A method and system for estimating the type of a seller user in a second-hand e-commerce scenario

Abstract

The application relates to a method and system for estimating the type of a seller user in a second-hand e-commerce scene, wherein the method comprises the following steps: obtaining the behavior data of a seller user in a period of time, using a machine learning model to predict a first prediction score of the type of the seller user as an individual based on the behavior data of the seller user; obtaining text data used by the seller user when publishing a commodity, using a machine learning model to predict a second prediction score of the type of the seller user as an individual based on the text data; obtaining image data used by the seller user when publishing a commodity, using a machine learning model to predict a third prediction score of the type of the seller user as an individual based on the image data; and comprehensively determining the type of the seller user as an individual or a merchant by comprehensively considering the first prediction score, the second prediction score and the third prediction score. The application comprehensively considers various data to distinguish the type of a seller user on a platform as an individual or a merchant, and has high prediction capability, high accuracy and high processing efficiency.

Description

Technical Field

[0001] This invention relates to the field of Internet application technology, and in particular to a method and system for predicting seller user types in a second-hand e-commerce scenario. Background Technology

[0002] Sellers on secondhand e-commerce platforms include both individuals and businesses (such as companies). Transaction models include both C2C (individual sellers selling to individual buyers) and C2B2C (individual sellers selling to businesses, who then sell to individual buyers). To improve transaction speed and volume, the platform has developed a service that recommends individual sellers to businesses. The platform needs to identify individual sellers from its large user base, ensuring high recall and accuracy rates to effectively serve both businesses and sellers, enhance user experience, facilitate transactions, and ultimately increase the platform's GMV (Gross Merchandise Volume). Summary of the Invention

[0003] To address the technical problems existing in the prior art, this invention proposes a method and system for predicting seller user types in a second-hand e-commerce scenario, in order to identify individual sellers among platform users.

[0004] To address the aforementioned technical problems, according to one aspect of the present invention, a method for predicting seller user type in a second-hand e-commerce scenario is provided. The method includes: acquiring seller user behavior data over a period of time; based on the seller user behavior data, using a machine learning model to predict a first predicted score indicating the seller user type as an individual; acquiring text data used by the seller user when posting products; based on the text data, using a machine learning model to predict a second predicted score indicating the seller user type as an individual; acquiring image data used by the seller user when posting products; based on the image data, using a machine learning model to predict a third predicted score indicating the seller user type as an individual; and combining the first predicted score, the second predicted score, and the third predicted score to determine whether the seller user type is an individual or a merchant.

[0005] To address the aforementioned technical problems, according to another aspect of the present invention, a seller user type prediction system for a second-hand e-commerce scenario is provided. This system includes a data acquisition module, a behavior module, a text module, an image module, and a prediction module. The data acquisition module is configured to acquire seller user behavior data over a period of time, text data used by seller users when posting products, and image data. The behavior data includes the time and type of the behavior, and the text data includes the product titles posted by the seller users and / or the nicknames used by the seller users. The behavior module is connected to the data acquisition module and configured to predict a first predicted score indicating the seller user type as an individual based on the seller user's behavior data using a machine learning model. The text module is connected to the data acquisition module and configured to predict a second predicted score indicating the seller user type as an individual based on the text data using a machine learning model. The image module is connected to the data acquisition module and configured to predict a third predicted score indicating the seller user type as an individual based on the image data using a machine learning model. The prediction module is connected to the behavior module, text module, and image module and configured to combine the first, second, and third predicted scores to determine whether the seller user type is an individual or a merchant.

[0006] This invention integrates data from multiple sources to identify whether platform sellers are individuals or merchants, demonstrating strong predictive capabilities, high accuracy, and high processing efficiency. The method described in this invention improves the recall rate and accuracy when recommending individual sellers to merchants downstream, thereby effectively enhancing the platform's user experience. This facilitates transactions for both parties and also increases the platform's revenue. Attached Figure Description

[0007] The preferred embodiments of the present invention will now be described in further detail with reference to the accompanying drawings, wherein:

[0008] Figure 1 This is a flowchart of a method for predicting seller user types in a second-hand e-commerce scenario according to an embodiment of the present invention;

