Goods search optimization method and system based on historical behavior data

Abstract

The invention discloses a commodity search optimization method and system based on historical behavior data, and relates to the technical field of commodity search optimization. Constructing a path signal dynamic adjustment mechanism for performing different sort weight adjustment on the dominant path, the dispersed path and the mixed path source path; in the process of adjusting each source path by the path signal dynamic adjustment mechanism, obtaining path dynamic feedback information of each source path in real time, analyzing after obtaining, and evaluating whether the adjustment effect of the path signal dynamic adjustment mechanism on each source path can reach the expectation; and a path signal dynamic adjustment mechanism is optimized according to an evaluation result. According to the method, the problem of signal misjudgment of multi-path repeated clicks is solved, dynamic balance of recommendation accuracy and diversity is realized, and the response capability and user satisfaction of a recommendation system are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of commodity search optimization, and in particular to a commodity search optimization method and system based on historical behavior data. Background Art

[0002] Product search is a core function used by users of e-commerce platforms when looking for specific products. By entering keywords or filtering conditions, users can find results that meet their needs from a large number of products. Common implementation methods include keyword-based text matching, product attribute-based filtering, and personalized recommendations combined with machine learning. However, traditional search methods usually only rely on keyword matching and product attribute screening, which easily ignores users' actual needs and preferences, resulting in a mismatch between search results and user expectations, thereby reducing user experience and conversion rate. Therefore, it is particularly important to optimize product search in order to improve the accuracy and relevance of search results and meet users' personalized needs. The optimization method based on historical behavior data can deeply explore users' potential intentions and build a personalized recommendation model by collecting and analyzing users' search, browsing, clicking, purchasing and other behavioral information, so that search results are not only based on the input keywords, but also can be intelligently sorted and recommended in combination with users' past behaviors, thereby improving search accuracy and user satisfaction, and ultimately improving the transaction conversion rate of the platform.

[0003] Existing product search optimization technology based on historical behavior data mainly collects multi-dimensional behavior data of users on e-commerce platforms, such as search keywords, product browsing records, click behaviors, favorite and add-to-cart behaviors, and actual purchase records, etc., uses these data to build user feature portraits, and adopts machine learning and deep learning models (such as RNN, Transformer) to predict users' potential search intentions and preferences. When a user initiates a search request, the historical behavior feature vector is extracted and the intent is analyzed in combination with the user's current search context. A candidate set of products that highly match the user's preferences is calculated, and the products are sorted through a comprehensive scoring model. The scoring model will comprehensively evaluate the relevance of products based on multiple dimensions such as user intention index and behavior feature information, so as to give priority to recommending products that better meet user needs. At the same time, this optimization process continuously adjusts model parameters through real-time monitoring of search click-through rate, conversion rate and user interaction behavior feedback to dynamically adapt to changes in user needs, continuously improve the accuracy of search results and personalized recommendation effects, thereby improving the user satisfaction and commercial conversion rate of the platform.

[0004] The prior art has the following deficiencies: When a user clicks on the same product multiple times through multiple recommendation source paths (such as search results, recommendation lists, or ad slots), the system will simply superimpose these click behaviors as a strong signal of interest in the product. Since the existing product search optimization technology based on historical behavior data does not independently model and evaluate the click source path when extracting click behavior features, it is unable to distinguish that these click behaviors may be operations by users to verify the consistency of different recommendation sources, rather than true expressions of preference, which will cause the interest weight of the product to be erroneously amplified. Due to the lack of refined analysis of the source of clicks, the existing technology cannot correctly identify the signal nature of repeated paths, resulting in excessive priority display of the product in the recommendation sorting, ignoring the user's potential interest in other products. In turn, this problem will increase the redundancy of the recommendation results, making users feel that the recommended content is single, reducing the satisfaction of the search experience, and at the same time, the platform's waste of resources on low-demand products will also affect the overall recommendation efficiency.

[0005] The above information disclosed in this Background section is only for enhancement of understanding of the background of the present disclosure and therefore it may contain information that does not constitute the prior art that is already known to one of ordinary skill in the art. Summary of the invention

[0006] The purpose of the present invention is to provide a commodity search optimization method and system based on historical behavior data to solve the problems in the above-mentioned background technology.

[0007] In order to achieve the above object, the present invention provides the following technical solution: a commodity search optimization method based on historical behavior data, specifically comprising the following steps: By detecting the user's click behavior in real time, it is identified whether the user has clicked on the same product multiple times through multiple recommendation source paths. If it is identified that the user has clicked on the same product multiple times through multiple recommendation source paths, the multiple source paths generated by the user during the time period of searching for the product are classified and marked to distinguish the click behavior characteristics of different source paths. Obtain the click behavior information of each source path in real time, analyze it after acquisition, evaluate the contribution strength of each source path to the user's real interest signal, and divide each source path into dominant path, decentralized path and mixed path according to the evaluation results; According to the division results of each source path, a path signal dynamic adjustment mechanism is constructed to adjust the dominant path, the decentralized path and the mixed path source paths with different sorting weights respectively; In the process of adjusting each source path by the path signal dynamic adjustment mechanism, the path dynamic feedback information of each source path is obtained in real time, and analyzed after acquisition to evaluate whether the adjustment effect of the path signal dynamic adjustment mechanism on each source path can meet expectations, and optimize the path signal dynamic adjustment mechanism according to the evaluation results; A comprehensive analysis is conducted on the application effect, adjustment records and optimization results of the path signal dynamic adjustment mechanism, and the path signal processing capability and dynamic adjustment mechanism of the recommendation ranking model are continuously optimized in combination with newly added user behavior data.

[0008] Preferably, the click behavior information of each source path is obtained in real time, and analyzed after acquisition to evaluate the contribution strength of each source path to the user's real interest signal, and each source path is divided into a dominant path, a decentralized path and a mixed path according to the evaluation result, which specifically includes the following steps: Obtain click behavior information of each source path in real time and perform preprocessing after acquisition; Extracting the path click distribution information and path click feature information from the preprocessed click behavior information of each source path, and analyzing them after extraction to generate the click concentration coefficient and click sparse index of each source path respectively; The click concentration coefficient and click sparse index of each source path are used to construct a path contribution intensity model, and the contribution evaluation coefficient of each source path is generated. The contribution evaluation coefficient of each source path is compared with the pre-set contribution evaluation coefficient threshold range, and the contribution intensity of each source path to the user's real interest signal is evaluated based on the comparison results. According to the evaluation results, each source path is divided into a dominant path, a dispersed path and a mixed path.

[0009] Preferably, the logic for obtaining the click concentration coefficient and click sparse index of each source path is as follows: Extract the path click distribution information from the preprocessed click behavior information of each source path, including the number of clicks on each product in each source path at different times within a period of time, the total number of clicks, and the deviation between the number of clicks on each product and the total number of clicks in the source path, and use the function , and To express, For time point, Indicates that within a period of time At the moment the user The source path to Number of clicks on products, Indicates that within a period of time At the moment the user The total number of clicks within source paths, Indicates that within a period of time At the moment the user The source path is The deviation between the number of clicks on a product and the total number of clicks in the source path. , , and are all positive integers, defining the time period as ; Calculate the click concentration coefficient of each source path. The specific calculation formula is as follows: ; In the formula, For the Click concentration coefficient of source paths; Extract the path click feature information from the preprocessed click behavior information of each source path, including the total number of different products clicked by users in each source path at different times within a period of time, the number of product categories, and the average number of clicks on all products by users in each source path, and use the function to calculate the click feature information according to the time series. , and To express, For time point, Indicates that within a period of time At the moment the user The total number of different products clicked within the source path, Indicates that within a period of time At the moment the user The number of product categories for different products clicked within the source path, Indicates that within a period of time At the moment the user The average number of clicks on all products in a source path; Calculate the click sparse index of each source path. The specific calculation formula is as follows: ; In the formula, For the The click sparseness index of the source path.

[0010] Preferably, the click concentration coefficients of each source path generated and click sparsity index Construct a path contribution intensity model and generate the contribution evaluation coefficient of each source path through weighted summation , and the contribution evaluation coefficients of each source path generated The contribution evaluation coefficient threshold interval is set A comparison is made, and the contribution strength of each source path to the user's real interest signal is evaluated based on the comparison results. Based on the evaluation results, each source path is divided into a dominant path, a dispersed path, and a mixed path. The specific comparison analysis and division are as follows: like , the contribution intensity of the source path to the user's real interest signal is low, then the source path is divided into a dispersed path; like , the contribution intensity of the source path to the user's real interest signal is medium, then the source path is classified as a mixed path; like , the contribution intensity of the source path to the user's real interest signal is high, then the source path is classified as a dominant path.

[0011] Preferably, according to the division results of each source path, a path signal dynamic adjustment mechanism is constructed for adjusting the dominant path, the decentralized path and the hybrid path source paths with different sorting weights, specifically: According to the division results of dominant path, decentralized path and mixed path, different ranking weight adjustment rules are set respectively to form a dynamic adjustment mechanism of path signal; this adjustment mechanism is based on the contribution evaluation coefficient of each source path and the optimization goal of the recommendation system, and automatically determines the ranking weight adjustment method and amplitude of the products in each path through pre-set rules; Different sorting strategy adjustments are made to the dominant path, dispersed path and mixed path source paths respectively, including: for the dominant path, the weight enhancement rule in the adjustment mechanism is used to increase the sorting weight of the clicked products in the path, and give priority to displaying the products that the user is clearly interested in; for the dispersed path, the weight dispersion rule in the adjustment mechanism is used to reduce the sorting weight of repeatedly clicked products, and at the same time increase the display weight of other unclicked products in the path to increase the diversity of recommended content; for the mixed path, the weight balance rule in the adjustment mechanism is used to maintain the sorting weight of the clicked products in the path to balance the accuracy and diversity of the recommendations.

