Independent station advertisement display method and device, equipment and medium

Abstract

This application relates to a method, apparatus, and medium for displaying advertisements on an independent website in the field of e-commerce technology. The method includes: responding to a target page loading event of an independent website store and obtaining a user interaction sequence corresponding to the event; using a preset interest evaluation model to determine the user's interest rating for each product in the independent website store based on the user interaction sequence; for each product whose interest rating meets preset conditions, using a preset conversion evaluation model to determine the corresponding conversion rate based on the product profile, user profile, and user interaction sequence; and selecting the product with the highest conversion rate as the advertised product and constructing a target advertisement corresponding to the advertised product for display on the target page. This application enables the push of product advertisements that attract users' attention and is expected to stimulate users to interact with the product for conversion.

Description

Technical Field

[0001] This application relates to the field of e-commerce technology, and in particular to a method for displaying advertisements on an independent website and the corresponding apparatus, computer equipment, and computer-readable storage medium. Background Technology

[0002] With the booming development of internet commerce, independent websites (i.e., e-commerce websites built and operated independently by merchants) have become an important channel for brand operation and product sales. Currently, the product recommendation methods commonly used in independent website stores are relatively static and fixed. For example, a typical approach is based on association rule analysis (such as "product A and product B can be purchased together, for example, a dress and a belt"), directly displaying pre-analyzed matching or related products in fixed recommendation positions on the page. Another common approach is to sort products based on global indicators such as historical sales volume, user reviews, or listing time, and display the top-ranked products as recommended content to all visitors.

[0003] The effectiveness of existing recommendation methods is primarily due to two factors: First, they fail to deeply consider the real-time intent and changing interests of users during a single visit, resulting in insufficient accuracy and dynamic responsiveness. Second, while selected recommended products have designated display areas on the page, their presentation is essentially no different from ordinary product displays, leading to weak user appeal and a low probability of them actually being noticed. In other words, due to the lack of responsiveness to dynamic user intent and more attractive presentation methods, the actual exposure of these recommended products is unsatisfactory, making it difficult to effectively stimulate user consumption. Therefore, overcoming the operational limitations of existing independent website stores is a technical problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0004] The primary objective of this application is to solve at least one of the aforementioned problems by providing a method for displaying advertisements on an independent website, and the corresponding apparatus, computer equipment, and computer-readable storage medium.

[0005] To achieve the various objectives of this application, the following technical solution is adopted: A method for displaying advertisements on an independent website, provided for one of the purposes of this application, includes the following steps: Respond to the target page load event of the independent website store and obtain the corresponding user interaction sequence for that event; A preset interest assessment model is used to determine the user's interest rating for each product in the independent website store based on the user interaction sequence; For each product whose interest score meets the preset conditions, a preset conversion evaluation model is used to determine the corresponding conversion rate based on the product profile, the user profile, and the user interaction sequence. The product with the highest conversion rate is selected as the advertising product, and a corresponding target ad is constructed and displayed on the target page.

[0006] On the other hand, an independent website advertising display device provided to meet one of the purposes of this application includes an event response module, an interest evaluation module, a conversion evaluation module, and an advertising display module. The event response module is used to respond to a target page loading event of an independent website store and obtain the corresponding user interaction sequence. The interest evaluation module is used to determine the user's interest rating for each product in the independent website store based on the user interaction sequence using a preset interest evaluation model. The conversion evaluation module is used to determine the corresponding conversion rate for each product whose interest rating meets preset conditions, based on the product's profile, the user's profile, and the user interaction sequence using a preset conversion evaluation model. The advertising display module is used to select the product with the highest conversion rate as the advertising product and construct a target advertisement corresponding to the advertising product for display on the target page.

[0007] On another front, a computer device provided for one of the purposes of this application includes a central processing unit and a memory, the central processing unit being used to invoke and run a computer program stored in the memory to perform the steps of the independent website advertising display method described in this application.

[0008] In another aspect, a computer program product provided for another purpose of this application includes a computer program / instructions that, when executed by a processor, implement the steps of the method described in any embodiment of this application. Attached Figure Description

[0009] The above and / or additional aspects and advantages of this application will become apparent and readily understood from the following description of the embodiments taken in conjunction with the accompanying drawings, wherein: Figure 1 The network architecture of the e-commerce platform exemplified in this application; Figure 2 This is a flowchart illustrating a typical embodiment of the independent website advertising display method of this application; Figure 3 This is a schematic diagram of the independent website advertising display device of this application; Figure 4 This is a schematic diagram of the structure of a computer device used in this application. Detailed Implementation

[0010] The following describes in detail Embodiment 1 of this application. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain this application, and should not be construed as limiting this application.

[0011] Those skilled in the art will understand that, unless specifically stated otherwise, the singular forms “a,” “an,” “the,” and “the” used herein may also include the plural forms. It should be further understood that the term “comprising” as used in this application means the presence of the stated features, integers, steps, operations, elements, and / or components, but does not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and / or groups thereof. It should be understood that when we say an element is “connected” or “coupled” to another element, it can be directly connected or coupled to the other element, or there may be intermediate elements. Furthermore, “connected” or “coupled” as used herein can include wireless connections or wireless coupling. The term “and / or” as used herein includes all or any units and all combinations of one or more associated listed items.

[0012] Those skilled in the art will understand that, unless otherwise defined, all terms used herein (including technical and scientific terms) have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains. It should also be understood that terms such as those defined in general dictionaries should be understood to have the same meaning as in the context of the prior art, and should not be interpreted in an idealized or overly formal sense unless specifically defined as herein.

[0013] like Figure 1 In the network architecture shown, the e-commerce platform 82 is deployed on the Internet to provide corresponding services to its users. Similarly, the devices 80 of the merchant users and the devices 81 of the consumer users of the e-commerce platform 82 are also connected to the Internet to use the services provided by the e-commerce platform.

[0014] An exemplary e-commerce platform 82 provides supply and demand matching of products and / or services to the general public through the Internet infrastructure. In e-commerce platform 82, products and / or services are provided as commodity information. For the sake of simplicity, the concepts of commodity and product are used in this application to refer to the products and / or services in e-commerce platform 82. Specifically, these may be physical products, digital products, tickets, service subscriptions, other offline services, etc.

[0015] In reality, various entities can access e-commerce platform 82 as users and utilize its online services to participate in the business activities facilitated by the platform. These entities can be natural persons, legal persons, or social organizations. Corresponding to the two types of entities in business activities—merchants and consumers—e-commerce platform 82 has two corresponding categories of users: merchant users and consumer users. Entities involved in the product distribution chain in business activities, including manufacturers, sellers, retailers, and logistics providers, can all use online services on e-commerce platform 82 as merchant users. Similarly, consumers in business activities, including actual or potential consumers, can use online services on e-commerce platform 82 as consumer users. In actual business activities, the same entity can operate as both a merchant user and a consumer user; this should be interpreted flexibly.

[0016] The infrastructure used to deploy the e-commerce platform 82 mainly includes the backend architecture and frontend devices. The backend architecture runs various online services through a service cluster, including middleware or frontend services for the platform, services for consumers, and services for merchants, to enrich and improve its service functions. The frontend devices mainly cover the terminal devices used by users as clients to access the e-commerce platform 82, including but not limited to various mobile terminals, personal computers, and point-of-sale devices. For example, merchant users can use their terminal device 80 to enter product information for their online stores or use the interfaces opened by the e-commerce platform to generate their product information; consumer users can use their terminal device 81 to access the webpage of the online store implemented by the e-commerce platform 82, trigger the shopping process by clicking the shopping button provided on the webpage, and call various online services provided by the e-commerce platform 82 during the shopping process to achieve the purpose of placing an order.

[0017] In some embodiments, the e-commerce platform 82 may be implemented via a processing facility including a processor and memory, which stores a set of instructions that, when executed, cause the e-commerce platform 82 to perform the e-commerce and support functions as described in this application. The processing facility may be part of a server, client, network infrastructure, mobile computing platform, cloud computing platform, fixed computing platform, or other computing platform, and may provide electronic components, merchant devices, payment gateways, application developers, marketing channels, transportation providers, customer devices, point-of-sale devices, etc., for the e-commerce platform 82.

