An automated search for customers method and system

Abstract

The application discloses an automatic search customer acquisition method and system, which comprises the following steps: S1, generating a plurality of search units for customer acquisition scene requirements based on a search strategy comprising a plurality of search parameters; S2, generating artificial feedback scores and initial scores for the search units; S3, constructing an initial probability distribution model for each search parameter based on the initial scores and generating an initial benchmark score, and calculating an initial comprehensive feedback score for each search unit based on the initial benchmark score; S4, training the probability distribution model to update the benchmark score under the corresponding search parameter; performing next page search on the search unit with the highest comprehensive feedback score and manually scoring the obtained page data to update the artificial feedback score of the search unit; updating the comprehensive feedback scores of the search units; stopping training if a preset condition is met; and S5, screening the plurality of search units based on the comprehensive feedback scores and distributing them to different time periods for activation, and users search through the activated search units.

Description

Technical Field

[0001] This invention relates to the field of automated marketing technology, and in particular to an automated search customer acquisition method and system. Background Technology

[0002] Automated search customer acquisition refers to the process by which businesses use programmatic search logic to uncover the specific needs and preferences of each potential customer across various platforms such as web pages and electronic maps, and then deliver marketing information to customers who are likely to purchase products or services, thereby achieving the goal of customer acquisition.

[0003] Currently, the main automated search customer acquisition solutions include the following two types: Option 1: Search-based customer acquisition solution based on traditional rules. This solution relies solely on literal matching of basic words to generate search terms, lacking semantic association analysis and expansion, resulting in limited and singular search terms with limited coverage. Furthermore, this solution often employs a pre-defined fixed-weight retrieval strategy (e.g., assuming the first page of search results is always more important than the tenth page) to calculate the priority of search units (including search terms, device type, etc.). This static retrieval strategy is difficult to adjust in real-time based on customer behavior data, market dynamics, and other variables, affecting the accuracy and recall rate of potential customer discovery and failing to adapt to the rapidly changing business environment and differentiated customer acquisition needs.

[0004] Option 2: AI-assisted search customer acquisition solution. While AI technology is introduced, it is typically only applied to specific parts of traditional processes, failing to achieve system-level optimization. The main drawbacks are: ① Insufficient search term reliability. Search terms generated by AI models may deviate from core business intent and lack effective real-time feedback and verification mechanisms, affecting search targeting. ② Search term priority calculation also relies on preset static retrieval strategies. ③ Lack of model self-updating and iteration capabilities. The model lacks the ability to automatically learn and continuously optimize based on historical customer acquisition results and user interaction feedback; over the long term, model performance and customer acquisition accuracy will gradually decline. ④ Lack of personalized targeting capabilities. Fixed model parameters prevent the understanding or adaptation to the personalized customer acquisition preferences and needs of specific users, resulting in a lack of targeting in the customer acquisition process and difficulty in meeting differentiated marketing goals.

[0005] In conclusion, providing a smarter, more efficient, and self-learning and continuously optimizing automated search customer acquisition method is of urgent need and significant importance for addressing current industry pain points and improving the intelligence level of enterprise marketing.

[0006] The statements herein provide only background information in relation to this invention and do not necessarily constitute prior art. Summary of the Invention

[0007] The purpose of this invention is to provide an automated search-based customer acquisition method and system. This system can automatically generate multiple search units highly relevant to the needs of customer acquisition scenarios. Based on initial scores randomly generated for each search unit, an initial probability distribution model is constructed for each search parameter. The probability model is then trained multiple times based on human feedback scores generated for each search unit, enabling the trained probability distribution model to calculate accurate comprehensive feedback scores for the search units. By activating multiple search units with high comprehensive feedback scores at different time periods, the matching degree between search units and users is greatly improved. Within a corresponding time period, when a user inputs a customer acquisition scenario requirement, a search is performed on the internet based on the activated search units corresponding to that time period, returning search results with a high proportion of valid entries, thereby achieving efficient user outreach.

[0008] To achieve the above objectives, this invention provides an automated search and customer acquisition method, comprising the following steps: S1. In historical search scenarios, input customer acquisition scenario requirements into the Large Language Model (LLM); the Large Language Model (LLM) generates multiple search units based on the customer acquisition scenario requirements and preset search strategies; the search strategies include multiple types of search parameters; S2. Based on the search unit, search online, obtain the first page of paginated data of the search results, and perform manual scoring and random scoring respectively to obtain the manual feedback score and initial score of the search unit; S3. Based on the initial scores, establish an initial probability distribution model for each search parameter; generate an initial benchmark score for each search parameter based on the initial probability distribution model; and calculate an initial comprehensive feedback score for each search unit based on the initial benchmark scores. S4. Based on the comprehensive feedback score and the manual feedback score, train the probability distribution model using the gradient boosting algorithm to update the baseline score under the corresponding search parameters; perform a next page search on the search unit with the highest comprehensive feedback score, and manually score the paginated data obtained from the search to update the manual feedback score of the corresponding search unit; update the comprehensive feedback score of each search unit based on the updated baseline score. Calculate the sum of the differences between the comprehensive feedback score and the corresponding human feedback score of all search units; if the sum of the differences is less than a preset difference threshold, or the number of training iterations reaches a preset number threshold, stop training the probability distribution model; S5. Sort all search units from high to low according to the comprehensive feedback score; select the top-ranked search units and allocate them to different time periods for activation; During the corresponding time period, when the user inputs the customer acquisition scenario requirement, a search is performed on the Internet based on the activated search unit corresponding to that time period, and the search results are returned.

