User demand recommendation method

Abstract

The invention relates to the field of big data, and discloses a user demand recommendation method which comprises the steps that entry buttons clicked by a user in a first unit time are acquired, the entry buttons comprise click heat, a browsing sequence, purchase data and complaint data, and character information of the entry buttons is converted into classification labels through a natural language model; configuring the central angle direction, the field angle direction and the radius of a sector according to the classification label, the browsing sequence and the click popularity of the entry button; and outputting the radius of a circle, the eccentric direction of the circle and the eccentric distance according to the radius of the sector, the central angle direction of the sector, the purchase data and the complaint data, and outputting a first recommendation result according to the overlapping area of the sector and the circle. According to the method, the calculation process is visualized, and multivariate data is unified. In addition, data classification is more accurate through the NLP model. And untrusted requests are rejected, so that the transaction security is improved.

Description

Technical Field

[0001] This invention relates to the field of big data, and in particular to a user demand recommendation method. Background Technology

[0002] The trend of online shopping is not only about buying goods, but also about buying services.

[0003] Similar to purchasing goods, users tend to buy similar products. On one hand, we analyze user behavior data such as category tags, browsing order, and click frequency, using user behavior as a reference for recommendations; on the other hand, we analyze click frequency, category tags, purchase data, and complaint data, using product attributes as a reference for recommendations. The types and units of these data differ. Therefore, traditional calculation methods, being purely algebraic, suffer from poor visualization and slow response times.

[0004] Furthermore, for online shopping platforms, every transaction may require substantive review to reduce transaction risks. For users, the platform's recommendation algorithm not only has the task of encouraging them to buy more, but also of verifying whether user behavior is abnormal or unsafe. However, there is currently no mechanism for using recommendation algorithms to verify transaction security. Introducing a separate security review mechanism would slow down data processing and waste more computing resources.

[0005] Furthermore, configuring the classification labels for the aforementioned data only involves a simple comparison query with a regular database, which may lead to inaccurate classification.

[0006] Therefore, there is a need for a method that can normalize multivariate data, visualize calculation results, make classification labels more accurate, and reduce data computation load and improve transaction security by using the recommendation results of recommendation algorithms as the object of security review. Summary of the Invention

[0007] The technical problem to be solved by the present invention is to provide a method that can normalize multivariate data, visualize calculation results, make classification labels more accurate, and reduce data computing load and improve transaction security by using the recommendation results of recommendation algorithms as the object of security review.

[0008] In a first aspect, the present invention provides a user demand recommendation method, comprising: Obtain the first transaction request; The algorithm acquires the keyword buttons clicked by users within a first unit of time. The keyword buttons include click popularity, browsing order, purchase data, and complaint data. The algorithm then uses a natural language model to convert the text information of the keyword buttons into classification tags. Configure the central angle direction, angular direction, and radius of the sector based on the category tags, browsing order, and click popularity of the entry button; output the radius, eccentric direction, and eccentric distance of the circle based on the radius, central angle direction, purchase data, and complaint data; and output the first recommendation result based on the overlapping area of ​​the sector and the circle. The first recommended result is output as the second recommended result based on the browsing order of each term button; Determine whether the term button of the first transaction request is within the first preset number of the second recommendation results. If yes, execute the first transaction request; otherwise, reject the first transaction request.

[0009] This invention discloses a user demand recommendation method, wherein the step of converting the text information of term buttons into classification tags using a natural language model includes: S201. Text information preprocessing: After splitting the text information of the term button using a word segmentation tool, the text information after synonym processing is output using a thesaurus. S202, Model Prompt Word Generation: Output prompt words based on the text information after synonym processing and the classification labels; S203, NLP Model Inference and Label Generation: After inputting prompt words using a pre-trained NLP model, output the classification labels of the word buttons.

[0010] This invention provides a user demand recommendation method, wherein: the central angle direction, angular direction, and radius of a sector are configured based on the category tags, browsing order, and click popularity of the keyword button; the radius, eccentric direction, and eccentric distance of a circle are output based on the radius, central angle direction, purchase data, and complaint data; and a first recommendation result is output based on the overlapping area of ​​the sector and the circle, including: Direction of the central angle of the sector: If the user clicks only one term button in the first unit of time, then the central angle direction is constructed based on the center angle of the term button; If a user clicks multiple term buttons within the first unit of time, the average of the center angles of the multiple term buttons will be used. Construct the central angle direction; The angle of the sector: output according to the following formula: Zhang Jiao = Among them, the above This represents the center angle of each term button in the browsing sequence; This represents the average center angle of multiple term buttons in the browsing order; 'a' represents the initial angle; and 'n' represents the number of term buttons in the browsing order.

[0011] The radius of the sector: The radius of the sector = ; The radius of the circle is half the radius of the sector. The eccentricity of the circle is in the same direction as the central angle of the sector; Eccentricity of a circle: Eccentricity = Purchase data × c - Complaint data × b; After initially aligning the center of the sector with the center of the circle, the circle is moved according to the eccentric direction and eccentric distance. The overlapping area of ​​the term buttons is configured according to the overlapping area of ​​the sector and the circle. The first recommendation result is output in descending order of the overlapping area.