[0012] Preferably, in the process of adjusting each source path by the path signal dynamic adjustment mechanism, the path dynamic feedback information of each source path is obtained in real time, and analyzed after acquisition to evaluate whether the adjustment effect of the path signal dynamic adjustment mechanism on each source path can meet expectations, and optimize the path signal dynamic adjustment mechanism according to the evaluation result, which specifically includes the following steps: In the process of adjusting each source path by the path signal dynamic adjustment mechanism, the path dynamic feedback information of each source path is obtained in real time, and preprocessed after being obtained; Extracting the ranking adjustment information and behavior deviation information from the pre-processed path dynamic feedback information of each source path, and analyzing them after extraction to generate a ranking adjustment amplitude index and a click behavior deviation index of each source path respectively; An adjustment effect evaluation model is constructed for the ranking adjustment amplitude index and click behavior deviation index generated for each source path, and an adjustment evaluation coefficient for each source path is generated. The generated adjustment evaluation coefficient for each source path is compared with the pre-set adjustment evaluation coefficient threshold for each source path. Based on the comparison results, it is evaluated whether the adjustment effect of the path signal dynamic adjustment mechanism on each source path can meet expectations, and the path signal dynamic adjustment mechanism is optimized based on the evaluation results.

[0013] Preferably, the logic for obtaining the ranking adjustment index and click behavior deviation index of each source path is as follows: Extract the sorting adjustment information from the preprocessed dynamic feedback information of each source path, including the sorting weight value of each product in each source path before dynamic adjustment, the sorting weight value after dynamic adjustment, and the rate of change of the number of clicks of each product in each source path before and after adjustment, and mark them as , and , Indicates Source path The ranking weight value of a product before dynamic adjustment. Indicates Source path The ranking weight value of each product after dynamic adjustment. Indicates that the user is The source path is The rate of change of the number of clicks on a product per unit time before and after adjustment, , , and are all positive integers; Calculate the ranking adjustment index of each source path. The specific calculation formula is as follows: ; In the formula, For the The ranking adjustment index of the source paths; The behavior deviation information in the pre-processed path dynamic feedback information of each source path is extracted, including the distribution ratio of the number of clicks of users on each product in each source path before adjustment, the distribution ratio after adjustment, and the change in the display weight value of each product in each source path before and after adjustment, and they are marked as , and , Indicates that the user is The source path is The distribution ratio of the number of clicks on each product before adjustment, Indicates that the user is The source path is The distribution ratio of the number of clicks on the products after adjustment, Indicates Source path The change in display weight of a product before and after adjustment; Calculate the click behavior deviation index of each source path. The specific calculation formula is as follows: ; In the formula, For the The click behavior deviation index of the source path.

[0014] Preferably, the ranking of each source path generated is adjusted by the amplitude index and click behavior deviation index Construct an adjustment effect evaluation model and generate the adjustment evaluation coefficients of each source path through weighted summation , and the adjusted evaluation coefficients of each source path generated Adjustment evaluation coefficient thresholds for each source path that are set in advance Compare and evaluate whether the adjustment effect of the path signal dynamic adjustment mechanism on each source path can meet expectations based on the comparison results, and optimize the path signal dynamic adjustment mechanism based on the evaluation results. The specific comparison and analysis are as follows: like , the adjustment effect of the path signal dynamic adjustment mechanism on the source path cannot meet the expectations, and the path signal dynamic adjustment mechanism needs to be optimized, including: re-evaluating the ranking weight adjustment rules of the products in the path, increasing the real-time data collection and analysis of the user's latest click behavior and display preference; dynamically adjusting the weighting parameters in the ranking adjustment amplitude index and the click behavior deviation index to optimize the balance between ranking and click distribution; re-setting the boundary conditions and priority rules for ranking weight adjustment; like , the adjustment effect of the path signal dynamic adjustment mechanism on the source path can achieve the expected effect, and there is no need to optimize the path signal dynamic adjustment mechanism.

[0015] Preferably, the commodity search optimization system based on historical behavior data includes a user click behavior detection module, a path contribution evaluation module, a path signal adjustment mechanism construction module, a path dynamic feedback evaluation module, and a path signal comprehensive optimization module; The user click behavior detection module detects the user's click behavior in real time to identify whether the user has clicked on the same product multiple times through multiple recommendation source paths. If it is identified that the user has clicked on the same product multiple times through multiple recommendation source paths, the multiple source paths generated by the user during the time period of searching for the product are classified and marked to distinguish the click behavior characteristics of different source paths. The path contribution evaluation module obtains the click behavior information of each source path in real time, analyzes it after acquisition, evaluates the contribution strength of each source path to the user's real interest signal, and divides each source path into dominant path, decentralized path and mixed path according to the evaluation results; A path signal adjustment mechanism construction module is used to construct a path signal dynamic adjustment mechanism for adjusting different ranking weights of dominant path, decentralized path and mixed path source paths according to the division results of each source path; The path dynamic feedback evaluation module obtains the path dynamic feedback information of each source path in real time during the process of the path signal dynamic adjustment mechanism adjusting each source path, and analyzes it after acquisition to evaluate whether the adjustment effect of the path signal dynamic adjustment mechanism on each source path can meet expectations, and optimizes the path signal dynamic adjustment mechanism according to the evaluation results; The path signal comprehensive optimization module conducts a comprehensive analysis of the application effect, adjustment records and optimization results of the path signal dynamic adjustment mechanism, and continuously optimizes the path signal processing capability and dynamic adjustment mechanism of the recommendation ranking model based on the newly added user behavior data.

[0016] In the above technical solution, the technical effects and advantages provided by the present invention are: 1. The present invention can accurately distinguish the click behavior characteristics of users in multiple recommendation source paths by introducing a dynamic adjustment mechanism for path signals, thus solving the technical problem that repeated click signals of paths cannot be identified in the prior art. By detecting user click behaviors in real time, classification and marking are performed at the path level, and the contribution of each path to the user's real interest signal is quantified through the click concentration coefficient and click sparse index. This fine-grained path modeling enables the system to filter out interference signals from duplicate sources and accurately identify the user's real interests. As a result, recommendation rankings are more accurate, and products that users are interested in can be displayed first, thereby significantly improving the click-through rate and conversion rate of the recommendation system.

[0017] 2. The present invention combines the ranking adjustment amplitude index and the click behavior deviation index to comprehensively evaluate the effect of the path signal dynamic adjustment mechanism on ranking optimization and click behavior guidance. By constructing a dynamic adjustment mechanism, differentiated ranking weight adjustment rules are set for different paths according to the path classification results. This mechanism gives priority to recommending products that users are clearly interested in in dominant paths, increases the weight of unclicked products in decentralized paths to improve the diversity of recommended content, and achieves a dynamic balance between accuracy and diversity in hybrid paths. In this way, the system can not only meet users' immediate interest needs, but also guide users to explore more potential interests, thereby improving the content coverage of the recommendation system and the overall satisfaction of users.

[0018] 3. The present invention constructs the adaptive optimization capability of the recommendation system through the real-time collection of dynamic feedback information of the path and the optimization of the adjustment mechanism. During the application of the adjustment mechanism, the system can dynamically monitor the user's new click behavior and path signal changes, and continuously optimize the parameters and rules of the path signal dynamic adjustment mechanism by adjusting the effect evaluation model. At the same time, the recommendation sorting model is continuously optimized in combination with the newly added user behavior data, so that the recommendation system can quickly respond to the changing trends of user interests and avoid the obsolescence and singularity of recommendation results. As a result, the recommendation system shows higher flexibility and continuous optimization capabilities in complex and changeable user demand scenarios, bringing higher user retention rate and commercial value to the platform. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in the present invention. For ordinary technicians in this field, other drawings can also be obtained based on these drawings.

[0020] Figure 1 It is a flowchart of the commodity search optimization method and system based on historical behavior data of the present invention.

[0021] Figure 2Schematic diagram of the modules of the commodity search optimization method and system based on historical behavior data of the present invention. DETAILED DESCRIPTION

[0022] Example embodiments will now be described more fully with reference to the accompanying drawings. However, example embodiments can be implemented in a variety of forms and should not be construed as limited to the examples set forth herein; rather, these example embodiments are provided so that the description of the present disclosure will be more comprehensive and complete, and the concept of the example embodiments will be fully conveyed to those skilled in the art.

[0023] The present invention provides Figure 1 The commodity search optimization method based on historical behavior data shown in the figure specifically includes the following steps: By detecting the user's click behavior in real time, it is identified whether the user has clicked on the same product multiple times through multiple recommendation source paths. If it is identified that the user has clicked on the same product multiple times through multiple recommendation source paths, the multiple source paths generated by the user during the time period of searching for the product are classified and marked to distinguish the click behavior characteristics of different source paths. By detecting user click behavior in real time and identifying whether the user has clicked on the same product multiple times through multiple recommendation source paths, this can be achieved by recording the product ID, click source path (such as search results, recommendation lists, ad space, etc.) and click timestamp in the user's click data. The system updates the user's click sequence in real time through the click log in the session, and when it detects that the same product ID appears in different source paths, it uses the timestamp to determine whether it complies with the logic of multi-path clicks (such as the click time interval is within the preset range). When the product ID appears repeatedly in multiple paths and the time interval is reasonable, the system determines that the user has clicked on the same product multiple times through multiple source paths. This process can be completed by automatically matching the path attributes and time attributes of the click records without manual intervention.

