Data analysis method and system based on user behavior sequence segmentation modeling

Abstract

The application provides a data analysis method and system based on user behavior sequence segmentation modeling, and relates to the technical field of data processing.The method comprises the following steps: acquiring multi-source heterogeneous original customer behavior data; performing quality cleaning on the original customer behavior data to obtain standardized customer behavior data; identifying a unified customer identifier from the standardized customer behavior data, dividing data with the same unified customer identifier into the same user data group; acquiring standard timestamps of each customer interaction record in the user data group, sorting the customer interaction records in the user data group according to the sequence of the standard timestamps to obtain sorted customer interaction records; sequentially connecting the sorted customer interaction records to form a behavior event list arranged in chronological order, and determining the behavior event list as a user behavior sequence.The application realizes unified collection and standardization of global behavior data of multi-service systems, and improves the accuracy of user behavior mining.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to a data analysis method and system based on user behavior sequence segmentation modeling. Background Technology

[0002] As enterprise information systems continue to improve, customer behavior data is typically stored across multiple heterogeneous data sources, such as sales systems, customer service systems, and website log systems. Current technologies for analyzing this multi-source heterogeneous data usually employ linear log aggregation to construct user behavior sequences, and then segment these sequences based on fixed time windows or simple event frequency thresholds for subsequent modeling and analysis.

[0003] However, existing technical solutions have the following drawbacks in practical applications:

[0004] For example, existing technologies often use fixed-length time windows or static event density thresholds to segment behavior fragments, ignoring the inherent semantic correlation and intensity variation characteristics between user behavior events. In real-world business scenarios, user behavior often exhibits a clustered distribution, with clear semantic boundaries between different behavior clusters. Fixed-window segmentation can easily forcibly cut off a complete semantic behavior process or incorrectly merge multiple unrelated behavior processes, resulting in a lack of semantic independence in the segmented behavior fragments. Existing segmentation methods typically treat behavior sequences as one-dimensional linear time data, failing to map the temporal and intensity attributes of behavior events into high-dimensional spatial structural features. Due to the lack of modeling the spatial morphology of behavior sequences, existing technologies may be unable to capture the volumetric variation characteristics of behavior density in multidimensional space, making it difficult to identify nodes where behavior patterns abruptly change. This leads to a lack of adaptive adjustment capability in determining segmentation boundaries, making it impossible to dynamically adjust the segmentation window size based on actual changes in behavior density, ultimately resulting in low accuracy of modeling and analysis results. Summary of the Invention

[0005] This invention provides a data analysis method and system based on user behavior sequence segmentation modeling, which realizes unified collection and standardization of behavior data across multiple business systems, and improves the accuracy of user behavior mining.

[0006] To solve the above-mentioned technical problems, the technical solution of the present invention is as follows:

[0007] Firstly, a data analysis method based on user behavior sequence segmentation modeling, the method comprising:

[0008] Step 1: Obtain raw customer behavior data from multiple heterogeneous sources, including customer interaction records from the sales system, customer service system, and website log system.

[0009] Step 2: Perform quality cleaning on the raw customer behavior data to obtain standardized customer behavior data;

[0010] Step 3: Identify a unified customer identifier from standardized customer behavior data, and divide data with the same unified customer identifier into the same user data group; obtain the standard timestamp of each customer interaction record in the user data group, sort the customer interaction records in the user data group according to the order of the standard timestamps, and obtain the sorted customer interaction records; connect the sorted customer interaction records in sequence to form a list of behavioral events arranged in chronological order, and determine the list of behavioral events as the user behavior sequence;

[0011] Step 4: Based on the user behavior sequence, extract the time attribute and behavior intensity attribute of each behavior event, and construct a geometric triangle with three consecutive behavior events as vertices; calculate the area value of each geometric triangle and use the area value as the spatial mapping weight; combine the time attribute and behavior intensity attribute to map the behavior event into spatial point cloud data in a three-dimensional coordinate system; based on the distribution density of the spatial point cloud data, stack along the time dimension to construct a behavior virtual model.

[0012] Step 5: In the behavior virtual model, select spatial point cloud data of adjacent time slices to construct triangular pyramidal geometric structures, calculate the volume values ​​of each triangular pyramidal geometric structure, and obtain the volume change sequence; identify the abrupt change nodes of volume values ​​in the volume change sequence, determine the time position corresponding to the abrupt change node as the segmentation boundary, and divide the user behavior sequence into multiple semantically independent behavior segments according to the segmentation boundary. The dynamic determination of the segmentation boundary includes adaptively adjusting the segmentation window size based on the threshold of the volume change rate.

[0013] Step 6: Perform modeling and analysis based on multiple semantically independent behavioral fragments to generate customer behavior analysis results. These results are used to characterize the accuracy of the analysis and modeling.

[0014] Furthermore, acquire multi-source heterogeneous raw customer behavior data, including customer interaction records from the sales system, customer service system, and website log system, including:

[0015] Establish connections with the sales system, the customer service system, and the website log system respectively through data interface calls or database synchronization.

[0016] Extract customer interaction records from the sales system, including transaction time, transaction amount, and product category;

[0017] Extract customer interaction records from the customer service system, including communication time, communication channel and service content, and extract customer interaction records from the website log system, including access time, access page path and dwell time.

[0018] The extracted customer interaction records from different systems are collected and initially integrated to form multi-source heterogeneous raw customer behavior data.

[0019] Furthermore, the raw customer behavior data undergoes quality cleaning to obtain standardized customer behavior data, including:

[0020] The integrity of the original customer behavior data from multiple sources and heterogeneous sources is verified, and abnormal records with missing key time attributes and user identifiers are removed to obtain the data after integrity verification. The key time attribute is a time field used to determine the time when the customer interaction record occurs and as the basis for time-series sorting.

[0021] Duplicate customer interaction records in the data after integrity verification are deduplicated to obtain the deduplicated data.

[0022] The time fields from different sources in the deduplicated data are uniformly converted into a standard timestamp format, and the user identifiers from different sources are mapped to a unified customer identifier, resulting in data with a unified format.

[0023] After standardizing the format, the data was determined to be standardized customer behavior data.

[0024] Furthermore, based on the user behavior sequence, the temporal and intensity attributes of each behavioral event are extracted, and three consecutive behavioral events are used as vertices to construct geometric triangles; the area of ​​each geometric triangle is calculated, and the area value is used as the spatial mapping weight; combining the temporal and intensity attributes, the behavioral events are mapped to spatial point cloud data in a three-dimensional coordinate system; based on the distribution density of the spatial point cloud data, a virtual behavioral model is constructed by stacking along the time dimension, including:

[0025] Traverse each behavioral event in the user behavior sequence, parse the standard timestamp of each behavioral event as the time attribute, and determine the behavior intensity attribute based on the duration of each behavioral event to obtain the time attribute and behavior intensity attribute of each behavioral event.

[0026] Based on the time and intensity attributes of each behavioral event, a sliding window method is used to select three consecutive behavioral events as vertices to construct a geometric triangle in the user behavior sequence. According to the relative positional relationship of the time and intensity attributes of the three consecutive behavioral events on the two-dimensional plane, the area value of each geometric triangle is calculated, and the area value is determined as the spatial mapping weight.

[0027] Based on time attributes, behavior intensity attributes, and spatial mapping weights, a three-dimensional coordinate system is constructed. The time attributes are mapped to the first coordinate axis data of the three-dimensional coordinate system, the behavior intensity attributes are mapped to the second coordinate axis data of the three-dimensional coordinate system, and the spatial mapping weights are mapped to the third coordinate axis data of the three-dimensional coordinate system, so as to obtain the spatial coordinates of each behavior event in the three-dimensional coordinate system.

[0028] The spatial coordinates of each behavioral event are collected to form spatial point cloud data in a three-dimensional coordinate system. The spatial point cloud data is divided along the first coordinate axis of the three-dimensional coordinate system to obtain multiple time slices, and the distribution density of the spatial point cloud data in each time slice is calculated.

