A user behavior attribution method and related products

Abstract

The application discloses a user behavior attribution method and related products. In the scheme, based on the multi-modal behavior data of multiple users within a preset time window, the behavior sequence of each user is constructed; based on the behavior sequence of each user, the corresponding natural language semantic label is determined, and the corresponding semantic vector is converted from the corresponding natural language semantic label of multiple users; the semantic vectors of multiple users in the current time window and the semantic vectors in the historical time window are respectively clustered to obtain the clustering results of the current time window and the clustering results of the historical time window, and the clustering results of the two time windows are compared to determine the semantic drift trajectory of each user group; based on the semantic drift trajectory of multiple user groups, the contribution of each user group to the fluctuation of the preset business index is quantified to obtain the quantification result and generate an attribution report. Compared with the existing technology, the precision of user behavior attribution is low, and the application has obvious advantages.

Description

Technical Field

[0001] This application relates to the field of data processing technology, and in particular to a user behavior attribution method and related products. Background Technology

[0002] In industries that rely on user operations, such as e-commerce, finance, and retail, accurate attribution of user behavior is a core support for companies to optimize operational strategies, enhance user value, and cope with fluctuations in business metrics.

[0003] Traditional user behavior attribution methods often rely on manually preset rules or simple statistical models to group users, such as dividing users into tiers based on recent purchase time, purchase frequency, and purchase amount, or classifying users through fixed tags, such as active users or churned users. However, these methods can only achieve a superficial division of user behavior results and cannot capture the deeper semantic motivations behind user behavior. This leads to a superficial interpretation of changes in user behavior and makes it difficult to meet enterprises' needs for accurate attribution of fluctuations in user behavior.

[0004] Improving the accuracy of user behavior attribution is a technical problem that urgently needs to be solved. Summary of the Invention

[0005] To address the aforementioned issues, this application provides a user behavior attribution method and related products, with the aim of improving the accuracy of user behavior attribution.

[0006] The embodiments of this application disclose the following technical solutions:

[0007] The first aspect of this application provides a user behavior attribution method, the method comprising:

[0008] Based on multimodal behavior data of multiple users within a preset time window, a behavior sequence for each user is constructed; the preset time window includes the current time window and historical time windows.

[0009] Based on the behavior sequence of each user, the corresponding natural language semantic tags are determined, and the natural language semantic tags corresponding to multiple users are converted into corresponding semantic vectors respectively;

[0010] The semantic vectors of multiple users in the current time window and the semantic vectors of the same multiple users in historical time windows are clustered to obtain the clustering results of the current time window and the clustering results of the historical time windows. The semantic drift trajectory of each user group is determined by comparing the clustering results of the current time window and the clustering results of the historical time windows. The clustering results contain multiple user groups.

[0011] Based on the semantic drift trajectories of multiple user groups, the contribution of each user group to the fluctuation of preset business indicators is quantified to obtain the quantification results, and an attribution report is generated based on the quantification results.

[0012] Optionally, determining the corresponding natural language semantic label based on the behavior sequence of each user includes:

[0013] Based on the behavior sequence of each user, a hierarchical update mechanism is used to determine the corresponding natural language semantic label; the hierarchical update mechanism includes atomic feature extraction and natural language semantic label determination.

[0014] Optionally, determining the corresponding natural language semantic label based on the behavior sequence of each user using a hierarchical update mechanism includes:

[0015] For any user:

[0016] Atomic features of the user's behavior sequence in the current time window and atomic features of the user's behavior sequence in the historical time window are extracted respectively to obtain the atomic feature extraction results of the user in the current time window and the atomic feature extraction results of the user in the historical time window.

[0017] If the user's behavior changes compared to the atomic feature extraction results in the current time window and the atomic feature extraction results in the historical time window, then the corresponding natural language semantic label is determined based on the atomic feature extraction results in the current time window using the large language model.

[0018] If the user's behavior has not changed compared to the atomic feature extraction results in the current time window and the atomic feature extraction results in the historical time window, then the corresponding historical natural language semantic label is used; the historical natural language semantic label is determined based on the user's atomic feature extraction results in the historical time window.

