Personalized customer portrait construction and service recommendation method based on large language model

Abstract

The invention provides a personalized customer portrait construction and service recommendation method based on a large language model, and belongs to the field of personalized services. The problem that personalized services are difficult to realize is solved; the method specifically comprises the following steps: S1, acquiring seat information; dialogue data, behavior data and static data of the user are obtained; s2, calculating a call completing rate, an intention rate, an output rate and a complaint rate of the user, and constructing a user hierarchical model to obtain a user quality hierarchy; s3, classifying seats to obtain seat levels, and arranging the seats for the user; s4, based on an initial static portrait; combining dynamic behavior feature tags; forming a multi-dimensional user portrait, classifying the multi-dimensional user portrait to obtain a user group, and setting a dialing strategy according to the user group; s5, matching the arranged seats according to the crowd dialing strategy; according to the method, personalized portrait construction is carried out on the user, service recommendation is carried out on the user, and the service suitability is optimized.

Description

Technical Field

[0001] This invention relates to a method for constructing personalized customer profiles and recommending services based on a large language model, and pertains to the field of personalized services. Background Technology

[0002] Existing methods for building personalized customer profiles and recommending services have the following shortcomings: Insufficient personalization: Existing methods, by simply classifying users, fail to fully and accurately reflect users' true needs and preferences. This results in customer profiles that may be too one-sided and unable to provide personalized service recommendations to users.

[0003] Insufficient dynamism: Existing methods lack dynamism in the process of building user profiles, and cannot capture changes in user behavior and new needs in a timely manner; as user behavior and preferences continue to change, static customer profiles are difficult to maintain their effectiveness and accuracy.

[0004] Limited matching: Existing methods often rely on simple rule matching or manual intervention when recommending services, resulting in low matching efficiency and insufficient suitability between users and agents; they also consume more manpower and cannot meet users' personalized needs. Summary of the Invention

[0005] In view of the shortcomings of existing technologies, the purpose of this invention is to provide a method for constructing personalized customer profiles and recommending services based on a large language model, aiming to solve the complex problems of personalized services.

[0006] To achieve the above objectives, the present invention provides a method for constructing personalized customer profiles and recommending services based on a large language model, characterized in that the recommendation method includes: Step S1: Obtain agent information; obtain user dialogue data, behavioral data, and static data; Step S2: Calculate the user's connection rate and intent rate using user dialogue data; calculate the user's output rate and complaint rate using user behavior data; construct a user segmentation model based on the connection rate, intent rate, output rate, and complaint rate, segment users, and obtain user quality levels. Step S3: Based on the seat information and the user quality level, classify the seats to obtain the seat level, and assign seats to users according to the seat level and the user quality level. Step S4: Construct basic feature tags based on the user's static data to generate an initial static profile; synchronously parse the user's real-time dialogue data to extract dynamic behavior feature tags; dynamically correct the static profile through the dynamic behavior feature tags to form a multi-dimensional user profile; classify the user profile according to the user quality level to obtain the user group to which it belongs, and set different dialing strategies for different user groups. Step S5: Personalize the agent assignment based on the call strategy for different user groups, and the agents provide services to users based on user profiles.

[0007] Furthermore, the specific steps of step S2 are as follows: Step S21: Based on the user's dialogue data, obtain the number of dialogues initiated and the number of successful dialogues, and calculate the user's connection rate; through the number of successful dialogues, analyze the dialogue content to determine whether the user has a clear purchase intention, and calculate the user's intention rate. Step S22: Obtain user behavior data, obtain the number of products purchased by the user based on the user behavior data, calculate the user's output rate, obtain the number of user complaint behaviors, and calculate the complaint rate by combining the number of successful conversations; Step S23: Construct a user segmentation model based on connection rate, intent rate, output rate, and complaint rate; score users based on the user segmentation model, customize segmentation levels, and segment users according to the scoring results.