[0018] E-commerce platform 82 can provide online services such as cloud computing services, Software as a Service (SaaS), Infrastructure as a Service (IaaS), Platform as a Service (PaaS), Desktop as a Service (DaaS), Hosted Software as a Service, Mobile Backend as a Service (MBaaS), and Information Technology Management as a Service (ITMaaS). In some embodiments, the various functional components of e-commerce platform 82 can be implemented to operate on various platforms and operating systems. For example, for an online store, its administrator user enjoys the same or similar functions regardless of whether it is on iOS, Android, HomonyOS, or a web page.

[0019] E-commerce platform 82 enables merchants to create their own independent websites to run their online stores. It provides merchants with corresponding business management engine instances, allowing them to establish, maintain, and operate one or more online stores across these independent websites. The business management engine instance can be used for content management, task automation, and data management for one or more online stores. It can be configured through interfaces or built-in components to support various specific business processes in the online store, supporting business activities. Independent websites are the infrastructure of e-commerce platform 82, which offers cross-border services. Merchants can maintain their online stores relatively independently and centrally based on these independent websites. Independent websites typically have dedicated domain names and storage space, and different independent websites are relatively independent. E-commerce platform 82 can provide standardized or customized technical support for a large number of independent websites, allowing merchants to customize a business management engine instance that suits their needs and use it to maintain one or more online stores.

[0020] Online stores can be configured and maintained in the backend by merchant users logging into their Business Management Engine instance as administrators. Supported by the various online services provided by the e-commerce platform 82's infrastructure, merchant users can configure various functions within their online stores and view various data as administrators. For example, merchant users can manage various aspects of their online stores, such as viewing recent online store activities, updating the online store's product catalog, managing orders, recent visit activity, and total order activity. Merchant users can also view more detailed information about their business and visitors to their online store by obtaining reports or metrics, such as displaying a sales summary of the merchant's overall business, specific sales and engagement data from promotional sales and marketing channels, etc.

[0021] E-commerce platforms 82 can provide communication facilities and associated merchant interfaces for electronic communication and marketing. For example, they can utilize electronic messaging aggregation facilities to collect and analyze communication interactions between merchants, consumers, merchant devices, customer devices, point-of-sale devices, etc., aggregating and analyzing communications to increase the potential for product sales. For instance, a consumer may have product-related questions, which could lead to a dialogue between the consumer and the merchant (or an automated processor-based agent representing the merchant), where the communication facilities handle the interaction and provide the merchant with analysis on how to increase the probability of a sale.

[0022] In some embodiments, applications suitable for installation on devices can be provided to serve the access needs of different users, enabling various users to access the e-commerce platform 82 by running the application on their terminal devices. Examples include the merchant backend module of online stores within the e-commerce platform 82. During the process of conducting business activities through these functions, the e-commerce platform 82 can implement various functions related to business activities as middleware or online services and expose corresponding interfaces. Then, toolkits corresponding to the interface access functions are embedded into the application to achieve functional expansion and task completion. The business management engine can include a series of basic functions and expose these functions to online services and / or applications via APIs. Online services and applications use the corresponding functions by remotely calling the corresponding APIs.

[0023] With the support of various components of the Business Management Engine instance, E-commerce Platform 82 can provide online shopping functionality, enabling merchants to connect with customers in a flexible and transparent manner. Consumers can select items online, create orders, provide delivery addresses in the orders, and complete payment confirmation. Merchants can then review and complete or cancel orders.

[0024] The independent website advertising display method of this application can be programmed into a computer program product and deployed on a client or server to run. For example, in the exemplary application scenario of this application, it can be deployed on the server of an e-commerce customer service platform. In this way, the method can be executed by human-computer interaction with the process of the computer program product through a graphical user interface by accessing the interface opened after the computer program product runs.

[0025] Please see Figure 2 The independent website advertising display method of this application, in its typical embodiment, includes the following steps: Step S1100: Respond to the target page loading event of the independent website store and obtain the user interaction sequence corresponding to the event; When a user visits a target page on an independent website, such as the store homepage or product order page (those skilled in the art can also set the target page as needed to display advertisements), the loading behavior of that page triggers a target page loading event. In response to this event, historical interaction data generated by the user's interactions with various products since entering the independent website store is extracted from the user's behavior log in the current session. This historical interaction data is organized in chronological order of the interactions, forming a user interaction sequence. This sequence completely records the user's behavioral trajectory during this visit, such as the user clicking on product A, adding product B to their favorites, and adding product C to their shopping cart. The historical interaction data corresponding to each interaction contains information in multiple dimensions, specifically including any of the following: behavior type weight, interactive product image, interactive product text, interactive product price, interaction freshness, and listing freshness. Behavior type weight is used to quantify the differences in user interest intensity reflected by different interactive behaviors; for example, the weight of browsing behavior is lower than the weight of adding to the shopping cart behavior. Interactive product image refers to the image data of the product that the user interacted with. Interactive product text includes text information such as the product's title, category, and attribute description. The price of an interactive product is its selling price. Interaction freshness refers to the time interval between the occurrence of the interaction and the current moment, used to measure the timeliness of the interest reflected by the interaction. Listing freshness refers to the time interval between the product's listing date and the current moment, used to assess the impact of the product's newness or staleness on user interest.

[0026] After acquiring this historical interaction data, the interaction features corresponding to each interaction are extracted. These interaction features include image feature vectors obtained by vectorizing the interactive product image, text semantic vectors obtained by semantically encoding the interactive product text, and various feature vectors obtained by embedding the behavior type weight, interactive product price, interaction freshness, and listing freshness. All interaction features are combined in chronological order to construct a complete user interaction sequence, which will serve as input data for subsequent model inference.

[0027] Step S1200: Using a preset interest assessment model, determine the corresponding user's interest rating for each product in the independent website store based on the user interaction sequence; After constructing the user interaction sequence, the sequence is input into a pre-trained interest evaluation model that has reached convergence. This model performs multi-level deep modeling on the user interaction sequence and ultimately outputs an interest rating distribution covering all products in the independent website store, where each rating corresponds to the user's real-time interest level in a specific product.

[0028] In one embodiment, the interest evaluation model includes a first sub-network, a second sub-network, and a multi-layer fully connected sub-network. The first sub-network can employ an MLP (Multilayer Perceptron) or at least one self-attention layer followed by an MLP. The second sub-network can be implemented using a transformer. The multi-layer fully connected sub-network includes multiple fully connected layers. The interest evaluation model takes a user interaction sequence as input. For each user-product interaction in the sequence, the first sub-network models these interaction features to obtain single-interaction user interaction features corresponding to each interaction. These single-interaction features are arranged in the corresponding interaction order to form a first target sequence, achieving accurate capture of the cross-relationships between features within a single interaction, such as strong relationships between category and click, price and purchase. Furthermore, the second sub-network in the model performs feature interactions between each single user interaction feature in the first target sequence and other single user features, capturing the temporal relationships and long-term dependencies between interaction sequences to obtain the second target sequence. This second target sequence is input into a multi-layer fully connected network and, through layer-by-layer nonlinear transformations, is abstracted and compressed into a fixed-dimensional vector. This vector is the vectorized representation of the product that the user is most likely to be interested in at the current moment, as predicted by the model; it can be called the target interest product vector. This target interest product vector does not directly correspond to any specific product in the store; it represents an idealized interest direction inferred by the model. To associate it with actual products in the store, the model pre-defines a product vector vocabulary, which contains a unique vectorized representation of each product in the store. Subsequently, the similarity between the target interest product vector and each product vector in the product vector vocabulary is calculated. This similarity calculation can use methods such as vector dot product or cosine similarity. Each calculated similarity value is the user's interest rating for the corresponding product. After completing the similarity calculation for all products in the vocabulary, a set of interest ratings covering all products is obtained; this set of ratings constitutes the interest evaluation distribution. This distribution quantifies a user's real-time interest in each product within an independent website store in the current session context.

