Intent prediction methods, devices, electronic equipment and storage media

Abstract

This application relates to the field of computer technology, providing an intent prediction method, apparatus, electronic device, and storage medium. The intent prediction method includes: first, acquiring target user data, which includes data associated with a target product; then, determining comprehensive characteristics corresponding to the target user based on the target user data, including user profile features, sentiment features, and behavioral features; finally, obtaining a target probability based on the comprehensive characteristics of the target user and a target prediction strategy, whereby the target probability represents the probability that the target user will purchase the target product. The target prediction strategy is correlated with the amount of historical user data corresponding to the target product, and the amount of historical user data is positively correlated with the product's shelf life. This allows the target prediction strategy to evolve gradually over time, enabling it to match prediction needs at different stages and improving prediction accuracy.

Description

Technical Field

[0001] This application relates to the field of computer technology, and in particular to an intention prediction method, apparatus, electronic device, and storage medium. Background Technology

[0002] In recent years, although machine learning and multimodal data processing technologies have provided important support for predicting user intent, these technologies often use a single approach to prediction, which fails to meet the prediction needs in complex scenarios and results in low prediction accuracy. Summary of the Invention

[0003] This application provides an intent prediction method, apparatus, electronic device, and storage medium to improve the accuracy of intent prediction.

[0004] In a first aspect, this application provides an intent prediction method, the method comprising:

[0005] Obtain target user data for the target user, wherein the target user refers to a user who has not purchased any product on the target program, and the target user data includes data associated with the target product;

[0006] Based on the target user data, the comprehensive characteristics corresponding to the target user are determined, including user profile characteristics, emotional characteristics, and behavioral characteristics;

[0007] Based on the comprehensive characteristics of the target user and the target prediction strategy, a target probability is obtained. The target probability represents the probability that the target user will purchase the target product. The target prediction strategy is related to the amount of historical user data corresponding to the target product. The amount of historical user data is positively correlated with the listing duration of the target product.

[0008] In an optional implementation, if the number of historical user data does not exceed a first preset threshold or the number of reference user data with purchase results in the historical user data does not reach a first preset condition, the target prediction strategy is a business prediction rule, which is set according to the feature types included in the comprehensive features.

[0009] The step of obtaining the target probability based on the comprehensive characteristics of the target user and the target prediction strategy includes:

[0010] The target probability is calculated based on the comprehensive characteristics of the target user and the business prediction rules.

[0011] In an optional implementation, when the amount of historical user data exceeds a first preset threshold but does not exceed a second preset threshold, and the amount of reference user data with purchase results in the historical user data reaches a first preset condition, the target prediction strategy is at least one of the preset prediction strategies, the preset prediction strategy includes business prediction rules and a target semi-supervised model, and the first preset threshold is less than the second preset threshold.

[0012] Before acquiring the target user data of the target user, the method further includes:

[0013] Based on the reference user data, the initial semi-supervised model is trained to obtain the target semi-supervised model.

[0014] In an optional implementation, obtaining the target probability based on the comprehensive characteristics corresponding to the target user and the target prediction strategy includes:

[0015] The comprehensive features corresponding to the target user are input into the target semi-supervised model, and the target semi-supervised model is used to predict the target probability.

[0016] In an optional implementation, obtaining the target probability based on the comprehensive characteristics corresponding to the target user and the target prediction strategy includes:

[0017] Calculate the first probability based on the comprehensive characteristics of the target user and the business prediction rule;

[0018] The comprehensive features corresponding to the target user are input into the target semi-supervised model, and the target semi-supervised model is used to predict the second probability;

[0019] According to a preset selection rule, the target probability is selected from the first probability and the second probability.

[0020] In an optional implementation, when the amount of historical user data exceeds a second preset threshold and the number of reference user data with purchase results in the historical user data reaches a second preset condition, the target prediction strategy is a target supervised model.

[0021] Before acquiring the target user data of the target user, the method further includes:

[0022] According to different training methods, the initial supervised model is trained based on the reference user data to obtain multiple trained supervised models;

[0023] The predictive capabilities of the multiple trained supervised models are evaluated, and the model with the best predictive capability is selected as the target supervised model.

[0024] In an optional implementation, the method further includes:

[0025] Regularly collect data on newly added reference users;

[0026] The multiple trained supervised models are updated based on the newly added reference user data;

[0027] The predictive power of each updated trained supervised model is evaluated, and the updated trained supervised model with the best predictive power is selected as the target supervised model.

[0028] Secondly, this application provides an intent prediction device, the device comprising:

[0029] The acquisition module is used to acquire target user data of the target user, wherein the target user refers to a user who has not purchased any product on the target program, and the target user data includes data associated with the target product;

[0030] The determination module is used to determine the comprehensive characteristics corresponding to the target user based on the target user data, wherein the comprehensive characteristics include user profile characteristics, emotional characteristics, and behavioral characteristics;

[0031] The prediction module is used to obtain a target probability based on the comprehensive characteristics of the target user and the target prediction strategy. The target probability represents the probability that the target user will purchase the target product. The target prediction strategy is related to the amount of historical user data corresponding to the target product. The amount of historical user data is positively correlated with the listing duration of the target product.

[0032] Thirdly, this application provides an electronic device including a processor and a memory, wherein the memory stores a computer program, and when the processor executes the computer program, it implements the intent prediction method described in any of the foregoing embodiments.

[0033] Fourthly, this application provides a storage medium storing a computer program thereon, which, when executed by a processor, implements the intent prediction method described in any of the foregoing embodiments.

