Semantic analysis-based user portrait recognition method and website recommendation method

Abstract

The invention relates to a user portrait recognition method and a website recommendation method based on semantic analysis, and the user portrait recognition method comprises the steps: dividing a website which is browsed by a current user in a first preset time period and relates to the same category into a plurality of sub-categories, and taking each sub-category as a point of a portrait; and for each sub-category, determining the sub-category as a current sub-category, determining other sub-categories as adjacent sub-categories, constructing edges, pointing to all the adjacent sub-categories, of the current sub-category, and taking a correlation coefficient, pointing to the adjacent sub-categories, of the current sub-category as a weight of the edges to form a user portrait. Consumption habits of consumers and relevance among different commodities are fully considered, on the basis, recommendation weights of websites are corrected layer by layer through browsing information and consumption information of current users in combination with browsing information and consumption information of group users, personalized and high-accuracy website recommendation schemes are provided for the users, and user experience is improved. And the user experience is improved.

Description

Technical Field

[0001] This invention relates to the field of user profiling and information push technology, specifically to a user profiling recognition method and a website recommendation method based on semantic analysis. Background Technology

[0002] User profiling, as an effective tool for outlining the needs and preferences of target users, has become an important tool for e-commerce platforms in the field of digital marketing to acquire potential customers, achieve precise content delivery, and enhance brand influence.

[0003] Most existing user profiling tools analyze user behavior from multiple perspectives, such as user interests, hobbies, age, and occupation, and further predict subsequent purchasing behavior. However, current user profiling tools do not consider the correlation between different products in real-world application scenarios, nor do they take into account consumer habits: for example, after purchasing a specific product, consumers no longer need to browse similar products, yet many e-commerce platforms repeatedly recommend similar items, negatively impacting the consumer experience. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a user profile recognition method based on semantic analysis, comprising the following steps: S1: Divide the websites that the current user browses within the same category during the first predetermined time period into multiple subcategories. Each subcategory serves as a point in the profile, and each point stores the current user's browsing and behavior information under the corresponding subcategory. S2: For each subcategory, determine it as the current subcategory, determine other subcategories as adjacent subcategories, construct edges from the current subcategory to all adjacent subcategories, and use the correlation coefficient between the current subcategory and adjacent subcategories as the weight of the edges to form a user profile; The correlation coefficient between the current subcategory and adjacent subcategories is determined by the browsing and behavioral information of a group of users in the current and adjacent subcategories, using the following formula: Correlation coefficient = α1 H+α2 N; H is the browsing correlation coefficient from the current subcategory to the adjacent subcategory, N is the behavioral correlation coefficient from the current subcategory to the adjacent subcategory, α1 and α2 are the weights of the browsing correlation coefficient and the behavioral correlation coefficient, respectively, and the sum of α1 and α2 is 1.

[0005] The method for determining the browsing association coefficient H between the current subcategory and the adjacent subcategory is as follows: ; Where A represents the current subcategory and B represents the adjacent subcategory; This represents the total browsing time of a group of users on the current subcategory website. This represents the total number of clicks made by a group of users on a website within the current subcategory. The total browsing time on the current subcategory website is the total browsing time of a group of users browsing adjacent subcategory websites within a second predetermined time period after browsing the current subcategory website. This refers to the total number of clicks made by a group of users on the current subcategory website during a second predetermined time period after browsing the current subcategory website.

[0006] The method for determining the behavioral association coefficient N between the current subcategory and the adjacent subcategory is as follows: N=( ) / ; in, This represents the total number of times a group of users triggered a pre-order action on the current subcategory website. This refers to the total number of times a group of users trigger a reservation on an adjacent subcategory website within the third reservation period after triggering a reservation behavior on the current subcategory website. This refers to the total number of times a group of users did not trigger a reservation on an adjacent subcategory website within the third reservation period after triggering the reservation behavior on the current subcategory website. and The first row represents the correction parameters, and the second row represents the correction parameters, respectively.

[0007] The first row is the correction parameter. With the second line of correction parameters The sum is 1, and the first row is the correction parameter. Greater than the second line correction parameter .

[0008] The aforementioned pre-ordering behavior is a purchasing behavior.