[0029] Step S1300: For each product whose interest score meets the preset conditions, a preset conversion evaluation model is used to determine the corresponding conversion rate based on the product profile, the user profile, and the user interaction sequence. After obtaining user interest ratings for all products in the store, products whose interest ratings meet preset criteria are selected, such as the top dozens or hundreds of products in the interest rating ranking, as a set of high-potential products to be evaluated. For each product in this set, a conversion evaluation model pre-trained to a convergent state is used. Combining the product profile, the user profile, and the aforementioned user interaction sequence, the conversion rate of each product in the current context is predicted. The product profile includes static and statistical information such as product category, brand, price range, historical sales data, and attribute tags. The user profile includes relatively stable attribute information such as age, gender, geographical location, historical spending level, and long-term preferred categories.

[0030] The conversion evaluation model predicts conversion rates through the collaborative work of multiple sub-networks. First, the embedding encoding sub-network encodes and transforms the three types of input data. This sub-network includes a text encoder and a sequence encoder. The text encoder uses a pre-trained language model to encode product profiles and user profiles into fixed-dimensional product profile vectors and user profile vectors, respectively. The sequence encoder uses a bidirectional long short-term memory network to perform temporal modeling of user interaction sequences, outputting a sequence vector that captures behavioral context information. Next, the feature cross-network performs deep interactive fusion of these vectors. This sub-network employs a multi-head cross-attention mechanism. On one hand, it uses the product profile vector as the query and the sequence vector as the key and value, generating a product sequence vector through attention calculation. This vector integrates the inherent attributes of the product with the user's attention information regarding product features in real-time behavior. On the other hand, it uses the user profile vector as the query and the sequence vector as the key and value, generating a user sequence vector that integrates the user's static attributes and dynamic behavioral patterns. Finally, the multi-layer fully connected sub-network concatenates the product sequence vector with the user sequence vector to form a basic hidden vector, which is then input into a deep neural network composed of multiple fully connected layers. After layer-by-layer nonlinear transformation, the last layer uses the Sigmoid activation function to compress the output value to between 0 and 1. This output value is the conversion rate of the product predicted by the model in the current user session context.

[0031] Step S1400: Select the product with the highest conversion rate as the advertising product, and construct the target advertisement corresponding to the advertising product to be displayed on the target page.

[0032] For each product whose conversion rate is predicted by the aforementioned conversion evaluation model, the product with the highest conversion rate is selected as the advertising product to be displayed. After determining the advertising product, the corresponding target advertisement is constructed and rendered on the user's currently loaded target page so that it can be displayed together with the target page.

[0033] The process of constructing a targeted advertisement involves the dynamic combination of multiple visual materials for the advertised product. Specifically, it involves acquiring pre-prepared selling point text images and target product images. Selling point text images are graphically designed text images that edit the product's core reasons or promotional information onto a separate image layer using specific fonts, colors, and layouts. Target product images are typically transparent images of the product with the background removed. Simultaneously, it involves retrieving ad template images and their corresponding completion composition information from a pre-built ad template library that match the user's profile. The ad template library is a pre-built database containing multiple different ad template images. Each template image is a base image with a complete background and overall visual style, with specific areas reserved for filling. Completion composition information is structured data corresponding to each ad template image, precisely describing the placement and size of the selling point text images and target product images within that template image. To achieve matching between template images and user profiles, a pre-trained, converged image-text matching model is invoked to calculate the semantic matching score between the user profile and each ad template image in the ad template library. The ad template image with the highest score and its completed composition information are selected. Subsequently, following the instructions of the completed composition information, the selling point copy image and the target product image are scaled and overlaid onto the corresponding area of ​​the ad template image to synthesize a complete final ad image. This image is then packaged into a target ad displayed on the target page. This ad visually presents the synthesized image and is interactive; users can trigger subsequent events by clicking on the ad. This completes the entire process from real-time user intent recognition and conversion potential prediction to personalized ad generation and display.

[0034] As can be seen from the typical embodiments of this application, the technical solution of this application has many advantages, including but not limited to the following aspects: This application obtains the corresponding user interaction sequence during the loading process of the target page of an independent website store. First, it uses a pre-set interest assessment model to quantify the user's interest in each product within the store, achieving accurate identification and dynamic response to the user's real-time intent. Based on this, it further uses a conversion assessment model to predict the conversion rate of products that users are most interested in, selecting the products with the highest conversion potential as advertising products. It is evident that this process, through the progressive application of the interest assessment model and the conversion assessment model, constructs a complete decision-making link from interest identification to conversion prediction. This not only ensures the high relevance of advertising products to user interests but also considers their potential for commercial conversion, making each ad display an effective catalyst for driving users to complete a purchase. Furthermore, this process is entirely based on real-time user interaction data during the current visit, dynamically calculating and adjusting to ensure that the advertising content updates in real time according to changes in user interests. This achieves dynamic synchronization between ad display and user needs, providing strong support for the intelligent operation and maximization of traffic value of independent website stores.

[0035] Furthermore, displaying advertised products that have undergone dual screening based on interest and conversion rates as ads on the target page offers greater visual appeal and information delivery efficiency compared to conventional product displays, effectively increasing user attention and click-through rates. Since the advertised products have already been optimized based on both interest and conversion dimensions, their relevance to user needs is high. Combined with a prominent ad format, this creates a synergistic effect between precise content and eye-catching presentation, significantly enhancing the actual exposure and user reach of the products.

[0036] In a further embodiment, step S1100, obtaining the user interaction sequence corresponding to the event, includes the following steps: Step S1110: Obtain the historical interaction data generated by the user corresponding to the target page loading event interacting with the products in the independent website store multiple times. The historical interaction data includes any multiple of the following: behavior type weight, interactive product image, interactive product text, interactive product price, interaction freshness, and listing freshness. To achieve dynamic adaptation of advertising content, decisions should be based on real-time user behavior during the current visit, rather than solely relying on historical statistical data. Specifically, when a user visits a target page on an independent website, the page load event is monitored and responded to. At this point, historical interaction data generated from the user's interactions with products within the store during this session can be extracted from the user's behavior logs. This historical interaction data can include any number of factors such as behavior type weight, interacting product category, interacting product image, interacting product text, interacting product price, interaction freshness, and listing freshness.

[0037] Behavior type weights are used to characterize the differences in the contribution of different interactive behaviors to reflecting user interests. For example, the weight of a click behavior can be set to 0.3, the weight of adding to cart behavior can be set to 0.8, and the weight of completing a purchase behavior can be set to 1.0. Specific weight values ​​can be configured based on business experience or quantified through differentiated learning of different behaviors using machine learning models. Interactive product images refer to images of the products that users interact with, such as the main product image or other images showcasing the product. Interactive product text includes any textual information describing the product, such as its category, title, attributes, and selling points. Interactive product price records the price of the product interacted with. Interaction freshness refers to the time interval between the occurrence of the user's interaction and the current moment; this time interval can be used as a reference to measure the timeliness of the interest reflected by the behavior. Listing freshness refers to the time interval between the listing time of the interacted product and the current moment; this information can be used to assess the impact of the product's newness or aging on user interest.

[0038] The above historical interaction data can be obtained by real-time reporting through front-end event tracking technology or by reading from the back-end log system. Data preparation is usually completed during page loading to support real-time calculation of subsequent models.

[0039] Step S1120: Extract the interaction features corresponding to each historical interaction data to construct a user interaction sequence.