[0034] This application provides an intent prediction method, apparatus, electronic device, and storage medium. The intent prediction method includes: first, acquiring target user data, which includes data associated with a target product; then, determining comprehensive characteristics of the target user based on the target user data, including user profile features, sentiment features, and behavioral features; finally, obtaining a target probability based on the comprehensive characteristics of the target user and a target prediction strategy. The target probability represents the probability that the target user will purchase the target product. The target prediction strategy is correlated with the amount of historical user data corresponding to the target product, and the amount of historical user data is positively correlated with the product's shelf life. This allows the target prediction strategy to evolve gradually over time, enabling it to match prediction needs at different stages and improving prediction accuracy. Attached Figure Description

[0035] To more clearly illustrate the technical solutions in the embodiments of this application, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0036] Figure 1 This illustration shows one of the flowcharts of the intent prediction method provided in the embodiments of this application.

[0037] Figure 2 An example diagram of the intent prediction method provided in the embodiments of this application is shown.

[0038] Figure 3 The second schematic flowchart of the intent prediction method provided in the embodiments of this application is shown.

[0039] Figure 4 The third schematic diagram of the intent prediction method provided in the embodiments of this application is shown.

[0040] Figure 5 The fourth schematic flowchart of the intent prediction method provided in the embodiments of this application is shown.

[0041] Figure 6 The fifth schematic diagram of the intent prediction method provided in the embodiments of this application is shown.

[0042] Figure 7 A functional block diagram of the intent prediction device provided in an embodiment of this application is shown.

[0043] Figure 8 A schematic diagram of the structure of an electronic device provided in an embodiment of this application is shown. Detailed Implementation

[0044] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the specific technical solutions of this application will be further described in detail below with reference to the accompanying drawings of the embodiments of this application. The following embodiments are used to illustrate this application, but are not intended to limit the scope of this application.

[0045] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used herein is for the purpose of describing embodiments of this application only and is not intended to limit this application.

[0046] In the following description, references are made to “some embodiments,” which describe a subset of all possible embodiments. However, it is understood that “some embodiments” may be the same subset or different subsets of all possible embodiments and may be combined with each other without conflict.

[0047] In recent years, although machine learning and multimodal data processing technologies have provided important support for predicting user intent, these technologies often use a single approach to prediction. This makes it impossible to meet the prediction needs of complex scenarios in various industries, and it is also unable to adapt to changes in prediction needs caused by data changes, resulting in low prediction accuracy.

[0048] Therefore, this application provides an intent prediction method. By acquiring target user data, it determines the target user's user profile characteristics, emotional characteristics, and behavioral characteristics to obtain the target user's comprehensive characteristics. This comprehensive characteristic is then combined with a target prediction strategy to predict the probability of the target user purchasing a target product. Furthermore, the target prediction strategy is related to the amount of historical user data corresponding to the target product, and the amount of historical user data is positively correlated with the product's shelf life. This allows the prediction strategy to be dynamically updated based on the amount of historical user data accumulated over the product's shelf life. This enables the prediction strategy to be updated and optimized over time, thereby improving the accuracy of the prediction.

[0049] To make the objectives and technical solutions of this application clearer and more intuitive, the intent prediction method provided by the embodiments of this application will be described in detail below with reference to the accompanying drawings and examples. It should be understood that the specific embodiments described herein are merely illustrative of this application and are not intended to limit this application.

[0050] The intent prediction method provided in this application is applied to electronic devices, including but not limited to smartphones, personal computers, tablets, wearable devices, netbooks, and personal digital assistants. The functions implemented by this method can be achieved by a processor in the electronic device calling program code. Of course, the program code can be stored in a computer storage medium. Therefore, the electronic device includes at least a processor and a storage medium. The following will use an electronic device as the execution subject to describe the various steps and corresponding technical effects of the methods provided in this application.

[0051] Please see Figure 1 This is a flowchart illustrating an intent prediction method provided in an embodiment of this application.

[0052] Step S212: Obtain target user data. Target user refers to a user who has not purchased any product on the target program. Target user data includes data associated with the target product.

[0053] In this embodiment, the target program refers to a network platform used to display or provide goods in a specific industry. These goods can be actual items or services. For example, if the target program is a network platform for the automotive industry, then the goods could be various types of cars. If the target program is a network platform for the education industry, then the goods could be various teaching services. Furthermore, a target user can be understood as a user who has not yet purchased any goods within the target program. Purchasing can be understood as the process of transferring a product, such as transferring usage rights through resource exchange.

[0054] To facilitate understanding, the following example illustrates the scenario where the target application is an online platform in the education industry, and the target users are parents of students who have not yet purchased any educational services. For instance, text, voice, and behavioral data generated by these parents during their research into a particular educational service can be collected in various ways to obtain target user data associated with the target product.

[0055] In one embodiment, text-type data can be obtained through the interactive interface of the network platform. For example, the interactive interface of the network platform can be used to obtain online consultation information from the student's parents regarding the teaching service, such as the teaching content, teaching duration, and number of teaching sessions. In one embodiment, voice-type data can be collected through a user-authorized voice capture device. For example, assuming the student's parents have authorized the network platform to record their voice information, the sales personnel can use the network platform to make a voice call to the student's parents to introduce the detailed teaching content and record the conversation between the sales personnel and the student's parents. In one embodiment, behavioral-type data can be obtained through the tracking mechanism of the network platform. For example, the network platform's tracking mechanism can collect the student's parents' browsing behavior, such as browsing duration and number of views, and can also collect the student's parents' interaction behavior with the sales personnel, such as the duration and number of voice communications.

[0056] Step S214: Based on the target user data, determine the comprehensive characteristics corresponding to the target user. The comprehensive characteristics include user profile characteristics, emotional characteristics, and behavioral characteristics.

[0057] In this embodiment, the obtained target user data can be preprocessed, including text, voice, and behavioral data. For example, data cleaning and / or standardization operations can be performed on this data to improve data quality and ensure consistency. Furthermore, feature extraction is performed on the preprocessed target user data to obtain user profile features, emotional features, and behavioral features corresponding to the target user.

