WO2023103584A1 - 对象处理方法、装置、计算机设备和存储介质 - Google Patents

对象处理方法、装置、计算机设备和存储介质 Download PDF

Info

Publication number
WO2023103584A1
WO2023103584A1 PCT/CN2022/125251 CN2022125251W WO2023103584A1 WO 2023103584 A1 WO2023103584 A1 WO 2023103584A1 CN 2022125251 W CN2022125251 W CN 2022125251W WO 2023103584 A1 WO2023103584 A1 WO 2023103584A1
Authority
WO
WIPO (PCT)
Prior art keywords
feature
interaction
historical
status
features
Prior art date
Application number
PCT/CN2022/125251
Other languages
English (en)
French (fr)
Inventor
乔阳
陈亮
方高林
Original Assignee
腾讯科技（深圳）有限公司
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 腾讯科技（深圳）有限公司 filed Critical 腾讯科技（深圳）有限公司
Publication of WO2023103584A1 publication Critical patent/WO2023103584A1/zh
Priority to US18/215,303 priority Critical patent/US20230342797A1/en

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
    • G06Q30/00—Commerce
    • G06Q30/02—Marketing; Price estimation or determination; Fundraising
    • G06Q30/0201—Market modelling; Market analysis; Collecting market data
    • G06Q30/0202—Market predictions or forecasting for commercial activities
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/90—Details of database functions independent of the retrieved data types
    • G06F16/95—Retrieval from the web
    • G06F16/953—Querying, e.g. by the use of web search engines
    • G06F16/9535—Search customisation based on user profiles and personalisation
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
    • G06Q10/00—Administration; Management
    • G06Q10/04—Forecasting or optimisation specially adapted for administrative or management purposes, e.g. linear programming or "cutting stock problem"

Definitions

  • the present application relates to the field of computer technology, in particular to an object processing method, device, computer equipment and storage medium.
  • the target users are usually selected from the user group based on artificial experience, and the content is pushed to the target users, and the content is not pushed to the non-target users in the user group.
  • the selected The target user is not necessarily the audience of the pushed content, but the user who has not been pushed the content may be the audience of the content, which leads to a low accuracy of the processing method of the user.
  • an object processing method, device, computer equipment, storage medium, and computer program product are provided.
  • An object processing method executed by a computer device, the method includes: obtaining historical interaction characteristics of a user object with respect to a historical resource object; obtaining historical status characteristics of dynamic influencing factors of the historical resource object; using the dynamic influencing factors to Changes in resource attributes that dynamically affect the historical resource object; historical status characteristics are determined based on the historical status information of the dynamic influencing factors; based on the historical interaction characteristics and the historical status characteristics, determine the user object
  • the conversion prediction feature for the target resource object at the current time and, based on the conversion prediction feature, predict the conversion possibility of the user object for the target resource object at the current time, so as to determine the conversion for the target resource object based on the conversion possibility.
  • An object processing device comprising: an interaction feature acquisition module, used to acquire historical interaction features of a user object with respect to a historical resource object; a status feature acquisition module, used to acquire the historical status of dynamic influencing factors of the historical resource object feature; the dynamic influencing factor is used to dynamically affect the change of the resource attribute of the historical resource object; the historical status feature is determined based on the historical status information of the dynamic influencing factor; the predictive feature determination module is used to determine based on the The historical interaction feature and the historical status feature are used to determine the conversion prediction feature of the user object for the target resource object at the current time; and, the possibility prediction module is used to predict the user object based on the conversion prediction feature.
  • the conversion possibility of the target resource object at the current time so as to determine the processing manner for the user object for the target resource object based on the conversion possibility.
  • a computer device comprising a memory and one or more processors, the memory storing computer readable instructions which, when executed by the processor, cause the one or more processors to perform the above A step in an object's processing method.
  • One or more non-volatile readable storage media on which computer-readable instructions are stored, and when the computer-readable instructions are executed by one or more processors, the one or more processors realize the above object Steps in a processing method.
  • a computer program product including computer-readable instructions, when the computer-readable instructions are executed by a processor, the steps of the above object processing method are realized.
  • Fig. 1 is an application environment diagram of an object processing method in some embodiments
  • Fig. 2 is a schematic flowchart of an object processing method in some embodiments
  • Fig. 3 is a comparison chart of the Shanghai Stock Exchange Index and the delivery conversion rate in some embodiments
  • Fig. 4 is a structural diagram of a feature generation model and an object conversion prediction model in some embodiments
  • Figure 5 is a structural diagram of a feature processing network in some embodiments.
  • Fig. 6 is a schematic diagram of the sample space of the whole scene in some embodiments.
  • Fig. 7 is a schematic flowchart of an object processing method in some embodiments.
  • Fig. 8 is a structural block diagram of an object processing device in some embodiments.
  • Figure 9 is a diagram of the internal structure of a computer device in some embodiments.
  • Figure 10 is a diagram of the internal structure of a computer device in some embodiments.
  • the object processing method provided in this application can be applied to the application environment shown in FIG. 1 .
  • the terminal 102 communicates with the server 104 through the network.
  • the server 104 can obtain the historical interaction characteristics of the user object with respect to the historical resource object, obtain the historical status characteristics of the dynamic influencing factors of the historical resource object, the dynamic influencing factors are used to dynamically affect the change of the resource attribute of the target resource object, and the historical status characteristics It is determined based on the historical situation information of dynamic influencing factors. Based on the historical interaction characteristics and historical situation characteristics, the conversion prediction characteristics of the user object for the target resource object at the current time are determined, and the conversion possibility of the user object for the target resource object is predicted based on the conversion prediction characteristics. Degree to determine what to do with a user object based on conversion likelihood.
  • the object processing method provided in this application can be applied in the financial field.
  • the target resource object can be a fund
  • the user object can be a user who purchases or pays attention to a fund
  • the conversion possibility is the probability of a user purchasing a fund.
  • the probability of a user purchasing a fund can be determined.
  • the probability threshold can be preset or set as required, for example, 0.6.
  • the user information including but not limited to user equipment information, user personal information, etc.
  • data including but not limited to data used for analysis, stored data, displayed data, etc.
  • information and data authorized by the user or fully authorized by all parties are all It is information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of relevant data need to comply with relevant laws, regulations and standards of relevant countries and regions.
  • information such as user objects, user resource objects, interaction features, and status features involved in this application are all obtained under the condition of full authorization.
  • the terminal 102 can be, but not limited to, various personal computers, notebook computers, smart phones, tablet computers and portable wearable devices, and the server 104 can be realized by an independent server or a server cluster composed of multiple servers.
  • the object processing method provided in this application may be based on artificial intelligence.
