Potential user identification method, device, storage medium, and apparatus

Abstract

The application relates to the technical field of metaverse, and discloses a potential user identification method, device, storage medium and apparatus, which comprises the following steps: splitting a potential user mining scene into multiple potential user mining sub-scenes, constructing multiple sub-scene semi-supervised models corresponding to the multiple potential user mining sub-scenes based on a multi-model cooperative semi-supervised learning mechanism, and identifying potential users through the multiple sub-scene semi-supervised models. In the application, different models are used to identify potential users for different potential user mining sub-scenes, and the models are constructed based on the multi-model cooperative semi-supervised learning mechanism, so that the accuracy of potential user identification can be improved, the efficiency of potential user identification can be improved, and the business promotion effect can be improved.

Description

Technical Field

[0001] This invention relates to the field of metaverse technology, and in particular to a method, device, storage medium and apparatus for identifying potential users. Background Technology

[0002] Accurately identifying potential user groups is crucial for business promotion. However, existing methods for identifying potential users are limited by the current stage of business development, data samples, and technical conditions, resulting in low identification efficiency and poor business promotion effects. Summary of the Invention

[0003] The main objective of this invention is to provide a method, device, storage medium, and apparatus for identifying potential users, aiming to solve the technical problems that existing potential user identification methods are limited by the current stage of business development, data samples, and technical conditions, resulting in low identification efficiency and poor business promotion effects.

[0004] To achieve the above objectives, the present invention provides a method for identifying potential users, the method comprising:

[0005] The potential user mining scenario is broken down into multiple potential user mining sub-scenarios;

[0006] Based on a multi-model collaborative semi-supervised learning mechanism, we construct multiple sub-scenario semi-supervised models corresponding to multiple potential user mining sub-scenarios.

[0007] Potential users are identified using the aforementioned semi-supervised models for multiple sub-scenes.

[0008] Optionally, the multi-model collaborative semi-supervised learning mechanism includes: positive and negative sample equalization processing; the construction of multiple sub-scenario semi-supervised models corresponding to multiple potential user mining sub-scenarios based on the multi-model collaborative semi-supervised learning mechanism includes:

[0009] A labeled sample set is formed based on labeled sample data, and the labeled sample set is split to obtain a model training set and a model test set.

[0010] When the ratio of positive to negative samples in the model training set is unbalanced, the model training set is subjected to sample equalization processing based on the expansion coefficient to obtain a new model training set.

[0011] The base model is trained based on the new model training set;

[0012] Based on the trained base model and the model test set, construct multiple semi-supervised models for multiple potential user mining sub-scenarios.

[0013] Optionally, the multi-model collaborative semi-supervised learning mechanism further includes: multiple base model construction algorithms; the base models include: a first base model, a second base model, and a third base model, wherein the first base model uses a decision tree algorithm as the base model algorithm, the second base model uses a logistic regression algorithm as the base model algorithm, and the third base model uses a random forest algorithm as the base model algorithm. The first base model and the third base model use a weighted loss function, and the second base model uses a cross-entropy loss function.

[0014] Optionally, the multi-model collaborative semi-supervised learning mechanism further includes: pseudo-label annotation technology; the construction of multiple sub-scenario semi-supervised models corresponding to multiple potential user mining sub-scenarios based on the trained base model and model test set includes:

[0015] Based on the trained base model, base model inference is performed on unlabeled sample data and model test set respectively;

[0016] Based on the inference results, pseudo-labels were applied to the unlabeled sample data and the model test set, respectively.

[0017] Based on the pseudo-label screening threshold under the current parameter conditions, the sample screening of the unlabeled sample set that has been pseudo-labeled is carried out until the model iteration optimization under the current parameter conditions is terminated, and multiple sub-scenario semi-supervised models corresponding to multiple potential user mining sub-scenarios are obtained.

[0018] Optionally, the model-cooperative semi-supervised learning mechanism further includes: a multi-parameter global search mechanism; the step of obtaining multiple sub-scenario semi-supervised models corresponding to multiple potential user mining sub-scenarios includes:

[0019] The current parameter conditions are updated based on the parameter global search mechanism, and a new round of base model iterative optimization is performed based on the updated current parameter conditions.

[0020] Based on the base model and model evaluation results of each parameter combination output by the parameter global search mechanism, multiple semi-supervised models corresponding to multiple potential user mining sub-scenarios are selected.

[0021] Optionally, the construction of multiple sub-scenario semi-supervised models corresponding to multiple potential user mining sub-scenarios based on the multi-model collaborative semi-supervised learning mechanism includes:

[0022] Construct a modeling indicator set, which includes an auxiliary modeling indicator set and a core modeling indicator set. Therefore, the core modeling indicator set includes business characteristic indicators from multiple dimensions.

[0023] Based on the modeling index set and the multi-model collaborative semi-supervised learning mechanism, multiple sub-scenario semi-supervised models are constructed for multiple potential user mining sub-scenarios.

[0024] Optionally, the identification of potential users through the multiple sub-scene semi-supervised models includes:

[0025] Using the modeling fields of the semi-supervised models for each sub-scene as inference input features, inference is performed on the user to be identified through the semi-supervised models for each sub-scene to obtain the model matrix for each sub-scene.

[0026] Calculate the weight values ​​of the semi-supervised models for each sub-scene based on the model matrix;

[0027] Based on the weight values, potential users are identified using the semi-supervised model of the multiple sub-scenes.

[0028] Furthermore, to achieve the above objectives, the present invention also proposes a potential user identification device, the potential user identification device including a memory, a processor, and a potential user identification program stored in the memory and executable on the processor, the potential user identification program being configured to implement the potential user identification method as described above.

[0029] In addition, to achieve the above objectives, the present invention also proposes a storage medium storing a potential user identification program, which, when executed by a processor, implements the potential user identification method as described above.

[0030] In addition, to achieve the above objectives, the present invention also proposes a potential user identification device, which includes: a scene segmentation module, a model construction module, and a user identification module;

[0031] The scenario splitting module is used to split the potential user mining scenario into multiple potential user mining sub-scenarios;

[0032] The model building module is used to build multiple sub-scenario semi-supervised models corresponding to multiple potential user mining sub-scenarios based on a multi-model collaborative semi-supervised learning mechanism.