[0009] Optionally, step S1 includes: S11. In historical searches, input customer acquisition scenario requirements into the large language model; the large language model generates multiple basic words based on the customer acquisition scenario requirements, and generates multiple related words for the basic words based on semantic understanding; S12. Calculate the correlation coefficient between the base word and the corresponding related words using the cosine similarity algorithm, and filter out the related words whose correlation coefficient with the base word is higher than a preset coefficient threshold. S13. The selected related words are used as search terms, and multiple search units are generated for the search terms based on a preset search strategy; the search strategy includes multiple types of search parameters; the search parameters include at least one of the following: search terminal type parameters, search language parameters, search term parameters, and regional information parameters.

[0010] Optionally, in step S3, an initial probability distribution model is established for each search parameter, and an initial benchmark score is generated for each search parameter based on the initial probability distribution model, including the following steps: S31. The search parameter corresponds to multiple parameter values; for different parameter values ​​of the search parameter, benchmark data is calculated respectively; the benchmark data is the average of the initial scores of all corresponding search units under the corresponding parameter value; S32. Construct an initial probability distribution model using a log-Gaussian distribution as the baseline data: Let X represent the baseline data, which follows a log-Gaussian distribution; Y0 represents the initial probability distribution model, defined as Y0 = ln(X), where ln(·) is the natural logarithm function, and Y0 satisfies a normal distribution Y0 ~ N(μ,σ). 2 );μ、σ 2 Let μ0 and σ0 represent the mean and variance, respectively. Based on the initial baseline data, the initial mean μ0 and initial variance σ0 of the initial probability distribution model Y0 are obtained by solving the maximum likelihood estimation method. 2 ; S33. Obtain the log-Gaussian probability density function f0(X) corresponding to the probability distribution model Y0. ; exp(·) represents the natural exponential function; the benchmark data corresponding to the peak value of f0(X) is used as the initial benchmark score for the corresponding search parameter.

[0011] Optionally, in step S3, an initial comprehensive feedback score is calculated for each search unit based on the initial benchmark score, including the following steps: S34. Define the state set of a Markov chain. ;in: ~ This indicates that the pagination data from page 1 to page p of the search unit is open by the user; p represents the maximum number of pages the user can scroll down from the first page of the multi-page pagination data returned by the search unit. S35, Order Represents the j-th state from the state set S. Transform into the (j+1)th state The probability of; Perform iterative calculations

[0012] This is the sum of the initial baseline scores for each search parameter; As an intermediate variable; The weighted fusion baseline score is obtained from the initial probability distribution model; S36, Order This represents the real-time pagination score of the i-th search unit under the aforementioned customer acquisition scenario requirements. The percentage of valid entries in the first page of paginated data of the i-th search unit; M represents the total number of search units under the customer acquisition scenario requirements; The comprehensive feedback score of the i-th search unit obtained based on the initial probability distribution model is denoted as . ; ; These are the weighting coefficients. .

[0013] Optionally, in step S4, training the probability distribution model using the gradient boosting algorithm includes: Let Y n This represents the probability distribution model obtained after the nth training iteration; N is the preset threshold number of times; The search unit with the highest comprehensive feedback score in each round of training is recorded as the optimal search unit for that round. Loss n For Y n loss function, ; denoted as the comprehensive feedback score for the k-th page of paginated data of the optimal search unit obtained from the nth training, and m represents the total number of pages of paginated data for the optimal search unit; This represents the human feedback score for the k-th page of paginated data of the optimal search unit obtained from the nth training iteration; Calculate the mean increment Δμ. η is the learning rate; μ n+1 =μ n +Δμ; μ1 represents the result of updating μ0 during the first training iteration; μ n+1 For the (n+1)th training iteration, for μ n The result of the update.

[0014] Optionally, in step S4, updating the baseline score under the corresponding search parameters includes: Obtaining the probability distribution model Y n The corresponding log-Gaussian probability density function , ; Will The benchmark data corresponding to the peak value is used as the basis for the probability distribution model Y. n Update the baseline score for the corresponding search parameters.

[0015] Optionally, in step S4, updating the comprehensive feedback score of the search unit includes the following steps: S41. Perform iterative calculations.

[0016] For probability distribution model Y n The sum of the baseline scores for each search parameter obtained; As an intermediate variable; For probability distribution model Y n The obtained weighted fusion benchmark score; S42, Based on the probability distribution model Y n The comprehensive feedback score obtained for the i-th search unit is denoted as ; ; The percentage of valid entries in the latest paginated data of the i-th search unit.

[0017] Optionally, the automated search customer acquisition method further includes: limiting the increment value Δμ; when Δμ ≥ r1×μ n At that time, let Δμ = r1 × μ n When Δμ≤r2×μ n At that time, let Δμ = r2×μ n r1∈(0,1), r2∈(-1,0).

