Prediction methods, models, systems, and related equipment for target events

Abstract

This application relates to a method, model, system, and related equipment for predicting target events. The method for predicting target events includes: acquiring several interaction events of a target user, wherein the interaction events constitute a behavioral sequence of the target user; encoding each interaction event to form an event feature vector; performing feature extraction and serialization processing on each event feature vector to form a sequence feature vector of the behavioral sequence, which serves as the latent state of the target user; and fusing the latent state and static attributes of the target user to obtain a prediction result of the occurrence of the target event. This method can cover the interaction chain of the target user, improving the reliability of target event prediction.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of event prediction, in particular to a target event prediction method, an event prediction model, an event prediction system, a computer device and a computer readable storage medium. BACKGROUND

[0002] With the rapid development of insurance, e-commerce and other industries, it is necessary for each enterprise to improve user experience and reduce its own risk, so it is also necessary to predict events such as transactions, complaints and fraud in advance.

[0003] Taking a complaint event as an example, at present, it is generally passive waiting for customers to complain, and after receiving customer complaints, manual customer service is involved, and problems that cannot be solved by customer service are further handed over to officers for processing, causing long case processing flow invalidation, and even causing further complaints. For a company, the pain points of large number of complaints, high complaint cost, high processing difficulty and long processing time have existed for a long time. SUMMARY

[0004] Therefore, it is necessary to provide a target event prediction method, an event prediction model, an event prediction system, a computer device and a computer readable storage medium capable of covering the interaction link of a target user and improving the reliability of target event prediction.

[0005] In one aspect, a target event prediction method is provided, which includes: obtaining a plurality of interaction events of a target user, wherein the interaction events are used to constitute a behavior sequence of the target user; respectively encoding the interaction events to form event feature vectors; performing feature extraction and serialization processing on each event feature vector to form a sequence feature vector of the behavior sequence, and the sequence feature vector is used as a hidden state of the target user; and fusing the hidden state of the target user and a static attribute to obtain a prediction result of the target event.

[0006] In an embodiment of the present application, encoding the interaction events to form the event feature vectors includes: encoding position information of the interaction events to form a first sub-vector, and encoding attribute information of the interaction events to form a second sub-vector; and splicing the first sub-vector and the second sub-vector to form the event feature vector belonging to the current interaction event.

[0007] In an embodiment of the present application, the position information of the interaction events is the position of the interaction events in the behavior sequence determined based on the occurrence time of the interaction events, and the attribute information includes at least one of the type, the duration and the result of the interaction events.

[0008] In an embodiment of the present application, the feature extraction and serialization processing are performed on each event feature vector to form a sequence feature vector of the behavior sequence, including: sequentially performing feature extraction on the information of the previous interaction event combined with each event feature vector based on the occurrence order of the interaction events to obtain a behavior feature vector; and performing recursive calling on the sequence feature vector formed previously by using the behavior feature vector to form a new sequence feature vector as the hidden state.

[0009] In an embodiment of the present application, the fusion of the hidden state and the static attribute of the target user is used to obtain the prediction result of the occurrence of the target event, including: performing feature transformation on the static attribute of the target user to form a static attribute feature vector of the target user; and performing fusion processing on the sequence feature vector and the static attribute feature vector of the target user to obtain the prediction result of the occurrence of the target event.

[0010] In an embodiment of the present application, the obtaining of the interaction events of the target user includes: periodically obtaining the interaction events occurred within a preset time length at a preset interval; and / or, in response to the occurrence of a new interaction event, obtaining the interaction events of the target user.

[0011] On the other hand, an event prediction model is provided, including: an encoding module, a sequence model and a static model; the encoding module is used to encode the obtained interaction events of the target user respectively to form event feature vectors; the interaction events are used to constitute a behavior sequence of the target user; the sequence model is used to perform feature extraction and serialization processing on each event feature vector to form a sequence feature vector of the behavior sequence, and the sequence feature vector is used as a hidden state of the target user; and the static model is used to obtain a static attribute of the target user; and the static attribute feature vector is used to fuse with the hidden state to obtain a prediction result of the occurrence of a target event.

[0012] On the other hand, an event prediction system is provided, including: an event prediction model and a fusion model; the event prediction model is used to obtain a plurality of interaction events of a target user, wherein the interaction events are used to constitute a behavior sequence of the target user; the interaction events are encoded respectively to form event feature vectors; feature extraction and serialization processing are performed on each event feature vector to form a sequence feature vector of the behavior sequence, and the sequence feature vector is used as a hidden state of the target user; and the fusion model is used to fuse the hidden state and a static attribute of the target user to obtain a prediction result of the occurrence of a target event.

[0013] In still another aspect, a computer device is provided, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, the processor implementing the following steps when executing the computer program: obtaining a plurality of interaction events of a target user, wherein the interaction events are used to constitute a behavior sequence of the target user; respectively encoding the interaction events to form event feature vectors; performing feature extraction and serialization processing on the event feature vectors to form a sequence feature vector of the behavior sequence, the sequence feature vector being a hidden state of the target user; fusing the hidden state of the target user and a static attribute to obtain a prediction result of occurrence of a target event.

[0014] In still another aspect, a computer readable storage medium is provided, having a computer program stored thereon, the computer program implementing the following steps when executed by a processor: obtaining a plurality of interaction events of a target user, wherein the interaction events are used to constitute a behavior sequence of the target user; respectively encoding the interaction events to form event feature vectors; performing feature extraction and serialization processing on the event feature vectors to form a sequence feature vector of the behavior sequence, the sequence feature vector being a hidden state of the target user; fusing the hidden state of the target user and a static attribute to obtain a prediction result of occurrence of a target event.