[0058] In one embodiment, a large language model can be used to perform semantic analysis on the text-type data in the preprocessed target user data to generate discrete or numerical feature values, thereby obtaining user profile features. These user profile features may include numerical metrics such as product acceptance, price sensitivity, urgency of need, willingness to meet in person, willingness to communicate again, spending power, prediction range, and decision-making factors. In another embodiment, feature construction can be performed on the behavioral type data in the preprocessed target user data to statistically analyze different behaviors, thus obtaining behavioral features. These behavioral features may include browsing statistics, communication statistics, etc.

[0059] In one embodiment, the SER (Speech Emotion Recognition) module in a speech recognition model such as Sense Voice can be used to perform sentiment analysis on the preprocessed speech type data of the target user data to obtain a sentiment score. This sentiment score is then normalized to obtain sentiment features consistent with the dimensions of user profile features and behavioral features. Furthermore, the sentiment features can represent the target user's emotional state towards the target product, such as positive, neutral, negative, excited, or hesitant. Moreover, the sentiment score can be normalized according to a preset formula to obtain the sentiment features. The preset formula is as follows:

[0060] y = (x / 100).clip(0, 1);

[0061] Where y represents the sentiment trait, x represents the sentiment score, .clip() represents the clipping function, and clip(0, 1) means that the value is forced to be 0 when it is less than 0 and forced to be 1 when it is greater than 1. It is used to ensure that the value of the sentiment trait is between 0 and 1.

[0062] Subsequently, feature fusion is performed on the user profile features, sentiment features, and behavioral features corresponding to the target user to obtain the comprehensive features corresponding to the target user. In one embodiment, fusion weights can be pre-set for the user profile features, sentiment features, and behavioral features. Then, a weighted sum can be performed based on the user profile features and their fusion weights, the sentiment features and their fusion weights, and the behavioral features and their fusion weights corresponding to the target user to obtain the comprehensive features corresponding to the target user. In another embodiment, the user profile features, sentiment features, and behavioral features corresponding to the target user can be input into a large language model, and the large language model can be used to generate the comprehensive features corresponding to the target user.

[0063] It can be understood that the embodiments of this application improve the accuracy of prediction by collecting multiple types of data as user data, that is, by using multimodal data to predict intent.

[0064] Step S216: Based on the comprehensive characteristics of the target user and the target prediction strategy, the target probability is obtained. The target probability represents the probability that the target user will purchase the target product. The target prediction strategy is related to the amount of historical user data corresponding to the target product. The amount of historical user data is positively correlated with the listing time of the target product.

[0065] In this embodiment, the target prediction strategy can be understood as the strategy currently used to predict intent. Furthermore, the target prediction strategy can be a pre-set rule or a pre-trained model. In one embodiment, the probability of a target user purchasing a target item can be calculated based on the comprehensive characteristics corresponding to the target user according to a preset rule, thus obtaining the target probability. In another embodiment, a preset model can be used to predict the probability of a target user purchasing a target item based on the comprehensive characteristics corresponding to the target user, thus obtaining the target probability. In yet another embodiment, both preset rules and a preset model can be used simultaneously to obtain their respective probabilities based on the comprehensive characteristics corresponding to the target user, and one of these probabilities can be selected as the target probability.

[0066] In this embodiment, as the listing duration of the target product gradually increases, the number of users inquiring about and / or purchasing the target product will gradually increase, leading to a gradual increase in the amount of historical user data corresponding to the target product. Consequently, the amount of user data that can be used as a basis for prediction will also gradually increase. Therefore, this embodiment utilizes this point to establish a correlation between the target prediction strategy and the listing duration of the target product through the amount of historical user data. This allows the target prediction strategy to gradually evolve over time, enabling it to match the prediction needs at different stages and thus improving the accuracy of predictions. The listing duration can be understood as the time from the moment the target product is first displayed on the online platform to the current moment.

[0067] It is understandable that as the listing period of the target product gradually increases, the amount of historical user data corresponding to the target product will accumulate to different stages. Therefore, the prediction strategies that this application embodiment can provide at different stages will also be different. In order to better understand this application, the prediction strategies that can be provided at each stage will be described below.

[0068] If the number of historical user data points corresponding to the target product, such as S, does not exceed the first preset threshold, such as 200 (i.e., S≤200), it indicates that the target product has been on the market for a relatively short time and has a small number of historical user data points, meaning it is in stage T1. Furthermore, only one prediction strategy can be provided in stage T1: the business prediction rule. This business prediction rule is set based on the user profile features, sentiment features, and behavioral features included in the comprehensive features.

[0069] If the number of historical user data points (S) corresponding to the target product exceeds a first preset threshold (e.g., 200), but the number of reference user data points with purchase results (C) within the historical user data does not meet a first preset condition (e.g., the number of reference user data points (C) does not reach a preset reference threshold (e.g., 10), or the percentage of reference user data to historical user data (C / S*100%) does not reach a preset percentage range (e.g., [5%, 20%]). That is, S > 200 and C < 10, or S > 200 and (C / S*100%) < 5%, it indicates that the target product has been listed for a period of time, the number of historical user data points has increased, but the number of user data points with purchase results is relatively small; this indicates that the product is in stage T2. Furthermore, the only predictive strategy available in stage T2 is the business prediction rule.

[0070] If the number of historical user data corresponding to the target product, such as S, exceeds a first preset threshold (e.g., 200) but does not exceed a second preset threshold (e.g., 2000), and the number of reference user data with purchase results, such as C, reaches a first preset condition (e.g., the number of reference user data, such as C, reaches a preset reference threshold (e.g., 10), or the percentage of reference user data to historical user data (e.g., (C / S*100%)) reaches a preset percentage range (e.g., [5%, 20%]), that is, 200 < S ≤ 2000 and C ≥ 10, or 200 < S ≤ 2000 and 5% ≤ (C / S*100%) ≤ 20%, it indicates that the target product has been listed for a relatively long time, and the number of historical user data and the number of user data with purchase results have increased, i.e., it is in stage T3. In stage T3, this embodiment of the application will use the reference user data and train the model through a semi-supervised learning training method to obtain the target semi-supervised model. Then, there are two prediction strategies that can be provided in stage T3: the first is the business prediction rule, and the second is the target semi-supervised model. The specific implementation method for obtaining the target semi-supervised model will be introduced in subsequent embodiments.