[0033] The user identification module is used to identify potential users through the multiple sub-scenario semi-supervised models.

[0034] This invention discloses a method of dividing a potential user mining scenario into multiple potential user mining sub-scenarios, constructing multiple sub-scenarios semi-supervised models corresponding to these sub-scenarios based on a multi-model collaborative semi-supervised learning mechanism, and identifying potential users through these multiple sub-scenarios semi-supervised models. Because this invention utilizes different models for different potential user mining sub-scenarios and constructs models based on a multi-model collaborative semi-supervised learning mechanism, it can improve the accuracy of potential user identification, thereby increasing the efficiency of potential user identification and enhancing the effectiveness of business promotion. Attached Figure Description

[0035] Figure 1 This is a schematic diagram of the structure of a potential user identification device in the hardware operating environment involved in the embodiments of the present invention;

[0036] Figure 2 This is a flowchart illustrating the first embodiment of the potential user identification method of the present invention;

[0037] Figure 3 This is a schematic diagram of search suggestions from an embodiment of the potential user identification method of the present invention;

[0038] Figure 4 This is a flowchart illustrating a second embodiment of the potential user identification method of the present invention;

[0039] Figure 5 This is a schematic diagram of a multi-model collaborative semi-supervised learning mechanism in an embodiment of the potential user identification method of the present invention;

[0040] Figure 6 This is a schematic diagram of a pseudo-labeling method according to an embodiment of the potential user identification method of the present invention;

[0041] Figure 7 This is a flowchart illustrating the third embodiment of the potential user identification method of the present invention;

[0042] Figure 8 This is a flowchart illustrating the third embodiment of the potential user identification method of the present invention;

[0043] Figure 9 This is a structural block diagram of the first embodiment of the potential user identification device of the present invention.

[0044] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0045] It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0046] Reference Figure 1 , Figure 1 This is a schematic diagram of the potential user identification device structure in the hardware operating environment involved in the embodiments of the present invention.

[0047] like Figure 1As shown, the potential user identification device may include: a processor 1001, such as a central processing unit (CPU), a communication bus 1002, a user interface 1003, a network interface 1004, and a memory 1005. The communication bus 1002 is used to enable communication between these components. The user interface 1003 may include a display screen, and optionally, the user interface 1003 may also include a standard wired interface or a wireless interface. In this invention, the wired interface of the user interface 1003 may be a USB interface. The network interface 1004 may optionally include a standard wired interface or a wireless interface (such as a Wireless-Fidelity (Wi-Fi) interface). The memory 1005 may be a high-speed random access memory (RAM) or a non-volatile memory (NVM), such as a disk storage device. The memory 1005 may also optionally be a storage device independent of the aforementioned processor 1001.

[0048] Those skilled in the art will understand that Figure 1 The structure shown does not constitute a limitation on potential user identification devices and may include more or fewer components than illustrated, or combine certain components, or have different component arrangements.

[0049] like Figure 1 As shown, the memory 1005, which is identified as a computer storage medium, may include an operating system, a network communication module, a user interface module, and a potential user identification program.

[0050] exist Figure 1 In the potential user identification device shown, the network interface 1004 is mainly used to connect to the backend server and communicate data with the backend server; the user interface 1003 is mainly used to connect to the user equipment; the potential user identification device calls the potential user identification program stored in the memory 1005 through the processor 1001 and executes the potential user identification method provided in the embodiment of the present invention.

[0051] Based on the above hardware structure, an embodiment of the potential user identification method of the present invention is proposed.

[0052] Reference Figure 2 , Figure 2 This is a flowchart illustrating the first embodiment of the potential user identification method of the present invention, which presents the first embodiment of the potential user identification method of the present invention.

[0053] In a first embodiment, the potential user identification method includes the following steps:

[0054] Step S10: Divide the potential user mining scenario into multiple potential user mining sub-scenarios.

[0055] It should be understood that Non-Fungible Tokens (NFTs) are unique cryptocurrency tokens derived from blockchain technology to represent data assets (including images and videos). NFTs allow any rights in the cloud universe to be easily financialized and materialized, thus providing underlying support for the generation, ownership confirmation, pricing, circulation, and traceability of digital assets within the metaverse. The tradable nature of NFTs has led to their recent surge in popularity, especially digital collectible NFTs. Mainstream internet companies, including major telecom operators, are actively developing their own NFT-related businesses and have successively released their own digital collectibles on several digital collectible platforms. Therefore, accurately identifying potential user groups and operating their own digital collectible businesses is a key concern for all digital collectible issuers. Currently, the metaverse NFT-related businesses of various telecom operators are still in their initial stages, with a lack of sample users. Therefore, the construction of potential user identification models mainly follows two methods: 1. Conduct feature analysis on NFT business subscribers, using clustering algorithms to cluster subscribers into customer groups and summarize the characteristics of each customer group. Marketing efforts include: 1. Reaching potential customers and identifying competitors based on the characteristics and rules of each customer group; 2. Using a limited number of NFT service subscribers as positive samples and randomly selecting a certain number of users as negative samples to form a modeling dataset. Applying simple and easy-to-understand machine learning algorithms, such as decision tree algorithms and logistic regression algorithms, to build machine learning models. Providing operators' technical specialists with several sample learning rules to support their development of relevant potential customer identification models. Overall, due to limitations in the current stage of business development, data samples, and technical conditions, the efficiency of NFT potential user identification models used by various telecom operators is currently low, and the business promotion effect is generally average.

[0056] To overcome the above-mentioned shortcomings, this embodiment uses different models for potential user identification for different potential user mining sub-scenarios, and builds models based on a multi-model collaborative semi-supervised learning mechanism, thereby improving the accuracy of potential user identification, improving the efficiency of potential user identification, and enhancing the business promotion effect.

[0057] It is understood that the execution subject of the method in this embodiment may be a potential user identification device with data processing, network communication and program running functions, such as a server, or other electronic devices that can achieve the same or similar functions. This embodiment does not limit this.