[0019] Optionally, the step of comparing the clustering results of the current time window and the clustering results of the historical time windows to determine the semantic drift trajectory of each user group includes:

[0020] For any user group:

[0021] Based on the semantic vectors of multiple users included in the user group, determine the center vector of the user group in the current time window;

[0022] Based on the semantic vectors of multiple users included in the user group, determine the center vector of the user group in the historical time window;

[0023] By comparing the center vector of the user group in the current time window with the center vector of the user group in historical time windows, the semantic drift trajectory of the user group is determined.

[0024] Optionally, determining the semantic drift trajectory of the user group by comparing the center vector of the user group in the current time window with the center vector of the user group in historical time windows includes:

[0025] Calculate the bulldozer distance between the center vector of the user group in the current time window and the center vector of the user group in the historical time window, obtain the bulldozer distance calculation result, and determine the semantic drift trajectory of the user group based on the bulldozer distance calculation result.

[0026] Optionally, determining the semantic drift trajectory of the user group by comparing the center vector of the user group in the current time window with the center vector of the user group in historical time windows includes:

[0027] Calculate the cosine similarity between the center vector of the user group in the current time window and the center vector of the user group in the historical time window, obtain the cosine similarity calculation result, and determine the semantic drift trajectory of the user group based on the cosine similarity calculation result.

[0028] Optionally, the contribution of each user group to the fluctuation of preset business indicators is quantified based on the semantic drift trajectories of multiple user groups to obtain quantification results, including:

[0029] Based on the semantic drift trajectories of multiple user groups, the contribution of each user group to the fluctuation of preset business indicators is quantified using the Shapley value attribution algorithm, and the quantification results are obtained.

[0030] A second aspect of this application provides a user behavior attribution device, the device comprising:

[0031] The behavior sequence construction module is used to construct the behavior sequence of each user based on the multimodal behavior data of multiple users within a preset time window; the preset time window includes the current time window and the historical time window;

[0032] The semantic label and semantic vector determination module is used to determine the corresponding natural language semantic label based on the behavior sequence of each user, and to convert the natural language semantic labels corresponding to multiple users into corresponding semantic vectors respectively;

[0033] The semantic drift trajectory determination module is used to cluster the semantic vectors of multiple users in the current time window and the semantic vectors of the same multiple users in historical time windows, respectively, to obtain the clustering results of the current time window and the clustering results of the historical time windows, and to compare the clustering results of the current time window and the clustering results of the historical time windows to determine the semantic drift trajectory of each user group; the clustering results contain multiple user groups.

[0034] The attribution report determination module is used to quantify the contribution of each user group to the fluctuation of preset business indicators based on the semantic drift trajectory of multiple user groups, obtain the quantification results, and generate an attribution report based on the quantification results.

[0035] A third aspect of this application provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements a user behavior attribution method as provided in any implementation of the first aspect.

[0036] Compared with the prior art, this application has the following beneficial effects:

[0037] The user behavior attribution method provided in this application includes: constructing a behavior sequence for each user based on multimodal behavior data of multiple users within a preset time window; the preset time window includes a current time window and a historical time window; determining corresponding natural language semantic tags based on the behavior sequence of each user, and converting the natural language semantic tags corresponding to multiple users into corresponding semantic vectors; clustering the semantic vectors of multiple users in the current time window and the semantic vectors of multiple users in the historical time window to obtain the clustering results of the current time window and the clustering results of the historical time window, and comparing the clustering results of the current time window and the clustering results of the historical time window to determine the semantic drift trajectory of each user group; the clustering results include multiple user groups; quantifying the contribution of each user group to the fluctuation of preset business indicators based on the semantic drift trajectories of multiple user groups to obtain quantification results, and generating an attribution report based on the quantification results.

