User circle selection method and related product
By converting natural language instructions into query vectors in the user selection method and performing semantic retrieval in the user vector library, combined with user behavior data and brand knowledge base, the problem of missing tag semantics and data sparsity in traditional methods is solved, and more accurate user selection is achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- XIAMEN NANXUN CO LTD
- Filing Date
- 2026-01-29
- Publication Date
- 2026-05-01
AI Technical Summary
Existing user selection methods rely on traditional ETL labeling systems or collaborative filtering algorithms, which suffer from problems such as missing label semantics and sparse behavioral data, resulting in low accuracy of user selection results.
By acquiring user selection instructions in natural language form, converting them into query vectors, and performing semantic retrieval in a user vector library, which is built based on user behavior data and brand knowledge base, and using a large language model and brand knowledge base for weighted fusion to improve semantic consistency matching.
It significantly improves the accuracy of user selection results, especially in cold start and behavioral data sparse situations, and can accurately recall target users, eliminating the semantic gap between business language and traditional selection rules, and the results are closer to the true business intent.
Smart Images
Figure CN121958313A_ABST
Abstract
Description
A user selection method and related products Technical Field
[0001] This application relates to the field of data processing technology, and in particular to a user selection method and related products. Background Technology
[0002] In the e-commerce and retail industry, user segmentation is the core foundation for achieving precise marketing, improving user conversion rates, and optimizing operational strategies.
[0003] Existing user segmentation technologies mainly rely on traditional Extract-Transform-Load (ETL) tagging systems or collaborative filtering algorithms. These methods collect user behavior data, such as browsing, clicking, or purchasing, to construct user profile tags, such as "purchased more than 3 times in the last 30 days" or "active in the last 7 days," and use these user profile tags as the basis for user segmentation and marketing decisions. However, there are certain shortcomings in practical applications: (1) the semantics of the tags are missing, making it difficult to capture the user's deeper intentions; (2) the behavioral data is sparse, making it difficult to form accurate user profiles. These shortcomings result in low accuracy of user segmentation results.
[0004] Improving the accuracy of user selection results is a technical problem that urgently needs to be solved. Summary of the Invention
[0005] To address the aforementioned issues, this application provides a user selection method and related products, aiming to improve the accuracy of user selection results.
[0006] This application discloses the following technical solutions: In its first aspect, this application provides a user selection method, which includes: obtaining a user selection instruction in natural language form and converting the user selection instruction into a query vector; performing semantic retrieval in a user vector library based on the query vector to obtain a target user set; wherein the user vectors in the user vector library are constructed based on user behavior data and a brand knowledge base.
[0007] Optionally, the brand knowledge base is constructed by: for any brand: clustering based on the brand's historical user behavior data to obtain multiple central vectors; using each central vector as a brand behavior template for the corresponding user group; and constructing a brand knowledge base based on the multiple brand behavior templates.
[0008] Optionally, the construction of the user vector library includes: collecting behavioral data from multiple users; the behavioral data includes browsing data, add-to-cart data, order data, return / exchange data, and customer dialogue data; for any user's behavioral data: using a large language model to convert the user's behavioral data into user notes in natural language form, and converting the user notes into an initial feature vector; weightedly fusing the initial feature vector and the brand behavior template fusion vector to obtain the user-enhanced feature vector; the brand behavior template fusion vector is obtained from the initial feature vector based on a brand knowledge base; and the user vector library is obtained based on the enhanced feature vectors of multiple users.
[0009] Optionally, the method for determining the brand behavior template fusion vector includes: calculating the cosine similarity between the initial feature vector and each brand behavior template in the brand knowledge base to obtain multiple similarity scores; normalizing the multiple similarity scores to obtain multiple weight values that correspond one-to-one with the multiple similarity scores; and performing weighted fusion on the multiple brand behavior templates in the brand knowledge base based on the multiple weight values to obtain the brand behavior template fusion vector.