[0058] It should be noted that currently, in order to adapt to the development needs of metaverse NFTs, operators are increasingly diversifying and subdividing NFT product types, specifically into "perpetually held NFTs," "limited-trade NFTs," and "token NFTs." Each type of NFT has its unique attributes and value, and its target customer group also has unique consumption behaviors, app usage habits, communication behaviors, investment tendencies, and personal preferences. To effectively support operators in operating various types of NFT products, this embodiment can break down the potential user mining scenario into multiple potential user mining sub-scenarios. For example, the NFT potential user mining scenario can be broken down into a sub-scenario for identifying potential users of permanently held NFTs, a sub-scenario for identifying potential users of limited-trade NFTs, and a sub-scenario for identifying potential users of token NFTs.

[0059] Step S20: Construct multiple sub-scenario semi-supervised models corresponding to multiple potential user mining sub-scenarios based on a multi-model collaborative semi-supervised learning mechanism.

[0060] It is understandable that this embodiment introduces a multi-model semi-supervised learning mechanism to train the semi-supervised models for each sub-scene, thereby improving the accuracy of the semi-supervised models for each sub-scene.

[0061] Step S30: Identify potential users using the multiple sub-scenario semi-supervised models.

[0062] For ease of understanding, please refer to Figure 3 This explanation does not limit the scope of this solution. Figure 3 As a schematic diagram of search suggestions, the potential user mining scenario is first broken down into multiple potential user mining sub-scenarios, namely, the potential user identification sub-scenarios of permanently held NFTs, the potential user identification sub-scenarios of limited-trade NFTs, and the potential user identification sub-scenarios of token NFTs. Then, a model for each sub-scenarios is constructed based on a multi-model collaborative semi-supervised learning mechanism, namely, the potential user identification sub-model of permanently held NFTs, the potential user identification model of limited-trade NFTs, and the potential user identification sub-model of token NFTs, as shown in the diagram. Finally, potential users are identified through each sub-scenarios model.

[0063] In the first embodiment, a method is disclosed that the potential user mining scenario is divided into multiple potential user mining sub-scenarios. Based on a multi-model collaborative semi-supervised learning mechanism, multiple sub-scenarios semi-supervised models are constructed for each potential user mining sub-scenarios. Potential users are identified through these multiple sub-scenarios semi-supervised models. Since this embodiment uses different models for potential user identification for different potential user mining sub-scenarios and constructs models based on a multi-model collaborative semi-supervised learning mechanism, the accuracy of potential user identification can be improved, thereby increasing the efficiency of potential user identification and enhancing the effectiveness of business promotion.

[0064] Reference Figure 4 , Figure 4 This is a flowchart illustrating the second embodiment of the search suggestion method of the present invention, based on the above. Figure 2 The first embodiment shown presents a second embodiment of the potential user identification method of the present invention.

[0065] In the second embodiment, the multi-model collaborative semi-supervised learning mechanism includes: positive and negative sample equalization processing, and step S20 includes:

[0066] Step S201: Based on the labeled sample data, form a labeled sample set, and split the labeled sample set to obtain the model training set and the model test set.

[0067] It should be understood that existing technologies lack effective modeling methods when modeling scenarios with extremely imbalanced positive and negative samples. Traditional and crude "oversampling" or "undersampling" methods can lead to overfitting or underfitting of the model, resulting in poor model performance.

[0068] Therefore, in order to overcome the above-mentioned defects, in this embodiment, an expansion coefficient is used to balance the positive and negative samples for samples with an unbalanced ratio of positive and negative samples, thereby improving the accuracy of model training.

[0069] Step S202: When the ratio of positive to negative samples in the model training set is unbalanced, perform sample equalization processing on the model training set based on the expansion coefficient to obtain a new model training set.

[0070] It should be noted that the modeling dataset in this scheme consists of two parts: a labeled sample dataset and an unlabeled sample dataset. The labeled sample dataset has explicit positive and negative sample labels, and similarly, the unlabeled sample dataset does not have such labels.

[0071] The model aims to identify potential NFT business subscribers of various types in the metaverse, therefore, actual NFT business subscribers are selected as positive samples when constructing the model.

[0072] In this scheme, all registered users of several digital collection apps up to the present are used as positive samples for modeling, and unregistered users in normal status are used as negative samples to construct a labeled sample set for modeling.

[0073] The positive sample consists of mobile phone users who actually registered for the digital collectibles app. Therefore, the positive sample user indicator data is obtained by taking the registration time of the digital collectibles app as the time node and obtaining indicator data according to the indicator definition. For example, the data extraction method for the indicator "total number of visits to NFT-related competing products in the past week" is "taking the registration time as the time node and calculating the total number of visits to competing NFT products by the mobile phone user in the week before registration". The data extraction method for the indicator "time interval since the last search for content related to "Bitcoin", "Metaverse", "Data Assets" and "Non-fungible Tokens" is "taking the registration time as the time node, calculating the search date of the last search for content related to "Bitcoin", "Metaverse", "Data Assets", and "Non-fungible Tokens" before the registration time and calculating the number of days between the search date and the registration time". Negative sample users are mobile phone users who are in normal status but have not registered for related digital collectible apps. The extraction of negative sample indicator data is based on the billing period T and the corresponding indicator data is extracted according to the indicator definition. For example, if the billing period is selected as April 20, the data extraction scope of the indicator "number of NFT apps accessed in the last three days" is "with April 20 as the time node, calculate the total number of NFT apps logged in by the user from April 18 to April 20".

[0074] The extraction method for unlabeled sample datasets is the same as the method for obtaining negative samples in labeled sample datasets, but there are two differences:

[0075] The billing period for the unlabeled sample dataset should be different from that for the labeled negative samples. That is, user data with a different billing period should be selected as the unlabeled sample. In this scheme, T is the billing period for the labeled negative samples and T+1 is the billing period for the unlabeled sample data. For example, if April 20 is used as the billing period for the labeled negative samples, then the billing period for the unlabeled sample dataset should be April 21.