[0038] The aforementioned method constructs complete user behavior sequences using multimodal behavioral data, avoiding the attribution bias caused by relying on single-dimensional data. Transforming user behavior sequences into interpretable natural language semantic labels prevents attribution from remaining merely on the surface of the data, helping to capture the deeper semantic motivations behind user behavior. Converting natural language semantic labels into semantic vectors ensures that subsequent clustering accurately reflects the semantic similarity of user groups, resulting in more precise attribution results. By clustering the semantic vectors of the current time window and historical time windows separately, and then comparing the clustering results of the two time windows, the semantic drift trajectory of user groups over time can be accurately identified, providing a reliable data foundation for quantifying the contribution of user groups. Quantifying the contribution of each user group to the fluctuation of preset business indicators based on semantic drift trajectories clearly distinguishes the impact of different user groups on the fluctuation of preset business indicators, ensuring that the attribution is based on the impact of changes in user behavior on business indicators. This avoids the bias caused by misjudging accidental correlations as causal relationships, guaranteeing the accuracy of the attribution results. Attached Figure Description

[0039] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0040] Figure 1 A flowchart of a user behavior attribution method provided in an embodiment of this application;

[0041] Figure 2 A schematic diagram illustrating the evolution and attribution of a user group as provided in an embodiment of this application;

[0042] Figure 3 This application provides a panoramic schematic diagram of customer flow during a major sales promotion.

[0043] Figure 4 This is a schematic diagram of a user behavior attribution device provided in an embodiment of this application. Detailed Implementation

[0044] As described earlier, current user behavior attribution methods mostly rely on manually preset rules or simple statistical models to group users, such as dividing users into tiers based on recent purchase time, purchase frequency, and purchase amount, or classifying users through fixed tags, such as active users or churned users. However, these methods can only achieve a superficial division of user behavior results and cannot capture the deep semantic motivations behind user behavior. This leads to a superficial interpretation of changes in user behavior and makes it difficult to meet enterprises' needs for accurate attribution of fluctuations in user behavior.

[0045] In view of the above problems, this application proposes a user behavior attribution method and related products. Based on multimodal behavior data of multiple users within a preset time window, a behavior sequence for each user is constructed. The preset time window includes a current time window and a historical time window. Based on the behavior sequence of each user, corresponding natural language semantic tags are determined, and the natural language semantic tags corresponding to multiple users are converted into corresponding semantic vectors. The semantic vectors of multiple users in the current time window and the semantic vectors of multiple users in the historical time window are clustered to obtain the clustering results for the current time window and the historical time window. The clustering results of the current time window and the historical time window are compared to determine the semantic drift trajectory of each user group. The clustering results contain multiple user groups. Based on the semantic drift trajectories of multiple user groups, the contribution of each user group to the fluctuation of preset business indicators is quantified to obtain quantification results, and an attribution report is generated based on the quantification results.

[0046] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present application, and not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present application.

[0047] See Figure 1 This figure is a flowchart of a user behavior attribution method provided in an embodiment of this application. Figure 1 As shown, the method includes the following steps:

[0048] S101. Based on the multimodal behavior data of multiple users within a preset time window, construct the behavior sequence of each user.

[0049] The preset time window includes the current time window and the historical time window.

[0050] The multimodal behavioral data here includes transaction data, interaction data, and external context data. Transaction data includes order amounts, etc.; interaction data includes browsing logs and customer service conversations, etc.; external context data includes competitor activities, etc.

[0051] Based on the preprocessing and serialization of multimodal behavioral data from multiple users within a preset time window, the behavioral sequence for each user is obtained by timestamp. For example, multimodal behavioral data of users over the past 30 days is extracted from the data warehouse, including order table data, event tracking table data, and customer service records. The order table data reflects user purchase information, such as a user purchasing product A; the event tracking table data reflects products viewed but not ordered, such as a user viewing product B 3 times without placing an order; and the customer service records represent user conversations with customer service, such as a user inquiring about slow delivery. Using a preset JSON template, the multimodal behavioral data is transformed into structured behavioral sequences, such as {"Action": "Browse", "Object": "Product B", "Frequency": "High", "Result": "Not Converted"}, and {"Action": "Inquiry", "Content": "Delivery Timeliness", "Mood": "Anxiety"}.

[0052] S102. Based on the behavior sequence of each user, determine the corresponding natural language semantic label, and convert the natural language semantic labels of multiple users into corresponding semantic vectors respectively.

[0053] In one feasible implementation, based on each user's behavioral sequence, the corresponding natural language semantic tags are determined, including:

[0054] Based on each user's behavior sequence, a hierarchical update mechanism is used to determine the corresponding natural language semantic labels; the hierarchical update mechanism includes atomic feature extraction and natural language semantic label determination.

