continue to flow into the current evaluation cycle or the next evaluation cycle in the same way, so as to continuously update the intention level of the same candidate target.

[0084] The anchor point generation module is used to determine the terminal anchor point sequence and the visual anchor point sequence, including: dividing the terminal motion data and visual gait data into multiple analysis windows.

[0085] According to the preset gait frequency band, the vertical acceleration component in the terminal motion data is bandpass processed, and the terminal peak extraction threshold is determined based on the statistical characteristics of the mean, standard deviation and peak distribution within the current analysis window to generate the terminal anchor point sequence;

[0086] Based on the preset gait frequency band, the visual gait frequency data is smoothed and bandpass processed, and the visual peak extraction threshold is determined based on the statistical characteristics of the mean, standard deviation and peak distribution within the current analysis window to generate a visual anchor point sequence.

[0087] Based on the comparison between the deviation of the adjacent anchor point interval from the average interval and the preset stability threshold, the step frequency stability of the terminal anchor point sequence and the visual anchor point sequence within the current analysis window is determined respectively.

[0088] The time calibration module is used to determine the update time difference, including: acquiring the terminal anchor sequence and visual anchor sequence within each analysis window;

[0089] The terminal anchor sequence and visual anchor sequence are discretized into corresponding terminal pulse sequences and visual pulse sequences;

[0090] The relative time difference of each analysis window is determined based on the cross-correlation results of the terminal pulse sequence and the visual pulse sequence;

[0091] The window weights of each analysis window are determined based on the peak intensity of cross-correlation and the continuity of anchor points. The relative time differences of each analysis window are then fused based on the window weights of each analysis window to obtain the global relative time difference.

[0092] The update time difference is determined based on the current global relative time difference and the effective time difference that meets the stability condition in the previous analysis period.

[0093] This implementation describes the specific processing steps of the anchor point generation module and the time calibration module. The process first extracts anchor point sequences that can characterize the gait beat of the same user from terminal motion data and visual gait frequency data. Then, it discretizes and cross-corresponds the anchor point sequences in different analysis windows to obtain the relative time difference. The updated time difference is formed by combining the window weight with the effective time difference in the previous analysis period. In this way, stable time calibration of online behavior trajectories can be performed without relying on high-precision network time synchronization.

[0094] Based on the above implementation method, the vertical acceleration component in the terminal motion data is obtained, and the vertical acceleration component is bandpass processed according to the preset gait frequency band to generate a terminal anchor point sequence; the vertical acceleration component is the acceleration sampling value reflecting the up and down movement fluctuations in the terminal coordinate system or gravity-aligned coordinate system, such as the periodic fluctuations caused by body undulations when walking;

[0095] Bandpass processing aims to retain only the signal components within the frequency band of the dominant gait frequency in order to suppress slow posture changes and high-frequency jitter. The preset gait frequency band is a frequency range pre-set based on the characteristics of normal human walking beats, preferably ranging from 1.0Hz to 2.5Hz. The terminal anchor sequence is a set of peak moments extracted from the bandpass-processed terminal beat signal.

[0096] Specifically, the terminal peak extraction threshold is jointly determined by the mean, standard deviation, and preset terminal peak sensitivity coefficient of the filtered terminal beat signal within the current analysis window; the extraction threshold is calculated by adding the mean to the product of the terminal peak sensitivity coefficient and the standard deviation, which is used to limit the amplitude limit of the terminal beat signal that can be identified as a valid gait peak.

[0097] This calculation is used to distinguish between normal gait fluctuations and terminal random non-gait vibrations; the mean reflects the overall vibration baseline within the current window, and the standard deviation reflects the dispersion of gait fluctuations relative to the baseline; the terminal peak sensitivity coefficient is used to dynamically balance missed detections and false detections within the current window, and its value can be from 0.8 to 1.5, with an example value of 1.1; based on this calculation logic, local peaks are searched within the analysis window and the corresponding peak times are extracted to obtain the terminal anchor point sequence;

[0098] In this embodiment of the invention, the threshold is dynamically and adaptively adjusted based on the mean and standard deviation within the current analysis window to reduce the impact of asynchronous gait characteristics and sensor vibration noise on the peak extraction accuracy; the terminal anchor point sequence is generated based on bandpass processing and adaptive threshold to shield the interference of terminal non-gait vibration on peak extraction;

[0099] Based on the above implementation method, visual gait frequency data is acquired, and smoothing and bandpass processing are performed on the visual gait frequency data according to the local mean smoothing rule based on the sliding time window and the gait frequency band. The window length of the sliding time window is configured to be 3 to 5 sampling periods to generate a visual anchor point sequence.