  • a feature generation model may be used to process historical interaction features and historical status features, so as to determine the conversion prediction of a user object for a target resource object at the current time feature.
  • a feature generation model is an artificial intelligence-based model, such as a trained neural network model, used to generate conversion prediction features.
  • an object conversion prediction model may be used to process the conversion prediction features to obtain the conversion possibility of the user object for the target resource object.
  • the object conversion prediction model is an artificial intelligence-based model, such as a trained neural network model, used to predict the conversion possibility.
  • the above application scenario is only an example, and does not constitute a limitation on the object processing method provided by the embodiment of this application.
  • the method provided by this embodiment of this application can also be applied in other application scenarios, such as the object processing method provided by this application
  • the method may be executed by the terminal 102, and the terminal 102 may upload the obtained conversion possibility degree corresponding to the user object to the server 104, and the server 104 may store the conversion possibility degree corresponding to the user object, or forward the conversion possibility degree corresponding to the user object to other end devices.
  • an object processing method is provided.
  • the method may be executed by a terminal, may also be executed by a server, and may also be executed jointly by the terminal and the server.
  • This method is applied in FIG. 1 server 104 as an example for description, including the following steps:
  • Step 202 acquiring historical interaction features of the user object with respect to the historical resource object.
  • the user object may be any natural person, such as a user who uses an application program.
  • Application programs include but are not limited to shopping software or financial software, for example, financial management software.
  • Resource objects can be virtual resources, including but not limited to game equipment, game pets, electronic coupons or electronic red envelopes, etc. Resource objects can also be real resources, including but not limited to gifts of cash or in-kind.
  • Historical resource objects refer to resource objects that have interacted with user objects in historical time. Historical resource objects can include resource objects that have interacted with user objects in all times before the current time, or can be a period of time before the current time.
  • a resource object that interacts with a user object. The interaction behavior between the user object and the resource object may include purchasing, following, clicking, or purchasing.
  • purchasing means that the user object purchases the resource object, which may be an offline purchase or an online purchase.
  • Following may refer to an operation of a user object following a resource object through the Internet, for example, it may be an operation of following a financial product displayed in a financial application software.
  • a click refers to a click operation performed by a user object on a resource object through the Internet.
  • Subscription refers to the operation in which a user object applies to purchase a resource object through the Internet. For example, when a user object purchases resource object A in historical time, resource object A is a historical resource object.
  • the historical interaction feature is the feature corresponding to the historical interaction data.
  • the historical interaction data can correspond to an interaction moment, which is the moment when the historical interaction data is generated, that is, the moment when the user object interacts with the resource object, for example, the user object visits Fund A's moment. Since the historical interaction features are obtained based on the historical interaction data, the interaction time corresponding to the historical interaction data is also the time when the historical interaction features are generated. Therefore, the interaction time corresponding to the historical interaction data is also the interaction time corresponding to the historical interaction features. For example, when the historical interaction data is the interaction data generated by the user object at time A, the time corresponding to the historical interaction feature is time A.
  • the historical interaction data may be stored in the server, or obtained by the server from other devices.
  • the historical interaction data includes information or interaction behavior types of historical resource objects.
  • the information of the historical resource object includes, but is not limited to, the identifier of the historical resource object or the price of the historical resource object.
  • the type of interactive behavior can be any behavior in the conversion link.
  • the conversion link includes the interactive behavior that needs to occur during the conversion process of the user object for the resource object.
  • the interactive behavior in the conversion link includes but is not limited to access (ie visit), click or subscribe.
  • Each resource object may correspond to a conversion link, and the conversion links corresponding to different resource objects may be the same or different.
  • the interactive behaviors in the conversion link are arranged according to the order in which the behaviors occur, and the higher the order of the behaviors, the higher the order of the interactive behaviors in the conversion link.
  • the interaction behaviors corresponding to the later behavior sequence will only occur when the interaction behavior corresponding to the earlier behavior sequence occurs.
  • the conversion link can be "visit ⁇ click ⁇ subscribe".
  • the occurrence of conversion means that the user object has generated all the interactive behaviors in the conversion link to the resource object, that is, it has generated the last interaction behavior in the conversion link. For example, if the conversion link of resource object A is "Visit ⁇ Click ⁇ Purchase", when a purchase behavior occurs between the user object and the resource object, it is determined that the user object has been converted to the resource object A.
  • the server can obtain the historical interaction data of the user object, and encode the historical interaction data to obtain the historical interaction features.
  • the historical interaction data includes data of multiple dimensions, the data of each dimension can be encoded separately to obtain the The coding features corresponding to the data of each data are combined, such as splicing, to obtain the historical interaction features. For example, if the historical interaction data is "Fund A, Fund A's price, purchase", then "Fund A", “Fund A's price” and “Purchase” are coded respectively to obtain the coding features corresponding to these three.
  • the encoding method may be any encoding algorithm, including but not limited to One-Hot Encoding.
  • encoding features can also be obtained by inputting data into an embedding layer.
  • the server can acquire the historical interaction data sequence of the user within the preset time range, the historical interaction data sequence includes multiple historical interaction data, and each historical interaction data corresponds to a different interaction moment, that is, each historical interaction data is For the interaction data generated by user objects at different times, the historical interaction data in the historical interaction data sequence are arranged according to the interaction time. The earlier the interaction time is, the higher the historical interaction data is in the interaction data sequence.
  • the server may encode each historical interaction data in the historical interaction data sequence to obtain historical interaction features corresponding to each historical interaction data.
  • the preset time range is the time range before the current time
  • the resource objects corresponding to the historical interaction data generated by the user object at different times can be different or the same, for example, in the historical interaction data sequence, time T 1
  • the corresponding historical interaction data is the data of the user object accessing the fund
  • the corresponding historical interaction data at the time T2 is the data of the user object purchasing the computer
  • the resource object at the time T1 is the fund
  • the resource object at the time T2 is the computer.
  • the server can arrange each historical interaction feature according to the interaction time to obtain a historical interaction feature sequence. The earlier the interaction time is, the higher the historical interaction feature is arranged in the historical interaction feature sequence.
  • the historical interaction feature sequence includes the historical interaction features generated by the user object at multiple moments, so as to reflect the user's interest preference.
  • Step 204 acquire the historical status characteristics of the dynamic influencing factors of the historical resource object; the dynamic influencing factors are used to dynamically affect changes in the resource attributes of the historical resource object; the historical status characteristics are determined based on the historical status information of the dynamic influencing factors.
  • the dynamic influencing factors refer to the factors that dynamically affect the changes of the resource attributes of the historical resource objects, including but not limited to resource factors or time factors.
  • Resource factors refer to factors related to resources. Resource factors correspond to resource information.