[0055] Atomic feature extraction can be achieved using lightweight models. Natural semantic labels can be determined using large language models.

[0056] In one feasible implementation, based on each user's behavior sequence, a hierarchical update mechanism is used to determine the corresponding natural language semantic tags, including:

[0057] For any user:

[0058] Atomic features of the user's behavior sequence in the current time window and atomic features of the user's behavior sequence in the historical time window are extracted respectively to obtain the atomic feature extraction results of the user in the current time window and the atomic feature extraction results of the user in the historical time window.

[0059] If the user's behavior changes compared to the atomic feature extraction results in the current time window and the atomic feature extraction results in the historical time window, then the corresponding natural language semantic label is determined based on the atomic feature extraction results in the current time window using a large language model.

[0060] If the user's behavior has not changed compared to the atomic feature extraction results in the current time window and the atomic feature extraction results in the historical time window, then the corresponding historical natural language semantic label is used; the historical natural language semantic label is determined based on the user's atomic feature extraction results in the historical time window.

[0061] For example, the constructed prompt words are input into the large language model based on the preset prompt word template. For example, from the perspective of a senior user researcher, the user's natural language semantic label is generated based on the user's behavior sequence, which includes three dimensions: purchasing power, preferences and potential risks. Data fabrication is prohibited. The result output by the large language model is that the user has medium to high purchasing power, has a strong intention to purchase product B but is in the price comparison stage, and is currently at risk of being lost due to logistics timeliness issues.

[0062] Here, the large language model only triggers the regeneration of natural language semantic labels for users whose behavior has changed. For users whose behavior has not changed, the natural language semantic labels are reused. The hierarchical update mechanism reduces the consumption of computing power.

[0063] This section utilizes a pre-trained semantic encoding model to transform the user's natural language labels into high-dimensional semantic vectors.

[0064] S103. Cluster the semantic vectors of multiple users in the current time window and the semantic vectors of multiple users in the historical time window respectively to obtain the clustering results of the current time window and the clustering results of the historical time window. Compare the clustering results of the current time window and the clustering results of the historical time window to determine the semantic drift trajectory of each user group.

[0065] The clustering results include multiple user groups.

[0066] In one feasible implementation, the semantic drift trajectory of each user group is determined by comparing the clustering results of the current time window and the clustering results of the historical time windows, including:

[0067] For any user group:

[0068] Based on the semantic vectors of multiple users included in the user group, the center vector of the user group in the current time window is determined.

[0069] Based on the semantic vectors of multiple users included in the user group, the center vector of the user group in the historical time window is determined.

[0070] By comparing the center vector of the user group in the current time window with the center vector of the user group in historical time windows, the semantic drift trajectory of the user group is determined.

[0071] For example, T1 is the historical time window and T2 is the current time window. After clustering the user group in these two time windows, the clustering results are compared and it is found that the central semantic of the user group in T1 is price sensitivity and the central semantic of T2 is service dissatisfaction. This indicates that the core focus of the user group has shifted from price to service experience, and a significant semantic drift has occurred.

[0072] In one feasible implementation, determining the semantic drift trajectory of the user group by comparing the center vector of the user group in the current time window with the center vector of the user group in historical time windows includes:

[0073] Calculate the bulldozer distance between the center vector of the user group in the current time window and the center vector of the user group in the historical time window, obtain the bulldozer distance calculation result, and determine the semantic drift trajectory of the user group based on the bulldozer distance calculation result.

[0074] In one feasible implementation, determining the semantic drift trajectory of the user group by comparing the center vector of the user group in the current time window with the center vector of the user group in historical time windows includes:

[0075] Calculate the cosine similarity between the center vector of the user group in the current time window and the center vector of the user group in the historical time window, obtain the cosine similarity calculation result, and determine the semantic drift trajectory of the user group based on the cosine similarity calculation result.

[0076] In one feasible implementation, the semantic vectors of multiple users in the current time window and the semantic vectors of the multiple users in historical time windows are clustered respectively to obtain the clustering results for the current time window and the clustering results for the historical time windows, including:

[0077] Unsupervised density clustering is performed on the semantic vectors of multiple users in the current time window and the semantic vectors of the same multiple users in the historical time window, respectively, to obtain the clustering results for the current time window and the clustering results for the historical time window.