[0015] The above target event prediction method, event prediction model, event prediction system, computer device, and computer readable storage medium regard the interaction events of the target user as a behavior sequence of the user. Based on the behavior sequence of the target user, feature extraction and serialization processing are performed on each interaction event to obtain a sequence feature vector, the sequence feature vector being a hidden state of the target user, the hidden state representing a predicted state of the target user, so that the hidden state can cover the interaction link of the target user. The prediction result of occurrence of a target event is obtained in combination with the static attribute and the hidden state of the target user, thereby improving the reliability of target event prediction. BRIEF DESCRIPTION OF DRAWINGS

[0016] Figure 1 is an application environment schematic diagram of an embodiment of the target event prediction method of the present application;

[0017] Figure 2 is a flow schematic diagram of an embodiment of the target event prediction method of the present application;

[0018] Figure 3 is a flow schematic diagram of another embodiment of the target event prediction method of the present application;

[0019] Figure 4 is Figure 3 is a flow schematic diagram of another perspective of the target event prediction method;

[0020] Figure 5 is a structural schematic diagram of an embodiment of the event prediction system of the present application;

[0021] Figure 6 is a structural schematic diagram of an embodiment of an event prediction model of the present application;

[0022] Figure 7 is Figure 6 is a structural schematic diagram of an embodiment of an event prediction system formed by the event prediction model shown in FIG. 1;

[0023] Figure 8 is a structural schematic diagram of an embodiment of a computer device of the present application. DETAILED DESCRIPTION

[0024] In order to make the purpose, technical solutions and advantages of the present application clearer, the present application will be further described in detail below in combination with the drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and do not limit the present application.

[0025] Typical machine learning techniques in the related art can be used to predict target events. The target event can be risk prediction such as complaint behavior, fraudulent behavior, etc., or prediction of events such as transaction behavior, praise behavior, etc.

[0026] The prediction model in the related art is usually to predict the risk level of a user in a key environment. The prediction model can be a traditional machine learning model represented by a generalized linear regression model, a tree model, and a kernel model, or a deep learning model represented by a deep network. At the same time, some Wide&Deep structure models that combine the advantages of traditional machine learning models and deep learning models, and multi-modal fusion models are gradually applied in actual business.

[0027] In the Wide&Deep structure model, the Deep part usually includes a multi-layer feedforward neural network, which mainly encodes sparse features such as categories, IDs, etc. that are difficult to process by traditional machine learning models. The encoding result of the sparse features is spliced with the dense numerical type features in the structured features, and finally fused through the traditional model (such as logistic regression) of the Wide part.

[0028] At the same time, in addition to the core algorithm module, a complete prediction system also needs data integration, processing modules, storage modules, etc. to meet the needs of service response speed, data backflow, model iteration, etc. in actual application.

[0029] Taking the interaction between a user and an insurance company as an example, and taking a complaint event as a target event, with the rapid development of the insurance industry, the user's requirements for insurance services are also getting higher and higher, and dissatisfaction in the service process may trigger the user's complaint behavior, among which there are also malicious complaint users. Therefore, the trend of large amount of complaints and high difficulty in complaint processing is becoming more and more obvious. In order to reduce the long time efficiency of case processing flow, the related technology will predict the possibility / probability of the occurrence of the complaint event. However, the related technology has the following two shortcomings:

[0030] Firstly, the related technology is difficult to cover the entire interaction life cycle of the user. Since the interaction process between the user and the insurance company is variable, each different type of interaction behavior and result may affect the user's complaint risk. In view of this, in the related risk assessment system technology, the prediction model usually models and predicts a single link, and multiple model developments are required to cover multiple links, which easily leads to the fact that the system can only focus on the so-called "key" interaction link and ignore some "weak" interaction events under normal circumstances. For example, in some related technologies, only the prior complaint event is learned and analyzed. Therefore, the timeliness of the system response is limited, and the timeliness of the response is often an important factor that affects the final result in similar risk events.

[0031] Secondly, the related technology is difficult to comprehensively utilize the data of the entire interaction link. The prediction model focusing on a single node is more difficult to comprehensively utilize the data of the entire interaction link. The processing of the interaction event data by the prediction model can usually only construct the so-called recency, frequency, and monetary (Recency, Frequency, Monetary) features related to the amount / quantity, and the expression learning ability of the prediction model is limited. In view of this, since the input of each link prediction model is different from the input of other link prediction models, that is, they are heterogeneous and independently trained, the coverage of the training data is different, and the data of the entire interaction link cannot be fully utilized to improve the accuracy of the prediction model.

[0032] The present application can also be applied to the financial insurance business scenario. Through machine learning, deep learning and other computer technologies, combined with user interaction behavior data and historical customer risk labels, the behavior (i.e., target event) of the user can be evaluated / early warned in each link of the entire life cycle of the user interacting with the financial institution. Further, the present application can assist the financial institution to intervene in the user's behavior as early as possible, reduce the risk level of the financial institution, and improve the user's experience and satisfaction.

[0033] The application aims to pre-position the identification ability of complaint risk and other target events, collect internal and external data labels, dynamic change factors and the like of the user in the interaction process with the user. For example, real-time complaint risk rating is performed on the claim customers, different response strategies / claim strategies are adopted for customers with different complaint risk levels in combination with the objective situation, thereby the number of complaints can be reduced, malicious complaint customers can be identified, and claim disputes can be properly solved as much as possible. In other words, based on the prediction result of the prediction of the target event, the risk management input of the financial institution can be optimized according to different risk levels, which is beneficial to improve the risk management level and user satisfaction.