[0100] Visual gait data is a time series composed of changes in the vertical displacement of the human body, such as changes in the height of the head and shoulder contour center and vertical fluctuations in the center of the human body frame; smoothing is used to suppress spike interference caused by single-frame jitter or local detection errors;

[0101] The visual anchor sequence is the set of peak moments extracted from the smoothed and bandpassed visual beat signal; specifically, the visual peak extraction threshold is determined by the mean, standard deviation and preset visual peak sensitivity coefficient of the visual beat signal in the current analysis window.

[0102] The threshold is calculated by adding the mean to the product of the visual peak sensitivity coefficient and the standard deviation, in order to filter out the fluctuations with gait rhythm significance from visual displacement fluctuations; the mean reflects the central horizontal of the overall vertical position change of the human body within this window.

[0103] The standard deviation reflects the degree of fluctuation in the vertical undulation of the human body; the visual peak sensitivity coefficient is used to control the sensitivity of visual peak extraction, and its value can be from 0.6 to 1.2, with an example value of 0.9; based on this calculation logic, local peak moments are extracted within the current analysis window to obtain the visual anchor point sequence;

[0104] This implementation uses a combination of smoothing and bandpass processing to first reduce local jitter in visual detection, and then highlights periodic fluctuations through adaptive thresholding, making visual beat features less likely to be covered by short-term occlusion or frame jumping; it can generate visual anchor point sequences based on the smoothing bandpass and adaptive thresholding of visual sequences, avoiding direct interference of single-frame visual noise with the alignment basis.

[0105] Based on the above implementation method, the terminal anchor sequence and the visual anchor sequence are obtained. According to the deviation of the interval between adjacent anchor points from the average interval, the step frequency stability of the terminal anchor sequence and the visual anchor sequence are determined to determine whether the current analysis window meets the subsequent time difference solution conditions.

[0106] Step frequency stability reflects whether the interval between adjacent anchor points within the same analysis window maintains a small fluctuation around its average interval. The deviation is the absolute amount of the difference between each adjacent anchor point interval and the corresponding average interval. The stability threshold limits the maximum allowable fluctuation range, which can be set according to a certain proportion of the average step interval of the corresponding window. For example, the proportion can be 25%.

[0107] When determining stability, for the terminal side, the time difference of each adjacent terminal anchor point in the current analysis window is traversed, and the absolute deviation of the time difference between the terminal anchor point and the average terminal anchor point interval in the window is calculated. All absolute deviations are accumulated to obtain the cumulative deviation of the terminal side, so as to determine whether the gait rhythm of the terminal side is continuous.

[0108] For the visual side, the same calculation logic is used to calculate and accumulate the absolute deviation between the time interval of adjacent visual anchor points and the average visual anchor point interval to obtain the cumulative deviation degree of the visual side, so as to determine whether the gait rhythm of the visual side is continuous; when the cumulative deviation degree of the terminal side and the visual side is less than the corresponding preset stability threshold, the current analysis window is determined to be stable.

[0109] This embodiment adds stability judgment rules to eliminate interference data with non-periodic beat characteristics, so that cross-correlation calculation is based on data windows whose feature stability meets the set threshold; step frequency stability window can be filtered based on anchor point interval deviation to avoid unstable windows directly entering time difference calculation;

[0110] Based on the above implementation method, the terminal anchor sequence and visual anchor sequence in each analysis window are obtained. According to the unified discrete time step, the terminal anchor sequence and visual anchor sequence are discretized into the corresponding terminal pulse sequence and visual pulse sequence. The relative time difference of each analysis window is determined according to the cross-correlation result of the two.

[0111] The terminal pulse sequence and visual pulse sequence obtained after discretization are sequences on a unified discrete time axis, where the position of the anchor point is recorded as a pulse and the other positions are recorded as zero values. The discrete time step used is... The time resolution used to control the cross-correlation calculation can be on the order of 20 milliseconds to balance computational load and alignment accuracy.

[0112] When calculating the relative time difference, the current analysis window at the offset index is determined using the cross-correlation formula. Lower terminal pulse sequence With visual pulse sequence degree of overlap This is used to measure whether shifting the anchor point sequence on one side along the time axis makes the gait beats on both sides more consistent:

[0113]

[0114] in, To analyze the index of the window, This represents the number of discrete sampling points in the current window. For discrete-time offset indexes, their physical meaning represents the difference in the number of steps the terminal pulse sequence and the visual pulse sequence slide along the time axis. The index for traversing the discrete sampling point sequence; the degree of overlap. Offset index at which the maximum value is reached And based on this, determine the relative time difference of the window. ;

[0115] This implementation method uses the gait beats that coexist on both the terminal and the visual side to solve the cross-correlation problem, so that the time difference estimation no longer depends on whether the terminal system clock and the edge side clock are naturally consistent; it can realize the window-level relative time difference solution based on pulse discretization and cross-correlation calculation, avoiding the direct transmission of timing error to the behavior alignment result;

[0116] Based on the above implementation method, the relative time difference of each analysis window is obtained, the window weight of each analysis window is determined according to the cross-correlation peak intensity and anchor point continuity, and the relative time difference of each analysis window is fused based on the window weight of each analysis window to obtain the global relative time difference.

