A social e-commerce product selection and promotion method and device, computer equipment and storage medium

Abstract

The application provides a product selection and promotion method and device for social e-commerce, computer equipment and a storage medium. The method generates a portrait based on user historical promotion information, comprehensively characterizes user promotion features, and provides a data basis for subsequent operations. Secondly, the product description is matched with the user portrait, the target product is selected from the newly online products, and the target product is accurately pushed to the user. The whole scheme realizes the whole process closed loop from user feature analysis, product screening and matching to accurate pushing, compared with the traditional promotion method, reduces the transmission of invalid promotion information, improves the matching degree of product and user promotion demand, reduces the platform product selection and promotion cost, improves the promotion efficiency and conversion rate, can effectively enhance the user promotion enthusiasm and user stickiness, and at the same time helps the platform to optimize the resource allocation, provides strong technical support for the sustainable development of social e-commerce.

Description

Technical Field

[0001] This application relates to the field of social e-commerce technology, and in particular to a method, apparatus, computer equipment, and storage medium for product selection and promotion in social e-commerce. Background Technology

[0002] In the social e-commerce sector, the accuracy of product selection and promotion directly impacts user experience and platform operational efficiency. Currently, product selection and promotion in social e-commerce largely rely on manual experience or simple data statistics, lacking in-depth analysis and dynamic matching of users' personalized needs. This results in low relevance between promotional content and users' actual situations, leading to a large number of ineffective push notifications, wasting platform resources, and easily provoking user resentment. Summary of the Invention

[0003] The purpose of this application is to at least address one of the aforementioned technical deficiencies, particularly the deficiency in the existing technology of products pushed to users having a low degree of matching with users.

[0004] The first aspect is a product selection and promotion method for social e-commerce, including:

[0005] Generate a user profile of the target user based on the target user's historical promotional information set;

[0006] After any product is launched, it is matched with user profiles based on product description information;

[0007] If the match is successful, the newly launched product will be identified as the target product.

[0008] Push the target product to the target users.

[0009] In one embodiment, the historical promotion information set includes multiple historical promotion information entries, which contain record data across various dimensions. Based on the target user's historical promotion information set, a user profile of the target user is generated, including:

[0010] For any historical promotional information, convert each record of data into promotional tags;

[0011] Group historical promotional information with the same promotional tag into one promotional tag group;

[0012] Determine the promotion success rate for each promotion tag group separately;

[0013] Select the target tag group from the promotion tag group based on the success rate of each promotion;

[0014] Generate user profiles based on target tag groups.

[0015] In one embodiment, the promotion success rate of each promotion tag group is determined, including:

[0016] Get the total number of executions for all historical promotional information within the promotional tag group;

[0017] Each completed order's historical promotional information is assigned a time decay weight; the greater the time difference between the current time and the order placement time of the completed order's historical promotional information, the smaller the time decay weight.

[0018] The weighted success rate is calculated by counting the number of historical promotional messages that resulted in orders based on the time decay weight.

[0019] The weighted ratio of the number of successful executions to the total number of executions is determined as the promotion success rate.

[0020] In one embodiment, the dimensions of the recorded data include average order value, promotion channels, product categories, whether an order is completed, order time, and commission rate.

[0021] In one embodiment, the product selection and promotion method further includes:

[0022] Update the historical promotion information set based on the real-time promotion records of the target users;

[0023] If the conditions for updating the user profile are met, the user profile will be updated based on the updated historical promotion information set.

[0024] In one embodiment, the historical promotion information set is updated based on the target user's real-time promotion records, including:

[0025] After the promotional link corresponding to the target user is clicked, the operation record after the promotional link is clicked is monitored, and a historical promotional information is generated based on the operation record and updated to the historical promotional information set.

[0026] In one embodiment, matching product description information with user profiles includes:

[0027] Product features are extracted from product description information to obtain multiple product feature keywords;

[0028] Convert the key features of each product into product tags to obtain a set of product tags;

[0029] Perform similarity matching between the product tag set and each target tag group;

[0030] If any similarity score is greater than the first threshold, the match is considered successful.

[0031] Otherwise, the match is deemed to have failed.

[0032] In one embodiment, pushing the target product to the target user includes:

[0033] Based on the target tag group that matches the target product, identify one or more target promotion channels;

[0034] Generate corresponding promotional copy based on the target promotion channels and product description information;

[0035] Package the target product's promotional link, target promotional channels, and corresponding promotional copy and send them to the target users.

[0036] In one embodiment, corresponding promotional copy is generated based on the target promotion channel and product description information, including:

[0037] Based on the style prompts and product description information corresponding to the target promotion channels, we obtain the prompts for generating channel copy.

[0038] Input the prompts generated from the channel copy into the first model to obtain the promotional copy corresponding to the target promotional channel.

[0039] In one embodiment, style cue words correspond one-to-one with selectable promotion channels, and the generation process of each style cue word includes:

[0040] Obtain the collection of promotional copy for the corresponding available promotional channels;

[0041] Style-extracted prompts are generated based on the set of promotional copy; these style-extracted prompts are used to indicate style prompts generated by the second major model based on the set of promotional copy.

[0042] Input the style extraction prompts into the second model to obtain the corresponding style prompts.

[0043] In one embodiment, the pre-training process of the second large model includes:

[0044] Obtain the collection of historical promotional copy corresponding to the available promotional channels, and annotate each historical promotional copy to obtain the corresponding standard style prompts;

[0045] The initial second-largest model was fine-tuned using a set of annotated historical promotional copy.

[0046] In one embodiment, the standard style cue words include tone analysis items, sentiment analysis items, and sentence structure analysis items. The annotated historical promotional copy includes tone tags, sentiment tags, and sentence structure tags. The initial second-largest model is fine-tuned using the annotated set of historical promotional copy, including:

[0047] Each historical promotional copy is input into the initial second model, and the second model is instructed to output the corresponding predicted tone label, predicted sentiment label, and predicted sentence structure label;

[0048] The first loss term is obtained based on the difference between the tone label and the predicted tone label, the second loss term is obtained based on the difference between the sentiment label and the predicted sentiment label, and the third loss term is obtained based on the difference between the sentence structure label and the predicted sentence structure label.

[0049] The objective function is obtained by weighted summation of the first, second, and third loss terms;

[0050] With the goal of reducing the objective function, the parameters of the second large model are fine-tuned, and the steps of obtaining the first loss term based on the difference between the tone label and the predicted tone label, the second loss term based on the difference between the sentiment label and the predicted sentiment label, and the third loss term based on the sentence structure label and the predicted sentence structure label are returned, until the fine-tuning termination condition is met.

[0051] In one embodiment, the objective function is obtained by weighted summation of the first loss term, the second loss term, and the third loss term, and the method further includes:

[0052] For any one of the first loss term, the second loss term, and the third loss term, if the rate of change for a first number of consecutive periods is lower than the second threshold, then the corresponding weight is reduced.

[0053] For any one of the first, second, and third loss terms, if the absolute value is greater than the third threshold, the corresponding weight is increased.

[0054] Secondly, this application provides a product selection and promotion device for social e-commerce, comprising:

[0055] The profile generation module is used to generate user profiles for target users based on their historical promotional information.