[0029] Based on the order of the time slices and their corresponding distribution density, the time slices are stacked along the first coordinate axis to construct a behavioral virtual model.

[0030] Furthermore, in the behavior virtual model, spatial point cloud data from adjacent time slices are selected to construct triangular pyramidal geometric structures. The volume values ​​of each triangular pyramidal geometric structure are calculated to obtain a volume change sequence. Abrupt change nodes in the volume change sequence are identified, and the time positions corresponding to these nodes are determined as segmentation boundaries. Based on these segmentation boundaries, the user behavior sequence is divided into multiple semantically independent behavior segments. Dynamically determining the segmentation boundaries includes adaptively adjusting the segmentation window size based on a threshold of the volume change rate.

[0031] In the behavioral virtual model, spatial point cloud data of adjacent time slices are selected along the time dimension to obtain a set of spatial point cloud data of adjacent time slices.

[0032] Based on the spatial point cloud data set of adjacent time slices, feature points between adjacent time slices are selected as vertices to construct triangular pyramidal geometric structures, and the spatial coordinates of the vertices of each triangular pyramidal geometric structure are determined.

[0033] The volume values ​​of each triangular pyramid geometry are calculated based on the spatial coordinates of the vertices. The calculated volume values ​​are then arranged in chronological order according to the time slices to obtain a volume change sequence.

[0034] Based on the volume change sequence, the volume change rate between adjacent volume values ​​is calculated, and the volume change rate is compared with a preset change rate threshold to identify abrupt change nodes in the volume change sequence.

[0035] In the process of determining the time position corresponding to the mutation node as the segmentation boundary, the size of the segmentation window is adaptively adjusted according to the threshold of the volume change rate. If the volume change rate exceeds the change rate threshold, the size of the segmentation window is reduced to improve the positioning accuracy. If the volume change rate is lower than the change rate threshold, the size of the segmentation window is expanded to smooth the noise. The time position corresponding to the mutation node after adaptive adjustment is determined as the segmentation boundary.

[0036] Based on the segmentation boundaries, the user behavior sequence is divided into multiple semantically independent behavior segments.

[0037] Furthermore, modeling and analysis are performed based on multiple semantically independent behavioral fragments to generate customer behavior analysis results. These results are used to characterize the accuracy of the analytical modeling, including:

[0038] Obtain multiple semantically independent behavioral segments; perform modeling and analysis based on multiple semantically independent behavioral segments, and extract key features of each behavioral segment;

[0039] Customer behavior analysis results are generated based on key features, and these results include customer behavior pattern recognition information.

[0040] Customer behavior analysis results are used as evaluation metrics to characterize the accuracy of analytical modeling.

[0041] Secondly, a data analysis system based on user behavior sequence segmentation modeling includes:

[0042] The acquisition module is used to acquire multi-source heterogeneous raw customer behavior data, which includes customer interaction records from the sales system, customer service system, and website log system.

[0043] The cleaning module is used to perform quality cleaning on the raw customer behavior data to obtain standardized customer behavior data.

[0044] The module is used to identify a unified customer identifier from standardized customer behavior data, and to divide data with the same unified customer identifier into the same user data group; to obtain the standard timestamp of each customer interaction record in the user data group, and to sort the customer interaction records in the user data group according to the order of the standard timestamps, so as to obtain the sorted customer interaction records; to connect the sorted customer interaction records in sequence to form a list of behavioral events arranged in chronological order, and to determine the list of behavioral events as the user behavior sequence;

[0045] The mapping module is used to extract the time attributes and behavior intensity attributes of each behavior event based on the user behavior sequence, construct a geometric triangle with three consecutive behavior events as vertices; calculate the area value of each geometric triangle and use the area value as the spatial mapping weight; combine the time attributes and behavior intensity attributes to map the behavior events into spatial point cloud data in a three-dimensional coordinate system; and construct a behavior virtual model by stacking along the time dimension based on the distribution density of the spatial point cloud data.

[0046] The calculation module is used to construct triangular pyramidal geometric structures from spatial point cloud data of adjacent time slices in the behavior virtual model, calculate the volume values ​​of each triangular pyramidal geometric structure, and obtain the volume change sequence; identify the abrupt change nodes of volume values ​​in the volume change sequence, determine the time position corresponding to the abrupt change node as the segmentation boundary, and divide the user behavior sequence into multiple semantically independent behavior segments according to the segmentation boundary. The dynamic determination of the segmentation boundary includes adaptively adjusting the segmentation window size based on the threshold of the volume change rate.

[0047] The analysis module is used to perform modeling and analysis based on multiple semantically independent behavioral fragments, generating customer behavior analysis results. These results are used to characterize the accuracy of the analysis and modeling.

[0048] Thirdly, a computing device includes:

[0049] One or more processors;

[0050] A storage device for storing one or more programs that, when executed by one or more processors, cause the one or more processors to implement the method.

[0051] Fourthly, a computer-readable storage medium storing a program that, when executed by a processor, implements the method.

[0052] The above-described solution of the present invention has at least the following beneficial effects:

[0053] By acquiring multi-source heterogeneous raw customer behavior data and performing quality cleaning and multi-source data alignment, noisy data, logical conflicts, and temporal offsets and semantic gaps between different systems are eliminated, constructing continuous and standardized user behavior sequences and providing a highly reliable and continuous data foundation. The triangle area algorithm is used to map the temporal and intensity attributes of behavioral events into spatial point cloud data in a three-dimensional coordinate system, which can quantitatively capture the local correlation features between behavioral events and realize the structured representation of behavioral sequences in three-dimensional space.

[0054] By constructing a virtual behavioral model and calculating the triangular pyramidal volume change sequence of adjacent time slices, abrupt change nodes in volume values ​​are identified to determine the segmentation boundary. This achieves adaptive adjustment of the segmentation window size based on behavioral density changes, ensuring that the segmented behavioral fragments are semantically independent and solving the problem of semantic fragmentation or confusion caused by fixed window segmentation. Modeling and analysis based on semantically independent behavioral fragments improves the accuracy of data analysis and modeling. Attached Figure Description

[0055] Figure 1 This is a flowchart illustrating the data analysis method based on user behavior sequence segmentation modeling provided in an embodiment of the present invention.

[0056] Figure 2 This is a schematic diagram of a data analysis system based on user behavior sequence segmentation modeling provided in an embodiment of the present invention. Detailed Implementation

[0057] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.

[0058] like Figure 1 As shown, embodiments of the present invention propose a data analysis method based on user behavior sequence segmentation modeling, the method comprising the following steps:

[0059] Step 1: Obtain raw customer behavior data from multiple heterogeneous sources, including customer interaction records from the sales system, customer service system, and website log system.

[0060] Step 2: Perform quality cleaning on the raw customer behavior data to obtain standardized customer behavior data;

[0061] Step 3: Identify a unified customer identifier from standardized customer behavior data, and divide data with the same unified customer identifier into the same user data group; obtain the standard timestamp of each customer interaction record in the user data group, sort the customer interaction records in the user data group according to the order of the standard timestamps, and obtain the sorted customer interaction records; connect the sorted customer interaction records in sequence to form a list of behavioral events arranged in chronological order, and determine the list of behavioral events as the user behavior sequence;

[0062] Step 4: Based on the user behavior sequence, extract the time attribute and behavior intensity attribute of each behavior event, and construct a geometric triangle with three consecutive behavior events as vertices; calculate the area value of each geometric triangle and use the area value as the spatial mapping weight; combine the time attribute and behavior intensity attribute to map the behavior event into spatial point cloud data in a three-dimensional coordinate system; based on the distribution density of the spatial point cloud data, stack along the time dimension to construct a behavior virtual model.

