An intelligent guidance method for e-commerce platforms based on deep learning multi-dimensional user portraits

Abstract

The present invention relates to the technical field of e-commerce platforms, and specifically relates to an intelligent guidance method for an e-commerce platform with a deep learning multi-dimensional user portrait. By constructing demand tags and interest tags for users, combining with the behavioral data of the target user group, it accurately predicts user demands and conducts personalized product recommendations. Demand tags and interest tags are constructed through deep learning algorithms. By calculating the correlation degree of the target user group, a demand budget for the user group is constructed, and different levels of interest guidance are carried out according to the interest intensity of users in specific products or services. The system also makes dynamic adjustments based on market feedback, such as adjusting product pricing and promotion strategies through positive feedback. This method can effectively improve the user purchase conversion rate, enhance the profitability of the platform, and has a high level of intelligence, and is applicable to the application of e-commerce platforms in personalized marketing.

Description

Technical Field

[0001] The present invention relates to the technical field of e-commerce platforms, and specifically relates to an intelligent guidance method for an e-commerce platform with a deep learning multi-dimensional user portrait. Background Art

[0002] An e-commerce platform refers to a third-party organization that provides a complete set of e-commerce service solutions for merchants on the Internet, including e-commerce website construction and subsequent operation services. Common APPs such as Taobao, JD.com, Douyin, and Xiaohongshu are all e-commerce platforms; with the development of e-commerce enterprises, the daily basic maintenance of websites has also become an issue that cannot be ignored in the daily operations of e-commerce enterprises.

[0003] During the development of a website, data resources, content resources, user resources, etc. will become increasingly rich, and thus the situation will become increasingly complex. During the daily operation of an e-commerce platform website, the maintenance and optimization of tens of thousands of web pages have gradually increased the work difficulty; in the prior art, operation management is usually carried out through a pre-established middle platform, and the middle platform will bind sticky data with relevance to users. When the sticky data deviates, the deviation value will be recorded, usually calculated using a distance function.

[0004] Technical Routes and Existing Problems of the Prior Art:

[0005] Traditional User Portrait Construction Methods: The prior art mostly adopts traditional user portrait construction methods based on rules and manual annotation, mainly constructing user portraits by collecting users' basic information, historical behavior data, and purchase preferences. Its technical route mainly relies on feature engineering and simple machine learning algorithms, and users are divided through clustering or classification models. Although these methods can reflect users' needs and interests to a certain extent, the following problems exist:

[0006] 1. Data Sparsity: Due to the relatively static data processing method of traditional methods, it is difficult to effectively capture the dynamic changes of user behavior, resulting in data sparsity problems and affecting the accuracy of the portrait.

[0007] 2. Inefficient Label Construction: Label construction usually relies on manual design, unable to deeply explore users' potential complex needs and interests, and unable to perform personalized and refined recommendations.

[0008] User Portrait Construction Methods Based on Deep Learning: With the development of deep learning, more and more e-commerce platforms have begun to adopt deep learning methods based on neural networks to construct user portraits. Its technical route usually includes training through large-scale user data to automatically learn users' potential features and needs. However, the problems faced by this method are:

[0009] 1. High computing resource requirements: Deep learning models require a large amount of computing resources and time for training. Especially when facing a huge amount of user data, they pose relatively high requirements for hardware configuration and storage.

[0010] 2. Complexity of label construction: Although deep learning can automatically learn features, how to design a reasonable label construction mechanism, especially in the construction of multi-dimensional labels, is still a difficult problem, which is likely to lead to model overfitting or label failure.

[0011] Personalized recommendation algorithms: Most existing personalized recommendation systems adopt traditional methods such as collaborative filtering and content-based recommendation. These methods can provide accurate recommendations in some cases. However, these methods usually only focus on single-dimensional user behavior and fail to comprehensively consider the multi-dimensional features of user interests. The existing recommendation systems have the following problems:

[0012] 1. Cold start problem of interests: For new users or new products, the system cannot obtain enough historical data, resulting in poor recommendation effects.