[0012] This invention provides a user demand recommendation method, wherein a first recommendation result is output as a second recommendation result based on the browsing order of each term button, including: Search for category tags that appear consecutively more than a preset number of times in the browsing order, and increase the opening angle of the corresponding term button for the category tag; or, Search for category tags that appear consecutively more than a preset number of times in the browsing order, and increase the radius of the fan-shaped area of ​​the corresponding term button for the category tag; or, Search for category tags that appear consecutively more than a preset number of times in the browsing order, and delete the category tag that is furthest from the center angle of the tag that appears consecutively more than the preset number of times in the browsing order; or, Determine whether a split tag appears more than a preset number of consecutive times in the browsing order. If it does, increase the opening angle of the entry button corresponding to the category tag. If it does not, merge similar category tags in the step of converting the text information of the entry button into category tags using a natural language model. Then jump to the step of converting the text information of the entry button into category tags using a natural language model. or, Determine whether a split tag appears more than a preset number of consecutive times in the browsing order. If it does, increase the radius of the sector of the entry button corresponding to the category tag. If it does not, merge similar category tags by using a natural language model to convert the text information of the entry button into category tags. Then jump to the step of using a natural language model to convert the text information of the entry button into category tags. or, Determine whether a split tag appears more than a preset number of consecutive occurrences in the browsing order. If it does, delete the category tag that is furthest from the center angle of the tag that appears more than the preset number of consecutive occurrences in the browsing order. If it does not appear, merge similar category tags in the step of converting the text information of the term button into category tags using a natural language model; and jump to the step of converting the text information of the term button into category tags using a natural language model.

[0013] The present invention provides a user demand recommendation method, wherein, if not, the first transaction request is rejected, including: If not, the first transaction request is rejected, and the step of converting the text information of the term button into category labels using a natural language model is merged with similar category labels; and the process jumps to the step of converting the text information of the term button into category labels using a natural language model.

[0014] Secondly, a user demand recommendation system, comprising interconnected user terminals and platform terminals, includes the following steps: The system terminal obtains the first transaction request from the user terminal; The system terminal acquires the keyword buttons clicked by the user within a first unit of time. These keyword buttons include click popularity, browsing order, purchase data, and complaint data. A natural language model is used to convert the text information of the keyword buttons into category tags. Based on the category tags, browsing order, and click popularity of the keyword buttons, the central angle direction, the angular direction, and the radius of the sector are configured. Based on the radius, central angle direction, purchase data, and complaint data of the sector, the radius, eccentric direction, and eccentric distance of the circle are output. A first recommendation result is output based on the overlapping area of ​​the sector and the circle. A second recommendation result is output based on the browsing order of each keyword button. It is determined whether the keyword button for the first transaction request is within the first preset number of positions in the second recommendation result. If so, the first transaction request from the user terminal is executed; otherwise, the first transaction request from the user terminal is rejected.

[0015] This invention discloses a user demand recommendation system, wherein the step of converting the text information of term buttons into classification tags using a natural language model includes: S201. Text information preprocessing: After splitting the text information of the term button using a word segmentation tool, the text information after synonym processing is output using a thesaurus. S202, Model Prompt Word Generation: Output prompt words based on the text information after synonym processing and the classification labels; S203, NLP Model Inference and Label Generation: After inputting prompt words using a pre-trained NLP model, output the classification labels of the word buttons.

[0016] This invention discloses a user demand recommendation system, wherein the system configures the central angle direction, angular direction, and radius of a sector based on the category tags, browsing order, and click popularity of the term buttons; it outputs the radius, eccentric direction, and eccentric distance of a circle based on the radius, central angle direction, purchase data, and complaint data; and it outputs a first recommendation result based on the overlapping area of ​​the sector and the circle, including: Direction of the central angle of the sector: If the user clicks only one term button in the first unit of time, then the central angle direction is constructed based on the center angle of the term button; If a user clicks multiple term buttons within the first unit of time, the average of the center angles of the multiple term buttons will be used. Construct the central angle direction; The angle of the sector: output according to the following formula: Zhang Jiao = Among them, the above This represents the center angle of each term button in the browsing sequence; This represents the average center angle of multiple term buttons in the browsing order; 'a' represents the initial angle; and 'n' represents the number of term buttons in the browsing order.

[0017] The radius of the sector: The radius of the sector = ; The radius of the circle is half the radius of the sector. The eccentricity of the circle is in the same direction as the central angle of the sector; Eccentricity of a circle: Eccentricity = Purchase data × c - Complaint data × b; After initially aligning the center of the sector with the center of the circle, the circle is moved according to the eccentric direction and eccentric distance. The overlapping area of ​​the term buttons is configured according to the overlapping area of ​​the sector and the circle. The first recommendation result is output in descending order of the overlapping area.

[0018] Thirdly, a computer-readable storage medium The computer-readable storage medium stores instructions that, when the computer-readable storage medium is run on a computer, cause the computer to execute a user demand recommendation method.

[0019] Fourthly, a computer program product containing instructions. When the computer program product is run on a computer, it causes the computer to execute a user demand recommendation method.

[0020] The user demand recommendation method of this invention differs from existing technologies in that, upon receiving a first transaction request, this invention, in its response, uses a natural language processing (NLP) model to transform the complex user-clickable keyword buttons into standardized category labels. Based on the category labels corresponding to the keyword buttons, click popularity, browsing order, purchase data, and complaint data, it outputs a sector and a circle, and uses a combination of numerical and graphical methods to output the first recommendation result through overlapping area, making the calculation process visual and unifying diverse data. Furthermore, the NLP model makes data classification more accurate. Additionally, it clusters browsed items according to browsing order to output a risk penalty function for each keyword button, thereby adjusting the first recommendation result to a second recommendation result. Finally, it uses the first preset number of characters of the second transaction request as a safe and trustworthy transaction request, rejecting untrusted requests, thereby increasing transaction security. That is, the present invention evaluates user behavior through a sector shape, evaluates the quality of the term buttons through a circle, outputs the first recommendation result based on the overlapping area, and adjusts the second recommendation result according to the browsing order. This adds an auxiliary verification function to the security of the first transaction request and also provides corresponding sorting guidance for the order of the subsequent product recommendation list.