[0018] The present invention also provides an automated search customer acquisition system for implementing the automated search customer acquisition method described in the present invention, comprising: a search unit generation module, a probability distribution model generation and optimization module, and a comprehensive feedback score calculation module; The search unit generation module is configured with a large language model; the large language model generates multiple search units based on a preset search strategy for the input customer acquisition scenario requirements; the search strategy includes multiple types of search parameters; The probability distribution model generation and optimization module generates initial probability distribution models for various search parameters based on the initial scores of each search unit. The comprehensive feedback score calculation module calculates an initial baseline score for each search parameter based on the initial probability distribution model, and calculates a weighted fusion baseline score corresponding to the search strategy based on the baseline score; by summing the real-time pagination data score of the search unit with the weighted fusion baseline score, the comprehensive feedback score of the search unit is obtained. The probability distribution model generation and optimization module also trains the probability distribution model based on the human feedback scores and comprehensive feedback scores of each search unit; the comprehensive feedback score calculation module also updates the baseline score and the comprehensive feedback score based on the probability distribution model updated in each round of training during the training process. The initial score and the manual feedback score are the results of random scoring and manual scoring of the first page of paginated data obtained by the search unit from the web search, respectively; the real-time paginated data score is the proportion of valid entries in the latest search paginated data of the search unit.

[0019] Optionally, the comprehensive feedback score calculation module calculates a weighted fusion benchmark score corresponding to the search strategy based on the Markov chain and the benchmark score, and sums the real-time pagination data score of the search unit with the weighted fusion benchmark score to obtain the corresponding comprehensive feedback score.

[0020] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention leverages a large language model to automatically generate basic and related words adapted to customer acquisition scenarios. It calculates the correlation coefficient between related words and basic words using cosine similarity, thereby selecting related words with high correlation coefficients to the basic words as search terms. Through the dual guarantee of "semantic association + data support," this invention overcomes the limitations of traditional manual search term setting, significantly improving the semantic matching degree between search terms and customer acquisition needs, reducing the generation rate of invalid search units, and laying the foundation for subsequent precise customer acquisition.

[0021] This invention generates random initial scores and human feedback scores from the first page of data returned by the search unit. Based on these initial scores, an initial probability distribution model is constructed for each search parameter of the search strategy. Multiple rounds of model training are then conducted using the human feedback scores to achieve continuous optimization and iteration of the probability distribution model, thereby generating benchmark scores for each search parameter. The calculation of the benchmark scores conforms to the skewed characteristics of the data, significantly improving its accuracy. The probability distribution model of this invention has strong self-updating and iterative capabilities, possessing the ability to automatically learn and continuously optimize based on historical customer acquisition results and user interaction feedback; the model performance does not decay over time.

[0022] In this invention, during model training, the optimal search unit is selected in each round to search for the next page of paginated data. The newly obtained paginated data is then manually scored to update the corresponding search unit's human feedback score. This invention eliminates the need for manual scoring of all search units in each training round, significantly reducing the number of manual scoring operations and paginated data searches, thus improving model training efficiency while ensuring model effectiveness.

[0023] This invention uses the baseline scores of each search parameter to obtain a weighted fusion baseline score under the corresponding search strategy through Markov state transition probabilities. This weighted fusion baseline score reflects the dynamic relationship between the depth of user page browsing and customer acquisition, enabling dynamic adaptation between pagination depth and pagination data quality.

[0024] The weighted fusion benchmark score is linked to each search parameter. Compared to a single scoring method, the comprehensive feedback score obtained based on the weighted fusion benchmark score is more in line with the actual customer acquisition needs. Search parameters can be freely configured, which not only meets personalized customer acquisition needs, but also greatly improves the flexibility and dynamic adjustment capabilities of customer acquisition, making it suitable for achieving differentiated marketing goals.

[0025] This invention calculates the comprehensive feedback score of each search unit by summing the real-time pagination data score of each search unit with the weighted fusion benchmark score, balancing historical accumulated search results with real-time search results, thereby obtaining better search units and improving the effective reach rate of push pages.

[0026] Based on comprehensive feedback scores, high-quality search units are activated in different time periods, enabling dynamic and differentiated configuration of search resources. This effectively improves the matching accuracy between search units and user needs at different times, avoids interference from invalid search units, and enhances the targeting and flexibility of customer acquisition matching.

[0027] The automated search and customer acquisition method of this invention operates intelligently throughout the entire process, requiring minimal manual intervention, thus reducing customer acquisition operation costs. It not only improves the overall stability and sustainable optimization capabilities of the customer acquisition process but also adapts to diverse and dynamically changing customer acquisition scenarios. Attached Figure Description

[0028] Figure 1 This is a flowchart of an automated search customer acquisition method in an embodiment of the present invention.

[0029] Figure 2 This is a schematic diagram of an automated search and customer acquisition system in an embodiment of the present invention. Detailed Implementation

[0030] The automated search and customer acquisition method and system proposed in this invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. The advantages and features of this invention will become clearer from the following description. It should be noted that the drawings are in a very simplified form and use non-precise proportions, only for the purpose of conveniently and clearly illustrating the embodiments of this invention. Please refer to the drawings to make the objectives, features, and advantages of this invention more apparent and understandable. It should be understood that the structures, proportions, sizes, etc., depicted in the accompanying drawings are only for the purpose of assisting those skilled in the art in understanding and reading the content disclosed in the specification, and are not intended to limit the implementation conditions of this invention. Therefore, they have no substantial technical significance. Any modifications to the structure, changes in the proportional relationships, or adjustments to the size, without affecting the effects and objectives achieved by this invention, should still fall within the scope of the technical content disclosed in this invention.

[0031] Customer acquisition is the process by which businesses attract potential customers and convert them into actual customers through various means and channels. The core objectives are to increase the number of customers and improve their quality. Existing search-based customer acquisition methods generally have several shortcomings: weak correlation between search terms and marketing objectives, simplistic search strategies that fail to meet personalized marketing goals, and a lack of dynamic adaptability in search unit priority calculation strategies. These shortcomings directly result in a low percentage of effective entries in the paginated results obtained by customers through search terms, ultimately leading to low customer acquisition efficiency and poor conversion rates.