[0056] The product matching module is used to match any product with user profiles based on product description information after the product is launched.

[0057] The target product determination module is used to identify the newly launched product as the target product if a match is successful.

[0058] The push module is used to push target products to target users.

[0059] Thirdly, this application provides a computer device including one or more processors and a memory storing computer-readable instructions. When the computer-readable instructions are executed by one or more processors, they perform the steps of the product selection and promotion method in any of the above embodiments.

[0060] Fourthly, this application 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 product selection and promotion method in any of the above embodiments.

[0061] As can be seen from the above technical solutions, the embodiments of this application have the following advantages:

[0062] Based on the product selection and promotion methods in this solution, a series of technical means are used to achieve efficient and precise social e-commerce product selection and promotion. First, user profiles are generated based on historical promotion information to comprehensively depict user promotion characteristics, providing a data foundation for subsequent operations. Second, product descriptions are matched with user profiles to filter target products from newly launched products, and then these target products are precisely pushed to users. The entire solution achieves a closed-loop process from user characteristic analysis and product selection and matching to precise push notifications. Compared with traditional promotion methods, it reduces the transmission of invalid promotional information, improves the alignment between products and user promotion needs, reduces platform product selection and promotion costs, and improves promotion efficiency and conversion rates. It effectively enhances user promotion enthusiasm and user stickiness, while also helping the platform optimize resource allocation and providing strong technical support for the sustainable development of social e-commerce. Attached Figure Description

[0063] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0064] Figure 1 A flowchart illustrating a product selection and promotion method provided in one embodiment of this application;

[0065] Figure 2 This is a schematic diagram of the process for generating a user profile in one embodiment of this application;

[0066] Figure 3 This is a flowchart illustrating the process of determining the success rate of promotion in one embodiment of this application;

[0067] Figure 4 This is a schematic diagram illustrating the process of matching a target product with a user profile in one embodiment of this application;

[0068] Figure 5 This is a schematic diagram of the process of pushing a target product to a target user in one embodiment of this application;

[0069] Figure 6 This is a schematic diagram of the process for generating style cue words in one embodiment of this application;

[0070] Figure 7 This is a flowchart illustrating the process of fine-tuning the second major model in one embodiment of this application;

[0071] Figure 8 This is an internal structural diagram of a computer device provided in one embodiment of this application. Detailed Implementation

[0072] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0073] This application provides a product selection and promotion method for social e-commerce. Please refer to [link / reference]. Figure 1 This includes steps S102 to S108.

[0074] S102, Generate a user profile of the target user based on the target user's historical promotion information set.

[0075] It is understandable that target users refer to individuals or groups engaged in product promotion and sales activities on social e-commerce platforms, who generate revenue through their own social networks and other channels. The historical promotion information set includes comprehensive data on target users' past product promotions on the platform, such as the product categories promoted, promotion periods, distribution of promotion channels (WeChat Moments, Douyin short videos, Xiaohongshu notes, etc.), sales data during the promotion period (transaction volume, sales revenue, return rate), and may also include user-audience interaction data (number of comments, likes, and shares). User profiles are constructed based on the aforementioned multi-source heterogeneous data of target users, analyzing and integrating the characteristics of users' promotional behavior patterns.

[0076] This step relies on big data analytics and machine learning technologies. First, data acquisition techniques are used to obtain historical promotional information sets from target users. Data cleaning techniques remove duplicate, erroneous, and invalid data. Then, data integration techniques are used to consolidate data scattered across different systems (such as order systems, marketing systems, and user behavior analysis systems) to form a structured dataset. Next, this structured dataset is analyzed to identify the types and characteristics of products with a high probability of successful promotion to users.

[0077] S104: After any product is launched, it matches the product description information with the user profile.

[0078] It's understandable that product description information is comprehensive data accompanying a product when it's launched on a social e-commerce platform. This includes structured and unstructured information such as basic product attributes (brand, origin, material, specifications), core selling points (functional features, technological advantages), and market positioning (target consumer group, price range). This step, based on information retrieval and semantic matching principles, compares the product description information with user profiles using multi-dimensional features. By calculating the similarity or matching degree of the two features, it determines whether the product is suitable for the target user's promotional capabilities and audience needs. For example, if the user profile shows that a promoter specializes in promoting maternal and infant products, targeting primarily young mothers who prefer high-commission items, and a newly launched maternal and infant product boasts high cost-effectiveness and its promotional copy emphasizes parent-child interaction, then a high degree of matching with the user profile indicates that the product might be suitable for this promoter.

[0079] First, feature extraction can be performed on product description information and user profiles, converting them into vector forms that can be processed by computers. Structured data is encoded numerically, while unstructured text data is vectorized using natural language processing techniques (such as Word2Vec and BERT). Then, algorithms such as cosine similarity and edit distance are used to calculate the similarity between vectors. User profiles can be expressed in the form of tags, or product description information can be converted into tags, and matching can be performed using the similarity between the tags.

[0080] S106. If the match is successful, the newly launched product will be identified as the target product.

[0081] It can be understood that a successful match means that the product description information and the user profile, after feature comparison and similarity calculation in step S104, meet the pre-set matching criteria (such as similarity thresholds and rule matching conditions). The target product is the product that has been screened and identified as meeting the promotional advantages and audience needs of the target user, possessing high promotional value and market potential. This step is based on decision theory, judging the fit between the product and the user according to the established matching criteria. When the product and the user profile match successfully, it indicates that the product aligns with the user's promotional direction and audience preferences, possessing a high promotional success rate, and is therefore identified as the target product; conversely, if the matching conditions are not met, the product is deemed unsuitable for the current user's promotion and is excluded. This step is a further screening and confirmation of the matching results in step S104, providing a clear target for accurate product delivery and is a key decision-making step for achieving efficient promotion. Specifically, this step can use a rule engine (such as the Drools rule engine) to define the threshold and rules for successful matching, and implement the conditional judgment logic through programming. For example, setting the similarity threshold between the product and the user profile to 0.7, when the calculated result is greater than or equal to this threshold, a successful match is determined, and the product is marked as the target product.

[0082] S108 pushes the target product to the target user.

[0083] This step is understandable; it involves social e-commerce platforms using various channels and methods to deliver promotional links for target products to target users, guiding them to engage in product promotion activities. Common push channels include platform in-app messages, dedicated promotion task centers, personalized recommendation pages, and instant messaging tools. Based on the principles of information dissemination and user incentives, after identifying the target product, to enhance user engagement and promotional effectiveness, it's necessary to select appropriate push timing, channels, and content formats based on user behavior habits, communication preferences, and promotional needs, ensuring users receive valuable promotional information in a timely manner.

[0084] Based on the product selection and promotion methods in this solution, a series of technical means are used to achieve efficient and precise social e-commerce product selection and promotion. First, user profiles are generated based on historical promotion information to comprehensively depict user promotion characteristics, providing a data foundation for subsequent operations. Second, product descriptions are matched with user profiles to filter target products from newly launched products, and then these target products are precisely pushed to users. The entire solution achieves a closed-loop process from user characteristic analysis and product selection and matching to precise push notifications. Compared with traditional promotion methods, it reduces the transmission of invalid promotional information, improves the alignment between products and user promotion needs, reduces platform product selection and promotion costs, and improves promotion efficiency and conversion rates. It effectively enhances user promotion enthusiasm and user stickiness, while also helping the platform optimize resource allocation and providing strong technical support for the sustainable development of social e-commerce.