[0076] Unlabeled sample users should be all users, that is, both positive and negative sample users (without positive or negative sample labels), while negative sample users in labeled datasets refer only to negative sample users with labeled labels.

[0077] Step S203: Train the base model based on the new model training set.

[0078] In the second embodiment, the multi-model collaborative semi-supervised learning mechanism further includes: multiple base model construction algorithms; the base models include: a first base model, a second base model, and a third base model, wherein the first base model uses a decision tree algorithm as the base model algorithm, the second base model uses a logistic regression algorithm as the base model algorithm, and the third base model uses a random forest algorithm as the base model algorithm. The first base model and the third base model use a weighted loss function, and the second base model uses a cross-entropy loss function.

[0079] It is understood that in this embodiment, a custom loss function is used to improve the learning performance of each base model during the semi-supervised training process.

[0080] In the second embodiment, the multi-model collaborative semi-supervised learning mechanism further includes: pseudo-label annotation technology; step S20 further includes:

[0081] Step S204: Perform base model inference on the unlabeled sample data and the model test set based on the trained base model.

[0082] It should be understood that in this embodiment, pseudo-labeling technology can also be used to further optimize the training of the base model under the current parameters.

[0083] Step S205: Based on the inference results, perform pseudo-label annotation on the unlabeled sample data and the model test set respectively.

[0084] Step S206: Based on the pseudo-label screening threshold under the current parameter conditions, perform sample screening on the unlabeled sample set that has been pseudo-labeled, until the model iterative optimization under the current parameter conditions terminates.

[0085] In the second embodiment, the model collaborative semi-supervised learning mechanism further includes a multi-parameter global search mechanism, and step S20 further includes:

[0086] Step S207: Update the current parameter conditions based on the parameter global search mechanism, and perform a new round of base model iterative optimization based on the updated current parameter conditions.

[0087] It is understood that, based on the optimal base model and model evaluation results of each parameter combination output by the parameter global search mechanism in this embodiment, the optimal model combination can be selected and used as the final base model combination.

[0088] Step S208: Based on the base model and model evaluation results of each parameter combination output by the parameter global search mechanism, select multiple sub-scenario semi-supervised models corresponding to multiple potential user mining sub-scenarios.

[0089] It should be understood that the primary purpose of existing technologies, whether clustering or classification models, is mostly to provide technical professionals with data-level learning rules to support the formulation of rules for selecting potential target user groups. However, due to the extreme scarcity of modeling samples, various algorithms cannot be adequately trained when learning the characteristics of NFT business subscribers, resulting in the final learning rules having certain sample limitations, which has a significant impact on the formulation of subsequent business rules and the use of the model.

[0090] Therefore, to overcome the aforementioned shortcomings, this embodiment proposes a novel modeling method for scenarios with extreme imbalance between positive and negative samples: a modeling method based on a novel multi-model collaborative semi-supervised learning mechanism. This method, based on a limited number of labeled samples, utilizes multiple base model construction algorithms, novel pseudo-label annotation techniques, a global parameter search mechanism, and a model iteration termination condition set based on a comprehensive base model performance evaluation strategy. It iteratively optimizes the base model using a large amount of unlabeled data and ultimately outputs the optimal version of the base model. This improved semi-supervised learning strategy differs from traditional fully supervised machine learning models; it can effectively solve the modeling problem of extreme imbalance between positive and negative samples in the NFT potential user identification model, thereby improving model performance.

[0091] For ease of understanding, please refer to Figure 5 This explanation does not limit the scope of this solution. Figure 5 This is a schematic diagram of a multi-model collaborative semi-supervised learning mechanism. The specific steps of the multi-model collaborative semi-supervised learning mechanism are as follows:

[0092] Step 1: Split the modeling dataset based on the labeled sample set, dividing it into training and test sets according to a certain ratio. In this scheme, the total number of labeled positive samples for the sub-scenarios of "permanent holding NFTs", "limited transaction NFTs", and "token NFTs" are 462, 1223, and 185, respectively. For each sub-scenarios, 30,000 negative sample users are randomly selected to form the labeled sample set, resulting in a significant imbalance between the positive and negative sample sizes. The training and test sets are split in a 1:3 ratio.

[0093] Step 2: Based on the actual condition of the training set, consider whether sample equalization is needed to create a new training set. An extremely imbalanced ratio of positive to negative samples will cause skewed model learning, resulting in poor model performance and unusable models. In this scheme, the original training sets of each sub-model have a severely imbalanced positive and negative sample ratio, requiring sample equalization to create a new training set. This scheme uses an adaptive sample synthesis algorithm to achieve sample equalization of the original training set. The specific algorithm implementation is as follows:

[0094] ① Calculate the class imbalance d of positive and negative samples.

[0095] d=n s / n l

[0096] Where, n s n is the number of samples in the minority class. l This represents the number of samples in the majority category.

[0097] ② Set the maximum class imbalance d th (Threshold), if d <d thThen the following sample synthesis step will be executed:

[0098] I. Calculate the number of data samples to be synthesized for the minority class samples (positive samples in this scheme):

[0099] G = (n l -n s )*λ

[0100] λ is the balance coefficient. For example, if λ = 1, it means that a perfectly balanced set of data samples is created.

[0101] II. Calculate the proportionality coefficient η. For each sample belonging to the minority class, calculate k neighbors using Euclidean distance, where Δ is the number of samples belonging to the majority class among the k neighbors, η. i =Δ i / k,i=1,...,n s The Euclidean distance formula is as follows: x p Let x be the p-th sample in the positive sample dataset. n x is the nth sample in the negative sample dataset. p1 ...x pn For sample x p The n-dimensional index value, x n1 ...x nn For x n The n-dimensional index value.

[0102] Ⅲ. Normalize the proportionality coefficient η.

[0103]

[0104] IV. Calculate for x i The number of samples to be generated.