[0008] Furthermore, the specific steps of step S21 are as follows: Step S211: Obtain the number of conversations initiated (dh) and the number of successful conversations (cdh). Calculate the user's connection rate based on the number of conversations initiated (dh) and the number of successful conversations (cdh), and denote the user's connection rate as α. Step S212: Based on the number of successful conversations, extract the text of the conversations from the conversation data to obtain the conversation text. Process the conversation text using a large language model to determine whether the user has a clear purchase intention and obtain the number of conversations with clear intentions, mdh. Calculate the intention rate based on the number of successful conversations, cdh and the number of conversations with clear intentions, mdh, and denote the intention rate as β.

[0009] Furthermore, the specific steps of step S212 are as follows: The dialogue content in the dialogue data is extracted into text. The user's dialogue content is obtained by retrieval. The user's dialogue text is counted and denoted as wb(i). Combined with the number of successful dialogues, the dialogue text corresponding to each successful dialogue is obtained as wb(1), wb(2), ..., wb(cdh). Obtain the intention keywords, match the dialogue text with the intention keywords, determine whether the dialogue text expresses the purchase intention, and record the dialogue text that expresses the purchase intention as the dialogue text with clear intention; judge the dialogue texts wb(1) to wb(cdh), count the number of dialogue texts with clear intention, and obtain the number of dialogues with clear intention mdh.

[0010] Furthermore, the specific steps of step S22 are as follows: Step S221: Obtain the number of users y, and combine the user behavior data to obtain the quantity of goods purchased by each user, denoted as gm(1), gm(2), ..., gm(y). Extract the maximum value gmx and the minimum value gmi from the quantity of goods purchased by each user. Normalize the quantity of goods purchased by each user based on the maximum value gmx and the minimum value gmi from the quantity of goods purchased by each user to obtain the user's output rate θ. ; Where: gm(g) refers to the quantity of goods purchased by the g-th user; Step S222: Based on user behavior data, obtain the number of user complaint behaviors. Combined with the number of successful conversations, if the user's complaint behavior occurs after a successful conversation, the complaint is deemed valid; otherwise, the complaint is deemed invalid. Count the number of valid complaints, denoted as the complaint count ts. Calculate the complaint rate γ based on the proportion of the complaint count to the number of successful conversations.

[0011] Furthermore, the specific steps of step S23 are as follows: Step S231: Obtain the connection rate, intent rate, output rate, and complaint rate. Assign weights to these rates to obtain the connection rate weight qz1, intent rate weight qz2, output rate weight qz3, and complaint rate weight qz4, where qz1 + qz2 + qz3 + qz4 = 1. Based on the connection rate, intent rate, output rate, and complaint rate, and their corresponding weights qz1, qz2, qz3, and qz4, construct a user segmentation model. The specific model is as follows: ; Step S232: Substitute the user's connection rate, intention rate, output rate, and complaint rate into the user segmentation model to obtain the user score pf; define a custom segmentation level n, with each level having a value range of 1 / n. Set [n-1 / n, n] as the value range of the first level. Map the segmentation levels according to the user score. When pf∈[(n-t) / n, (n-t-1) / n], the user belongs to the t-th level, where t is a positive integer from 1 to n, thus obtaining the user quality level.

[0012] Furthermore, the specific steps of step S3 are as follows: Step S31: Obtain the number of agents through the agent information, obtain the complaint rate, user rating and effective communication number of each agent, score the agents according to the complaint rate, user rating and effective communication number of each agent, and sort the agents in descending order according to the scores; obtain the sorted list. Step S32: Obtain the user quality level, and map the seats to different user quality levels based on the number of seats. Classify the seats using the seat ranking list to obtain the seat level. Specifically: Obtain the user quality level tier n, obtain the number of seats zx, and divide the seats into equal levels according to the tier, with each user quality level corresponding to a seat number of zx / n; sort the seats in descending order of their ratings, and map the seats to different user quality levels to obtain the seat level; provide service to users of the same user quality level according to their seat level.

[0013] Furthermore, the specific steps of step S31 are as follows: Step S311: Based on the agent information, obtain the agent's complaint rate, user rating, and number of effective communications. Calculate the agent rating zpf based on the agent's complaint rate, user rating, and number of effective communications. Step S312: Obtain the number of seats zx, obtain the seat score for each seat, denoted as zpf(z), sort the seats according to the seat scores, obtain the first seat score as the comparison value bj, perform a loop comparison of the seat scores, and move the seat scores greater than the comparison value forward to obtain the descending order of the seat scores.