[0009] Figure 2 This is a flowchart of the XGBoost modeling process according to an embodiment of the present invention;

[0010] Figure 3 This is a flowchart for predicting seller user types according to an embodiment of the present invention;

[0011] Figure 4 This is a block diagram of a seller user type prediction system in a second-hand e-commerce scenario provided by an embodiment of the present invention;

[0012] Figure 5This is a schematic block diagram of a data acquisition module provided according to an embodiment of the present invention;

[0013] Figure 6 This is a schematic diagram of a behavior module provided according to an embodiment of the present invention;

[0014] Figure 7 This is a schematic diagram of a text module provided according to an embodiment of the present invention;

[0015] Figure 8 This is a schematic diagram of a text module provided according to another embodiment of the present invention;

[0016] Figure 9 This is a schematic block diagram of an image module provided according to an embodiment of the present invention;

[0017] Figure 10 This is a schematic block diagram of a prediction module provided according to an embodiment of the present invention; and

[0018] Figure 11 This is a flowchart of another method for predicting seller user types in a second-hand e-commerce scenario, provided by an embodiment of the present invention. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] In the following detailed description, reference can be made to the accompanying drawings, which form part of this application and illustrate specific embodiments of the present application. In the drawings, similar reference numerals describe substantially similar components in different figures. Specific embodiments of the present application are described in sufficient detail below to enable those skilled in the art to implement the technical solutions of the present application. It should be understood that other embodiments may also be utilized, or structural, logical, or electrical changes may be made to the embodiments of the present application.

[0021] This invention provides a method and system for predicting seller user types in a second-hand e-commerce scenario. It predicts the type of a seller user from multiple perspectives, thereby accurately identifying the seller user type and enabling accurate recall of individual seller users when recommending them to downstream merchants. This invention identifies seller user types from three aspects: user behavior, the text used by the user when posting products, and the product images. It also uses a machine learning model to predict the seller user type. Therefore, referring to… Figure 1 , Figure 1 This is a flowchart of the seller user type prediction method in the second-hand e-commerce scenario provided in this embodiment. The prediction method includes the following steps:

[0022] Step S1a: Collect data to obtain the raw dataset. The secondhand e-commerce platform maintains user logs in its database. The platform detects users' daily behaviors and records them in real time in the user logs. Therefore, the user logs record various user behaviors in log form, such as logging in, viewing, searching, collecting, posting products, inquiring, or answering inquiries. Whenever a user performs any of these behaviors, it is recorded as a log entry. Each log entry includes the time of the behavior, the type of behavior, and the corresponding operation content. In one embodiment, the platform's user logs for the most recent 30, 60, or 90 days are read based on the platform's data volume, thereby obtaining multiple logs from multiple users to form the raw dataset.

[0023] Step S2a involves performing dimensional analysis and tagging on the logs. The dimensions include the time of occurrence of the behavior, the end time, the type of behavior, and the corresponding operation content. Operation content could include, for example, the title of a published product, the title of a favorited product, or the nickname of a favorite seller. Tagging each log entry with these dimensions facilitates subsequent data processing.

[0024] Step S3a involves acquiring the behavior dataset, text dataset, and image dataset. To meet the input requirements of each model, data from the original dataset is extracted and calculated to obtain different datasets. Specifically, to predict user behavior, the data in the original dataset is calculated according to preset behavioral feature indicators, resulting in multiple behavioral features for each seller user, thus forming the behavior dataset. Feature indicators include quantity-based and frequency-based indicators. For example, counting the number of times the same behavior occurs corresponds to the quantity of corresponding operations, such as the number of products posted, product sales volume, and the number of inquiries or responses. Another example is calculating the frequency of a certain behavior in different periods, such as the cumulative login time in the past week, the number of logins in the past week, the number of purchases in the past month / week, the number of inquiries in the past month / week, the number of favorites in the past month / week, and the number of products posted each month / week. These feature data constitute the behavior dataset. User nicknames and product titles are obtained from the original dataset as the text dataset. Images used by users when posting products are obtained from the original dataset as the image dataset. When a user uses multiple images, the first image is extracted and stored in the image dataset.