[0078] S104. Based on the semantic drift trajectories of multiple user groups, quantify the contribution of each user group to the fluctuation of preset business indicators, obtain the quantification results, and generate an attribution report based on the quantification results.

[0079] In one feasible implementation:

[0080] Based on the semantic drift trajectories of multiple user groups, the contribution of each user group to the fluctuation of preset business indicators is quantified using the Shapley value attribution algorithm, and the quantification results are obtained.

[0081] Preset macro business metrics, such as the fluctuation value of 30-day retention rate, are decomposed into various user groups that have experienced semantic drift using the Shapley value attribution algorithm. The Shapley value quantifies the marginal contribution of each user drift group to the fluctuation of business metrics, ensuring that the sum of the contributions of all user groups equals the total fluctuation value of business metrics, thus mathematically guaranteeing the fairness and accuracy of attribution.

[0082] We use a large language model to summarize the semantic changes of key drifting groups and generate attribution reports in natural language form. For example: This week's Gross Merchandise Volume (GMV) decreased by 5%, mainly attributed to user group C. Phenomenon: This user group showed a significant semantic drift from hesitant price comparison to service dissatisfaction. Reason: High-frequency word analysis showed a surge in the weight of slow delivery and unresponsive customer service. Recommendation: Send logistics coupons to this user group and prioritize customer service.

[0083] Natural language attribution reports generated by large language models directly link technical data with business logic, solving the pain point of traditional user behavior attribution methods that cannot explain behavioral motivations, and achieving white-boxing of user behavior fluctuations. Attribution reports can directly guide the operations team to formulate precise strategies, such as launching matching promotional activities for loyal users lost due to competitors' low prices, thereby improving operational efficiency and business metrics.

[0084] like Figure 2 As shown, Figure 2 This is a schematic diagram of user group evolution and attribution provided in an embodiment of this application, illustrating how dynamic changes in the user group cause fluctuations in the core business metric R30, where R30 represents the 30-day retention rate. Figure 2The study compared the user distribution across two time windows: T1, representing the historical baseline, and T2, representing the current fluctuation period. The R30 index was normal at 45% in T1, while it decreased by 15% to 30% in T2. ​​The red dots represent high-value, stable customers. This high concentration in both time windows indicates relatively stable retention rates and behavioral characteristics, and is not the cause of the R30 decline. The blue dots represent price-sensitive, fluctuating customers, which also maintained their original size without significant churn. The large-scale disappearance of high-value, stable customers in T2 is the direct cause of the overall R30 decline. Temporal clustering revealed that the semantic features of high-value, stable customers shifted from their product focus in T1 to their focus on low-priced competitors in T2. ​​Quantitative calculations confirmed that this group's churn had the largest marginal contribution to the R30 decline, concluding that the original high-value, stable customer base was lost due to the impact of low-priced competitors. Figure 2 Transforming abstract user group changes and attribution logic into intuitive point cluster distribution changes allows business personnel to quickly understand which groups have changed, how these changes affect core metrics, and the specific reasons behind them. This lowers the barrier to understanding technical conclusions and allows attribution results to be implemented into operational actions more quickly.

[0085] By capturing changes in user behavior characteristics through time-series differential analysis and directly linking these changes to fluctuations in business metrics, we can accurately identify the core user groups and reasons affecting the business, which helps support precise strategy formulation.