[0013] 2. Interest drift problem: Users' interests change over time, and existing methods cannot flexibly adapt to the dynamic changes of users' interests, resulting in poor long-term effectiveness of recommendations. Summary of the Invention

[0014] In view of the above problems, the present invention provides an effective solution. The object of the present invention is to provide an intelligent guidance method for an e-commerce platform with a deep learning multi-dimensional user portrait, including the following steps:

[0015] S1. Label construction: Use deep learning methods to construct a label set;

[0016] S2. Calculate the relevance of target users: For the target user group, construct a target user vector group and a feature label vector, and perform weighted summation to obtain a target user association function, which is used as the attention value of the target user for the e-commerce platform;

[0017] S3. Calculate the demand budget of the target user group: Calculate the group user association degree and construct a group user association function, and combine it with the target user association function to calculate the proportion of the target user group's interest intensity;

[0018] S4. Interest guidance response: Based on the proportion of the target user group's interest intensity, conduct guidance for the target user.

[0019] Further, the S1 includes:

[0020] S101. Collect user data and divide the user data into subjective data and objective data;

[0021] S102. Based on the subjective data, construct an initial demand label set ; where the subscript i is the user group serial number;

[0022] S103. Substitute the initial demand label set into the matrix to construct an initial demand label matrix ; where the subscript i is the user group serial number;

[0023] S104. Based on objective data, construct an interest label set ;

[0024] S105. Substitute the interest label set into the matrix to construct an interest label matrix ;

[0025] S106. Based on the deep learning method, perform cyclic training on the initial demand label matrix and the interest label matrix to obtain a convergent demand label matrix and a convergent interest label matrix ;

[0026] S107. According to the convergent demand label matrix and the convergent interest label matrix , construct a generalized composite label set ;

[0027] Furthermore, the deep learning method in S106 is completed through the gradient descent algorithm.

[0028] Furthermore, the target users in S2 are users for the same specific commodity or service, and the attention of the target users to the target commodity or service is determined by a target user correlation function, and the target user correlation function satisfies:

[0029]

[0030] where represents the attention of the th user group as the target user group to the target commodity, and respectively represent the convergent demand label matrix value and the convergent interest label matrix value of the th user in the th user group; represents the number of users in the th user group;

[0031] Furthermore, S2 specifically includes:

[0032] S201. Construct a target user vector group for the target user group , where represents the target user vector of the th user group;

[0033] S202. Construct a target user label set based on the target user vector, and construct a feature label vector for each target user vector , and assign a weight parameter to obtain the target user label set ;

[0034] S203. Calculate the target user association function; find the included angle between the target user vector group and the feature label vector , calculate the user correlation value, and perform weighted summation on the user correlation values of the same user group to obtain the target user association function .

[0035] Further, the specific steps of S3 include:

[0036] S301. According to the same target time, for the target user group, convert the target user vector group into a target user group vector group , satisfying:

[0037]

[0038] where is the time span for performing the smoothing function operation on , is the smoothing function, represents performing smoothing processing on the target user vector within the time span T, using the time window sliding algorithm; represents multiplying the attention as a multiple coefficient by the smoothed target user vector to obtain the target user group vector ;

[0039] S302. Calculate the target user group association degree, sum up the target user group vector group to obtain the target user group vector ;

[0040] S303. Calculate the correlation between the target user group vector and the feature label vector to obtain the target user group association function ;

[0041] S304. The target user association function and the target user group association function Determine the proportional value to obtain the proportion of the interest intensity of the target user group, satisfying:

[0042]

[0043] Among them, represents the proportion of the interest intensity of the target user group, that is, the proportion of the interest intensity of the target user group in the total user group interest intensity.

[0044] Furthermore, the S4 includes:

[0045] When the proportion of the interest intensity of the target user group reaches a preset first threshold, make a first-level guidance reaction;

[0046] When the proportion of the interest intensity of the target user group reaches a preset second threshold, make a second-level guidance reaction;

[0047] When the proportion of the interest intensity of the target user group reaches a preset third threshold, make a third-level guidance reaction;

[0048] Among them, the third threshold is greater than the second threshold, and the second threshold is greater than the first threshold.

[0049] Furthermore, in the third-level guidance reaction, targeted guidance is carried out for individual users, and target goods or services are directly pushed.