[0063] Step 5: In the behavior virtual model, select spatial point cloud data of adjacent time slices to construct triangular pyramidal geometric structures, calculate the volume values ​​of each triangular pyramidal geometric structure, and obtain the volume change sequence; identify the abrupt change nodes of volume values ​​in the volume change sequence, determine the time position corresponding to the abrupt change node as the segmentation boundary, and divide the user behavior sequence into multiple semantically independent behavior segments according to the segmentation boundary. The dynamic determination of the segmentation boundary includes adaptively adjusting the segmentation window size based on the threshold of the volume change rate.

[0064] Step 6: Perform modeling and analysis based on multiple semantically independent behavioral fragments to generate customer behavior analysis results. These results are used to characterize the accuracy of the analysis and modeling.

[0065] In this embodiment of the invention, multi-source heterogeneous raw customer behavior data from sales systems, customer service systems, and website log systems are acquired, cleaned to obtain standardized data, and a continuous user behavior sequence is constructed based on the standardized data. Then, by extracting the time and intensity attributes of behavior events, constructing geometric triangles with three consecutive behavior events as vertices and using their areas as spatial mapping weights, and combining the above attributes, the behavior events are mapped to spatial point cloud data in a three-dimensional coordinate system and stacked along the time dimension to construct a virtual behavior model. Finally, in this model, triangular pyramids are constructed using spatial point cloud data of adjacent time slices, abrupt nodes are identified by calculating volume change sequences, and the segmentation window is adaptively adjusted according to the volume change rate threshold to determine the segmentation boundary. This process divides semantically independent behavior segments and performs modeling analysis. Therefore, this method overcomes the technical problems in existing user behavior data analysis, such as fragmented multi-source data, non-standardized data, discontinuous behavior sequences, difficulty in quantifying and representing behavior features, inaccurate positioning of sequence segmentation boundaries and inability to dynamically adapt to behavior changes, and insufficient accuracy in modeling analysis. This achieves the standardized integration and continuous sequence construction of multi-source customer behavior data, accurate quantification of behavior features and realization of visual spatial representation, dynamic and accurate division of semantically independent behavior segments, and improves the accuracy and reliability of customer behavior modeling analysis.

[0066] In a preferred embodiment of the present invention, step 1 above may include:

[0067] Step 1.1: Establish connections with the sales system, customer service system, and website log system respectively through data interface calls or database synchronization. Specifically, this includes: building data transmission paths with the three types of business systems based on standardized data interface technology, using data interface calls as the connection method, configuring interface communication protocols and access permissions according to general network communication specifications, and sequentially completing interface registration and link testing with the sales system, customer service system, and website log system to ensure stable initiation of data retrieval commands and achieve bidirectional data request and response interaction.

[0068] The integration method is database synchronization. The database data source storage address and access verification information of each business system are manually configured, the corresponding data synchronization mapping channel is built, and fixed data transmission verification rules are set to ensure that the original interactive data inside the database is output in real time without loss or tampering. Through the cooperation of the integration method, a stable connection with the sales system, customer service system, website log system is completely completed.

[0069] Step 1.2: Extract customer interaction records containing transaction time, transaction amount, and product category from the sales system. Specifically, this includes: after completing the connection setup of the sales system, selectively capturing the original customer interaction records retained within the system, focusing on extracting data containing three core information categories: transaction time, transaction amount, and product category; performing integrity verification calculations on each sales-related customer interaction record to be retained, setting the total number of core fields for a single record to three, corresponding to transaction time, transaction amount, and product category respectively; the field integrity of a single record is equal to the number of valid non-empty core fields in the single record divided by the total number of core fields specified, retaining only compliant customer interaction records with a field integrity calculation result of 100%; following the work mode of capturing and verifying in batches, completing the unified collection and organization of all compliant sales-related customer interaction records to form a dedicated original interaction data set for the sales system.

[0070] Step 1.3 involves extracting customer interaction records from the customer service system, including communication time, communication channel, and service content, and from the website log system, including access time, access page path, and dwell time. Specifically, this includes: First, data extraction is performed on the customer service system, collecting customer interaction information related to communication time, communication channel, and service content recorded within the system. Using a unified field completeness calculation method, the total number of core fields for customer service data is set to three. The completeness of a single record is obtained by dividing the number of valid non-empty core fields by the total number of core fields. Customer service communication interaction records that meet the completeness standard are then selected and independently aggregated. Next, data extraction is performed on the website log system, accurately obtaining information related to the three types of browsing behavior: access time, access page path, and dwell time. Completeness verification is also performed according to the three fixed core fields, eliminating invalid log records with missing key fields, and aggregating all compliant web browsing customer interaction records.

[0071] Step 1.4 involves collecting and initially integrating the extracted customer interaction records from different systems to form multi-source heterogeneous raw customer behavior data. Specifically, this includes: first, counting the total number of valid records collected from the three business systems. The total number of raw data collected is equal to the total number of compliant interaction records from the sales system, the customer service system, and the website log system. Using the original business time of each data entry as a temporary correlation benchmark, the three independent data sets from the sales system, customer service system, and website log system are fully aggregated. Preliminary fusion and arrangement of cross-system data is completed according to unified data splicing rules, eliminating the problem of fragmented data formats and scattered data entries between different heterogeneous systems. Finally, the integrated customer interaction data is uniformly archived and categorized to form multi-source heterogeneous raw customer behavior data covering all business scenarios including transaction communication and web browsing.

[0072] In this embodiment of the invention, data connections are established with the sales system, customer service system, and website log system through data interface calls or database synchronization. Core customer interaction records corresponding to each system are extracted in a targeted manner. Then, customer interaction records from different sources are uniformly collected and initially integrated to form multi-source heterogeneous original customer behavior data. Therefore, this technology overcomes the technical problems of existing technologies, such as single data source, fragmented and scattered storage of customer behavior data, inability to collect full-scenario interaction information, difficulty in reconstructing complete user behavior trajectories, and lack of basic analytical data dimensions. This enables the comprehensive collection and integration of customer behavior data for all business types, including transaction communication and web browsing, enriching the information dimensions of the original data and consolidating a complete and detailed foundation of original data with full coverage.

[0073] In a preferred embodiment of the present invention, step 2 above may include:

[0074] Step 2.1 involves performing integrity verification on the multi-source heterogeneous original customer behavior data, removing abnormal records lacking key time attributes and user identifiers to obtain the integrity-verified data. The key time attribute is a time field used to determine the occurrence time of customer interaction records and serves as the basis for time-series sorting. This key time attribute has multi-source heterogeneous characteristics, specifically defined as follows:

[0075] For customer interaction records from the sales system, the key time attribute refers to the transaction time;

[0076] For customer interaction records from the customer service system, the key time attribute refers to the communication time;

[0077] For customer interaction records from the website log system, the key time attribute refers to the access time.

[0078] The specific verification process includes:

[0079] The raw customer behavior data for integrity verification is divided into two categories: the first category consists of key time attribute fields, and the second category consists of user identification fields. The total number of key fields for overall verification is set to R = 2 (the sum of the two categories of fields). Each customer interaction record in the multi-source heterogeneous raw customer behavior data is traversed, and the number of complete and non-blank valid key fields in that record, b (within the range of 0, 1, or 2), is counted. The field completeness ratio of that record is calculated as P = b / R.

[0080] Threshold-based elimination: The integrity threshold is set to 1. If P < 1, the record is considered to have missing key time attributes or user identifiers, and is therefore classified as an abnormal record and eliminated directly. If P = 1, the record is considered to have compliant field information, and is temporarily retained and proceeds to the next processing stage. Through a process of verifying each record and eliminating records in batches, the integrity of all raw data is screened to eliminate the potential risks of data gaps, temporal disorder, and inability to associate user identities when constructing subsequent behavior sequences due to missing key time-series fields or identity identifiers.