[0086] like Figure 3 As shown, Figure 3 This is a panoramic diagram of customer flow during a major promotional event, provided in an embodiment of this application. CustomerFlow represents customer flow. Figure 3This chart exemplifies the changes in customer flow, key factors contributing to performance fluctuations, and corresponding business decision recommendations during e-commerce promotional events. The final net GMV loss for this promotion was 140,000, with the loss of high-value customers directly contributing 580,000. The left side of the chart uses different colored lines and blocks to illustrate the migration paths of different customer groups during the promotion. The red line represents the loss of high-net-worth customers due to unsegmented or diluted benefits, while the green line represents the successful conversion of price-sensitive customers due to the bonus strategy. The performance fluctuation contribution chart on the right quantifies the net impact of the promotion strategy on GMV for different customer segments. The 56.9% failure rate for high-value customers indicates the loss of loyal high-diamond or high-value customers who left due to dissatisfaction with their benefits, directly causing a performance loss of 138,000, which is the primary reason for the overall GMV decline of 140,000. The 41.2% success rate indicates the growth of price-sensitive, hesitant customers who converted through bonuses, bringing some performance gains and partially offsetting the negative impact of the high-value customer loss. Based on data analysis, a decision-making suggestion was given: a general benefit and free gift strategy, where a lot of samples were given with the purchase of full-size products. This led high-value users to feel that the benefits were not differentiated and that the brand was cheapened. High-V users felt that they had been taken advantage of, which triggered a retaliatory return of goods.

[0087] The user behavior attribution method provided in this application includes: constructing a behavior sequence for each user based on multimodal behavior data of multiple users within a preset time window; the preset time window includes a current time window and a historical time window; determining corresponding natural language semantic tags based on the behavior sequence of each user, and converting the natural language semantic tags corresponding to multiple users into corresponding semantic vectors; clustering the semantic vectors of multiple users in the current time window and the semantic vectors of multiple users in the historical time window to obtain the clustering results of the current time window and the clustering results of the historical time window, and comparing the clustering results of the current time window and the clustering results of the historical time window to determine the semantic drift trajectory of each user group; the clustering results include multiple user groups; quantifying the contribution of each user group to the fluctuation of preset business indicators based on the semantic drift trajectories of multiple user groups to obtain quantification results, and generating an attribution report based on the quantification results.

[0088] The aforementioned method constructs complete user behavior sequences using multimodal behavioral data, avoiding the attribution bias caused by relying on single-dimensional data. Transforming user behavior sequences into interpretable natural language semantic labels prevents attribution from remaining merely on the surface of the data, helping to capture the deeper semantic motivations behind user behavior. Converting natural language semantic labels into semantic vectors ensures that subsequent clustering accurately reflects the semantic similarity of user groups, resulting in more precise attribution results. By clustering the semantic vectors of the current time window and historical time windows separately, and then comparing the clustering results of the two time windows, the semantic drift trajectory of user groups over time can be accurately identified, providing a reliable data foundation for quantifying the contribution of user groups. Quantifying the contribution of each user group to the fluctuation of preset business indicators based on semantic drift trajectories clearly distinguishes the impact of different user groups on the fluctuation of preset business indicators, ensuring that the attribution is based on the impact of changes in user behavior on business indicators. This avoids the bias caused by misjudging accidental correlations as causal relationships, guaranteeing the accuracy of the attribution results.

[0089] Based on the user behavior attribution method described in the preceding embodiments, this application also provides a user behavior attribution device. Figure 4 This is a schematic diagram of the device. Figure 4 As shown, the user behavior attribution device includes:

[0090] The behavior sequence construction module 401 is used to construct the behavior sequence of each user based on the multimodal behavior data of multiple users within a preset time window; the preset time window includes the current time window and the historical time window.

[0091] The semantic label and semantic vector determination module 402 is used to determine the corresponding natural language semantic label based on the behavior sequence of each user, and convert the natural language semantic labels corresponding to multiple users into corresponding semantic vectors respectively.

[0092] The semantic drift trajectory determination module 403 is used to cluster the semantic vectors of multiple users in the current time window and the semantic vectors of the multiple users in the historical time window, respectively, to obtain the clustering results of the current time window and the clustering results of the historical time window, and to compare the clustering results of the current time window and the clustering results of the historical time window to determine the semantic drift trajectory of each user group; the clustering results contain multiple user groups.

[0093] The attribution report determination module 404 is used to quantify the contribution of each user group to the fluctuation of preset business indicators based on the semantic drift trajectory of multiple user groups, obtain the quantification result, and generate an attribution report based on the quantification result.

[0094] Optionally, the semantic label and semantic vector determination module is used for:

[0095] Based on the behavior sequence of each user, a hierarchical update mechanism is used to determine the corresponding natural language semantic label; the hierarchical update mechanism includes atomic feature extraction and natural language semantic label determination.