[0025] Step S4a, Model Training. Since the purpose of this invention is to identify whether a user is an individual or a merchant, which is a classification problem, a classification model from machine learning can be used for prediction. First, a training set is constructed. In one embodiment, a certain number (e.g., 1000) of user log data with known user types are collected from the original dataset. After processing in steps 2 and 3 above, three training sets are obtained. Samples in the three training sets are labeled as individuals or merchants, such as 1 representing an individual seller and 0 representing a merchant. Then, models are built to obtain a first classification model, a second classification model, and a third classification model. The first classification model is obtained using an XGBoost (Extreme Gradient Boosted Decision Trees) model or a deep learning model. Taking the XGBoost model as an example, its construction process is as follows... Figure 2 As shown, it includes the following steps:

[0026] Step S41a: Import the XGBoost model data package.

[0027] Step S42a: Read the behavior dataset required for modeling, divide the behavior dataset into a training set for training and a test set for testing, and ensure that the ratio of the number of samples in the training set to the number of samples in the test set is 7:3. The samples in the behavior dataset are various user behavior features labeled with 1 or 0.

[0028] Step S43a: Invoke the XGBoost model and train it based on the samples in the training set. If the current XGBoost model is not suitable for this invention, it is modified, for example, by setting the maximum tree depth, learning rate or shrinkage step size, defining the learning task as a binary classification problem, and the number of iterations (i.e., the number of trees the model will use). The XGBoost model is a forward iterative model that trains multiple CART trees, each CART tree obtains a prediction value, and finally the prediction values ​​of multiple CART trees are added together as the final prediction value.

[0029] For the t-th tree and the i-th sample, the model's predicted value is:

[0030]

[0031] in, It is the prediction result of sample i after the t-th iteration; f t (x i ) is the model prediction result for the t-th tree; This is the model prediction result for the (t-1)th tree.

[0032] The original objective function is further obtained as follows:

[0033]

[0034] in, It is the model's loss function. y is the predicted value of the entire model for the i-th sample. i It is the true value of the i-th sample. This is the sum of the complexities of all t trees, which is treated as a regularization term in the function; i is the number of samples, and the first part on the right side of the equals sign is the loss, which represents the difference between the predicted value and the actual value of the t-th tree. j is the cumulative variable, representing the number of trees, indicating that the complexity of each tree is accumulated.

[0035] Step S44a involves making predictions using the trained model based on the test set. During testing, the `predict` function can be called, with the test samples as parameters. Finally, the output prediction percentage is rounded to either 0 or 1 to obtain the predicted value.

[0036] Step S45a: Evaluate the current model based on the prediction results. For example, evaluate the model using metrics such as accuracy, precision, and F1 score.

[0037] Step S46a: Determine if the evaluation is passed, such as whether each indicator meets the requirements. If it does, the model training is complete, the first classification model is obtained, and the training process ends. If the evaluation indicator requirements are not met, then in step S47a, the model is fine-tuned, such as modifying the number of iterations or adjusting the learning rate, until all evaluation indicators meet the requirements.

[0038] Since the second classification model processes text data, in one embodiment of the present invention, the second classification model includes a natural language processing model and a DNN network for performing the binary classification task. In one embodiment, based on the Google open-source language model BERT, product titles or user nicknames from the text dataset are used as corpus for fine-tuning to obtain a language model capable of expressing product titles or user nicknames. The model output vector is fed to the binary classification model composed of the DNN network. After passing through the DNN network, the probability that the seller / user is an individual is output, and after rounding, a predicted value of 0 or 1 is obtained. There can be one or two second classification models. For example, one second classification model can be trained using product titles and user nicknames as samples. Alternatively, they can be trained separately, i.e., trained using product titles and user nicknames as samples respectively, to obtain two second classification models.

[0039] The third classification model is an image-based binary classification model. Before training the third classification model, the images need to be preprocessed, such as normalizing the images to a uniform size and converting color images to grayscale images (binarization) and storing them as samples. The image binary classification model, for example, uses a neural convolutional network (CNN) structure. The samples are pre-labeled and associated. By inputting the samples into the model, the model learns binary classification: a score greater than 0.5 is 1, representing an individual seller; a score less than 0.5 is 0, representing a merchant.