[0117] The update time difference is determined based on the current global relative time difference and the effective time difference that meets the stability condition in the previous analysis period; the window weight determines the contribution ratio of the current analysis window in the global time difference fusion, so that windows whose peak feature distribution meets the preset clarity and whose continuity parameters meet the standard are given higher calculation distribution weights;

[0118] Global relative time difference is the combined time difference result given by multiple effective analysis windows; effective time difference is the time difference result that has passed the stability test and been retained in the previous analysis period.

[0119] In multi-window fusion and update calculations, the global relative time difference obtained by fusing multiple effective windows in the current analysis period is calculated using the global relative time difference calculation formula. This characterizes the overall offset of the current terminal clock relative to the offline clock.

[0120]

[0121] in, For the set of valid windows, For the first Window weights for each analysis window. For the first The relative time difference of each analysis window;

[0122] The window weight is determined by the product of the beat sharpness weight, which corresponds to the degree of distinction between the main and secondary peaks of cross-correlation, and the behavioral continuity weight, which reflects the continuity of anchor points within the window. The beat sharpness weight is used to measure whether the main cross-correlation peak is significantly stronger than other candidate peaks. The behavioral continuity weight is used to measure whether the terminal anchor point and the visual anchor point within the window maintain continuous and stable gait characteristics.

[0123] In practice, the method for determining the beat sharpness weight is as follows: Extract the largest peak from the cross-correlation results as the primary peak, and extract the second largest peak as the secondary peak. Calculate the ratio of the primary peak to the secondary peak. When this ratio is greater than a preset ratio threshold, such as 1.5, map the ratio to between 0 and 1 using a preset parameterized S-shaped function. The specific mapping formula is as follows:

[0124]

[0125] in, Weighted by beat clarity, The ratio of the primary peak value to the secondary peak value. The ratio threshold, The gain coefficient is used to control the steepness of the mapping curve. The natural constant is used to calculate the beat sharpness weight;

[0126] The method for determining the continuity weight of behavior is as follows: calculate the variance of the terminal anchor point interval and the visual anchor point interval within the current analysis window respectively, and use the sum of the two variances as the continuity deviation.

[0127] To ensure dimensional consistency, a reference standard variance constant with the same time square dimension as the continuity deviation is introduced. The dimensionless behavioral continuity weight between 0 and 1 is obtained by dividing the reference standard variance constant by the sum of the continuity deviation and the reference standard variance constant. The above method does not require the introduction of a complex implicit mapping model, and can achieve accurate quantification of window weights simply by decomposing and aligning the dimensions of the statistical indicators.

[0128] To avoid abrupt changes in update time difference caused by random biases in the current analysis period, the update time difference is calculated by balancing the current analysis period's estimation result with the previous effective time difference result using the update time difference calculation formula, thus obtaining the update time difference used for actual time calibration. :

[0129]

[0130] in, The effective time difference that satisfies the stability condition in the previous analysis period. The update coefficient, used to control the impact of the current result on the update time difference, can range from 0.3 to 0.6, with an example value of 0.5. If the current analysis window set does not meet the stability condition, the previous effective time difference remains unchanged.

[0131] This implementation method uses window weight fusion and update coefficient balancing to ensure that the time difference results can follow the actual drift changes while being less susceptible to interference from single abnormal window data. It can achieve stable output of updated time difference based on multi-window fusion and effective time difference update, avoiding significant fluctuations in relative time difference.

[0132] This embodiment uses the collaborative data features of terminal anchor sequence and visual anchor sequence for dual-sided constraints, and filters out non-compliant data windows based on stability judgment and window weight parameters, thereby improving the accuracy of time difference calculation in scenarios with free drift of heterogeneous clocks and variable gait states.

[0133] For scenarios where the user's gait is weak but the rhythm is continuous, the adaptive threshold can preserve the effective peak value; for scenarios where there are local jumps on the visual side but the overall rhythm is still recognizable, smoothing and multi-window fusion can reduce the disturbance of isolated outliers.