[0081] Step 2.2 involves deduplicating duplicate customer interaction records in the data after integrity verification to obtain deduplicated data. Specifically, this includes: based on the compliant data retained after integrity verification, establishing a unified standard for determining duplicate records. If two or more customer interaction records have the same user identifier content, the same key time attribute content, and completely identical core interaction business content, then such records are identified as duplicate customer interaction records. The total number of compliant records to be deduplicated is counted, and then the total number of all duplicate records identified through screening is counted. The data duplication rate is calculated as the total number of all duplicate records divided by the total number of compliant records to be deduplicated. For all interaction records determined to be duplicates, only the earliest original single record generated by the data is retained, and all other redundant records with the same source are deleted, completing the simplification and optimization of the entire data. This processing eliminates the data superposition and redundancy problem generated during the synchronous collection of data from multiple systems.

[0082] Step 2.3 involves converting time fields from different sources in the deduplicated data into a unified standard timestamp format and mapping user identifiers from different sources to a unified customer identifier, resulting in data with a unified format. Specifically, this includes: standardizing the conversion of differentiated time representations retained by different business systems; converting and integrating all custom time formats from the sales system, customer service system, and website log system into a unified standard timestamp format; the cumulative value of the standard timestamp for a single record equals the annual fixed time base plus the monthly time difference plus the daily time difference, and finally summing the global time difference corresponding to the hour, minute, and second of the day, achieving complete uniformity in the time dimension of all data; and establishing a cross-system user identifier mapping ledger, matching and binding unique globally unified customer identifiers to the local user identifiers independently set by the sales system, customer service system, and website log system, completing the full-coverage replacement and rewriting of local identifiers within the entire data set to unified customer identifiers.

[0083] Step 2.4: After unifying the format, the data is determined as standardized customer behavior data. This includes: calculating the quantity changes throughout the entire data processing process, counting the total number of compliant records retained after integrity verification, and counting the total number of redundant and duplicate records removed during the deduplication process. The final total number of valid data records is equal to the total number of compliant records retained after integrity verification minus the total number of redundant and duplicate records removed during the deduplication process. All valid data that has completed integrity verification and removed abnormal records, completed deduplication and cleaned up redundant data, and completed time format unification and user identifier mapping rewriting are uniformly arranged according to a fixed field order. The overall data storage structure and field definitions are solidified, and the data is integrated to form standardized customer behavior data with a unified format, unique identity, and complete information.

[0084] In this embodiment of the invention, by sequentially performing integrity checks on multi-source heterogeneous raw customer behavior data to remove abnormal records lacking key time attributes and user identifiers, deduplicating duplicate customer interaction records, uniformly converting time formats from different sources to a standard timestamp format, and mapping user identifiers from different systems to a unified customer identifier, the technical means of ultimately determining the data after the above processing as standardized customer behavior data overcome the technical problems of multi-source heterogeneous raw customer behavior data having missing key information, duplicate records, inconsistent time formats and user identifiers, resulting in data being disorganized, unable to adapt to subsequent behavior sequence construction and modeling analysis, and easily interfering with the overall analysis accuracy. Thus, the standardization processing of multi-source heterogeneous customer behavior data is achieved, ensuring the integrity, uniqueness, and consistency of the data, and generating usable standardized customer behavior data.

[0085] In a preferred embodiment of the present invention, step 3 above, which involves identifying a unified customer identifier from standardized customer behavior data and grouping data with the same unified customer identifier into the same user data group, specifically includes:

[0086] After completing the full data standardization process, all standardized customer behavior data is retrieved, and the unified customer identifier embedded within the data is locked as the sole basis for aggregation and classification. Each standardized customer interaction record is identified and matched with the identifier, and the entire data is assigned to the corresponding unified customer identifier. All interaction content under the same unified customer identifier is aggregated. The total number of records in a single user data group is equal to the sum of the number of all valid customer interaction records associated with that unified customer identifier. Throughout the process, the mixing and superposition of behavior data from different users is completely eliminated, thus breaking the problem of multi-user data mixing in the original linear log aggregation mode and achieving precise splitting and aggregation from mixed data across the entire domain to exclusive data for a single user.

[0087] Obtain the standard timestamps of each customer interaction record in the user data group, and sort the customer interaction records in the user data group according to the order of the standard timestamps to obtain the sorted customer interaction records, specifically including:

[0088] For each segmented independent user data group, the standard timestamp value of each customer interaction record within the group is extracted one by one. The standard timestamp value is used as the sole time-series sorting reference standard. A basic sorting logic is set: the time difference between any two interaction records is equal to the standard timestamp value of the later interaction record minus the standard timestamp value of the earlier interaction record. If the calculated time difference is positive, the earlier record is determined to have occurred earlier; if the time difference is negative, the two records are swapped. According to this difference comparison method, a full-domain round-by-round comparison and sorting is performed on all customer interaction records within the entire user data group. The process is repeatedly verified and adjusted until all interaction records within the group are arranged in a strict chronological order, ultimately obtaining a sorted customer interaction record with a rigorous and coherent time-series logic.

[0089] The sorted customer interaction records are concatenated sequentially to form a list of behavioral events arranged in chronological order. This list of behavioral events is then defined as a sequence of user behaviors, specifically including:

[0090] All customer interaction records that have been sorted chronologically within a single user data group are sequentially linked together, without adding any extra blank data or redundant separation information. The total number of entries in the final behavioral event list is calculated, and the total number of entries in the behavioral event list is equal to the sum of the number of sorted customer interaction records in the corresponding user data group. After all interaction records are linked together, a complete, coherent, and standardized behavioral event list that strictly follows the time progression pattern is formed. This chronological behavioral event list is directly defined as a unique user behavior sequence.

[0091] In this embodiment of the invention, a technical means is used to identify a unified customer identifier from standardized customer behavior data and divide data with the same identifier into the same user data group. The standard timestamps of each interaction record in the group are extracted and sorted according to time sequence. The sorted interaction records are then connected in sequence to form a time-series list of behavioral events to determine the user behavior sequence. This overcomes the technical problems of standardized data being difficult to accurately collect according to user dimensions, the time sequence of single user behavior records being messy and scattered, and the inability to form a coherent and complete behavioral trajectory, which affects the accuracy of subsequent feature extraction and geometric modeling. As a result, it can accurately aggregate the full amount of single user behavior data and strictly follow the time logic to sort and connect them, constructing a user behavior sequence that is clear in time sequence, logically continuous, and complete closed loop.

[0092] In a preferred embodiment of the present invention, step 4 above may include:

[0093] Step 4.1: Traverse each behavioral event in the user behavior sequence, parse the standard timestamp of each behavioral event as a time attribute, and determine the behavior intensity attribute based on the duration of each behavioral event. This process specifically includes: retrieving the constructed user behavior sequence, traversing each item sequentially, parsing the attribute of each behavioral event, parsing the fixed standard timestamp in each behavioral event, and directly defining the standard timestamp value as the time attribute, denoted as... Ensure time attributes It is completely consistent with the unified standard timestamp format, without any additional conversion, ensuring the consistency of the time sequence dimension.

[0094] Based on the actual duration of each behavioral event, the behavioral intensity attribute is calculated and determined, denoted as . Define the duration of the behavioral event as (Unit: seconds), the behavior type weighting coefficient is... The behavioral intensity attribute is calculated using a fixed quantization formula: The weighting coefficient for behavior type is set to a fixed value based on the system to which the behavior event belongs: Transaction-related behaviors in the sales system =1.2, Customer service system communication behaviors =1.0, Browsing behavior in the website log system =0.8.

[0095] This formula quantifies the intensity of each behavioral event, ultimately matching each event with a corresponding time attribute T and behavioral intensity attribute. .