[0105]

[0106] V.g i Each sample generation executes the following loop:

[0107] Step 1: Determine the single-sample expansion coefficient λ, where λ should be less than g. i The minimum value;

[0108] Step 2: From x i λ minority class samples x are randomly selected from the k nearest neighbors. p1 ...x pλ λ synthetic samples are synthesized based on the following formula for synthesizing data samples:

[0109] s i =x i +(x pi -xi )*β, where β is a random number between 0 and 1

[0110] Step 3: The undersampling mechanism is used to reverse-pick up the fitted samples near the decision boundary and retain the high-quality synthetic samples;

[0111] Step 4: Calculate the ratio of positive to negative sample sizes to determine if the synthesis termination condition is met.

[0112] Step 3: Initialize the base model parameters, the number of iterations to terminate, the pseudo-label screening threshold, and other parameter settings. Then, formulate the iteration termination conditions based on the number of iterations to terminate, the model performance evaluation results of the previous iteration, and the model performance evaluation status of the current iteration.

[0113] Step 4: Based on the new model training set formed after the equalization process in Step 2, select multiple model algorithms of different categories as base model algorithms, and train the base models using the base model parameters from Step 3 (using the idea of ​​single-view collaborative training). The base model parameters in Step 3 should correspond to the base model algorithms in this step. In this scheme, the decision tree algorithm (LightGBM), logistic regression algorithm (LogisticRegression), and random forest algorithm (RandomForest) are selected as base model algorithms. Therefore, the base model parameters in Step 3 are Learning rate and n_estimators (LightGBM), penalty and C (LogisticRegression), and n_estimators and max_depth (RandomForest), respectively. Custom loss functions are introduced for the three base models to optimize performance. The specific loss functions are as follows:

[0114] Due to the imbalance of the modeling samples, LightGBM and RandomForest in this scheme use a custom weighted loss function, the specific formula of which is as follows:

[0115]

[0116] In the formula, weight i The penalty weights set based on the imbalance of the modeled samples, y i For actual user tags, This represents the output probability predicted by the model. Compared to traditional loss functions, using this weighted loss function can significantly improve the model's F1 score, allowing for more thorough learning of sample features and resulting in better model performance.

[0117] LogisticRegression uses cross-entropy to define its loss function. Compared with the conventional least squares loss function, this custom loss function can significantly improve the model's prediction accuracy without reducing the model's recall. The specific formula is as follows:

[0118]

[0119] Step 5: Based on the N base models trained, perform base model inference on unlabeled samples and model test sets respectively.

[0120] Step 6: Introduce a novel pseudo-labeling method, and perform pseudo-labeling on the unlabeled samples and model test set samples based on Step 5. This scheme uses a novel model fusion strategy to fuse the base models and uses the output as the sample pseudo-label value. The specific model fusion method is as follows: Figure 6 As shown, Figure 6 This diagram illustrates the pseudo-labeling method. Taking sample X from the unlabeled sample set as an example, firstly, each base model "votes" for sample X through a hard voting mechanism, calculating the percentage of base models voting as positive samples and using this percentage as the inflation coefficient (the inflation coefficient can further stratify the accuracy of the soft voting fusion results of the base models, which is beneficial to improving model performance). Secondly, based on the inference results of the model test set in step 5, the performance of each base model is evaluated (each base model selects the same probability threshold and uses hard voting to infer labels from the model test set samples). In this scheme, F1 score is used to quantify the base model performance; the larger the F1 score, the better the performance of the base model. The base model performance ratio is output, for example, 1:1.5:3. In this scheme, the performance evaluation ratio of each base model is used as the weight value of each base model. Finally, based on the probability value of each base model predicting sample X as a potential user of a certain type of NFT, the weight allocation value of each base model, and the inflation coefficient, the final comprehensive value is output, which is used to measure the overall evaluation of several base models on whether sample X is a potential user of a certain type of NFT. The normalized overall evaluation value is the pseudo-label value for the unlabeled samples. The overall evaluation value is calculated as follows:

[0121] P = (z1×p1 + z2×p2 + ... + z n ×p n )×λ

[0122] In the formula, z1, z2, ..., z n Let p1, p2, ..., p be the weights of the n base models. n λ represents the predicted probability values ​​of the n base models for the unlabeled samples, and λ is the inflation coefficient.

[0123] Step 7: Based on the base model inference results of the unlabeled sample set and model test set in Step 5, apply the pseudo-labeling method from Step 6 to perform pseudo-labeling on the unlabeled sample set and model test set respectively (the resulting pseudo-label values ​​are the P-values ​​in Step 6). Based on the original labeled values ​​and pseudo-labeled values ​​of the model test set, and using a 0.5 threshold, apply multi-dimensional model performance evaluation methods such as accuracy, ROC, and F1 Score to evaluate the model performance of the base model in the current iteration round, and store the model evaluation results for this round.

[0124] Step 8: Based on the pseudo-label filtering threshold set in Step 3, filter the unlabeled sample set that has been pseudo-labeled. Select unlabeled sample sets that exceed the pseudo-label filtering threshold as positive samples, and use the remaining unlabeled samples as the unlabeled sample set for the next iteration for further filtering. The selected unlabeled positive sample dataset is merged with the model training sample set from the previous round to generate a new model training sample set.

[0125] Step 9: Based on the current base model and parameters such as the pseudo-label screening threshold, and combined with the updated new modeling sample set, execute steps 4 to 8 again to further optimize the training of the base model under the current parameters.

[0126] Step 10: Based on the evaluation results of the base model performance in the previous and current rounds and the termination iteration conditions set in Step 3, determine whether the model iteration optimization under the current parameter conditions should be terminated. If the conditions are met, terminate the iteration under the current parameter conditions, output the base model training results from the previous round as the optimal version under the current parameters, and execute Step 11. Otherwise, execute Step 9.

[0127] Step 11: Update the input parameters from Step 3 based on the parameter global search mechanism, perform a new round of base model iterative optimization under the new parameters, and continue executing Steps 4-10. It is important to note that the unlabeled dataset under the new model parameter conditions must be the original, full unlabeled dataset.