  • resource information can include market information, which includes but is not limited to Shanghai Stock Exchange Index, Dow Jones Index, US Dollar Index or new energy industry index.
  • Resource The information may also include information about changes in market information, including but not limited to changes in the Shanghai Composite Index, the Dow Jones Index, the US Dollar Index, or the new energy industry index.
  • a time factor refers to a time-related factor. The time factor corresponds to time information, and the time information includes but is not limited to week, month, day, first transaction identifier or second transaction identifier.
  • the first transaction identifier is any one of a transaction day identifier or a non-trade day identifier.
  • the second transaction identifier is any one of a transaction moment identifier or a non-transaction moment identifier.
  • a trading day flag is used to indicate a trading day
  • a non-trading day flag is used to indicate a non-trading day.
  • the trading time mark is used to indicate that it is a trading time
  • the non-trading time mark is used to indicate that it is not a trading time. For example, 1 is used as a trading day mark, and 0 is used as a non-trading mark.
  • the historical status information of dynamic influencing factors corresponds to historical moments, and different historical status information corresponds to different historical moments.
  • the historical status information of the dynamic influencing factors is used to represent the status of the dynamic influencing factors at the historical moment or a period of time before the historical moment, for example, the historical status information corresponding to the T1 time is used to represent the status of the dynamic influencing factors at T1 or The situation during a period of time before time T1 .
  • the historical situation information is the time information corresponding to the historical moment, which may include at least one of the week/month/day, hour, trading day identifier or trading period identifier corresponding to the historical moment
  • the historical situation information can be the resource information at the historical moment, for example, the Shanghai stock index at the historical moment, or the change amount of resource information in a period of time before the historical moment, for example, it can be the historical Changes in major global markets and industry indexes such as Shanghai Composite Index, Dow Jones Index, US Dollar Index or New Energy Industry Index in the past 1 day/7 days/30 days corresponding to the moment.
  • the historical status information of the dynamic influencing factors may be stored in the server, or may be acquired by the server from other devices.
  • the historical status characteristics of the dynamic influencing factors are obtained by encoding the historical status information of the dynamic influencing factors.
  • the historical moment corresponding to the historical status feature of the dynamic influencing factor is consistent with the historical moment corresponding to the historical status information of the dynamic influencing factor.
  • the historical moment corresponding to the historical situation feature may be the interaction time corresponding to the historical interaction feature.
  • the server may determine the interaction time corresponding to the historical interaction feature, determine the historical resource object corresponding to the historical interaction feature, and obtain the historical status characteristics of the dynamic influencing factors of the historical resource object at the interaction moment corresponding to the historical interaction feature.
  • the historical interaction feature is the feature generated by the data of the user object buying the fund at the time T1
  • the time T1 is the interaction time
  • the server can obtain the historical status characteristics of the dynamic influencing factors of the fund at the time T1 . Therefore, the time corresponding to the historical interaction feature and the historical situation feature is consistent.
  • the server can encode the historical status feature to obtain the historical status feature.
  • the server can determine the historical resource object and the interaction time corresponding to the historical interaction feature, and obtain the historical status characteristics of the dynamic influencing factors of the historical resource object at the interaction moment, for example, obtain the dynamic influencing factors of the historical resource object at the interaction time
  • the historical situation information at the time is encoded to obtain the historical situation characteristics. Therefore, a historical interaction feature can obtain a historical status feature, and the time corresponding to the historical status feature obtained according to the historical interaction feature is the interaction time of the historical interaction feature.
  • time information and resource information are as shown in Table 1.
  • Step 206 based on the historical interaction features and historical status features, determine the conversion prediction features of the user object for the target resource object at the current time.
  • the target resource object may be any resource object, and the target resource object may be the same as or different from the historical resource object, for example, it may be a fund.
  • the resource attribute of the resource object is an attribute related to the resource, for example, it may be the price of the resource object, for example, it may be the price of the fund.
  • the time corresponding to the historical interaction feature and the historical situation feature are consistent, and both are the interaction time of the historical interaction feature.
  • the conversion prediction feature is a feature used to predict the conversion possibility of the user object for the target resource object, and the conversion possibility refers to the probability of conversion.
  • the server may perform feature fusion based on historical interaction features and historical status features to obtain conversion prediction features of the user object for the target resource object at the current time.
  • feature fusion may be at least one of feature splicing, feature addition or feature multiplication.
  • the server can perform feature calculations based on historical interaction features to obtain incremental features.
  • the server can combine historical interaction features and historical status features to obtain historical splicing features, and perform feature operations on historical splicing features to obtain incremental features.
  • the incremental filtering feature corresponding to the feature is used to filter the incremental feature to obtain the conversion prediction feature of the user object for the target resource object at the current time.
  • the feature operation includes at least one of linear operation or nonlinear operation, linear operation includes but not limited to multiplication or addition operation, nonlinear operation includes but not limited to exponential operation, logarithmic operation or hyperbolic tangent (tanh function) operations.
  • the filtering process can be realized by feature multiplication.
  • the server can multiply the incremental filter feature with the value at the corresponding position in the incremental feature to obtain the user object
  • the server can first unify the dimensions of the incremental filter feature and the incremental feature, and then unify the dimensions of the incremental feature The filter feature is multiplied by the incremental feature to obtain the conversion prediction feature corresponding to the target resource of the user object at the current time.
  • the server can obtain a trained feature generation model, and the feature generation model is used to generate conversion prediction features.
  • the server can input historical interaction features and historical status features into the feature generation model, and obtain conversion prediction features.
  • the target resource object may be a resource object to be pushed.
  • the terminal may send a resource object push request for the target resource object to the server, and the resource object push request may carry an identifier of the target resource object.
  • the server may respond to the resource object push request to obtain a user object collection, for each user object in the user object collection, obtain the historical interaction characteristics of the user object with respect to the historical resource object, obtain the historical status characteristics of the dynamic influencing factors of the historical resource object, To determine whether the user object is the audience of the target resource object according to the acquired data, when it is determined that the user object is the audience of the target resource object, push the target resource object to the user object, when it is determined that the user object is not the audience of the target resource object, then The target resource object is not pushed to the user object.
  • Step 208 Predict the conversion possibility of the user object for the target resource object based on the conversion prediction feature, so as to determine the processing method for the user object for the target resource object based on the conversion possibility.
  • the processing methods include but are not limited to incentives or neglect.
  • Incentives refer to performing incentive operations on user objects, so as to urge user objects to convert to target resource objects.
  • Ignoring refers to not performing incentive actions on user objects for target resource objects.
  • Incentive operations include, but are not limited to, pushing target resource objects, pushing coupons for purchasing target resource objects, and the like.