[0050] Furthermore, in the second-level guidance reaction, the guidance reaction module conducts comprehensive guidance for the target users, guides individual users to compare multiple similar products, and guides individual users to complete the purchase.

[0051] Furthermore, in the first-level guidance reaction, category guidance is carried out for the target users, the original other product labels of individual users are weighted, and individual users are guided to turn to other products.

[0052] Compared with the prior art, the beneficial effects of the present invention:

[0053] 1. Based on the deep learning of user portraits, the present invention constructs the individual-to-individual correlation degree and the individual-to-group correlation degree, and uses them as the basic parameters for e-commerce guidance. It can make three-level linkage intelligent guidance according to the changing individual-to-individual correlation degree and individual-to-group correlation degree, which not only efficiently realizes the accurate capture of needs, but also is conducive to the implementation of long-term marketing strategies.

[0054] 2. The deep learning of the present invention constructs a cyclic system for learning with subjective data and objective data, which not only improves the efficiency of deep learning, but also makes it possible to achieve accurate guidance with the constructed multi-dimensional label system.

[0055] 3. Based on the real-time behavior habits of different users, the present invention conducts statistical analysis to construct the group-to-group correlation degree and the individual-to-individual correlation degree. When the e-commerce platform performs precise push or guidance, the guidance content parameters will be transmitted back to the advertising side, which not only helps optimize the system itself, but also helps the advertising side optimize the advertising content. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] Figure 1 It is a flowchart of the method for the specific embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0057] The present invention will be further described in detail below in conjunction with the drawings and specific embodiments:

[0058] In this embodiment, an intelligent guidance method for an e-commerce platform with a deep learning multi-dimensional user portrait is disclosed, including the following steps:

[0059] S1. Tag construction: Use deep learning methods to construct a tag set; specifically including:

[0060] S101. Collect user data and divide the user data into subjective data and objective data;

[0061] S102. Based on the subjective data, construct an initial demand tag set ; where the subscript i is the user group serial number;

[0062] S103. Substitute the initial demand tag set into the matrix to construct an initial demand tag matrix ; where the subscript i is the user group serial number;

[0063] S104. Based on the objective data, construct an interest tag set ;

[0064] S105. Substitute the interest tag set into the matrix to construct an interest tag matrix ;

[0065] S106. Based on the gradient descent algorithm, perform cyclic training on the initial demand tag matrix and the interest tag matrix to obtain a convergent demand tag matrix and a convergent interest tag matrix ;

[0066] S107. According to the convergent demand tag matrix and the convergent interest tag matrix , construct a generalized composite tag set .

[0067] S2. Calculate the relevance of target users: For the target user group, construct a target user vector group and a feature label vector, and perform weighted summation to obtain the target user association function, which is used as the attention value of the target user for the e-commerce platform;

[0068] The target users are users for the same specific commodity or service. The attention of the target users to the target commodity or service is determined by the target user association function, and the target user association function satisfies:

[0069]

[0070] where, represents the attention of the th user group in the target user group to the target commodity, and respectively represent the convergent demand label matrix value and the convergent interest label matrix value of the th user in the th user group; represents the number of users in the th user group;

[0071] S2 specifically includes:

[0072] S201. Construct a target user vector group for the target user group , where represents the target user vector of the th user group;

[0073] S202. Based on the target user vector, construct a target user label set, and for each target user vector, construct a feature label vector , and assign a weight parameter to obtain the target user label set ;

[0074] S203. Calculate the target user association function; find the included angle between the target user vector group and the feature label vector to calculate the user relevance value, and perform weighted summation on the user relevance values of the users in the same user group to obtain the target user association function .