[0071] If the number of historical user data corresponding to the target product, such as S, exceeds a second preset threshold, such as 2000, and the number of reference user data with purchase results, such as C, in the historical user data reaches a second preset condition, for example, the number of reference user data, such as C, reaches a preset reference threshold, such as 10, or the percentage of reference user data to historical user data, such as (C / S*100%), exceeds a preset percentage threshold of 20% (i.e., S > 2000 and C ≥ 10, or S > 2000 and (C / S*100%) > 20%), it indicates that the target product has been listed for a long time, and the number of historical user data and the number of user data with purchase results are both very large, i.e., it is in stage T4. In stage T4, this embodiment of the application will use reference user data and train the model through supervised learning training methods to obtain a target supervised model. Then, there are three prediction strategies that can be provided in stage T4: the first is business prediction rules, the second is a target semi-supervised model, and the third is a target supervised model. The specific implementation method of obtaining the target supervised model will be introduced in subsequent embodiments.

[0072] For ease of understanding, this application provides a table to illustrate the relationship between the amount of user data at each stage in the above examples and the available prediction strategies.

[0073]

[0074] In the table above, S represents the amount of historical user data, C represents the amount of reference user data, and (C / S*100%) represents the percentage of reference user data to historical user data. It should be understood that the first preset threshold 200, the second preset threshold 2000, the preset reference threshold 10, the preset percentage range [5%, 20%], and the preset percentage threshold 20% in the above embodiments and tables are merely examples. These threshold parameters can be set according to actual applications, and this application does not limit them.

[0075] It can be understood that, in this embodiment of the application, as the time for which a product is listed gradually increases, the amount of user data corresponding to the product will also gradually increase. In this way, the available prediction strategy can be updated and optimized based on the amount of user data at different stages, thereby realizing the gradual evolution of the prediction strategy with time characteristics and providing technical support for improving the accuracy of prediction.

[0076] The following describes the implementation method for obtaining the target semi-supervised model. That is, before step S212, the intent prediction method provided in this application embodiment may further include: step S200, training the initial semi-supervised model based on reference user data to obtain the target semi-supervised model.

[0077] For ease of understanding, this application provides an example diagram, please refer to it. Figure 2 The system can obtain the number of historical user data points corresponding to the target product and the number of reference user data points with purchase results, and then evaluate these two types of user data. If the number of historical user data points does not exceed a second preset threshold (e.g., 2000), and the percentage of reference user data points to historical user data points does not exceed a preset percentage threshold (e.g., 20%), then it indicates that this is stage T3. It is understandable that although a certain number of parameter user data points with purchase results are accumulated in stage T3, the number of these reference user data points has not yet reached the amount required for supervised learning. Therefore, in this embodiment, a semi-supervised learning approach is used to train the model in stage T3.

[0078] In this implementation, firstly, an autoencoder can be used to reconstruct features from parameter user data with actual purchase results and other user data from historical user data without actual purchase results. That is, the autoencoder compresses the input user data into a low-dimensional representation using an encoder, and then reconstructs it using a decoder, thereby learning the latent feature structure and distribution patterns in the input user data. This helps extract representative user data features and reduces the impact of noise, thus providing high-quality data for subsequent model training.

[0079] Then, pseudo-labels are determined for the reconstructed user data, which involves constructing virtual purchase results for these other user data to expand the amount of user data with purchase results. In one embodiment, K-means clustering can be used to determine the pseudo-labels for the reconstructed user data. This involves dividing the reconstructed user data into multiple groups and then determining pseudo-labels for each group of other user data based on the similarity between the parameter user data and each group. In another embodiment, label propagation can be used to determine the pseudo-labels for the reconstructed user data. This involves constructing a graph model based on the similarity between the reconstructed user data and the reference user data, and then using the graph model to transmit the actual purchase results of the reference user data to other similar user data to determine the pseudo-label for each other user data.

[0080] Finally, the reference user data with actual purchase results is combined with other user data with virtual purchase results to obtain a sample set. This sample set is then used to train the initial semi-supervised model, resulting in the target semi-supervised model. This allows for the provision of two prediction strategies in stage T3. That is, as the amount of user data increases, the prediction strategies provided by this embodiment also increase, thereby realizing the evolution of prediction strategies over time.

[0081] The following describes the implementation method for obtaining the target supervised model. Specifically, before step S212, the intent prediction method provided in this embodiment may further include: step S202, training the initial supervised model according to different training methods and reference user data to obtain multiple trained supervised models. Step S204, evaluating the predictive capabilities of the multiple trained supervised models, and selecting the model with the best predictive capability as the target supervised model.

[0082] For better understanding, please continue reading. Figure 2 The system can obtain the number of historical user data points corresponding to the target product and the number of reference user data points with purchase results, and then evaluate these two types of user data. If the number of historical user data points exceeds a second preset threshold (e.g., 2000), and the percentage of reference user data to historical user data exceeds a preset percentage threshold (e.g., 20%), then it indicates that this is stage T4. It is understandable that stage T4 accumulates a large number of parameter user data points with purchase results, reaching the quantity required for supervised learning. Therefore, in this embodiment, supervised learning is used to train the model in stage T4.

[0083] In the embodiments of this application, the initial supervised model can be trained using different training methods and reference user data. For example, the initial supervised model can be trained using XG Boost, Light GBM, and Transformer methods respectively, based on reference user data, resulting in three trained supervised models. It should be understood that, in addition to these three training methods, other algorithms can also be used to train the model.