[0040] For each interaction, a subsequence of interaction features can be extracted from the historical interaction data. This subsequence consists of features from various dimensions involved in that interaction. For example, for a single click, the subsequence of interaction features may include the behavior type vector, product category embedding vector, product image feature vector, product text semantic vector, product price encoding vector, interaction freshness vector, and listing freshness vector. The internal order of each feature item in these subsequences can be set by those skilled in the art according to model input requirements and business characteristics. After extracting the subsequence of interaction features for each interaction, these subsequences can be arranged sequentially according to the chronological order of the interactions, thus forming a complete user interaction sequence. The construction of this user interaction sequence provides temporal context information for subsequent interest evaluation models, enabling the model to capture the dynamic patterns of user interest evolution over time. For example, if the sequence first shows the subsequence of clicking on a mobile phone and then shows the subsequence of browsing phone cases, this sequential relationship may suggest that the user is considering purchasing a mobile phone and related accessories. Through this sequential feature representation, the interest assessment model can learn the evolution of user behavior patterns as interaction sequence unfolds, thereby improving the accuracy of interest prediction. The length of the user interaction sequence can be set to a fixed value or dynamically changed according to actual needs. Sequences that are too short can be padded, and sequences that are too long can be truncated to ensure the standardization of the model input.

[0041] In practical implementation, a pre-trained, converged open-source convolutional neural network model (e.g., ResNet50; those skilled in the art can also use other models suitable for image feature extraction and quantization representation) can be used. After forward inference with the interactive product image as input, the feature vector output from the last fully connected layer is taken as the product image feature vector. Similarly, a pre-trained, converged open-source language model (e.g., BERT; those skilled in the art can also use other models suitable for text feature extraction and quantization representation) can be used. After forward inference with the interactive product text as input, the feature vector output from the last fully connected layer is taken as the product text semantic vector. For behavior type weights and interactive product prices, they can be represented as corresponding behavior type vectors and product price vectors through binned discretization embedding. The corresponding interaction freshness decay value and listing freshness decay value can be calculated using a pre-configured time decay function with interaction freshness and listing freshness as input parameters, and then represented as corresponding interaction freshness vectors and listing freshness vectors through binned discretization embedding. Of course, these vector representations of interactive features are ultimately processed into the same dimension.

[0042] In this embodiment, historical interaction data containing multi-dimensional information such as behavior type weights, interactive product images, interactive product text, interactive product prices, interaction freshness, and listing freshness is acquired, and a structured user interaction sequence is constructed based on this data, providing rich and refined input features for subsequent interest assessment. Introducing behavior type weights quantifies the differences in user interest intensity reflected by different interactive behaviors (such as clicks, adding to cart, and purchasing), enabling the model to more accurately distinguish between users' superficial curiosity and deep purchase intentions. Timeliness features such as interaction freshness and listing freshness are also incorporated, allowing the model to dynamically perceive real-time changes in user interests and the product's lifecycle status, avoiding interference from outdated behaviors in current interest judgments and significantly improving the timeliness and accuracy of interest assessment. Furthermore, the fusion of multi-modal features such as images, text, and prices makes the representation of products and user behavior more comprehensive, laying a solid data foundation for the subsequent model to capture deep-level feature interactions.

[0043] In a further embodiment, step S1200, determining the user's interest rating for each product in the independent website store based on the user interaction sequence using a preset interest evaluation model, includes the following steps: Step S1210: Using the heterogeneous modeling subnetwork in the preset interest evaluation model, perform feature relationship modeling on the user interaction sequence to obtain the first feature sequence; In the heterogeneous modeling sub-network, feature relationships are modeled on user interaction sequences. This sub-network is implemented using the Transformer architecture. Each element in the user interaction sequence corresponds to a vector combination of features generated by the user in a particular interaction, consisting of multiple different dimensions. These feature dimensions may include behavior type embedding, product image features, product text semantic features, product price embedding, etc. Using these vector combinations as the input sequence, the Transformer's self-attention mechanism calculates the attention weight between each element in the sequence and all other elements in the sequence, and fuses the information of other elements into the current element through weighted summation. This calculation process is repeated in each layer of the Transformer, enabling the model to fully capture the deep correlations between different feature dimensions. For example, a click on a high-priced product may be highly correlated with another browsing behavior on a product of a certain category. This complex relationship across feature categories and across interaction sequences is captured and encoded by the self-attention mechanism. After processing by the heterogeneous modeling sub-network, the original input sequence is transformed into a first feature sequence, in which each feature vector has been fused with its relational features after interacting with all other features in the sequence.

[0044] Step S1220: The first feature sequence is modeled using the order modeling subnetwork in the interest evaluation model to obtain the second feature sequence. After completing the feature relationship modeling, the first feature sequence is input into the sequence modeling subnetwork to model the interaction sequence relationship. This subnetwork consists of a feature connection layer and a Transformer layer. The feature connection layer first operates on each feature vector in the first feature sequence, horizontally concatenating multiple vectors belonging to the same user history and product interaction into a single, higher-dimensional vector, which is the single user interaction feature of that interaction. All single user interaction features are organized according to the chronological order of the interactions to form a new sequence. This sequence is input into a Transformer network, where its self-attention mechanism comes into play again, but this time the modeling object is the temporal dependency between each interaction. This Transformer can learn the interest evolution pattern inherent in a series of interactions. The output of the sequence modeling subnetwork is the second feature sequence, where each element (i.e., the single user interaction feature of each interaction) has incorporated its sequence relationship information within the overall interaction sequence context.

[0045] Step S1230: The target inference subnetwork in the interest evaluation model performs inference mapping modeling on the second feature sequence for target interest products to obtain the corresponding interest evaluation distribution, which includes the interest rating of the corresponding user for each product in the independent website store.

[0046] After modeling the interaction order relationship, it is input into the target inference sub-network to model the inference mapping of target interest products. This sub-network consists of a multi-layer fully connected network (which can be a multi-layer perceptron, i.e., MLP). The second feature sequence is first input into the multi-layer fully connected network, and through layer-by-layer nonlinear transformation, it is abstracted and compressed into a fixed-dimensional vector. This vector is the vectorized representation of the product that the model predicts the user is most likely to be interested in at the current moment, which can be called the target interest product vector. This target interest product vector does not directly correspond to any specific product in the store; it represents an idealized interest direction inferred by the model. To associate it with the actual products in the store, the model pre-sets a product vector vocabulary, which contains the unique vectorized representation of each product in the store. Subsequently, the similarity between the target interest product vector and each product vector in the product vector vocabulary is calculated. This similarity calculation can use methods such as vector dot product or cosine similarity. Each calculated similarity value is the user's interest rating for the corresponding product. After completing the similarity calculation for all products in the vocabulary, a set of interest ratings covering all products is obtained. This set of ratings constitutes the interest evaluation distribution. This distribution quantifies a user's real-time interest in each product within an independent website store in the current session context.

[0047] In this embodiment, firstly, a heterogeneous modeling sub-network (such as Transformer) can fully cross-integrate features of different dimensions (such as images, text, and prices) within a single interaction, accurately capturing fine-grained interest signals such as "high-priced items are viewed for a long time" or "items in a certain category are clicked quickly." Secondly, the sequence modeling sub-network organizes the features of a single interaction into a behavioral sequence through feature connections and temporal modeling (such as another Transformer), effectively learning the evolution of user interests with the order of interactions, such as the progressive relationship of intent from "browsing a phone" to "searching for phone cases," thereby dynamically grasping the migration of interests. Finally, the target inference sub-network compresses and maps complex temporal behavioral patterns into a target interest product vector, and performs similarity matching with a pre-set product vector vocabulary, ultimately outputting an interest rating distribution covering all products in the store. This process translates abstract user interests into quantifiable ratings associated with specific products, achieving end-to-end accurate inference from behavioral sequences to the interests of all products in the store, providing a highly reliable basis for subsequent screening of conversion potential.

[0048] In a further embodiment, step S1300, determining the corresponding conversion rate using a preset conversion evaluation model based on the product profile, the user profile, and the user interaction sequence, includes: Step S1310: The embedded encoding subnetwork in the preset conversion evaluation model encodes the product profile, the user profile, and the user interaction sequence to obtain the corresponding product profile vector, user profile vector, and sequence vector. In the conversion evaluation model, the embedding encoding sub-network is responsible for converting the raw input data into high-dimensional dense vectors that can be processed by subsequent deep networks. This sub-network contains two types of encoders: a text encoder and a sequence encoder. The text encoder uses a pre-trained language model based on the Transformer architecture (such as BERT) to process product profiles and user profiles respectively. It extracts deep semantic features through multi-layer self-attention mechanisms and feedforward networks, and finally outputs fixed-dimensional product profile vectors and user profile vectors. The sequence encoder uses a recurrent neural network such as a bidirectional long short-term memory network (Bi-LSTM) to specifically process user interaction sequences, modeling the temporal behavior patterns of users in the current session, and outputting a context-aware sequence vector. The product profile vector, user profile vector, and sequence vector are all mapped to the same vector dimension, but represent their respective semantic features.