[0009] The present invention also provides a website recommendation method based on the user profile recognition method described in any one of the above claims, comprising the following steps: S3: Based on the current user's browsing information in the current subcategory, identify each website in the current subcategory as the target website, determine the first recommendation coefficient of the target website, and use it as its recommendation weight; S4: Sort the recommendation weights of all target websites in all subcategories in descending order, and recommend the website ranked first.

[0010] In S3, the first recommendation coefficient is determined by the following formula: = LM+ P; in, It is the highest recommendation rating; LM is the subcategory recommendation coefficient, which is determined by the current user's browsing behavior in the current subcategory and the current user's browsing behavior in all subcategories in the user profile. P is the similarity coefficient, which is determined by the product feature information of the target website and the product feature information of the websites that the current user is browsing in the current subcategory. and These are the weights of the subcategory recommendation coefficient and the similarity coefficient, respectively.

[0011] The method for determining the subcategory recommendation coefficient (LM) is as follows: LM= ; in, This refers to the total browsing time of the current user on the current subcategory website within the first predetermined time period. This represents the total number of clicks made by the current user on the current subcategory website within the first predetermined time period. This represents the total browsing time of the current user across all subcategories within the profile during the first predetermined time period. This represents the total number of clicks made by the current user across all subcategories of websites within the first predetermined time period.

[0012] The methods for determining the similarity coefficient P include: S31: Obtain the set of websites within the current subcategory that the current user has browsed during the first predetermined time period, and determine the first product feature set vector using the following formula: ; Where i represents a specific website i in the current subcategory website set, and k represents the number of websites in the current subcategory website set that the current user browses within the first predetermined time period; This indicates the duration of a user's browsing time on a specific website i within the first predetermined time period; This indicates the number of clicks made by the current user on a specific website i within the first predetermined time period; This represents the vector of product feature set for a specific website i, obtained through semantic analysis of the product features of that specific website i. S32: Perform semantic analysis on the product feature information of the target website to obtain the second product feature set vector C'; S33: Will The semantic similarity between C and C' is determined as the similarity coefficient P.

[0013] S3 further includes: updating the first recommendation coefficient by using the correlation coefficient between the adjacent subcategories of the current subcategory and the current subcategory to obtain the second recommendation coefficient, which serves as the recommendation weight of the target website.

[0014] The user profile recognition method and website recommendation method based on semantic analysis provided by this invention fully consider consumers' consumption habits and the correlation between different products. On this basis, by combining the current user's browsing and consumption information with the browsing and consumption information of group users, the recommendation weight of the website is adjusted layer by layer to provide users with personalized and highly accurate website recommendation solutions, thereby improving the user experience. Attached Figure Description

[0015] Figure 1 The flowchart illustrates the implementation of the semantic analysis-based user profiling recognition method and website recommendation method of this invention.

[0016] Figure 2 Example of a user profile constructed in this invention. Detailed Implementation

[0017] To gain a better understanding of the technical solution and beneficial effects of the present invention, the technical solution of the present invention and its beneficial effects are described in detail below with reference to the accompanying drawings.

[0018] This invention identifies the current user's browsing and consumption behavior, categorizes the websites involved in these behaviors, and constructs user profile nodes. Then, based on the browsing and consumption behavior of a group of users on the categorized websites, it constructs edges between these user profile nodes to link multiple subcategories within the same main category. The weights of these edges reflect the relevance of products on websites within each subcategory. Based on this foundation, website recommendations are made, ensuring that the recommended websites fully consider product attributes and the current user's consumption behavior. Please refer to [the relevant documentation / contact information]. Figure 1 As shown, it mainly includes the following steps S1 to S10.

[0019] Step S1: Divide the websites related to the same category that the current user browsed within the first predetermined time period into multiple subcategories. Each subcategory serves as a point in the user profile. Each point stores the current user's browsing and behavioral information within the corresponding subcategory. The browsing information specifically includes the browsing duration and number of clicks of the current user on the websites within the corresponding subcategory within the first predetermined time period. The behavioral information includes whether the current user performed any related consumption behaviors, such as purchasing, on the products of the websites within the corresponding subcategory within the first predetermined time period.