  • the server may compare the conversion possibility with the possibility threshold, and when it is determined that the conversion possibility is greater than the possibility threshold, determine that the processing method for the user object is incentive, otherwise, determine that the processing method for the user object is ignore.
  • the processing method is incentive
  • the possibility threshold may be preset or set as required, for example, it may be 60%.
  • the server may obtain a trained object conversion prediction model, and the object conversion prediction model is used to predict the conversion possibility based on the conversion prediction features.
  • the object transformation prediction model and the feature generation model can be trained independently or jointly.
  • the server can obtain training samples, input the training samples into the feature generation model, and use the output of the feature generation model as a sample transformation Prediction features, input the sample conversion prediction features into the object conversion prediction model, use the output of the object conversion prediction model as the sample conversion probability, determine the model loss function based on the sample conversion probability, and use the model loss value to adjust the feature generation model and object conversion
  • the model parameters of the prediction model are iteratively trained until both the feature generation model and the object conversion prediction model converge, and the trained feature generation model and the trained object conversion prediction model are obtained.
  • the target resource object is a resource object to be pushed.
  • the server may acquire a user object set, and the user object set includes multiple user objects, and the user objects in the user object set may be stored in the server or acquired by the server from other devices.
  • the server can use the method from step 202 to step 208 to determine the processing method for the user object, and obtain the user object whose processing method is incentive from the user object collection, as the target user object, and send the target user object The user object pushes the target resource object.
  • the target user object is the audience user of the target resource object.
  • the historical interaction characteristics of the user object with respect to the historical resource object are obtained, the historical status characteristics of the dynamic influencing factors of the historical resource object are obtained, and based on the historical interaction characteristics and historical status characteristics, it is determined that the user object interacts with the target resource object at the current time.
  • the conversion prediction feature of the conversion prediction feature predicts the conversion possibility of the user object for the target resource object based on the conversion prediction feature, so as to determine the processing method for the user object based on the conversion possibility.
  • historical status features are determined based on the historical status information of dynamic influencing factors, so historical status features can reflect the historical status of dynamic influencing factors of target resource objects
  • the historical interaction feature can reflect the interaction situation of the user object with respect to the historical resource object
  • the conversion prediction feature of the user object at the current time for the target resource object is determined, so that when the conversion prediction feature is obtained , which not only considers the interaction between the user object and the resource object in the historical time, but also considers the status of the dynamic influencing factors of the resource object in the historical time, thereby improving the accuracy of the conversion prediction feature, and further improving the conversion prediction feature.
  • the accuracy of the resulting conversion likelihood Therefore, when the processing method of the user object is determined according to the conversion possibility, a processing method consistent with the user object can be obtained, which improves the accuracy of the processing method of the user object. It can be understood that the processing method of the user object determined by the traditional method is not accurate enough, which will cause the computer to perform some ineffective processing, thereby causing a waste of computer resources. The accuracy of the processing method of the present application is improved, so it can reduce the Invalidation processing, thus saving computer resources.
  • the user behavior sequence is a sequence of user behavior information in a period of time arranged in chronological order, corresponding to the historical interaction feature sequence. It can be seen from this that market conditions and time in financial scenarios have a greater relationship with user conversion.
  • the object processing method provided in this application can be applied in the field of precision marketing in a financial scenario, and can improve the conversion rate. For example, users in financial scenarios can be used as user objects, and funds can be used as resource objects.
  • historical resource objects and target resource objects can be funds, and market conditions and time can be used as dynamic factors affecting funds.
  • the processing method is to obtain the conversion probability of the users in the financial scene for the fund, and deliver information about the fund to users with a higher conversion probability, such as sending fund marketing text messages, thereby increasing the conversion rate.
  • the dynamic influencing factors include at least one of resource factors or time factors; the resource factors are resource factors that change dynamically in resource scenarios; obtaining the historical status characteristics of the dynamic influencing factors of historical resource objects includes: determining the generation history At the interaction moment of the interaction feature, determine the time status characteristics corresponding to the time factor based on the time information of the time factor at the interaction moment; obtain the resource information of the resource factor at the interaction moment, and determine the resource status characteristics corresponding to the resource factor based on the resource information; At least one of a condition characteristic or a resource condition characteristic, determines a historical condition characteristic.
  • the resource factor is a factor related to resources, and the resource value of the resource factor changes dynamically with time.
  • the resource value refers to the value of the resource, for example, the price of a fund.
  • the resource scene refers to the scene where the resource is located.
  • the resource scene is, for example, a market, for example, it may be at least one of a domestic market or a global market.
  • a condition feature is a feature that reflects a condition.
  • the time status feature refers to the status feature corresponding to the time factor, and is used to reflect the status of the time factor.
  • the resource status feature refers to the status feature corresponding to the resource factor, and is used to reflect the status of the resource factor.
  • Historical condition characteristics may include at least one of temporal condition characteristics or resource condition characteristics.
  • the server may obtain time information of the time factor at the interaction moment corresponding to the historical interaction feature, encode the time information, and obtain the time status feature of the time factor at the interaction moment.
  • the server can obtain resource information of resource factors at an interaction moment corresponding to historical interaction characteristics, encode the resource information, and obtain resource status characteristics of resource factors at this interaction moment.
  • the server may use at least one of the time status feature or the resource status feature as the historical status feature.
  • the server may use the time status feature as the historical status feature, or the resource status feature as the historical status feature, or the time status feature and the resource status feature as the historical status feature.
  • the historical status feature is determined based on the time status feature corresponding to the time factor or the resource status feature corresponding to the resource factor, The historical status feature can be made more in line with the actual situation, thereby improving the accuracy of the historical status feature.
  • determining the conversion prediction features of the user object for the target resource object at the current time includes: The feature of the degree of attention; based on the feature of the degree of attention, the conversion prediction feature of the user object for the target resource object at the current time is determined.
  • the attention degree feature is used to reflect the attention status of the user object to the target resource object. Since the attention of the user object to the target resource object has a great influence on the conversion of the user object to the target resource object, the accuracy of the conversion prediction feature can be improved by determining the conversion prediction feature of the user object based on the attention degree feature of the user object .
  • the server may perform feature fusion of the historical interaction features and the historical status features to obtain the feature of the degree of attention of the user object to the target resource object at the current time.
  • the server can determine the interaction moment of the historical interaction feature, and obtain the historical status characteristics of the dynamic influencing factors of the historical resource object corresponding to the historical interaction feature at the interaction moment , based on the historical interaction feature and the historical status feature, determine the user object's attention status feature on the target resource object at the interaction moment. Since the interaction moments corresponding to the historical interaction features are different, it is possible to obtain the characteristics of the user object's attention to the target resource object at different interaction moments.
  • the server can perform feature fusion on the attention status features at each interaction moment to obtain the attention degree features.