[0134] In the application example, the fusion computing node reads the corresponding segments from the terminal motion array and visual step frequency array in memory within each analysis window, and completes filtering, peak extraction, stability judgment and cross-correlation calculation; the relative time difference and window weight of each window are written into the time difference buffer of the current analysis cycle, and fusion is performed and updated time difference is generated after the end of the analysis cycle;

[0135] The updated time difference is written to the time difference status bit corresponding to the current user. After reading the status bit, the time calibration module performs time shifting on the online behavior trajectory in the current evaluation period to form the calibrated online behavior trajectory required for subsequent scoring.

[0136] When determining a user's intention score for a candidate target object, the intention calculation module's calculation process includes determining an online interest score; determining the online interest score includes: extracting online interaction event data associated with the candidate target object; and generating an online interest score based on the action weights corresponding to the online interaction event data and the time decay characteristics based on the recording time.

[0137] The intent calculation module is used to determine the offline contact score, including: extracting offline behavioral event data associated with the candidate target object;

[0138] The offline baseline value is determined based on the action weight and action confidence corresponding to the offline action type; the offline decay value is determined based on the time difference between the event time of the offline behavioral event data and the end time of the current evaluation period; and the offline contact score is generated based on the offline baseline value and the offline decay value.

[0139] The intention calculation module is used to determine the synchronization score, including: screening candidate event pairs from the calibrated online and offline behavior trajectories according to preset synchronization judgment rules. The candidate event pairs meet the following conditions:

[0140] The following criteria are met: the online interactive event data and the offline behavioral event data have a time difference of less than the synchronization threshold; the online action type and the offline action type satisfy the preset correspondence; the synchronization decay value of each candidate event pair is determined based on the time difference and the preset decay function; the synchronization decay value of each candidate event pair is accumulated by combining the synchronization weights pre-configured by the event type combination to obtain the synchronization score.

[0141] The intention calculation module is used to determine the deduction value, including: extracting the offline dwelling events corresponding to the candidate target objects according to the preset invalid dwelling judgment rules. The invalid dwelling judgment rules include at least the dwelling duration threshold, target area association conditions and abnormal behavior conditions; and determining the deduction value based on the abnormal behavior characteristics and dwelling duration.

[0142] The hierarchical management module is used to output corresponding service action instructions, including: determining the first level, second level, third level or fourth level based on the intention score and three preset hierarchical thresholds, wherein the three preset hierarchical thresholds satisfy the following conditions in sequence: the first hierarchical threshold is less than the second hierarchical threshold, and the second hierarchical threshold is less than the third hierarchical threshold.

[0143] The score change is determined based on the intention score of the current evaluation period and the previous evaluation period; whether to output a key attention mark is determined based on whether the user's intention score for the candidate target object reaches the preset level and whether the score change meets the preset conditions; and the corresponding service action instructions, such as sales guide intervention, discount push, experience reminder, continued observation, or regular record, are output based on the final determined intention level.

[0144] This implementation describes the specific processing steps of the intention calculation module and the hierarchical management module. The process revolves around candidate target objects, calculating online interest scores, offline contact scores, synchronization scores, and deduction scores, and generating an intention score accordingly. The intention level is determined based on the score changes between the current and previous evaluation periods, and corresponding service action instructions are output. In this way, multi-source behavioral evidence under a unified timeline can be converted into executable hierarchical results.

[0145] Based on the above implementation method, candidate target objects are obtained. Based on the candidate target objects, online interaction event data associated with the candidate target objects are extracted from the calibrated online behavior trajectory. Action weight accumulation and time decay processing are performed on the online interaction event data to generate online interest scores.

[0146] The calculated online interest score reflects the intensity of online attention a user generates around a candidate target object, such as the higher intensity of attention formed by continuing to save the object after viewing details.

[0147] The online base value is the original behavior weight accumulation value obtained by summing the weights of multiple online actions; the online decay value is the amount of time impact that needs to be deducted as the time of the event gradually moves away from the end of the current evaluation period.

[0148] When calculating online interest scores, based on candidate target objects Related online event index collection Calculate the candidate target object based on the various online interaction events. Online interest score To measure the effective impact of users' online attention retention on the object during the current evaluation period, the calculation formula is as follows:

[0149]

[0150] Among them, for the traversal index is Online interactive events, The action weights are used to differentiate the contribution of actions such as browsing, staying, saving, and adding to cart to the intention; The online event recording time is calibrated. This is the end time of the current evaluation period; The online interest time decay constant, which controls the rate at which the influence of online behavior weakens over time, can be adjusted within a preset range according to the scenario. Since the online interest score is specified in the embodiment as the online base value minus the online decay value, the above index term can be understood as the amount retained after attenuating the original weight.