[0021] The following description, in conjunction with the accompanying drawings, further illustrates a user demand recommendation method of the present invention. Attached Figure Description

[0022] Figure 1 This is a flowchart of a user demand recommendation method; Figure 2 This is a diagram illustrating the combination of numbers and shapes. Detailed Implementation

[0023] like Figure 1 , 2 As shown, the user demand recommendation method of the present invention includes... Obtain the first transaction request; The algorithm acquires the keyword buttons clicked by users within a first unit of time. The keyword buttons include click popularity, browsing order, purchase data, and complaint data. The algorithm then uses a natural language model to convert the text information of the keyword buttons into classification tags. See Figure 2 Configure the central angle direction, opening angle direction, and radius of the sector based on the category tags, browsing order, and click popularity of the entry button; output the radius, eccentric direction, and eccentric distance of the circle based on the radius, central angle direction, purchase data, and complaint data; and output the first recommendation result based on the overlapping area of ​​the sector and the circle. The first recommended result is output as the second recommended result based on the browsing order of each term button; Determine whether the term button of the first transaction request is within the first preset number of the second recommendation results. If yes, execute the first transaction request; otherwise, reject the first transaction request.

[0024] Upon receiving a first transaction request, this invention, in its response, uses the aforementioned natural language model to transform the complex user-clickable keyword buttons into standardized category tags. Based on the category tags corresponding to the keyword buttons, click popularity, browsing order, purchase data, and complaint data, it outputs a sector and a circle, and uses a combination of numerical and graphical methods to output the first recommendation result through overlapping area, making the calculation process visual and unifying diverse data. Furthermore, the NLP model enhances data classification accuracy. Additionally, it clusters browsed items according to browsing order to output a risk penalty function for each keyword button, thereby adjusting the first recommendation result to a second recommendation result. Finally, it uses the first preset number of characters of the second transaction request as a secure and trustworthy transaction request, rejecting untrusted requests, thus increasing transaction security.

[0025] That is, the present invention evaluates user behavior through a sector shape, evaluates the quality of the term buttons through a circle, outputs the first recommendation result based on the overlapping area, and adjusts the second recommendation result according to the browsing order. This adds an auxiliary verification function to the security of the first transaction request and also provides corresponding sorting guidance for the order of the subsequent product recommendation list.

[0026] Among them, the aforementioned independent claims tend to output a recommendation list based on the second recommendation result.

[0027] The acquisition of the first transaction request can be understood as follows: within the current first unit of time, the first transaction request obtained needs to be considered in conjunction with the keyword button data from the previous first unit of time, including click popularity, browsing order, purchase data, complaint data, and text information. The first transaction request could be: purchasing a housekeeper for in-home cleaning once. Alternatively, the time when the first transaction request is obtained can be taken as the end of the first unit of time, and the aforementioned click popularity, browsing order, purchase data, complaint data, and text information from the first unit of time preceding the acquisition of the first transaction request can be analyzed.

[0028] The process involves acquiring user-clicked keyword buttons within a first unit of time. These keyword buttons include click popularity, browsing order, purchase data, and complaint data. A natural language model is used to convert the text information of the keyword buttons into classification tags. Specifically, the first unit of time is the unit preceding the first unit of time for acquiring the first transaction request. For example, if the first unit of time is 10 minutes, and each unit of time is fixed by dividing the hour into six units, for instance, if the first transaction request is acquired at 8:05, we should analyze the click popularity, browsing order, purchase data, complaint data, and text information from 7:50 to 8:00. Alternatively, we can also use the click popularity, browsing order, purchase data, complaint data, and text information acquired from 7:55 to 8:05 as the analysis object. Both of these methods can be considered as aspects of this invention.

[0029] The term button can be, for example, the product / service button on the 58 Daojia app. The term button not only includes product photos but also text information to assist in searching for keywords.

[0030] The click popularity of the keyword button can be defined as the number of times the button is clicked by every 100 people within the first unit of time. Click popularity when the number of app users per hour is less than 1000 can be ignored. That is, if the number of app users per hour in the first unit of time is less than 1000, the click popularity should be based on the first unit of time when the number of users exceeds 1000. This is to avoid the calculation results being affected by an excessively small data set.

[0031] The browsing order of the term buttons can be as follows: In the first unit of time, the order in which users click is: 1. Nanny comes to tidy up the house; 2. Nanny comes to clean the newly renovated house; 3. Nanny comes to tidy up the house; 4. Nanny comes to tidy up the house (corresponding to the relevant button).

[0032] In other words, it reflects the user's behavioral trajectory or browsing footprint. The browsing order not only reflects the dispersion of the user's browsing, but also influences the recommendation results through the standard deviation and clustering of consecutively browsed items reflected by the browsing order.

[0033] The purchase data for the keyword button can be: purchase count, that is, the total number of purchases by all users of the app within the first unit of time. The more purchases, the better the product is and the more worth recommending it.

[0034] The complaint data for the term button can be: the number of complaints, that is, the total number of complaints from all users of the app within the first unit of time. The more complaints, the worse the product is and the less worth recommending it.

[0035] Among them, the text information of the entry button is converted into classification labels by using a natural language model, including: S201. Preprocessing of text information: The text information of the entry button: A beautiful nanny cleans the house.

[0036] Data cleaning of the thesaurus: Use a word segmentation tool (Jieba) for data cleaning. Utilize the thesaurus to output "beautiful nanny" as its synonym "nanny", thereby performing preliminary cleaning on the data and removing data noise. Output "nanny cleans the house" as "nanny comes to tidy up the house". This facilitates classification as the same label for processing.

[0037] The text information after synonym processing: Nanny comes to tidy up the house.

[0038] S202. Generation of model prompt words: Output prompt words according to the text information after synonym processing in combination with classification labels; Based on the output of the thesaurus, design a dynamic prompt template, for example: "Analyze the following service descriptions 'thesaurus', generate classification labels, such as 'food delivery', 'nanny comes to cook', 'nanny comes to tidy up the house', 'nanny comes to clean a newly renovated house','repair and replace the toilet','repair wires'. This guides the model to generate relatively unified classification labels. This ensures that the central angles corresponding to the pre-stored classification labels in the database can be directly output.

[0039] Thesaurus output: Nanny comes to tidy up the house.

[0040] Prompt words: Analyze the following service description 'nanny comes to tidy up the house', output classification label: 'nanny comes to tidy up the house', generate classification labels, such as 'food delivery', 'nanny comes to cook', 'nanny comes to tidy up the house', 'nanny comes to clean a newly renovated house','repair and replace the toilet','repair wires'.