  • the result of the calculation is used as the degree of attention feature.
  • the attention status feature at the interaction moment is used to reflect the attention status of the user object on the target resource object at the interaction moment.
  • the feature generation model may include multiple feature processing networks, and each feature processing network in the feature generation model may be connected, for example, the output data of one feature processing network is input to another feature processing network
  • each feature processing network corresponds to a connection order
  • the output data of the feature processing network corresponding to the earlier connection order is input into the feature processing network corresponding to the later connection order.
  • a feature generation model is shown, which includes n feature processing networks, the n feature processing networks have a connection relationship, and the output data of the first feature processing network is input to the second In the feature processing network, the connection order of the first feature processing network is higher than that of the first feature processing network, and the connection order of the second feature processing network is later than that of the first feature processing network.
  • the feature processing network is used to generate attention status features, and each feature processing network can be used to generate attention status features at an interaction moment.
  • the server can determine the feature processing network corresponding to the interaction moment, and use the The historical interaction feature corresponding to the interaction moment and the historical status feature corresponding to the interaction moment are input into the feature processing network corresponding to the interaction moment, and the feature processing network is used to obtain the attention status feature at the interaction moment.
  • the feature processing network corresponding to the interaction moment may be determined based on the interaction moment, for example, the earlier the interaction moment is, the earlier the connection order of the feature processing network corresponding to the interaction moment is.
  • T 1- T n are n interaction moments
  • T j-1 time is the time before T j time
  • the feature processing network corresponding to T j time is the jth feature processing network.
  • x 1 -x n are historical interaction features corresponding to each interaction moment in T 1 -T n
  • x j is a historical interaction feature corresponding to T j .
  • the feature extraction network is a self-defined structure, which may be a network obtained by improving an existing network structure, for example, it may be a network obtained by improving a long short-term memory neural network (LSTM, Long short-term memory).
  • LSTM long short-term memory neural network
  • the feature extraction network When the feature extraction network is improved based on the long-term short-term memory neural network, the feature extraction network can also be called an improved long-term short-term memory neural network (FLSTM, Financial Long short-term memory). Of course, it can also be based on the Transformer model.
  • FLSTM Long-term short-term memory neural network
  • the server may use the feature of the degree of attention of the user object on the target resource object at the current time as the conversion prediction feature of the user object on the target resource object at the current time.
  • the server may obtain object information of the user object, encode the object information to obtain object encoding features of the user object, and obtain conversion prediction features based on the object encoding features and attention degree features of the user object. For example, the server may perform feature fusion such as feature concatenation processing on the object encoding feature and the degree of attention feature, and use the processing result as the conversion prediction feature.
  • the object information may include the attribute information of the user object, and may also include the resource interaction information of the user object.
  • the interaction information generated between the user object and the resource object before the current time may include, for example, the interaction information generated by the user object within a specified time range before the current time.
  • the conversion prediction feature is obtained by combining the object information and attention degree features of the user object. Since the object information of the user object can reflect the characteristics of the user object, the characteristics of the user object are integrated into the conversion prediction feature, which improves the conversion prediction. expressiveness of features.
  • the conversion prediction features may include attention degree features, and the conversion prediction features may also include object features.
  • the server can also perform feature extraction on object encoding features to obtain object extraction features, and splicing object extraction features and attention degree features to obtain conversion prediction features, so that conversion prediction features include object features.
  • the feature generation model may also include an object feature extraction network, which is used to extract and obtain object extraction features, and the object feature extraction network may use a fully connected neural network with any number of layers, with two layers of fully connected Networks
  • Figure 4 shows the object feature extraction network in the feature generative model.
  • Y1 refers to object information.
  • the object feature extraction network includes a feature encoding layer, a first feature extraction layer, and a second feature extraction layer.
  • the first feature extraction layer and the second feature extraction layer are respectively a fully connected neural network, and the feature encoding layer is used to process object information.
  • the first feature extraction layer is used for feature extraction of object coding features to obtain first extraction features
  • the second feature extraction layer is used for feature extraction of first extraction features to obtain object extraction features.
  • the conversion prediction feature of the user object is determined based on the user object's attention degree feature, which improves the conversion rate. Accuracy of predicted features.
  • each historical interaction feature corresponds to an interaction moment
  • the interaction moment is the moment when the historical interaction feature is generated
  • the historical status feature is the status feature of the dynamic influencing factors of the historical resource object at the interaction moment
  • the interaction moment is the moment within the preset time range before the current time
  • determining the characteristics of the degree of attention of the user object to the target resource object at the current time includes: for each historical interaction characteristic Corresponding interaction Moment, determine the previous moment of the interaction moment, obtain the attention status feature of the user object at the previous moment, and obtain the prior attention status feature; the prior attention status feature is used to represent the user object’s attention to the target resource object at the previous time Situation;
  • the incremental feature is processed to obtain the incremental feature at the interaction moment; the incremental feature is the increase of the historical interaction feature compared with the prior attention status feature Based on the historical interaction characteristics and historical status characteristics at the interaction moment, the attention status characteristics at
  • the interaction time is the time when the historical interaction data used to obtain the historical interaction features is generated.
  • the preset time range can be preset as required, for example, it can be the last 3 months or the last half year.
  • the interaction moment is a moment in a preset time range.
  • the previous moment of the interaction moment includes at least one of the interaction moments before the interaction moment, for example, the previous moment of the interaction moment is the interaction moment closest to the interaction moment before the interaction moment.
  • the historical interaction feature sequence includes multiple historical interaction features, each historical interaction feature is arranged according to the interaction time, then the interaction time corresponding to the previous historical interaction feature is the interaction time corresponding to the last historical interaction feature.
  • the sequence of historical interaction features is "historical interaction features at time T 1 , historical interaction features at time T 2 , and historical interaction features at time T 3 ", although time T 1 and time T 2 are both at time T 3 Before, but because time T2 is closest to time T3 , time T2 is the previous time of time T3 .
  • the attention status feature is used to represent the attention status of the user object to the target resource object. Incremental features can reflect the new features brought by the historical interaction features at the interaction moment relative to the attention status features at the previous moment.
  • the server may determine a preset time range, obtain resource interaction information of user objects within the preset time range, and the resource interaction information includes historical interaction data generated by user objects at multiple moments in the preset time range, and generate historical The moment of the interaction data is determined as the interaction moment, and the historical interaction data is encoded to obtain the historical interaction characteristics at the interaction moment, and the historical interaction characteristics are arranged according to the interaction moment to obtain the historical interaction characteristic sequence.
  • the historical interaction feature sequence is, for example, "x 1 , x 2 , x 3 ,..., x n ", and the interaction time corresponding to x j is T j , where 1 ⁇ j ⁇ n.