[0024] When it is identified that a user has clicked on the same product multiple times through multiple recommendation source paths, the system will classify and mark the multiple source paths generated by the user during the time period of searching for the product to distinguish the click behavior characteristics of different source paths. This process is achieved by grouping the click logs by source path. After grouping, the system marks the click frequency, time distribution, product attribute characteristics, etc. of each path, and generates a path classification table to reflect the independent characteristics of each path. Classification marking provides support for the calculation of subsequent path parameters, such as evaluating the contribution of the path to the user's interest signal, laying the foundation for optimizing the recommendation ranking.

[0025] The purpose of doing this is to address the defect of the existing technology that the click source path is not independently modeled. By classifying and modeling the path signal, the problem of click behavior signal confusion is solved. The core of the technical problem is that repeated clicks on the path are mistakenly superimposed as a strong signal of interest in the product, causing the recommendation system to amplify the weight of the product and ignore the potential interest in other products. This design ensures that the system can independently analyze the path contribution signal by accurately identifying and marking the click behavior characteristics of different source paths, providing data support for the subsequent dynamic adjustment mechanism, thereby improving the relevance and diversity of the recommendation ranking, and effectively solving the problems of redundant recommended content and unbalanced ranking.

[0026] Obtain the click behavior information of each source path in real time, analyze it after acquisition, evaluate the contribution strength of each source path to the user's real interest signal, and divide each source path into dominant path, decentralized path and mixed path according to the evaluation results; In this embodiment, the click behavior information of each source path is obtained in real time, and analyzed after acquisition to evaluate the contribution strength of each source path to the user's real interest signal, and each source path is divided into a dominant path, a decentralized path and a mixed path according to the evaluation result, which specifically includes the following steps: Obtain click behavior information of each source path in real time and perform preprocessing after acquisition; Real-time acquisition of click behavior information for each source path can be achieved by deploying real-time data collection modules in the front-end and back-end of the recommendation system. The front-end module captures click behavior information in real time through the user's interaction log, including the clicked product ID, source path identifier (such as search results, recommendation lists, advertising space), click timestamp, and user session ID, and then sends the data to the back-end data processing pipeline through event streaming technology (such as Kafka or other message queues). In the background, the system will preliminarily classify the received click events according to the source path identifier and store them in a distributed database. In order to achieve real-time performance, the back-end can use an in-memory computing framework (such as Spark Streaming) to process streaming data to ensure that the path click behavior is captured and recorded immediately after the user click occurs.

[0027] The purpose of preprocessing is to provide consistency and integrity for subsequent data extraction and analysis, and to avoid noise or outliers in the original data from interfering with the path contribution evaluation. Preprocessing includes the following steps: First, perform integrity checks on the acquired click data, such as detecting whether there are missing key fields (such as product ID or source path); second, perform data deduplication processing to ensure that repeated click records will not affect subsequent analysis; third, standardize the timestamp format of click behavior, and unify time data from different sources into a consistent time zone and format; finally, remove abnormal data through filtering algorithms, such as detecting whether multiple repeated clicks in an ultra-short period of time are misoperations or automated behaviors. The implementation of preprocessing can be completed in the background through software logic, such as embedding data verification and cleaning modules in the data pipeline to verify and correct each click record before entering the next stage of the analysis process.

[0028] Extracting the path click distribution information and path click feature information from the preprocessed click behavior information of each source path, and analyzing them after extraction to generate the click concentration coefficient and click sparse index of each source path respectively; The click concentration coefficient and click sparse index of each source path are used to construct a path contribution intensity model, and the contribution evaluation coefficient of each source path is generated. The contribution evaluation coefficient of each source path is compared with the pre-set contribution evaluation coefficient threshold range, and the contribution intensity of each source path to the user's real interest signal is evaluated based on the comparison results. According to the evaluation results, each source path is divided into a dominant path, a dispersed path and a mixed path.

[0029] The pre-set contribution evaluation coefficient threshold interval can be dynamically determined in a data-driven manner by combining historical click behavior data and model optimization results. The specific method is to first collect a large amount of historical click data, covering the click behavior characteristics of users in different source paths, including the distribution of path click concentration coefficient and path click sparse index. By analyzing these historical data, clustering algorithms (such as K-means) are used to group the source paths, and the evaluation coefficient ranges corresponding to dominant paths, dispersed paths, and mixed paths are automatically divided. After the grouping is completed, the statistical characteristics (such as mean, standard deviation, etc.) of the evaluation coefficient of each group of paths are calculated as a preliminary reference for the threshold interval. In addition, through optimization algorithms (such as Bayesian optimization or grid search), combined with the actual effect of the recommendation ranking model, the upper and lower limits of the threshold are dynamically adjusted to improve the accuracy of classification and recommendation performance. This process can be completed through the background offline training pipeline, and the threshold interval is updated regularly to ensure that the system can adapt to changes in user behavior patterns.

[0030] In this embodiment, the logic for obtaining the click concentration coefficient and click sparse index of each source path is as follows: Extract the path click distribution information from the preprocessed click behavior information of each source path, including the number of clicks on each product in each source path at different times within a period of time, the total number of clicks, and the deviation between the number of clicks on each product and the total number of clicks in the source path, and use the function , and To express, For time point, Indicates that within a period of time At the moment the user The source path to Number of clicks on products, Indicates that within a period of time At the moment the user The total number of clicks within source paths, Indicates that within a period of time At the moment the user The source path is The deviation between the number of clicks on a product and the total number of clicks in the source path. , , and are all positive integers, defining the time period as ; Extracting the path click distribution information from the preprocessed click behavior information of each source path can be achieved through a click log analysis process based on a distributed data processing framework (such as Apache Spark or Flink). First, the click logs collected in real time are grouped by source path, and each click record is divided into corresponding time periods (such as every minute, every hour) according to the timestamp. In each source path group, the products clicked by the user are counted, the number of clicks on each product in a specific time period is extracted, and the total number of clicks in the time period is accumulated and calculated. In addition, by calculating the ratio of the number of clicks on each product to the total number of clicks in the source path, the relative distribution of product clicks can be further obtained, thereby generating a deviation value between the number of clicks and the total number of clicks. The entire extraction process is implemented through an automated data stream processing pipeline to ensure real-time performance and accuracy.

[0031] The quantitative data required to obtain the path click distribution information can be collected and stored in real time from the user interaction data through the pre-deployed logging system. Specifically, the number of clicks is obtained by recording and counting the product ID of each click behavior; the total number of clicks is obtained by accumulating and counting all click records in the source path over a period of time; the deviation value is calculated based on the difference between the number of clicks for each product and the total number of clicks in the source path, specifically |Product clicks−Total number of clicks / Total number of products|. In order to ensure the accuracy and consistency of this data, the logging system needs to record the path, timestamp, and product information of each click event in real time, and use distributed storage for maintenance, so as to ensure efficient calculation and dynamic update during the data extraction process.

[0032] Calculate the click concentration coefficient of each source path. The specific calculation formula is as follows: ; In the formula, For the Click concentration coefficient of source paths; This formula calculates the click concentration coefficient by integration in order to comprehensively and dynamically reflect the concentration of user click behaviors in each source path over a period of time. The design of the formula combines multiple dimensions of data such as the number of clicks, total clicks, and deviation values, and integrates the characteristics of these dimensions into a quantifiable concentration coefficient through step-by-step calculations. First, Indicates at time Neidi The number of clicks on a product, which is the basic data of click behavior within the path, is multiplied by The impact of high deviation products on centralization is further reduced (the larger the deviation value, the greater the difference in uniformity between the product clicks and the total distribution of path clicks). The total number of clicks in the normalized path ensures that the formula is applicable to source paths with different click magnitudes, and the time period integration operation dynamically captures the concentration changes in the time dimension by calculating the data at consecutive time points, thereby generating a click concentration coefficient that comprehensively reflects the concentration of the path. This design can accurately distinguish whether the user's interest in different paths is concentrated, especially for paths with high-frequency clicks, and helps to evaluate the true level of path contribution strength.

[0033] No. Click concentration coefficient of source paths The size of directly reflects the concentration of users' click behaviors in the path, and is closely related to the contribution strength of the path to the user's real interest signal. When the click concentration coefficient is large, it means that users' click behaviors are more concentrated on a few products, which usually indicates that users have a high interest in these products, and the contribution of the path to the user's real interest signal is more significant; on the contrary, when the click concentration coefficient is small, it means that users' click behaviors are scattered on multiple products, which may reflect that users' exploration behaviors in the path are more than clear expressions of interest, and the contribution of the path to the user's real interest signal is relatively low. Therefore, the click concentration coefficient provides a quantitative means to help the system identify whether the path reflects the user's real interest, and thus provide data support for optimizing the recommendation ranking model.

[0034] Extract the path click feature information from the preprocessed click behavior information of each source path, including the total number of different products clicked by users in each source path at different times within a period of time, the number of product categories, and the average number of clicks on all products by users in each source path, and use the function to calculate the click feature information according to the time series. , and To express, For time point, Indicates that within a period of time At the moment the user The total number of different products clicked within the source path, Indicates that within a period of time At the moment the user The number of product categories for different products clicked within the source path, Indicates that within a period of time At the moment the user The average number of clicks on all products in a source path; Extracting the path click feature information from the preprocessed click behavior information of each source path can be achieved through a distributed data processing framework (such as Apache Spark or Flink) combined with grouping and aggregation operations. First, the click log data collected in real time is grouped according to the source path and time period (such as per hour or per minute). In each source path group, the total number of different products clicked by the user is counted by deduplicating the product ID, and the total number of product categories is counted by deduplicating the product category field. In addition, the average number of clicks is calculated by accumulating the total number of clicks in the source path and dividing it by the number of different products. The extraction process relies on the fields such as product ID, product category, and timestamp contained in the click log, and updates the path click feature information in real time through pipeline processing to ensure that the system dynamically extracts the user's click behavior characteristics in different time periods.