[0034] The target event prediction method provided by the application can be applied to the application environment as shown in Figure 1 Figure 1 is a schematic diagram of an application environment of an embodiment of the target event prediction method of the application.

[0035] The terminal 102 communicates with the server 104 through the network.

[0036] The terminal 102 sends a prediction instruction of occurrence of a target event to the server 104. The server 104 performs prediction by executing the target event prediction method in response to receiving the prediction instruction, and feeds back the prediction result of the target event to the terminal 102.

[0037] The terminal 102 can be, but is not limited to, various personal computers, notebook computers, smart phones, tablet computers and portable wearable devices, and the server 104 can be implemented by an independent server or a server cluster composed of multiple servers.

[0038] In one embodiment, as shown in Figure 2 a target event prediction method is provided, Figure 2 is a flowchart of an embodiment of the target event prediction method of the application. Taking the server in Figure 1 as an example, the method comprises the following steps:

[0039] S201: Obtain a plurality of interaction events of a target user, wherein the interaction events are used to constitute a behavior sequence of the target user.

[0040] In this embodiment, the interaction events are interactions between the user and the merchant / platform, such as exposure, click, deposit, conversion, retention, purchase, complaint, etc., which can all be used as interaction events. The types of the plurality of interaction events can be different. For example, the plurality of interaction events obtained at the same time can include click events, deposit events, etc.

[0041] ​Optionally, when the plurality of interaction events of the target user are acquired, the interaction events are arranged according to the occurrence time, so as to facilitate the analysis of the user behavior in combination with the information of the occurrence sequence of the interaction events.

[0042] The behavior sequence can include all interaction events of the user after the interaction event occurs between the user and the platform, that is, the plurality of acquired interaction events are part of the behavior sequence. Alternatively, the plurality of acquired interaction events constitute the behavior sequence.

[0043] The target user is a specified user for which a prediction result of the occurrence of a target event is needed to be predicted. Optionally, the target user can be only one user.

[0044] The plurality of interaction events of the target user are acquired, and the interaction events are regarded as components of the behavior sequence in combination with the idea of the natural language processing technology, so as to analyze the behavior of the target user according to the interaction events, and to achieve the prediction of the occurrence possibility of the target event to obtain the prediction result.

[0045] At the same time, the plurality of interaction events of the target user can exist independently of the business link, which is beneficial to improve the real-time performance and coverage of the prediction, and to reduce the influence on the data processing and response speed of the business link.

[0046] S202: Encode the interaction events respectively to form event feature vectors.

[0047] In this embodiment, each interaction event is encoded respectively to form each event feature vector corresponding to each interaction event, so as to facilitate the analysis of the user behavior in combination with each interaction event.

[0048] S203: Perform feature extraction and serialization processing on each event feature vector to form a sequence feature vector of the behavior sequence, and the sequence feature vector is taken as the hidden state of the target user.

[0049] In this embodiment, the feature extraction and serialization processing are performed on each event feature vector to form a sequence feature vector of the behavior sequence. This means that the sequence feature vector can contain the information of each interaction event of the target user, and is related to the target event which needs to be predicted, so the sequence feature vector is taken as the hidden state of the target user.

[0050] The hidden state is obtained by using the model / system for prediction, and can be used to represent the state of the target user. The hidden state is the state that the target user can be in which is obtained by prediction, rather than the real state of the target user, so this state is named as the hidden state.

[0051] S204: Fuse the hidden state of the target user and the static attribute to obtain the prediction result of the occurrence of the target event.

[0052] In the embodiment, to improve the accuracy and reliability of the prediction result, the target user's static attributes are further combined on the basis of the target user's hidden state to widen the width of the process data of the prediction result.

[0053] In this way, the target user's hidden state and static attributes are fused, the available data range is expanded, the prediction result of the target user's occurrence of the target event is obtained, the obtained prediction result can cover the target user's interaction link, and more user's associated information is combined, which is beneficial to significantly improve the reliability of the prediction result.

[0054] Optionally, the prediction result can be in the form of probability to represent the possibility of the occurrence of the target event, or can be in the form of output of the possible occurrence of the target event or the possible non-occurrence of the target event to represent the possibility of the occurrence of the target event, or can preset multiple target event occurrence possibility levels, and the prediction result is one of the levels. The prediction result can also represent the possibility of the occurrence of the target event in other ways, which will not be described here.

[0055] The static attributes are attributes associated with the target user himself / herself, and include at least one of portrait information, associated relationship information, and transaction / invoice information of the target user. For example, user portrait information such as gender and age, associated relationship information such as associated business personnel and relatives, and transaction / invoice information such as policies and orders between the company / platform, which will not be exemplified and described here.

[0056] Optionally, the number of target events can be one or more, which is not limited here. That is, in the prediction method of the target event, the hidden state and the static attributes of the target user can be used to predict the occurrence possibility of different target events.