[0041] S203. NLP model inference and label generation: After using the pre-trained NLP model to input the prompt words, output the classification label of the entry button.

[0042] Use the pre-trained NLP model for zero-shot or few-shot inference, input the prompt words, and output a list of candidate labels and their confidence levels.

[0043] The model calculates the label probability. If the confidence level is low, iterate the prompt.

[0044] Output classification label: Nanny comes to tidy up the house.

[0045] Output the central angle according to the classification label and the database: 90 degrees.

[0046] Experiments show (1000 terms tested) that the label accuracy is 92% (22% higher than the baseline) and the processing time is 0.4 seconds per term. Combined with the subsequent circle eccentricity distance, it improves the efficiency of calculating the overlapping area of ​​recommendations by 15%, making it suitable for scenarios such as park service apps.

[0047] It's important to emphasize that the text information in terminology buttons is often designed to complement the platform's keyword search function. This text is typically lengthy and complex, causing significant difficulties in configuring category tags, resulting in low accuracy and efficiency. However, leveraging NLP models with dynamically changing model hints can further improve the accuracy of category tags. Furthermore, synonym data is first cleaned and normalized, facilitating the continuous generation and training of accurate NLP models.

[0048] Among them, see Figure 2 Based on the category tags, browsing order, and click popularity of the entry buttons, configure the central angle direction, angular direction, and radius of the sector; based on the radius, central angle direction, purchase data, and complaint data, output the radius, eccentric direction, and eccentric distance of the circle; and output the first recommendation result based on the overlapping area of ​​the sector and the circle. This can be understood as: The combination of numbers and shapes, with a sector representing a vector and a circle representing the vector and its changes.

[0049] That is, in the sector: In a two-dimensional coordinate system, the central angle represents the park's category label. For example, 0-60 degrees represents repair, 61-120 degrees represents cleaning, and 121-180 degrees represents catering. Each specific category corresponds to a unique central angle. The central angle of each category label is pre-stored in the database. The more complex the label, the more it is biased towards the corresponding position. For example, food delivery is biased towards 180 degrees, cooking by a nanny is biased towards 120 degrees, house cleaning by a nanny is biased towards 90 degrees, cleaning of a newly renovated house by a nanny is biased towards 61 degrees, repairs requiring cleaning, such as toilet replacement, are biased towards 60 degrees, and repairs not requiring cleaning, such as electrical wiring repair, are biased towards 0 degrees. The above is just a bias orientation; the specific service item can be determined based on its pre-stored central angle.

[0050] Therefore, the direction of the central angle of the sector is: If the user clicks only one term button in the first unit of time, then the central angle direction is constructed based on the center angle of the term button; If a user clicks multiple term buttons within the first unit of time, the average of the center angles of the multiple term buttons will be used. Construct the central angle direction; The direction of the central angle of the sector represents the direction of its centerline. The center of the sector is located at the origin of the two-dimensional coordinate system.

[0051] Therefore, the fan-shaped angle, i.e., the dispersion of the browsing order of multiple term buttons, i.e., the standard deviation, is output according to the following formula: Zhang Jiao = Among them, the above This represents the center angle of each term button in the browsing sequence; This represents the average center angle of multiple term buttons in the browsing sequence; 'a' represents the initial angle; and 'n' represents the number of term buttons in the browsing sequence. 'a' can be, for example, 30 degrees to 200 degrees. That is, 'a' should be greater than the largest standard deviation historically observed. In other words, the angle cannot be negative, and 'a' can be chosen to be as small as possible based on the actual situation, thus making the change in the angle more sensitive. Furthermore, since the maximum angle of the sector is a circle, it does not prevent the consideration of the overlapping area with the circle.

[0052] Among the aforementioned angles, larger angles create wider fan-shaped sectors, which better reflect the user's focused attention and clear intent within the first unit of time, resembling a searchlight-shaped sector. Smaller angles create narrower fan-shaped sectors, representing a highly exploratory user with ambiguous intent, resembling a laser pointer-shaped sector.

[0053] The radius of the sector, in a two-dimensional coordinate system, represents the intensity of demand, i.e., the click intensity; click intensity is the average number of clicks per 100 people within the first unit of time preceding the first unit of time. The unit is centimeters.

[0054] For example, the first unit of time is 1 hour, and each whole hour is counted as one unit of time. For instance, 7:00 to 8:00 is one unit of time, and 8:00 to 9:00 is another unit of time. The app mentioned above has over 1000 users per hour. If the current user is at 8:30, then the number of times all users clicked this button between 7:00 and 8:00 should be converted to the number of clicks per 100 users, outputting the click popularity; and the radius of the sector should be output using the following formula: The radius of the sector = That is, the radius of the sector increases with the increase of click popularity, and the image shows an upward trend that gradually decreases.

[0055] That is, within a circle: Therefore, the radius of the circle is half the radius of the sector; Therefore, the direction of the eccentricity of the circle is the same as the direction of the central angle of the sector; Therefore, the eccentricity of the circle is: based on the complaint data and purchase data output, where the complaint data is configured with a larger weight coefficient b, and the purchase data is configured with a smaller weight c according to the number of purchases.

[0056] That is, eccentricity distance = purchase data × c - complaint data × b; Where c can be, for example, 0.01~1, and b can be, for example, 2~10. The maximum eccentricity should not exceed 1 / 3 of the sector radius.

[0057] It should be noted that the center of the circle is initially located at the origin of the two-dimensional coordinate system, that is, it coincides with the center of the sector. However, in order to adjust the overlapping area by the specific position of the eccentricity, we introduced specific data to change the position of the circle, thereby changing the overlapping area. That is, the larger the purchase data, the larger the overlapping area; the larger the complaint data, the smaller the overlapping area.