[0049] Step S1320: Using the feature cross subnetwork in the conversion evaluation model, the product profile vector and the sequence vector are fused with the product profile vector and the user profile vector respectively to obtain the corresponding product sequence vector and user sequence vector. After obtaining the product profile vector, user profile vector, and sequence vector, the feature cross-network performs deep interaction and fusion of these vectors through a multi-head cross-attention layer. This sub-network contains multiple parallel attention heads that have learned different weight matrices, enabling each head to capture different types of interaction relationships. For the fusion of the product profile vector and the sequence vector, each attention head uses the product profile vector as the query and the sequence vector as the key and value. It calculates the weighted contribution of each position in the sequence to the product features through scaled dot product attention, generating a fusion vector that incorporates behavioral context information. After concatenating the fusion vectors from all heads and performing a linear transformation, the final fusion result, the product sequence vector, is obtained. This vector contains both the inherent attributes of the product itself and the degree of attention given to it by the user's real-time behavior sequence. Similarly, to obtain the user sequence vector, each attention head uses the user profile vector as the query and the sequence vector as the key and value, calculates the interaction relationship between the user's static attributes and dynamic behaviors, generates a corresponding fusion vector, and then obtains the user sequence vector through multi-head fusion. Through this mechanism, the static characteristics of goods and users can be deeply integrated with the real-time behavioral sequences of users, forming two highly correlated comprehensive representations.

[0050] Step S1330: The conversion rate of the product is obtained by performing a conversion mapping on the underlying latent vector obtained by splicing the product sequence vector and the user sequence vector using the multi-layer fully connected sub-network in the conversion evaluation model.

[0051] After obtaining the product sequence vector and user sequence vector, a multi-layer fully connected sub-network is responsible for jointly modeling the two and ultimately outputting the conversion rate. This sub-network first concatenates the product sequence vector and user sequence vector to form a basic latent vector, which fully integrates information about product characteristics, user attributes, and real-time behavioral patterns. Subsequently, this basic latent vector is input into a deep neural network composed of multiple stacked fully connected layers. Each fully connected layer contains linear transformations and non-linear activation functions (such as ReLU) to progressively extract higher-order features and learn complex combinations of features. After several layers of propagation and transformation, the final fully connected layer uses the Sigmoid activation function to compress the network's final output to the range of 0 to 1. This output value is the model's predicted conversion rate for the product. This conversion rate quantifies the probability that a user will convert and interact with the product in the current user and current session context.

[0052] In this embodiment, firstly, the embedding encoding sub-network employs a text encoder (such as a pre-trained language model) and a sequence encoder (such as Bi-LSTM) to uniformly encode heterogeneous product profiles, user profiles, and user interaction sequences into high-dimensional semantic vectors. This preserves the deep semantics of static attributes while capturing the temporal context of dynamic behaviors. Secondly, the feature cross-network introduces a multi-head cross-attention mechanism. Instead of simply concatenating features, it allows deep interaction between product profiles and behavior sequences, and between user profiles and behavior sequences, generating product sequence vectors and user sequence vectors respectively. This design enables the model to accurately learn the probability of a "product with a certain attribute" being converted in the eyes of a "user with a specific behavioral pattern," achieving high-order, non-linear fusion between features. Finally, these two deeply fused vectors are concatenated and input into a multi-layer fully connected network for conversion mapping. This fully utilizes the powerful fitting ability of deep networks to output an accurate conversion rate estimate between 0 and 1. The entire process assigns a conversion probability reflecting the commercial value of each high-potential product, providing a crucial decision dimension for ultimately selecting the most valuable advertising products.

[0053] In a further embodiment, step S1400, constructing the target advertisement corresponding to the advertised product, includes the following steps: Step S1410: Obtain the selling point copywriting images and target product images of the advertised product, and obtain the advertising template image and its completion composition information that match the user's profile from the preset advertising template library; After selecting the highest-converting products from the store's numerous items through the aforementioned steps, the key to successful advertising lies in effectively presenting these products to users. This ensures the products attract attention while efficiently conveying their value. Even carefully selected products may be overlooked by users if displayed only as standard items (e.g., showing only the main image and price). To maximize exposure and appeal, a more visually impactful and efficient presentation is needed. Furthermore, different user groups have varying aesthetic preferences and information focuses. Younger users may favor vibrant and trendy designs, while more mature users may prefer a simpler, more professional approach. Therefore, matching ad templates to user profiles and dynamically adding personalized product information allows for a deep alignment between ad content and user preferences, effectively increasing click-through rates and conversion potential.

[0054] In practice, pre-prepared selling point text images and target product images are acquired. The selling point text images are graphically processed text images; they are not simply text, but rather images that edit the core reasons for the product (such as "limited-time 50% off," "intelligent noise reduction," "pure cotton and skin-friendly") onto a separate image layer using specific fonts, colors, effects (such as outlines and shadows), and layout methods. The purpose is to enhance the visual expressiveness and persuasiveness of the text. The target product image is typically a transparent image of the product with the background removed, allowing for flexible integration with various backgrounds during subsequent compositing and preventing background interference with the presentation of the product. These two types of materials can be pre-uploaded by the merchant when creating the product or automatically generated by an image generation model. For example, a pre-trained, converged open-source multimodal generation model can be used to automatically generate diverse selling point text images based on the product's category, attributes, and pre-set promotional information.

[0055] Simultaneously, a pre-built database of advertising templates is invoked, containing multiple different advertising template images and their corresponding completion information. Each advertising template image is an image with a relatively complete background and overall visual style, presenting the basic visual content of the advertisement. For example, it could be a warm home scene background, a cool technological background, a refreshing natural landscape background, or a simple solid color gradient background. Before being filled with content, these template images already constitute a visually complete image, but some areas are reserved blank or can be covered by other layers. These areas are designed for the subsequent filling of selling point copy images and target product images. The completion information is a structured data description corresponding to each advertising template image, precisely defining the specific location and size of the selling point copy images and target product images within the advertising template image. The completion of the composition information can be achieved in various forms. For example, it can be a JSON data object containing the coordinates of the top left corner, width, and height of the "product image area" and the coordinates of the center point, width, and height of the "text image area". Alternatively, it can be a single-channel mask image of the same size as the advertising template image, where the area with a pixel value of 1 represents the area that the product image should cover, the area with a pixel value of 2 represents the area that the text image should cover, and the other areas are 0.

[0056] Subsequently, to match ad template images with user profiles, a pre-trained, converged image-text matching model, such as the CLIP model, is invoked to calculate the semantic matching score between the ad template image and the user profile. The inference process of the image-text matching model involves encoding the user's profile tags and each ad template image in the ad template library into the same vector space, and then calculating the cosine similarity between the corresponding encoded text vector and image vector. This similarity is the matching score between the ad template image and the user. By calculating the matching scores of all template images in the ad template library and determining the ad template image with the highest score, the corresponding completion composition information of the ad template image that best matches the current user profile is then retrieved from the library.

[0057] Step S1420: According to the completed composition information, the selling point copy image and the target product image are added to the advertising template image to obtain the target advertisement for the advertised product.