[0085] In one embodiment, the historical promotion information set includes multiple historical promotion information entries. These historical promotion information entries include record data from various dimensions. Based on the target user's historical promotion information set, a user profile of the target user is generated. Please refer to [link to relevant documentation]. Figure 2 This includes steps S202 to S210.

[0086] S202, for any historical promotional information, convert each record data into promotional tags.

[0087] Historical promotional information refers to various data records generated when target users previously promoted products on social e-commerce platforms. This data can be categorized into multiple dimensions, such as average order value, promotion channel, product category, order status, order time, and commission rate. These recorded data serve as the building blocks of historical promotional information, representing the specific data values ​​for each of the aforementioned dimensions, such as an average order value of 200 yuan, the promotion channel being Douyin (TikTok), and the product category being beauty and skincare. Promotional tags, on the other hand, are semantic identifiers abstracted and refined from the recorded data, used to summarize the core characteristics of historical promotional information. This step is based on the theory of data feature extraction and abstraction. Differentiated processing strategies are employed for data recorded across different dimensions. For numerical data such as average order value (AOV), discretization is achieved by setting thresholds or using data binning techniques. For example, AOV can be categorized into "low AOV (< 100 yuan)," "medium AOV (100-500 yuan)," and "high AOV (≥ 500 yuan)" and assigned corresponding tags. Promotion channels and product categories are categorized data and can be directly mapped to tags based on a pre-defined classification system. For instance, the promotion channel "Douyin (TikTok)" is mapped to the tag "Douyin Channel Promotion," and the product category "Beauty & Skincare" is mapped to the tag "Beauty & Skincare Category Promotion." The "whether a sale was completed" tag is a Boolean value and is directly converted into "Sold Promotion" or "Unsold Promotion" tags. Order time can be extracted based on time characteristics, such as categorized by quarter, month, or time period, generating tags like "Spring Promotion" and "Evening Promotion." Commission rates are also categorized using thresholds, forming tags like "Low Commission Rate" and "High Commission Rate." This step transforms complex recorded data into promotional tags, providing a standardized data format for subsequent data grouping, analysis, and user profile construction, and is the foundation of the entire process. Structured data can be processed using Python's pandas library. Data can be transformed into labels through custom mapping rules and binning functions, such as using the `cut` function to bin data based on average order value and commission rate. For time-series data, pandas' time series processing capabilities can be used to extract features and generate labels.

[0088] S204, group historical promotional information with the same promotional tag into a promotional tag group.

[0089] It can be understood that a promotion tag group is a data set aggregated from multiple historical promotional messages with the same promotion tag. The core principle of this step is data clustering. Based on the promotion tags generated in step S202, historical promotional information is categorized and integrated, grouping promotional information with the same characteristic identifier into one group, which facilitates subsequent centralized analysis of similar promotional activities. This grouping method reduces data complexity, organizes scattered historical promotional information into a structured organization according to key features, and is conducive to observing and comparing the characteristics and effects of different types of promotional activities. It provides an ordered data structure for calculating promotion success rate and screening target tag groups, and is an important intermediate step in building user profiles. A hash table grouping method can be used, with promotion tags as keys and corresponding historical promotional information lists as values ​​to build a hash table, and data can be traversed and inserted into the corresponding list to complete the grouping; alternatively, relational databases such as MySQL can be used to achieve grouped queries through the "GROUP BY promotion tag" statement.

[0090] S206, determine the promotion success rate of each promotion tag group.

[0091] It is understandable that the promotion success rate is a quantitative indicator that measures the promotional effectiveness of historical promotional information within each promotional tag group. It can be determined based on the ratio of the number of completed orders within the group to the total number of promotions, or by weighting multiple dimensions such as average order value, sales volume, and commission rate to arrive at a comprehensive value. This step, based on data analysis and evaluation theory, evaluates the effectiveness of the promotional tag groups formed in step S204, quantifies the effectiveness of different types of promotional activities, and clarifies which tag groups correspond to promotional strategies that are more recognized by the market and can bring higher returns. This provides a decision-making basis for selecting target tag groups and reflects the user's promotional capabilities and market adaptability in different fields. It is a key data support step for building accurate user profiles.

[0092] S208: Select the target tag group from the promotion tag group based on the success rate of each promotion.

[0093] It is understandable that the target tag group is selected from all promotion tag groups, representing those with a high promotion success rate. Based on the promotion success rate calculated in step S206, the tag group that best reflects the user's promotional advantages, preferences, and market adaptability is selected by setting screening criteria or algorithms. A threshold screening method can be used, with a preset success rate threshold to determine the target tag group; alternatively, algorithms such as quicksort or heapsort can be used to sort the tags in descending order of success rate, selecting the top-ranked tag groups. The selection of the target tag group directly affects the accuracy and usability of the user profile. Selecting representative tag groups can highlight the user's core promotional characteristics, helping the platform to accurately select products and formulate promotional strategies for the user's advantageous areas.

[0094] S210, Generate user profiles based on target tag groups.

[0095] It is understandable that a user profile is a comprehensive digital description of a target user in the field of social e-commerce promotion. By integrating target tag group information, it presents the user's promotion characteristics, advantages, audience preferences and market adaptability in a tagged form.

[0096] In one embodiment, the promotion success rate of each promotion tag group is determined separately. Please refer to [link to relevant documentation]. Figure 3 This includes steps S302 to S308.

[0097] S302, retrieve the total number of executions of all historical promotional information within the promotional tag group.

[0098] In the social e-commerce user profile building system, the total execution count refers to the total number of records of all historical promotional information within a promotional tag group. In other words, it's the sum of the number of times all promotional activities under that tag group were implemented, regardless of whether the promotion was successful or not. This metric is the basic denominator for calculating the promotion success rate, used to measure the scale of the promotional activity. This step obtains the raw data scale by simply counting the historical promotional information within the promotional tag group, providing benchmark data for subsequent weighted calculations. Its principle lies in the basic counting logic of data statistics: by traversing all records within the promotional tag group, the total execution count is accumulated. This step works closely with subsequent steps; the total execution count, as the denominator in the calculation of the promotion success rate, directly affects the accuracy of the final result.

[0099] S304 assigns a time decay weight to each completed historical promotional message. The greater the time difference between the current time and the order time of the completed historical promotional message, the smaller the time decay weight.

[0100] As we can understand, time decay weight is a numerical indicator used to measure the timeliness of historical promotional information. It reflects the degree to which historical promotional information from a completed order contributes to the current promotional effect evaluation over time. Considering that factors such as market environment, user demand, and competitive landscape change over time, earlier order promotional information has relatively low reference value for the current promotional strategy. Therefore, by constructing a time decay model, the weight of each historical promotional information from a completed order is dynamically adjusted based on the time difference between the current time and the order placement time, so that recent promotional results have a greater weight in the evaluation, more accurately reflecting the user's current promotional capabilities and market adaptability. The time decay function here can use an exponential decay function, such as... ,in, For time decay weight, The preset decay factor is t, which is the time difference between the current time and the order time of the historical promotional information for completed orders.