[0128] Global search involves iterating through all the input parameters of a semi-supervised learning framework, comparing the optimal application effect of the base model under different parameter combinations, and then selecting the optimal parameter combination among all parameter combinations and outputting the optimal base model under this combination as the optimal learning result of the semi-supervised model. The semi-supervised learning framework for the NFT latent user identification model in this scheme has the following parameters: pseudo-label screening threshold (Threshold), base model LightGBM and RandomForest hyperparameters (number of weak classifiers n_estimators), base model LightGBM hyperparameter (Learning rate), base model RandomForest hyperparameter (maximum depth of the tree max_depth), and LogisticRegression hyperparameter (the reciprocal of the regularization coefficient C). The initial values ​​and step sizes of each parameter are set as follows: Threshold (initial value 0.5, step size 0.01), n_estimators (initial value 200, step size 50), Learning rate (initial value 0.001, step size 0.001), max_depth (initial value 1, step size 1), and C (initial value 0.01, step size 0.01). The optimal solution is searched by traversing all parameter combinations.

[0129] Step 12: Based on the optimal base model and model evaluation results of each parameter combination output by the parameter global search mechanism, select the optimal model combination and use it as the final base model combination.

[0130] Reference Figure 7 , Figure 7 This is a flowchart illustrating the third embodiment of the search suggestion method of the present invention, based on the above. Figure 2 The first embodiment shown presents a third embodiment of the potential user identification method of the present invention.

[0131] In the third embodiment, step S20 includes:

[0132] Step S201': Construct a modeling indicator set, which includes an auxiliary modeling indicator set and a core modeling indicator set. Therefore, the core modeling indicator set includes business characteristic indicators of multiple dimensions.

[0133] It should be understood that existing technologies are all built on a simple set of modeling metrics, and rarely involve important metrics that reflect the characteristics of NFT users, such as customers' search preferences, customers' app behavior habits, customers' investment enthusiasm, and customers' role attributes, resulting in poor overall model performance.

[0134] Therefore, to overcome the aforementioned shortcomings, this embodiment analyzes the NFT business subscription customer group and discovers that these users exhibit characteristics such as investment preferences and targeted attention preferences. From a business perspective, several core user behavior characteristic indicators are constructed using NFT market popularity data from the operator sector, NFT form data, NFT category information, user APP usage data, search engine content parsing data, and user signaling trajectory data to support the construction of a potential NFT user identification model.

[0135] Step S202': Based on the modeling index set and the multi-model collaborative semi-supervised learning mechanism, construct multiple sub-scenario semi-supervised models corresponding to multiple potential user mining sub-scenarios.

[0136] The construction of the modeling indicator set is a crucial step in modeling. In this solution, the modeling indicator sets for each sub-scenario are designed from dimensions such as operator NFT product market popularity data, NFT form data, user basic attributes, user communication information, NFT category information, consumption information, social circle information, traffic information, search and attention information, and APP usage information. Unlike existing modeling techniques, the construction of the modeling indicator set in this solution is divided into two parts: the construction of an auxiliary modeling indicator set and the construction of a core modeling indicator set. The core indicators are several business characteristic indicators based on feature analysis and summarization of registered users of various NFT digital collectible APPs. These indicators can effectively capture and distinguish high-probability potential users, improving the overall performance of the model. For ease of understanding, please refer to Tables 1 and 2, but this does not limit this solution. Table 1 shows the auxiliary modeling indicator set for the NFT potential user identification model, and Table 2 shows the core modeling indicator set for the NFT potential user identification model.

[0137] Table 1. Set of Auxiliary Modeling Indicators for NFT Potential User Identification Model

[0138]

[0139]

[0140] Table 2. Core Modeling Indicators for the NFT Potential User Identification Model

[0141]

[0142]

[0143] Reference Figure 8 , Figure 8 This is a flowchart illustrating the fourth embodiment of the search suggestion method of the present invention, based on the above. Figure 2 The first embodiment shown presents a fourth embodiment of the potential user identification method of the present invention.

[0144] In the fourth embodiment, step S30 includes:

[0145] Step S301: Using the modeling fields of the semi-supervised models of each sub-scene as inference input features, inference is performed on the user to be identified through the semi-supervised models of each sub-scene to obtain the model matrix of each sub-scene.

[0146] It should be understood that in order to reasonably output the target customers (i.e., identify potential users) for the three sub-scenarios, this solution adopts a self-developed weight evaluation strategy, which outputs the optimal selection weights for each sub-model through a dynamic parameter search mechanism.

[0147] Step S302: Calculate the weight values ​​of the semi-supervised models for each sub-scene based on the model matrix.

[0148] For ease of understanding, the following examples are provided, but they do not limit the scope of this solution. The specific steps for calculating the customer group range for each sub-scenario fusion selection are as follows:

[0149] 1. Construct the model matrix for each sub-scene

[0150] Users whose accounts are in good standing for the new billing cycle and who have not ordered any NFT-related products in the past month are considered as potential users to be predicted. The modeling fields of three sub-scenario models—the permanent NFT holding potential user identification sub-model, the limited-transaction NFT potential user identification sub-model, and the token NFT potential user identification sub-model—are used as inference input features. The trained sub-scenario models are applied to predict users for the new billing cycle. Now, let's assume the inference result for scenario 1 (the permanent NFT holding potential user identification sub-model) is X1, where x... 11 ,x 12 ,x 13 ,...,x 1n The model predicts the probability that each user is a potential user of a permanently held NFT. Similarly, the model inference results for scenarios 2 and 3 are X2 and X3, respectively. The model inference matrix is ​​constructed as follows:

[0151]

[0152] 2. Calculate the weights of each sub-model.

[0153] ① Data Transformation

[0154] in

[0155]

[0156] ② Calculate the correlation coefficient

[0157] in

[0158] ③ Calculate the eigenvalues ​​λ1, λ2, and λ3 of the correlation coefficient matrix respectively.

[0159] ④ A self-developed weight evaluation method is introduced to construct the final selection weights for the three scenarios. The formula for the weight evaluation method is as follows:

[0160]

[0161] Where θ is the direction of the variable parameter update, and α, β and γ are the input parameters.

[0162] The final combined weights for the selection magnitudes of the three scenarios are as follows:

[0163]

[0164]

[0165]

[0166] In the formula, C represents the Cridit gradient search with E1 and E2 as search parameters.