[0014] Furthermore, the subsequent steps of step S311 are as follows: The system collects the number of user ratings, analyzes each user rating, and combines it with the number of effective interactions with agents. If a user ratings an agent after an effective interaction, the rating is valid. If a user ratings an agent multiple times consecutively or without any interaction, the rating is invalid. The system records the number of valid ratings and the rating scores. Based on the number of valid ratings and the rating scores, the average user rating pjz is obtained. The complaint rate bts and the number of valid interactions jcs of the agents are obtained. The average user rating pjz is combined with the agent rating to calculate the agent rating zpf. .

[0015] Furthermore, the specific steps of step S4 are as follows: Step S41: Obtain the user's static data, extract keywords from the user's static data to obtain the user's static tags, count the user's static tags to obtain a static tag list, and generate an initial static profile from the static tag list. Step S42: Obtain user dialogue data, perform intent recognition on the dialogue data to obtain user intent keywords, use user intent keywords as dynamic behavioral feature labels to supplement the initial static profile, and obtain a multi-dimensional user profile. Step S43: Based on the multidimensional user profile, classify users, match the tags in the multidimensional user profile, obtain a multidimensional user profile as the standard profile, match the multidimensional user profile with the standard profile to obtain the matching value. If the matching value is greater than 0.5, treat it as a user group. If the matching value is less than 0.5, treat the multidimensional user profile as a new group and match other multidimensional user profiles.

[0016] Compared with the prior art, the beneficial effects of the present invention are: Strengthen matching rules: Use multi-faceted data to segment users by quality, and classify agents according to the user quality segmentation. Provide higher-level agents to high-quality users to provide better service and ensure the retention of high-quality users. Provide more suitable services based on the agent's service direction and the user's personalized profile.

[0017] Detailed personalized profile construction: By analyzing users' dialogue data, behavioral data, and static data, user profiles are constructed, reflecting users' basic attributes and revealing their potential needs and preferences, providing a solid data foundation for subsequent service recommendations; and providing strong support for improving user experience and service quality.

[0018] Real-time optimization: By continuously learning and updating the latest user data through user dialogue data, we ensure the timeliness and accuracy of personalized profiles; we correct deviations in the profiles in a timely manner through real-time optimization, avoiding service mismatch problems caused by information lag; and we enable service recommendations to keep up with changes in user needs and provide service solutions that always meet user expectations. Attached Figure Description

[0019] Other features, objects, and advantages of the invention will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings: Figure 1 This is a schematic diagram of the method of the present invention; Figure 2 This is a schematic diagram illustrating the data classification of the present invention; Figure 3 This is a schematic diagram of data matching in this invention. Detailed Implementation

[0020] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0021] Example 1 Please see Figure 1 Methods for building personalized customer profiles and recommending services based on large language models include: Step S1: Obtain agent information; obtain user dialogue data, behavioral data, and static data; It should be noted that: user dialogue data refers to real-time interaction records between users and the system / customer service, including text and voice. Behavioral data mainly refers to user actions within the product or service, such as clicks, browsing, purchases, and complaints. User static data refers to a user's basic attributes or long-term stable characteristics, which typically do not change frequently with interactive behavior.

[0022] It should be noted that a customer service representative, also known as a call center agent or operator, is a dedicated position established by a company's customer service department. They communicate with customers through channels such as telephone and the internet to provide product consultation, after-sales support, and problem solutions.