[0040] The training process for the second and third classification models is similar to that of the first classification model. The difference lies in the setting and fine-tuning of the model parameters. For the training and fine-tuning of the language model, binary classification model, and CNN image binary classification model, please refer to the relevant field descriptions, which will not be repeated here.

[0041] Step S5a: Predict the target seller users based on the trained model. The process of predicting seller users in the original dataset is as follows: Figure 3 As shown:

[0042] Step S51a: Collect data on seller users of unknown types from the original dataset as the second dataset. Since the data collected above represents all users on the platform, the types of some users are known. Users of known types can be used as the dataset for model training, and the corresponding data of the remaining users can be used as the second dataset.

[0043] Step S52a: Select a seller user from the second dataset as the target seller user.

[0044] Step S53a: Based on the behavioral characteristic data of the target seller user, a first predicted score (actModel) is predicted for the seller user type as an individual using a first classification model. The first predicted score (actModel) is either 0 or 1.

[0045] Step S54a: Based on the text data of the target seller user, a second prediction score (textModel) is predicted as "individual" using a second classification model. In one embodiment, the target seller user's nickname and the title of their published product are used as inputs to the second classification model to obtain the second prediction score (textModel). In another embodiment, when two second classification models are trained using the seller user's nickname and the title of their published product, the target seller user's nickname and the title of their published product are input into the two models respectively to obtain two prediction values. The average of the two prediction values ​​is then taken as the second prediction score (textModel). When both prediction values ​​are 0, the second prediction score (textModel) is 0; when both prediction values ​​are 1, the second prediction score (textModel) is 1; when one prediction value is 0 and the other is 1, the second prediction score (textModel) is 0.5.

[0046] Step S55a: Based on the image data of the target seller user, use the third classification model to predict the third prediction score picModel for the seller user type as an individual; when the seller user type is predicted to be an individual, the third prediction score picModel is 0, and when the seller user type is predicted to be a merchant, the third prediction score picModel is 1.

[0047] Step S56a: Calculate the overall score of the target seller user according to the following formula:

[0048] Score=Sum(w1*actModel+w2*textModel+w3*picModel)

[0049] Wherein, Score is the overall score; actModel is the first predicted score, textModel is the second predicted score, picModel is the third predicted score, and w1, w2, and w3 are the corresponding weights; in one embodiment, w1 = w2 = w3 = 1.

[0050] Step S57a: Compare the overall score of the target seller user with a threshold. In one embodiment, the threshold is set to 1.5. That is, determine whether the Score is greater than or equal to 1.5. If the Score is greater than or equal to 1.5, in step S58a, determine that the target seller user type is an individual. If the Score is less than 1.5, in step S59a, determine that the target seller user type is a merchant.

[0051] Step S60a: Determine if there are any unpredicted users. If so, return to step S52a to retrieve new target seller users. If all users in the second dataset have been predicted, the prediction process ends.

[0052] In one specific embodiment, steps S56a-S60a can be implemented using a voting algorithm, that is, voting on the first predicted score, the second predicted score, and the third predicted score with equal or unequal weights. For example, when voting with equal weights, if the sum of the three predicted scores is greater than or equal to 2, then the individual seller is identified; otherwise, the merchant is identified. The thresholds and weights can be determined according to actual needs. For example, when behavioral data, text data, and image data are balanced, their weights are set to be equal. If the data for a certain type is significantly larger, the weight of that type of score is increased, and the weight of the type of score with a smaller data volume is decreased accordingly.

[0053] This invention also provides a seller user type prediction system in the context of second-hand e-commerce, such as... Figure 4 As shown, the system includes a data acquisition module 1, a behavior module 2, a text module 3, an image module 4, and a prediction module 5. The data acquisition module 1 is used to acquire log data of seller users over a period of time to form a raw dataset. The user log includes multiple entries, each recording the time of occurrence of a user behavior, the type of behavior, and the corresponding operation content. In one embodiment, such as... Figure 5As shown, the data acquisition module 1 includes a data reading unit 11 and a dimension analysis unit 12. The data reading unit 11 retrieves multiple log entries from seller user logs for a specific period, such as the last 90 days, the last 30 days, or the last two weeks. The dimension analysis unit 12 is connected to the data reading unit 11, performs dimension analysis on the seller user log data, and labels the corresponding dimensions. The dimensions include the seller user's nickname, the time of the behavior, the type of behavior, the content of the operation, the product title, and the corresponding product image. The behavior type includes logging in, posting products, viewing products / sellers, searching for products / sellers, adding products / sellers to favorites, asking or answering inquiries, etc.