[0035] The quantitative data in the path click feature information comes from the preprocessed click log. Specifically, the total number of different products clicked is obtained by counting the number of product IDs after deduplication in the same source path. The deduplication logic can be based on the uniqueness of the product ID field; the number of product categories is obtained by deduplicating the product category field in the click log. The product category field is usually the associated information of the product ID. The system can dynamically obtain and update it according to the metadata table of the product; the average number of clicks on all products by users in the path is calculated by accumulating the total number of all click records in the source path and dividing it by the total number of different products. The above data depends on the real-time recording of the click log, including core fields such as product ID, product category and click time. These fields are collected by the log module at the front end of the system and stored in the background. After statistics and aggregation by the distributed data processing framework, quantitative data of the path click feature information is formed.

[0036] Calculate the click sparse index of each source path. The specific calculation formula is as follows: ; In the formula, For the The click sparseness index of the source path.

[0037] This formula calculates the click sparsity index by integration, aiming to dynamically quantify the degree of dispersion of users' click behavior in a certain source path, and comprehensively describes the sparsity of clicks in the path by taking into account factors such as the total number of products, product category distribution, and click uniformity. First, represents the total number of items clicked within the path, and represents the number of product categories. The product of the two reflects the breadth of path clicks and category diversity. This part defines the important basis of sparsity. Next, the denominator introduces , the uniformity of user click behavior is measured by the absolute value of the deviation between the number of clicks and the average value. The larger the deviation, the lower the sparsity. The summation in the denominator and the overall normalization adjustment enable the calculation result to balance the differences in the number of clicks on different paths. At the same time, the integral operation dynamically captures the click distribution characteristics of the time dimension. The overall formula comprehensively reflects the sparsity characteristics of path clicks through multiple variables, making it more universal in different time periods, product categories and click levels, and can scientifically quantify the contribution of the path to user interest exploration behavior.

[0038] No. Click sparseness index of source paths The size of directly reflects the degree of dispersion of the user's click behavior in the path, which in turn affects the evaluation of the contribution strength of the path's real interest signal. When the click sparsity index is high, it indicates that the user clicks on more different products in the path, and the product categories are widely distributed. At the same time, the uniformity of the click behavior is high, which usually means that the path is more used for exploratory behavior, the user's interest signal is not clear enough, and the contribution to the real interest is low; on the contrary, when the click sparsity index is low, it indicates that the user's click behavior is less dispersed and concentrated on certain specific products or categories, which may reflect a clearer interest signal. Therefore, the contribution of this path to the user's real interest signal is relatively stronger. The click sparsity index provides a key dimension for the analysis of the path to the interest signal by quantifying dispersion. It complements the click concentration coefficient and jointly evaluates the overall contribution strength of the path to the interest signal.

[0039] In this embodiment, the click concentration coefficients of each source path generated are and click sparsity index Construct a path contribution intensity model and generate the contribution evaluation coefficient of each source path through weighted summation , and the contribution evaluation coefficients of each source path generated The contribution evaluation coefficient threshold interval is set A comparison is made, and the contribution strength of each source path to the user's real interest signal is evaluated based on the comparison results. Based on the evaluation results, each source path is divided into a dominant path, a dispersed path, and a mixed path. The specific comparison analysis and division are as follows: like , the contribution intensity of the source path to the user's real interest signal is low, then the source path is divided into a dispersed path; This means that the user's click behavior within this source path is relatively scattered, and may be more of an exploratory operation on different products or categories rather than a clear expression of interest. This path usually reflects that the user's interests have not yet been focused, or that the user is trying to learn more product options through extensive click behavior. Therefore, the impact of this path on the recommendation ranking is that the system needs to reduce the weight of repeated recommended products and increase product diversity to match the user's exploration needs, thereby improving the richness of the user experience.

[0040] like , the contribution intensity of the source path to the user's real interest signal is medium, then the source path is classified as a mixed path; This indicates that the user's click behavior in this path contains a certain degree of interest concentration and also shows a certain degree of exploration. The user's real interest signal is between clear and vague. This path usually means that the user has some preferences for specific products, but is still in the process of exploring other options. For this path, the recommendation system needs to increase the diversity of recommendations appropriately while maintaining a certain weight of the products of interest, so as to balance the user's demand for specific products and potential interest in new products, and optimize the coverage of the recommendation ranking and user satisfaction.

[0041] like , the contribution intensity of the source path to the user's real interest signal is high, then the source path is classified as a dominant path.

[0042] This means that the user's click behavior in the source path is highly concentrated on a few products or categories, indicating that the user's interest signal is very clear and the products in the path are very attractive to the user. The impact of this path on the recommendation ranking is that the system should give priority to improving the recommendation weight of the products in the path, focusing on displaying products that the user is obviously interested in, maximizing the match with the user's current needs, thereby improving the accuracy of the recommendation and the click-through rate. This path is of great significance in meeting the immediate needs of users, and it also highlights the commercial value of the platform.

[0043] The click concentration coefficient of each source path generated and click sparsity index Construct a path contribution intensity model and generate the contribution evaluation coefficient of each source path through weighted summation The specific method is: first, determine the weight coefficients of the click concentration coefficient and the click sparse index according to the goals of the recommendation system (such as accuracy and diversity requirements). and The selection of weight coefficients can be optimized in a historical data-driven way, such as regression analysis based on user click data or obtaining the optimal weight value through model training to ensure a reasonable balance between the two in contribution evaluation. Then, use the formula The click concentration coefficient and the click sparse index are weighted summed, where reflects the influence weight of centrality on the intensity of path contribution, while It reflects the weight of the impact of sparsity on the strength of path contribution. Generally speaking, if the recommendation system pays more attention to the user's explicit interests (such as quickly satisfying user needs), it can be appropriately increased. If more attention is paid to the user's exploration behavior and diversity needs, the weight can be appropriately increased. Through this method, the model can comprehensively reflect the centralization and sparsity characteristics of the path and effectively quantify the contribution of the path to the user's real interest signal.

[0044] According to the division results of each source path, a path signal dynamic adjustment mechanism is constructed to adjust the dominant path, the decentralized path and the mixed path source paths with different sorting weights respectively; In this embodiment, according to the division results of each source path, a path signal dynamic adjustment mechanism is constructed for adjusting the dominant path, the decentralized path and the hybrid path source paths with different sorting weights respectively, specifically: According to the division results of dominant path, decentralized path and mixed path, different ranking weight adjustment rules are set respectively to form a dynamic adjustment mechanism of path signal; this adjustment mechanism is based on the contribution evaluation coefficient of each source path and the optimization goal of the recommendation system, and automatically determines the ranking weight adjustment method and amplitude of the products in each path through pre-set rules; To achieve "setting different ranking weight adjustment rules according to the division results of dominant path, dispersed path and mixed path, and forming a dynamic adjustment mechanism of path signal", it can be realized through the dynamic rule engine combined with the real-time data processing module. The specific method is, first, a set of ranking weight adjustment rules are preset in the system, and these rules are defined based on the characteristics of path type (dominant path, dispersed path, mixed path). For example, the rules of dominant path give priority to improving the ranking weight of clicked products, while the rules of dispersed path reduce the weight of repeatedly clicked products and enhance the diversity weight of unclicked products. Secondly, in the background of the recommendation system, the type of each source path is determined in real time by the contribution evaluation coefficient, and it is mapped to the corresponding ranking weight adjustment rule. The adjustment rule uses the contribution evaluation coefficient as an input parameter, and combines the optimization goal of the recommendation system (such as the balance between accuracy and diversity) to automatically generate the adjustment strategy and amplitude of the ranking weight. Finally, the ranking weight of the products in the recommendation model is updated in real time through the dynamic rule engine, so that the characteristics of the path type and the user's behavior pattern are reflected in the recommendation results. The purpose of doing this is to make full use of the differences in the contributions of different paths to user interest signals, give priority to recommending clear interest products in the dominant path, enhance exploratory recommendations in the decentralized path, and achieve a dynamic balance between accuracy and diversity in the hybrid path, thereby improving the user satisfaction and overall efficiency of the recommendation system.

[0045] Different sorting strategy adjustments are made to the dominant path, dispersed path and mixed path source paths respectively, including: for the dominant path, the weight enhancement rule in the adjustment mechanism is used to increase the sorting weight of the clicked products in the path, and give priority to displaying the products that the user is clearly interested in; for the dispersed path, the weight dispersion rule in the adjustment mechanism is used to reduce the sorting weight of repeatedly clicked products, and at the same time increase the display weight of other unclicked products in the path to increase the diversity of recommended content; for the mixed path, the weight balance rule in the adjustment mechanism is used to maintain the sorting weight of the clicked products in the path to balance the accuracy and diversity of the recommendations.