[0096] Step 4.2: Based on the time and intensity attributes of each behavioral event, three consecutive behavioral events are selected as vertices in the user behavior sequence using a sliding window method to construct geometric triangles. The area of ​​each geometric triangle is calculated based on the relative positional relationship of the time and intensity attributes of the three consecutive behavioral events on the two-dimensional plane, and the area value is determined as the spatial mapping weight. Specifically, this includes: obtaining the time attribute T and intensity attribute of all behavioral events... Then, a sliding window method is used to traverse and select the user behavior sequence. The step size of the sliding window is set to 1. That is, every time the window is moved, the first behavior event of the previous window is removed and the next new behavior event is added to ensure that there are always three consecutive behavior events in the window.

[0097] Three consecutive behavioral events are denoted as , , The corresponding time attribute is , , The corresponding behavior intensity attribute is , , Using the three behavioral events as vertices, a geometric triangle is constructed in the time-intensity two-dimensional plane. The area of ​​the triangle is calculated using the determinant formula, and the area value is defined as the spatial mapping weight, denoted as . The calculation formula is: Following the sliding window traversal method, the area of ​​the triangle formed by each group of three consecutive behavioral events is calculated sequentially, and a corresponding spatial mapping weight is assigned to each group of behavioral events. This quantifies the strength of the correlation between behavioral events, thus avoiding the limitation of existing technologies that cannot quantify behavioral correlations.

[0098] Step 4.3: Based on the time attribute, behavior intensity attribute, and spatial mapping weights, construct a three-dimensional coordinate system. Map the time attribute to the first coordinate axis data of the three-dimensional coordinate system, map the behavior intensity attribute to the second coordinate axis data of the three-dimensional coordinate system, and map the spatial mapping weights to the third coordinate axis data of the three-dimensional coordinate system to obtain the spatial coordinates of each behavior event in the three-dimensional coordinate system. Specifically, this includes: based on the time attribute... Behavioral intensity attribute Spatial mapping weights Construct a three-dimensional Cartesian coordinate system and define the three axes: the first coordinate axis (x-axis) corresponds to the time attribute. The second coordinate axis (y-axis) corresponds to the behavior intensity attribute. The third coordinate axis (z-axis) corresponds to the spatial mapping weight. .

[0099] The numerical range of the three axes is adjusted according to the extreme values ​​of the attributes of all behavioral events to ensure complete data presentation; the three-dimensional features of a single behavioral event are mapped to spatial coordinates, denoted as... The coordinate mapping formula is: ;

[0100] Each behavioral event is mapped to a coordinate system to ensure that each behavioral event has a unique spatial coordinate in the three-dimensional coordinate system, thereby realizing the transformation of one-dimensional temporal behavioral data into high-dimensional spatial structural features.

[0101] Step 4.4: Collect the spatial coordinates of each behavioral event to form spatial point cloud data in a three-dimensional coordinate system; divide the spatial point cloud data along the first coordinate axis of the three-dimensional coordinate system to obtain multiple time slices, and calculate the distribution density of the spatial point cloud data within each time slice. Specifically, this includes: collecting the spatial coordinates of all behavioral events... The entire domain is aggregated to form spatial point cloud data in a three-dimensional coordinate system. The total number of spatial point clouds is equal to the total number of behavioral events in the user's behavioral sequence.

[0102] Divide the time along the x-axis (time axis) and set fixed time intervals. =1800 seconds; Define the maximum time attribute in the sequence as The minimum time attribute is denoted as N, and the number of time slices is calculated using the following formula: , in the formula The rounding sign is used to ensure that the time slice completely covers all behavioral data.

[0103] After the partitioning is completed, the number of spatial point clouds within a single time slice is counted and denoted as . Define the spatial point cloud distribution density as The formula for calculating the distribution density is: The distribution density ρ is used to characterize the density of user behavior within the corresponding time slice, capturing the clustered distribution characteristics of user behavior.

[0104] Step 4.5: Based on the chronological order and corresponding distribution density of the time slices, stack the time slices along the first coordinate axis to construct the behavioral virtual model. Specifically, this includes: sorting all time slices by their starting timestamp values ​​in ascending order to ensure the timing matches the order in which user actions occur; defining the stacking height of the time slices as... Set the proportional coefficient =1000, the stacking height is positively correlated with the distribution density, and the calculation formula is: = × .

[0105] Following the sorted order, the time slices are stacked layer by layer along the x-axis (time axis), keeping the horizontal range of each slice consistent with the range of the y-axis and z-axis, forming a complete three-dimensional structure, and finally constructing a behavioral virtual model.

[0106] In this embodiment of the invention, the standard timestamps of each behavioral event are traversed and parsed as time attributes, and the behavioral intensity attribute is determined based on the duration of the event. A geometric triangle is constructed by selecting three consecutive behavioral events through a sliding window, and the area of ​​the triangle is calculated by combining the relative positions of the two-dimensional plane and set as the spatial mapping weight. A three-dimensional coordinate system is built to map the time attribute, behavioral intensity attribute, and spatial mapping weight to the corresponding coordinate axes to obtain the spatial coordinates of each behavioral event. The coordinates are summarized to form three-dimensional spatial point cloud data, and time slices are divided along the time axis. The point cloud distribution density within the slices is statistically analyzed. Then, the time slices are stacked according to the temporal sequence to construct a behavioral virtual model. Therefore, this method overcomes the technical problems of traditional user behavior temporal analysis, which relies only on one-dimensional data to represent features, cannot quantify the strength of the correlation between behaviors, finds it difficult to achieve spatial three-dimensional representation of abstract behaviors, and cannot intuitively reflect the distribution pattern of behaviors. This easily leads to weak subsequent feature extraction and insufficient recognition. In this way, fragmented linear user behaviors can be transformed into quantifiable and three-dimensionally presentable three-dimensional structured data, accurately labeling the behavioral correlation weights and behavioral clustering features at different time periods, and constructing a behavioral virtual model with rich feature dimensions and high spatial recognition.

[0107] In a preferred embodiment of the present invention, step 5 above may include:

[0108] Step 5.1: In the behavior virtual model, select spatial point cloud data of adjacent time slices along the time dimension to obtain a set of spatial point cloud data of adjacent time slices. Specifically, this includes: retrieving the completed behavior virtual model, which is a three-dimensional structure formed by stacking multiple time slices layer by layer along the time axis. All time slices have been sorted in chronological order. Along the time axis of the first coordinate axis of the three-dimensional coordinate system, select two adjacent time slices after sorting in sequence. That is, first select the first and second time slices, then select the second and third time slices, and so on, until the last two adjacent time slices are selected.

[0109] For each pair of adjacent time slices, extract all the spatial point cloud data contained within them. Integrate the spatial point cloud data of the two slices in the same adjacent slice group to form a spatial point cloud data set corresponding to the adjacent time slices. Each spatial point cloud data set contains all the spatial point cloud coordinates of the two adjacent slices.

[0110] Step 5.2: Based on the spatial point cloud data set of adjacent time slices, feature points between adjacent time slices are selected as vertices to construct triangular pyramidal geometric structures. The spatial coordinates of the vertices of each triangular pyramidal geometric structure are determined. Specifically, for each set of spatial point cloud data sets of adjacent time slices, feature point screening is carried out. The screening rules are clear and uniform to ensure that the feature points can accurately represent the behavioral distribution characteristics of the corresponding time slices. The specific screening method is as follows: from two adjacent time slices, three key feature points are selected respectively. The first feature point is the point corresponding to the position with the highest spatial point cloud distribution density in the slice. The second feature point is the spatial point with the smallest temporal attribute in the slice, i.e., the earliest occurrence. The third feature point is the spatial point with the largest temporal attribute in the slice, i.e., the latest occurrence. In this way, a total of 6 feature points are selected for each set of adjacent time slices. From these, 4 core feature points are selected as vertices to construct triangular pyramidal geometric structures. The selection rule is: select the feature point with the highest distribution density in the previous time slice and the 3 feature points in the next time slice. These 4 feature points together serve as the 4 vertices of the triangular pyramid to ensure that the triangular pyramid can reflect the behavioral density changes and temporal correlations between adjacent slices.