[0054] The behavior module 2 is connected to the data acquisition module 1. Based on the seller user's behavior data, it uses a machine learning model to predict a first predicted score for the seller user type as "individual". In one embodiment, such as... Figure 6As shown, the behavior module 2 includes a behavior feature extraction unit 21, a first model training unit 22, and a first prediction unit 23. The behavior feature extraction unit 21 is connected to the data acquisition module 1 and is configured to use seller user log data and one or more dimensional information to calculate various behavior features based on preset behavior feature indicators. For example, it can count the total number of products posted / purchased, the total number of times users favorited / inquired / logged in, and the total login time based on quantity indicators; then, it can calculate corresponding frequency data based on frequency indicators, where a time constant is set to a month, week, or day, and different time constants can be set for different behavior types. For example, it can calculate the number of products posted in the most recent month, the number of products posted each week, the login time in the most recent week, and the number of times users favorited / inquired / purchased / logged in the most recent week, etc. The first model training unit 22 trains a first classification model based on seller user behavior features labeled as individuals or merchants as training set samples. In one embodiment, when the data acquisition module 1 acquires log data from all users, the types of some users are known. Therefore, the first prediction unit 23 can separate the behavioral feature data of users with known user types from the behavioral dataset obtained after feature extraction to form a sample set. In another embodiment, when the data acquisition module 1 acquires user log data of unknown user types, the first prediction unit 23 notifies the data acquisition module 1 to read user log data of known user types, and the behavioral feature extraction unit 21 obtains various behavioral features to form a sample set. The sample set is divided into a training set and a test set, and each sample is associated with its label, which is labeled as 0 or 1. In this embodiment, 1 represents an individual and 0 represents a merchant. The first prediction unit 23 trains the model with samples in the training set, tests the model with samples in the test set, and evaluates the model according to preset evaluation indicators based on the test results. When the model meets the requirements, the model training is completed and provided to the first prediction unit 23. To distinguish it from the models used by the text module and the image module, the model trained and used by the behavioral module is referred to as the first classification model, and the models trained and used by the text module and the image module are referred to as the second and third classification models, respectively. In one embodiment, the first classification model is an XGBoost model or a 3-layer or multi-layer DNN network model. The first prediction unit 23 is connected to the first model training unit 22. It sequentially inputs the behavioral characteristics of users with unknown user types from the behavioral database into the first classification model. The first classification model then predicts a first prediction score for the seller user type as "individual." This score is either 0 or 1. A score of 0 indicates that the seller user type is predicted to be a merchant, while a score of 1 indicates that the seller user type is predicted to be an individual. The process is as follows... Figure 3 As shown, it will not be elaborated further here.

[0055] Text module 3 is connected to data acquisition module 1. Based on the acquired text data, it uses a machine learning model to predict a second predicted score for the seller's user type as "individual". In one embodiment, such as... Figure 7 As shown, the text module 3 includes a text reading unit 31a, a second model training unit 32a, and a second prediction unit 33a. In one embodiment, the text reading unit 31a is connected to the data acquisition module 1, and reads the seller user's nickname and product title from the original dataset to form a text dataset. The construction of the text sample set is similar to the construction of the behavior sample set, and will not be described again here. In this embodiment, the second model training unit 32 trains a second classification model based on natural language semantics using seller user nicknames and product titles labeled as individuals or merchants as training set samples. In this embodiment, the second classification model based on natural language semantics is based on the open-source BERT model and a CNN model for binary classification tasks. In this embodiment, the seller user's nickname and product title in a sample are input text to the BERT model, and a vector representing the seller user is output to the CNN model performing the binary classification task. The CNN model obtains a probability percentage representing that the user is an individual, which is rounded to obtain a predicted value of 0 or 1. After training, the model is tested, evaluated, and fine-tuned using samples in the test set until the evaluation criteria are met. The second prediction unit 33 uses the second classification model to input the user features and product features of each seller user in the text dataset into the second classification model, and obtains a second prediction score for the seller user type as an individual through the second classification model. The second prediction score is 0 or 1.