[0046] To achieve different sorting strategy adjustments for dominant paths, dispersed paths, and hybrid paths, the dynamic sorting weight adjustment module can be combined with real-time behavior data analysis. The specific method is to first design a weight adjustment function library in the recommendation system, which contains three adjustment rules for path types. For dominant paths, the system identifies products with highly concentrated user interests through contribution evaluation coefficients, and uses the "weight enhancement rule" to combine the click weights of these products with user behavior data (such as click counts and time frequency) to increase their sorting weights, thereby giving priority to displaying products that users are clearly interested in. For dispersed paths, the system adopts the "weight dispersion rule" to reduce the sorting weight of repeatedly clicked products and increase the display weight of other unclicked products in the path by analyzing the number of repeated clicks of products in the path and the attribute data of unclicked products (such as product category and similarity), thereby improving the diversity and coverage of recommended content. For hybrid paths, the system uses the "weight balance rule" to weight and smooth the click weights of all products in the path (such as assigning importance weights based on product categories), while moderately balancing the display frequency of clicked and unclicked products to take into account both accuracy and diversity. During the implementation process, the weight adjustment module continuously updates the ranking weight of each product through real-time data streams (such as user click logs and product attributes), and dynamically applies these weights in the recommendation generation stage, so that the recommendation results always reflect the user's latest behavior patterns and path characteristics. The purpose of this design is to maximize user satisfaction, improve the accuracy of recommendations through dominant paths, optimize the diversity of recommended content through decentralized paths, and achieve a dynamic balance between diversity and accuracy through hybrid paths, thereby comprehensively improving the overall effect of the recommendation system.

[0047] In the process of adjusting each source path by the path signal dynamic adjustment mechanism, the path dynamic feedback information of each source path is obtained in real time, and analyzed after acquisition to evaluate whether the adjustment effect of the path signal dynamic adjustment mechanism on each source path can meet expectations, and optimize the path signal dynamic adjustment mechanism according to the evaluation results; In this embodiment, in the process of adjusting each source path by the path signal dynamic adjustment mechanism, the path dynamic feedback information of each source path is obtained in real time, and analyzed after acquisition to evaluate whether the adjustment effect of the path signal dynamic adjustment mechanism on each source path can meet expectations, and optimize the path signal dynamic adjustment mechanism according to the evaluation result, specifically including the following steps: In the process of adjusting each source path by the path signal dynamic adjustment mechanism, the path dynamic feedback information of each source path is obtained in real time, and preprocessed after being obtained; Real-time acquisition of dynamic feedback information of each source path can be achieved through the front-end behavior monitoring module and the back-end real-time data processing system. The front-end monitoring module is responsible for real-time capture of user behavior data in each path, including the timestamp of clicking on the product, the number of clicks, the product ID, the number of impressions, and the sorting weight, etc. These data are recorded through the embedding technology or the event log system. The back-end data processing system receives and integrates the real-time data of the front end through a streaming data processing framework (such as Apache Kafka or Flink), and transmits the click behavior, display weight adjustment, and sorting change logs of the products in the path in real time to the central data storage in the form of a streaming pipeline for subsequent processing and analysis. In this way, the immediate effect of the path signal adjustment can be dynamically tracked, providing high-quality original data for subsequent evaluation.

[0048] The purpose of preprocessing is to improve the consistency, integrity and analysis efficiency of data, and ensure that dynamic feedback information can accurately reflect the actual effect of path adjustment in subsequent steps. Preprocessing mainly includes the following steps: First, perform data deduplication to filter out redundant data caused by network delays or repeated system transmission; second, perform integrity verification to ensure that each feedback record contains necessary fields (such as product ID, timestamp, sorting weight, etc.), and discard or complete the data of missing fields; third, standardize the timestamp to standard time zone format, and normalize the sorting weight to the same numerical range; fourth, remove outliers, and filter out obviously abnormal data such as click frequency and display number (such as a large number of clicks or displays in a very short time) by setting a reasonable range. These preprocessing steps can be completed through the background automated data pipeline, combined with data validation rules and cleaning algorithms, to process dynamic feedback information one by one in the real-time data stream to ensure that the data entering the subsequent analysis steps is accurate and reliable.

[0049] Extracting the ranking adjustment information and behavior deviation information from the pre-processed path dynamic feedback information of each source path, and analyzing them after extraction to generate a ranking adjustment amplitude index and a click behavior deviation index of each source path respectively; An adjustment effect evaluation model is constructed for the ranking adjustment amplitude index and click behavior deviation index generated for each source path, and an adjustment evaluation coefficient for each source path is generated. The generated adjustment evaluation coefficient for each source path is compared with the pre-set adjustment evaluation coefficient threshold for each source path. Based on the comparison results, it is evaluated whether the adjustment effect of the path signal dynamic adjustment mechanism on each source path can meet expectations, and the path signal dynamic adjustment mechanism is optimized based on the evaluation results.

[0050] The pre-set thresholds of adjustment evaluation coefficients for each source path can be dynamically determined by combining historical data analysis and model optimization. The specific method is to first collect a large amount of historical data, including the ranking adjustment amplitude index and click behavior deviation index generated by each source path under different adjustment mechanisms, as well as the adjustment evaluation coefficients and recommendation effects (such as click-through rate and conversion rate) of these paths. Then, cluster analysis algorithms (such as K-means) are used to group historical path data, and the path evaluation coefficients are divided into high-efficiency areas, low-efficiency areas, and intermediate areas according to the quality of the recommendation effect. The statistical characteristics of each group of evaluation coefficients (such as mean and standard deviation) are calculated to determine the preliminary threshold range. In addition, the threshold range is further adjusted through machine learning models (such as Bayesian optimization) to keep it consistent with the actual goals of the recommendation system (such as accuracy and diversity balance) when evaluating the effect of the adjustment mechanism. The entire process can be achieved through the background offline analysis module and real-time tuning module. Offline analysis is used to generate the basic threshold range, and real-time tuning dynamically updates the threshold in combination with the newly added data to ensure that the adjustment evaluation coefficient threshold can continuously reflect the system optimization needs.

[0051] In this embodiment, the logic for obtaining the ranking adjustment index and click behavior deviation index of each source path is as follows: Extract the sorting adjustment information from the preprocessed dynamic feedback information of each source path, including the sorting weight value of each product in each source path before dynamic adjustment, the sorting weight value after dynamic adjustment, and the rate of change of the number of clicks of each product in each source path before and after adjustment, and mark them as , and , Indicates Source path The ranking weight value of a product before dynamic adjustment. Indicates Source path The ranking weight value of each product after dynamic adjustment. Indicates that the user is The source path is The rate of change of the number of clicks on a product per unit time before and after adjustment, , , and are all positive integers; Extracting the sorting adjustment information from the preprocessed path dynamic feedback information of each source path can be achieved by combining the front-end user behavior tracking and the back-end sorting weight log record with the real-time data processing framework. Specifically, the front-end tracking records the user's interactive behavior (such as clicks, dwell time, etc.) within the path, and sends these behavior data to the back-end in real time. When the path dynamic adjustment mechanism takes effect, the back-end sorting model will update the sorting weight of each product in real time, and record the sorting weight value before and after the adjustment in the log. At the same time, through streaming data processing tools (such as Apache Kafka or Flink), the number of clicks of each product in the real-time log and its corresponding timestamp are integrated, the click change rate per unit time before and after the adjustment is calculated, and these data are marked as "sorting adjustment information". These extraction steps are completed in the background through an automated pipeline to ensure real-time data processing and synchronization with the dynamic adjustment mechanism.

[0052] The sorting adjustment information in the dynamic feedback information of the path consists of the following quantitative data: 1. Sorting weight value (before and after adjustment): obtained through the weight distribution result of the sorting model before and after each path adjustment. Specifically, the sorting weight value is recorded in the system log file. Each time the model calls the sorting algorithm, the generated weight value will be synchronously written to the log to mark the priority of each product in the path; 2. Number of clicks: through real-time statistics of user click behavior tracking points, each click event will be recorded, including product ID, click timestamp, etc.; 3. Click change rate: calculated in real time by the background, the formula is change rate = (number of clicks after adjustment − number of clicks before adjustment) / time interval. The time interval is calculated by the timestamps before and after, and the number of clicks is directly extracted from the tracking data. These data are calculated and stored in real time through the streaming processing pipeline, providing accurate input for the subsequent sorting adjustment amplitude index calculation.

[0053] Calculate the ranking adjustment index of each source path. The specific calculation formula is as follows: ; In the formula, For the The ranking adjustment index of the source paths; This formula calculates the sort adjustment amplitude index Quantify the degree of change in the ranking weight of products in the path before and after dynamic adjustment, and the correlation between this change and the dynamic change of user click behavior. Specifically, the core of the formula design is to comprehensively and dynamically capture the magnitude of the ranking weight adjustment, while combining the click change rate to reflect the effect of the adjustment. In the formula, Indicates the path The absolute value of the change in the ranking weight of a product before and after the adjustment is used to measure the impact of the adjustment mechanism on the product; the denominator It is the normalization term of the sorting weight, ensuring that the formula is applicable to products at different weight levels and avoiding excessive influence of weights with large absolute values ​​on the results. Introducing click rate change , through nonlinear amplification of logarithmic functions, the change of user click behavior has a moderate amplification effect on the contribution of the index, while avoiding the damage of extreme click rate values ​​to the formula calculation. The sum of all products in the formula reflects the overall ranking adjustment of the path, dynamically capturing the comprehensive optimization effect of the path after adjustment. Through this calculation method, It can accurately evaluate the actual extent of the ranking weight adjustment and whether it is consistent with the changes in user behavior, thereby providing a scientific basis for subsequent optimization mechanisms.