[0010] Optionally, the step of performing semantic retrieval in the user vector library based on the query vector to obtain the target user set includes: performing semantic retrieval in the user vector library based on the query vector using a marketing conversion prediction model to obtain the target user set; the marketing conversion prediction model is obtained after training on an industry knowledge base; the industry knowledge base is constructed based on historical industry experience.
[0011] A second aspect of this application provides a user selection device, comprising: a user selection instruction acquisition module, configured to acquire user selection instructions in natural language form and convert the user selection instructions into query vectors; and a user selection module, configured to perform semantic retrieval in a user vector library based on the query vectors to obtain a target user set; wherein the user vectors in the user vector library are constructed based on user behavior data and a brand knowledge base.
[0012] Optionally, the construction of the user vector library includes: collecting behavioral data from multiple users; the behavioral data includes browsing data, add-to-cart data, order data, return / exchange data, and customer dialogue data; for any user's behavioral data: using a large language model to convert the user's behavioral data into user notes in natural language form, and converting the user notes into an initial feature vector; weightedly fusing the initial feature vector and the brand behavior template fusion vector to obtain the user-enhanced feature vector; the brand behavior template fusion vector is obtained from the initial feature vector based on a brand knowledge base; and the user vector library is obtained based on the enhanced feature vectors of multiple users.
[0013] Optionally, the brand knowledge base is constructed by: for any brand: clustering based on the brand's historical user behavior data to obtain multiple central vectors; using each central vector as a brand behavior template for the corresponding user group; and constructing a brand knowledge base based on the multiple brand behavior templates.
[0014] A third aspect of this application provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the user selection method provided in any implementation of the first aspect.
[0015] The fourth aspect of this application provides a processor for running a computer program that, when running, executes a user selection method as provided in any implementation of the first aspect.
[0016] Compared to existing technologies, this application offers the following advantages: The user selection method provided in this application includes: obtaining user selection instructions in natural language form; converting the user selection instructions into query vectors; and performing semantic retrieval in a user vector library based on the query vectors to obtain a target user set. The user vectors in the user vector library are constructed based on user behavior data and a brand knowledge base. This method converts user selection instructions in natural language form into query vectors and performs semantic retrieval in the user vector library. It matches the target user set based on the semantic consistency of the user selection instructions and user vectors, eliminating the semantic gap between business language and traditional selection rules. This makes the user selection results closer to the true intent of the business and improves the accuracy of the user selection results. By constructing user vectors based on user behavior data and brand knowledge, even when user behavior data is sparse, the integration of brand knowledge into the user behavior data compensates for the sparsity of the behavior data, enabling accurate recall of users with sparse behavior data during semantic retrieval, significantly improving the accuracy of user selection results in cold start situations. Attached Figure Description
[0017] 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.
[0018] Figure 1 is a flowchart of a user selection method provided in an embodiment of this application; Figure 2 is a flowchart of another user selection method provided in an embodiment of this application; Figure 3 is a structural schematic diagram of a user selection device provided in an embodiment of this application. Detailed Implementation
[0019] As described above, current user segmentation mainly relies on traditional ETL tagging systems or collaborative filtering algorithms. These methods collect user behavior data, such as browsing, clicking, or purchasing, to construct user profile tags, such as "purchased more than 3 times in the last 30 days" or "active in the last 7 days," and use these user profile tags as the basis for user segmentation and marketing decisions. However, there are certain shortcomings in practical applications: (1) the semantics of the tags are missing, making it difficult to capture the user's deeper intentions; (2) the behavioral data is sparse, making it difficult to form accurate user profiles. These shortcomings result in low accuracy of user segmentation results.
[0020] In view of the above problems, this application proposes a user selection method and related products, which obtain user selection instructions in natural language form, convert the user selection instructions into query vectors, and perform semantic retrieval in a user vector library based on the query vectors to obtain a target user set; the user vectors in the user vector library are constructed based on user behavior data and brand knowledge base.
[0021] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present application, and not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present application.
[0022] Referring to Figure 1, which is a flowchart of a user selection method provided in an embodiment of this application. As shown in Figure 1, the method includes the following steps: S101, obtaining user selection instructions in natural language form, and converting the user selection instructions into a query vector.