[0056] In another embodiment, such as Figure 8 As shown, the text module 3 includes a text reading unit 31b, a second model training unit 32b, and a second prediction unit 33b. In this embodiment, the text set obtained by the text reading unit 31b includes a product title set and a user nickname set, and the sample set includes a product title sample set and a user nickname sample set. The second model training unit 32b trains two prediction models based on the two sample sets, namely a fourth model and a fifth model. The second prediction unit uses the fourth model and the fifth model to make predictions respectively, and calculates the average of their respective prediction values ​​to obtain a second prediction value. This value is 0, 1, or 0.5.

[0057] The image module 4 is connected to the data acquisition module 1, and uses a machine learning model to predict a third predicted score for seller users whose user type is "individual" based on the image data. In one embodiment, such as... Figure 9As shown, the image module 4 includes an image extraction unit 41, an image preprocessing unit 42, a third model training unit 43, and a third prediction unit 44. The image extraction unit 41 determines a target image from multiple images used by the user when posting products in the original dataset. For example, it may use the first image among multiple images as the target image, or an image with a complete product image among multiple images as the target image. The image preprocessing unit 42 preprocesses the target image, for example, normalizing the target image to a uniform size, or converting color images to black and white images to simplify processing. The sample set construction process in this module is similar to that in the previous two modules and will not be repeated here. The third model training unit 43 trains a third classification model using product images labeled as individuals or merchants as training set samples. The third classification model is an image binary classification model with a CNN network structure. Its training and fine-tuning can be found in relevant field descriptions and will not be repeated here. The third prediction unit 44 inputs the preprocessed target image of a user into the third classification model to obtain a third prediction score for the seller user type as an individual. The score is 0 or 1, which respectively represent the user type as a merchant and an individual.

[0058] The prediction module 5 is connected to the behavior module 2, text module 3, and image module 4, and combines the first prediction score, the second prediction score, and the third prediction score to determine whether the seller user is an individual or a merchant. In one embodiment, such as Figure 10 The estimation module 5 includes a scoring unit 51 and a decision-making unit 52. The scoring unit 51 receives a first predicted score, a second predicted score, and a third predicted score from the behavior module 2, the text module 3, and the image module 4, and calculates the seller user's overall score according to the following formula:

[0059] Score=Sum(w1*actModel+w2*textModel+w3*picModel)

[0060] Wherein, Score is the overall score; actModel is the first predicted score, textModel is the second predicted score, picModel is the third predicted score, and w1, w2, and w3 are the corresponding weights. In this embodiment, w1 = w2 = w3 = 1. The scoring unit 51 sends the calculated overall score to the decision unit 52. The decision unit 52 compares the seller user's overall score with a threshold, and determines the seller user's type as an individual or a merchant based on the comparison result. For example, setting the threshold to 1.5, when Score ≥ 1.5, the seller user's type is determined to be an individual; when Score < 1.5, the seller user's type is determined to be a merchant.

[0061] Figure 11This is a flowchart of a method for predicting seller user types in a second-hand e-commerce scenario according to another embodiment of the present invention. In this embodiment, the method includes the following steps:

[0062] Step S1b involves collecting user logs for known user types to obtain the original sample set. For example, collecting user logs for known user types from the past month, ensuring the ratio of individual users to merchant users is between 4:6 and 6:4. This guarantees a balanced sample with labels of 0 and 1.

[0063] Step S2b involves processing the original sample set by processing each user's log data into behavioral features, user text, and product images, thereby obtaining three sample sets: behavioral sample set, text sample set, and image sample set.

[0064] Step S3b: Train the first classification model, the second classification model, and the third classification model respectively.

[0065] Step S4b: Collect user logs of unknown user types in the platform to obtain the raw dataset.

[0066] Step S5b: Process the original dataset to obtain the behavior dataset, text dataset, and image dataset.

[0067] Step S6b: Select one user as the target seller user.