[0058] After acquiring the selling point text image, the target product image, and the selected advertising template image and its complete composition information, the image compositing operation is performed according to the instructions in the complete composition information. Taking the use of coordinate region definition for the complete composition information as an example, first, a new canvas with the same size as the advertising template image is created, and the advertising template image is copied onto this canvas as the bottom background. Then, according to the coordinates and size of the "product image area" defined in the complete composition information, the target product image is scaled and cropped as necessary to fit the area perfectly, and placed on top of the original image content in the template image. Similarly, according to the coordinates and size of the "text image area" defined in the complete composition information, the selling point text image is scaled accordingly and placed in the corresponding area. Finally, a complete image is synthesized, including the background atmosphere, the main product, and the prominent selling point text. Furthermore, the advertising component is invoked to encapsulate the image into a target advertisement displayed on the target page. The target advertisement visually displays the image and is interactive. The specific interaction can be triggered by the user clicking on the target advertisement to execute corresponding events.

[0059] In this embodiment, the semantic similarity between user profiles and advertising template images is calculated using an image-text matching model. This allows for the automatic matching of the most suitable advertising background and overall style to user groups with different aesthetic preferences and information needs. For example, a template with a strong promotional feel is matched to users seeking cost-effectiveness, while a high-end and minimalist template is matched to users who value quality. This achieves a personalized advertising visual presentation, effectively enhancing the affinity and attractiveness of the advertisement to the target user. Based on the completed composition information, the selling point copy images and product transparency images are precisely synthesized into the template image. This not only ensures the professionalism and aesthetics of the advertising composition but also conveys the core reasons for the product to the user in the most intuitive and impactful way through graphically designed selling point copy, significantly enhancing the efficiency of advertising information delivery and click-through rate. This dynamic and intelligent advertising synthesis method presents the selected advertising products in the best visual form, achieving dual optimization of content and form, maximizing the exposure effect and conversion potential of the advertisement.

[0060] In a further embodiment, after step S1400, which involves constructing a target advertisement corresponding to the advertised product for display on the target page, the following steps are included: Step S1500: Respond to the user's trigger event on the target advertisement, jump to the product details page of the advertised product, and create a dialogue window between the user and the customer service guide; After successfully constructing and displaying targeted ads that highly match users' real-time interests, the next step is to further shorten the decision-making path from when users notice a product to when they ultimately complete a purchase, providing them with a seamless and efficient conversion channel. While simple ad displays can attract user attention, if users click on an ad and only enter a regular product details page, they may still face information overload or decision-making hesitation, leading to the loss of potential conversions. Therefore, at the crucial moment when users express strong purchase intentions by clicking on an ad, immediate customer service intervention, providing refined shopping guides designed to eliminate purchase concerns, can effectively capture the traffic brought by the ad and convert users' instantaneous interest into actual purchase behavior. This series of operations aims to build a complete conversion chain from "ad exposure" to "instant consultation" to "facilitating decision-making," maximizing the commercial value of advertising.

[0061] To achieve this, when a user browses and notices the displayed target ad on the target page, their click constitutes a trigger event for the target ad. An event listener pre-configured on the target page captures this trigger event in real time. In response to this event, a page redirection is first executed, directing the user's browser window or in-app view from the current target page to the product details page corresponding to the advertised product. This redirection can be achieved by modifying the URL in the browser's address bar or updating the routing state within the application, ensuring that the user can immediately view the product's complete information, specifications, user reviews, and other detailed content.

[0062] Simultaneously with the page redirection, a background service is launched to create a chat window between the user and the customer service representative. This representative can be a human or an intelligent chatbot built using natural language processing technology. The method of creating the chat window depends on the platform architecture: on web pages, a floating chat box component can be dynamically generated using JavaScript code, or it can slide out from the edge of the page as a sidebar; on mobile applications, the application's built-in instant messaging module can be invoked. Regardless of the implementation method, the chat window will proactively appear in a preset location on the user interface when the product details page finishes loading, usually in the lower right corner or bottom of the page, to provide immediate consultation services to the user without completely obscuring product information. The creation of this chat window involves interaction with an instant messaging server, assigning a unique session identifier to the user and the designated customer service representative for subsequent message routing and transmission.

[0063] Step S1510: Generate product shopping guide information to guide the user to make a quick decision. The shopping guide customer service will then publish the product shopping guide information to the chat window for the user to view.

[0064] After the dialogue window is successfully established, the customer service robot or human customer service representative will not passively wait for the user to ask questions, but will proactively generate and push a product recommendation message aimed at guiding the user to make a quick decision. This product recommendation message can be generated by calling a pre-trained multimodal model or large language model that has reached a convergence state. It uses the user's profile, the characteristics of the advertised products, and historical interaction data generated from the user's multiple interactions with products within the independent website store as input for inference. These models can be fine-tuned and trained using historical dialogue flow data generated from human customer service representatives providing product recommendation services to users on the independent website store, constructing a corresponding dataset for fine-tuning until convergence is achieved, thus enabling the generation of product recommendation messages. The dataset construction described here can be flexibly implemented by those skilled in the art based on the input and output of the model's inference and the capabilities the model needs to learn, as revealed here.

[0065] In other embodiments, the specific content format of product guide information can be diverse and not limited to plain text. It can be a rich text message containing a summary of the product's core selling points, such as highlighting key decision-making factors like "limited-time discount," "genuine product guarantee," and "free installation" using bold or colored fonts. It can also be an infographic pre-designed by the operators of an independent website store, integrating information such as the product's core advantages, promotional activities, and user reviews into a single image for clear and concise information delivery. Alternatively, it can be designed as a short video or an animated GIF demonstrating the product's usage scenarios or core functions.

[0066] After generating the aforementioned product recommendation information, the information is sent to the user as a message through the established chat window, posing as a customer service representative. The user will see this proactively pushed message in their chat with the customer service representative.

[0067] In this embodiment, by first responding to the user's click event and immediately executing page redirection and dialog window creation in parallel, the user's decision-making path is greatly shortened. While users enter the product details page to obtain in-depth product information, they can receive immediate customer service access without any additional operation, effectively capturing the high-intent traffic brought by advertising and reducing the risk of churn due to insufficient information or decision-making doubts. Secondly, customer service (human or AI) proactively pushes refined product guidance information, rather than passively waiting for user questions. This proactive service model focuses on the core advantages of the product, accurately answers potential user doubts, and guides users to make quick purchase decisions. This guidance information can be dynamically generated based on user profiles and real-time behavior, ensuring its relevance and persuasiveness. It is evident that deeply binding advertising display, information acquisition, and instant consultation forms a proactive, efficient, and coherent user service loop, greatly improving the efficiency from ad click to final conversion and fully leveraging the commercial value of advertising.

[0068] In a further embodiment, before step S1200, which uses a preset interest evaluation model to determine the corresponding user's interest rating for each product in the independent website store based on the user interaction sequence, the following steps are included: Step S2200: Obtain the dataset, which includes multiple samples and their supervision labels. The samples and their supervision labels correspond to sample interaction sequences and their corresponding target interest evaluation distributions. To train the interest assessment model, a dedicated dataset for model learning is first required. This dataset is built upon a massive amount of historical user interaction data collected from the log system of independent online stores, thereby constructing the corresponding original interaction sequences. The method for constructing these sequences is consistent with the method used to obtain user interaction sequences during the online inference phase, ensuring consistency in data distribution for model training and use.

[0069] After acquiring a large number of raw interaction sequences, a corresponding supervised label needs to be constructed for each raw interaction sequence, i.e., the target interest evaluation distribution corresponding to that sample. The target interest evaluation distribution is a vector with the same dimension as the total number of products in the independent website store. Each component in the vector represents the user's true interest preference for the next interactive product in the sequence, given the context of the sample's interaction sequence. In one embodiment of constructing this label, the last interactive product in the sequence is taken as a positive sample, and its label value at the corresponding position in the vector is set to 1, while the label values ​​at the corresponding positions of all other products are set to 0, thus forming a hard label distribution in the form of one-hot encoding. In other embodiments, the diversity of user interests can also be considered. Not only is the last interactive product regarded as a positive sample, but other products belonging to the same category as the last interactive product, having similar attributes, or frequently purchased together are also given a certain positive label weight, forming a soft label distribution. For example, the similarity between other products and the last interactive product can be calculated, and the similarity can be normalized and used as the label value. In practice, the original interaction sequence is typically processed using a sliding window. For example, a fixed-length window is set, with the first part of the window used as the sample interaction sequence for model input, and the item immediately following the window used as the prediction target to construct the supervision label. After constructing all samples, the dataset is randomly divided into a training set and a validation set according to a preset ratio, such as 7:3. The training set is used to drive the model to learn parameters, while the validation set is used to evaluate model performance and adjust hyperparameters during training.