[0032] The purpose of this invention is to provide an automated search-based customer acquisition method and system. This system can automatically generate multiple search units highly relevant to the needs of customer acquisition scenarios. Based on initial scores randomly generated for each search unit, an initial probability distribution model is constructed for each search parameter. The probability model is then trained multiple times based on human feedback scores generated for each search unit, enabling the trained probability distribution model to calculate accurate comprehensive feedback scores for the search units. By activating multiple search units with high comprehensive feedback scores at different time periods, the matching degree between search units and users is greatly improved. Within a corresponding time period, when a user inputs a customer acquisition scenario requirement, a search is performed on the internet based on the activated search units corresponding to that time period, returning search results with a high proportion of valid entries, thereby achieving efficient user outreach.

[0033] like Figure 1As shown, the present invention provides an automated search customer acquisition method, including the following steps: S1. In historical search scenarios, input customer acquisition scenario requirements into the large language model; the large language model generates multiple search units based on the customer acquisition scenario requirements and preset search strategies; the search strategies include multiple types of search parameters.

[0034] In this embodiment, step S1 includes: S11. In historical searches, input customer acquisition scenario requirements into the large language model; the large language model generates multiple basic words based on the customer acquisition scenario requirements, and generates multiple related words for the basic words based on semantic understanding; S12. Calculate the correlation coefficient between the base word and the corresponding related words using the cosine similarity algorithm, and filter out the related words whose correlation coefficient with the base word is higher than a preset coefficient threshold. S13. The selected related words are used as search terms, and multiple search units are generated for the search terms based on a preset search strategy; the search strategy includes multiple types of search parameters.

[0035] This invention leverages a large language model to automatically generate basic and related words adapted to customer acquisition scenarios. It calculates the correlation coefficient between related words and basic words using cosine similarity, thereby selecting related words with high correlation coefficients to the basic words as search terms. Through the dual guarantee of "semantic association + data support," this invention overcomes the limitations of traditional manual search term setting, significantly improving the semantic matching degree between search terms and customer acquisition needs, reducing the generation rate of invalid search units, and laying the foundation for subsequent precise customer acquisition.

[0036] Search parameters can be configured according to actual needs. In this embodiment, the search parameters include at least one of the following: search terminal type parameter, search language parameter, search term parameter, and geographic information parameter. Each search parameter has multiple parameter values. For example, the parameter values ​​for the search terminal type parameter include "Taobao", "Xiaohongshu", and "Baidu Maps"; the parameter values ​​for the search language parameter include "Chinese", "English", and "French"; and the parameter values ​​for the geographic information parameter include "Shanghai", "Beijing", and "London".

[0037] In this invention, the number of basic words, the number of related words, and the number of parameter values ​​for each search parameter can all be freely set according to requirements.

[0038] S2. Based on the search unit, perform a search online, obtain the first page of search results, and then manually score and randomly score the paginated data to obtain the manual feedback score and initial score of the search unit.

[0039] In one embodiment, a predefined user feedback indicator system is used for manual scoring. This scoring includes: scoring the effectiveness of customer matching (F1: 0-10 points, higher score for higher matching degree), scoring the satisfaction with search results (F2: 0-10 points, higher score for higher satisfaction), and scoring the percentage of invalid searches (F3: 0-10 points, higher score for lower percentage). Users submit feedback through the system's interactive interface (e.g., by checking "effective match" or "unsatisfactory"), and the system automatically records the scores F1, F2, and F3 for each search unit. Weights can also be assigned to each score, resulting in the manual feedback score F = F1 × 0.5 + F2 × 0.3 + F3 × 0.2 (the weights of F1, F2, and F3 are set according to business needs; this is only an example).

[0040] Table 1

[0041] Table 1 shows multiple search units in one embodiment, along with their corresponding initial scores and human feedback scores. In this embodiment, the input customer acquisition scenario requirement is "auto parts supplier," extracting the basic keyword "auto parts," and then generating related keywords such as "new energy vehicle parts," "off-road vehicle parts," "commercial vehicle parts," and "bus parts." The search strategy in this embodiment includes three search parameters (search terminal type parameter, search language parameter, and search term parameter). The search unit is a combination of "search term + search terminal type + search language," such as "off-road vehicle parts + Taobao + Japanese." A search unit can be understood as a search path. Searching online based on the search unit can return multiple pages of paginated data, with each page containing multiple entries.

[0042] S3. Based on the initial scores, establish an initial probability distribution model for each search parameter; generate an initial benchmark score for each search parameter based on the initial probability distribution model; and calculate an initial comprehensive feedback score for each search unit based on the initial benchmark scores.

[0043] In step S3, an initial probability distribution model is established for each search parameter, and an initial benchmark score is generated for each search parameter based on the initial probability distribution model, including the following steps: S31. The search parameter corresponds to multiple parameter values; for different parameter values ​​of the search parameter, benchmark data is calculated respectively; the benchmark data is the average of the initial scores of all corresponding search units under the corresponding parameter value.

[0044] Taking Table 1 as an example, when the value of the "Search Terminal Type Parameter" is "Taobao", the baseline data for the parameter value "Taobao" is: The parameter value "Taobao" received a human feedback score of [score missing]. .