[0054] No. The ranking adjustment index of source paths The size of directly reflects the adjustment range of the path signal dynamic adjustment mechanism on the ranking weight of the products in the path and its correlation with the change of user click behavior, and is closely related to whether the adjustment effect meets expectations. When is large, it means that the adjustment mechanism has made a large adjustment to the ranking weights of the products in the path, and this adjustment is closely related to the dynamic changes in the user's click behavior, indicating that the adjustment mechanism is relatively successful in optimizing the path ranking and can significantly improve the accuracy or diversity of recommendations; and when If it is small, it means that the adjustment is small or the adjusted ranking does not fully reflect the changes in user click behavior, which may mean that the adjustment mechanism does not work effectively. By analyzing the size of the path signal, we can accurately evaluate whether the path signal dynamic adjustment mechanism has achieved the expected optimization effect on the path, and adjust the mechanism parameters or optimization strategy accordingly.

[0055] The behavior deviation information in the pre-processed path dynamic feedback information of each source path is extracted, including the distribution ratio of the number of clicks of users on each product in each source path before adjustment, the distribution ratio after adjustment, and the change in the display weight value of each product in each source path before and after adjustment, and they are marked as , and , Indicates that the user is The source path is The distribution ratio of the number of clicks on each product before adjustment, Indicates that the user is The source path is The distribution ratio of the number of clicks on the products after adjustment, Indicates Source path The change in display weight of a product before and after adjustment; Extracting behavior deviation information from the path dynamic feedback information of each source path that has been preprocessed can be achieved through front-end tracking and back-end real-time log processing combined with a streaming data processing framework. The front-end tracking records each click behavior of the user in the path, captures key fields such as product ID, click timestamp, and source path identifier, and sends them to the back-end in real time; at the same time, the back-end records the display weight of each product in the path through the log of the sorting model, and regularly synchronizes the display weight values ​​before and after adjustment. Through streaming data processing tools (such as Apache Kafka or Flink), the system can associate the front-end user click behavior data with the display weight change data in the back-end. Before and after the execution of the path adjustment mechanism, the number of clicks on the products in the path is counted, and the distribution ratio of the number of clicks on each product is calculated (that is, the ratio of the number of product clicks to the total number of clicks in the path). In addition, by comparing the sorting model log, the change in display weight before and after the adjustment is extracted. These extraction steps are completed in the back-end automated data flow pipeline to ensure the real-time and accuracy of the data.

[0056] The quantitative data in the behavior deviation information includes three categories: 1. Click distribution ratio before adjustment: calculated by counting the number of clicks on each product in the path and dividing it by the total number of clicks in the path. For example, the number of clicks on each product is recorded in the user click log, and the system groups them according to the click timestamp and path identifier and calculates the distribution ratio; 2. Click distribution ratio after adjustment: also calculated by counting the number of clicks on each product in the adjusted path in the same way; 3. Display weight value change: calculated by comparing the display weight of each product in the path before and after the adjustment (recorded in the sorting model log), the formula is change = weight after adjustment - weight before adjustment. The acquisition of these quantitative data depends on the real-time recording of user behavior embedding data and sorting model logs, and the click behavior and display weight are grouped, counted and compared through the background data pipeline to achieve efficient behavior deviation information extraction and processing.

[0057] Calculate the click behavior deviation index of each source path. The specific calculation formula is as follows: ; In the formula, For the The click behavior deviation index of the source path.

[0058] This click behavior deviation index The calculation formula aims to quantify the change in the distribution of user click behavior within the path before and after the adjustment, and comprehensively evaluate the optimization effect of the dynamic adjustment mechanism of path signals on user click behavior by combining the impact of the change in display weight. Capture the change in the click distribution ratio of each product, and clearly measure the guiding role of the adjustment mechanism on the click behavior of each product in the path. Introduced display weight variation , the effect of display weight changes on click behavior deviation is amplified nonlinearly through logarithmic functions, while suppressing the interference of extreme weight changes on the formula results. This normalization process ensures that the index still has good stability and comparability in scenarios with large differences in path scale and display weight. In addition, the formula reflects the distribution deviation of path click behavior as a whole by weighted summation of all product deviation values, providing a scientific basis for quantifying the effect of dynamic adjustment. Through this calculation method, it is possible to clearly evaluate the effectiveness of the dynamic adjustment mechanism in optimizing the distribution of user click behavior and improving the accuracy and diversity of recommendations.

[0059] No. Click behavior deviation index of source paths The size of directly reflects the guiding and optimizing effect of the dynamic adjustment mechanism of path signals on the distribution of user click behaviors within the path, and is closely related to whether the adjustment effect meets expectations. When is large, it means that the distribution of user click behaviors in the path has changed significantly before and after the adjustment. This may indicate that the adjustment mechanism effectively redistributes user attention, such as reducing the proportion of repeated clicks on products and increasing the attention to unclicked products, thereby optimizing the diversity of recommended content. When it is small, it means that the distribution of click behaviors before and after the adjustment has changed little, which may mean that the adjustment mechanism fails to fully guide user behavior and the optimization effect of the path signal is limited. The size of the path signal dynamic adjustment mechanism can be used to quantitatively evaluate whether the path signal dynamic adjustment mechanism has successfully achieved the recommendation goals (such as the balance between accuracy and diversity) and provide a basis for further optimization of the mechanism.

[0060] In this embodiment, the ranking of each source path generated is adjusted by the amplitude index and click behavior deviation index Construct an adjustment effect evaluation model and generate the adjustment evaluation coefficients of each source path through weighted summation , and the adjusted evaluation coefficients of each source path generated Adjustment evaluation coefficient thresholds for each source path that are set in advance Compare and evaluate whether the adjustment effect of the path signal dynamic adjustment mechanism on each source path can meet expectations based on the comparison results, and optimize the path signal dynamic adjustment mechanism based on the evaluation results. The specific comparison and analysis are as follows: like , the adjustment effect of the path signal dynamic adjustment mechanism on the source path cannot meet the expectations, and the path signal dynamic adjustment mechanism needs to be optimized, including: re-evaluating the ranking weight adjustment rules of the products in the path, increasing the real-time data collection and analysis of the user's latest click behavior and display preference; dynamically adjusting the weighting parameters in the ranking adjustment amplitude index and the click behavior deviation index to optimize the balance between ranking and click distribution; re-setting the boundary conditions and priority rules for ranking weight adjustment; This situation means that the current dynamic adjustment mechanism of path signals has failed to effectively optimize the ranking weights or user click behavior distribution of products in the source path, resulting in a low match between the recommendation results and user interest signals. Specifically, the ranking adjustment range index and the click behavior deviation index show that the adjustment range is insufficient or the distribution change is not significant, which may cause the repeatedly clicked products in the path to be over-prioritized without fully considering the recommendation potential of other products. In this case, the recommendation results may cause users to feel that the content is monotonous or not relevant enough, reducing the user satisfaction and click-through conversion rate of the recommendation system. In addition, system resources may be wasted on recommending low-priority products, and fail to maximize the potential interest needs of users. Therefore, the mechanism must be optimized by re-evaluating rules, dynamically adjusting parameters, etc. to improve the recommendation effect.

[0061] Re-evaluating the ranking weight adjustment rules of the products in the path and real-time data collection and analysis can be achieved through a dynamic analysis model based on user behavior and a real-time data stream processing system. Specifically, it is necessary to first use a real-time behavior monitoring system to collect the user's latest click behavior and display preference data, which include click time, click frequency, display times, and product category preferences. Through streaming processing tools (such as Kafka or Flink), these data are transmitted to the background analysis module in real time. The background analysis module builds a behavioral feature model based on the user's click history, and re-evaluates the ranking weight adjustment rules of the products in the path through algorithms such as cluster analysis or association rule mining. For example, the initial ranking weight of the product is dynamically adjusted according to the user's click tendency for certain categories or brands. For the ranking adjustment amplitude index (PMI) and the click behavior deviation index (PCB), it is necessary to optimize their weighted parameters through experimental data. Specifically, the weight ratio of these two indexes in the ranking weight optimization and click distribution optimization can be dynamically adjusted through a multi-objective optimization algorithm to achieve a balance between accuracy and diversity. In addition, resetting the boundary conditions and priority rules for sorting weight adjustment can be accomplished by setting reasonable thresholds and priority algorithms, such as dynamically defining the minimum and maximum limits of weight adjustment based on user historical preference data to prevent extreme adjustments from causing imbalance in recommendation results.

[0062] By re-evaluating the ranking weight adjustment rules of the products in the path, we can ensure that the ranking rules can timely reflect the user's latest click behavior and display preferences, thereby improving the accuracy of the recommended content. At the same time, increasing the real-time data collection and analysis of the user's latest click behavior and display preferences can dynamically capture changes in user needs and avoid inaccurate recommendation results caused by data delays. Dynamically adjusting the weighted parameters in the ranking adjustment amplitude index and the click behavior deviation index can optimize the balance between accuracy and diversity in the recommendation system, thereby avoiding the problem of too single or too scattered recommended content. Resetting the boundary conditions and priority rules for ranking weight adjustment can avoid extreme situations where the ranking adjustment amplitude is too large or too small, ensuring that the recommendation results are optimized within a controllable range, thereby improving the stability of the recommendation system and user satisfaction. These optimization measures can comprehensively improve the responsiveness and adaptability of the dynamic adjustment mechanism of path signals, and ultimately achieve rapid response and accurate matching of the recommendation system to user needs.

[0063] like , the adjustment effect of the path signal dynamic adjustment mechanism on the source path can achieve the expected effect, and there is no need to optimize the path signal dynamic adjustment mechanism.