[0070] Step S2210: Train the interest evaluation model to convergence based on the dataset, and learn the ability to infer the corresponding interest evaluation distribution based on any user interaction sequence.

[0071] The constructed training set is input into the interest evaluation model, which has not yet converged. The interest evaluation model adopts a multi-sub-network cascade design. The input layer of the model receives a sample interaction sequence, where each element corresponds to the raw data generated by the user in a certain interaction. This raw data is first converted into a dense vector representation through the embedding layer. For behavior types, they are one-hot encoded and then mapped to behavior embedding vectors through an embedding matrix; for product images, a pre-trained convolutional neural network (e.g., ResNet50) can be used to extract fixed-dimensional image feature vectors; for product text, a pre-trained language model (e.g., BERT) can be used to extract text semantic vectors; for product price and two types of freshness information, they can be first discretized by binning and then mapped to the corresponding price embedding vector, interaction freshness embedding vector, and shelf freshness embedding vector through the embedding layer. After processing by the embedding layer, each interaction corresponds to a set of vectors, which together constitute the feature set at that moment of interaction.

[0072] These feature vectors are then fed into a heterogeneous modeling sub-network. This sub-network employs a Transformer architecture, with its core being a multi-head self-attention mechanism. In this sub-network, all vectors from different interaction times and different feature dimensions are combined into a sequence. The self-attention mechanism calculates the attention weights between each vector in the sequence and all other vectors in the sequence. Through this calculation, the model can learn deep relationships between different types of features. For example, the product image features corresponding to a click might establish high attention weights with the product text features corresponding to another purchase, thus encoding this cross-feature dimension relationship into the first feature sequence of the output. Each output feature vector contains relational information from other features.

[0073] After obtaining the first feature sequence, it is input into the sequence modeling sub-network. This sub-network first contains a feature connection layer, which horizontally concatenates multiple feature vectors belonging to the same interaction to form a single vector that integrates all information from this interaction—the single user interaction feature. Subsequently, the single user interaction features corresponding to all interaction moments are organized into a new sequence according to their original chronological order and input into another Transformer network. This Transformer network focuses on modeling the temporal dependencies between interactive behaviors. Its self-attention mechanism can capture the evolution of user interests as the interaction sequence unfolds; for example, if a user looks at a phone and then immediately browses phone cases, this sequential relationship is encoded by the model. The output of this sub-network is the second feature sequence, where each vector incorporates its sequential contextual information within the overall interaction sequence.

[0074] Next, the second feature sequence is passed to the target inference sub-network. This sub-network first consists of a multi-layer fully connected network, whose function is to compress and map the entire second feature sequence into a fixed-dimensional vector. This vector can be regarded as the vector representation of the idealized product that the user is most likely to be interested in at the current moment, extracted and inferred by the model from the user's historical behavior, i.e., the target interest product vector. Subsequently, the model uses a pre-generated product vector vocabulary, which contains a unique vector representation corresponding to each product in the store. The similarity between the target interest product vector and each product vector in the product vector vocabulary is calculated, for example, by calculating their dot product or cosine similarity. Each calculated similarity value represents the user's predicted level of interest in the corresponding product, i.e., the interest score. After completing the similarity calculation for all products in the vocabulary, an interest score vector covering all products is obtained. This vector is the interest evaluation distribution output by the model for the current input sample.

[0075] The model's output interest evaluation distribution and the pre-constructed supervision labels (target interest evaluation distribution) are input into the loss function to calculate the loss value. The loss function can be the cross-entropy loss function, used to measure the difference between the model's predicted interest distribution and the true label distribution. Through backpropagation, the loss value is passed forward layer by layer from the output layer, and an optimizer (e.g., Adam) updates the parameters of all sub-networks in the entire interest evaluation model based on the calculated gradients. This process is repeated on the training set, processing a batch of samples and updating the parameters once per iteration. After each training epoch, the performance of the current model is evaluated using the validation set, monitoring whether the loss value decreases and whether other evaluation metrics (e.g., recall, hit rate) improve. Hyperparameters such as the learning rate, number of network layers, and number of attention heads are adjusted based on the validation set performance. The training process terminates when the model's performance on the validation set no longer improves or reaches the preset number of training epochs. Ultimately, after multiple rounds of iterative optimization on the training set and hyperparameter tuning on the validation set, the interest assessment model learned the ability to accurately infer users' real-time interest ratings for all products in the store from any user interaction sequence, and can be deployed to process user interaction sequences generated online in real time.

[0076] This embodiment reveals the training process of the interest assessment model, ensuring that the model acquires reliable and accurate reasoning capabilities to serve the real-time user interaction sequences generated online, and quantitatively assess the corresponding user's interest in all products within the independent website store.

[0077] Please see Figure 3This application provides an independent website advertising display device, which is a functional embodiment of the independent website advertising display method of this application. On another note, this independent website advertising display device, also provided to fulfill one of the purposes of this application, includes an event response module 1100, an interest evaluation module 1200, a conversion evaluation module 1300, and an advertising display module 1400. The event response module 1100 is used to respond to the target page loading event of the independent website store and obtain the corresponding user interaction sequence. The interest evaluation module 1200 is used to determine the user's interest rating for each product in the independent website store based on the user interaction sequence using a preset interest evaluation model. The conversion evaluation module 1300 is used to determine the corresponding conversion rate for each product whose interest rating meets preset conditions, based on the product profile, the user profile, and the user interaction sequence using a preset conversion evaluation model. The advertising display module 1400 is used to select the product with the highest conversion rate as the advertising product and construct a target advertisement corresponding to the advertising product for display on the target page.

[0078] In a further embodiment, the event response module 1100 includes: a first data acquisition submodule, used to acquire historical interaction data generated by the user corresponding to the target page loading event interacting with products in the independent website store multiple times, the historical interaction data including any multiple of behavior type weight, interactive product image, interactive product text, interactive product price, interaction freshness, and listing freshness; and a sequence component submodule, used to extract the interaction features corresponding to each historical interaction data, used to construct a user interaction sequence.

[0079] In a further embodiment, the interest assessment module 1200 includes: a first modeling submodule, used to perform feature relationship modeling on the user interaction sequence using a heterogeneous modeling subnetwork in a preset interest assessment model to obtain a first feature sequence; a second modeling submodule, used to perform interaction order relationship modeling on the first feature sequence using an order modeling subnetwork in the interest assessment model to obtain a second feature sequence; and a third modeling submodule, used to perform target interest product inference mapping modeling on the second feature sequence using a target inference subnetwork in the interest assessment model to obtain a corresponding interest assessment distribution, wherein the interest assessment distribution includes the corresponding user's interest rating for each product in the independent website store.

[0080] In a further embodiment, the conversion evaluation module 1300 includes: an embedding encoding submodule, used to encode the product profile, the user profile, and the user interaction sequence using an embedding encoding subnetwork in a preset conversion evaluation model to obtain corresponding product profile vectors, user profile vectors, and sequence vectors; a feature cross submodule, used to fuse the product profile vector and the sequence vector with the product profile vector and user profile vector respectively using a feature cross subnetwork in the conversion evaluation model to obtain corresponding product sequence vectors and user sequence vectors; and a fully connected mapping submodule, used to perform a conversion mapping on the underlying latent vector obtained by concatenating the product sequence vector and the user sequence vector using a multi-layer fully connected subnetwork in the conversion evaluation model to obtain the conversion rate of the product.