[0101] S306: The number of historical promotional messages that resulted in orders is calculated based on the time decay weight, resulting in a weighted number of successful transactions.

[0102] It is understandable that the weighted success count is a value obtained by weighting the number of historical promotional messages that resulted in a sale, taking into account the time decay weight. It more accurately reflects the effective results of the promotional campaign across different time dimensions. The time decay weight of each historical promotional message that resulted in a sale, calculated in step S304, is multiplied by 1 (representing one successful promotion), and then the weighted values ​​of all historical promotional messages that resulted in a sale are summed. This method considers both the fact of successful promotions and assigns different importance to successful promotions in different historical periods based on the time factor, avoiding treating all sales records equally and making the statistical results more timely and valuable for reference. The weighted success count, as the numerator in calculating the promotional success rate, together with the total number of executions, determines the final value of the promotional success rate. The accuracy of its calculation directly affects the evaluation of the promotional effect.

[0103] S308 defines the weighted ratio of the number of successful executions to the total number of executions as the promotion success rate.

[0104] As is understandable, the promotion success rate is a core indicator used to measure the effectiveness of promotional activities within a promotional tag group. The weighted ratio calculation method comprehensively considers the total execution scale of the promotional activity and the success results across different time dimensions, making it more scientific and reasonable than traditional, simple success rate calculation methods. This step is based on ratio assessment theory, using the total number of executions obtained in step S302 as the denominator and the weighted success rate calculated in step S306 as the numerator. The ratio of these two values ​​yields a value between 0 and 1, intuitively reflecting the success level of the promotional activity within the promotional tag group.

[0105] In one embodiment, the product selection and promotion method further includes: updating the historical promotion information set based on the real-time promotion records of the target user. If the profile update conditions are met, the user profile is updated based on the updated historical promotion information set.

[0106] In the dynamic data management system for product selection and promotion in social e-commerce, the real-time promotion records of target users refer to the instantaneous data records generated by the promotional activities that target users are currently conducting or have just completed on the social e-commerce platform. In the rapidly changing operating environment of social e-commerce, user promotional behavior and market feedback are constantly evolving. Real-time acquisition and updating of promotion records ensures that the historical promotion information set always contains the latest data, providing timely support for subsequent analysis. Message queue technologies such as Apache Kafka can be used to achieve efficient transmission and processing of real-time promotion records. When a target user generates a new promotion record, the data is sent to a Kafka topic in the form of a message. The data processing system reads the message in real-time through consumer groups and writes it to the storage system corresponding to the historical promotion information set (such as a relational database like MySQL or a distributed database like Cassandra). Using the streaming computing framework Apache Flink, real-time promotion records can be cleaned, transformed, and aggregated in real-time, directly updating the historical promotion information set.

[0107] User profile update conditions are pre-defined rules and standards used to determine whether user profiles need to be updated. These conditions can be based on various factors such as time intervals (e.g., checking every 24 hours), data change volume (e.g., the amount of new data added to the historical promotion information set exceeds a certain threshold), and business event triggers (e.g., users promoting new product categories or entering new promotion channels). This step is based on dynamic model optimization theory. A user's promotional ability, preferences, and market adaptability change over time and with business changes. When the profile update conditions are met, promotional tags are recalculated, promotional tag groups are divided, promotional success rates are determined, and target tag groups are selected based on the updated historical promotion information set. This updates the user profile, making it more accurately reflect the user's current situation and providing a more realistic basis for subsequent product selection and promotion. It works in conjunction with the previous historical promotion information set update step; the former provides the updated data foundation for the latter, while the profile update further optimizes the decision-making basis for product selection and promotion, forming a dynamic optimization closed loop.

[0108] In one embodiment, the historical promotion information set is updated based on the target user's real-time promotion records. This includes: monitoring the operation records after the promotion link corresponding to the target user is clicked, generating a historical promotion information record based on the operation records, and updating it to the historical promotion information set. It can be understood that the promotion link corresponding to the target user refers to the exclusive link used by the target user to promote products on the social e-commerce platform. It is unique and allows for precise tracking of the user's promotional effectiveness. The operation records refer to the data records generated during a series of actions such as browsing products, adding items to the cart, and placing orders after the promoted user clicks the promotion link. These records cover multiple dimensions of information, including click time, duration of browsing product details pages, time of adding items to the cart, time of placing an order, quantity of purchased items, and payment amount. User behavior after clicking the promotion link can directly reflect the promotional effect and user interests. By monitoring the operation records, the entire process data from clicking the promotion link to completing a purchase (or other actions) can be obtained. Integrating this data to generate historical promotion information and updating it to the historical promotion information set provides more detailed and targeted data support for subsequent analysis.

[0109] In one embodiment, product description information is matched with user profiles; please refer to [link to relevant documentation]. Figure 4 This includes steps S402 to S410.

[0110] S402, extract product features from product description information to obtain multiple product feature keywords.

[0111] Product feature keywords are representative words or phrases extracted from descriptive information. For example, for a product description like "a moisturizing and repairing serum suitable for sensitive skin," keywords such as "sensitive skin," "moisturizing and repairing," and "serum" can be extracted. These keywords form the basis for generating product tags and directly affect the accuracy of matching. For structured product descriptions, such as brand and material, extraction can be performed directly based on a pre-defined classification system and mapping rules. For unstructured text information, such as product promotional materials and user reviews, natural language processing techniques such as lexical analysis, syntactic analysis, and semantic analysis are required. First, the text is segmented into words using word segmentation technology. Then, methods such as part-of-speech tagging and named entity recognition are used to filter out nouns, verbs, adjectives, and other words with practical meaning. Finally, keyword extraction algorithms (such as TF-IDF and TextRank algorithms) are used to extract the keywords that best represent the product features from these words.

[0112] S404 converts the key features of each product into product tags, resulting in a set of product tags.

[0113] It can be understood that a product tag set is a collection of multiple product tags that comprehensively and systematically describe the various characteristics and attributes of a product. This step is based on semantic mapping and knowledge integration theories. After extracting product feature keywords in step S402, these keywords need to be further abstracted and categorized, converting them into product tags with clear semantic and business meanings. This process requires referencing a pre-established product tag system, which contains standard tag definitions for various product attributes and characteristics. By matching and mapping product feature keywords with tags in the tag system, or by summarizing and merging keywords based on their semantics, the keywords are converted into corresponding product tags. The tag system used here is the same as that used for the target tag group, facilitating subsequent matching.

[0114] The conversion can be achieved by establishing a keyword-tag mapping table. The mapping table predefines the correspondence between various keywords and product tags. Programmatically, iterates through the list of product feature keywords, searching for matching tags in the mapping table for conversion. For keywords that cannot be directly matched, semantic similarity calculation methods from natural language processing (such as cosine similarity) can be used to calculate the semantic similarity between the keyword and each tag in the tag system, selecting the tag with the highest similarity as the conversion result. Alternatively, a rule engine (such as the Drools rule engine) can be used to automatically convert keywords to tags by writing rule statements.