[0045] S32. Construct an initial probability distribution model using a log-Gaussian distribution as the baseline data: Let X represent the baseline data, which follows a log-Gaussian distribution; Y0 represents the initial probability distribution model, defined as Y0 = ln(X), where ln(·) is the natural logarithm function, and Y0 satisfies a normal distribution Y0 ~ N(μ,σ). 2 );μ、σ 2 Let μ0 and σ0 represent the mean and variance, respectively. Based on the initial baseline data, the initial mean μ0 and initial variance σ0 of the initial probability distribution model Y0 are obtained by solving the maximum likelihood estimation method. 2 .

[0046] S33. Obtain the log-Gaussian probability density function f0(X) corresponding to the probability distribution model Y0. ; exp(·) represents the natural exponential function; the benchmark data corresponding to the peak value of f0(X) is used as the initial benchmark score for the corresponding search parameter.

[0047] In this embodiment, the initial baseline values ​​for the search terminal type parameter, search language parameter, and search term parameter are denoted as R_n0, R_m0, and R_w0, respectively.

[0048] In step S3, an initial comprehensive feedback score is calculated for each search unit based on the initial benchmark score, including the following steps: S34. Define the state set of a Markov chain. ;in: ~ This indicates that the pagination data from page 1 to page p of the search unit is open by the user; p represents the maximum number of pages the user can scroll down from the first page of the multi-page pagination data returned by the search unit. The specific value of p is obtained based on experience; in this embodiment, p=13.

[0049] S35, Order Represents the j-th state from the state set S. Transform into the (j+1)th state The probability of; Perform iterative calculations .

[0050] In this embodiment, the sum of the initial baseline scores for each search parameter is used. =R_n0+R_m0+R_w0. As an intermediate variable; The weighted fusion baseline score is obtained from the initial probability distribution model.

[0051] In this embodiment, Based on experience. In one embodiment, -0.5= .For example, =0.65, =0.6. This indicates that the deeper the pagination, the lower the probability of scrolling down to the next page.

[0052] S36, Order This represents the real-time pagination score of the i-th search unit under the aforementioned customer acquisition scenario requirements. The percentage of valid entries in the first page of paginated data of the i-th search unit; M represents the total number of search units under the customer acquisition scenario requirements; The comprehensive feedback score of the i-th search unit obtained based on the initial probability distribution model is denoted as . ; ; These are the weighting coefficients. In this embodiment, .

[0053] In one embodiment, It can be obtained through existing AI tools (such as Doubao) or through other means. No limitations are imposed in this invention.

[0054] S4. Based on the comprehensive feedback score and the manual feedback score, train the probability distribution model using the gradient boosting algorithm to update the baseline score under the corresponding search parameters; perform a next page search on the search unit with the highest comprehensive feedback score, and manually score the paginated data obtained from the search to update the manual feedback score of the corresponding search unit; update the comprehensive feedback score of each search unit based on the updated baseline score. Calculate the sum of the differences between the comprehensive feedback score and the corresponding human feedback score for all search units; if the sum of the differences is less than a preset difference threshold, or the number of training iterations reaches a preset number threshold, stop training the probability distribution model.

[0055] In step S4, updating the baseline score under the corresponding search parameters includes: Obtaining the probability distribution model Y n The corresponding log-Gaussian probability density function , ; Will The benchmark data corresponding to the peak value is used as the basis for the probability distribution model Y. nUpdate the baseline score for the corresponding search parameters.

[0056] In this embodiment, the probability distribution model Y n The initial baselines for the updated search terminal type parameter, search language parameter, and search term parameter are denoted as R_n. n 、R_m n 、R_w n .

[0057] In step S4, the comprehensive feedback score of the search unit is updated, including the following steps: S41. Perform iterative calculations. .

[0058] For probability distribution model Y n The sum of the baseline scores for each search parameter. In this embodiment, R_n n +R_m n +R_w n .

[0059] As an intermediate variable; For probability distribution model Y n The obtained weighted fusion benchmark score.

[0060] S42, Based on the probability distribution model Y n The comprehensive feedback score obtained for the i-th search unit is denoted as ; ; This represents the percentage of valid entries in the latest search pagination data of the i-th search unit.

[0061] In step S4, training the probability distribution model using the gradient boosting algorithm includes: Let Y n Let represent the probability distribution model obtained after the nth training iteration; the search unit with the largest comprehensive feedback score in each training round is denoted as the optimal search unit. Loss n For Y n loss function, , This represents the overall feedback score for the k-th page of data of the optimal search unit (which has m pages of page data) obtained from the nth training iteration; This represents the human feedback score for the k-th page of paginated data of the optimal search unit during the nth training iteration.

[0062] In another embodiment, , It is a constant. .

[0063] Calculate the mean increment Δμ. η is the learning rate, which ranges from 0.01 to 0.09. In this embodiment, η = 0.04.

[0064] μ n+1 =μ n +Δμ. μ1 is the result of updating μ0 in the first training iteration. n+1 For the (n+1)th training iteration, for μ n The result of the update.

[0065] In a preferred embodiment, the increment value Δμ is also limited: when Δμ ≥ r1 × μ n At that time, let Δμ = r1 × μ n When Δμ≤r2×μ n At that time, let Δμ = r2×μ n r1∈(0,1), r2∈(-1,0). Ensure μ n+1 The updated result still conforms to the statistical properties of a log-Gaussian distribution, avoiding extreme values ​​that could distort the benchmark score. In one embodiment, r1 = 0.15 and r2 = -0.1.