[0058] The first recommendation result is output based on the overlapping area of ​​the sector and the circle. This can be understood as outputting the first recommendation result in descending order of overlapping area. In other words, the first recommendation result is a list of each keyword button, with the keyword button with the largest overlap appearing first in the first recommendation result and being the most recommended.

[0059] That is, after initially aligning the center of the sector with the center of the circle, the circle is moved according to the eccentric direction and eccentric distance. The overlapping area of ​​the term button is configured according to the overlapping area of ​​the sector and the circle. The first recommendation result is then output according to the overlapping area from largest to smallest.

[0060] The process of outputting the second recommendation result based on the browsing order of each term button can be understood as follows: Since the browsing order represents the sequence in which a user views the keywords of different category tags within the first unit of time, the consecutive browsing of the same category tags naturally indicates a stronger purchase intention from the user at that moment. For example, the browsing order might be: 1. A nanny comes to clean a newly renovated house; 2. A nanny comes to tidy up a house; 3. A nanny comes to tidy up a house; 4. A nanny comes to tidy up a house; 5. Food delivery, corresponding to the relevant keywords / buttons.

[0061] Therefore, the button representing the category tag that users most want for a housekeeper to come and clean their house will naturally be displayed. So, if more than a preset number of consecutive category tags appear in the browsing sequence within the first unit of time, their order can be adjusted by increasing their fan-shaped angle.

[0062] That is, it includes the following steps: Search for category tags that appear consecutively more than a preset number of times in the browsing order, and increase the angle of the corresponding term button for the category tag.

[0063] Specifically, the preset consecutive number of times can be 3, and the angle can be increased by 10%. For example, if the category tag for "house cleaning service" appears more than 3 times consecutively in the browsing sequence within the first unit of time, then the angle of the corresponding entry button for "house cleaning service" will be increased from the original 30 degrees to 33 degrees.

[0064] or, Search for category tags that appear consecutively more than a preset number of times in the browsing order, and increase the radius of the fan-shaped area of ​​the corresponding term button for the category tag.

[0065] Specifically, the preset consecutive number of times can be 3. Increasing the angle can be achieved by increasing the radius by 10%, but the radius of the circle remains the same, that is, half the radius of the sector before the 10% increase. For example, if the category tag for "house cleaning services" appears more than 3 times consecutively in the browsing sequence within the first unit of time, then the radius of the sector of the button for the corresponding entry "house cleaning services" will be increased from the original 30 centimeters to 33 centimeters.

[0066] or, Search for category tags that appear consecutively more than a preset number of times in the browsing order, and delete the category tag that is furthest from the center angle of the tag that appears consecutively more than the preset number of times in the browsing order.

[0067] Specifically, the preset consecutive occurrence count can be 3 times. If a housekeeper's home cleaning service (center angle biased towards 90 degrees) and food delivery (center angle biased towards 180 degrees) appear more than 3 times consecutively, then the category tag "electrical wiring repair" (center angle biased towards 0 degrees) appears 2 times. Therefore, to reduce the standard deviation and better align with the trend of focus, we delete the 2 browsing records of "electrical wiring repair," which has the farthest center angle distance. This increases the angle of the "food delivery" and "housekeeper's home cleaning" tag buttons, thereby increasing their overlap area and increasing the likelihood of them ranking higher.

[0068] or, Determine whether a split tag appears more than a preset number of consecutive times in the browsing order. If it does, increase the opening angle of the entry button corresponding to the category tag. If it does not, merge similar category tags by using a natural language model to convert the text information of the entry button into a category tag. Then jump to the step of using a natural language model to convert the text information of the entry button into a category tag.

[0069] Specifically, in S202, because the dynamic prompt template pre-sets the category tags, it cannot display category tags more than a preset consecutive number of times. Therefore, it cannot further highlight the category tags that the user wants to purchase based on the browsing order. This may be because the granularity of the category tags is too small. We can appropriately increase their granularity. Specifically, in the browsing order, "housekeeper service" and "housekeeper service for newly renovated houses" are initially configured as two category tags. However, the database presets that these two tags are similar and can be merged into "housekeeper service". Therefore, the method of this invention is re-executed until a split tag with a consecutive occurrence exceeding the preset consecutive number is found, and its position in the second recommendation result is adjusted, thereby more accurately recommending term buttons to the user and further ensuring transaction security.

[0070] or, If a split tag appears more than a preset number of consecutive occurrences in the browsing order, the radius of the sector of the entry button corresponding to the category tag is increased. If no split tag appears, the step of converting the text information of the entry button into category tags using a natural language model is performed to merge similar category tags. Then, the process jumps to the step of converting the text information of the entry button into category tags using a natural language model.

[0071] Specifically, in S202, because the dynamic prompt template pre-sets the category tags, it cannot display category tags more than a preset consecutive number of times. Therefore, it cannot further highlight the category tags that the user wants to purchase based on the browsing order. This may be because the granularity of the category tags is too small. We can appropriately increase their granularity. Specifically, in the browsing order, "housekeeper service" and "housekeeper service for newly renovated houses" are initially configured as two category tags. However, the database presets that these two tags are similar and can be merged into "housekeeper service". Therefore, the method of this invention is re-executed until a split tag with a consecutive occurrence exceeding the preset consecutive number is found, and its position in the second recommendation result is adjusted, thereby more accurately recommending term buttons to the user and further ensuring transaction security.

[0072] or, Determine whether a split tag appears more than a preset number of consecutive occurrences in the browsing order. If it does, delete the category tag that is furthest from the center angle of the tag that appears more than the preset number of consecutive occurrences in the browsing order. If it does not appear, merge similar category tags in the step of converting the text information of the term button into category tags using a natural language model; and jump to the step of converting the text information of the term button into category tags using a natural language model.