[0115] S406, perform similarity matching between the product tag set and each target tag group.

[0116] It's understandable that matching the product tag set with the target tag set aims to determine whether the product meets the promotional needs and capabilities of the target user by quantifying the degree of similarity between the two. Specific similarity calculation methods can employ algorithms such as cosine similarity, Euclidean distance, and Jaccard coefficient. First, the product tag set and target tag set are converted into computer-processable vector forms, for example, using a bag-of-words model or word vector model (such as Word2Vec or GloVe) to represent tags as vectors. Then, the selected similarity calculation algorithm is used to calculate the similarity value between the vectors. A higher similarity value indicates a better match between the product and the target user's promotional characteristics, making the product more likely to be suitable for promotion. This step is the core of product-user profile matching, providing a quantitative basis for subsequent judgments on product matching success. It connects the generation of the product tag set with the determination of matching results, playing a screening and evaluation role in the entire product selection and promotion process.

[0117] S408: If any similarity score is greater than the first threshold, the match is considered successful.

[0118] S410, otherwise, the match is deemed to have failed.

[0119] It is understandable that the first threshold is a pre-set critical value used to determine whether a product and a target user are successfully matched. It is determined comprehensively based on factors such as the business needs of the social e-commerce platform, historical matching data, and promotional effects. After calculating the similarity between the product tag set and each target tag group in step S406, these similarity values ​​are compared with the first threshold. As long as there is at least one similarity value greater than the first threshold, it indicates that the product has a high degree of fit with the target user's promotional characteristics in some aspects and can meet at least some of the target user's promotional needs. Therefore, the product is determined to be successfully matched with the target user. The determination of successful matching provides a decision-making basis for subsequently identifying the product as the target product and pushing it to the target user. It is a key decision-making step for achieving accurate product selection and effective promotion. If all similarity values ​​are less than the first threshold, it indicates that the product and the target user's promotional characteristics are significantly different and not suitable for promotion to the target user. In this case, step S410 is required to determine that the matching has failed.

[0120] In one embodiment, the target product is pushed to the target user; please refer to [link to relevant documentation]. Figure 5 This includes steps S502 to S506.

[0121] S502: Determine one or more target promotion channels based on the target tag group that matches the target product.

[0122] As we can understand it, target promotion channels refer to social e-commerce promotion methods that are selected based on the characteristics of the target product and target users, and can effectively reach potential audiences and improve promotion results. Common ones include WeChat Moments, Douyin, Xiaohongshu, and Weibo. Target tag groups contain key characteristics of the target user's past successful promotion experience, such as which channels the user is good at promoting specific product categories, and promotion performance data on different channels. By analyzing target tag groups that match the target product, we can uncover channels that have yielded good results for the target user when promoting similar products. Simultaneously, by combining the attributes and market positioning of the target product, we can select the channel combination most likely to achieve efficient promotion. For example, if the target tag group includes promotion channel tags, it means that the target user has a high conversion rate on the promotion channels corresponding to those tags, and these can be considered as target promotion channels.

[0123] S504 generates corresponding promotional copy based on the target promotion channels and product description information.

[0124] As we can understand it, promotional copywriting refers to written promotional content created to advertise a target product on specific promotional channels, attract the attention of potential audiences, and encourage them to make a purchase. This step is based on the theory of channel characteristic adaptation and product selling point extraction. Different target promotional channels have different user group characteristics, content dissemination patterns, and format requirements. For example, WeChat Moments users prefer concise, lifestyle-oriented copywriting, while Douyin (TikTok) short video platforms require lively, interesting, and rhythmic video copywriting. Meanwhile, product descriptions contain key information such as the product's core selling points, functionalities, and market positioning. By analyzing the characteristics of the target promotional channels and the product descriptions, the most attractive selling points of the product on that channel are extracted, and expressed in a way that aligns with the channel's style and user preferences, targeted promotional copywriting is generated. This step acts as a bridge connecting the target promotional channels and the target product. By generating adapted promotional copy, the target product can better integrate into the ecosystem of each promotional channel, increasing user acceptance and interest in the product, thereby improving promotional effectiveness.

[0125] S506: Package the promotional link of the target product, the target promotional channels, and the corresponding promotional copy and send them to the target users.

[0126] As we can understand it, a promotional link is a unique link for a target product to target users on a social e-commerce platform, used to guide them to click through to the product details page to purchase or learn more. This step is based on the theory of information integration and precise push notifications. The promotional link of the target product, the selected target promotion channels, and the corresponding promotional copy generated for each channel are integrated and packaged into a complete promotional data package, which is then pushed to the target users. This allows them to conduct more targeted promotions, enabling them to carry out subsequent product promotion activities and achieve the promotional goals of the social e-commerce platform. Message queues (such as RabbitMQ and Kafka) can be used to achieve efficient transmission and asynchronous processing of the promotional data package. First, the promotional link of the target product, the target promotion channels, and their corresponding promotional copy are encapsulated into a message of a specific format and sent to the message queue. Then, the push service system reads the message from the message queue and, based on the target user's device identifier and push settings, sends the message to the corresponding terminal device.

[0127] In one embodiment, generating corresponding promotional copy based on the target promotion channel and product description information includes: obtaining channel copy generation prompts based on style hints for the target promotion channel and product description information; inputting the channel copy generation prompts into the first main model to obtain the promotional copy corresponding to the target promotion channel.

[0128] It's understandable that style cue words for target promotion channels are a set of keywords or phrases describing the content style, language characteristics, and user preferences of a specific promotion channel. Their purpose is to guide the large-scale model in generating copy that matches the characteristics of that channel. For example, style cue words for the Douyin (TikTok) short video platform might include "lively and interesting," "rhythmic," "conversational," and "visually impactful"; while style cue words for the Xiaohongshu (Little Red Book) image and text community might include "product recommendation," "detailed review," "emotional resonance," and "lifestyle." Channel copy generation cue words are formed by combining the style cue words of the target promotion channel with product description information, providing comprehensive guidance for the large-scale model to generate promotional copy. By combining the style characteristics of the target promotion channel with the core information of the product, the large-scale model can be provided with clearer and more targeted generation guidance. In practice, the product description information is first used to extract keywords to obtain the product's core selling points and key attributes; then, these keywords are integrated with the style cue words of the target promotion channel, and appropriate templates or structures are used to organize them into complete cue words. For example, for a smartwatch and the Douyin platform, the prompt for generating promotional copy might be: "In a vivid, interesting, and conversational way, combine the smartwatch's AMOLED high-definition screen, heart rate monitoring, sleep monitoring and other functions to highlight its suitability for sports enthusiasts and create a visually impactful and story-driven promotional copy."

[0129] The first major model refers to the large language model used to generate promotional copy, which possesses natural language understanding and generation capabilities. This step is based on text generation technology using a large language model. By learning from massive amounts of text data, the large model can understand the semantic information in prompt words and generate text content that meets the requirements. When the prompt words for channel copy generation are input into the first major model, the model analyzes the target promotional channel's style characteristics and product information contained in the prompt words. Utilizing its own knowledge and reasoning abilities, it generates promotional copy that matches the style of the target promotional channel and accurately conveys the product's selling points.