  • the server can determine the feature processing network corresponding to each interaction moment.
  • the interaction time When there is no previous interaction moment, such as interaction time
  • the attention status feature corresponding to the time for example, the aggregation feature corresponding to the interaction time may be obtained first, and the attention status feature at the interaction time is generated according to the aggregation feature corresponding to the interaction time.
  • the historical interaction feature sequence is “x 1 , x 2 , x 3 ,..., x n ”,
  • c j is the aggregation feature at T j time
  • h j is the attention status feature at T j time
  • x j is the historical interaction feature at time T j
  • t j is the time status feature at T j time
  • m j is the resource status feature at T j time
  • x 1 is the historical interaction feature at T 1 time
  • t 1 is the time status feature at T 1 time
  • m 1 is the resource status feature at T 1 time
  • the attention status feature h 1 at the time T 1 is obtained.
  • the interaction moment has a previous moment
  • the server can obtain the previous moment of the interaction moment, obtain the attention status feature at the previous moment, obtain the prior attention status feature, and combine the prior attention status feature with the interaction time
  • the following historical interaction features are spliced, and the feature operation is performed on the features obtained after splicing to obtain incremental features. Based on the historical interaction characteristics and historical status characteristics at the interaction moment, the incremental features are processed to obtain the attention status characteristics at the interaction moment.
  • Feature processing network can be used to obtain the attention status features, as shown in Figure 4, T1 time is the previous time T2 time, x2 is the historical interaction feature at T2 time, t2 is the time status feature at T2 time, m2 is the resource status feature at T 2 time, the attention status feature h 1 at T 1 time obtained by the first feature processing network is input to the second feature processing network, and x 2 , t 2 and m 2 are input to the second In the first feature processing network, the second feature processing network obtains the incremental features at T 2 based on the attention status features h 1 and x 2 at T 1 , and performs incremental feature based on x 2 , t 2 and m 2 Processing to obtain the attention status feature h 2 at time T 2 .
  • the feature processing network may also include an incremental feature generation network, and the incremental feature generation network is used to generate incremental features, and the server may combine the historical status features at the interaction time with the attention status features at the previous time Carry out splicing, input the spliced features into the incremental feature generation network, use the parameters of the incremental feature production network and the activation function to perform feature operations on the spliced features, and obtain the incremental features at the moment of interaction.
  • a feature processing network corresponding to time T is shown.
  • the feature processing network includes an incremental feature generation network.
  • the input data of the incremental feature generation network includes historical interaction features x t and T-1 Attention status feature h t-1 at time
  • the output of the incremental feature generation network is the incremental feature CS t at time T.
  • incremental feature CS t tanh(W c [h t-1 ,x t ]+b c ), where W c and b c are the parameters of the incremental feature generation network, and tanh is the hyperbolic tangent function, which is Activation functions for incremental feature generation networks.
  • the previous attention status feature is used to represent the user object’s attention to the target resource object at the previous moment
  • the increase in the interaction time is obtained.
  • Quantitative features, based on the historical interaction features and historical status features at the interaction moment the incremental features are processed to obtain the attention status characteristics at the interaction moment, and based on the attention status characteristics at each interaction moment, it is determined that the user object is targeted at the target at the current time. Therefore, the process of determining the attention degree feature involves the user's attention to the target resource object at the previous moment, thereby further improving the accuracy of the attention degree feature.
  • processing the incremental features to obtain the attention status features at the interaction moment includes: obtaining the aggregated features of the user object at the previous moment, and obtaining the First aggregate features; based on historical interaction features and historical status features, determine the incremental weights corresponding to incremental features; determine the aggregation weights corresponding to previous aggregation features, based on incremental weights and aggregation weights, for incremental features and previous aggregation features Perform weighted calculations to obtain the aggregation features at the interaction moment; determine the attention status characteristics at the interaction moment based on the aggregation characteristics at the interaction moment.
  • the aggregation feature corresponding to the interaction moment can be calculated by using the method of this embodiment.
  • the interaction moment does not have a previous moment in the historical interaction feature sequence, that is, when the interaction moment is the interaction moment corresponding to the historical interaction feature ranked first in the historical interaction feature sequence, the aggregation feature corresponding to the interaction moment is based on the interaction The historical interaction characteristics and historical situation characteristics of the moment are obtained.
  • the server can combine the historical interaction features at the interaction moment with the historical status features at the interaction moment to obtain the historical splicing features, determine the incremental weight corresponding to the incremental feature based on the historical splicing feature, and determine the incremental weight corresponding to the incremental feature based on the incremental weight and the aggregation weight. , carry out weighted calculation on the incremental feature and the previous aggregated feature, and obtain the aggregated feature at the interaction moment.
  • the aggregated features at the interaction moment are obtained through weighted calculations, and the weights can be used to determine how much of the features in the aggregated features at the interaction moment are from the previous aggregated features, and how much is from the incremental features.
  • the larger the aggregation weight the greater the degree from the previously aggregated features, and the larger the incremental weight, the greater the degree from the incremental features. In this way, the aggregation feature at the moment of interaction can be more in line with the real situation.
  • the server can combine the historical interaction features at the interaction moment with the attention status features at the previous time to obtain the first stitching feature, and determine the increment corresponding to the incremental feature based on the first stitching feature and the historical stitching feature feature.
  • the server may determine the aggregation weight corresponding to the previous aggregation feature based on the attention status feature at the previous moment and the historical interaction feature at the interaction moment. For example, the attention status feature at the previous moment and the historical interaction feature at the interaction moment can be concatenated to obtain the first concatenated feature, and the feature operation can be performed on the first concatenated feature to obtain the aggregation weight corresponding to the previous aggregated feature.
  • the server can process the aggregated features at the interaction moment with the attention status feature at the previous moment and the historical interaction feature at the interaction moment to obtain the attention status feature at the interaction moment. For example, the server can use the previous moment Concatenate the features of attention status and the historical interaction features of the interaction moment to obtain the first splicing feature, and process the aggregated features at the interaction moment based on the first splicing feature to obtain the attention status feature at the interaction moment.
  • the first splicing feature can be obtained Perform feature operation on spliced features to obtain the first spliced feature after feature operation, perform nonlinear operation on the aggregated features at the interaction moment, obtain aggregated features after nonlinear operation, combine the first spliced feature after feature operation with nonlinear operation The final aggregation features are multiplied to obtain the attention status features at the moment of interaction.
  • the aggregated features at the interaction moment can be obtained by using the feature processing network, as shown in Figure 4, taking the aggregated feature at time T2 as an example, the first feature processing network will obtain the time T1 Aggregate feature c 1 and attention status feature h 1 at time T 1 are input to the second feature processing network, and the second feature processing network can determine c 1 based on h 1 (that is, the attention status feature at the previous moment) and x 2 (that is, the aggregation weight corresponding to the previously aggregated feature), determine the incremental weight corresponding to the incremental feature based on x 2 , t 2 and m 2 , and then obtain the aggregated feature c 2 at T 2 through weighted calculation.