[0020] In this invention, websites under the same category include various shopping websites belonging to the same product category, such as digital products, clothing, home furnishings, and books. Subcategories correspond to shopping websites that are further subdivided under the same product category. For example, digital office products include photography equipment, mobile phones and computers, office equipment, and video equipment; clothing can include women's and men's clothing, as well as tops, trousers, shoes and hats, and winter, summer, and spring / autumn clothing; home furnishings can include hard furnishings and soft furnishings, as well as bedding and cabinets; and books can include military, political, economic, cultural, and entertainment books, as well as children's and adult reading books.

[0021] Therefore, subcategories can involve various different classification principles. The core invention point of this invention is to recommend products on websites based on the correlation of consumer behavior between different subcategories. Therefore, when dividing subcategories, it is preferable to divide them from the perspective of highly correlated consumer behavior. For example, the probability of the same consumer buying both winter and summer clothes is less than the probability of buying both tops and pants. Therefore, when dividing clothing websites into multiple subcategories, it is preferable to divide them according to the principle of tops, pants, and shoes / hats. Similarly, digital office products can preferably be divided into photography equipment, photography accessories, video equipment, video accessories, office equipment, office supplies, etc. The subcategories of a website can be directly extracted from the website attributes, or a brief analysis can be performed using a Bayesian classification model. When the classification attributes of two subcategories are similar, the two subcategories can be merged.

[0022] Step S2: For each subcategory, determine it as the current subcategory, determine other subcategories as adjacent subcategories, construct edges from the current subcategory to all adjacent subcategories, and use the correlation coefficient between the current subcategory and adjacent subcategories as the weight of the edges to form a user profile.

[0023] Figure 2This is a schematic diagram of a user profile constructed according to the present invention: The websites browsed by the current user within a first predetermined time period that involve the same major category include three subcategories: A, B, and C. For subcategory A, the correlation coefficients pointing to subcategory B and subcategory C are 0.4 and 0.9, respectively. For subcategory B, the correlation coefficients pointing to subcategory A and subcategory C are 0.7 and 0.8, respectively. For subcategory C, the correlation coefficients pointing to subcategory A and subcategory B are 0.9 and 0.2, respectively.

[0024] In this invention, the correlation coefficient is used to characterize the dependency relationship between two subcategories. The higher the correlation coefficient from the current subcategory to the adjacent subcategory, the more likely consumers are to purchase products from the adjacent subcategory website after purchasing products from the current subcategory website. For subcategories A and B in this invention, the correlation coefficient from subcategory A to subcategory B is 0.4, while the correlation coefficient from subcategory B to subcategory A is 0.7. This indicates that consumers are not very willing to purchase products from subcategory B after purchasing products from subcategory A, but are more willing to purchase products from subcategory A after purchasing products from subcategory B. Therefore, in general scenarios, products from subcategory A are often used to support the function of products from subcategory B, or the product update cycle of products from subcategory A is shorter than that of products from subcategory B. For example, products from subcategory B involve photographic equipment, while products from subcategory A involve tripods, which falls into this category.

[0025] Similarly, for subcategories A and C, the correlation coefficient between them is 0.9, indicating that consumers usually purchase products from both subcategory websites at the same time. In general, the products from these two subcategory websites are usually highly compatible; for example, mattresses and beds belong to this scenario.

[0026] Therefore, by constructing a system that divides products into subcategories and determines the correlation coefficients between each subcategory, this invention can concretely display the types of products a user browses within a certain time period and the dependencies between these product types, and make subsequent product website recommendations in a correlated manner.

[0027] In this invention, with subcategory A as the current subcategory and subcategory B as the adjacent subcategory, the method for determining the association coefficient between the current subcategory A and the adjacent subcategory B is as follows: Correlation coefficient = α1 { / 2}+α2 {( ) / }; in,{ / 2} and {( ) / } represents the browsing correlation coefficient and behavioral correlation coefficient of the current subcategory pointing to the adjacent subcategory, respectively; α1 and α2 are the weights of the browsing correlation coefficient and behavioral correlation coefficient, respectively, and the sum of α1 and α2 is 1; A represents the current subcategory, and B represents the adjacent subcategory; This represents the total browsing time of a group of users on website A within the current subcategory. This represents the total number of clicks made by a group of users on website A within the current subcategory. The total browsing time on the current subcategory A website is the total browsing time of a group of users browsing the adjacent subcategory B website during the second predetermined time period after browsing the current subcategory A website. This refers to the total number of clicks made by a group of users on the current subcategory A website during a second predetermined time period after browsing the current subcategory A website.