[0066] S5. Sort all search units from high to low according to the comprehensive feedback score; select the top-ranked search units and allocate them to different time periods for activation; During the corresponding time period, when the user inputs the customer acquisition scenario requirement, a search is performed on the Internet based on the activated search unit corresponding to that time period, and the search results are returned.

[0067] This invention activates high-quality search units in different time periods based on comprehensive feedback scores. It also incorporates user activity time characteristics (based on historical data statistics: core active time periods for cross-border e-commerce are 9:00-12:00, 14:00-18:00, and 20:00-22:00) to formulate a time-segmented activation strategy, avoiding system peak pressure caused by concentrated activation. This invention enables dynamic and differentiated configuration of search resources, effectively improving the matching accuracy between search units and user needs at different times, avoiding interference from invalid search units, and enhancing the targeting and flexibility of customer acquisition matching.

[0068] In a preferred embodiment, a variety of search unit selection strategies can also be implemented: ① A maximum of 80 search units can be activated for a single terminal type (to avoid concentrating resources on a single terminal type); ② No more than 15 search units can be activated for a single search term (to avoid over-focusing on a certain semantic direction); ③ Invalid search units with a comprehensive feedback score of less than 3 points in the past 3 rounds of training are removed (based on feedback closed-loop data, to reduce the resource consumption of low-value units).

[0069] In another embodiment, 600 search units are selected. The top 500 are selected as search units to be activated that day, and the remaining 100 high-potential units are included in a "reserve pool" (used to replace units marked as invalid in customer acquisition that day).

[0070] This invention generates random initial scores and human feedback scores from the first page of data returned by the search unit. Based on these initial scores, an initial probability distribution model is constructed for each search parameter of the search strategy. Multiple rounds of model training are then conducted using the human feedback scores to continuously optimize and iterate the probability distribution model, thereby generating benchmark scores for each search parameter. The calculation of the benchmark scores conforms to the data skewness characteristics, significantly improving the accuracy of the benchmark scores. The probability distribution model of this invention has strong self-updating and iterative capabilities, possessing the ability to automatically learn and continuously optimize based on historical customer acquisition results and user interaction feedback; the model performance does not decay over time.

[0071] This invention uses the baseline scores of each search parameter to obtain a weighted fusion baseline score under the corresponding search strategy through Markov state transition probabilities. This weighted fusion baseline score reflects the dynamic relationship between the depth of user page browsing and customer acquisition, enabling dynamic adaptation between pagination depth and pagination data quality.

[0072] The weighted fusion benchmark score is linked to each search parameter. Compared to a single scoring method, the comprehensive feedback score obtained based on the weighted fusion benchmark score is more in line with the actual customer acquisition needs. Search parameters can be freely configured, which not only meets personalized customer acquisition needs, but also greatly improves the flexibility and dynamic adjustment capabilities of customer acquisition, making it suitable for achieving differentiated marketing goals.

[0073] This invention calculates the comprehensive feedback score of each search unit by summing the real-time pagination data score of each search unit with the weighted fusion benchmark score, balancing historical accumulated search results with real-time search results, thereby obtaining better search units and improving the effective reach rate of push pages.

[0074] The automated search and customer acquisition method of this invention operates intelligently throughout the entire process, requiring minimal manual intervention, thus reducing customer acquisition operation costs. It not only improves the overall stability and sustainable optimization capabilities of the customer acquisition process but also adapts to diverse and dynamically changing customer acquisition scenarios.

[0075] In this invention, during model training, the optimal search unit is selected in each round to search for the next page of paginated data. The newly obtained paginated data is then manually scored to update the corresponding search unit's human feedback score. This invention eliminates the need for manual scoring of all search units in each training round, significantly reducing the number of manual scoring operations and paginated data searches, thus improving model training efficiency while ensuring model effectiveness.

[0076] In one embodiment, three search units are generated using a large language model and a search strategy. First, the first page of pagination data for each search unit is obtained. Then, based on the updated model, an optimal search unit (with the highest overall feedback score) is found in each training round, and the next page of pagination data for that search unit is obtained. The model's training limit is set to 10 iterations. In the first training round, if the first search unit has the highest overall feedback score, its second page of pagination data is obtained; no new pagination data is obtained for the second and third search units (figuratively, the first search unit moves forward one step, while the second and third search units remain unchanged). In the second training round, if the third search unit has the highest overall feedback score, its second page of pagination data is searched; no new pagination data is obtained for the first and third search units (figuratively, the second search unit moves forward one step, while the first and third search units remain unchanged). After ten rounds of model updates, the first, second, and third search units have obtained 3, 3, and 4 pages of pagination data, respectively. In other words, in ten rounds of search, only the newly obtained 10 pages of paginated data need to be manually scored, and the model is updated based on these 10 pages of paginated data, which greatly improves the efficiency of model iteration.