[0111] After determining the four vertices, the spatial coordinates corresponding to each feature point that has been calculated are directly retrieved to clarify the three-dimensional spatial coordinates of each vertex. The coordinate values ​​are completely consistent with the mapped spatial coordinates without any additional conversion, ensuring that the construction of the triangular pyramid geometry is based on accurate and unified spatial coordinate data.

[0112] Step 5.3: Calculate the volume values ​​of each triangular pyramid geometric structure based on the vertex spatial coordinates. Arrange the calculated volume values ​​according to the order of time slices to obtain the volume change sequence. Specifically, this includes: performing calculations based on the determined spatial coordinates of the four vertices, where the horizontal coordinate represents the time attribute and the vertical coordinate represents the behavior intensity attribute. Define the three feature points corresponding to the next time slice as the base vertices of the triangular pyramid. Calculate the area of ​​the base triangle using a unified cross-difference calculation method. First, multiply the differences between two sets of time attributes, then multiply the differences between two sets of behavior intensity attributes. Subtract the results of the two multiplications, take the absolute value of the difference obtained from the calculation, and divide the absolute value into two equal parts to obtain the actual area of ​​the base triangle.

[0113] The vertical height of the triangular pyramid is calculated by taking the vertical distance from the feature point with the highest distribution density in the previous time slice to the triangular plane formed by the three feature points in the next slice. During the calculation, the time attribute and behavior intensity attribute corresponding to the four vertices are cross-combined in turn. After integrating all the calculated values, the absolute value of the whole is taken. Then, the difference between the three sets of behavior intensity attributes is squared. The squared values ​​are summed up and the square root of the sum is taken. Finally, the absolute value obtained in the previous step is divided by the square root value to obtain the accurate vertical height.

[0114] After calculating the base area and vertical height, the base area and vertical height values ​​are multiplied together, and the resulting value is divided into three equal parts to obtain the volume of a single triangular pyramid. Following this calculation method, the volume of the triangular pyramid corresponding to each group of adjacent time slices is calculated one by one. Then, all the calculated volume results are arranged and linked together in strict accordance with the chronological order of adjacent time slices to form a complete volume change sequence.

[0115] Step 5.4: Based on the volume change sequence, calculate the volume change rate between adjacent volume values, compare the volume change rate with a preset change rate threshold, and identify abrupt change nodes in the volume values ​​in the volume change sequence. Specifically, this includes: relying on the generated volume change sequence, sequentially selecting two adjacent volume values ​​in the sequence to perform change rate calculations; the conventional calculation method is to subtract the volume value of the preceding position from the volume value of the following position, divide the difference obtained by subtraction by the original volume value of the preceding position, and then convert the final result into a percentage to obtain the corresponding volume change rate; if the volume value of the preceding position is empty, directly convert the volume value of the following position into a percentage and use it as the volume change rate of the current position.

[0116] A fixed comparison standard value is set throughout the process, with 30% as the basic judgment threshold. The calculated volume change rate is checked one by one. When the change rate deviates from the positive and negative range and the actual value exceeds 30%, a sudden volume change is determined at the corresponding time position and marked as a mutation node. When the actual value of the change rate is less than or equal to 30%, the volume change at the current position is considered stable, and no sudden change occurs. Through segment-by-segment verification and screening, all mutation nodes within the volume change sequence are comprehensively identified. Each marked mutation node corresponds to a precise time of occurrence of the behavior.

[0117] Step 5.5, in the process of determining the time position corresponding to the mutation node as the segmentation boundary, the segmentation window size is adaptively adjusted according to the threshold of the volume change rate. If the volume change rate exceeds the change rate threshold, the segmentation window size is reduced to improve positioning accuracy. If the volume change rate is lower than the change rate threshold, the segmentation window size is expanded to smooth noise. The time position corresponding to the mutation node after adaptive adjustment is determined as the segmentation boundary. Specifically, in the process of determining the time position corresponding to the mutation node as the segmentation boundary, the segmentation window size is adaptively adjusted according to the set change rate threshold of 30% to ensure the positioning accuracy of the segmentation boundary and smooth data noise. First, the initial size of the segmentation window is set to the length of one time slice, i.e., 1800 seconds, consistent with the fixed time interval of the time slice in step 4.4. The adjustment rules are clear: if the calculated volume change rate exceeds the preset 30% threshold, it indicates that the behavior at that location is changing drastically, and the positioning accuracy of the segmentation boundary needs to be improved. The segmentation window size is reduced to half of the initial window size, i.e., 1800 seconds ÷ 2 = 900 seconds. By reducing the window range, the specific time position corresponding to the mutation node can be accurately located, avoiding boundary offset due to an excessively large window. If the calculated volume change rate is lower than the preset 30% threshold, it indicates that the behavior at that location is changing gently, with data noise interference. The segmentation window size needs to be expanded to twice the initial window size, i.e., 1800 seconds × 2 = 3600 seconds. By expanding the window range, data noise is smoothed, avoiding misjudging normal behavior fluctuations as mutation nodes.

[0118] After the adjustment is completed, the time position corresponding to the mutation node after adaptive adjustment is used as the final segmentation boundary. Each mutation node corresponds to a segmentation boundary, and the time position of the segmentation boundary is completely consistent with the timestamp corresponding to the mutation node, ensuring that the segmentation boundary accurately corresponds to the mutation point of the user behavior pattern.

[0119] Step 5.6: Based on the segmentation boundaries, divide the user behavior sequence into multiple semantically independent behavior segments. Specifically, this includes: retrieving the constructed user behavior sequence, which is a list of continuous behavior events arranged in chronological order; combining this with all determined segmentation boundaries, splitting the user behavior sequence according to the following rules: using the time position corresponding to each segmentation boundary as the splitting point, the user behavior sequence is split from the splitting point, such that the behavior events between every two adjacent segmentation boundaries form an independent behavior segment; if there are behavior events between the start position of the sequence and the first segmentation boundary, they are treated as a separate behavior segment; if there are behavior events between the last segmentation boundary and the end position of the sequence, they are also treated as a separate behavior segment.

[0120] Each segmented behavior event is located between two adjacent segmentation boundaries, with a clearly defined time range. The behavior events within each segment have inherent semantic relationships, belonging to the same behavior cluster, and there is no mixing of unrelated behaviors. For example, if the first segmentation boundary timestamp is 1680001800 and the second segmentation boundary timestamp is 1680005400, then the behavior events between timestamps 1680000000 and 1680001800 are classified as the first behavior segment, and the behavior events between timestamps 1680001800 and 1680005400 are classified as the second behavior segment, and so on. Through this splitting method, the user behavior sequence is ultimately divided into multiple semantically independent behavior segments.

[0121] In this embodiment of the invention, spatial point cloud data of adjacent time slices are selected in the behavior virtual model. A triangular pyramid geometric structure is constructed based on the feature points between slices and the spatial coordinates of the vertices are determined. The volume of the triangular pyramid is calculated and arranged in time sequence to obtain a volume change sequence. The change rate of adjacent volumes is calculated and combined with a preset threshold to identify abrupt change nodes. At the same time, the size of the segmentation window is adaptively adjusted according to the volume change rate threshold. When the volume change rate exceeds the standard, the window is reduced to improve positioning accuracy, and when it is below the threshold, the window is expanded to smooth noise. Finally, the optimized abrupt change nodes are used to delineate the segmentation boundary and split into multiple semantically independent behavior segments. Therefore, this method overcomes the technical problems of traditional fixed time window segmentation, which cannot adapt to dynamic behavior fluctuations, is difficult to accurately capture business semantic inflection points, is easily affected by data noise, causing boundary positioning deviations, and has poor semantic correlation of segmented segments. Thus, it can accurately identify key turning points of user behavior based on three-dimensional geometric change features, taking into account segmentation accuracy and noise resistance, and efficiently segmenting independent behavior segments that fit the actual business logic and have clear semantic boundaries.