[0075] S3. Calculate the demand budget of the target user group: Calculate the group user association degree and construct the group user association function, and combine the target user association function to calculate the proportion of the interest intensity of the target user group; specifically including:

[0076] S301. According to the same target time, for the target user group, convert the target user vector group into the target user group vector group , satisfying:

[0077]

[0078] Among them, is the time span for performing a smoothing function operation on , is the smoothing function, represents smoothing the target user vector within the time span T, using a time window sliding algorithm; represents multiplying the attention as a multiple coefficient by the smoothed target user vector to obtain the target user group vector ;

[0079] S302. Calculate the correlation degree of the target user group, sum up the target user group vector group to obtain the target user group vector ;

[0080] S303. Calculate the correlation between the target user group vector and the feature label vector to obtain the target user group correlation function ;

[0081] S304. Determine the proportional value between the target user correlation function and the target user group correlation function to obtain the proportion of the target user group interest intensity, satisfying:

[0082]

[0083] Among them, represents the proportion of the target user group interest intensity, that is, the proportion of the interest intensity of the target user group in the total user group interest intensity.

[0084] S4. Interest guidance response: Based on the proportion of the target user group interest intensity, guide the target user; specifically including:

[0085] When the proportion of the target user group interest intensity reaches the preset first threshold, make a first-level guidance response; in the first-level guidance response, perform category guidance on the target user, weight the original other product labels of the individual user, and guide the individual user to turn to other products;

[0086] When the proportion of the target user group interest intensity reaches the preset second threshold, make a second-level guidance response; in the second-level guidance response, the guidance response module performs comprehensive guidance on the target user, guides the individual user to compare multiple similar products, and guides the individual user to complete the purchase;

[0087] When the proportion of the interest intensity of the target user group reaches a preset third threshold, a third-level guidance response is made; in the third-level guidance response, targeted guidance is provided to individual users, and target goods or services are directly pushed.

[0088] Among them, the third threshold is greater than the second threshold, and the second threshold is greater than the first threshold.

[0089] The market monitoring system will make different degrees of reverse operations according to the feedback of different market users. When the monitoring system detects positive feedback such as "99+ positive reviews, 0 negative reviews, and 99+ inventory reduction", the system administrator will recommend raising the commodity price.

[0090] Target users: Registered users of the e-commerce platform.

[0091] User group with the same target: Multiple users who have made behavior records such as browsing, consulting, evaluating, forwarding, or purchasing for the same commodity or service.

[0092] For each user in the user group with the same target, their own information and all behavior records are used as elements to construct user vectors, forming a target user group vector group. Specific Embodiment 2

[0094] An intelligent guidance method for an e-commerce platform with a deep learning multi-dimensional user portrait:

[0095] Background: An e-commerce platform hopes to improve the user purchase conversion rate through intelligent guidance. The platform uses the method of deep learning multi-dimensional user portraits to provide personalized commodity recommendations for users. In order to achieve precise guidance, the platform adopts the following intelligent guidance method.

[0096] Example: Practical operation steps of an e-commerce platform applying this method

[0097] S1. Label construction:

[0098] Collect user data: The e-commerce platform obtains the behavior data of each user through data such as the user's browsing history, search records, purchase behavior, and evaluation content. These data are divided into two categories: subjective data and objective data. Subjective data includes user evaluations, likes, or comment content on commodities; objective data includes the user's basic information (such as age, gender, geographical location, etc.) and behavior data (such as the time of browsing commodities, the number of clicks, etc.).

[0099] Construct an initial demand label set: Based on the user's subjective evaluation data, the system constructs demand labels for each user. For example, user A may mention in the comment "like mobile phones with high cost performance" or "pay more attention to the camera performance of mobile phones", which constructs labels such as "cost performance" and "photography" for user A.

[0100] Construct an interest tag set: Based on the objective data of users (such as browsing history and purchase behavior), the system constructs interest tags. For example, if user A frequently browses mobile phones and purchases a smartphone with high cost performance, then the system will label user A with interest tags such as "smartphone" or "high cost performance".

[0101] Perform gradient descent training: Through deep learning algorithms, the system trains these tags to identify the potential needs and interests of each user and generate a converged tag matrix. For example, the demand tag matrix of user A may show that he is more inclined to choose a mobile phone with higher cost performance, while the interest tag matrix may show that he is particularly interested in the camera function.

[0102] Construct a generalized composite tag set: On this basis, the system combines the demand tags and interest tags to generate a comprehensive tag set. For example, the generalized composite tags of user A may include: "high cost performance mobile phone", "strong camera performance", "smartphone", etc.