[0057] In the above prediction method of the target event, the interaction event of the target user is regarded as a behavior sequence of the user. Based on the behavior sequence of the target user, the feature extraction and the sequence processing are performed on each interaction event to obtain a sequence feature vector, the sequence feature vector is taken as the hidden state of the target user, the hidden state represents the predicted state of the target user, that is, the hidden state can cover the interaction link of the target user. The prediction result of the occurrence of the target event is obtained by combining the static attributes and the hidden state of the target user, the available data for obtaining the prediction result is widened, and on the basis that the hidden state can cover the interaction link of the target user, the risk of insufficient or unbalanced data in a single link is effectively reduced, which is beneficial to improve the reliability of the prediction of the target event. At the same time, the embodiment can exist independently of the business link, which is beneficial to reduce the influence on the business link while improving the real-time performance and coverage of the prediction.

[0058] It can be seen that the present application combines the idea of natural language processing technology. Natural language processing (NLP) is a science integrating linguistics, computer science and mathematics, and the research of natural language processing involves natural language, i.e. the language used in daily life. Natural language processing is mainly applied to machine translation, public opinion monitoring, automatic abstract, opinion extraction, text classification, question answering, text semantic comparison, speech recognition, Chinese OCR, etc.

[0059] In a natural language processing model based on deep learning technology, the general processing steps include the following two steps:

[0060] First, the basic data unit is a natural language token, for example, a Chinese character for Chinese and a word for English. The basic data unit needs to be encoded to form a vector corresponding to the token.

[0061] Second, the vectorized data is used to train the target task by a deep sequence model to obtain a model that can extract information of the whole orange or even a chapter. The deep sequence model can be RNN (Recurrent Neural Network), LSTM (Long Short-Term Memory), attention model, etc.

[0062] Specifically, in the present application, for each user, an interaction event (login, browsing, online consultation, incoming call, purchase, claim settlement, etc.) with the insurance company is regarded as a "Chinese character", and the whole interaction sequence of the user, i.e. the behavior sequence, is regarded as a "sentence". By properly encoding the interaction events of the target user, the behavior sequence of the target user is modeled by using a sequence model such as a deep sequence model to obtain a feature vector related to the target event of the whole interaction sequence of the user, so that the subsequent static model can be fused to improve the accuracy of the prediction of the event prediction model, and thus the reliability of the event prediction system is improved.

[0063] In one embodiment, as shown in Figure 3 and Figure 4 , another method for predicting a target event is provided, Figure 3 is a flowchart of another embodiment of the method for predicting a target event of the present application, Figure 4 is a flowchart of another perspective of the method for predicting a target event shown in Figure 3 , which includes the following steps:

[0064] S301: Obtain a plurality of interaction events of a target user.

[0065] In the embodiment, a plurality of interaction events of a target user are obtained. The interaction events are used to constitute a behavior sequence of the target user. The target user is one of the users.

[0066] Optionally, the hidden state of each user can be obtained in a pre-computed manner. When the probability of the target event occurring to the target user needs to be obtained, the hidden state of the target user can be directly called, i.e., only incremental calling can be performed online, so as to improve the real-time requirement of an interactive application (such as telephone call, online consultation, etc.). That is, the embodiment can improve the timeliness without significantly increasing the engineering cost.

[0067] Further, the plurality of interaction events of the target user can be obtained periodically and / or triggered. That is, the plurality of interaction events of the target user can be obtained periodically, or the interaction events of the target user can be obtained triggered, or the interaction events of the target user can be obtained periodically and triggered in combination, which will not be described herein.

[0068] Specifically, the periodicity means that the interaction events occurring within a preset time length are obtained periodically at intervals of the preset time length. The preset time length can be in units of days, hours, or minutes. The specific preset time length can be determined according to actual computing power and requirements, which is not limited herein.

[0069] The triggering means that the interaction event of the user is obtained in response to the occurrence of a new interaction event. In other words, when a new event occurs to the user, the event prediction model is triggered to update the hidden state of the user, i.e., to update the sequence feature vector.

[0070] S302: Encode the interaction events respectively to form event feature vectors.

[0071] In the embodiment, each interaction event is encoded respectively to form an event feature vector corresponding to each interaction event. Optionally, each interaction event can be encoded in sequence according to the occurrence order of the interaction events, so as to further improve the reliability of the prediction result.

[0072] For example, in a user behavior sequence, the information contained in the user first browsing a page and then placing an order can be different from the information contained in the user first placing an order and then browsing a page.

[0073] Specifically, encoding an interaction event to form an event feature vector can be:

[0074] The interaction event is split into two parts, one of which is the position information of the interaction event, and the other of which is the attribute information of the interaction event.

[0075] The position information of the interaction event is encoded to form a first sub-vector, and the attribute information of the interaction event is encoded to form a second sub-vector. The first sub-vector and the second sub-vector are spliced to form an event feature vector belonging to the current interaction event.

[0076] The position information of the interaction event is determined based on the occurrence time of the interaction event, such as a timestamp.

[0077] The attribute information includes at least one of the type, duration, and result of the interaction event.

[0078] The position information and the attribute information of the interaction event are encoded separately, which can reduce the interference between the position information and the attribute information. As can be easily understood, since the occurrence times of the events are different, the position information of the same type of interaction event occurring at different times must be different, but the attribute information is similar or even the same. Encoding the two separately can preserve the difference between the position information and the correlation and similarity between the attribute information.

[0079] The event feature vector obtained by splicing the first sub-vector and the second sub-vector extracts the information of the interaction event, further makes the hidden state of the target user close to the real state of the target user, and thus is beneficial to improve the accuracy and reliability of the prediction result of the target event.

[0080] S303: Feature extraction is performed on the event feature vector to obtain a behavior feature vector.