[0077] In one embodiment, the present invention achieves the following: (1) The matching degree between search terms and customer needs is improved by more than 30% compared with the prior art (the matching degree of the prior art is about 45%, the initial state of the present invention is about 60%, and it is improved to 75% after 10 rounds of feedback iteration). (2) The accuracy of weight calculation is continuously optimized: the logarithmic Gaussian distribution modeling improves the initial correlation coefficient between the benchmark score and the customer conversion probability from 0.5 in the prior art to 0.8, and further improves to 0.92 after feedback closed-loop iteration. (3) The dynamic adaptation capability is enhanced, and the Markov chain state transition improves the pagination data utilization rate by 25%. (4) The customer acquisition stability is improved. The weighted fusion of benchmark data and real-time feedback reduces the fluctuation range of the search unit priority score from ±20% in the prior art to ±8%, and further reduces it to ±5% after feedback iteration, making the customer acquisition effect more stable. (5) The system operation efficiency is improved: targeted activation of several search units can effectively avoid peak traffic that is prone to system crash. (6) Significant closed-loop iteration effect: After every 10 rounds of feedback iteration, the overall customer acquisition accuracy of the system is improved by 12%, and the customer conversion rate is improved by 20% (compared to existing technologies), achieving a self-reinforcing effect of "the more you use it, the stronger it becomes".

[0078] like Figure 2 As shown, the present invention also provides an automated search customer acquisition system for implementing the automated search customer acquisition method described in the present invention. The system includes: a search unit generation module 10, a probability distribution model generation and optimization module 30, and a comprehensive feedback score calculation module 20.

[0079] The search unit generation module 10 is configured with a large language model; the large language model generates multiple search units based on a preset search strategy for the input customer acquisition scenario requirements; the search strategy includes multiple types of search parameters.

[0080] The probability distribution model generation and optimization module 30 generates an initial probability distribution model for various search parameters based on the initial scores of each search unit.

[0081] The comprehensive feedback score calculation module 20 calculates an initial baseline score for each search parameter based on the initial probability distribution model, and calculates a weighted fusion baseline score corresponding to the search strategy based on the baseline score; by summing the real-time pagination data score of the search unit with the weighted fusion baseline score, the comprehensive feedback score of the search unit is obtained.

[0082] The probability distribution model generation and optimization module 30 also trains the probability distribution model based on the human feedback scores and comprehensive feedback scores of each search unit; the comprehensive feedback score calculation module 20 also updates the benchmark score and the comprehensive feedback score based on the probability distribution model updated in each round of training during the training process.

[0083] The initial score and the manual feedback score are the results of random scoring and manual scoring of the first page of paginated data obtained by the search unit from the web search, respectively; the real-time paginated data score is the proportion of valid entries in the latest search paginated data of the search unit.

[0084] In one embodiment, the comprehensive feedback score calculation module 20 calculates a weighted fusion benchmark score corresponding to the search strategy based on the Markov chain and the benchmark score, and sums the real-time pagination data score of the search unit with the weighted fusion benchmark score to obtain the corresponding comprehensive feedback score.

[0085] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0086] Although the present invention has been described in detail through the preferred embodiments above, it should be understood that the above description should not be considered as a limitation of the present invention. Various modifications and substitutions to the present invention will be apparent to those skilled in the art after reading the above description. Therefore, the scope of protection of the present invention should be defined by the appended claims.

Claims

1. An automated search-based customer acquisition method, characterized in that, Including the following steps: S1. In historical search scenarios, input customer acquisition scenario requirements into the large language model; the large language model generates multiple search units based on the customer acquisition scenario requirements and preset search strategies; the search strategies include multiple types of search parameters. S2. Based on the search unit, search online, obtain the first page of paginated data of the search results, and perform manual scoring and random scoring respectively to obtain the manual feedback score and initial score of the search unit; S3. Based on the initial scores, establish an initial probability distribution model for each search parameter; generate an initial benchmark score for each search parameter based on the initial probability distribution model; and calculate an initial comprehensive feedback score for each search unit based on the initial benchmark scores. In step S3, an initial probability distribution model is established for each search parameter, and an initial benchmark score is generated for each search parameter based on the initial probability distribution model, including the following steps: S31. The search parameter corresponds to multiple parameter values; for different parameter values ​​of the search parameter, benchmark data is calculated respectively; the benchmark data is the average of the initial scores of all corresponding search units under the corresponding parameter value; S32. Construct an initial probability distribution model using a log-Gaussian distribution as the baseline data: Let X represent the baseline data, which follows a log-Gaussian distribution; Y0 represents the initial probability distribution model, defined as Y0 = ln(X), where ln(·) is the natural logarithm function, and Y0 satisfies a normal distribution Y0 ~ N(μ,σ). 2 );μ、σ 2 Let μ0 and σ0 represent the mean and variance, respectively. Based on the initial baseline data, the initial mean μ0 and initial variance σ0 of the initial probability distribution model Y0 are obtained by solving the maximum likelihood estimation method. 2 ; S33. Obtain the log-Gaussian probability density function f0(X) corresponding to the probability distribution model Y0. ； exp(·) represents the natural exponential function; the benchmark data corresponding to the peak value of f0(X) is used as the initial benchmark score for the corresponding search parameter; In step S3, an initial comprehensive feedback score is calculated for each search unit based on the initial benchmark score, including the following steps: S34. Define the state set of a Markov chain. ;in: ~ This indicates that the pagination data from page 1 to page p of the search unit is open by the user; p represents the maximum number of pages the user can scroll down from the first page of the multi-page pagination data returned by the search unit. S35, Order Represents the j-th state from the state set S. Transform into the (j+1)th state The probability of; Perform iterative calculations This is the sum of the initial baseline scores for each search parameter; As an intermediate variable; The weighted fusion baseline score is obtained from the initial probability distribution model; S36, Order This represents the real-time pagination score of the i-th search unit under the aforementioned customer acquisition scenario requirements. The percentage of valid entries in the first page of paginated data of the i-th search unit; M represents the total number of search units under the customer acquisition scenario requirements; The comprehensive feedback score of the i-th search unit obtained based on the initial probability distribution model is denoted as . ; ; These are the weighting coefficients. ; S4. Based on the comprehensive feedback score and the manual feedback score, train the probability distribution model using the gradient boosting algorithm to update the baseline score under the corresponding search parameters; perform a next page search on the search unit with the highest comprehensive feedback score, and manually score the paginated data obtained from the search to update the manual feedback score of the corresponding search unit; update the comprehensive feedback score of each search unit based on the updated baseline score. Calculate the sum of the differences between the comprehensive feedback score and the corresponding human feedback score of all search units; if the sum of the differences is less than a preset difference threshold, or the number of training iterations reaches a preset number threshold, stop training the probability distribution model; S5. Sort all search units from high to low according to the comprehensive feedback score; select the top-ranked search units and allocate them to different time periods for activation; During the corresponding time period, when the user inputs the customer acquisition scenario requirement, a search is performed on the Internet based on the activated search unit corresponding to that time period, and the search results are returned.