[0068] Step S7b: Based on the behavioral data, text data, and image data of the target seller user, predictions are made using the first classification model, the second classification model, and the third classification model to obtain the first predicted score, the second predicted score, and the third predicted score, respectively.

[0069] Step S8b: Calculate the comprehensive score of the target seller user based on the first predicted score, the second predicted score, and the third predicted score.

[0070] Step S9b: Determine whether the comprehensive score is greater than or equal to the threshold. If it is greater, then in step S10b, confirm that the target seller user is an individual. If the comprehensive score is less than the threshold, then in step S11b, confirm that the target seller user is a merchant.

[0071] Step S12b: Determine if there are any users of unknown type. If so, return to step S6b. If not, end the process.

[0072] In this embodiment, various classification models are first trained, and then user logs of unknown user types are read uniformly. This allows for targeted identification of users with unknown user types, resulting in high processing efficiency. The invention uses a large number of balanced samples for model training, resulting in strong predictive ability and high accuracy. The method described in this invention improves the recall and accuracy when recommending individual sellers to merchants, effectively enhancing the platform's user experience and increasing platform revenue while facilitating transactions between the two parties.

[0073] The above embodiments are for illustrative purposes only and are not intended to limit the invention. Those skilled in the art can make various changes and modifications without departing from the scope of the invention. Therefore, all equivalent technical solutions should also fall within the scope of the invention.

Claims

1. A method for predicting seller user types in a second-hand e-commerce scenario, including: Based on the amount of data from the second-hand e-commerce platform, read the user logs of the second-hand e-commerce platform to obtain multiple logs from multiple users to form the original dataset; The user logs record various user behaviors on the second-hand e-commerce platform in log form. The data in the original dataset are calculated according to preset behavioral feature indicators to obtain multiple behavioral features for each seller user. The behavioral feature indicators include quantity indicators and frequency indicators. The quantity indicators include the number of times the same behavior occurs and the quantity of corresponding operation content. The frequency indicators include the frequency of a certain behavior occurring in different periods. The behavioral feature data of seller users over a period of time is obtained. Based on the behavioral feature data of seller users, a machine learning model is used to predict the first predicted score of seller user type as individual. Obtain the text data used by the seller user when posting products, and based on the text data, use a machine learning model to predict a second predicted score for the seller user type as an individual; the text data includes the product title posted by the seller user and / or the nickname used by the seller user. The system acquires image data used by sellers when posting products, and uses a machine learning model to predict a third prediction score for sellers who are individuals; the image data includes images with complete product images used by sellers when posting products. The first predicted score, the second predicted score, and the third predicted score are combined to determine whether the seller user is an individual or a merchant.

2. The method according to claim 1, further comprising the step of obtaining the behavioral characteristics of the seller user based on the seller user's behavioral data: Obtain multiple dimensions of information for each piece of behavioral data, including the time of the behavior, the type of behavior, and the corresponding operation content; and Multiple behavioral characteristics are obtained by using the information from the multiple dimensions and calculating based on preset behavioral characteristic indicators.

3. The method according to claim 2, wherein the type of behavior includes one or more of the following: logging in, posting products, viewing seller information, searching seller information, adding seller items to favorites, and consulting or replying to inquiries.

4. The method according to claim 1, further comprising: The fourth prediction score is based on the product title, which indicates that the seller user type is an individual. The fifth prediction score is given based on the seller user's nickname, indicating that the seller user type is an individual. as well as Calculate the weighted average of the fourth and fifth predicted scores, and use the weighted average as the second predicted score.

5. The method according to claim 1 or 4, further comprising: The seller user's overall rating is calculated using the following formula: Score=Sum(w1 actModel+w2 textModel+w3 picModel) Where Score is the overall score; actModel is the first predicted score, textModel is the second predicted score, picModel is the third predicted score, and w1, w2, and w3 are the corresponding weights; and The seller user's overall score is compared with a threshold. If the overall score is greater than or equal to the threshold, the seller user is determined to be an individual. If the overall score is less than the threshold, the seller user is determined to be a merchant.