[0167] Step S303: Based on the weight values, identify potential users using the multiple sub-scenario semi-supervised models.

[0168] To facilitate understanding, examples are provided below, but these do not limit the scope of this solution. The specific steps for selecting the target customer group are as follows:

[0169] First, the inference customer groups are sorted from highest to lowest according to the inference probabilities of Model Scenario 1, Model Scenario 2, and Model Scenario 3. Target customers for each scenario are then selected from highest to lowest probability based on the size of the selected customers. If the total number of inference customers is N, then the size of the selected target customers for Scenario 1 is w1*N, for Scenario 2 it is w2*N, and for Scenario 3 it is w3*N. For the overlapping scenarios, the final target contact scenario is selected based on the corresponding inference probabilities of the three scenarios.

[0170] Furthermore, embodiments of the present invention also propose a storage medium storing a potential user identification program, which, when executed by a processor, implements the potential user identification method as described above.

[0171] In addition, refer to Figure 9The present invention also proposes a potential user identification device, which includes: a scene segmentation module 10, a model construction module 20, and a user identification module 30.

[0172] The scenario splitting module 10 is used to split the potential user mining scenario into multiple potential user mining sub-scenarios.

[0173] It should be understood that Non-Fungible Tokens (NFTs) are unique cryptocurrency tokens derived from blockchain technology to represent data assets (including images and videos). NFTs allow any rights in the cloud universe to be easily financialized and materialized, thus providing underlying support for the generation, ownership confirmation, pricing, circulation, and traceability of digital assets within the metaverse. The tradable nature of NFTs has led to their recent surge in popularity, especially digital collectible NFTs. Mainstream internet companies, including major telecom operators, are actively developing their own NFT-related businesses and have successively released their own digital collectibles on several digital collectible platforms. Therefore, accurately identifying potential user groups and operating their own digital collectible businesses is a key concern for all digital collectible issuers. Currently, the metaverse NFT-related businesses of various telecom operators are still in their initial stages, with a lack of sample users. Therefore, the construction of potential user identification models mainly follows two methods: 1. Conduct feature analysis on NFT business subscribers, using clustering algorithms to cluster subscribers into customer groups and summarize the characteristics of each customer group. Marketing efforts include: 1. Reaching potential customers and identifying competitors based on the characteristics and rules of each customer group; 2. Using a limited number of NFT service subscribers as positive samples and randomly selecting a certain number of users as negative samples to form a modeling dataset. Applying simple and easy-to-understand machine learning algorithms, such as decision tree algorithms and logistic regression algorithms, to build machine learning models. Providing operators' technical specialists with several sample learning rules to support their development of relevant potential customer identification models. Overall, due to limitations in the current stage of business development, data samples, and technical conditions, the efficiency of NFT potential user identification models used by various telecom operators is currently low, and the business promotion effect is generally average.

[0174] To overcome the above-mentioned shortcomings, this embodiment uses different models for potential user identification for different potential user mining sub-scenarios, and builds models based on a multi-model collaborative semi-supervised learning mechanism, thereby improving the accuracy of potential user identification, improving the efficiency of potential user identification, and enhancing the business promotion effect.

[0175] It should be noted that currently, in order to adapt to the development needs of metaverse NFTs, operators are increasingly diversifying and subdividing NFT product types, specifically into "perpetually held NFTs," "limited-trade NFTs," and "token NFTs." Each type of NFT has its unique attributes and value, and its target customer group also has unique consumption behaviors, app usage habits, communication behaviors, investment tendencies, and personal preferences. To effectively support operators in operating various types of NFT products, this embodiment can break down the potential user mining scenario into multiple potential user mining sub-scenarios. For example, the NFT potential user mining scenario can be broken down into a sub-scenario for identifying potential users of permanently held NFTs, a sub-scenario for identifying potential users of limited-trade NFTs, and a sub-scenario for identifying potential users of token NFTs.

[0176] The model building module 20 is used to build multiple sub-scenario semi-supervised models corresponding to multiple potential user mining sub-scenarios based on a multi-model collaborative semi-supervised learning mechanism.

[0177] It is understandable that this embodiment introduces a multi-model semi-supervised learning mechanism to train the semi-supervised models for each sub-scene, thereby improving the accuracy of the semi-supervised models for each sub-scene.

[0178] The user identification module 30 is used to identify potential users through the multiple sub-scene semi-supervised models.

[0179] For ease of understanding, please refer to Figure 3 This explanation does not limit the scope of this solution. Figure 3 As a schematic diagram of search suggestions, the potential user mining scenario is first broken down into multiple potential user mining sub-scenarios, namely, the potential user identification sub-scenarios of permanently held NFTs, the potential user identification sub-scenarios of limited-trade NFTs, and the potential user identification sub-scenarios of token NFTs. Then, a model for each sub-scenarios is constructed based on a multi-model collaborative semi-supervised learning mechanism, namely, the potential user identification sub-model of permanently held NFTs, the potential user identification model of limited-trade NFTs, and the potential user identification sub-model of token NFTs, as shown in the diagram. Finally, potential users are identified through each sub-scenarios model.

[0180] In this embodiment, a method is disclosed that the potential user mining scenario is divided into multiple potential user mining sub-scenarios. Based on a multi-model collaborative semi-supervised learning mechanism, multiple sub-scenarios semi-supervised models are constructed for each potential user mining sub-scenarios. Potential users are identified through these multiple sub-scenarios semi-supervised models. Since this embodiment uses different models for potential user identification for different potential user mining sub-scenarios and constructs models based on a multi-model collaborative semi-supervised learning mechanism, the accuracy of potential user identification can be improved, thereby increasing the efficiency of potential user identification and enhancing the effectiveness of business promotion.

[0181] Other embodiments or specific implementations of the potential user identification device described in this invention can be found in the above-described method embodiments, and will not be repeated here.

[0182] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or system 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 system. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or system that includes that element.

[0183] The sequence numbers of the above embodiments of the present invention are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0184] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as a read-only memory image (ROM) / random access memory (RAM), magnetic disk, optical disk), and includes several instructions to cause a terminal device (which may be a mobile phone, computer, server, air conditioner, or network device, etc.) to execute the methods described in the various embodiments of the present invention.