  • the first feature processing network will obtain the time T1 Aggregate feature c 1 and attention status feature h 1 at time T 1 are input to the second feature processing network, and the second feature processing network can determine c 1 based on h 1 (that is, the attention status feature at the previous moment) and x 2 (that is, the aggregat
  • the features of the state of attention at the moment of interaction can also be obtained by using a feature processing network.
  • the feature processing network can perform feature operations on the aggregated features at the moment of interaction to obtain the features of the state of attention at the moment of interaction, for example
  • the feature processing network may include an adjustment value generation network.
  • the adjustment value generation network is used to generate an aggregated adjustment value for adjusting the aggregated feature.
  • the aggregated adjustment value may be based on the attention status feature at the previous moment and the historical interaction feature at the interaction moment. Generated, for example, the attention status feature at the previous moment and the historical interaction feature at the interaction moment can be input into the adjustment value generation network to obtain the aggregated adjustment value corresponding to the aggregated feature at the interaction moment.
  • a feature processing network corresponding to time T is shown.
  • the feature processing network includes an adjustment value generation network, and the input of the adjustment value generation network includes historical interaction features x t at time T and time T-1. Focusing on the status features, the output of the adjustment value generation network is the aggregated adjustment value O t .
  • Aggregated adjustment value O t ⁇ (W o [h t-1 ,x t ]+b o ), where W o and b o are parameters of the adjustment value generation network.
  • [h t-1 , x t ] means splicing h t-1 and x t .
  • the server can use the obtained aggregation adjustment value to adjust the aggregation feature at the interaction moment to obtain the attention status feature at the interaction moment.
  • the server can also perform a nonlinear operation on the aggregated feature at the interaction moment, and adjust the aggregated feature after the nonlinear operation by using the aggregation adjustment value to obtain the attention status feature at the interaction moment.
  • the feature processing network also includes a nonlinear operation layer.
  • the input of the nonlinear operation layer is the aggregation feature c t at time T
  • the output of the nonlinear operation layer is the aggregation feature c after the nonlinear operation t and the aggregation adjustment value O t are input into the multiplication module.
  • a circle with a " ⁇ " in the figure represents the multiplication module.
  • the multiplication module is used for feature multiplication. Aggregate the adjusted value to perform feature multiplication to obtain the attention status feature h t at the interaction moment.
  • the dimension of the aggregation adjustment value may be the same as the dimension of the aggregation feature.
  • the incremental weight is determined based on the historical interaction characteristics and historical status characteristics at the interaction moment, the incremental weight is more in line with the real situation, and the aggregation feature reflects the characteristics left over before the interaction moment, based on the incremental weight Perform weighted calculations on the incremental features and the previous aggregated features to obtain the aggregated features at the interaction time, so that the aggregated features at the interaction time are generated based on the features at the interaction time and the features between the interaction moments, so that they can be inherited Some of the features before the interaction moment include the features added at the interaction moment, which improves the accuracy of the aggregated features at the interaction moment.
  • the historical status feature includes at least one of the time status feature at the interaction moment or the resource status feature at the interaction moment.
  • determining the incremental weight corresponding to the incremental feature includes : Based on the historical interaction characteristics at the interaction moment and the time status characteristics at the interaction moment, the first weight corresponding to the incremental feature is obtained; based on the historical interaction characteristics at the interaction moment and the resource status characteristics at the interaction moment, the incremental feature corresponding to the second weight; based on at least one of the first weight or the second weight, determine the incremental weight corresponding to the incremental feature.
  • the server can splice the historical interaction features at the interaction moment with the time status features at the interaction moment to obtain the first historical splicing feature, and splice the historical interaction features at the interaction moment with the resource status features at the interaction moment, Get the second history splice feature.
  • the server may determine the incremental weight corresponding to the incremental feature based on at least one of the first historical splicing feature or the second historical splicing feature. For example, the server may determine the first weight corresponding to the incremental feature based on the first historical splicing feature, determine the second weight corresponding to the incremental feature based on the second historical splicing feature, and obtain the incremental feature based on at least one of the first weight or the second weight.
  • the incremental weight corresponding to the quantity feature for example, the first weight can be used as the incremental weight, or the second weight can be used as the incremental weight, or the first weight and the second weight can be summed, and the summed The results are used as incremental weights.
  • the server may determine the third weight corresponding to the incremental feature based on the first splicing feature, and add at least one of the first weight or the second weight to the third weight to obtain the Incremental weight.
  • the first weight, the second weight, and the third weight can be summed, and the result of the summation can be used as the incremental weight corresponding to the incremental feature, and the aggregated feature at the interaction moment can be obtained by weighting the incremental weight .
  • the aggregation feature c t at the interaction moment f t *c t-1 +(i t +M t +T t )*CS t
  • c t is the aggregation feature at the interaction moment, that is, T
  • c t -1 is the aggregation feature at the previous time, namely T-1 time
  • f t is the aggregation weight corresponding to the previous aggregation feature
  • M t is the second weight
  • T t is the first weight
  • CS t is the incremental feature at time T.
  • the module with a "ten" in a circle is the summing module, and the summing model is used to perform summing operations, that is, the model of summing operations.
  • the first weight T t , the second weight M t and the second weight The three weights i t are input to the summation model for summing to obtain (i t +M t +T t ).
  • the feature processing network in the feature generation network may be an improved long short-term memory neural network.
  • f t can also be the forgetting gate in the long-term short-term memory neural network, which is used to determine the extent to which the last state c t-1 is passed back, and it can also be the input gate in the long-term short-term memory neural network, which determines The extent to which update information at step t is introduced.
  • M t can be called the market bias gate added on the basis of the long-term short-term memory neural network
  • T t can be called the time bias added on the basis of the long-term short-term memory neural network Door.
  • the first weight is determined based on the historical interaction characteristics at the interaction moment and the time status characteristics at the interaction moment
  • the second weight is determined based on the historical interaction characteristics at the interaction moment and the resource status characteristics at the interaction moment
  • the first weight conforms to the time condition
  • the second weight conforms to the resource condition, that is, the first weight and the second weight conform to the real situation, so that based on at least one of the first weight or the second weight, the incremental weight corresponding to the incremental feature is determined , so that the incremental weight conforms to the actual situation and improves the accuracy of the incremental weight.