[0084] For XG Boost training, the learning rate can be set to 0.1, the maximum depth to 6, and the number of iterations to 100. For Light GBM training, the learning rate can be set to 0.05, the feature binning to 100, and the number of iterations to 100. For both training methods, cross-entropy (log loss) is used as the loss function, and the model parameters are optimized using their built-in gradient boosting mechanism or stochastic gradient descent algorithm. For Transformer training, the model can be configured with 6 encoder layers and 8 attention heads, with the number of iterations set to 100. Parameter updates are performed using the cross-entropy loss function and the Adam optimizer (initial learning rate 0.001).

[0085] Subsequently, the three trained supervised models are evaluated, and the model with the best predictive ability is selected as the target supervised model based on the evaluation results. For example, a 5-fold cross-validation method can be used for model evaluation. It is understood that the model evaluation method for each trained supervised model in this embodiment is similar; for brevity, the following description uses a single trained supervised model.

[0086] For example, a sample set consisting of reference user data can be divided into five parts. Four parts are used as the training set to train the model, and the remaining part is used as the validation set to validate the trained model. Evaluation metrics used to assess the predictive ability of the trained supervised model are recorded. Furthermore, evaluation metrics can include one or a combination of accuracy, precision, recall, F1 score, AUC-ROC (Area Under the Receiver Operating Characteristic Curve), log loss, and PR-AUC (Precision-Recall Area Under Curve).

[0087] Accuracy, representing the ratio of correctly predicted samples to the total sample size, is simple and intuitive, making it useful for preliminary evaluation of a model's overall performance. Precision, representing the proportion of samples predicted to have purchased a product that actually did, is expressed as: Precision = (True Positive Samples / (True Positive Samples + False Positive Samples)). True positive samples represent those predicted to have purchased a product and actually did, while false positive samples represent those predicted to have purchased a product but did not. A high precision rate indicates fewer users who received ineffective follow-ups, making it suitable for scenarios prioritizing high-quality predictions. Recall rate represents the proportion of correctly predicted samples out of the actual product purchases. It can be expressed as: Recall Rate = (True Positive Samples / (True Positive Samples + False Negative Samples)). True positive samples represent the number of samples that were predicted to purchase the product and actually did so, while false negative samples represent the number of samples that were predicted not to purchase the product but actually did. A high recall rate indicates that fewer potential customers have been missed by sales personnel, making it suitable for scenarios emphasizing user coverage.

[0088] The F1 score is the harmonic mean of precision and recall, expressed as: F1 score = 2 * (precision * recall) / (precision + recall). The F1 score balances precision and recall, and is suitable for scenarios where both precision and recall are critical. AUC-ROC (Area Under the Receiver Operating Characteristic Curve) represents the area under the curve of a model's ability to distinguish between purchasing and not purchasing. Its value ranges from 0 to 1, with higher values ​​indicating better ability. AUC-ROC is suitable for evaluating the robustness of a model to probability predictions, especially when the predicted probability distribution is uneven. The log loss value measures the difference between the predicted probability and the actual label, expressed as: Log loss = -(1 / n) * Σ[y * log(p) + (1-y) * log(1-p)], where y represents the true label of the sample, and p represents the predicted probability of the sample. The log loss value is suitable for evaluating the accuracy of probability predictions. PR-AUC represents the area under the precision-recall curve, and is particularly suitable for situations where the sample set is class-imbalanced (such as when there are few samples in the reference user data). PR-AUC focuses on the predictive performance of the positive class (i.e., purchased goods), and is suitable for scenarios where the business focuses more on the positive class.

[0089] In one embodiment, the F1 score and AUC-ROC can be used as the core evaluation metrics for assessing the model's predictive ability, while accuracy, precision, recall, log loss, and PR-AUC, or a combination thereof, can be used as auxiliary evaluation metrics. For example, a first average of the F1 score and AUC-ROC for each trained supervised model can be calculated. If the difference between the first averages of these multiple trained supervised models does not fall within a preset range, then the trained supervised model with the highest first average is considered the model with the best predictive ability, i.e., the target supervised model. If the difference between the first averages of these multiple trained supervised models falls within a preset range, then multiple evaluation metrics are selected from the auxiliary evaluation metrics, and a second average is calculated. The trained supervised model with the highest second average is considered the model with the best predictive ability, i.e., the target supervised model.

[0090] It is understood that after obtaining the target supervised model, the embodiments of this application will periodically update it. That is, after step S204, the intent prediction method provided by the embodiments of this application may further include: step S206, periodically collecting newly added reference user data; step S208, updating multiple trained supervised models based on the newly added reference user data; step S210, evaluating the prediction ability of each updated trained supervised model, and selecting the updated trained supervised model with the best prediction ability as the target supervised model.

[0091] It is understandable that after reaching the T4 stage, the number of reference user data with purchase results will gradually increase with the listing time of the target product. Therefore, after obtaining the target supervised model for the first time, this application embodiment will also periodically collect newly added reference user data and update multiple trained supervised models based on the newly added reference user data.

[0092] In one embodiment, newly added reference user data can be combined with existing reference user data to obtain a sample set. This sample set is then used to retrain each previously trained model to obtain an updated trained supervised model. In another embodiment, newly added reference data can be used as a separate sample set, and this sample set is used to retrain each previously trained model to obtain an updated trained supervised model.

[0093] Furthermore, based on sample updates within the sample set, the training parameters can be adjusted according to the training method when each trained supervised model is retrained. For example, for a trained supervised model trained using XG Boost, the learning rate can be reduced to 0.05, and the maximum depth increased to 8. For a trained supervised model trained using Light GBM, the number of feature bins can be increased to 150. For a trained supervised model trained using Transformer, the number of layers can be increased to 8, and the number of attention heads increased to 10, thereby enhancing the model's expressive power and enabling it to adapt to more complex data distributions.