[0028] This represents the total number of times a group of users triggered a purchase on the current subcategory A website. The total number of times a group of users triggers a purchase on the adjacent subcategory B website within a third predetermined time period after triggering a purchase on the current subcategory A website. The total number of times a group of users did not trigger a purchase on the adjacent subcategory B website within the third predetermined time period after triggering a purchase on the current subcategory A website; and The first row represents the correction parameters and the second row represents the correction parameters, respectively. The first row represents the correction parameters. With the second line of correction parameters The sum is 1, and the first row is the correction parameter. Greater than the second line correction parameter .

[0029] For example, suppose , Regarding the current subcategory A: A group of users spent a total of 5 million hours browsing websites in subcategory A. Of these 5 million hours, 2 million hours of browsing involved users who, within a second predetermined time period after completing their initial browsing session, also browsed websites in the adjacent subcategory B. Therefore... =5 million =2 million; A group of users clicked on websites in subcategory A a total of 10 million times. Of these 10 million clicks, 4 million of these users also clicked on websites in the adjacent subcategory B within a second predetermined time period after their initial click. Therefore... =10 million =4 million; A group of users made 2 million purchases on website A (subcategory A). Of these 2 million purchases, 500,000 involved users who also purchased items from website B (adjacent subcategory B) within a third predetermined time period after their initial purchase. The remaining 1.5 million involved users who did not purchase any items from website B within the third predetermined time period after their initial purchase. =2 million =500,000 =1.5 million; Therefore, the correlation coefficient = α1 { / 2}+α2 ( ) = α1 0.4+α2 0.45.

[0030] Therefore, this invention visualizes the browsing and consumption behavior of a group of users between two subcategories as a correlation coefficient pointing from one subcategory to another. In this way, the correlation coefficient represents the degree of dependence of products in adjacent subcategory websites on products in the current subcategory website, laying the foundation for subsequent recommendations of related product websites.

[0031] In this invention, the so-called first predetermined time period, second predetermined time period, and third predetermined time period can be the same or different.

[0032] Step S3: Based on the current user's browsing information in the current subcategory, identify each website belonging to the current subcategory as the target website and determine the recommendation weight of the target website.

[0033] In a first preferred embodiment of the present invention, a first recommendation coefficient for the target website is determined by the following formula, and the first recommendation coefficient is assigned to the recommendation weight of the target website: = LM+ P; in, It is the highest recommendation rating; LM is the subcategory recommendation coefficient, which is determined by the current user's browsing behavior in the current subcategory and the current user's browsing behavior in all subcategories in the user profile. P is the similarity coefficient, which is determined by the product feature information of the target website and the product feature information of the websites that the current user is browsing in the current subcategory. and These are the weights of the subcategory recommendation coefficient and the similarity coefficient, respectively. and The sum is 1.

[0034] The method for determining the subcategory recommendation coefficient (LM) is as follows: LM= ; in, This refers to the total browsing time of the current user on the current subcategory website within the first predetermined time period. This represents the total number of clicks made by the current user on the current subcategory website within the first predetermined time period. This represents the total browsing time of the current user across all subcategories within the profile during the first predetermined time period. This represents the total number of clicks made by the current user across all subcategories of websites within the first predetermined time period.

[0035] For example, suppose a user browses websites in subcategories A, B, and C for a total of 1 hour, 2 hours, and 2 hours respectively within a first predetermined time period, and clicks on websites in subcategories A, B, and C 4, 5, and 7 times respectively within the first predetermined time period. Then, the recommendation coefficient LM for subcategory A is ( ). ) / 2=0.225.

[0036] The method for determining the similarity coefficient P includes the following steps S31 to S33.