[0081] In a further embodiment, the advertising display module 1400 includes: a second data acquisition submodule, used to acquire the selling point copywriting image and the target product image of the advertised product, and to acquire an advertising template image and its completion composition information that match the user's profile from a preset advertising template library; and a template filling submodule, used to fill the selling point copywriting image and the target product image into the advertising template image according to the completion composition information to obtain the target advertisement of the advertised product.

[0082] In a further embodiment, after the advertising display module 1400, there are: an event response submodule, used to respond to the user's trigger event on the target advertisement, jump to the product details page of the advertised product, and create a dialogue window between the user and the customer service guide; and a message publishing submodule, used to generate product shopping guide information to guide the user to make a quick decision, and the customer service guide publishes the product shopping guide information to the dialogue window for the user to view.

[0083] In a further embodiment, before the interest evaluation module 1200, there are: a training preparation submodule, used to acquire a dataset, the dataset including multiple samples and their supervision labels, the samples and their supervision labels corresponding to sample interaction sequences and their corresponding target interest evaluation distributions; and a model convergence submodule, used to train an interest evaluation model to a convergent state based on the dataset, and learn the ability to infer the corresponding interest evaluation distribution based on any user interaction sequence.

[0084] To address the aforementioned technical problems, embodiments of this application also provide computer equipment. For example... Figure 4The diagram shows the internal structure of a computer device. The computer device includes a processor, a computer-readable storage medium, a memory, and a network interface connected via a system bus. The computer-readable storage medium stores an operating system, a database, and computer-readable instructions. The database may store a sequence of control information. When the computer-readable instructions are executed by the processor, the processor can implement a method for displaying advertisements on an independent website. The processor of the computer device provides computing and control capabilities to support the operation of the entire computer device. The memory of the computer device may store computer-readable instructions. When the computer-readable instructions are executed by the processor, the processor can execute the method for displaying advertisements on an independent website as described in this application. The network interface of the computer device is used for communication with a terminal. Those skilled in the art will understand that… Figure 4 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0085] In this embodiment, the processor is used to execute... Figure 3 The system contains the specific functions of each module and its sub-modules. The memory stores the program code and various data required to execute these modules or sub-modules. The network interface is used for data transmission between the user terminal and the server. In this embodiment, the memory stores the program code and data required to execute all modules / sub-modules in the independent website advertising display device of this application. The server can call the server's program code and data to execute the functions of all sub-modules.

[0086] This application also provides a storage medium storing computer-readable instructions, which, when executed by one or more processors, cause the one or more processors to perform the steps of the independent website advertising display method of any embodiment of this application.

[0087] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments of this application can be implemented by a computer program instructing related hardware. This computer program can be stored in a computer-readable storage medium, and when executed, it can include the processes of the embodiments of the methods described above. The aforementioned storage medium can be a magnetic disk, optical disk, read-only memory (ROM), or random access memory (RAM), etc.

[0088] In summary, this application enables the push of product advertisements that attract users' attention, and is expected to stimulate users to interact with the product and make a conversion.

[0089] Those skilled in the art will understand that the steps, measures, and solutions in the various operations, methods, and processes discussed in this application can be alternated, modified, combined, or deleted. Furthermore, other steps, measures, and solutions in the various operations, methods, and processes discussed in this application can also be alternated, modified, rearranged, decomposed, combined, or deleted. Furthermore, steps, measures, and solutions in the prior art that are similar to those in the open-source operations, methods, and processes of this application can also be alternated, modified, rearranged, decomposed, combined, or deleted.

[0090] The above description is only a partial embodiment of this application. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of this application, and these improvements and modifications should also be considered within the scope of protection of this application.

Claims

1. A method for displaying advertisements on an independent website, characterized in that, Includes the following steps: Respond to the target page load event of the independent website store and obtain the corresponding user interaction sequence for that event; A preset interest assessment model is used to determine the user's interest rating for each product in the independent website store based on the user interaction sequence; For each product whose interest score meets the preset conditions, a preset conversion evaluation model is used to determine the corresponding conversion rate based on the product profile, the user profile, and the user interaction sequence. The product with the highest conversion rate is selected as the advertising product, and a corresponding target ad is constructed and displayed on the target page.

2. The independent website advertising display method according to claim 1, characterized in that, Obtain the user interaction sequence corresponding to the event, including the following steps: Obtain historical interaction data generated by the user corresponding to the target page loading event interacting with products in the independent website store multiple times. The historical interaction data includes any multiple items from behavior type weight, interactive product image, interactive product text, interactive product price, interaction freshness, and listing freshness. Extract the interaction features corresponding to each historical interaction data point to construct a user interaction sequence.

3. The method for displaying advertisements on an independent website according to claim 1, characterized in that, Using a pre-defined interest assessment model, based on the user interaction sequence, determine the corresponding user's interest rating for each product in the independent website store, including the following steps: The user interaction sequence is modeled using a heterogeneous modeling subnetwork in a preset interest evaluation model to obtain a first feature sequence. The first feature sequence is modeled using the order modeling subnetwork in the interest evaluation model to obtain the second feature sequence; The target inference subnetwork in the interest evaluation model models the inference mapping of the second feature sequence to target interest products, and obtains the corresponding interest evaluation distribution, which includes the interest rating of the corresponding user for each product in the independent website store.

4. The method for displaying advertisements on an independent website according to claim 1, characterized in that, A preset conversion evaluation model is used to determine the corresponding conversion rate based on the product profile, the user profile, and the user interaction sequence, including: The embedded encoding subnetwork in the preset conversion evaluation model encodes the product profile, the user profile, and the user interaction sequence to obtain the corresponding product profile vector, user profile vector, and sequence vector. The product profile vector and the sequence vector are fused together by the feature cross subnetwork in the conversion evaluation model to obtain the corresponding product sequence vector and user sequence vector. The conversion rate of the product is obtained by performing a conversion mapping on the underlying latent vector obtained by concatenating the product sequence vector and the user sequence vector through the multi-layer fully connected subnetwork in the conversion evaluation model.

5. The method for displaying advertisements on an independent website according to claim 1, characterized in that, Constructing a target ad for a product includes the following steps: The system obtains the selling point copy images and target product images of the advertised products, and retrieves the advertising template images and their completion composition information that match the user's profile from the preset advertising template library. The selling point text image and the target product image are added to the advertising template image according to the completed composition information to obtain the target advertisement for the advertised product.

6. The method for displaying advertisements on an independent website according to claim 1, characterized in that, After constructing the target ad corresponding to the advertised product for display on the target page, the following steps are included: In response to the user's trigger event on the target advertisement, the user is redirected to the product details page of the advertised product, and a chat window is created between the user and the customer service representative. Product recommendation information is generated to guide the user in making a quick decision. The customer service representative then publishes this product recommendation information to the chat window for the user to view.

7. The method for displaying advertisements on an independent website according to claim 1, characterized in that, Before determining the user's interest rating for each product in the independent website store based on the user interaction sequence using a preset interest assessment model, the following steps are included: Obtain a dataset, which includes multiple samples and their supervision labels, wherein the samples and their supervision labels correspond to sample interaction sequences and their corresponding target interest evaluation distributions; The interest evaluation model is trained to convergence based on the dataset, and the model learns the ability to infer the corresponding interest evaluation distribution based on any user interaction sequence.

8. An independent website advertising display device, characterized in that, include: The event response module is used to respond to the target page loading event of an independent website store and obtain the corresponding user interaction sequence for the event. The interest assessment module is used to determine the user's interest rating for each product in the independent website store based on the user interaction sequence using a preset interest assessment model. The conversion evaluation module is used to determine the corresponding conversion rate for each product whose interest score meets the preset conditions, based on the product profile, the user profile, and the user interaction sequence using a preset conversion evaluation model. The advertising display module is used to select the products with the highest conversion rates as advertising products and construct corresponding target advertisements for display on the target page.

9. A computer device comprising a central processing unit and a memory, characterized in that, The central processing unit is used to invoke and run a computer program stored in the memory to perform the steps of the method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, It stores, in the form of computer-readable instructions, a computer program implemented according to any one of claims 1 to 7, which, when invoked by a computer, executes the steps included in the corresponding method.

Images