[0073] Specifically, in S202, because the dynamic prompt template pre-sets the category tags, it cannot display category tags more than a preset consecutive number of times. Therefore, it cannot further highlight the category tags that the user wants to purchase based on the browsing order. This may be because the granularity of the category tags is too small. We can appropriately increase their granularity. Specifically, in the browsing order, "housekeeper service" and "housekeeper service for newly renovated houses" are initially configured as two category tags. However, the database presets that these two tags are similar and can be merged into "housekeeper service". Therefore, the method of this invention is re-executed until a split tag with a consecutive occurrence exceeding the preset consecutive number is found, and its position in the second recommendation result is adjusted, thereby more accurately recommending term buttons to the user and further ensuring transaction security.

[0074] Specifically, determining whether the term button in the first transaction request is within the first preset number of the second recommendation results; if so, executing the first transaction request; otherwise, rejecting the first transaction request, includes: The initial preset number of buttons can be 80% of the total number of views in the browsing order, or 80% of the total number of keyword buttons. For example, if the total number of views in the browsing order is 100, then the initial preset number of buttons is 80. In other words, if the first transaction request requests the transaction of the first 80 keyword buttons in the second recommendation result, the first transaction request can be executed; otherwise, it is rejected. In this way, we utilize the second recommendation result output by the recommendation algorithm not only as the second recommendation result for the display order of keyword buttons in subsequent advertisements to sort the keyword buttons, but also use the recommendation result to filter the rationality of purchases based on user behavior and product behavior, thereby avoiding high-risk transactions such as platform fraud, hacker attacks, and transaction scams, and ensuring transaction security.

[0075] If not, the first transaction request will be rejected, including: If not, the first transaction request is rejected, and the step of converting the text information of the term button into category labels using a natural language model is merged with similar category labels; and the process jumps to the step of converting the text information of the term button into category labels using a natural language model.

[0076] Specifically, in S202, because the dynamic prompt template pre-sets the category tags, it cannot display category tags more than a preset consecutive number of times. Therefore, it cannot further highlight the category tags that the user wants to purchase based on the browsing order. This may be because the granularity of the category tags is too small. We can appropriately increase their granularity. Specifically, in the browsing order, "housekeeper service" and "housekeeper service for newly renovated houses" are initially configured as two category tags. However, the database presets that these two tags are similar and can be merged into "housekeeper service". Therefore, the method of this invention is re-executed until a split tag with a consecutive occurrence exceeding the preset consecutive number is found, and its position in the second recommendation result is adjusted, thereby more accurately recommending term buttons to the user and further ensuring transaction security.

[0077] In some embodiments, see Figure 1 The process of converting the text information of the term buttons into classification labels using a natural language model includes: S201. Text information preprocessing: After splitting the text information of the term button using a word segmentation tool, the text information after synonym processing is output using a thesaurus. S202, Model Prompt Word Generation: Output prompt words based on the text information after synonym processing and the classification labels; S203, NLP Model Inference and Label Generation: After inputting prompt words using a pre-trained NLP model, output the classification labels of the word buttons.

[0078] In some embodiments, see Figure 2 Based on the category tags, browsing order, and click popularity of the entry buttons, configure the central angle direction, angular direction, and radius of the sector; based on the radius, central angle direction, purchase data, and complaint data, output the radius, eccentric direction, and eccentric distance of the circle; and output the first recommendation result based on the overlapping area of ​​the sector and the circle, including: Direction of the central angle of the sector: If the user clicks only one term button in the first unit of time, then the central angle direction is constructed based on the center angle of the term button; If a user clicks multiple term buttons within the first unit of time, the average of the center angles of the multiple term buttons will be used. Construct the central angle direction; The angle of the sector: output according to the following formula: Zhang Jiao = Among them, the above This represents the center angle of each term button in the browsing sequence; This represents the average center angle of multiple term buttons in the browsing order; 'a' represents the initial angle; and 'n' represents the number of term buttons in the browsing order.

[0079] The radius of the sector: The radius of the sector = ; The radius of the circle is half the radius of the sector. The eccentricity of the circle is in the same direction as the central angle of the sector; Eccentricity of a circle: Eccentricity = Purchase data × c - Complaint data × b; After initially aligning the center of the sector with the center of the circle, the circle is moved according to the eccentric direction and eccentric distance. The overlapping area of ​​the term buttons is configured according to the overlapping area of ​​the sector and the circle. The first recommendation result is output in descending order of the overlapping area.

[0080] In some embodiments, see Figure 1 Based on the browsing order of each term button, the first recommended result is output as the second recommended result, including: Search for category tags that appear consecutively more than a preset number of times in the browsing order, and increase the opening angle of the corresponding term button for the category tag; or, Search for category tags that appear consecutively more than a preset number of times in the browsing order, and increase the radius of the fan-shaped area of ​​the corresponding term button for the category tag; or, Search for category tags that appear consecutively more than a preset number of times in the browsing order, and delete the category tag that is furthest from the center angle of the tag that appears consecutively more than the preset number of times in the browsing order; or, Determine whether a split tag appears more than a preset number of consecutive times in the browsing order. If it does, increase the opening angle of the entry button corresponding to the category tag. If it does not, merge similar category tags in the step of converting the text information of the entry button into category tags using a natural language model. Then jump to the step of converting the text information of the entry button into category tags using a natural language model. or, Determine whether a split tag appears more than a preset number of consecutive times in the browsing order. If it does, increase the radius of the sector of the entry button corresponding to the category tag. If it does not, merge similar category tags by using a natural language model to convert the text information of the entry button into category tags. Then jump to the step of using a natural language model to convert the text information of the entry button into category tags. or, Determine whether a split tag appears more than a preset number of consecutive occurrences in the browsing order. If it does, delete the category tag that is furthest from the center angle of the tag that appears more than the preset number of consecutive occurrences in the browsing order. If it does not appear, merge similar category tags in the step of converting the text information of the term button into category tags using a natural language model; and jump to the step of converting the text information of the term button into category tags using a natural language model.

[0081] In some embodiments, see Figure 1 If not, the first transaction request shall be rejected, including: If not, the first transaction request is rejected, and the step of converting the text information of the term button into category labels using a natural language model is merged with similar category labels; and the process jumps to the step of converting the text information of the term button into category labels using a natural language model.