2. The automated search and customer acquisition method as described in claim 1, characterized in that, Step S1 includes: S11. In historical searches, input customer acquisition scenario requirements into the large language model; the large language model generates multiple basic words based on the customer acquisition scenario requirements, and generates multiple related words for the basic words based on semantic understanding; S12. Calculate the correlation coefficient between the base word and the corresponding related words using the cosine similarity algorithm, and filter out the related words whose correlation coefficient with the base word is higher than a preset coefficient threshold. S13. The selected related words are used as search terms, and multiple search units are generated for the search terms based on a preset search strategy; the search strategy includes multiple types of search parameters; the search parameters include at least one of the following: search terminal type parameters, search language parameters, search term parameters, and regional information parameters.

3. The automated search and customer acquisition method as described in claim 1, characterized in that, In step S4, training the probability distribution model using the gradient boosting algorithm includes: Let Y n This represents the probability distribution model obtained after the nth training iteration; N is the preset threshold number of times; The search unit with the highest comprehensive feedback score in each round of training is recorded as the optimal search unit for that round. Loss n For Y n loss function, ; denoted as the comprehensive feedback score for the k-th page of paginated data of the optimal search unit obtained from the nth training, and m represents the total number of pages of paginated data for the optimal search unit; This represents the human feedback score for the k-th page of paginated data of the optimal search unit obtained from the nth training iteration; Calculate the mean increment Δμ. η is the learning rate; m n+1 =m n +Dm; μ1 represents the result of updating μ0 during the first training iteration; μ n+1 For the (n+1)th training iteration, for μ n The result of the update.

4. The automated search and customer acquisition method as described in claim 3, characterized in that, In step S4, updating the baseline score under the corresponding search parameters includes: Obtaining the probability distribution model Y n The corresponding log-Gaussian probability density function , ； Will The benchmark data corresponding to the peak value is used as the basis for the probability distribution model Y. n Update the baseline score for the corresponding search parameters.

5. The automated search and customer acquisition method as described in claim 4, characterized in that, In step S4, the comprehensive feedback score of the search unit is updated, including the following steps: S41. Perform iterative calculations. ; For probability distribution model Y n The sum of the baseline scores for each search parameter obtained; As an intermediate variable; For probability distribution model Y n The obtained weighted fusion benchmark score; S42, Based on the probability distribution model Y n The comprehensive feedback score obtained for the i-th search unit is denoted as ; ; This represents the percentage of valid entries in the latest search pagination data of the i-th search unit.

6. The automated search and customer acquisition method as described in claim 3, characterized in that, Also includes: Limit the increment value Δμ; when Δμ ≥ r1×μ n At that time, let Δμ = r1 × μ n When Δμ≤r2×μ n At that time, let Δμ = r2×μ n r1∈(0,1), r2∈(-1,0).

7. An automated search customer acquisition system, used to implement the automated search customer acquisition method as described in any one of claims 1 to 6, characterized in that, include: The module includes a search unit generation module, a probability distribution model generation and optimization module, and a comprehensive feedback score calculation module. The search unit generation module is configured with a large language model; the large language model generates multiple search units based on a preset search strategy for the input customer acquisition scenario requirements; the search strategy includes multiple types of search parameters; The probability distribution model generation and optimization module generates initial probability distribution models for various search parameters based on the initial scores of each search unit. The comprehensive feedback score calculation module calculates an initial baseline score for each search parameter based on the initial probability distribution model, and calculates a weighted fusion baseline score corresponding to the search strategy based on the baseline score; by summing the real-time pagination data score of the search unit with the weighted fusion baseline score, the comprehensive feedback score of the search unit is obtained. The probability distribution model generation and optimization module also trains the probability distribution model based on the human feedback scores and comprehensive feedback scores of each search unit; the comprehensive feedback score calculation module also updates the baseline score and the comprehensive feedback score based on the probability distribution model updated in each round of training during the training process. The initial score and the manual feedback score are the results of random scoring and manual scoring of the first page of paginated data obtained by the search unit from the web search, respectively; the real-time paginated data score is the proportion of valid entries in the latest search paginated data of the search unit.

8. The automated search and customer acquisition system as described in claim 7, characterized in that, The comprehensive feedback score calculation module calculates the weighted fusion benchmark score corresponding to the search strategy based on the Markov chain and the benchmark score, and sums the real-time pagination data score of the search unit with the weighted fusion benchmark score to obtain the corresponding comprehensive feedback score.

Images