6. A seller user type prediction system for a second-hand e-commerce scenario, including: The data acquisition module reads user logs from the secondhand e-commerce platform based on the amount of data available, thus obtaining multiple logs from multiple users to form the original dataset. The user logs record various user behaviors on the secondhand e-commerce platform. The data in the original dataset is calculated according to preset behavioral characteristic indicators, resulting in multiple behavioral characteristics for each seller user. These behavioral characteristic indicators include quantity-based indicators and frequency-based indicators. Quantity-based indicators include the number of times the same behavior occurs and the quantity of corresponding operation content, while frequency-based indicators include the frequency of a certain behavior occurring at different times. The module also acquires text data used by seller users when posting products, including the product title and / or the seller user's nickname. Finally, it acquires images with complete product images used by seller users when posting products as image data. The behavior module, which is connected to the data acquisition module, is configured to use a machine learning model to predict the first predicted score of the seller user type as an individual based on the seller user's behavioral characteristic data. The text module, which is connected to the data acquisition module, is configured to use a machine learning model to predict a second predicted score for seller users of the type "individual" based on the text data. An image module, connected to a data acquisition module, is configured to use a machine learning model to predict a third prediction score for seller users of the type "individual" based on the image data. as well as The prediction module, which is connected to the behavior module, text module and image module, is configured to combine the first prediction score, the second prediction score and the third prediction score to determine whether the seller user is an individual or a merchant.

7. The system according to claim 6, wherein the data acquisition module comprises: The data reading unit is configured to retrieve multiple seller user log data entries from the seller user logs. as well as The dimensional analysis unit, which is connected to the data reading unit, is configured to perform dimensional analysis on seller user log data and mark the corresponding dimensions, including the time of occurrence of seller behavior, behavior type, and operation content.

8. The system of claim 7, wherein the behavior module comprises: The behavioral feature extraction unit is connected to the data acquisition module and is configured to use seller user log data and multiple dimensions of information to calculate various behavioral features based on preset behavioral feature indicators. The first model training unit is configured to use seller user behavior features labeled as individuals or merchants as training set samples to train the first classification model. as well as The first prediction unit is connected to the first model training unit and the behavior feature extraction unit. It is configured to input the behavior features of seller users into the first classification model, and obtain a first prediction score for the seller user type as an individual through the first classification model.

9. The system according to claim 8, wherein the first classification model is an XGBoost model, a three-layer or multi-layer DNN network model.

10. The system of claim 7, wherein the text module comprises: A text reading unit, connected to the data acquisition module, is configured to read the seller's nickname and product title; The second model training unit, which is connected to the text reading unit, is configured to train a second classification model based on natural language semantics using seller user nicknames and / or product titles labeled as individuals or merchants as training set samples. as well as The second prediction unit is connected to the text reading unit and the second model training unit. It is configured to input the seller user's nickname and / or product title into the second classification model, and obtain a second prediction score for the seller user type as an individual through the second classification model.

11. The system according to claim 10, wherein the second model training unit uses the seller user nickname labeled as an individual or a merchant and the product title labeled as an individual or a merchant as samples in the training set to train two second classification models respectively; correspondingly, the second prediction unit uses the seller user nickname and the product title published by the seller as inputs to the two second classification models to obtain the corresponding fourth prediction score and fifth prediction score, and uses the weighted average of the fourth prediction score and the fifth prediction score as the second prediction score.

12. The system of claim 7, wherein the image module comprises: An image extraction unit, connected to the data acquisition module, is configured to determine a target image from multiple images used by the user when publishing a product. An image preprocessing unit, connected to the image extraction unit, is configured to preprocess the target image; The third model training unit is configured to use product images labeled as individuals or merchants as training set samples to train the third classification model; as well as The third prediction unit, which is connected to the third model training unit and the image preprocessing unit, is configured to input the preprocessed target image into the third classification model to obtain a third prediction score for the seller user type as an individual.

13. The system of claim 6, wherein the estimation module comprises: The rating unit is configured to calculate the seller user's overall rating according to the following formula: Score=Sum(w1 actModel+w2 textModel+w3 picModel) Where Score is the overall score; actModel is the first predicted score, textModel is the second predicted score, picModel is the third predicted score, and w1, w2, and w3 are the corresponding weights; and The decision unit, which is connected to the scoring unit, is configured to determine whether the seller user is an individual or a merchant based on the seller user's overall score and threshold.

Images