[0023] For example, the user selection command can be used to select high-potential customers who have anti-aging needs in the past three months and are not price-sensitive. User selection supports natural language interaction, allowing business personnel to complete complex and highly accurate user selection without needing to understand coding.
[0024] Converting user selection instructions into query vectors can be achieved using pre-trained language models, such as BERT or Sentence-BERT.
[0025] S102. Based on the query vector, perform semantic retrieval in the user vector database to obtain the target user set.
[0026] The user vectors in the user vector library are constructed based on user behavior data and brand knowledge base.
[0027] For example, each user vector in the user vector library is an enhanced user representation generated by modeling the multi-dimensional behavioral feature data of the corresponding user and integrating brand behavior template knowledge obtained from brand historical customer clustering. User vectors not only reflect individual behavior but also incorporate brand-level group cognition, significantly improving the selection accuracy in cold start and sparse scenarios.
[0028] The user selection method provided in this application includes: obtaining user selection instructions in natural language form and converting the user selection instructions into query vectors; performing semantic retrieval in a user vector library based on the query vectors to obtain a target user set; the user vectors in the user vector library are constructed based on user behavior data and a brand knowledge base. This method converts user selection instructions in natural language form into query vectors and performs semantic retrieval in the user vector library. It matches the target user set based on the semantic consistency of the user selection instructions and user vectors, eliminating the semantic gap between business language and traditional selection rules, making the user selection results closer to the true intent of the business and improving the accuracy of the user selection results. By constructing user vectors based on user behavior data and brand knowledge, even when user behavior data is relatively sparse, the method compensates for the sparsity of the behavior data by integrating brand knowledge into the user behavior data, still accurately recalling users with sparse behavior data during semantic retrieval, significantly improving the accuracy of user selection results in cold start situations.
[0029] To further improve the user selection method described in the above embodiments, a step of constructing a user vector library is added based on the above embodiments. The method of obtaining the target user set by performing semantic retrieval in the user vector library based on the query vector is further refined into obtaining the target user set by performing semantic retrieval in the user vector library based on the query vector using a marketing conversion prediction model.
[0030] Referring to Figure 2, this figure is a flowchart of another user selection method provided in an embodiment of this application. As shown in Figure 2, the method includes the following steps: S201, constructing a user vector library.
[0031] In one feasible implementation: collect behavioral data from multiple users; the behavioral data includes browsing data, add-to-cart data, order data, return / exchange data, and customer conversation data.
[0032] For any user's behavior data: the user's behavior data is converted into user notes in natural language form using a large language model, and the user notes are then converted into an initial feature vector.
[0033] The initial feature vector and the brand behavior template fusion vector are weighted and fused to obtain the user-enhanced feature vector; the brand behavior template fusion vector is obtained from the initial feature vector based on the brand knowledge base.
[0034] The user vector library is obtained based on the enhanced feature vectors of multiple users.
[0035] Data sources for collecting behavioral data from multiple users include Customer Relationship Management (CRM) order flows, event tracking logs, and instant messaging (IM) records. CRM order flows record user purchasing behavior, such as order time, product, amount, and frequency, reflecting spending power and preferences. Event tracking logs collect user actions on applications or websites through front-end code, such as browsing pages, clicking buttons, and adding items to cart, revealing interest patterns. Customer IM records contain chat content between users and customer service, including implicit intent information such as sentiment, complaint feedback, and price sensitivity.
[0036] The collected user behavior data undergoes time decay processing. Specifically, for each behavior event in the user behavior sequence, the time decay is calculated based on the interval between the behavior occurrence time and the current time. Calculate the time decay weight The formula is as follows: ;in, represents the attenuation factor, for example, it can be 0.05; e is the natural constant.
[0037] The large language model is used to convert user behavior data into user notes in natural language form. For example, the prompt word template provided to the large language model is: a user-identified behavior sequence, ordered chronologically in descending order, generating a profile summary of no more than 200 characters, focusing on price sensitivity, category preference, and potential churn risk. User notes can be converted into initial feature vectors using embedding models, such as BGE-M3 or OpenAI-Embeddings.