[0023] Step S2: Calculate the user's connection rate and intent rate using user dialogue data; calculate the user's output rate and complaint rate using user behavior data; construct a user segmentation model based on the connection rate, intent rate, output rate, and complaint rate, segment users, and obtain user quality levels. Step S21: Based on the user's dialogue data, obtain the number of dialogues initiated and the number of successful dialogues, and calculate the user's connection rate; through the number of successful dialogues, analyze the dialogue content to determine whether the user has a clear purchase intention, and calculate the user's intention rate. Step S211: Obtain the number of conversations initiated (dh) and the number of successful conversations (cdh). Calculate the user's connection rate based on the number of conversations initiated (dh) and the number of successful conversations (cdh), and denote the user's connection rate as α. ; Step S212: Based on the number of successful conversations, extract the text of the conversations from the conversation data to obtain the conversation text. Process the conversation text using a large language model to determine whether the user has a clear purchase intention and obtain the number of conversations with clear intentions, mdh. Calculate the intention rate based on the number of successful conversations, cdh and the number of conversations with clear intentions, mdh, and denote the intention rate as β. ; Step S2121: Extract the dialogue content from the dialogue data, retrieve the user's dialogue content by searching, and count the user's dialogue text, denoted as wb(i); combine the number of successful dialogues to obtain the dialogue text corresponding to each successful dialogue: wb(1), wb(2), ..., wb(cdh). Step S2122: Obtain intention keywords. Based on the large language model, match the dialogue text with the intention keywords and determine whether the dialogue text expresses purchase intention. Record the dialogue text that expresses purchase intention as the dialogue text with clear intention. Judge the dialogue texts wb(1), wb(2), ..., wb(cdh) and count the number of dialogue texts with clear intention to obtain the number of dialogues with clear intention mdh. It should be noted that: Intent keywords refer to words that can express the user's purchase intention during the conversation, such as: "buy", "place an order", "purchase", "how much", "how to buy"; Step S22: Obtain user behavior data, obtain the number of products purchased by the user based on the user behavior data, calculate the user's output rate, obtain the number of user complaint behaviors, and calculate the complaint rate by combining the number of successful conversations; Step S221: Obtain the number of users y, and combine the user behavior data to obtain the quantity of goods purchased by each user, denoted as gm(1), gm(2), ..., gm(y). Extract the maximum value gmx and the minimum value gmi from the quantity of goods purchased by each user. Normalize the quantity of goods purchased by each user based on the maximum value gmx and the minimum value gmi from the quantity of goods purchased by each user to obtain the user's output rate θ. ; It should be noted that gm(g) refers to any one of gm(1), gm(2), ..., gm(y), where g∈[1, y].

[0024] Step S222: Based on the user's behavior data, obtain the number of user complaint behaviors. Combined with the number of successful conversations, if the user's complaint behavior occurs after a successful conversation, the complaint is deemed valid; otherwise, the complaint is deemed invalid. Count the number of valid complaints, denoted as the number of complaints ts. Calculate the complaint rate γ based on the proportion of the number of complaints to the number of successful conversations. ; Where: cdh represents the number of successful conversations; Step S23: Construct a user segmentation model based on connection rate, intent rate, output rate, and complaint rate; score users based on the user segmentation model, customize segmentation levels, and segment users according to the scoring results; Step S231: Obtain the connection rate, intent rate, output rate, and complaint rate. Assign weights to these rates to obtain the connection rate weight qz1, intent rate weight qz2, output rate weight qz3, and complaint rate weight qz4, where qz1 + qz2 + qz3 + qz4 = 1. Based on the connection rate, intent rate, output rate, and complaint rate, and their corresponding weights qz1, qz2, qz3, and qz4, construct a user segmentation model. The specific model is as follows: ; Please see Figure 2 Step S232: Substitute the user's connection rate, intention rate, output rate, and complaint rate into the user segmentation model to obtain the user score pf; define a segmentation level n, with each level having a value range of 1 / n. Set [n-1 / n, n] as the value range of the first level. Map the segmentation levels according to the user score. When pf∈[(n-t) / n, (n-t-1) / n], the user belongs to the t-th level, where t is a positive integer from 1 to n, thus obtaining the user quality level. Step S3: Based on the seat information and the user quality level, classify the seats to obtain the seat level, and assign seats to users according to the seat level and the user quality level. Step S31: Obtain the number of agents through the agent information, obtain the complaint rate, user rating and effective communication number of each agent, score the agents according to the complaint rate, user rating and effective communication number of each agent, and sort the agents in descending order according to the scores; obtain the sorted list. Step S311: Based on the agent information, obtain the agent's complaint rate, user rating, and number of effective communications. Calculate the agent rating zpf based on the agent's complaint rate, user rating, and number of effective communications. Step S3111: Obtain the number of user ratings, analyze each user rating, and combine it with the number of effective interactions with the agent. If the user rating is given after an effective interaction, the rating is valid. If the user rating is given multiple times consecutively or without any interaction, the rating is invalid. Record the number of valid ratings and the rating score. Step S3112: Based on the number of valid ratings and the rating scores, obtain the average user rating pjz, obtain the agent's complaint rate bts and the number of valid interactions jcs, and combine the average user rating pjz to calculate the agent rating zpf. ; Step S312: Obtain the number of seats zx, obtain the seat score for each seat, denoted as zpf(z), sort the seats according to the seat scores, obtain the first seat score as the comparison value bj, compare the seat scores in a loop, and move the seat scores greater than the comparison value forward to obtain the descending order of the seat scores. Please see Figure 3 Step S32: Obtain the user quality level, combine it with the number of seats, map the seats to different user quality levels, and classify the seats through the seat sorting list to obtain the seat level; specifically as follows: obtain the user quality level tier n, obtain the number of seats zx, divide the seats into the same number of levels according to the tier, and the number of seats corresponding to each user quality level is zx / n; sort the seats in descending order according to the seat rating, map the seats to different user quality levels to obtain the seat level; provide services to users of the same user quality level according to the seat level.