[0082] like Figure 1 , 2 As shown, the present invention provides a user demand recommendation system, comprising interconnected user terminals and platform terminals, and includes the following steps: The system terminal obtains the first transaction request from the user terminal; The system terminal acquires the keyword buttons clicked by the user within a first unit of time. These keyword buttons include click popularity, browsing order, purchase data, and complaint data. A natural language model is used to convert the text information of the keyword buttons into category tags. Based on the category tags, browsing order, and click popularity of the keyword buttons, the central angle direction, the angular direction, and the radius of the sector are configured. Based on the radius, central angle direction, purchase data, and complaint data of the sector, the radius, eccentric direction, and eccentric distance of the circle are output. A first recommendation result is output based on the overlapping area of ​​the sector and the circle. A second recommendation result is output based on the browsing order of each keyword button. It is determined whether the keyword button for the first transaction request is within the first preset number of positions in the second recommendation result. If so, the first transaction request from the user terminal is executed; otherwise, the first transaction request from the user terminal is rejected.

[0083] Upon receiving a first transaction request, this invention, in its response, uses the aforementioned natural language model to transform the complex user-clickable keyword buttons into standardized category tags. Based on the category tags corresponding to the keyword buttons, click popularity, browsing order, purchase data, and complaint data, it outputs a sector and a circle, and uses a combination of numerical and graphical methods to output the first recommendation result through overlapping area, making the calculation process visual and unifying diverse data. Furthermore, the NLP model enhances data classification accuracy. Additionally, it clusters browsed items according to browsing order to output a risk penalty function for each keyword button, thereby adjusting the first recommendation result to a second recommendation result. Finally, it uses the first preset number of characters of the second transaction request as a secure and trustworthy transaction request, rejecting untrusted requests, thus increasing transaction security.

[0084] Among them, the aforementioned independent claims tend to identify transaction security based on the second recommendation result.

[0085] In some embodiments, see Figure 1 The process of converting the text information of the term buttons into classification labels using a natural language model includes: S201. Text information preprocessing: After splitting the text information of the term button using a word segmentation tool, the text information after synonym processing is output using a thesaurus. S202, Model Prompt Word Generation: Output prompt words based on the text information after synonym processing and the classification labels; S203, NLP Model Inference and Label Generation: After inputting prompt words using a pre-trained NLP model, output the classification labels of the word buttons.

[0086] In some embodiments, see Figure 2 Based on the category tags, browsing order, and click popularity of the entry buttons, configure the central angle direction, angular direction, and radius of the sector; based on the radius, central angle direction, purchase data, and complaint data, output the radius, eccentric direction, and eccentric distance of the circle; and output the first recommendation result based on the overlapping area of ​​the sector and the circle, including: Direction of the central angle of the sector: If the user clicks only one term button in the first unit of time, then the central angle direction is constructed based on the center angle of the term button; If a user clicks multiple term buttons within the first unit of time, the average of the center angles of the multiple term buttons will be used. Construct the central angle direction; The angle of the sector: output according to the following formula: Zhang Jiao = Among them, the above This represents the center angle of each term button in the browsing sequence; This represents the average center angle of multiple term buttons in the browsing order; 'a' represents the initial angle; and 'n' represents the number of term buttons in the browsing order.

[0087] The radius of the sector: The radius of the sector = ; The radius of the circle is half the radius of the sector. The eccentricity of the circle is in the same direction as the central angle of the sector; Eccentricity of a circle: Eccentricity = Purchase data × c - Complaint data × b; After initially aligning the center of the sector with the center of the circle, the circle is moved according to the eccentric direction and eccentric distance. The overlapping area of ​​the term buttons is configured according to the overlapping area of ​​the sector and the circle. The first recommendation result is output in descending order of the overlapping area.

[0088] This invention provides a computer-readable storage medium. The computer-readable storage medium stores instructions that, when the computer-readable storage medium is run on a computer, cause the computer to execute a user demand recommendation method.

[0089] This invention provides a computer program product containing instructions. When the computer program product is run on a computer, it causes the computer to execute the user demand recommendation method.

[0090] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made by those skilled in the art to the technical solutions of the present invention without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A user demand recommendation method, characterized in that: include Obtain the first transaction request; The algorithm acquires the keyword buttons clicked by users within a first unit of time. The keyword buttons include click popularity, browsing order, purchase data, and complaint data. The algorithm then uses a natural language model to convert the text information of the keyword buttons into classification tags. Configure the central angle direction, angular direction, and radius of the sector based on the category tags, browsing order, and click popularity of the entry button; output the radius, eccentric direction, and eccentric distance of the circle based on the radius, central angle direction, purchase data, and complaint data; and output the first recommendation result based on the overlapping area of ​​the sector and the circle. The first recommended result is output as the second recommended result based on the browsing order of each term button; Determine whether the term button of the first transaction request is within the first preset number of the second recommendation results. If yes, execute the first transaction request; otherwise, reject the first transaction request.

2. The user demand recommendation method according to claim 1, characterized in that: The process of converting the text information of the term buttons into classification labels using a natural language model includes: S201. Text information preprocessing: After splitting the text information of the term button using a word segmentation tool, the text information after synonym processing is output using a thesaurus. S202, Model Prompt Word Generation: Output prompt words based on the text information after synonym processing and the classification labels; S203, NLP Model Inference and Label Generation: After inputting prompt words using a pre-trained NLP model, output the classification labels of the word buttons.