[0094] Subsequently, the predictive power of each updated trained supervised model is evaluated, and the updated trained supervised model with the best predictive power is selected as the target supervised model. It is understood that the method for evaluating the predictive power of each updated trained supervised model is similar to the method for evaluating the predictive power of multiple trained supervised models described above. Therefore, for the specific method of evaluating the predictive power of each updated trained supervised model, please refer to the above embodiments.

[0095] Understandably, different prediction strategies can be offered for different stages. The following section will introduce appropriate prediction strategies for different stages to predict the probability of target users purchasing target products.

[0096] In stages T1 and T2, only one forecasting strategy, namely the business forecasting rule, is available. Therefore, in these two stages, this application embodiment provides an implementation method for step S216 described above. Please refer to... Figure 3 This includes step S216A, which calculates the target probability based on the comprehensive characteristics of the target user and the business prediction rules.

[0097] In this embodiment of the application, the business prediction rule includes prediction weights corresponding to user profile features, emotional features, and behavioral features. Then, the prediction weights of the user profile features, emotional features, and behavioral features in the comprehensive features corresponding to the target user can be weighted and summed to obtain the target probability, that is, the probability that the target user will purchase the target product.

[0098] In stage T3, two prediction strategies are available: business prediction rules and target semi-supervised models. Therefore, in this application embodiment, two implementation methods are provided for step S216 above. The first implementation method is as follows: Figure 4 As shown, this includes step S216B, which involves inputting the comprehensive features corresponding to the target user into the target semi-supervised model and using the semi-supervised model to predict the target probability.

[0099] Understandably, compared to stages T2 and T1, stage T3 introduces a new prediction strategy: a target semi-supervised model. Therefore, in stage T3, only the target semi-supervised model can be used to determine the target probability. This involves inputting the comprehensive features corresponding to the target user into the target semi-supervised model, using the model for prediction, and then using the output of the target semi-supervised model as the target probability, i.e., the probability that the target user will purchase the target product.

[0100] The second implementation method is as follows: Figure 5 As shown, the process includes: Step S216C-1, calculating the first probability based on the comprehensive characteristics of the target user and the business prediction rules. Step S216C-2, inputting the comprehensive characteristics of the target user into the target semi-supervised model, and using the target semi-supervised model to predict the second probability. Step S216C-3, selecting the target probability from the first and second probabilities according to a preset selection rule.

[0101] Understandably, compared to stages T2 and T1, stage T3 introduces a new prediction strategy—the target semi-supervised model. Therefore, in stage T3, both prediction strategies can be used to jointly determine the target probability. Specifically, a weighted sum is performed based on the comprehensive characteristics of the target user and the business prediction rules, and the resulting probability is used as the first probability. Then, the comprehensive characteristics of the target user are input into the target semi-supervised model, and prediction is performed using the target semi-supervised model. The output of the target semi-supervised model is used as the second probability.

[0102] Next, according to a preset selection rule, a target probability is selected from a first probability and a second probability. In one embodiment, the preset selection rule may be to select the smaller of the first probability and the second probability as the target probability. This reminds business personnel to have a deeper understanding of the target user's purchasing needs in order to provide more comprehensive and complete services. In another embodiment, the preset selection rule may be to select the larger of the first probability and the second probability as the target probability. This reminds business personnel to continuously follow up with the target user.

[0103] In stage T4, three prediction strategies are available: business prediction rules, target semi-supervised models, and target supervised models. Therefore, in this embodiment, the present application provides an implementation method for step S216 described above. Please refer to... Figure 6 This includes step S216D, which involves inputting the comprehensive features corresponding to the target user into the target supervision model and using the target supervision model to predict the target probability.

[0104] Understandably, although three prediction strategies can be provided in stage T4, the target-supervised model prediction strategy uses the most reference user data with purchase results compared to the other two strategies, and its training methods are more diverse. Therefore, this embodiment of the application considers the target-supervised model prediction strategy to have the highest reliability. Thus, in stage T4, this embodiment of the application inputs the comprehensive features corresponding to the target user into the target-supervised model, uses the target-supervised model for prediction, and uses the output of the target-supervised model as the target probability, i.e., the probability that the target user will purchase the target product.

[0105] Furthermore, to further improve the accuracy of predictions using the target-supervised model, business forecasting rules can be used to calculate probabilities each time the target-supervised model is used for prediction, and these probabilities can be used to evaluate and optimize the target-supervised model.

[0106] To perform the corresponding steps in the above embodiments and various possible methods, an implementation of an intent prediction device is given below. Please refer to... Figure 7 This is a functional block diagram of the intent prediction device 300 provided in this embodiment of the invention. It should be noted that the basic principle and technical effects of the intent prediction device 300 provided in this embodiment are the same as those in the above embodiments. For the sake of brevity, any parts not mentioned in this embodiment can be referred to the corresponding content in the above embodiments. The intent prediction device 300 includes:

[0107] The acquisition module 310 is used to acquire target user data of the target user. The target user refers to a user who has not purchased any product on the target program. The target user data includes data associated with the target product.

[0108] The determination module 320 is used to determine the comprehensive characteristics of the target user based on the target user data. The comprehensive characteristics include user profile characteristics, emotional characteristics, and behavioral characteristics.

[0109] The prediction module 330 is used to obtain the target probability based on the comprehensive characteristics of the target user and the target prediction strategy. The target probability represents the probability that the target user will purchase the target product. The target prediction strategy is related to the amount of historical user data corresponding to the target product. The amount of historical user data is positively correlated with the listing time of the target product.