[0025] Step S4: Construct basic feature tags based on the user's static data to generate an initial static profile; synchronously parse the user's real-time dialogue data to extract dynamic behavior feature tags; dynamically correct the static profile through the dynamic behavior feature tags to form a multi-dimensional user profile; classify the user profile according to the user quality level to obtain the user group to which it belongs, and set different dialing strategies for different user groups. Step S41: Obtain the user's static data, extract keywords from the user's static data to obtain the user's static tags, count the user's static tags to obtain a static tag list, and generate an initial static profile from the static tag list. Step S42: Obtain user dialogue data, perform intent recognition using an NLP model to obtain user intent keywords, use user intent keywords as dynamic behavioral feature labels to supplement the initial static profile, and obtain a multi-dimensional user profile. Step S43: Based on the multidimensional user profile, classify users, match the tags in the multidimensional user profile through word vector matching, obtain a multidimensional user profile as the standard profile, match the multidimensional user profile with the standard profile to obtain the matching value. If the matching value is greater than 0.5, it is regarded as a user group. If the matching value is less than 0.5, the multidimensional user profile is regarded as a new group, and other multidimensional user profiles are matched. Step S5: Personalize the agent assignment based on the call strategy for the target audience, and the agents provide services to users based on user profiles; Matching is performed based on the seat level and the user quality level, that is, matching a seat with a seat level of 1 with a user with a user quality level of 1. Based on the agent information, obtain the agent's service preferences, match the agent's service preferences with the dialing strategy for the target audience, ensure the efficient implementation of the strategy, and the agent provides specific services to the user based on the user profile; It should be noted that: agent service preferences refer to the agent's skill preferences, such as being good at proactively identifying needs, providing customized solutions, responding quickly, using standardized scripts, and proactively ending conversations.

[0026] The above formulas are all dimensionless calculations. The formulas are derived from software simulations using a large amount of collected data to obtain the most recent real-world results. The preset parameters in the formulas are set by those skilled in the art according to the actual situation. For example, there are weighting coefficients and proportional coefficients. The values ​​set are to quantify each parameter to obtain a specific value, which is convenient for subsequent comparison. The values ​​of the weighting coefficients and proportional coefficients are only required to not affect the proportional relationship between the parameters and the quantified values.