[0151] This embodiment extracts online feature scores based on action weight coefficients and time decay parameters to improve the accuracy of behavioral feature data in the time dimension in representing the evaluation results; it can calculate online interest scores based on action weights and time decay, avoiding the long-term inflation of scores due to outdated online behaviors;

[0152] Based on the above implementation method, candidate target objects are obtained. Based on the candidate target objects, offline behavior event data associated with the candidate target objects are extracted from the offline behavior trajectory. The offline behavior event data is calculated by superimposing offline action weights and action confidence, and the attenuation processing is performed by combining the time difference between the event time and the end of the current evaluation period to generate an offline contact score.

[0153] The calculated offline contact score reflects the intensity of contact between users and candidate target objects in offline physical scenarios, such as reaching out to touch or picking up / placing samples after staying in the area.

[0154] The offline baseline value is the original contact characteristic reference value determined by the offline action weight and action confidence; the offline decay value is the amount of influence that decreases as the offline event moves further away from the end of the current evaluation period.

[0155] When calculating offline contact scores, based on candidate target objects Related offline event index set Calculate candidate target objects from various offline behavioral events. Offline contact score The effective strength of a user's interaction with an object in a physical setting is measured by the following formula:

[0156]

[0157] Among them, for the traversal index is Offline behavioral events, Identify its action type The corresponding action weights are used to distinguish the differences in the impact of behaviors such as entering the area, staying, reaching out to touch, and picking up or placing samples on intentions. The action confidence score for this offline action reflects the reliability of the visual or behavioral recognition results. The event time for offline behavioral events; This is the end time of the current evaluation period; To control the impact of offline contact on the retention period, the offline contact time decay constant is used.

[0158] This implementation calculates the offline contact score by combining action type, action confidence, and time decay, thereby accurately extracting and distinguishing short-term high-weight action features from long-term low-weight features at the data processing level; the calculation of the offline contact score can be based on offline action weight, action confidence, and time decay, avoiding a rough judgment of offline interest based solely on dwell time.

[0159] Based on the above implementation method, the calibrated online behavior trajectory and offline behavior trajectory are obtained. According to the preset synchronization judgment rule, candidate event pairs that belong to the same candidate target object, have a time difference less than the synchronization threshold, and whose online action type and offline action type satisfy the preset correspondence are selected. The candidate event pairs are processed by time difference decay and synchronization weight accumulation to obtain the synchronization score.

[0160] Candidate event pairs are online and offline events that can be considered related to the same object under a unified timeline. For example, a user interacts with an object offline while viewing the object's details online. The synchronization threshold limits the maximum time difference boundary that allows online and offline events to be considered to occur synchronously.

[0161] Synchronization weights are pre-configured weights for event type combinations, used to reflect the differences in concurrency value of different paired behaviors. When determining the synchronization score, the absolute value of the time difference between the online event time and the offline event time in each candidate event pair is obtained by traversing each candidate event pair belonging to the candidate target object.

[0162] Using a preset synchronization decay constant, the absolute value of the time difference is subjected to negative exponential decay processing to determine the synchronization decay value of each candidate event pair. The calculation formula is as follows:

[0163]

[0164] in, This represents the absolute value of the time difference between the online event moment and the offline event moment. The synchronization decay constant is multiplied by the pre-configured synchronization weights for event type combinations; the results for each candidate event are then summed to obtain a comprehensive synchronization score. The above processing is used to measure the concurrency of users simultaneously viewing or comparing objects online while interacting with them offline.

[0165] Synchronization weights are used to distinguish the differences in intentional feature contributions between different combinations of behaviors, such as offline hand-reaching and online detail viewing versus offline casual stay and brief online browsing. In practical applications, pre-configured synchronization weights can be implemented by a preset two-dimensional action weight matrix, where the row dimension corresponds to the set of online action types and the column dimension corresponds to the set of offline action types. Each element in the two-dimensional matrix corresponds to the synchronization weight value of a specific behavior combination. This two-dimensional correspondence is loaded into memory in the form of key-value pairs for fast matching during system initialization. The synchronization decay constant is used to control the rate at which the concurrency value decreases as the time difference increases.