[0130] In one embodiment, please refer to Figure 6 Each style prompt word corresponds to a selectable promotion channel, and the generation process of each style prompt word includes steps S602 to S606.

[0131] S602, obtain the set of promotional copy for the corresponding optional promotional channels.

[0132] It's understandable that optional promotional channels refer to various methods available for product promotion on social e-commerce platforms, such as WeChat Moments, Douyin, Xiaohongshu, and Weibo. These channels differ in user groups, content formats, and dissemination patterns. A "collection of promotional copy" is a large collection of promotional copy data for a specific optional promotional channel. It covers various forms of promotional content, including text, images, and videos, published by different products, at different times, and by different creators on that channel. It includes information such as the channel's style characteristics, user preferences, and content trends. This collection of promotional copy can be obtained using web scraping technology. For different optional promotional channels, corresponding web scraping programs are written, and data collection rules are set according to the channel's webpage structure and data interface rules to scrape promotional copy data from the platform. For example, using Python's Scrapy framework to write a web scraper can efficiently scrape promotional copy from Xiaohongshu.

[0133] S604, Generate style extraction prompts based on the set of promotional copy. Style extraction prompts are used to instruct the second major model to generate style prompts based on the set of promotional copy.

[0134] It's understandable that style extraction prompts are designed based on the promotional copy set, serving as instructional text information to guide the second-largest model in style analysis and extraction. The second-largest model is a large-scale language model based on deep learning, such as the GPT series, BERT, and LLaMA, possessing natural language understanding and generation capabilities, and able to perform semantic analysis and feature extraction on the input text data. This step is based on natural language processing prompting engineering techniques and data feature analysis theory. First, the promotional copy set is preprocessed, including text cleaning, word segmentation, and stop word removal, to extract effective words and phrases. Then, the overall characteristics of the promotional copy set are analyzed, such as word frequency, sentence structure characteristics, sentiment, and theme distribution. Combined with prior knowledge of the styles of the available promotional channels, style extraction prompts are constructed. For example, if it's found that many trending internet terms frequently appear in the promotional copy collection on the Douyin platform, and the sentence structures are mostly short, powerful, and rhythmic, with a predominantly positive and exaggerated emotional tone, style extraction prompts can be generated: "Analyze the following Douyin promotional copy collection, extract frequently used trending internet terms, unique sentence structures, and prominent emotional tones, and summarize the style keywords of Douyin platform promotional copy that are vivid, interesting, colloquial, and rhythmic." These style extraction prompts act as a bridge connecting the promotional copy collection and the second major model, providing clear task instructions and analytical directions for the major model, ensuring that it can accurately extract style prompts that match the characteristics of the channel.

[0135] S606: Input the style extraction prompts into the second model to obtain the corresponding style prompts.

[0136] Understandably, the second model, after being trained on a large amount of text data, possesses the ability to understand complex semantic instructions and generate relevant text. When style extraction prompts are input into the second model, it first performs semantic parsing on the prompts to understand the task requirements and analysis direction. Then, it performs deep semantic analysis and feature extraction on the set of promotional copy. By learning from the language patterns, vocabulary collocations, and content structures in the set of promotional copy, and combining its own language knowledge and reasoning ability, the model extracts keywords or phrases that accurately summarize the style characteristics of the available promotional channels—the style prompts.

[0137] In one embodiment, the pre-training process of the second major model includes: acquiring a set of historical promotional copy corresponding to the available promotional channels, and annotating each historical promotional copy to obtain corresponding standard style prompts. The initial second major model is then fine-tuned using the annotated set of historical promotional copy. It can be understood that the set of historical promotional copy refers to the collection of all promotional copy data generated over a period of time for various available promotional channels (such as WeChat Moments, Douyin short video platform, Xiaohongshu text and image community, etc.) in a social e-commerce scenario. These copy pieces cover promotional content from different product types, different promotional periods, and different creators. Annotation refers to the process of manually or using specific algorithms to identify, extract, and define the style characteristics of historical promotional copy, aiming to assign accurate and representative standard style prompts to each copy. Standard style prompts are a set of keywords or phrases that have been rigorously defined and reviewed, accurately summarizing the style characteristics of the corresponding historical promotional copy, and serve as an important reference standard for model training. The principle of this step is based on data annotation and supervised learning theory. By collecting historical promotional copy, a rich set of raw data samples is provided for model training. The annotation process assigns clear labels to this data, constructing a supervised learning dataset with input (historical promotional copy) and output (standard style cue words). Accurate annotation enables the model to learn the mapping relationship between promotional copy and style features during subsequent fine-tuning, thereby improving the accuracy of the model in extracting style cue words. For example, for a promotional short video copy on the Douyin platform, annotators need to analyze its rhythm, language style, etc. If the copy uses a large number of popular internet slang terms and has simple sentence structures, it can be labeled with standard style cue words such as "fast-paced" and "popular internet slang terms."

[0138] In one embodiment, the standard style prompts include tone analysis items, sentiment analysis items, and sentence structure analysis items. The annotated historical promotional copy includes tone tags, sentiment tags, and sentence structure tags. (See [link to relevant documentation]). Figure 7 The initial second model is fine-tuned using a set of labeled historical promotional copy, including steps S702 to S708.

[0139] S702 inputs each historical promotional copy into the initial second model and instructs the second model to output the corresponding predicted tone label, predicted sentiment label, and predicted sentence structure label.

[0140] It's understandable that the initial second-largest model refers to large-scale language models that have already been pre-trained on large-scale general corpora, such as the GPT series, BERT, and LLaMA. These models possess basic natural language understanding and generation capabilities, but have not yet been optimized for the task of extracting style features from social e-commerce promotional copy. Predicting tone labels involves the second-largest model predicting the tone type of the input historical promotional copy, such as formal, colloquial, or humorous. Predicting sentiment labels is the model's prediction of the sentiment expressed in the copy, such as positive, negative, or neutral. Predicting sentence structure labels is the model's judgment of the sentence structure characteristics of the copy, such as predominantly short sentences, complex long sentences, or parallel sentence structures. During general pre-training, the initial second-largest model learns extensive language knowledge and semantic understanding capabilities. When historical promotional copy is input into the model, it performs semantic analysis, feature extraction, and inference based on its internal parameters and learned language patterns, thereby outputting corresponding predicted labels. This process is similar to the model predicting style features of new input data based on its existing knowledge base. For example, if you input a lively beauty product promotion copy from the Douyin platform into the model, the model will analyze the words and sentences in the copy and output predictive tone tags such as "colloquial", "positive" predictive sentiment tags, and "short sentences as the main type" predictive sentence structure tags.

[0141] S704, the first loss term is obtained based on the difference between the tone label and the predicted tone label, the second loss term is obtained based on the difference between the sentiment label and the predicted sentiment label, and the third loss term is obtained based on the difference between the sentence structure label and the predicted sentence structure label.