[0096] Optionally, determining the corresponding natural language semantic label based on the behavior sequence of each user using a hierarchical update mechanism includes:

[0097] For any user:

[0098] Atomic features of the user's behavior sequence in the current time window and atomic features of the user's behavior sequence in the historical time window are extracted respectively to obtain the atomic feature extraction results of the user in the current time window and the atomic feature extraction results of the user in the historical time window.

[0099] If the user's behavior changes compared to the atomic feature extraction results in the current time window and the historical time window, then the corresponding natural language semantic label is determined based on the atomic feature extraction results in the current time window using a large language model.

[0100] If the user's behavior has not changed compared to the atomic feature extraction results in the current time window and the atomic feature extraction results in the historical time window, then the corresponding historical natural language semantic label is used; the historical natural language semantic label is determined based on the user's atomic feature extraction results in the historical time window.

[0101] Optionally, the semantic drift trajectory determination module is used for:

[0102] For any user group:

[0103] Based on the semantic vectors of multiple users included in the user group, the center vector of the user group in the current time window is determined.

[0104] Based on the semantic vectors of multiple users included in the user group, the center vector of the user group in the historical time window is determined.

[0105] By comparing the center vector of the user group in the current time window with the center vector of the user group in historical time windows, the semantic drift trajectory of the user group is determined.

[0106] Optionally, the semantic drift trajectory of the user group is determined by comparing the center vector of the user group in the current time window with the center vector of the user group in historical time windows, including:

[0107] Calculate the bulldozer distance between the center vector of the user group in the current time window and the center vector of the user group in the historical time window, obtain the bulldozer distance calculation result, and determine the semantic drift trajectory of the user group based on the bulldozer distance calculation result.

[0108] Optionally, the semantic drift trajectory of the user group is determined by comparing the center vector of the user group in the current time window with the center vector of the user group in historical time windows, including:

[0109] Calculate the cosine similarity between the center vector of the user group in the current time window and the center vector of the user group in the historical time window, obtain the cosine similarity calculation result, and determine the semantic drift trajectory of the user group based on the cosine similarity calculation result.

[0110] Optionally, the attribution report determination module is used for:

[0111] Based on the semantic drift trajectories of multiple user groups, the contribution of each user group to the fluctuation of preset business indicators is quantified using the Shapley value attribution algorithm, and the quantification results are obtained.

[0112] Optionally, the semantic drift trajectory determination module is used for:

[0113] Unsupervised density clustering is performed on the semantic vectors of multiple users in the current time window and the semantic vectors of the same multiple users in the historical time window, respectively, to obtain the clustering results for the current time window and the clustering results for the historical time window.

[0114] Furthermore, embodiments of this application also provide a computer-readable storage medium storing a computer program that, when executed by a processor, implements the user behavior attribution method as described in any of the method embodiments.

[0115] It should be noted that the various embodiments in this specification are described in a progressive manner, and the same or similar parts between the various embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, for the device embodiments, since they are basically similar to the method embodiments, the description is relatively simple, and the relevant parts can be referred to the description of the method embodiments. The device embodiments described above are merely illustrative, and the units described as separate components may or may not be physically separate. The components indicated as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment solution according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0116] The above description is merely one specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A user behavior attribution method, characterized in that, include: Based on multimodal behavioral data of multiple users within a preset time window, construct the behavioral sequence for each user; The preset time window includes the current time window and the historical time window; Based on the behavior sequence of each user, the corresponding natural language semantic tags are determined, and the natural language semantic tags corresponding to multiple users are converted into corresponding semantic vectors respectively; The semantic vectors of multiple users in the current time window and the semantic vectors of the same multiple users in historical time windows are clustered to obtain the clustering results of the current time window and the clustering results of the historical time windows. The semantic drift trajectory of each user group is determined by comparing the clustering results of the current time window and the clustering results of the historical time windows. The clustering results contain multiple user groups. Based on the semantic drift trajectories of multiple user groups, the contribution of each user group to the fluctuation of preset business indicators is quantified to obtain the quantification results, and an attribution report is generated based on the quantification results.

2. The method according to claim 1, characterized in that, The step of determining the corresponding natural language semantic label based on the behavior sequence of each user includes: Based on the behavior sequence of each user, a hierarchical update mechanism is used to determine the corresponding natural language semantic label; the hierarchical update mechanism includes atomic feature extraction and natural language semantic label determination.