[0027] Finally, it should be noted that the above-described embodiments are merely specific implementations of the present invention, used to illustrate the technical solutions of the present invention, and not to limit it. The scope of protection of the present invention is not limited thereto. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that any person skilled in the art can still modify or easily conceive of changes to the technical solutions described in the foregoing embodiments within the technical scope disclosed in the present invention, or make equivalent substitutions for some of the technical features; and these modifications, changes, or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for constructing personalized customer profiles and recommending services based on a large language model, characterized in that, The recommendation method includes: Step S1: Obtain agent information; obtain user dialogue data, behavioral data, and static data; Step S2: Calculate the user's connection rate and intent rate using user dialogue data; calculate the user's output rate and complaint rate using user behavior data; construct a user segmentation model based on the connection rate, intent rate, output rate, and complaint rate, segment users, and obtain user quality levels. Step S3: Based on the seat information and the user quality level, classify the seats to obtain the seat level, and assign seats to users according to the seat level and the user quality level. Step S4: Construct basic feature tags based on the user's static data to generate an initial static profile; synchronously parse the user's real-time dialogue data to extract dynamic behavior feature tags; dynamically correct the static profile through the dynamic behavior feature tags to form a multi-dimensional user profile; classify the user profile according to the user quality level to obtain the user group to which it belongs, and set different dialing strategies for different user groups. Step S5: Personalize the agent assignment based on the call strategy for different user groups, and the agents provide services to users based on user profiles.

2. The method for constructing personalized customer profiles and recommending services based on a large language model according to claim 1, characterized in that, The specific steps of step S2 are as follows: Step S21: Based on the user's dialogue data, obtain the number of dialogues initiated and the number of successful dialogues, and calculate the user's connection rate; through the number of successful dialogues, analyze the dialogue content to determine whether the user has a clear purchase intention, and calculate the user's intention rate. Step S22: Obtain user behavior data, obtain the number of products purchased by the user based on the user behavior data, calculate the user's output rate, obtain the number of user complaint behaviors, and calculate the complaint rate by combining the number of successful conversations; Step S23: Construct a user segmentation model based on connection rate, intent rate, output rate, and complaint rate; Users are scored based on a user segmentation model, and the segmentation levels are customized. Users are segmented based on the scoring results.

3. The method for constructing personalized customer profiles and recommending services based on a large language model according to claim 2, characterized in that, The specific steps of step S21 are as follows: Step S211: Obtain the number of conversations initiated (dh) and the number of successful conversations (cdh). Calculate the user's connection rate based on the number of conversations initiated (dh) and the number of successful conversations (cdh), and denote the user's connection rate as α. Step S212: Based on the number of successful conversations, extract the text of the conversations from the conversation data to obtain the conversation text. Process the conversation text using a large language model to determine whether the user has a clear purchase intention and obtain the number of conversations with clear intentions, mdh. Calculate the intention rate based on the number of successful conversations, cdh and the number of conversations with clear intentions, mdh, and denote the intention rate as β.

4. The method for constructing personalized customer profiles and recommending services based on a large language model according to claim 3, characterized in that, The specific steps of step S212 are as follows: The dialogue content in the dialogue data is extracted into text. The user's dialogue content is obtained by retrieval. The user's dialogue text is counted and denoted as wb(i). Combined with the number of successful dialogues, the dialogue text corresponding to each successful dialogue is obtained as wb(1), wb(2), ..., wb(cdh). Obtain intent keywords, match the dialogue text with intent keywords, determine whether the dialogue text expresses purchase intention, and record the dialogue text that expresses purchase intention as dialogue text with clear intention. Judge the dialogue texts wb(1) to wb(cdh), count the number of dialogue texts with clear intentions, and obtain the number of dialogues with clear intentions mdh.

5. The method for constructing personalized customer profiles and recommending services based on a large language model according to claim 4, characterized in that, The specific steps of step S22 are as follows: Step S221: Obtain the number of users y, and combine the user behavior data to obtain the quantity of goods purchased by each user, denoted as gm(1), gm(2), ..., gm(y). Extract the maximum value gmx and the minimum value gmi from the quantity of goods purchased by each user. Normalize the quantity of goods purchased by each user based on the maximum value gmx and the minimum value gmi from the quantity of goods purchased by each user to obtain the user's output rate θ. ； Where: gm(g) refers to the quantity of goods purchased by the g-th user; Step S222: Based on user behavior data, obtain the number of user complaint behaviors. Combined with the number of successful conversations, if the user's complaint behavior occurs after a successful conversation, the complaint is deemed valid; otherwise, the complaint is deemed invalid. Count the number of valid complaints, denoted as the complaint count ts. Calculate the complaint rate γ based on the proportion of the complaint count to the number of successful conversations.