[0142] It's understandable that tone, sentiment, and sentence structure labels are labels assigned to historical promotional copy by humans or through specific algorithms when annotating it. These labels represent the actual style attributes of the copy and serve as the standard and reference for model learning. The first, second, and third loss terms measure the degree of deviation between the model's predicted tone, sentiment, and sentence structure labels and the actual labels, respectively. They are quantitative indicators for evaluating the model's prediction accuracy. The loss function measures the difference between the model's prediction and the actual results, and the model parameters are optimized by minimizing the loss function. For each historical promotional copy, the differences between its predicted and actual labels in the three dimensions of tone, sentiment, and sentence structure are calculated. Common loss function calculation methods include cross-entropy loss function and mean squared error loss function. Taking the cross-entropy loss function to calculate the first loss term as an example, assuming that tone labeling is a classification problem (e.g., divided into 5 tone types), the predicted tone label output by the model is a probability distribution vector. The cross-entropy loss function calculates the difference between this probability distribution and the probability distribution of the actual tone label (usually in one-hot encoding form) to obtain the value of the first loss term. Similarly, calculate the second and third loss terms. These loss terms reflect the degree of error of the model in predicting different style features; the smaller the value, the more accurate the model prediction.

[0143] S706, the first loss term, the second loss term, and the third loss term are weighted and summed to obtain the objective function.

[0144] The objective function is understood to be a comprehensive measure of the model's overall error in predicting tone labels, sentiment labels, and sentence structure labels. It is obtained by weighted summation of the three loss terms and serves to guide the optimization of model parameters. This step is based on multi-objective optimization theory. In the task of extracting style features from social e-commerce promotional copy, the model needs to accurately predict multiple dimensions of features, including tone, sentiment, and sentence structure. Therefore, the loss terms of these three dimensions are integrated to construct a unified objective function for overall model optimization. The weight of each loss term represents the relative importance of its corresponding style feature in the task and can be set according to actual business needs and experience. For example, if sentiment labels are considered to have the greatest impact on promotional effectiveness, the second loss term can be given a higher weight; if sentence structure is more critical in certain promotional channels, the weight of the third loss term can be increased. By weighted summation, the three loss terms are merged into a single value. The smaller this value, the better the model's overall performance in predicting each style feature.

[0145] S708, with the goal of reducing the objective function, fine-tunes the parameters of the second large model, and returns the steps of obtaining the first loss term based on the difference between the tone label and the predicted tone label, the second loss term based on the difference between the sentiment label and the predicted sentiment label, and the third loss term based on the sentence structure label and the predicted sentence structure label, until the fine-tuning termination condition is met.

[0146] Fine-tuning, as we understand it, refers to the process of optimizing and improving the performance of a model on a specific task (i.e., the task of extracting style features from social e-commerce promotional copy) by adjusting the model's parameters (such as weights and biases in a neural network) based on the initial second-largest model. The termination condition for fine-tuning is a pre-defined standard used to determine whether the model training has reached an ideal state and whether training can be stopped. Common conditions include reaching a preset number of training epochs, the objective function value falling below a certain threshold, and the model's performance on the validation set no longer improving. The principle is based on gradient descent optimization algorithms and iterative learning theory. With the goal of reducing the objective function, gradient descent algorithms (such as stochastic gradient descent (SGD) and adaptive moment estimation algorithms like Adam) are used to calculate the gradient of the objective function with respect to the model parameters. The model parameters are adjusted according to the gradient direction, causing the objective function value to gradually decrease. After each parameter adjustment, the process returns to step S702, where historical promotional copy is re-inputted into the model to obtain predicted labels. Then, the loss term is calculated, the objective function is constructed, and the parameters are adjusted again, repeating this iterative cycle. During this process, the model continuously learns the mapping relationship between historical promotional copy and style feature labels, gradually correcting its prediction bias and improving its ability to extract style features of social e-commerce promotional copy. When the fine-tuning termination condition is met, it indicates that the model has reached a good performance level on the current task, and the training process ends.

[0147] In one embodiment, the objective function is obtained by weighted summation of the first, second, and third loss terms. The method further includes: for any one of the first, second, and third loss terms, if the rate of change for a first number of consecutive periods is lower than a second threshold, the corresponding weight is reduced. For any one of the first, second, and third loss terms, if the absolute value is greater than the third threshold, the corresponding weight is increased. It can be understood that the rate of change describes the magnitude of change of the loss term between adjacent periods. The first number is a pre-set threshold for the number of consecutive periods used to judge the stability of the loss term's change. The second threshold is a critical value for measuring the magnitude of the rate of change of the loss term, used to determine whether the change of the loss term tends to be stable. The learning difficulty and importance of different style features (tone, emotion, sentence structure) will change as the training process progresses. When the rate of change of a certain loss term is lower than the second threshold for several consecutive periods, it indicates that the model has tended to stabilize in predicting that style feature, and the learning effect improves slowly. At this point, lowering the corresponding weight can appropriately reduce the influence of this loss term in the objective function, guiding the model to allocate more learning resources to other loss terms that have not yet been fully optimized. This avoids overfitting to a particular style feature and achieves balanced optimization of the model in multi-style feature prediction tasks. For example, in the early stages of training, the model's prediction error for sentiment labels is relatively large. As training progresses, the rate of change of the second loss term (sentiment label prediction error) is less than 0.01 for 10 consecutive cycles (the first threshold is set to 10) (the second threshold is set to 0.01), indicating that the model has become relatively stable in terms of sentiment label prediction. At this point, lowering the weight of the second loss term encourages the model to strengthen its ability to predict tone and sentence structure labels.

[0148] The third threshold is a critical value used to measure the absolute value of the loss term, determining whether the model's error in predicting a certain style feature is too large. This step is based on the principle of prioritizing key problems in optimization. When the absolute value of a loss term exceeds the third threshold, it indicates a significant error in predicting the corresponding style feature (tone, sentiment, or sentence structure). Increasing the weight of this loss term increases its proportion in the objective function, making the model focus more on optimizing this style feature during subsequent parameter adjustments and increasing the correction of this part of the error. For example, if the absolute value of the third loss term (sentence structure label prediction error) reaches 0.8 (with the third threshold set to 0.5), it indicates poor performance in sentence structure prediction. Increasing the weight of the third loss term will make the objective function more focused on reducing this part of the error, guiding the model to learn more accurate sentence structure features, thereby improving performance on the sentence structure label prediction task. This step works in conjunction with other steps to adjust the optimization direction of the objective function by assigning higher weights to loss terms with larger errors. This ensures that the model can promptly identify and resolve key issues during fine-tuning, accelerates the overall convergence speed of the model on multi-style feature prediction tasks, and improves the overall performance of the model.

[0149] This application provides a product selection and promotion device for social e-commerce, including a profile generation module, a product matching module, a target product determination module, and a push module. The profile generation module generates a user profile for the target user based on their historical promotion information. The product matching module matches any newly launched product with the user profile based on its description. The target product determination module identifies the newly launched product as the target product if a match is successful. The push module pushes the target product to the target user.

[0150] This application provides a computer device including one or more processors and a memory storing computer-readable instructions. When the computer-readable instructions are executed by one or more processors, they perform the steps of the product selection and promotion method in any of the above embodiments.