[0110] Optionally, if the number of historical user data does not exceed the first preset threshold or the number of reference user data with purchase results in the historical user data does not reach the first preset condition, the target prediction strategy is a business prediction rule, which is set according to the feature types included in the comprehensive features; the prediction module 330 is also used to calculate the target probability according to the comprehensive features corresponding to the target user and the business prediction rule.

[0111] Optionally, when the amount of historical user data exceeds a first preset threshold but does not exceed a second preset threshold, and the amount of reference user data with purchase results in the historical user data reaches a first preset condition, the target prediction strategy is at least one of the preset prediction strategies. The preset prediction strategies include business prediction rules and a target semi-supervised model. The first preset threshold is less than the second preset threshold. The intent prediction device 300 also includes a training module 340, which is used to train the initial semi-supervised model based on the reference user data to obtain the target semi-supervised model.

[0112] Optionally, the prediction module 330 is also used to input the comprehensive features corresponding to the target user into the target semi-supervised model and use the target semi-supervised model to predict the target probability.

[0113] Optionally, the prediction module 330 is further configured to calculate a first probability based on the comprehensive characteristics of the target user and the business prediction rules; input the comprehensive characteristics of the target user into the target semi-supervised model and use the target semi-supervised model to predict a second probability; and select a target probability from the first probability and the second probability according to a preset selection rule.

[0114] Optionally, when the amount of historical user data exceeds a second preset threshold and the amount of reference user data with purchase results in the historical user data reaches a second preset condition, the target prediction strategy is a target supervised model; the training module 340 is also used to train the initial supervised model according to the reference user data in different training methods to obtain multiple trained supervised models; evaluate the prediction ability of multiple trained supervised models, and select the model with the best prediction ability as the target supervised model.

[0115] Optionally, the training module 340 is also used to periodically collect newly added reference user data; update multiple trained supervised models based on the newly added reference user data; evaluate the predictive ability of each updated trained supervised model, and select the updated trained supervised model with the best predictive ability as the target supervised model.

[0116] It should be noted that, in the embodiments of this application... Figure 7 The module division shown in the intent prediction device is illustrative and represents only one logical functional division; in actual implementation, other division methods may be used. Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, exist as separate physical units, or have two or more units integrated into one unit. The integrated units can be implemented in hardware, as software functional units, or a combination of both.

[0117] It should be noted that, in the embodiments of this application, if the above-described methods are implemented as software functional modules and sold or used as independent products, they can also be stored in a computer-readable storage medium. Based on this understanding, the technical solutions of the embodiments of this application, or the parts that contribute to related technologies, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause an electronic device to execute all or part of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), magnetic disks, or optical disks. Thus, the embodiments of this application are not limited to any specific hardware and software combination.

[0118] This application also provides an electronic device; please refer to [link / reference]. Figure 8 The electronic device 100 includes a processor 120, a memory, and a network interface 150 connected via a system bus 110. The processor 120 provides computing and control capabilities. The memory includes a non-volatile storage medium 130 and internal memory 140. The non-volatile storage medium 130 stores an operating system, computer programs, and a database. The internal memory 140 provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium 130. The database of the electronic device 100 stores data. The network interface 150 of the electronic device 100 communicates with external terminals via a network connection. When the computer program is executed by the processor 120, it implements the intent prediction method provided in the above embodiments.

[0119] It is understandable that, Figure 8 The structures shown are merely block diagrams of some structures related to the present application and do not constitute a limitation on the electronic devices to which the present application is applied. Specific electronic devices may include more or fewer components than shown in the figures, or combine certain components, or have different component arrangements. In one embodiment, the intent prediction device provided by this application can be implemented as a computer program, which can be implemented in... Figure 8 The device operates on the electronic device shown. The memory of the electronic device can store the various program modules that make up the above-described apparatus. The computer program composed of the various program modules causes the processor to execute the steps of the methods in the various embodiments of this application described in this specification.

[0120] This application also provides a storage medium storing a computer program thereon, which, when executed by a processor, implements the intent prediction method provided in the above embodiments.

[0121] It should be noted that the descriptions of the storage medium and device embodiments above are similar to the descriptions of the method embodiments above, and have similar beneficial effects. For technical details not disclosed in the storage medium, storage medium, and device embodiments of this application, please refer to the descriptions of the method embodiments of this application for understanding.

[0122] It should be understood that the phrases "one embodiment," "an embodiment," or "some embodiments" throughout the specification mean that a specific feature, structure, or characteristic related to an embodiment is included in at least one embodiment of this application. Therefore, "in one embodiment," "in one embodiment," or "in some embodiments" appearing throughout the specification do not necessarily refer to the same embodiment. Furthermore, these specific features, structures, or characteristics can be combined in any suitable manner in one or more embodiments. It should be understood that in the various embodiments of this application, the sequence numbers of the above-described processes do not imply a sequential order of execution; the execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application. The sequence numbers of the above-described embodiments are merely descriptive and do not represent the superiority or inferiority of the embodiments. The descriptions of the various embodiments above tend to emphasize the differences between the various embodiments; their similarities or commonalities can be referred to mutually, and for the sake of brevity, they will not be repeated here.

[0123] In this article, the term "and / or" is merely a description of the relationship between related objects, indicating that there can be three kinds of relationships. For example, object A and / or object B can represent three situations: object A exists alone, object A and object B exist simultaneously, and object B exists alone.

[0124] 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 apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. 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 apparatus that includes that element.

[0125] In the several embodiments provided in this application, it should be understood that the disclosed devices and methods can be implemented in other ways. The embodiments described above are merely illustrative. For example, the division of modules is only a logical functional division, and in actual implementation, there may be other division methods, such as: multiple modules or components can be combined, or integrated into another system, or some features can be ignored or not executed. In addition, the coupling, direct coupling, or communication connection between the various components shown or discussed can be through some interfaces, and the indirect coupling or communication connection between devices or modules can be electrical, mechanical, or other forms.