[0081] In this embodiment, based on the occurrence order of each interaction event, the information of each event feature vector combined with the previous interaction event is sequentially extracted to obtain a feature vector. In other words, the event feature vectors of each interaction event can be sequentially extracted based on the position information to extract the information contained in the event feature vectors, and the behavior feature vectors belonging to each interaction event are obtained to facilitate the analysis of the user's behavior habits based on the behavior feature vectors.

[0082] That is, the first sub-vector contains the position information of the interaction event, and the second sub-vector contains the attribute information of the interaction event. The event feature vector formed by splicing the first sub-vector and the second sub-vector contains the position information and the attribute information of the interaction event at the same time. The event feature vector has a position feature and an attribute feature respectively, and thus the behavior feature vector obtained by feature extraction using the event feature vector is beneficial to improve the accuracy of the prediction result.

[0083] Optionally, when the event feature vector is extracted, the information of other event feature vectors can be combined to enrich the expression of the behavior feature vector.

[0084] S304: The behavior feature vector is serialized to form a sequence feature vector as the hidden state of the target user.

[0085] In this embodiment, the behavior feature vector is serialized to form a sequence feature vector as the hidden state of the target user by recursively calling the sequence feature vector formed in advance by the behavior feature vector. That is, the sequence feature vector can be updated by the newly obtained behavior feature vector, i.e., the hidden state of the target user is updated.

[0086] Of course, if the interaction event of the target user is obtained for the first time in step S301, the behavior feature vector formed at present can be used as the sequence feature vector and as the hidden state of the target user. When the behavior feature vector is obtained next time, the sequence feature vector formed by the previous interaction event is recursively called to form a new sequence feature vector as the hidden state of the target user.

[0087] In this way, after feature extraction and serialization of each event feature vector, the sequence feature vector of the behavior sequence is formed, and the sequence feature vector is used as the hidden state of the target user to represent the state of the target user. The sequence feature vector contains information of the behavior sequence and information associated with the target event.

[0088] S305: Obtain the static attribute feature vector of the target user.

[0089] In this embodiment, the static attribute of the target user is transformed to form a static attribute feature vector of the target user to improve the richness of the static attribute of the target user.

[0090] The static attribute is structured information that does not belong to the user interaction event. The static attribute can include at least one of portrait information, association relationship information, and transaction / invoice information of the target user.

[0091] Optionally, the transformation manner of the feature transformation of the static attribute can be a machine learning feature processing method, such as quantization, standardization, binning, missing value processing, aggregation, etc., which is not limited herein.

[0092] S306: Fuse the sequence feature vector and the static attribute feature vector to obtain a prediction result of the occurrence of the target event.

[0093] In this embodiment, the fusion of the hidden state of the target user and the static attribute can be a fusion process of the sequence feature vector and the static attribute feature vector of the target user to obtain a prediction result of the occurrence of the target event.

[0094] As such, the prediction result combines information of both the hidden state of the target user and the static attribute. The hidden state is a behavior sequence formed by modeling the entire interaction event of the target user, and the obtained user behavior sequence is a feature vector related to the target event, which is fused with the static attribute. That is, the sequence feature vector can be used as a depth part, while the static attribute feature vector is used as a width part, which can effectively improve the accuracy of the prediction result, thereby improving the reliability of the event prediction model and the event prediction system.

[0095] In summary, the prediction method of the target event in the embodiment can pre-compute a part of the prediction calculation in advance offline. For example, the prediction user hidden state part is pre-computed, and the hidden state can be stored in a database or cache for retrieval by the user's ID. In this way, only incremental calls can be performed online, reducing the impact of high computational cost on timeliness. When the business system needs to call the event prediction system for scoring, the latest hidden state of the target user can be obtained through the user ID, and then spliced with the latest static attribute to directly call the fusion model to obtain the score, effectively reducing the real-time prediction time. Moreover, the embodiment can fully utilize the learning ability and expression ability of the event prediction model / event prediction system.

[0096] It should be understood that, although Figures 2-4 The steps in the flowchart of the embodiment are displayed in sequence according to the arrows, but these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise specified herein, the execution of these steps is not strictly limited in sequence, and these steps can be executed in other orders. Moreover, Figures 2-4 At least a part of the steps in the embodiment can include multiple sub-steps or multiple stages, which are not necessarily executed at the same time, but can be executed at different times, and the execution order of these sub-steps or stages is not necessarily sequential, but can be executed in rotation or alternation with at least a part of other steps or sub-steps or stages of other steps.

[0097] For example, steps S301-S304 are the process of obtaining the hidden state of the target user, and step S305 is the process of obtaining the static attribute of the user. This means that steps S301-S304 and step S305 can be performed simultaneously or sequentially. Alternatively, steps S301-S304 are pre-computation steps, which calculate the hidden state of each user, in response to receiving an instruction to obtain a prediction result, step S305 obtains the static attribute of the target user, and the hidden state of the target user is called, and then the hidden state and the static attribute of the target user are fused to output the prediction result of the target event of the target user.

[0098] In one embodiment, as shown in Figure 5 An event prediction system is provided, Figure 5 is a structural schematic diagram of an embodiment of the event prediction system of the present application.

[0099] The event prediction system comprises an event prediction model 51 and a fusion model 52.

[0100] The event prediction model 51 is configured to obtain a plurality of interaction events of a target user, wherein the interaction events are used to constitute a behavior sequence of the target user; encode the interaction events respectively to form event feature vectors; perform feature extraction and sequence processing on the event feature vectors to form a sequence feature vector of the behavior sequence, and the sequence feature vector is taken as a hidden state of the target user.