[0038] For example, the initial feature vector and the brand behavior template fusion vector are weighted and fused to obtain the user-enhanced feature vector. The calculation formula is as follows: ;in, This represents the enhanced feature vector of the user. Represents the initial eigenvector; This represents a vector for fusing brand behavior templates. This represents the fusion coefficient, which is determined by the sparsity of user behavior data. The less user behavior data, the lower the fusion coefficient. The larger the value, the more it relies on general brand knowledge, and the richer the user behavior data. The smaller the value, the more dependent it is on its own behavior.
[0039] The dynamic weighting mechanism described above allows new users with only a few behaviors to leverage the behavioral templates of established brand customers, significantly improving the accuracy of cold start recommendations.
[0040] In one feasible implementation, the construction of the brand knowledge base includes: for any brand, clustering based on the brand's historical user behavior data to obtain multiple center vectors.
[0041] Each central vector serves as a brand behavior template for the corresponding user group.
[0042] A brand knowledge base is built based on multiple brand behavior templates.
[0043] For example, behavioral data of a brand's historical high-value customers are obtained, including browsing data, add-to-cart data, purchase data, and customer service interaction data. K-Means clustering analysis is performed on the behavioral data of the brand's historical high-value customers to generate K customer group clusters. The central vector of each customer group cluster is extracted, and each central vector is defined as a brand behavior template. Based on multiple brand behavior templates, a brand knowledge base is constructed.
[0044] The values of the central vector across each behavioral dimension reflect the typical behavioral patterns of the corresponding user group. Each central vector is defined as a brand behavior template, such as a high-repurchase-rate mother and baby group template or a large-scale promotion stockpiling group template. Each template represents a high-value user group with similar consumption behaviors and can be seen as a general knowledge or experience model for the brand. All brand behavior templates together constitute the brand knowledge base, which is used for subsequent user profile enhancement.
[0045] In one feasible implementation, the method for determining the brand behavior template fusion vector includes: calculating the cosine similarity between the initial feature vector and each brand behavior template in the brand knowledge base to obtain multiple similarity scores.
[0046] The multiple similarity scores are normalized to obtain multiple weight values that correspond one-to-one with the multiple similarity scores.
[0047] Based on the multiple weight values, multiple brand behavior templates in the brand knowledge base are weighted and fused to obtain the brand behavior template fusion vector.
[0048] For example, the formula for calculating the brand behavior template fusion vector is as follows: ;in, This represents a vector for fusing brand behavior templates. This represents the i-th brand behavior template; This represents the similarity score between the initial feature vector and the i-th brand behavior template; Indicates will The activation function is transformed into a probability distribution.
[0049] By constructing brand behavior templates through clustering and then combining sparse user behavior with general brand knowledge through similarity alignment and dynamic weighted fusion, the quality of user profiles is improved. This allows new users with only a small amount of behavioral data to also benefit from brand behavior templates, significantly improving the accuracy of cold start recommendations.
[0050] S202. Obtain the user selection instruction in natural language form and convert the user selection instruction into a query vector.
[0051] It supports natural language interaction, allowing business personnel to complete complex and highly accurate user selection without needing to understand code.
[0052] S203. Based on the query vector, a marketing conversion prediction model is used to perform semantic retrieval in the user vector database to obtain the target user set.
[0053] The marketing conversion prediction model is obtained by training on an industry knowledge base, which is built based on historical industry experience.
[0054] For example, historical marketing data from an enterprise's internal Sales Force Automation (SFA) system is used as real business feedback. For each user, a binary label is constructed: y=1 indicates that the user successfully converted after the marketing campaign; y=0 indicates that the user did not convert. This label constitutes the supervisory signal for training the marketing conversion prediction model.
[0055] The loss function is optimized, consisting of two parts: semantic matching loss and cross-entropy loss between the predicted conversion rate and the industry benchmark conversion rate. The formula is as follows: ;in, This represents the total loss function; Indicates semantic matching loss; This represents the cross-entropy loss between the predicted conversion rate and the industry benchmark conversion rate; This represents the weighting coefficient.