[0122] In a preferred embodiment of the present invention, step 6 above may include:

[0123] Step 6.1: Obtain multiple semantically independent behavior segments; perform modeling and analysis based on multiple semantically independent behavior segments, and extract key features of each behavior segment. Specifically, this includes: retrieving all semantically independent behavior segments divided in Step 5.6, and performing integrity verification on each behavior segment one by one to ensure that a single segment only contains behavior events within two adjacent segmentation intervals, with no data omissions or content duplication throughout the process, while ensuring that the behavior events within the segment have a unified semantic association, conforming to the clustered distribution pattern of actual user behavior.

[0124] Based on each independent behavioral segment, targeted modeling and analysis are carried out, and key features are defined into three fixed categories: transaction, communication, and browsing. The classification standards are consistent with those in the previous data collection and feature analysis stages, so as to achieve uniform terminology and data continuity throughout the process.

[0125] Define the number of times a single feature appears in the current action segment as The cumulative total number of occurrences of all key features within the current behavior segment is Construct the formula for calculating feature importance weights: The calculation is completed by substituting the values ​​one by one into the formula, and the screening threshold is set to 15%. When the feature importance weight meets the requirements... When the feature importance is ≥15%, it is designated as the core key feature of the current segment. When the percentage is less than 15%, the feature is identified as a redundant feature and removed directly; the core features of each behavioral segment are accurately extracted through this quantitative screening method.

[0126] Step 6.2: Generate customer behavior analysis results based on key features. The customer behavior analysis results include customer behavior pattern recognition information, specifically including: based on the core key features of each segment obtained by screening, conduct a comprehensive integrated analysis and combine feature weight values ​​to generate standardized analysis results with customer behavior pattern recognition information.

[0127] Define the cumulative total weight of the core features of the transaction class within a single behavior segment as follows: The cumulative total weight of the core features of the communication category is The cumulative total weight of the core features of the browsing category is The total weight of all core features in the current segment is The formulas for calculating the proportion of the three types of features are as follows: The sum of the three categories is fixed at 100%, and the core behavioral attributes of a single behavioral segment are defined based on the percentage values.

[0128] Further integrating all behavioral segment data, four categories of identification information were identified: user consumption preferences, browsing habits, consultation characteristics, behavioral conversion patterns, and so on. The number of core features overlapping between two adjacent behavioral segments was defined as... The total number of core features of the previous behavioral segment is Construct the formula for calculating feature correlation: The correlation criterion is set at 50%. When the percentage is ≥50%, it is determined that there is a direct business conversion relationship between the two behaviors. All quantitative analysis conclusions are integrated to form a complete customer behavior analysis result.

[0129] Step 6.3 uses the customer behavior analysis results as evaluation indicators to characterize the accuracy of the analysis and modeling. Specifically, it includes using the output customer behavior analysis results as the core evaluation basis, setting two quantitative indicators: modeling accuracy rate and feature deviation rate, to achieve closed-loop verification of the modeling effect throughout the entire process.

[0130] The number of identification information entries that correspond to the user's actual behavior is defined as follows: The total number of all behavior recognition information entries in the analysis results is The formula for calculating modeling accuracy is: ;

[0131] Define the feature weights obtained from model analysis as follows: The weight of the actual feature corresponding to the user's actual behavior is The formula for calculating the characteristic deviation rate is: ;

[0132] Set unified verification standards; when the modeling accuracy meets the requirements... ≥80% and the deviation rate of all key features meets the requirements. A modeling accuracy of ≤20% indicates that the overall modeling process is accurate and effective. If the modeling accuracy is lower than the limit or the deviation rate of any feature exceeds the limit, it is determined that there is an error in the modeling, and it is necessary to backtrack and complete the optimization and adjustment of the preliminary steps such as data cleaning and sequence segmentation. This quantitative evaluation system can intuitively represent the accuracy of the overall modeling and analysis.

[0133] In this embodiment of the invention, by acquiring multiple semantically independent behavioral segments and conducting specialized modeling analysis on each segment, accurately extracting key features of each segment, and generating exclusive customer behavior analysis results containing customer behavior pattern recognition information based on the extracted key features, and setting the generated customer behavior analysis results as evaluation indicators to reversely characterize the overall accuracy of the analysis and modeling, this technical approach overcomes the technical problems of traditional global integrated modeling, which is prone to being mixed with redundant and invalid behavioral information, has vague and scattered core feature extraction, low recognition of customer behavior patterns, and lacks a closed-loop verification link to make it impossible to intuitively determine the reliability of the modeling analysis. Thus, it can accurately mine the core value features of each subdivided behavioral segment, form clear and complete customer behavior pattern recognition conclusions, and also realize the autonomous verification and judgment of the modeling analysis effect throughout the entire process, ensuring the accuracy, effectiveness, and business practical value of the final customer behavior analysis results.

[0134] like Figure 2 As shown, embodiments of the present invention also provide a data analysis system based on user behavior sequence segmentation modeling, including:

[0135] The acquisition module is used to acquire multi-source heterogeneous raw customer behavior data, which includes customer interaction records from the sales system, customer service system, and website log system.

[0136] The cleaning module is used to perform quality cleaning on the raw customer behavior data to obtain standardized customer behavior data.

[0137] The module is used to identify a unified customer identifier from standardized customer behavior data, and to divide data with the same unified customer identifier into the same user data group; to obtain the standard timestamp of each customer interaction record in the user data group, and to sort the customer interaction records in the user data group according to the order of the standard timestamps, so as to obtain the sorted customer interaction records; to connect the sorted customer interaction records in sequence to form a list of behavioral events arranged in chronological order, and to determine the list of behavioral events as the user behavior sequence;

[0138] The mapping module is used to extract the time attributes and behavior intensity attributes of each behavior event based on the user behavior sequence, construct a geometric triangle with three consecutive behavior events as vertices; calculate the area value of each geometric triangle and use the area value as the spatial mapping weight; combine the time attributes and behavior intensity attributes to map the behavior events into spatial point cloud data in a three-dimensional coordinate system; and construct a behavior virtual model by stacking along the time dimension based on the distribution density of the spatial point cloud data.

[0139] The calculation module is used to construct triangular pyramidal geometric structures from spatial point cloud data of adjacent time slices in the behavior virtual model, calculate the volume values ​​of each triangular pyramidal geometric structure, and obtain the volume change sequence; identify the abrupt change nodes of volume values ​​in the volume change sequence, determine the time position corresponding to the abrupt change node as the segmentation boundary, and divide the user behavior sequence into multiple semantically independent behavior segments according to the segmentation boundary. The dynamic determination of the segmentation boundary includes adaptively adjusting the segmentation window size based on the threshold of the volume change rate.

[0140] The analysis module is used to perform modeling and analysis based on multiple semantically independent behavioral fragments, generating customer behavior analysis results. These results are used to characterize the accuracy of the analysis and modeling.

[0141] It should be noted that this system is a system corresponding to the above method. All implementation methods in the above method embodiments are applicable to this embodiment and can achieve the same technical effect.

[0142] Embodiments of the present invention also provide a computing device, including: a processor and a memory storing a computer program, wherein the computer program, when executed by the processor, performs the method described above. All implementations in the above method embodiments are applicable to this embodiment and can achieve the same technical effects.

[0143] Embodiments of the present invention also provide a computer-readable storage medium storing instructions that, when executed on a computer, cause the computer to perform the method described above. All implementations in the above method embodiments are applicable to this embodiment and can achieve the same technical effects.