[0101] The fusion model 52 is connected with the event prediction model 51 and is configured to fuse the hidden state and the static attribute of the target user to obtain a prediction result of a target event.

[0102] Optionally, the fusion model 52 fuses the static attribute and the hidden state, that is, the fusion model 52 can be a machine learning model capable of fusing structured data features and deep sequence features. For example, the fusion model 52 can be a generalized linear model (such as LR), a tree model (such as a decision tree, a random forest, a GBDT, etc.), an SVM / FM, a neural network model, etc. The specific selection of the fusion model 52 can be selected according to actual data and environment, which is not limited herein. When the neural network model is selected as the fusion model 52, a shallow neural network model can be selected, and the output of the shallow neural network model is taken as the final result of the prediction, that is, the prediction result of the target event.

[0103] Further, the number of the fusion model 52 can be multiple, which are respectively used to output prediction results of different target events, so as to enrich the functionality of the event prediction system.

[0104] It can be seen that in the present embodiment, the event prediction system can be called based on the latest interaction data by using a unified event prediction system, and even can be called at any time. The event prediction system in the present embodiment covers the complete interaction link of the user, which is beneficial to timely response to the call. At the same time, the event prediction model can be uniformly trained based on the data of all users, which is beneficial to effectively reduce the risk of insufficient or unbalanced data in a single link. Moreover, the unified event prediction model / system can make full use of as much user behavior data as possible to train, improve the accuracy of the event prediction model / system, and thus effectively improve the reliability of the event prediction model / system.

[0105] In one embodiment, as shown in Figure 6 An event prediction model is provided, Figure 6FIG. 1 is a structural schematic diagram of an embodiment of an event prediction model of the present application.

[0106] The event prediction model comprises an encoding module 61, a sequence model 62 and a static model 63.

[0107] The encoding module 61 is configured to encode each of a plurality of target user interaction events to form an event feature vector. The interaction events are used to form a behavior sequence of the target user.

[0108] The sequence model 62 is configured to perform feature extraction and sequence processing on each event feature vector to form a sequence feature vector of the behavior sequence, and the sequence feature vector is used as a hidden state of the target user.

[0109] The static model 63 is configured to obtain a static attribute of the target user. The static attribute feature vector is used to fuse with the hidden state to obtain a prediction result of the target event.

[0110] Optionally, the sequence model 62 can be a deep sequence model 62, etc. For example, LSTM, GRU, attention network, etc. The specific selection can be made according to actual data and environmental requirements, which is not limited herein.

[0111] Further, the encoding module 61 comprises a first sub-encoding module and a second sub-encoding module, which are respectively configured to encode position information and attribute information of the target event.

[0112] The first sub-encoding module can convert the position information into a vector of a preset length, so as to facilitate simplifying the complexity of data processing. Optionally, the first sub-encoding module can be a sine transformation module, a fully connected trainable neural network module, etc.

[0113] The second sub-encoding module can be a fully connected neural network, etc. The weight of the fully connected neural network can be trained in the model training process.

[0114] The event prediction model in the embodiment is similar to the Wide&Deep model, but is different from the Wide&Deep model. In the embodiment, the data of the user behavior is encoded by using the sequence model 62, i.e. the event feature vector is used to perform feature extraction and recursive call by using the sequence model 62. Optionally, the sequence model 62 can be a deep sequence model 62, etc.

[0115] The event prediction model can be trained based on user historical portrait data, behavior data, and corresponding target event data. The specific training method and hyperparameter configuration can be adjusted according to actual application. For example, in the training process of the event prediction model, the depth part and the width part can be trained separately and fused in the width part; that is, the parts outputting hidden states and static attributes are trained separately and fused in the static attribute part. The depth part and the width part can also be trained end-to-end, for example, when the width part also adopts a neural network model.

[0116] The specific limitations of the event prediction system and the event prediction model can refer to the limitations of the prediction method of the target event in the above, which will not be repeated here. Each module in the above event prediction system and event prediction model can be realized by software, hardware, and a combination thereof in whole or in part. The above modules can be embedded in or independent of the processor in the computer device in hardware form, or can be stored in the memory in the computer device in software form, so as to be called and executed by the processor to perform the operations corresponding to each of the above modules.

[0117] In one embodiment, as shown in Figure 7 , an event prediction system is provided, Figure 7 , an event prediction system is provided, Figure 6 , an event prediction system is provided,

[0118] The interactive event input of the user is encoded by the encoding module to obtain an event feature vector. The encoding module can include a first sub-encoding module and a second sub-encoding module, which are respectively used for encoding position information and attribute information. The event feature vector is input into the sequence model to obtain a sequence feature vector for representing user behavior.

[0119] The static attribute of the user is input into the static model to obtain a static attribute feature vector for representing the user's own attributes.

[0120] The static attribute feature vector and the sequence feature vector are input into the fusion model to calculate the occurrence probability of the target event corresponding to the fusion model, i.e., the prediction result.

[0121] In one embodiment, a computer device is provided, which can be a server, and the internal structure diagram thereof can be as shown in Figure 8 , an event prediction system is provided, Figure 8 , an event prediction system is provided,

[0122] The computer device comprises a processor, a memory, a network interface and a database connected through a system bus. The processor of the computer device is configured to provide computing and control capabilities. The memory of the computer device comprises a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for running the operating system and the computer program in the non-volatile storage medium. The database of the computer device is configured to store prediction method related data of a target event. The network interface of the computer device is configured to communicate with an external terminal through a network connection. The computer program is configured to be executed by the processor to implement a prediction method of a target event.