[0056] By using historical industry experience as a supervisory signal, the user vectors generated by the marketing conversion prediction model are forced to be not only semantically relevant but also conform to prior patterns in industry conversion. Moreover, the introduction of SFA industry experience as supervision ensures that the model output is highly consistent with historical successful marketing cases, exhibiting strong interpretability.
[0057] To further illustrate the effectiveness of the user selection method provided in this embodiment, an experiment was conducted using historical datasets from cosmetics brand A and maternal and infant brand B as examples. The datasets were specifically designed to handle over 500,000 users, over 2,000 products, and over 12 million behavioral records. The validation set consisted of actual conversion data from 100 completed marketing campaigns within the company's SFA system.
[0058] The user selection instructions obtained in natural language form, such as selecting high-potential customers who have anti-aging needs in the past 3 months and are not price-sensitive, are compared with the overlap between the target user set included in the selection results and the users who actually made purchases. The comparison results are shown in Table 1.
[0059] Table 1
[0060] According to the comparison results shown in Table 1, Precision is the percentage of target users among the selected users. The closer the Precision value is to 1, the better. Recall is the percentage of target users who are successfully selected. The closer the Recall value is to 1, the better. F1-Score is the harmonic mean of Precision and Recall, which is used to comprehensively measure the model performance. Baseline 1 represents a selection method based on traditional rule-based tags. This method relies on manually defined rules, making it difficult to capture users' deeper intentions and resulting in low overall accuracy. Baseline 2 represents a selection method based solely on user behavior vectors. While modeling based on user behavior data improves semantic understanding, it still lacks contextual knowledge and cannot distinguish between high-value users and ordinary users. Baseline 3 represents a selection method based on user behavior and brand knowledge. This method can identify potential high-value users even when user behavior is sparse, and its selection results are significantly better than Baseline 1 and Baseline 2. The user selection method based on user behavior, brand knowledge, and industry experience provided in this application significantly outperforms other methods in various metrics. It can not only understand complex business intentions but also accurately identify high-value customers who are both interested and easily converted by combining brand knowledge and industry experience.
[0061] To verify the contribution of industry experience, an ablation experiment was also conducted in this embodiment. The results of the ablation experiment are shown in Table 2.
[0062] Table 2
[0063] Based on the ablation experiment results shown in Table 2, the AUC (Area Under ROC Curve) value measures the model's ability to distinguish between positive and negative samples; the closer to 1, the more accurate the recommendation results. The NDCG@10 (Normalized Discounted Cumulative Gain at 10) value measures the quality of the top 10 results in the recommendation list, taking into account the weight of the ranking position; the earlier the ranking, the more important it is, reflecting whether "good products are ranked first." The AUC value obtained from the complete marketing conversion prediction model is the highest, at 0.845. When industry experience is removed, the AUC value drops to 0.792, indicating that industry experience is helpful in judging user intent. After removing brand knowledge fusion, it drops to 0.755, indicating that the recommendation quality for cold-start users drops significantly after removal. The NDCG@10 of the complete model is 0.762. After removing industry experience validation, NDCG@10 drops to 0.710, indicating that the recommendation order deviates more from actual conversion behavior. After removing brand knowledge fusion, NDCG@10 drops to 0.675, indicating that the original behavior is sparse for new users or low-frequency users, making effective modeling impossible. Ablation experiments verified that industry experience and brand knowledge directly improved the AUC and NDCG@10 values of the final user selection results, ensuring that the user selection results were not only semantically relevant but also easy to convert.
[0064] Furthermore, in actual marketing activities, users are randomly divided into a control group and an experimental group. The control group uses the results of various indicators obtained based on traditional tagging methods, while the experimental group uses the results of various indicators obtained based on the user selection method provided in this application embodiment. Each group includes 50,000 users. Taking SMS marketing channels as an example, the experimental results are shown in Table 3.