[0037] Step S31: Obtain the set of websites within the current subcategory that the current user has browsed during the first predetermined time period, and determine the first product feature set vector using the following formula: ; Where i represents a specific website i in the current subcategory website set, and k represents the number of websites in the current subcategory website set that the current user browses within the first predetermined time period; This indicates the duration of a user's browsing time on a specific website i within the first predetermined time period; This indicates the number of clicks made by the current user on a specific website i within the first predetermined time period; As mentioned above, this represents the total browsing time of the current user on the current subcategory website within the first predetermined time period; As mentioned above, this represents the total number of clicks made by the current user on the current subcategory website within the first predetermined time period. This represents the vector of product feature set for a specific website i, obtained through semantic analysis of the product features of that specific website i. For example: For subcategory A, suppose the current user browsed four websites under it within a first predetermined time period: A1, A2, A3, and A4. After semantic analysis of the product features of the four websites, we obtain product feature set vectors C1, C2, C3, and C4 for each website. Within the first predetermined time period, the current user's browsing time on the four websites was 15 minutes, 10 minutes, 20 minutes, and 15 minutes respectively, and the number of clicks on each website was 1. Then: C={( ) / 2} C1+{( ) / 2} C2+{( ) / 2} C3+{( ) / 2} C4=0.25 C1+0.21 C2+0.29 C3+0.25 C 4。

[0038] Step S32: Perform semantic analysis on the product feature information of the target website to obtain the second product feature set vector C'; Step S33: ... The semantic similarity between C and C' is determined as the similarity coefficient P.

[0039] Therefore, the first recommendation coefficient of this invention takes into account: The product features of the target website (represented above as the second product feature set vector); When a user browses products in a target website's category, this reflects their tendency to browse all categories of websites within a certain time period (represented as a subcategory recommendation coefficient in the above text through user browsing behavior information). The characteristics of products viewed by a user within a certain time period; and the user's preference for each product when browsing products in the category of the target website within a certain time period (represented in the above as the first product feature set vector).

[0040] Thus, the information represented by the first recommendation coefficient is diverse and comprehensive, and therefore better meets the actual needs of users.

[0041] In a second preferred embodiment of the present invention, the first recommendation coefficient is updated by the correlation coefficient between the adjacent subcategories of the current subcategory and the current subcategory to obtain the second recommendation coefficient, and the second recommendation coefficient is assigned to the recommendation weight of the target website.

[0042] Specifically, it includes the following steps S34 to S35.

[0043] Step S34: If the current user triggers a purchase behavior for a product in an adjacent subcategory website, then the adjacent subcategory is identified as the target adjacent subcategory.

[0044] Step S35: Obtain the second recommendation coefficient using the following formula: = ; in, The second recommendation coefficient is denoted by j, which represents the target's adjacent subcategories, and m is the number of the target's adjacent subcategories. The correlation coefficient between the target adjacent subcategories and the current subcategory. This is the correction factor.

[0045] For example, with Figure 2 Based on the image shown, assuming The correlation coefficient is 0.5, and the current user has made purchases in both subcategory B and subcategory C websites. Both subcategory B and subcategory C belong to the target adjacent subcategories of subcategory A. m=2, and the correlation coefficient between the target adjacent subcategory B and the current subcategory A is [value missing]. =0.7, the correlation coefficient between the target adjacent subcategory C and the current subcategory A. =0.9, then the second recommendation coefficient for the current subcategory A is 0.9. = {( +1)+( +1)}= 5.61.

[0046] As mentioned above, the closer the correlation coefficient between adjacent subcategories and the current subcategory is to 1, the more likely a user is to purchase products from the current subcategory website after having purchased products from adjacent subcategory websites; furthermore, the correlation coefficient is always between 0 and 1; therefore, this invention introduces a constant correlation coefficient greater than 1: "{ / (1- "+1" makes it have a positive corrective effect on the second recommendation coefficient, and, The closer it is to 1, the more "{ / (1- The higher the value of ")+1", the more it passes through "{ / (1- The higher the second recommendation coefficient after correction ()+1), the higher the correlation coefficient is reflected in the second recommendation coefficient, so as to influence the recommendation value of the target website through the purchasing behavior of adjacent subcategories.