[0123] Those skilled in the art can understand that, Figure 8 The structure shown in the figure is only a block diagram of part of the structure related to the scheme of the present application, and does not constitute a limitation on the computer device to which the scheme of the present application is applied. The specific computer device can comprise more or fewer components than those shown in the figure, or combine certain components, or have a different arrangement of components.

[0124] In one embodiment, a computer device is provided, comprising a memory, a processor and a computer program stored in the memory and executable on the processor, and the processor implements the following steps when executing the computer program:

[0125] S201: Obtain a plurality of interaction events of a target user, wherein the interaction events are used to constitute a behavior sequence of the target user.

[0126] S202: Encode the interaction events respectively to form event feature vectors.

[0127] S203: Perform feature extraction and serialization processing on the event feature vectors to form a sequence feature vector of the behavior sequence, and the sequence feature vector is used as a hidden state of the target user.

[0128] S204: Fuse the hidden state of the target user and the static attribute to obtain a prediction result of the occurrence of a target event.

[0129] In one embodiment, the processor further implements the following steps when executing the computer program:

[0130] S301: Obtain a plurality of interaction events of a target user.

[0131] In this embodiment, a plurality of interaction events of a target user are obtained, wherein the interaction events are used to constitute a behavior sequence of the target user.

[0132] Periodically obtain the interaction events occurred within a preset time length at a preset time length interval; and / or, in response to the occurrence of a new interaction event, obtain the interaction events of the target user.

[0133] S302: encode the interaction events respectively to form event feature vectors.

[0134] In this embodiment, encoding an interaction event to form an event feature vector can be encoding position information of the interaction event to form a first sub-vector, and encoding attribute information of the interaction event to form a second sub-vector; splicing the first sub-vector and the second sub-vector to form an event feature vector belonging to the current interaction event.

[0135] The position information of the interaction event is the position of the interaction event in the behavior sequence determined based on the occurrence time of the interaction event.

[0136] The attribute information includes at least one of the type, duration, and result of the interaction event.

[0137] S303: feature extraction is performed on the event feature vectors to obtain a behavior feature vector.

[0138] In this embodiment, based on the occurrence order of each interaction event, the information of the previous interaction event is sequentially combined with each event feature vector for feature extraction to obtain a behavior feature vector.

[0139] S304: the behavior feature vector is subjected to serialization processing to form a sequence feature vector as a hidden state of the target user.

[0140] In this embodiment, the behavior feature vector recursively calls the previously formed sequence feature vector to form a sequence feature vector as a hidden state of the target user.

[0141] The feature extraction and serialization processing of each event feature vector form a sequence feature vector of the behavior sequence, and the sequence feature vector is a hidden state of the target user.

[0142] S305: obtaining a static attribute feature vector of the target user.

[0143] In this embodiment, the static attributes of the target user are subjected to feature transformation to form a static attribute feature vector of the target user.

[0144] S306: fusing the sequence feature vector and the static attribute feature vector to obtain a prediction result of the occurrence of the target event.

[0145] In this embodiment, the sequence feature vector and the static attribute feature vector of the target user are subjected to fusion processing to obtain a prediction result of the occurrence of the target event.

[0146] In this way, the hidden state and the static attributes of the target user can be fused to obtain a prediction result of the occurrence of the target event, and the reliability of the prediction result is improved.

[0147] In one embodiment, a computer readable storage medium is provided, and the computer readable storage medium stores a computer program, and the computer program is executed by a processor to implement the following steps:

[0148] S201: Obtain a plurality of interaction events of a target user, wherein the interaction events are used to constitute a behavior sequence of the target user.

[0149] S202: Encode the interaction events respectively to form event feature vectors.

[0150] S203: Perform feature extraction and serialization processing on the event feature vectors to form a sequence feature vector of the behavior sequence, and the sequence feature vector is used as a hidden state of the target user.

[0151] S204: Fuse the hidden state of the target user and a static attribute to obtain a prediction result of a target event.

[0152] In one embodiment, the computer program is executed by the processor to further implement the following steps:

[0153] S301: Obtain a plurality of interaction events of a target user.

[0154] In this embodiment, a plurality of interaction events of a target user are obtained, wherein the interaction events are used to constitute a behavior sequence of the target user.

[0155] Periodically obtain the interaction events occurred within a preset time length at intervals of the preset time length; and / or, in response to a new interaction event, obtain the interaction events of the target user.

[0156] S302: Encode the interaction events respectively to form event feature vectors.

[0157] In this embodiment, encoding an interaction event to form an event feature vector can be encoding position information of the interaction event to form a first sub-vector, and encoding attribute information of the interaction event to form a second sub-vector; and splicing the first sub-vector and the second sub-vector to form the event feature vector belonging to the current interaction event.

[0158] The position information of the interaction event is a position of the interaction event in the behavior sequence determined based on a time of occurrence of the interaction event.

[0159] The attribute information includes at least one of a type, a time length, and a result of the interaction event.

[0160] S303: Perform feature extraction on the event feature vectors to obtain a behavior feature vector.

[0161] In the embodiment, based on the occurrence sequence of each interaction event, the information of the previous interaction event is extracted for each event feature vector to obtain a behavior feature vector.

[0162] In the embodiment, the behavior feature vector is recursively called for the previously formed sequence feature vector to form a sequence feature vector as the hidden state of the target user.