[0065] Table 3
[0066] Based on the results of the above comparative experiment, from the perspective of SMS click-through rate, users in the experimental group were more interested in the SMS content, indicating that the users selected by the user selection method provided in this application embodiment had a higher degree of matching with the marketing theme and stronger message relevance. From the perspective of conversion rate, the proportion of users in the experimental group who actually placed orders after clicking was significantly higher than that in the control group, indicating that the users selected in the experimental group had stronger purchasing intentions and decision-making abilities, that is, high-value users were effectively identified. From the perspective of average order value, the selected users in the experimental group had higher average spending, and potential groups with high repurchase rates and high average order values could be identified, indicating that the selection results of the experimental group were not only numerous but also of high quality. From the perspective of return on investment, with the same input, the return of the experimental group was significantly improved, indicating that in actual business, the user selection method provided in this application embodiment can significantly improve the accuracy of marketing.
[0067] Another user selection method provided in this application includes: constructing a user vector library; obtaining user selection instructions in natural language form and converting the user selection instructions into query vectors; based on the query vectors, performing semantic retrieval in the user vector library using a marketing conversion prediction model to obtain a target user set; the user vectors in the user vector library are constructed based on user behavior data and a brand knowledge base; the marketing conversion prediction model is obtained after training on an industry knowledge base; and the industry knowledge base is constructed based on historical industry experience.
[0068] This method converts user selection instructions in natural language into query vectors and performs semantic retrieval in a user vector database. Based on the semantic consistency between the user selection instructions and user vectors, it matches the target user set, bridging the semantic gap between business language and traditional selection rules. This makes the user selection results closer to the true business intent, improving the accuracy of the user selection results. Furthermore, by constructing user vectors based on user behavior data and brand knowledge, even when user behavior data is sparse, brand knowledge is integrated into the user behavior data to compensate for the sparsity. This still allows for accurate recall of users with sparse behavior data during semantic retrieval, significantly improving the accuracy of user selection results in cold start scenarios.
[0069] Based on the user selection method described in the preceding embodiments, this application also provides a user selection device. Figure 3 is a schematic diagram of the device. As shown in Figure 3, the user selection device includes: a user selection instruction acquisition module 301, used to acquire user selection instructions in natural language form and convert the user selection instructions into a query vector.
[0070] The user selection module 302 is used to perform semantic retrieval in the user vector library based on the query vector to obtain a set of target users; the user vectors in the user vector library are constructed based on user behavior data and brand knowledge base.
[0071] Optionally, the brand knowledge base is constructed by: for any brand, clustering based on the brand's historical user behavior data to obtain multiple center vectors.
[0072] Each central vector serves as a brand behavior template for the corresponding user group.
[0073] A brand knowledge base is built based on multiple brand behavior templates.
[0074] Optionally, the user vector library is constructed by collecting behavioral data from multiple users; the behavioral data includes browsing data, add-to-cart data, order data, return / exchange data, and customer conversation data.
[0075] For any user's behavior data: the user's behavior data is converted into user notes in natural language form using a large language model, and the user notes are then converted into an initial feature vector.
[0076] The initial feature vector and the brand behavior template fusion vector are weighted and fused to obtain the user-enhanced feature vector; the brand behavior template fusion vector is obtained from the initial feature vector based on the brand knowledge base.
[0077] The user vector library is obtained based on the enhanced feature vectors of multiple users.
[0078] Optionally, the method for determining the brand behavior template fusion vector includes: calculating the cosine similarity between the initial feature vector and each brand behavior template in the brand knowledge base to obtain multiple similarity scores.
[0079] The multiple similarity scores are normalized to obtain multiple weight values that correspond one-to-one with the multiple similarity scores.
[0080] Based on the multiple weight values, multiple brand behavior templates in the brand knowledge base are weighted and fused to obtain the brand behavior template fusion vector.
[0081] Optionally, the user selection module is used to: based on the query vector, perform semantic retrieval in the user vector database using a marketing conversion prediction model to obtain a target user set; the marketing conversion prediction model is obtained after training on an industry knowledge base; the industry knowledge base is constructed based on historical industry experience.
[0082] Furthermore, embodiments of this application also provide a computer-readable storage medium storing a computer program that, when executed by a processor, implements the user selection method as described in any of the method embodiments.