[0103] S2. Calculate the relevance of target users:

[0104] Construct a target user vector group: The platform constructs a target user vector for each user who has interacted under a specific product. For example, users A, B, and C have all browsed and purchased "smartphones with high cost performance", and the platform constructs a target user group vector containing users A, B, and C.

[0105] Calculate the attention of target users: Based on the user's tag set and interest tags, the platform calculates the attention of each user. For example, the attention of users A and B is relatively high, indicating that they are very interested in the target product, while the attention of user C is relatively low, and he may just browse the product by accident.

[0106] Calculate the target user association function: The platform compares the target user vector with the product feature tags to calculate the relevance. For example, the relevance between the vector of user A and the tag of "high cost performance mobile phone" is relatively high, so the platform determines that he has a greater attention to this type of product.

[0107] S3. Calculate the demand budget of the target user group:

[0108] Convert to a target user group vector group: The platform performs time smoothing processing on the target user group vector. For example, if user A has continuously shown a high level of attention to a certain mobile phone in the past week, then his target user vector will be given a higher weight.

[0109] Calculating the relevance of the target user group: The platform sums up the demands of the target user group to calculate the overall demand intensity of the target user group. For example, if users A, B, and C have a high degree of attention, then their combined demand intensity will affect the promotion strategy of the product.

[0110] S4. Interest guidance response:

[0111] Primary guidance response: When the proportion of the interest intensity of the target user group reaches a preset first threshold, the system will recommend other similar products to the users. For example, if users A and B have a high degree of attention to "high-cost performance mobile phones", the platform will recommend other high-cost performance products of the same type to them, rather than only recommending the same model of mobile phone.

[0112] Secondary guidance response: When the proportion of the interest intensity of the target user group reaches the second threshold, the system will provide users with more choices and guide them to compare different products. For example, users A and B start to compare mobile phones of different brands, and the platform provides some detailed comparison data (such as cost performance, user reviews, photo-taking effects, etc.) to help users make decisions.

[0113] Tertiary guidance response: When the proportion of the interest intensity of the target user group reaches the third threshold, the platform will directly push the target product or service. For example, if user A has browsed a certain mobile phone many times and added it to the shopping cart, the platform will directly push the limited-time discount information of this mobile phone to encourage the user to place an order.

[0114] Market monitoring feedback and dynamic adjustment:

[0115] When the system detects that the positive feedback (such as positive reviews, repost volume, etc.) of a certain product remains high, the platform will dynamically adjust the price of the product. For example, if the sales volume of a certain mobile phone continues to rise, the platform may increase the price to increase profits, or conduct promotional activities when there is an overstock of inventory.

[0116] Example of actual application scenario:

[0117] User A: 30 years old, male, has recently browsed multiple smartphones and is particularly concerned about the photo-taking performance. The system constructs tags such as "strong photo-taking", "smartphone", and "high cost performance" for him. Through deep learning algorithms, the platform identifies that user A has a high degree of attention to mobile phones and pushes multiple high-cost performance smartphones with strong photo-taking performance.

[0118] User B: 25 years old, female, has recently browsed multiple smartphones and noticed the discount information of low-price mobile phones. The platform constructs tags such as "price-sensitive" and "smartphone" based on her behavioral data and pushes promotional information suitable for her needs, such as discounted smartphones.

[0119] Through such intelligent guidance, the e-commerce platform can more accurately predict and guide user behavior, improve the exposure rate and sales conversion rate of goods, thereby enhancing the profitability of the platform.

[0120] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than limiting it; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that the present invention can still be modified or equivalently replaced; and all modifications or equivalent replacements of technical solutions that do not depart from the essence and scope of the technical solutions of the present invention should also be covered by the scope of the claims of the present invention.