[0047] Correction factor The introduction of the denominator (1- The settings are all to avoid the influence of excessively high correlation coefficients on the primary recommendation coefficient, which would significantly weaken the impact of other factors on the target website's recommendation weight. For example, as shown above, the correlation coefficient between subcategory A and subcategory C is 0.9. Let's assume that substituting it into "1 / (1- If we use this value to correct the first recommendation coefficient, the subsequent updated recommendation weight will be 10 times the first recommendation coefficient. Therefore, this invention uses "{ / (1- The first recommendation coefficient is corrected by adding 1. The value is usually between 0.5 and 5. For example, when 0.5 When taken as 0.9 =3.63, compared to 1 / (1- =10, which significantly buffers the unreasonable correction effect of the correlation coefficient on the first recommendation coefficient.

[0048] In a third preferred embodiment of the present invention, the second recommendation coefficient is updated based on whether the current user has triggered a purchase behavior for the products in the current subcategory website, to obtain a third recommendation coefficient, and the third recommendation coefficient is assigned to the recommendation weight of the target website.

[0049] = ξ; in, It is the third recommendation coefficient; ξ is the third row correction parameter: when the current user has made a purchase for a product in the current subcategory website, this value is less than 1; conversely, when the current user has not made a purchase for a product in the current subcategory website, this value is greater than 1.

[0050] The principle behind this step is to take into account the purchasing and browsing habits of most consumers: triggering a purchase means that the consumer has already selected a product that suits their taste, therefore, there is no need to recommend too many related product websites.

[0051] S4: Sort the recommendation weights of all target websites in all subcategories in descending order, and recommend the website ranked first.

[0052] In a fourth preferred embodiment of the present invention, the third recommendation coefficient of the target website in the current subcategory is updated based on the specific website pointed to by the current user's purchase behavior in the adjacent subcategories of the current subcategory, to obtain a fourth recommendation coefficient. The fourth recommendation coefficient is then assigned to the recommendation weight of the target website, and the ranking of the target website is adjusted based on the recommendation weight.

[0053] One specific implementation includes the following steps S5 to S10.

[0054] For each target adjacent subcategory determined in step S34, perform the following steps S5 to S9.

[0055] Step S5: Identify the websites in the target adjacent subcategories where the current user triggered a purchase as the baseline websites.

[0056] Step S6: Perform semantic analysis on the product feature information of the benchmark website to obtain the benchmark product feature set vector D'.

[0057] Step S7: Obtain the set of websites where users in the group also triggered purchasing behavior on websites within the current subcategory within the first predetermined time period after purchasing products from the benchmark website, and use this as the comparison website set; perform semantic analysis on the product feature information of each website in the comparison website set to obtain the product feature set vector of each website in the comparison website set. .

[0058] Step S8: Obtain the feature set vector of the comparison website products: = ; in, The vector represents the set of product features of the comparison websites, where n is the number of websites in the comparison website set determined for the j-th target's adjacent subcategories.

[0059] For example, still using Figure 2Based on the profile shown, assuming that for the current subcategory A's adjacent subcategories B and C, the current user has made purchases on both subcategory B and subcategory C websites, and both subcategory B and subcategory C belong to the target adjacent subcategories of subcategory A, m=2; assuming that the current user has purchased goods on subcategory B's website B1 and subcategory C's website C1, then website B1 and website C1 are determined as the baseline websites.

[0060] For the first benchmark website B1, if historically, after a group of users purchased goods on B1, they also purchased goods from websites A1, A2, A3, and A4 within the current subcategory within the first predetermined time period, then n=4, and the feature set vectors of the goods from websites A1, A2, A3, and A4 are respectively... , , , ;but = + + / 4.

[0061] Similarly, for the second benchmark website C1, if historically, after purchasing goods on C1, a group of users purchased goods from websites A3, A4, and A5 within the first predetermined time period of the current subcategory, then n=3, and the feature set vectors of the goods on websites A3, A4, and A5 are respectively... , , ;but = + + / 3.

[0062] Step S9: Obtain the second product feature set vector C' of the target website and the product feature set vector of the comparison website from step S32. Semantic similarity, used as the calibration similarity of adjacent subcategories of the target. .

[0063] Following on from the previous section, the similarity of the target adjacent sub-category B and the target adjacent sub-category C is calibrated. These are respectively the second product feature set vector C' of the target website and the product feature set vector of the comparison website. Semantic similarity.