[0166] This embodiment filters candidate event pairs based on object and action combinations under a unified timeline and performs attenuation processing on time differences to accurately extract the correlation features of synchronous concurrent events, effectively filtering out unrelated online and offline independent events within the same system cycle; it can calculate the synchronization score based on a unified timeline and synchronization judgment rules to avoid false concurrent events from inflating the intention score;

[0167] Based on the above implementation method, the offline dwelling events corresponding to the candidate target objects are obtained, and the offline dwelling events are screened according to the preset invalid dwelling judgment rules to determine the deduction value; the invalid dwelling judgment rules include at least the dwelling time threshold, target area association conditions and abnormal behavior conditions;

[0168] Abnormal behavioral characteristics are reflected in behavioral data features that occur during the stay but do not meet the preset effective contact conditions, such as high-frequency fragmented jitter in terminal acceleration, continuous online events falling on non-object-related interactions, visual head orientation deviating from the target area, and no hand-reaching or contact enhancement behavior offline.

[0169] The deduction factor is the deduction amount corresponding to the duration of stay and the abnormal behavior characteristics. When determining the deduction score, the set of offline stay events that meet the candidate target object association and satisfy the invalid stay judgment rules are extracted. For each stay event in this set, in order to avoid the direct introduction of the time dimension and the resulting dimension conflict when the score is integrated, the duration is divided by the preset benchmark stay duration with the same time dimension to convert it into a dimensionless stay degree coefficient.

[0170] The total deduction score is obtained by multiplying the dwell time coefficient by a deduction factor reflecting the strength of abnormal behavior characteristics, and then summing the dimensionless products corresponding to each dwell time event. ;

[0171] This calculation effectively suppresses and reduces dwelling behavior that lacks evidence of genuine interest; the deduction factor comprehensively reflects behavioral signs that do not support genuine shopping interest during the dwelling period, and the dwelling degree coefficient is used to relatively represent the persistence of ineffective dwelling.

[0172] When determining the deduction factor, abnormal behavior characteristics are quantified into the number of abnormal items. For example, during the dwell period, if any of the following occurs, the number of abnormal items increases by 1: the terminal acceleration sliding variance in the frequency band above 5Hz exceeds the preset jitter threshold, the number of non-object-related interaction actions recorded in the online event exceeds the preset upper limit, or the head yaw angle identified by the visual side exceeds the preset deviation angle. The deduction factor is configured as the product of the preset base deduction and the number of abnormal items, so that the more abnormal items there are, the greater the deduction force for the dwell behavior.

[0173] This embodiment performs condition screening on the dwell data, quantitatively deducting data that does not match the contact action, does not hit the target object, or meets abnormal conditions from the total score, thereby eliminating the interference of invalid behavior data on the intention calculation; the deduction value can be calculated based on the invalid dwell judgment rule, avoiding simply increasing the intention by dwelling;

[0174] Based on the above implementation method, online interest score, offline contact score, synchronization score and deduction score are obtained. The scores are combined and calculated according to preset coefficients to determine the intention score. Based on the intention score and three preset stratification thresholds, as well as the score changes in the current evaluation period and the previous evaluation period, the first level, second level, third level or fourth level is determined, and the corresponding service action instructions are output.

[0175] The intention score represents the overall intention of the candidate target within the current evaluation period; three preset stratification thresholds are used to divide the intention score into four level ranges; the key attention mark is an enhanced reminder mark output when the score reaches a preset level and the score change meets preset conditions;

[0176] When calculating intention scores, candidate target objects are identified. Intention rating This combines online attention, offline contact, concurrent engagement, and invalid dwell time into a comparable comprehensive result, calculated using the following formula:

[0177]

[0178] in, , , and Interest scores on the control line Offline contact score Synchronization score and deduction of points The coefficients that influence contribution in the overall score respectively reflect the impact of online attention on the judgment of the current object, the importance of evidence of physical contact, the enhancing effect of synchronous behavior, and the inhibitory strength of ineffective dwell time;

[0179] Each coefficient satisfies the non-negativity constraint and their sum is 1. An example can be set as follows: , , , To highlight the importance of synchronized behavior; when determining the level, if the intention score is less than the first tier threshold, it can be determined as a lower level below the first level; if it reaches a higher threshold, it will be determined as a higher level.

[0180] The determination of the three preset stratification thresholds is based on statistical analysis of a large number of historical user intention rating samples. Specific percentage percentiles are extracted as boundaries according to the ascending order of the ratings. For example, the 40%, 70%, and 90% percentiles are respectively marked as the first, second, and third stratification thresholds, thereby ensuring the objectivity and effectiveness of the stratification.

[0181] When the growth rate of the current evaluation period score relative to the previous evaluation period score reaches the preset condition, specifically by calculating the current evaluation period score minus the previous evaluation period score, and then dividing by the previous evaluation period score to obtain the relative growth rate, if the relative growth rate is greater than the preset growth rate threshold and the current score is not lower than the first tier threshold, a key attention mark can be output.