[0144] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A data analysis method based on user behavior sequence segmentation modeling, characterized in that, The method includes: Step 1: Obtain raw customer behavior data from multiple heterogeneous sources, including customer interaction records from the sales system, customer service system, and website log system. Step 2: Perform quality cleaning on the raw customer behavior data to obtain standardized customer behavior data; Step 3: Identify a unified customer identifier from standardized customer behavior data, and divide data with the same unified customer identifier into the same user data group; obtain the standard timestamp of each customer interaction record in the user data group, sort the customer interaction records in the user data group according to the order of the standard timestamps, and obtain the sorted customer interaction records; connect the sorted customer interaction records in sequence to form a list of behavioral events arranged in chronological order, and determine the list of behavioral events as the user behavior sequence; Step 4: Traverse each behavioral event in the user behavior sequence, parse the standard timestamp of each behavioral event as the time attribute, and determine the behavior intensity attribute based on the duration of each behavioral event to obtain the time attribute and behavior intensity attribute of each behavioral event; based on the time attribute and behavior intensity attribute of each behavioral event, select three consecutive behavioral events as vertices in the user behavior sequence using a sliding window method to construct geometric triangles, calculate the area value of each geometric triangle based on the relative positional relationship of the time attribute and behavior intensity attribute of the three consecutive behavioral events on the two-dimensional plane, and determine the area value as the spatial mapping weight; Based on time attributes, behavior intensity attributes, and spatial mapping weights, a three-dimensional coordinate system is constructed. The time attributes are mapped to the first coordinate axis data of the three-dimensional coordinate system, the behavior intensity attributes are mapped to the second coordinate axis data of the three-dimensional coordinate system, and the spatial mapping weights are mapped to the third coordinate axis data of the three-dimensional coordinate system, so as to obtain the spatial coordinates of each behavior event in the three-dimensional coordinate system. The spatial coordinates of each behavioral event are collected to form spatial point cloud data in a three-dimensional coordinate system. The spatial point cloud data is divided along the first coordinate axis of the three-dimensional coordinate system to obtain multiple time slices, and the distribution density of the spatial point cloud data in each time slice is calculated. According to the chronological order of the time slices and their corresponding distribution densities, the time slices are stacked along the first coordinate axis to construct a behavioral virtual model. Step 5: In the behavior virtual model, spatial point cloud data of adjacent time slices are selected along the time dimension to obtain a set of spatial point cloud data of adjacent time slices; based on the set of spatial point cloud data of adjacent time slices, feature points between adjacent time slices are selected as vertices to construct triangular pyramidal geometric structures, and the spatial coordinates of the vertices of each triangular pyramidal geometric structure are determined; the volume values ​​of each triangular pyramidal geometric structure are calculated according to the spatial coordinates of the vertices, and the calculated volume values ​​are arranged according to the order of the time slices to obtain a volume change sequence; based on the volume change sequence, the volume change rate between adjacent volume values ​​is calculated, and the volume change rate is compared with a preset change rate threshold to identify abrupt change nodes in the volume change sequence; during the process of determining the time position corresponding to the abrupt change node as the segmentation boundary, the segmentation window size is adaptively adjusted according to the volume change rate threshold; if the volume change rate exceeds the change rate threshold, the segmentation window size is reduced to improve positioning accuracy; if the volume change rate is lower than the change rate threshold, the segmentation window size is expanded to smooth noise; the time position corresponding to the abrupt change node after adaptive adjustment is determined as the segmentation boundary; according to the segmentation boundary, the user behavior sequence is divided into multiple semantically independent behavior segments; Step 6: Perform modeling and analysis based on multiple semantically independent behavioral fragments to generate customer behavior analysis results. These results are used to characterize the accuracy of the analysis and modeling.

2. The data analysis method based on user behavior sequence segmentation modeling according to claim 1, characterized in that, Acquire multi-source, heterogeneous raw customer behavior data, including customer interaction records from the sales system, customer service system, and website log system, including: Establish connections with the sales system, the customer service system, and the website log system respectively through data interface calls or database synchronization. Extract customer interaction records from the sales system, including transaction time, transaction amount, and product category; Extract customer interaction records from the customer service system, including communication time, communication channel and service content, and extract customer interaction records from the website log system, including access time, access page path and dwell time. The extracted customer interaction records from different systems are collected and initially integrated to form multi-source heterogeneous raw customer behavior data.

3. The data analysis method based on user behavior sequence segmentation modeling according to claim 2, characterized in that, The raw customer behavior data is cleaned to obtain standardized customer behavior data, including: The integrity of the original customer behavior data from multiple sources and heterogeneous sources is verified, and abnormal records with missing key time attributes and user identifiers are removed to obtain the data after integrity verification. The key time attribute is a time field used to determine the time when the customer interaction record occurs and as the basis for time-series sorting. Duplicate customer interaction records in the data after integrity verification are deduplicated to obtain the deduplicated data. The time fields from different sources in the deduplicated data are uniformly converted into a standard timestamp format, and the user identifiers from different sources are mapped to a unified customer identifier, resulting in data with a unified format. After standardizing the format, the data was determined to be standardized customer behavior data.

4. The data analysis method based on user behavior sequence segmentation modeling according to claim 3, characterized in that, Modeling and analysis are performed based on multiple semantically independent behavioral fragments to generate customer behavior analysis results. These results are used to characterize the accuracy of the analytical modeling, including: Obtain multiple semantically independent behavioral segments; perform modeling and analysis based on multiple semantically independent behavioral segments, and extract key features of each behavioral segment; Customer behavior analysis results are generated based on key features, and these results include customer behavior pattern recognition information. Customer behavior analysis results are used as evaluation metrics to characterize the accuracy of analytical modeling.

5. A data analysis system based on user behavior sequence segmentation modeling, wherein the system implements the method as described in any one of claims 1 to 4, characterized in that, include: The acquisition module is used to acquire multi-source heterogeneous raw customer behavior data, which includes customer interaction records from the sales system, customer service system, and website log system. The cleaning module is used to perform quality cleaning on the raw customer behavior data to obtain standardized customer behavior data. The module is used to identify a unified customer identifier from standardized customer behavior data, and to divide data with the same unified customer identifier into the same user data group; to obtain the standard timestamp of each customer interaction record in the user data group, and to sort the customer interaction records in the user data group according to the order of the standard timestamps, so as to obtain the sorted customer interaction records; to connect the sorted customer interaction records in sequence to form a list of behavioral events arranged in chronological order, and to determine the list of behavioral events as the user behavior sequence; The mapping module is used to extract the time attributes and intensity attributes of each behavior event based on the user behavior sequence, and construct a geometric triangle using three consecutive behavior events as vertices. Calculate the area of ​​each geometric triangle and use the area value as the spatial mapping weight; By combining temporal and behavioral intensity attributes, behavioral events are mapped to spatial point cloud data in a three-dimensional coordinate system. Based on the distribution density of spatial point cloud data, a behavioral virtual model is constructed by stacking along the time dimension. The calculation module is used to construct triangular pyramidal geometric structures from spatial point cloud data of adjacent time slices in the behavior virtual model, calculate the volume values ​​of each triangular pyramidal geometric structure, and obtain the volume change sequence; identify the abrupt change nodes of volume values ​​in the volume change sequence, determine the time position corresponding to the abrupt change node as the segmentation boundary, and divide the user behavior sequence into multiple semantically independent behavior segments according to the segmentation boundary. The dynamic determination of the segmentation boundary includes adaptively adjusting the segmentation window size based on the threshold of the volume change rate. The analysis module is used to perform modeling and analysis based on multiple semantically independent behavioral fragments, generating customer behavior analysis results. These results are used to characterize the accuracy of the analysis and modeling.

6. A computing device, characterized in that, include: One or more processors; A storage device for storing one or more programs that, when executed by one or more processors, cause the one or more processors to implement the method as described in any one of claims 1 to 4.

7. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a program that, when executed by a processor, implements the method as described in any one of claims 1 to 4.

Images