6. The method for constructing personalized customer profiles and recommending services based on a large language model according to claim 5, characterized in that, The specific steps of step S23 are as follows: Step S231: Obtain the connection rate, intent rate, output rate, and complaint rate. Assign weights to these rates to obtain the connection rate weight qz1, intent rate weight qz2, output rate weight qz3, and complaint rate weight qz4, where qz1 + qz2 + qz3 + qz4 = 1. Based on the connection rate, intent rate, output rate, and complaint rate, and their corresponding weights qz1, qz2, qz3, and qz4, construct a user segmentation model. The specific model is as follows: ； Step S232: Substitute the user's connection rate, intention rate, output rate, and complaint rate into the user segmentation model to obtain the user score pf; define a custom segmentation level n, with each level having a value range of 1 / n. Set [n-1 / n, n] as the value range of the first level. Map the segmentation levels according to the user score. When pf∈[(n-t) / n, (n-t-1) / n], the user belongs to the t-th level, where t is a positive integer from 1 to n, thus obtaining the user quality level.

7. The method for constructing personalized customer profiles and recommending services based on a large language model according to claim 1, characterized in that, The specific steps of step S3 are as follows: Step S31: Obtain the number of agents through the agent information, obtain the complaint rate, user rating and effective communication number of each agent, score the agents according to the complaint rate, user rating and effective communication number of each agent, and sort the agents in descending order according to the scores; obtain the sorted list. Step S32: Obtain the user quality level, and map the seats to different user quality levels based on the number of seats. Classify the seats using the seat ranking list to obtain the seat level. Specifically: Obtain the user quality level tier n, obtain the number of seats zx, and divide the seats into equal levels according to the tier, with each user quality level corresponding to a seat number of zx / n; sort the seats in descending order of their ratings, and map the seats to different user quality levels to obtain the seat level; provide service to users of the same user quality level according to their seat level.

8. The method for constructing personalized customer profiles and recommending services based on a large language model according to claim 7, characterized in that, The specific steps of step S31 are as follows: Step S311: Based on the agent information, obtain the agent's complaint rate, user rating, and number of effective communications. Calculate the agent rating zpf based on the agent's complaint rate, user rating, and number of effective communications. Step S312: Obtain the number of seats zx, obtain the seat score for each seat, denoted as zpf(z), sort the seats according to the seat scores, obtain the first seat score as the comparison value bj, perform a loop comparison of the seat scores, and move the seat scores greater than the comparison value forward to obtain the descending order of the seat scores.

9. The method for constructing personalized customer profiles and recommending services based on a large language model according to claim 8, characterized in that, The subsequent steps of step S311 are as follows: The system collects the number of user ratings, analyzes each user rating, and combines it with the number of effective interactions with agents. If a user ratings an agent after an effective interaction, the rating is valid. If a user ratings an agent multiple times consecutively or without any interaction, the rating is invalid. The system records the number of valid ratings and the rating scores. Based on the number of valid ratings and the rating scores, the average user rating pjz is obtained. The complaint rate bts and the number of valid interactions jcs of the agents are obtained. The average user rating pjz is combined with the agent rating to calculate the agent rating zpf. 。 10. The method for constructing personalized customer profiles and recommending services based on a large language model according to claim 1, characterized in that, The specific steps of step S4 are as follows: Step S41: Obtain the user's static data, extract keywords from the user's static data to obtain the user's static tags, count the user's static tags to obtain a static tag list, and generate an initial static profile from the static tag list. Step S42: Obtain user dialogue data, perform intent recognition on the dialogue data to obtain user intent keywords, use user intent keywords as dynamic behavioral feature labels to supplement the initial static profile, and obtain a multi-dimensional user profile. Step S43: Based on the multidimensional user profile, classify users, match the tags in the multidimensional user profile, obtain a multidimensional user profile as the standard profile, match the multidimensional user profile with the standard profile to obtain the matching value. If the matching value is greater than 0.5, treat it as a user group. If the matching value is less than 0.5, treat the multidimensional user profile as a new group and match other multidimensional user profiles.

Images