[0182] Correspondingly, the first, second, third, and fourth levels can be mapped to the corresponding service action instructions in the shopping guide intervention, discount push, experience reminder, continued observation, or regular recording, respectively; the service action instructions can be encapsulated into structured instruction messages that are recognized by the business system, written into the message queue or business interface buffer, and asynchronously read and executed by the shopping guide terminal, marketing push component, or recording system;

[0183] This implementation method considers both the composition of the score and the changes in the score, making it easier to identify sudden increases in genuine intent, while short-term sporadic behavior is not easily escalated directly; it can output the level of intent and trigger service actions based on multiple score combinations and score changes, avoiding the disconnect between marketing actions and genuine intent.

[0184] This embodiment integrates online interest scores, offline contact scores, synchronization scores, and deduction scores into the same scoring calculation model, using synchronization behavior parameters as independent enhancement factors to improve the extraction and recognition of concurrent behavioral features such as physical contact and simultaneous online interaction. For scenarios with many adjacent objects and users switching their focus objects in a short period of time, the combined effect of synchronization scores and deduction scores can also reduce interference caused by object confusion and pseudo-stays.

[0185] In the application embodiment, after obtaining the calibrated online and offline behavior trajectories, the fusion computing node extracts the corresponding online events, offline events and candidate event pairs from the memory event set within the current evaluation period, taking candidate target objects as the unit, calculates the online interest score, offline contact score, synchronization score and deduction score, and combines them to form an intention score;

[0186] The intention level is determined based on the stratification threshold and historical scores in the previous evaluation period, and a structured instruction containing user identifier, candidate target object identifier, intention level, key focus identifier and service action type is generated.

[0187] Once structured instructions are written into the business execution channel, they can be displayed as reminders by the shopping guide terminal, triggered by the marketing push component to provide promotional information, or updated by the observation and recording system. Subsequent new online interaction events and offline behavioral events will continue to enter the next evaluation cycle to update the score and level of the same candidate target.

[0188] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. An intention-based hierarchical management system based on online and offline integrated behavioral trajectories, characterized in that, include: The data acquisition module is used to collect online interaction event data, terminal motion data, offline behavior event data and visual cadence data of the same user. Among them, multiple online interaction event data are sorted according to the recording time to form an online behavior trajectory, and multiple offline behavior event data are sorted according to the event time to form an offline behavior trajectory. The object association module is used to determine candidate target objects based on the area identifier in the offline behavioral event data and according to the preset mapping relationship between the offline area identifier and the online target object identifier. Anchor point generation module, used to generate terminal anchor point sequence and visual anchor point sequence based on the terminal motion data and the visual cadence data; The time calibration module is used to determine the update time difference based on the terminal anchor sequence and the visual anchor sequence, and to perform time calibration on the online behavior trajectory based on the update time difference to obtain the calibrated online behavior trajectory. The intention calculation module is used to determine the user's intention score for the candidate target object within the preset current evaluation period based on the calibrated online behavior trajectory, the offline behavior trajectory and the candidate target object; When calculating intention scores, candidate target objects are identified. Intention rating The calculation formula combines online attention, offline contact, concurrent events, and invalid stays into a single comprehensive result: ； in, , , and Interest scores on the control line Offline contact score Synchronization score and deduction of points The coefficients that influence contribution in the overall score respectively reflect the impact of online attention on the judgment of the current object, the importance of evidence of physical contact, the enhancing effect of synchronous behavior, and the inhibitory strength of ineffective dwell time; The hierarchical management module is used to determine the intention level based on the intention score in the current evaluation period and the change of the intention score relative to the previous evaluation period, and output the corresponding service action instructions. The intention calculation module is used to determine the deduction value, including: According to the preset invalid stay determination rules, the offline stay events corresponding to the candidate target objects are extracted. The invalid stay determination rules include at least a stay duration threshold, target area association conditions, and abnormal behavior conditions. The deduction value is determined based on the abnormal behavior characteristics and the duration of stay; the abnormal behavior characteristics reflect the behavioral data features that occur during the stay but do not meet the preset effective contact conditions.

2. The intention-based hierarchical management system based on online and offline integrated behavioral trajectories as described in claim 1, characterized in that, The online interaction event data includes user identifier, target object identifier, action type, recording time, and action weight; The offline behavior event data includes user identifier, area identifier, offline action type, event time, and action confidence level; The terminal motion data includes the sampling time and the vertical acceleration component; The visual cadence data includes visual sampling time and changes in the vertical displacement of the human body.