3. The user demand recommendation method according to claim 2, characterized in that: Configure the central angle direction, angular direction, and radius of the sector based on the category tags, browsing order, and click popularity of the entry buttons; output the radius, eccentric direction, and eccentric distance of the circle based on the radius, central angle direction, purchase data, and complaint data; and output the first recommendation result based on the overlapping area of ​​the sector and the circle, including: Direction of the central angle of the sector: If the user clicks only one term button in the first unit of time, then the central angle direction is constructed based on the center angle of the term button; If a user clicks multiple term buttons within the first unit of time, the average of the center angles of the multiple term buttons will be used. Construct the central angle direction; The angle of the sector: output according to the following formula: Zhang Jiao = Among them, the above This represents the center angle of each term button in the browsing sequence; This represents the average center angle of multiple term buttons in the browsing order; 'a' represents the initial angle; and 'n' represents the number of term buttons in the browsing order. The radius of the sector: The radius of the sector = ; The radius of the circle is half the radius of the sector; The eccentricity of the circle is in the same direction as the central angle of the sector; Eccentricity of a circle: Eccentricity = Purchase data × c - Complaint data × b; After initially aligning the center of the sector with the center of the circle, the circle is moved according to the eccentric direction and eccentric distance. The overlapping area of ​​the term buttons is configured according to the overlapping area of ​​the sector and the circle. The first recommendation result is output in descending order of the overlapping area.

4. The user demand recommendation method according to claim 3, characterized in that: The first recommended result is output as the second recommended result based on the browsing order of each term button, including: Search for category tags that appear consecutively more than a preset number of times in the browsing order, and increase the opening angle of the corresponding term button for the category tag; or, Search for category tags that appear consecutively more than a preset number of times in the browsing order, and increase the radius of the fan-shaped area of ​​the corresponding term button for the category tag; or, Search for category tags that appear consecutively more than a preset number of times in the browsing order, and delete the category tag that is furthest from the center angle of the tag that appears consecutively more than the preset number of times in the browsing order; or, Determine whether a split tag appears more than a preset number of consecutive times in the browsing order. If it does, increase the opening angle of the entry button corresponding to the category tag. If it does not, merge similar category tags in the step of converting the text information of the entry button into category tags using a natural language model. Then jump to the step of converting the text information of the entry button into category tags using a natural language model. or, Determine whether a split tag appears more than a preset number of consecutive times in the browsing order. If it does, increase the radius of the sector of the entry button corresponding to the category tag. If it does not, merge similar category tags by using a natural language model to convert the text information of the entry button into category tags. Then jump to the step of using a natural language model to convert the text information of the entry button into category tags. or, Determine whether a split tag appears more than a preset number of consecutive occurrences in the browsing order. If it does, delete the category tag that is furthest from the center angle of the tag that appears more than the preset number of consecutive occurrences in the browsing order. If it does not appear, merge similar category tags in the step of converting the text information of the term button into category tags using a natural language model; and jump to the step of converting the text information of the term button into category tags using a natural language model.

5. The user demand recommendation method according to claim 4, characterized in that: If not, the first transaction request shall be rejected, including: If not, the first transaction request is rejected, and the step of converting the text information of the term button into category labels using a natural language model is merged with similar category labels; and the process jumps to the step of converting the text information of the term button into category labels using a natural language model.

6. A user demand recommendation system, characterized in that: Including interconnected user terminals and platform terminals, the process includes the following steps: The system terminal obtains the first transaction request from the user terminal; The system terminal acquires the keyword buttons clicked by users within a first unit of time. These keyword buttons include click popularity, browsing order, purchase data, and complaint data. A natural language model is used to convert the text information of the keyword buttons into category tags. Based on the category tags, browsing order, and click popularity of the keyword buttons, the central angle direction, the angular direction, and the radius of the sector are configured. Based on the radius, central angle direction, purchase data, and complaint data, the radius, eccentricity direction, and eccentricity distance of the circle are output. A first recommendation result is output based on the overlapping area of ​​the sector and the circle. The first recommendation result is output as a second recommendation result based on the browsing order of each term button; it is determined whether the term button of the first transaction request is within the first preset number of the second recommendation result. If so, the first transaction request of the user terminal is executed; otherwise, the first transaction request of the user terminal is rejected.

7. A user demand recommendation system according to claim 6, characterized in that: The process of converting the text information of the term buttons into classification labels using a natural language model includes: S201. Text information preprocessing: After splitting the text information of the term button using a word segmentation tool, the text information after synonym processing is output using a thesaurus. S202, Model Prompt Word Generation: Output prompt words based on the text information after synonym processing and the classification labels; S203, NLP Model Inference and Label Generation: After inputting prompt words using a pre-trained NLP model, output the classification labels of the word buttons.

8. A user demand recommendation system according to claim 7, characterized in that: Configure the central angle direction, angular direction, and radius of the sector based on the category tags, browsing order, and click popularity of the entry buttons; output the radius, eccentric direction, and eccentric distance of the circle based on the radius, central angle direction, purchase data, and complaint data; and output the first recommendation result based on the overlapping area of ​​the sector and the circle, including: Direction of the central angle of the sector: If the user clicks only one term button in the first unit of time, then the central angle direction is constructed based on the center angle of the term button; If a user clicks multiple term buttons within the first unit of time, the average of the center angles of the multiple term buttons will be used. Construct the central angle direction; The angle of the sector: output according to the following formula: Zhang Jiao = Among them, the above This represents the center angle of each term button in the browsing sequence; This represents the average center angle of multiple term buttons in the browsing order; 'a' represents the initial angle; and 'n' represents the number of term buttons in the browsing order. The radius of the sector: The radius of the sector = ; The radius of the circle is half the radius of the sector; The eccentricity of the circle is in the same direction as the central angle of the sector; Eccentricity of a circle: Eccentricity = Purchase data × c - Complaint data × b; After initially aligning the center of the sector with the center of the circle, the circle is moved according to the eccentric direction and eccentric distance. The overlapping area of ​​the term buttons is configured according to the overlapping area of ​​the sector and the circle. The first recommendation result is output in descending order of the overlapping area.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores instructions that, when the computer-readable storage medium is run on a computer, cause the computer to perform a user demand recommendation method as described in any one of claims 1 to 5.

10. A computer program product containing instructions, characterized in that, When the computer program product is run on a computer, it causes the computer to perform a user demand recommendation method as described in any one of claims 1 to 5.

Images