[0064] Step S10: After all adjacent subcategories of the target have completed steps S5 to S9, determine the fourth recommendation coefficient using the following formula: ; Where m has the meaning as described in step S35 above, it represents the number of adjacent subcategories of the target. The calibration weights for adjacent subcategories are determined based on the current user's browsing information in each adjacent subcategory. These weights are positively correlated with the current user's browsing time and click count in each adjacent subcategory, similar to the method used to determine the subcategory recommendation coefficient (LM). ; in, This represents the total browsing time of the current user on the website of the j-th adjacent target subcategory within the first predetermined time period. This represents the total number of clicks made by the current user on the website of the adjacent subcategory of the j-th target within the first predetermined time period. This represents the total browsing time of the current user across all target adjacent subcategory websites within the first predetermined time period. This represents the total number of clicks made by the current user across all target adjacent subcategories within the first predetermined time period.

[0065] It should be noted that in this invention, the term " The purpose of this feature is to make the recommendation weight of the target website more accurate. However, since its value is also constant between 0 and 1, its introduction will inevitably lead to a decrease in the recommendation weight of the target website. Therefore, when adjusting the order of the target websites according to the updated recommendation weight, only the ranking position of the target websites belonging to the same subcategory is adjusted within their original ranking position range, and the ranking position of all target websites is not adjusted.

[0066] For example: Suppose that there are 5 target websites identified in subcategory A: A1, A2, A3, A4, A5; 5 target websites identified in subcategory B: B1, B2, B3, B4, B5; and 5 target websites identified in subcategory C: C1, C2, C3, C4, C5.

[0067] The recommendation weights of each item before steps S5 to S10 were executed, and the results of sorting them in descending order based on the recommendation weights, are as follows:

[0068] Subsequently, assuming that subcategory A has a target adjacent subcategory, the recommendation weights of A1, A2, A3, A4, and A5 are updated based on the product information of websites purchased by user A in the target adjacent subcategory: the recommendation weights of A1, A2, A3, A4, and A5 are updated from 0.84, 0.63, 0.61, 0.47, and 0.56 as shown in the table above to 0.73, 0.54, 0.60, 0.45, and 0.33 respectively. Since A1, A2, A3, A4, and A5 were ranked 1st, 3rd, 4th, 9th, and 6th in the recommendation ranking before the update, these five websites will still occupy these ranking positions after the update. The positions are only adjusted based on the updated recommendation weights of the five websites. However, the positions of websites under subcategory B and subcategory C are not adjusted, as follows:

[0069] Otherwise, if a global update is performed, website A5, with the lowest weight value (0.33), will be ranked 15th from the bottom. However, before the update, website A5 was ranked 6th, so its recommendation bias should be relatively strong. Only because the websites in subcategories B and C have not been corrected by the benchmark website, resulting in website A5 exhibiting a "false lowest recommendation weight," is it unfair to website A in subcategory A.

[0070] Although the present invention has been described using the above preferred embodiments, it is not intended to limit the scope of protection of the present invention. Any changes and modifications made by those skilled in the art to the above embodiments without departing from the spirit and scope of the present invention shall still fall within the scope of protection of the present invention. Therefore, the scope of protection of the present invention shall be defined by the claims.

Claims

1. A user profile recognition method based on semantic analysis, characterized in that, Includes the following steps: S1: Divide the websites that the current user browses within the same category during the first predetermined time period into multiple subcategories. Each subcategory serves as a point in the profile, and each point stores the current user's browsing and behavior information under the corresponding subcategory. S2: For each subcategory, determine it as the current subcategory, determine other subcategories as adjacent subcategories, construct edges from the current subcategory to all adjacent subcategories, and use the correlation coefficient between the current subcategory and adjacent subcategories as the weight of the edges to form a user profile; The correlation coefficient between the current subcategory and adjacent subcategories is determined by the browsing and behavioral information of a group of users in the current and adjacent subcategories, using the following formula: Correlation coefficient = α1 H+α2 N; H is the browsing correlation coefficient from the current subcategory to the adjacent subcategory, N is the behavioral correlation coefficient from the current subcategory to the adjacent subcategory, α1 and α2 are the weights of the browsing correlation coefficient and the behavioral correlation coefficient, respectively, and the sum of α1 and α2 is 1.