[0064] This situation shows that the current path signal dynamic adjustment mechanism has a significant adjustment effect on the source path and successfully optimizes the recommended content. Specifically, the ranking adjustment range index and click behavior deviation index show that the ranking weight adjustment range of the products in the path is reasonable and consistent with the changes in user behavior. The click distribution deviation shows that the adjustment mechanism effectively guides users to pay attention to more products or higher priority products. In this case, the recommendation results can accurately match the user's interest signals while providing diversified content, thereby improving user satisfaction and the system's click-through conversion rate. The optimized recommendation path effectively allocates system resources, ensures a positive feedback loop between user behavior and recommended content, and further improves the efficiency and overall value of the recommendation system.

[0065] The ranking adjustment index of each source path generated and click behavior deviation Construct an adjustment effect evaluation model and generate the adjustment evaluation coefficients of each source path through weighted summation The specific method is: first, set the weight coefficient according to the optimization goal of the recommendation system (such as accuracy first, diversity first, or a balance between the two). and . Weight Index reflecting the adjustment range of ranking Importance in evaluation, used to measure the direct guiding effect of adjusting the order of products in the path on user click behavior; weight Reflects click behavior deviation index The importance of is used to measure whether the change in click distribution within the path meets the optimization goal. Next, the and The weighted summation is performed according to the following formula: ;in , ensuring that the weights of the two are reasonable in the overall evaluation. The specific value of the weight coefficient can be trained through historical data or dynamically adjusted using a multi-objective optimization algorithm. For example, in a scenario where accuracy is the priority, The value of can be increased in scenarios that emphasize diversity. Through this weighted summation method, the adjustment effect evaluation model can comprehensively consider the role of the path signal dynamic adjustment mechanism in both ranking adjustment and click behavior optimization, providing a scientific and quantitative basis for evaluating the overall optimization effect of the path.

[0066] A comprehensive analysis is conducted on the application effect, adjustment records and optimization results of the path signal dynamic adjustment mechanism, and the path signal processing capability and dynamic adjustment mechanism of the recommendation ranking model are continuously optimized in combination with newly added user behavior data.

[0067] The comprehensive analysis of the application effect, adjustment records and optimization results of the dynamic adjustment mechanism of path signals can be achieved through the background data analysis module combined with the machine learning model. Specifically, the application effect data of the dynamic adjustment mechanism (such as the path evaluation coefficient TPG and the click-through rate and conversion rate of the recommendation system), adjustment records (including the changes in the path adjustment rules and parameters for each time) and optimization results (such as changes in recommendation accuracy and diversity) are first collected through the log recording system. These data are stored in a centralized database, and data integration and analysis are performed through periodic triggering of batch processing tasks or real-time streaming analysis tools (such as Apache Flink). Next, the collected data is modeled using data mining and machine learning methods, such as using decision trees or random forest models to analyze the impact of different adjustment strategies on user behavior; using time series analysis or regression models to predict the long-term optimization trend of the dynamic adjustment mechanism of path signals. In addition, the newly added user behavior data (such as click, browse and purchase records) are input into the recommendation sorting model in real time through the incremental learning method, and the path signal processing rules and adjustment mechanism parameters are dynamically adjusted to continuously improve the effect of the recommendation system.

[0068] By comprehensively analyzing the application effects, adjustment records, and optimization results, we can deeply understand the effect of the dynamic adjustment mechanism of path signals in actual applications, and provide a scientific basis for further optimizing the recommendation system. This analysis can help identify potential problems in the current adjustment strategy, such as whether the path signal processing rules are over-adjusted or under-adjusted, and whether the optimization results meet the dynamic trend of user needs. Combined with the newly added user behavior data, the dynamic response capability of the recommendation ranking model can be enhanced to ensure that the recommended content is always consistent with the user's latest interest preferences. At the same time, continuous optimization of path signal processing capabilities and dynamic adjustment mechanisms can help achieve adaptive learning of the recommendation system, avoid the problem of recommendation effect attenuation caused by fixed rules, and thus improve the long-term performance and user satisfaction of the recommendation system. This approach not only improves the accuracy and diversity of recommendations, but also effectively enhances the flexibility and continuous improvement capabilities of the recommendation system.

[0069] like Figure 2 The commodity search optimization system based on historical behavior data shown includes a user click behavior detection module, a path contribution evaluation module, a path signal adjustment mechanism construction module, a path dynamic feedback evaluation module, and a path signal comprehensive optimization module; The user click behavior detection module detects the user's click behavior in real time to identify whether the user has clicked on the same product multiple times through multiple recommendation source paths. If it is identified that the user has clicked on the same product multiple times through multiple recommendation source paths, the multiple source paths generated by the user during the time period of searching for the product are classified and marked to distinguish the click behavior characteristics of different source paths. The path contribution evaluation module obtains the click behavior information of each source path in real time, analyzes it after acquisition, evaluates the contribution strength of each source path to the user's real interest signal, and divides each source path into dominant path, decentralized path and mixed path according to the evaluation results; A path signal adjustment mechanism construction module is used to construct a path signal dynamic adjustment mechanism for adjusting different ranking weights of dominant path, decentralized path and mixed path source paths according to the division results of each source path; The path dynamic feedback evaluation module obtains the path dynamic feedback information of each source path in real time during the process of the path signal dynamic adjustment mechanism adjusting each source path, and analyzes it after acquisition to evaluate whether the adjustment effect of the path signal dynamic adjustment mechanism on each source path can meet expectations, and optimizes the path signal dynamic adjustment mechanism according to the evaluation results; The path signal comprehensive optimization module conducts a comprehensive analysis of the application effect, adjustment records and optimization results of the path signal dynamic adjustment mechanism, and continuously optimizes the path signal processing capability and dynamic adjustment mechanism of the recommendation ranking model based on the newly added user behavior data.

[0070] The above formulas are all dimensionless and numerical calculations. The formula is a formula for the most recent real situation obtained by collecting a large amount of data and performing software simulation. The preset parameters in the formula are set by technicians in this field according to actual conditions.

[0071] The above embodiments can be implemented in whole or in part by software, hardware, firmware or any other combination. When implemented by software, the above embodiments can be implemented in whole or in part in the form of a computer program product. The computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on a computer, the process or function described in the embodiment of the present application is generated in whole or in part. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium, or transmitted from one computer-readable storage medium to another computer-readable storage medium. For example, the computer instructions can be transmitted from one website, computer, server or data center to another website, computer, server or data center by wired or wireless (such as infrared, wireless, microwave, etc.). The computer-readable storage medium can be any available medium that can be accessed by a computer or a data storage device such as a server or data center that contains one or more available media sets. The available medium can be a magnetic medium (for example, a floppy disk, a hard disk, a tape), an optical medium (for example, a DVD), or a semiconductor medium. The semiconductor medium can be a solid-state hard disk.

[0072] It should be understood that in the various embodiments of the present application, the size of the serial numbers of the above-mentioned processes does not mean the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present application.

[0073] Those of ordinary skill in the art will appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of this application.

[0074] In the several embodiments provided in the present application, it should be understood that the disclosed systems and methods can be implemented in other ways. For example, the embodiments described above are only schematic. For example, the division of the units is only a logical function division. There may be other division methods in actual implementation, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be an indirect coupling or communication connection through some interfaces, devices or units, which can be electrical, mechanical or other forms.

[0075] The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed on multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0076] In addition, each functional unit in each embodiment of the present application may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit.

[0077] The above is only a specific implementation of the present application, but the protection scope of the present application is not limited thereto. Any person skilled in the art who is familiar with the present technical field can easily think of changes or substitutions within the technical scope disclosed in the present application, which should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.

Claims

1. A commodity search optimization method based on historical behavior data, characterized in that: The specific steps include: By detecting the user's click behavior in real time, it is identified whether the user has clicked on the same product multiple times through multiple recommendation source paths. If it is identified that the user has clicked on the same product multiple times through multiple recommendation source paths, the multiple source paths generated by the user during the time period of searching for the product are classified and marked to distinguish the click behavior characteristics of different source paths. Obtain the click behavior information of each source path in real time, analyze it after acquisition, evaluate the contribution strength of each source path to the user's real interest signal, and divide each source path into dominant path, decentralized path and mixed path according to the evaluation results; According to the division results of each source path, a path signal dynamic adjustment mechanism is constructed to adjust the dominant path, the decentralized path and the mixed path source paths with different sorting weights respectively; In the process of adjusting each source path by the path signal dynamic adjustment mechanism, the path dynamic feedback information of each source path is obtained in real time, and analyzed after acquisition to evaluate whether the adjustment effect of the path signal dynamic adjustment mechanism on each source path can meet expectations, and optimize the path signal dynamic adjustment mechanism according to the evaluation results; A comprehensive analysis is conducted on the application effect, adjustment records and optimization results of the path signal dynamic adjustment mechanism, and the path signal processing capability and dynamic adjustment mechanism of the recommendation ranking model are continuously optimized in combination with newly added user behavior data.

2. The commodity search optimization method based on historical behavior data according to claim 1, characterized in that: The click behavior information of each source path is obtained in real time, and analyzed after acquisition to evaluate the contribution strength of each source path to the user's real interest signal, and the source paths are divided into dominant paths, dispersed paths and mixed paths according to the evaluation results, which specifically includes the following steps: Obtain click behavior information of each source path in real time and perform preprocessing after acquisition; Extracting the path click distribution information and path click feature information from the preprocessed click behavior information of each source path, and analyzing them after extraction to generate the click concentration coefficient and click sparse index of each source path respectively; The click concentration coefficient and click sparse index of each source path are used to construct a path contribution intensity model, and the contribution evaluation coefficient of each source path is generated. The contribution evaluation coefficient of each source path is compared with the pre-set contribution evaluation coefficient threshold range, and the contribution intensity of each source path to the user's real interest signal is evaluated based on the comparison results. According to the evaluation results, each source path is divided into a dominant path, a dispersed path and a mixed path.