3. The intention-based hierarchical management system based on online and offline integrated behavioral trajectories as described in claim 1, characterized in that, The object association module is used to query the target object mapping relationship based on the offline area identifier and the online target object identifier in order to determine the candidate target object; The target object mapping relationship includes the mapping relationship corresponding to product code, electronic price tag code, or shelf area code; If no corresponding target object mapping relationship is found, perform one of the following operations: Default assignment is performed based on the region identifier; Perform downgrade matching based on the higher-level region; Generate a mapping failure record containing a region identifier, timestamp, and a miss status flag.

4. The intention-based hierarchical management system based on online and offline integrated behavioral trajectories as described in claim 1, characterized in that, The anchor point generation module is used to determine the terminal anchor point sequence and the visual anchor point sequence, including: dividing the terminal motion data and the visual cadence data into multiple analysis windows: According to the preset gait frequency band, the vertical acceleration component in the terminal motion data is bandpass processed, and the terminal peak extraction threshold is determined based on the statistical characteristics of the mean, standard deviation and peak distribution in the current analysis window to generate a terminal anchor point sequence. According to the preset gait frequency band, the visual gait frequency data is smoothed and bandpass processed, and the visual peak extraction threshold is determined based on the statistical characteristics of the mean, standard deviation and peak distribution within the current analysis window to generate a visual anchor point sequence; Based on the comparison between the deviation of the adjacent anchor point interval from the average interval and the preset stability threshold, the step frequency stability of the terminal anchor point sequence and the visual anchor point sequence within the current analysis window is determined respectively.

5. The intention-based hierarchical management system based on online and offline integrated behavioral trajectories according to claim 1, characterized in that, The time calibration module is used to determine the update time difference, including: Obtain the terminal anchor sequence and visual anchor sequence within each analysis window; The terminal anchor sequence and the visual anchor sequence are discretized into corresponding terminal pulse sequences and visual pulse sequences; The relative time difference of each analysis window is determined based on the cross-correlation results of the terminal pulse sequence and the visual pulse sequence; The window weights of each analysis window are determined based on the peak intensity of cross-correlation and the continuity of anchor points. The relative time differences of each analysis window are then fused based on the window weights of each analysis window to obtain the global relative time difference. The update time difference is determined based on the current global relative time difference and the effective time difference that satisfies the stability condition in the previous analysis period.

6. The intention-based hierarchical management system based on online and offline integrated behavioral trajectories according to claim 1, characterized in that, When determining the user's intention score for the candidate target object, the intention calculation module includes determining the online interest score. Determine online interest scores, including: Based on the candidate target object, extract the online interaction event data associated with that candidate target object; The online interest score is generated based on the action weights corresponding to the online interactive event data and the time decay characteristics based on the recording time.

7. The intention-based hierarchical management system based on online and offline integrated behavioral trajectories according to claim 1, characterized in that, The intention calculation module is used to determine the offline contact score, including: Based on the candidate target object, extract the offline behavioral event data associated with that candidate target object; The offline baseline value is determined based on the action weight and action confidence corresponding to the offline action type. The offline decay value is determined based on the time difference between the event time of the offline behavioral event data and the end time of the current evaluation period; The offline contact score is generated based on the offline baseline value and the offline attenuation value.

8. The intention-based hierarchical management system based on online and offline integrated behavioral trajectories according to claim 1, characterized in that, The intention calculation module is used to determine the synchronization score, including: According to preset synchronization determination rules, candidate event pairs are selected from the calibrated online and offline behavior trajectories. The candidate event pairs meet the following conditions: The following conditions must be met: belonging to the same candidate target object; the time difference between the recording time of online interaction event data and the event time of offline behavior event data is less than the synchronization threshold; and the online action type and the offline action type satisfy the preset correspondence. The synchronous decay value of each candidate event pair is determined based on the time difference and the preset decay function; The synchronization score is obtained by summing the synchronization attenuation values ​​of each candidate event pair by combining the synchronization weights pre-configured by the event type combination.

9. The intention-based hierarchical management system based on online and offline integrated behavioral trajectories according to claim 1, characterized in that, The hierarchical management module is used to output corresponding service action instructions, including: The intention score and three preset stratification thresholds determine the first level, the second level, the third level or the fourth level, wherein the three preset stratification thresholds satisfy the following conditions in sequence: the first stratification threshold is less than the second stratification threshold, and the second stratification threshold is less than the third stratification threshold. The score changes are determined based on the intention scores of the current assessment cycle and the previous assessment cycle. Based on the judgment results of whether the user's intention rating for the candidate target object reaches the preset level and whether the change of the rating meets the preset conditions, it is determined whether to output a key attention mark; Based on the final determined level of intent, output corresponding service action instructions in the form of sales guide intervention, promotional push, experience reminder, continued observation, or routine record.

Images