2. The user profile recognition method based on semantic analysis as described in claim 1, characterized in that, The method for determining the browsing association coefficient H between the current subcategory and the adjacent subcategory is as follows: ； Where A represents the current subcategory and B represents the adjacent subcategory; This represents the total browsing time of a group of users on the current subcategory website. This represents the total number of clicks made by a group of users on a website within the current subcategory. The total browsing time on the current subcategory website is the total browsing time of a group of users browsing adjacent subcategory websites within a second predetermined time period after browsing the current subcategory website. This refers to the total number of clicks made by a group of users on the current subcategory website during a second predetermined time period after browsing the current subcategory website.

3. The user profile recognition method based on semantic analysis as described in claim 1, characterized in that, The method for determining the behavioral association coefficient N between the current subcategory and the adjacent subcategory is as follows: N=（ ） / ； in, This represents the total number of times a user group triggers a pre-order action on the current subcategory website. This refers to the total number of times a group of users trigger a reservation on an adjacent subcategory website within the third reservation period after triggering a reservation behavior on the current subcategory website. This refers to the total number of times a group of users did not trigger a reservation on an adjacent subcategory website within the third reservation period after triggering the reservation behavior on the current subcategory website. and The first row represents the correction parameters, and the second row represents the correction parameters, respectively.

4. The user profile recognition method based on semantic analysis as described in claim 3, characterized in that, The first line contains correction parameters. With the second line of correction parameters The sum is 1, and the first row is the correction parameter. Greater than the second line correction parameter .

5. The user profile recognition method based on semantic analysis as described in claim 3, characterized in that, The aforementioned pre-ordering behavior is a purchasing behavior.

6. A website recommendation method based on the user profile recognition method according to any one of claims 1-5, characterized in that, Includes the following steps: S3: Based on the current user's browsing information in the current subcategory, identify each website in the current subcategory as the target website, determine the first recommendation coefficient of the target website, and use it as its recommendation weight; S4: Sort the recommendation weights of all target websites in all subcategories in descending order, and recommend the website ranked first.

7. The website recommendation method as described in claim 6, characterized in that: In S3, the first recommendation coefficient is determined by the following formula: = LM+ P; in, It is the highest recommendation rating; LM is the subcategory recommendation coefficient, which is determined by the current user's browsing behavior in the current subcategory and the current user's browsing behavior in all subcategories in the user profile. P is the similarity coefficient, which is determined by the product feature information of the target website and the product feature information of the websites that the current user is browsing in the current subcategory. and These are the weights of the subcategory recommendation coefficient and the similarity coefficient, respectively.

8. The website recommendation method as described in claim 7, characterized in that, The method for determining the subcategory recommendation coefficient (LM) is as follows: LM= ; in, This refers to the total browsing time of the current user on the current subcategory website within the first predetermined time period. This represents the total number of clicks made by the current user on the current subcategory website within the first predetermined time period. This represents the total browsing time of the current user across all subcategories within the profile during the first predetermined time period. This represents the total number of clicks made by the current user across all subcategories of websites within the first predetermined time period.

9. The website recommendation method as described in claim 8, characterized in that, Methods for determining the similarity coefficient P include: S31: Obtain the set of websites within the current subcategory that the current user has browsed during the first predetermined time period, and determine the first product feature set vector using the following formula: ； Where i represents a specific website i in the current subcategory website set, and k represents the number of websites in the current subcategory website set that the current user browses within the first predetermined time period; This indicates the duration of a user's browsing time on a specific website i within the first predetermined time period; This indicates the number of clicks made by the current user on a specific website i within the first predetermined time period; This represents the vector of product feature set for a specific website i, obtained through semantic analysis of the product features of that specific website i. S32: Perform semantic analysis on the product feature information of the target website to obtain the second product feature set vector C'; S33: Will The semantic similarity between C and C' is determined as the similarity coefficient P.

10. The website recommendation method as described in claim 6, characterized in that, S3 further includes: updating the first recommendation coefficient by using the correlation coefficient between the adjacent subcategories of the current subcategory and the current subcategory to obtain the second recommendation coefficient, which serves as the recommendation weight of the target website.

Images