[0151] Specific limitations regarding the product selection and promotion device can be found in the limitations of the product selection and promotion method described above, and will not be repeated here. Each module in the aforementioned product selection and promotion device can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device in hardware form, or stored in the memory of a computer device in software form, so that the processor can call and execute the operations corresponding to each module. It should be noted that the module division in this embodiment is illustrative and only represents a logical functional division; other division methods may be used in actual implementation.

[0152] Indicatively, such as Figure 8 As shown, Figure 8 This is a schematic diagram of the internal structure of a computer device provided in an embodiment of this application. (Refer to...) Figure 8 The computer device 800 includes a processing component 802, which further includes one or more processors, and memory resources represented by memory 801 for storing instructions, such as application programs, that can be executed by the processing component 802. The application programs stored in memory 801 may include one or more modules, each corresponding to a set of instructions. Furthermore, the processing component 802 is configured to execute instructions to perform the steps of the product selection promotion method of any of the above embodiments.

[0153] The computer device 800 may also include a power supply component 803 configured to perform power management of the computer device 800, a wired or wireless model interface 804 configured to connect the computer device 800 to a model, and an input / output (I / O) interface 805.

[0154] This application 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 product selection and promotion method in any of the above embodiments.

[0155] Finally, it should be noted that in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0156] The various embodiments in this specification are described in a progressive manner. Each embodiment focuses on the differences from other embodiments. The various embodiments can be combined as needed, and the same or similar parts can be referred to each other.

[0157] The above description of the disclosed embodiments enables those skilled in the art to make or use this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, this application is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A product selection and promotion method for social e-commerce, characterized in that, include: Generate a user profile of the target user based on the target user's historical promotional information set; After any product is launched, it is matched with the user profile based on the product description information; If the match is successful, the newly launched product will be identified as the target product. Push the target product to the target user; The step of pushing the target product to the target user includes: Based on the target tag group that matches the target product, determine one or more target promotion channels; Based on the style prompts corresponding to the target promotion channels and the product description information, channel copywriting generation prompts are obtained; Input the prompt words generated from the channel copy into the first model to obtain the promotion copy corresponding to the target promotion channel; The promotional link of the target product, the target promotional channel, and the corresponding promotional copy are packaged and sent to the target user; The style prompts correspond one-to-one with the available promotion channels, and the generation process of each style prompt includes: Obtain the set of promotional copy for the corresponding optional promotional channels; Style extraction prompts are generated based on the set of promotional copy; these style extraction prompts are used to instruct the second model to generate style prompts based on the set of promotional copy. The extracted style prompts are input into the second model to obtain the corresponding style prompts. The pre-training process of the second major model includes: Obtain a set of historical promotional copy corresponding to the optional promotion channels, and annotate each historical promotional copy to obtain the corresponding standard style prompt words; the standard style prompt words include tone analysis items, sentiment analysis items and sentence structure analysis items, and the annotated historical promotional copy includes tone tags, sentiment tags and sentence structure tags; Each of the aforementioned historical promotional copy texts is input into the initial second large model, and the second large model is instructed to output the corresponding predicted tone label, predicted sentiment label, and predicted sentence structure label; A first loss term is obtained based on the difference between the tone label and the predicted tone label; a second loss term is obtained based on the difference between the sentiment label and the predicted sentiment label; and a third loss term is obtained based on the difference between the sentence structure label and the predicted sentence structure label. The objective function is obtained by weighted summation of the first loss term, the second loss term, and the third loss term; With the objective function as the goal, the parameters of the second large model are fine-tuned, and the steps of obtaining the first loss term based on the difference between the tone label and the predicted tone label, obtaining the second loss term based on the difference between the sentiment label and the predicted sentiment label, and obtaining the third loss term based on the sentence structure label and the predicted sentence structure label are returned, until the fine-tuning termination condition is met.

2. The product selection and promotion method according to claim 1, characterized in that, The historical promotion information set includes multiple historical promotion information entries, which contain multi-dimensional recorded data. Generating a user profile for the target user based on the target user's historical promotion information set includes: For any of the aforementioned historical promotional information, each of the recorded data will be converted into a promotional tag; The historical promotional information with the same promotional tag is grouped into a promotional tag group; Determine the promotion success rate for each of the aforementioned promotional tag groups; Select the target tag group from the promotion tag group based on the promotion success rate described above; The user profile is generated based on the target tag group.

3. The product selection and promotion method according to claim 2, characterized in that, The determination of the promotion success rate for each of the aforementioned promotion tag groups includes: Obtain the total number of executions for all historical promotional information within the aforementioned promotional tag group; Each completed order is assigned a time decay weight to its historical promotional information. The greater the time difference between the current time and the order placement time of the completed order's historical promotional information, the smaller the time decay weight. The weighted success rate is obtained by calculating the number of historical promotional messages that result in a single order based on the time decay weight. The weighted ratio of the number of successful weighted results to the total number of executions is defined as the promotion success rate.

4. The product selection and promotion method according to claim 2, characterized in that, The dimensions of the recorded data include average order value, promotion channels, product categories, whether an order is completed, order time, and commission rate.

5. The product selection and promotion method according to claim 1, characterized in that, Also includes: Update the historical promotion information set based on the real-time promotion records of the target users; If the conditions for updating the user profile are met, the user profile is updated based on the updated set of historical promotional information.

6. The product selection and promotion method according to claim 5, characterized in that, The step of updating the historical promotion information set based on the real-time promotion records of the target user includes: After the promotional link corresponding to the target user is clicked, the operation record after the promotional link is clicked is monitored, and a historical promotional information is generated based on the operation record and updated to the historical promotional information set.

7. The product selection and promotion method according to claim 2, characterized in that, The matching of product description information with the user profile includes: Product tags are extracted from the product description information to obtain multiple product feature keywords; The product feature keywords are converted into product tags to obtain a set of product tags; Perform similarity matching between the product tag set and each of the target tag groups; If any similarity score is greater than the first threshold, the match is considered successful. Otherwise, the match is deemed to have failed.

8. The product selection and promotion method according to claim 1, characterized in that, The step of weighted summing of the first loss term, the second loss term, and the third loss term to obtain the objective function further includes: For any one of the first loss item, the second loss item, and the third loss item, if the rate of change for a first number of consecutive periods is lower than the second threshold, then the corresponding weight is reduced. For any one of the first loss term, the second loss term, and the third loss term, if the absolute value is greater than the third threshold, the corresponding weight is increased.

9. A product selection and promotion device for social e-commerce, characterized in that, The apparatus is used to implement the product selection and promotion method as described in any one of claims 1-8, the apparatus comprising: The profile generation module is used to generate a user profile of the target user based on the target user's historical promotion information set; The product matching module is used to match any product with the user profile based on the product description information after the product is launched. The target product determination module is used to identify the newly launched product as the target product if a match is successful. The push module is used to push the target product to the target user.

10. A computer device, characterized in that, It includes one or more processors and a memory storing computer-readable instructions, which, when executed by the one or more processors, perform the steps of the product selection and promotion method as described in any one of claims 1-8.

11. A storage medium, characterized in that, The storage medium stores computer-readable instructions, which, when executed by one or more processors, cause the one or more processors to perform the steps of the product selection and promotion method as described in any one of claims 1-8.

Images