3. The commodity search optimization method based on historical behavior data according to claim 2 is characterized in that: The logic for obtaining the click concentration coefficient and click sparse index of each source path is as follows: Extract the path click distribution information from the preprocessed click behavior information of each source path, including the number of clicks on each product in each source path at different times within a period of time, the total number of clicks, and the deviation between the number of clicks on each product and the total number of clicks in the source path, and use the function , and To express, For time point, Indicates that within a period of time At the moment the user The source path to Number of clicks on products, Indicates that within a period of time At the moment the user The total number of clicks within source paths, Indicates that within a period of time At the moment the user The source path is The deviation between the number of clicks on a product and the total number of clicks in the source path. , , and are all positive integers, defining the time period as ; Calculate the click concentration coefficient of each source path. The specific calculation formula is as follows: ； In the formula, For the Click concentration coefficient of source paths; Extract the path click feature information from the preprocessed click behavior information of each source path, including the total number of different products clicked by users in each source path at different times within a period of time, the number of product categories, and the average number of clicks on all products by users in each source path, and use the function to calculate the click feature information according to the time series. , and To express, For time point, Indicates that within a period of time At the moment the user The total number of different products clicked within the source path, Indicates that within a period of time At the moment the user The number of product categories for different products clicked within the source path, Indicates that within a period of time At the moment the user The average number of clicks on all products in a source path; Calculate the click sparse index of each source path. The specific calculation formula is as follows: ； In the formula, For the The click sparseness index of the source path.

4. The commodity search optimization method based on historical behavior data according to claim 3 is characterized in that: The click concentration coefficient of each source path generated and click sparsity index Construct a path contribution intensity model and generate the contribution evaluation coefficient of each source path through weighted summation , and the contribution evaluation coefficients of each source path generated The contribution evaluation coefficient threshold interval is set in advance. A comparison is made, and the contribution strength of each source path to the user's real interest signal is evaluated based on the comparison results. Based on the evaluation results, each source path is divided into a dominant path, a dispersed path, and a mixed path. The specific comparison analysis and division are as follows: like , the contribution intensity of the source path to the user's real interest signal is low, then the source path is divided into a dispersed path; like , the contribution intensity of the source path to the user's real interest signal is medium, then the source path is classified as a mixed path; like , the contribution intensity of the source path to the user's real interest signal is high, then the source path is classified as a dominant path.

5. The commodity search optimization method based on historical behavior data according to claim 4 is characterized in that: According to the division results of each source path, a path signal dynamic adjustment mechanism is constructed to adjust the dominant path, the decentralized path and the mixed path source paths with different sorting weights, specifically: According to the division results of dominant path, decentralized path and mixed path, different ranking weight adjustment rules are set respectively to form a dynamic adjustment mechanism of path signal; this adjustment mechanism is based on the contribution evaluation coefficient of each source path and the optimization goal of the recommendation system, and automatically determines the ranking weight adjustment method and amplitude of the products in each path through pre-set rules; Different sorting strategy adjustments are made to the dominant path, dispersed path and mixed path source paths respectively, including: for the dominant path, the weight enhancement rule in the adjustment mechanism is used to increase the sorting weight of the clicked products in the path, and give priority to displaying the products that the user is clearly interested in; for the dispersed path, the weight dispersion rule in the adjustment mechanism is used to reduce the sorting weight of repeatedly clicked products, and at the same time increase the display weight of other unclicked products in the path to increase the diversity of recommended content; for the mixed path, the weight balance rule in the adjustment mechanism is used to maintain the sorting weight of the clicked products in the path to balance the accuracy and diversity of the recommendations.

6. The commodity search optimization method based on historical behavior data according to claim 5, characterized in that: In the process of adjusting each source path by the path signal dynamic adjustment mechanism, the path dynamic feedback information of each source path is obtained in real time, and analyzed after acquisition to evaluate whether the adjustment effect of the path signal dynamic adjustment mechanism on each source path can meet expectations, and optimize the path signal dynamic adjustment mechanism according to the evaluation result, which specifically includes the following steps: In the process of adjusting each source path by the path signal dynamic adjustment mechanism, the path dynamic feedback information of each source path is obtained in real time, and preprocessed after being obtained; Extracting the ranking adjustment information and behavior deviation information from the pre-processed path dynamic feedback information of each source path, and analyzing them after extraction to generate a ranking adjustment amplitude index and a click behavior deviation index of each source path respectively; An adjustment effect evaluation model is constructed for the ranking adjustment amplitude index and click behavior deviation index generated for each source path, and an adjustment evaluation coefficient for each source path is generated. The generated adjustment evaluation coefficient for each source path is compared with the pre-set adjustment evaluation coefficient threshold for each source path. Based on the comparison results, it is evaluated whether the adjustment effect of the path signal dynamic adjustment mechanism on each source path can meet expectations, and the path signal dynamic adjustment mechanism is optimized based on the evaluation results.

7. The commodity search optimization method based on historical behavior data according to claim 6, characterized in that: The logic for obtaining the ranking adjustment index and click behavior deviation index of each source path is as follows: Extract the sorting adjustment information from the preprocessed dynamic feedback information of each source path, including the sorting weight value of each product in each source path before dynamic adjustment, the sorting weight value after dynamic adjustment, and the rate of change of the number of clicks of each product in each source path before and after adjustment, and mark them as , and , Indicates Source path The ranking weight value of a product before dynamic adjustment. Indicates Source path The ranking weight value of each product after dynamic adjustment. Indicates that the user The source path is The rate of change of the number of clicks on a product per unit time before and after adjustment, , , and All are positive integers; Calculate the ranking adjustment index of each source path. The specific calculation formula is as follows: ； In the formula, For the The ranking adjustment index of the source paths; The behavior deviation information in the pre-processed path dynamic feedback information of each source path is extracted, including the distribution ratio of the number of clicks of users on each product in each source path before adjustment, the distribution ratio after adjustment, and the change in the display weight value of each product in each source path before and after adjustment, and they are marked as , and , Indicates that the user The source path is The distribution ratio of the number of clicks on each product before adjustment, Indicates that the user The source path is The distribution ratio of the number of clicks on the products after adjustment, Indicates Source path The change in display weight of a product before and after adjustment; Calculate the click behavior deviation index of each source path. The specific calculation formula is as follows: ； In the formula, For the The click behavior deviation index of the source path.

8. The commodity search optimization method based on historical behavior data according to claim 7, characterized in that: The ranking adjustment index of each source path generated and click behavior deviation index Construct an adjustment effect evaluation model and generate the adjustment evaluation coefficients of each source path through weighted summation , and the adjusted evaluation coefficients of each source path generated Adjustment evaluation coefficient thresholds for each source path that are set in advance Compare and evaluate whether the adjustment effect of the path signal dynamic adjustment mechanism on each source path can meet expectations based on the comparison results, and optimize the path signal dynamic adjustment mechanism based on the evaluation results. The specific comparison and analysis are as follows: like , the adjustment effect of the path signal dynamic adjustment mechanism on the source path cannot meet the expectations, and the path signal dynamic adjustment mechanism needs to be optimized, including: re-evaluating the ranking weight adjustment rules of the products in the path, increasing the real-time data collection and analysis of the user's latest click behavior and display preference; dynamically adjusting the weighting parameters in the ranking adjustment amplitude index and the click behavior deviation index to optimize the balance between ranking and click distribution; re-setting the boundary conditions and priority rules for ranking weight adjustment; like , the adjustment effect of the path signal dynamic adjustment mechanism on the source path can achieve the expected effect, and there is no need to optimize the path signal dynamic adjustment mechanism.

9. A commodity search optimization system based on historical behavior data, used to implement the commodity search optimization method based on historical behavior data as described in any one of claims 1 to 8, characterized in that: It includes user click behavior detection module, path contribution evaluation module, path signal adjustment mechanism construction module, path dynamic feedback evaluation module and path signal comprehensive optimization module; The user click behavior detection module detects the user's click behavior in real time to identify whether the user has clicked on the same product multiple times through multiple recommendation source paths. If it is identified that the user has clicked on the same product multiple times through multiple recommendation source paths, the multiple source paths generated by the user during the time period of searching for the product are classified and marked to distinguish the click behavior characteristics of different source paths. The path contribution evaluation module obtains the click behavior information of each source path in real time, analyzes it after acquisition, evaluates the contribution strength of each source path to the user's real interest signal, and divides each source path into dominant path, decentralized path and mixed path according to the evaluation results; A path signal adjustment mechanism construction module is used to construct a path signal dynamic adjustment mechanism for adjusting different ranking weights of dominant path, decentralized path and mixed path source paths according to the division results of each source path; The path dynamic feedback evaluation module obtains the path dynamic feedback information of each source path in real time during the process of the path signal dynamic adjustment mechanism adjusting each source path, and analyzes it after acquisition to evaluate whether the adjustment effect of the path signal dynamic adjustment mechanism on each source path can meet expectations, and optimizes the path signal dynamic adjustment mechanism according to the evaluation results; The path signal comprehensive optimization module conducts a comprehensive analysis of the application effect, adjustment records and optimization results of the path signal dynamic adjustment mechanism, and continuously optimizes the path signal processing capability and dynamic adjustment mechanism of the recommendation ranking model based on the newly added user behavior data.

Images