[0126] The modules described above as separate components may or may not be physically separate. The components shown as modules may or may not be physical modules. They may be located in one place or distributed across multiple network units. Some or all of the modules may be selected to achieve the purpose of this embodiment according to actual needs.

[0127] In addition, each functional module in the various embodiments of this application can be integrated into one processing unit, or each module can be a separate unit, or two or more modules can be integrated into one unit; the integrated modules can be implemented in hardware or in the form of hardware plus software functional units.

[0128] Those skilled in the art will understand that all or part of the steps of the above method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When the program is executed, it performs the steps of the above method embodiments. The aforementioned storage medium includes various media that can store program code, such as mobile storage devices, read-only memory (ROM), magnetic disks, or optical disks.

[0129] Alternatively, if the integrated units described above are implemented as software functional modules and sold or used as independent products, they can also be stored in a computer-readable storage medium. Based on this understanding, the technical solutions of the embodiments of this application, or the parts that contribute to related technologies, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause an electronic device to execute all or part of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as mobile storage devices, ROMs, magnetic disks, or optical disks.

[0130] The methods disclosed in the several method embodiments provided in this application can be arbitrarily combined without conflict to obtain new method embodiments.

[0131] The above description is merely an embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. An intention prediction method, characterized by, The method comprises: obtaining target user data of a target user, the target user being a user who has not made any purchase of any product on a target program, the target user data comprising data associated with a target product; determining comprehensive features corresponding to the target user according to the target user data, the comprehensive features comprising user portrait features, emotional features and behavioral features; obtaining a target probability according to the comprehensive features corresponding to the target user and a target prediction strategy, the target probability representing a probability of the target user purchasing the target product, the target prediction strategy being associated with a quantity of historical user data corresponding to the target product, the quantity of historical user data being in a positive correlation with a listing duration of the target product; in a case where the quantity of historical user data does not exceed a first preset threshold or a quantity of reference user data with purchase results in the historical user data does not reach a first preset condition, the target prediction strategy is a business prediction rule, the business prediction rule being set according to a feature type included in the comprehensive features; in a case where the quantity of historical user data exceeds a first preset threshold and does not exceed a second preset threshold and the quantity of reference user data with purchase results in the historical user data reaches a first preset condition, the target prediction strategy is at least one of preset prediction strategies, the preset prediction strategies comprising the business prediction rule and a target semi-supervised model, the first preset threshold being less than the second preset threshold.

2. The intention prediction method of claim 1, wherein, The obtaining of the target probability according to the comprehensive features corresponding to the target user and the target prediction strategy comprises: calculating the target probability according to the comprehensive features corresponding to the target user and the business prediction rule.

3. The intention prediction method of claim 1, wherein, Before the obtaining of the target user data of the target user, the method further comprises: training an initial semi-supervised model according to the reference user data to obtain the target semi-supervised model.

4. The intent prediction method of claim 3, wherein, The obtaining of the target probability according to the comprehensive features corresponding to the target user and the target prediction strategy comprises: inputting the comprehensive features corresponding to the target user into the target semi-supervised model and predicting the target probability by using the target semi-supervised model.

5. The intent prediction method of claim 3, wherein, The obtaining of the target probability according to the comprehensive features corresponding to the target user and the target prediction strategy comprises: calculating a first probability according to the comprehensive features corresponding to the target user and the business prediction rule; inputting the comprehensive features corresponding to the target user into the target semi-supervised model and predicting a second probability by using the target semi-supervised model; selecting the target probability from the first probability and the second probability according to a preset selection rule.

6. The intent prediction method of claim 1, wherein, in a case where the quantity of historical user data exceeds a second preset threshold and the quantity of reference user data with purchase results in the historical user data reaches a second preset condition, the target prediction strategy is a target supervised model; Before the obtaining of the target user data of the target user, the method further comprises: training an initial supervised model according to the reference user data in different training modes to obtain a plurality of trained supervised models; The prediction abilities of the plurality of trained supervised models are evaluated, and a model with the best prediction ability is selected as the target supervised model.

7. The intent prediction method of claim 6, wherein, The method further comprises: periodically collecting newly added reference user data; updating the plurality of trained supervised models according to the newly added reference user data; evaluating the prediction abilities of each updated trained supervised model, and selecting an updated trained supervised model with the best prediction ability as the target supervised model.

8. An intention prediction apparatus characterized by comprising: The device comprises: an acquisition module configured to acquire target user data of a target user, the target user being a user who has not made any purchase of any product on a target program, and the target user data including data associated with a target product; a determination module configured to determine, according to the target user data, a comprehensive feature corresponding to the target user, the comprehensive feature including a user portrait feature, an emotional feature, and a behavioral feature; a prediction module configured to obtain a target probability according to the comprehensive feature corresponding to the target user and a target prediction strategy, the target probability representing a probability of the target user purchasing the target product, the target prediction strategy being associated with a number of historical user data corresponding to the target product, the number of historical user data being in a positive correlation with an online time length of the target product; in a case where the number of historical user data does not exceed a first preset threshold or a number of reference user data with purchase results in the historical user data does not reach a first preset condition, the target prediction strategy is a business prediction rule, the business prediction rule being set according to a feature type included in the comprehensive feature; in a case where the number of historical user data exceeds the first preset threshold and does not exceed a second preset threshold and the number of reference user data with purchase results in the historical user data reaches the first preset condition, the target prediction strategy is at least one of preset prediction strategies, the preset prediction strategies including the business prediction rule and a target semi-supervised model, the first preset threshold being smaller than the second preset threshold.

9. An electronic device, comprising: A processor and a memory are included, the memory stores a computer program, and the processor, when executing the computer program, implements the intent prediction method of any one of claims 1-7.

10. A storage medium, characterized by A computer program is stored thereon, and the computer program, when executed by a processor, implements the intent prediction method of any one of claims 1-7.

Images