[0083] Furthermore, this application embodiment also provides a processor for running a computer program, which executes the user selection method as described in any of the foregoing method embodiments.
[0084] It should be noted that the various embodiments in this specification are described in a progressive manner, and the same or similar parts between the various embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, for the device embodiments, since they are basically similar to the method embodiments, the description is relatively simple, and the relevant parts can be referred to the description of the method embodiments. The device embodiments described above are merely illustrative, and the units described as separate components may or may not be physically separate. The components indicated as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment solution according to actual needs. Those skilled in the art can understand and implement this without creative effort.
[0085] The above description is merely one specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.
Claims
1. A user selection method, characterized in that, include: Obtain the user selection instruction in natural language form, and convert the user selection instruction into a query vector; Based on the query vector, semantic retrieval is performed in the user vector library to obtain the target user set; the user vectors in the user vector library are constructed based on user behavior data and brand knowledge base.
2. The method according to claim 1, characterized in that, The construction method of the brand knowledge base includes: for any brand: clustering based on the brand's historical user behavior data to obtain multiple central vectors; using each central vector as a brand behavior template for the corresponding user group; and constructing a brand knowledge base based on multiple brand behavior templates.
3. The method according to claim 1, characterized in that, The method for constructing the user vector library includes: collecting behavioral data from multiple users; the behavioral data includes browsing data, add-to-cart data, order data, return / exchange data, and customer dialogue data; for any user's behavioral data: using a large language model to convert the user's behavioral data into user notes in natural language form, and converting the user notes into an initial feature vector; weightedly fusing the initial feature vector and the brand behavior template fusion vector to obtain the user-enhanced feature vector; the brand behavior template fusion vector is obtained from the initial feature vector based on the brand knowledge base; and the user vector library is obtained based on the enhanced feature vectors from multiple users.
4. The method according to claim 3, characterized in that, The method for determining the brand behavior template fusion vector includes: calculating the cosine similarity between the initial feature vector and each brand behavior template in the brand knowledge base to obtain multiple similarity scores; normalizing the multiple similarity scores to obtain multiple weight values that correspond one-to-one with the multiple similarity scores; and performing weighted fusion on the multiple brand behavior templates in the brand knowledge base based on the multiple weight values to obtain the brand behavior template fusion vector.
5. The method according to claim 1, characterized in that, The step of performing semantic retrieval in the user vector database based on the query vector to obtain the target user set includes: performing semantic retrieval in the user vector database based on the query vector using a marketing conversion prediction model to obtain the target user set; the marketing conversion prediction model is obtained after training on an industry knowledge base; the industry knowledge base is constructed based on historical industry experience.
6. A user selection device, characterized in that, include: The user selection instruction acquisition module is used to acquire user selection instructions in natural language form and convert the user selection instructions into query vectors. The user selection module is used to perform semantic retrieval in the user vector library based on the query vector to obtain a set of target users; the user vectors in the user vector library are constructed based on user behavior data and brand knowledge base.
7. The apparatus according to claim 6, characterized in that, The construction method of the brand knowledge base includes: for any brand: clustering based on the brand's historical user behavior data to obtain multiple central vectors; using each central vector as a brand behavior template for the corresponding user group; and constructing a brand knowledge base based on multiple brand behavior templates.
8. The apparatus according to claim 6, characterized in that, The method for constructing the user vector library includes: collecting behavioral data from multiple users; the behavioral data includes browsing data, add-to-cart data, order data, return / exchange data, and customer dialogue data; for any user's behavioral data: using a large language model to convert the user's behavioral data into user notes in natural language form, and converting the user notes into an initial feature vector; weightedly fusing the initial feature vector and the brand behavior template fusion vector to obtain the user-enhanced feature vector; the brand behavior template fusion vector is obtained from the initial feature vector based on the brand knowledge base; and the user vector library is obtained based on the enhanced feature vectors from multiple users.
9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the user selection method as described in any one of claims 1-5.
10. A processor, characterized in that, Used to run a computer program, which, when running, performs the user selection method as described in any one of claims 1-5.