[0163] In the embodiment, the behavior feature vector is recursively called for the previously formed sequence feature vector to form a sequence feature vector as the hidden state of the target user.

[0164] The sequence feature vector of the behavior sequence is formed by the feature extraction and the sequence processing of each event feature vector, and the sequence feature vector is the hidden state of the target user.

[0165] S305: Obtain a static attribute feature vector of the target user.

[0166] In the embodiment, the static attribute of the target user is transformed to form a static attribute feature vector of the target user.

[0167] S306: Fuse the sequence feature vector and the static attribute feature vector to obtain a prediction result of the target event.

[0168] In the embodiment, the sequence feature vector and the static attribute feature vector of the target user are fused to obtain a prediction result of the target event.

[0169] In this way, the hidden state and the static attribute of the target user can be fused to obtain a prediction result of the target event, and the reliability of the prediction result is improved.

[0170] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program. The computer program can be stored in a non-volatile computer readable storage medium, and when the computer program is executed, the processes of the above-mentioned embodiments of the methods can be included. Any reference to memory, storage, databases, or other media in the embodiments provided by the present application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. As an illustration but not limitation, RAM is available in many forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDR SDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0171] The technical features of the above embodiments can be combined in any way. In order to make the description simple, not all possible combinations of the technical features in the above embodiments are described, but as long as the combinations of the technical features do not exist, they should be considered as the scope of the present application.

[0172] The above embodiments only express several implementation manners of the present application, and the description is more specific and detailed, but it should not be understood as a limitation on the scope of the patent. It should be pointed out that for ordinary skilled in the art, without departing from the concept of the present application, some modifications and improvements can be made, which are all within the scope of the present application. Therefore, the scope of the patent of the present application should be subject to the appended claims.

Claims

1. A method of predicting a target event, characterized by, The method comprises the following steps: obtaining a plurality of interaction events of a target user, wherein the interaction events are used to constitute a behavior sequence of the target user; encoding the interaction events respectively to form event feature vectors; extracting features from each of the event feature vectors and performing serialization processing to form a sequence feature vector of the behavior sequence, the sequence feature vector being a hidden state of the target user; the hidden state is obtained through pre-computation, and the hidden state is used to represent the state of the target user; obtaining static attributes of the target user; the static attributes are structured information not belonging to user interaction events; when it is necessary to obtain the probability of the target event occurring to the target user, fusing the hidden state and the static attributes of the target user to obtain a prediction result of the target event occurring; the step of extracting features from each of the event feature vectors and performing serialization processing to form a sequence feature vector of the behavior sequence comprises the following steps: based on the occurrence sequence of each of the interaction events, sequentially extracting features from each of the event feature vectors in combination with information of a previous interaction event to obtain a behavior feature vector; recursively calling the sequence feature vector formed previously by using the behavior feature vector to form a new sequence feature vector as the hidden state.

2. The method of claim 1, wherein the target event is a target event of a user. the step of encoding the interaction events to form event feature vectors comprises the following steps: encoding position information of the interaction events to form a first sub-vector, and encoding attribute information of the interaction events to form a second sub-vector; splicing the first sub-vector and the second sub-vector to form the event feature vector belonging to the current interaction event.

3. The method of claim 2, wherein the target event is a target event of a user. The position information of the interaction event is the position of the interaction event in the behavior sequence based on the occurrence time of the interaction event. The attribute information comprises at least one of the type, duration and result of the interaction event.

4. The method of claim 1, wherein the target event is a target event of a user. The step of fusing the hidden state and the static attributes of the target user to obtain a prediction result of the target event occurring comprises the following steps: performing feature transformation on the static attributes of the target user to form a static attribute feature vector of the target user; fusing the sequence feature vector and the static attribute feature vector of the target user to obtain the prediction result of the target event occurring.

5. The method of claim 1, wherein the target event is a target event of a user. The step of obtaining a plurality of interaction events of a target user comprises the following steps: periodically obtaining interaction events occurring within a preset time interval at a preset time interval; and / or, in response to a new interaction event occurring, obtaining interaction events of the target user.

6. An event prediction model implementing the method of any one of claims 1 to 5, characterized in that, The event prediction model comprises: an encoding module configured to encode a plurality of target user interaction events obtained respectively to form event feature vectors; wherein the interaction events are used to constitute a behavior sequence of the target user; a sequence model configured to extract features from each of the event feature vectors and perform serialization processing to form a sequence feature vector of the behavior sequence, the sequence feature vector being a hidden state of the target user; a static model configured to obtain static attributes of the target user; the static attributes are used to be fused with the hidden state to obtain a prediction result of the target event occurring.

7. An event prediction system for implementing the target event prediction method according to any one of claims 1 to 5, characterized by The event prediction system comprises: An event prediction model is configured to obtain a plurality of interaction events of a target user, wherein the interaction events are used to constitute a behavior sequence of the target user; the interaction events are encoded respectively to form event feature vectors; feature extraction and serialization processing are performed on the event feature vectors to form a sequence feature vector of the behavior sequence, and the sequence feature vector is used as a hidden state of the target user; A fusion model is configured to fuse the hidden state of the target user and a static attribute to obtain a prediction result of a target event.

8. A computer device comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that, The processor executes the computer program to implement the steps of the target event prediction method in any one of claims 1 to 5.

9. A computer readable storage medium having stored thereon a computer program, characterized in that, The computer program is executed by the processor to implement the steps of the target event prediction method in any one of claims 1 to 5.

Images