Claims

1. An e-commerce platform intelligent guidance method based on deep learning multi-dimensional user portraits, characterized in that: The following steps are involved: S1. Label construction: Use deep learning methods to construct a label set; specifically including: S101, collecting user data, and dividing the user data into subjective data and objective data; S102. Construct an initial demand label set based on subjective data ; Wherein, the subscript i is the user group number; S103: Initial requirement tag set Substitute into the matrix to construct the initial demand label matrix ; Wherein, the subscript i is the user group number; S104. Constructing interest tag sets based on objective data ; S105: Collect interest tags Substitute into the matrix to construct the interest tag matrix ; S106. Based on the deep learning method, the initial demand label matrix and interest tag matrix Perform cyclic training to obtain the convergence requirement label matrix and convergent interest label matrix ; S107, label matrix according to convergence requirements and convergent interest label matrix , construct a generalized composite tag set ; S2. Calculate the relevance of target users: For the target user group, construct a target user vector group and a feature label vector, and perform weighted summation to obtain a target user association function as the target user's attention value for the e-commerce platform; specifically, it includes: S201: Construct a target user vector group for the target user group ,in, Indicates Target user vector for user groups; S202: construct a target user tag set based on the target user vector, and construct a feature tag vector for each target user vector , and assign weight parameters to obtain the target user tag set ; S203, calculate the target user association function; group the target user vector and feature label vector Find the angle, calculate the user relevance value, perform weighted summation on the user relevance values ​​of the same user group, and obtain the target user association function ; S3. Calculate the target user group demand budget: Calculate the group user association and construct the group user association function. Combined with the target user association function, calculate the interest intensity ratio of the target user group; specifically include: S301, according to the same target time, for the target user group, the target user vector group Convert to target user group vector group ,satisfy: ； in, For The time span over which the smoothing function is performed, is the smoothing function, Indicates that the target user vector Smoothing is performed within the time span T, using a time window sliding algorithm; Indicates that attention will be paid As a multiplier to multiply the smoothed target user vector , get the target user group vector ; S302, calculate the relevance of the target user group, and group the target user group vectors Add and get the target user group vector ; S303, target user group vector and feature label vector Calculate the relevance and obtain the target user group association function ; S304, target user association function And the target user group association function Determine the ratio value and obtain the interest intensity ratio of the target user group, satisfying: ； in, Indicates the interest intensity ratio of the target user group, that is, the interest intensity ratio of the target user group to the interest intensity ratio of the total user group; S4. Interest-guided response: Provide guidance to target users based on the interest intensity ratio of the target user group.

2. According to claim 1, a method for intelligent guidance of an e-commerce platform based on deep learning of multi-dimensional user portraits is characterized by: The deep learning method in S106 is implemented by a gradient descent algorithm.

3. According to claim 2, a method for intelligent guidance of an e-commerce platform based on deep learning of multi-dimensional user portraits is characterized by: The target user in S2 is a user who is interested in the same specific product or service. The target user's attention to the target product or service is determined by a target user association function. The target user association function satisfies: ； in, Indicates the target user group The attention of each user group to the target product. and Respectively represent of user groups The convergent demand label matrix value and the convergent interest label matrix value of each user; Indicates The number of users in user groups; where Indicates that the convergence requirement label matrix value and the convergence interest label matrix value are multiplied as the value of the continuous addition operation.

4. According to claim 3, the e-commerce platform intelligent guidance method based on deep learning multi-dimensional user portraits is characterized by: The S4 includes: When the interest intensity ratio of the target user group reaches the preset first threshold, a first-level guidance response is made; When the interest intensity ratio of the target user group reaches the preset second threshold, a secondary guidance response is made; When the interest intensity ratio of the target user group reaches the preset third threshold, a third-level guidance response is made; The third threshold is greater than the second threshold, and the second threshold is greater than the first threshold.

5. According to claim 4, the method for intelligent guidance of an e-commerce platform based on deep learning of multi-dimensional user portraits is characterized by: In the three-level guidance response, targeted guidance is provided to individual users, and target products or services are directly pushed.

6. The e-commerce platform intelligent guidance method of deep learning multi-dimensional user portrait according to claim 5 is characterized in that: In the secondary guidance response, the guidance response module performs comprehensive guidance on the target user, guides the individual user to compare a variety of similar products, and guides the individual user to complete the purchase.

7. The e-commerce platform intelligent guidance method of deep learning multi-dimensional user portrait according to claim 6 is characterized by: In the first-level guidance response, category guidance is performed for target users, and the original other product tags of individual users are weighted to guide individual users to switch to other products.

Images