[0185] The above are merely preferred embodiments of the present invention and do not limit the scope of the patent. Any equivalent structural or procedural transformations made based on the description and drawings of the present invention, or direct or indirect applications in other related technical fields, are similarly included within the scope of patent protection of the present invention.

Claims

1. A method for identifying potential users, characterized in that, The potential user identification method includes: The potential user mining scenario is divided into multiple potential user mining sub-scenarios, including the potential user identification sub-scenarios of permanently holding NFTs, the potential user identification sub-scenarios of limited-trade NFTs, and the potential user identification sub-scenarios of token NFTs. Based on a multi-model collaborative semi-supervised learning mechanism, multiple sub-scenario semi-supervised models corresponding to multiple potential user mining sub-scenarios are constructed. The multi-model collaborative semi-supervised learning mechanism includes positive and negative sample equalization processing, multiple base model construction algorithms, pseudo-label annotation technology, and multi-parameter global search mechanism. Potential users are identified using the aforementioned semi-supervised models for multiple sub-scenes.

2. The potential user identification method as described in claim 1, characterized in that, The multi-model collaborative semi-supervised learning mechanism includes: positive and negative sample equalization processing; the construction of multiple sub-scenario semi-supervised models corresponding to multiple potential user mining sub-scenarios based on the multi-model collaborative semi-supervised learning mechanism includes: A labeled sample set is formed based on labeled sample data, and the labeled sample set is split to obtain a model training set and a model test set. When the ratio of positive to negative samples in the model training set is unbalanced, the model training set is subjected to sample equalization processing based on the expansion coefficient to obtain a new model training set. The base model is trained based on the new model training set; Based on the trained base model and the model test set, construct multiple semi-supervised models for multiple potential user mining sub-scenarios.

3. The potential user identification method as described in claim 2, characterized in that, The multi-model collaborative semi-supervised learning mechanism also includes: multiple base model construction algorithms; the base models include: a first base model, a second base model, and a third base model. The first base model uses a decision tree algorithm as the base model algorithm, the second base model uses a logistic regression algorithm as the base model algorithm, and the third base model uses a random forest algorithm as the base model algorithm. The first base model and the third base model use a weighted loss function, so the second base model uses a cross-entropy loss function.

4. The potential user identification method as described in claim 2, characterized in that, The multi-model collaborative semi-supervised learning mechanism also includes: pseudo-label annotation technology; the construction of multiple sub-scenario semi-supervised models corresponding to multiple potential user mining sub-scenarios based on the trained base model and model test set includes: Based on the trained base model, base model inference is performed on unlabeled sample data and model test set respectively; Based on the inference results, pseudo-labels were applied to the unlabeled sample data and the model test set, respectively. Based on the pseudo-label screening threshold under the current parameter conditions, the sample screening of the unlabeled sample set that has been pseudo-labeled is carried out until the model iteration optimization under the current parameter conditions is terminated, and multiple sub-scenario semi-supervised models corresponding to multiple potential user mining sub-scenarios are obtained.

5. The potential user identification method as described in claim 4, characterized in that, The model-cooperative semi-supervised learning mechanism also includes: a multi-parameter global search mechanism; the acquisition of multiple sub-scenario semi-supervised models corresponding to multiple potential user mining sub-scenarios includes: The current parameter conditions are updated based on the parameter global search mechanism, and a new round of base model iterative optimization is performed based on the updated current parameter conditions. Based on the base model and model evaluation results of each parameter combination output by the parameter global search mechanism, multiple semi-supervised models corresponding to multiple potential user mining sub-scenarios are selected.

6. The potential user identification method as described in any one of claims 1 to 5, characterized in that, The method for constructing multiple sub-scenario semi-supervised models corresponding to multiple potential user mining sub-scenarios based on a multi-model collaborative semi-supervised learning mechanism includes: Construct a modeling indicator set, which includes an auxiliary modeling indicator set and a core modeling indicator set. Therefore, the core modeling indicator set includes business characteristic indicators from multiple dimensions. Based on the modeling index set and the multi-model collaborative semi-supervised learning mechanism, multiple sub-scenario semi-supervised models are constructed for multiple potential user mining sub-scenarios.

7. The potential user identification method as described in any one of claims 1 to 5, characterized in that, The process of identifying potential users through the multiple sub-scenario semi-supervised models includes: Using the modeling fields of the semi-supervised models for each sub-scene as inference input features, inference is performed on the user to be identified through the semi-supervised models for each sub-scene to obtain the model matrix for each sub-scene. Calculate the weight values ​​of the semi-supervised models for each sub-scene based on the model matrix; Based on the weight values, potential users are identified using the semi-supervised model of the multiple sub-scenes.

8. A potential user identification device, characterized in that, The latent user identification device includes: a memory, a processor, and a latent user identification program stored in the memory and executable on the processor, wherein the latent user identification program, when executed by the processor, implements the latent user identification method as described in any one of claims 1 to 7.

9. A storage medium, characterized in that, The storage medium stores a potential user identification program, which, when executed by a processor, implements the potential user identification method as described in any one of claims 1 to 7.

10. A potential user identification device, characterized in that, The potential user identification device includes: a scene segmentation module, a model building module, and a user identification module; The scenario splitting module is used to split the potential user mining scenario into multiple potential user mining sub-scenarios, wherein the potential user mining sub-scenarios include a potential user identification sub-scenarios for permanently holding NFTs, a potential user identification sub-scenarios for limited-trade NFTs, and a potential user identification sub-scenarios for token NFTs. The model building module is used to build multiple sub-scenario semi-supervised models corresponding to multiple potential user mining sub-scenarios based on a multi-model collaborative semi-supervised learning mechanism. The multi-model collaborative semi-supervised learning mechanism includes positive and negative sample equalization processing, multiple base model construction algorithms, pseudo-label annotation technology, and a multi-parameter global search mechanism. The user identification module is used to identify potential users through the multiple sub-scenario semi-supervised models.

Images