3. The method according to claim 2, characterized in that, The step of determining the corresponding natural language semantic labels based on the behavior sequence of each user using a hierarchical update mechanism includes: For any user: Atomic features of the user's behavior sequence in the current time window and atomic features of the user's behavior sequence in the historical time window are extracted respectively to obtain the atomic feature extraction results of the user in the current time window and the atomic feature extraction results of the user in the historical time window. If the user's behavior changes compared to the atomic feature extraction results in the current time window and the atomic feature extraction results in the historical time window, then the corresponding natural language semantic label is determined based on the atomic feature extraction results in the current time window using the large language model. If the user's behavior has not changed compared to the atomic feature extraction results in the current time window and the atomic feature extraction results in the historical time window, then the corresponding historical natural language semantic label is used; the historical natural language semantic label is determined based on the user's atomic feature extraction results in the historical time window.

4. The method according to claim 1, characterized in that, The step of comparing the clustering results of the current time window with the clustering results of the historical time windows to determine the semantic drift trajectory of each user group includes: For any user group: Based on the semantic vectors of multiple users included in the user group, determine the center vector of the user group in the current time window; Based on the semantic vectors of multiple users included in the user group, determine the center vector of the user group in the historical time window; By comparing the center vector of the user group in the current time window with the center vector of the user group in historical time windows, the semantic drift trajectory of the user group is determined.

5. The method according to claim 4, characterized in that, The step of comparing the center vector of the user group in the current time window with the center vector of the user group in historical time windows to determine the semantic drift trajectory of the user group includes: Calculate the bulldozer distance between the center vector of the user group in the current time window and the center vector of the user group in the historical time window, obtain the bulldozer distance calculation result, and determine the semantic drift trajectory of the user group based on the bulldozer distance calculation result.

6. The method according to claim 4, characterized in that, The step of comparing the center vector of the user group in the current time window with the center vector of the user group in historical time windows to determine the semantic drift trajectory of the user group includes: Calculate the cosine similarity between the center vector of the user group in the current time window and the center vector of the user group in the historical time window, obtain the cosine similarity calculation result, and determine the semantic drift trajectory of the user group based on the cosine similarity calculation result.

7. The method according to claim 1, characterized in that, The semantic drift trajectories based on multiple user groups are used to quantify the contribution of each user group to the fluctuation of preset business indicators, resulting in quantification results, including: Based on the semantic drift trajectories of multiple user groups, the contribution of each user group to the fluctuation of preset business indicators is quantified using the Shapley value attribution algorithm, and the quantification results are obtained.

8. The method according to claim 1, characterized in that, The step of clustering the semantic vectors of multiple users in the current time window and the semantic vectors of the multiple users in historical time windows to obtain the clustering results for the current time window and the clustering results for the historical time windows includes: Unsupervised density clustering is performed on the semantic vectors of multiple users in the current time window and the semantic vectors of the same multiple users in the historical time window, respectively, to obtain the clustering results for the current time window and the clustering results for the historical time window.

9. A user behavior attribution device, characterized in that, include: The behavior sequence construction module is used to construct the behavior sequence of each user based on the multimodal behavior data of multiple users within a preset time window; The preset time window includes the current time window and the historical time window; The semantic label and semantic vector determination module is used to determine the corresponding natural language semantic label based on the behavior sequence of each user, and to convert the natural language semantic labels corresponding to multiple users into corresponding semantic vectors respectively; The semantic drift trajectory determination module is used to cluster the semantic vectors of multiple users in the current time window and the semantic vectors of the same multiple users in historical time windows, respectively, to obtain the clustering results of the current time window and the clustering results of the historical time windows, and to compare the clustering results of the current time window and the clustering results of the historical time windows to determine the semantic drift trajectory of each user group; the clustering results contain multiple user groups. The attribution report determination module is used to quantify the contribution of each user group to the fluctuation of preset business indicators based on the semantic drift trajectory of multiple user groups, obtain the quantification results, and generate an attribution report based on the quantification results.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the user behavior attribution method as described in any one of claims 1-8.

Images