  • the attention status feature is generated by inputting the historical interaction feature and the historical status feature into the feature processing network corresponding to the interaction moment; the feature processing network includes an incremental weight prediction network; based on the historical interaction feature and the historical status feature , determining the incremental weight corresponding to the incremental feature includes: inputting the historical interaction feature and the historical status feature into the incremental weight prediction network, and predicting the incremental weight corresponding to the incremental feature.
  • the server can input the historical interaction features and historical status features into the incremental weight prediction network, and predict the incremental weight corresponding to the incremental features.
  • the incremental weight prediction network can also There are multiple, for example, each historical status feature can correspond to an incremental weight prediction network.
  • the server can input the historical status feature and historical interaction feature into the incremental weight corresponding to the historical status feature In the prediction network, the weight predicted by the historical condition feature is obtained.
  • a feature processing network corresponding to time T is shown.
  • the feature processing network includes the first incremental weight prediction network and the second incremental weight prediction network.
  • the weight prediction network, the first incremental weight prediction network is an incremental weight prediction network corresponding to time status characteristics
  • the second incremental weight prediction network is an incremental weight prediction network corresponding to resource status characteristics.
  • the input data of the feature processing network corresponding to time T includes historical interaction feature x t at time T, time status feature Time t at time T , resource status feature Market t at time T, aggregation feature c t-1 at time T-1 , and Attention state feature h t- 1 at time T-1
  • the output data of feature processing network corresponding to time T includes aggregation feature c t at time T and attention state feature h t at time T.
  • the input of the first incremental weight prediction network includes the historical interaction feature x t and the time status feature Time t
  • the output of the first incremental weight prediction network is the first weight T t corresponding to the incremental feature at time T
  • the second The output result of the incremental weight prediction network is the second weight M t corresponding to the incremental feature at time T.
  • the first weight T t ⁇ (W t [Time t ,x t ]+b t )
  • the second weight M t ⁇ (W m [Market t ,x t ]+b m )
  • W t And b t is a parameter of the first incremental weight prediction network
  • W m and b m are parameters of the second incremental weight prediction network.
  • [Time t , x t ] means splicing Time t and x t .
  • ⁇ is the activation function of the network.
  • the activation function used when obtaining M t and T t can be tanh, since the value range of the output result of the tanh activation function is [-1, +1], which can better reflect the impact of time T and duration factors on time Whether the information added by T (that is, the historical interaction feature) has a positive or negative effect.
  • the feature processing network further includes a third incremental weight prediction network, and the third incremental weight prediction network is used to predict and obtain the interaction time The incremental feature corresponds to the third weight.
  • the feature processing network corresponding to time T includes the third incremental weight prediction network
  • the input data of the third incremental weight prediction network includes the historical interaction feature x t at time T and the attention status at time T-1
  • the output data of the third incremental weight prediction network is the third weight it corresponding to the incremental feature at time T.
  • the third weight it ⁇ (W i [h t-1 , x t ]+bi )
  • W i and bi are parameters of the third incremental weight prediction network.
  • ⁇ is the activation function of the network.
  • the historical interaction features and historical status features are input into the incremental weight prediction network, and the incremental weight corresponding to the incremental feature is predicted, so that the incremental weight can be accurately and quickly predicted, and the incremental weight is improved. accuracy and predictive efficiency.
  • the feature processing network also includes an aggregation weight prediction network, and determining the aggregation weight corresponding to the previous aggregation feature includes: inputting the previous attention status feature and the historical status feature at the interaction moment into the aggregation weight prediction network , to predict the aggregation weights corresponding to the previously aggregated features.
  • the server can concatenate the previous attention status features with the historical status features at the moment of interaction, input the spliced features into the aggregation weight prediction network, and use the aggregation weight prediction network network parameters and activation functions to predict the spliced
  • the features are subjected to feature operations to obtain the aggregation weights corresponding to the previously aggregated features.
  • the feature processing network at time T includes an aggregation weight prediction network.
  • the input of the aggregation weight prediction network includes the attention status feature h t-1 at time T-1 and the historical interaction feature x t at time T, and the output result is the aggregation weight f t corresponding to the aggregation feature c t-1 at time T-1.
  • aggregation weight f t ⁇ (W f [h t-1 , x t ]+b f ), where W f and b f are parameters of the aggregation weight prediction network.
  • ⁇ represents the activation function.
  • the entire feature processing network in Figure 5 can be called the User behavior part, which is a model for extracting and representing user historical behavior interests.
  • the weight prediction feature generation network can be called the Query part, which is the module used to generate the query vector in the attention mechanism.
  • the object feature extraction network can be called a DNN part, which is a module for extracting and representing user features.
  • the object transformation prediction model can also be called a full-space multi-objective module (Multi-task part).
  • DNN is the abbreviation of Deep Neural Networks, which means deep neural network in Chinese.
  • the prior attention status feature and the historical status feature at the interaction moment are input into the aggregation weight prediction network to obtain the aggregation weight corresponding to the prior aggregation feature, which improves the efficiency and accuracy of predicting the aggregation weight.
  • determining the characteristics of the degree of attention of the user object to the target resource object at the current time includes: obtaining the object characteristics of the user object and the current status of the dynamic influencing factors of the target resource object at the current time.
  • Situation characteristics based on the object characteristics and current situation characteristics, determine the weights corresponding to the attention situation characteristics at each interaction moment; use the weights corresponding to each attention situation characteristics to perform weighted calculations on each attention situation characteristics, and determine the user object at the current time.
  • the degree of interest characteristic for the target resource object includes: obtaining the object characteristics of the user object and the current status of the dynamic influencing factors of the target resource object at the current time.
  • Situation characteristics based on the object characteristics and current situation characteristics, determine the weights corresponding to the attention situation characteristics at each interaction moment; use the weights corresponding to each attention situation characteristics to perform weighted calculations on each attention situation characteristics, and determine the user object at the current time.
  • the degree of interest characteristic for the target resource object includes: obtaining the object characteristics of the user object and the current status of the dynamic
  • the current status feature is used to characterize the status of the dynamic influencing factors of the target resource object at the current time.
  • the current status features include the time status features of the time factors at the current time
  • the current status features include the resource status features of the resource factors at the current time.
  • the server may concatenate the object feature and the current situation feature to obtain a second concatenation feature, and obtain a weight prediction feature based on the second concatenation feature, and the weight prediction feature is used to predict the respective weights corresponding to the attention situation characteristics at each interaction moment.
  • the server may concatenate the time status feature at the current time, the resource status feature at the current time, and the object feature to obtain the second spliced feature.
  • the server may use the second spliced feature as the weight prediction feature, or perform a feature operation on the weight prediction feature to obtain the weight prediction feature.
  • the server can perform weight prediction based on the weight prediction feature and the attention status feature, and obtain the weight corresponding to the attention status feature. After obtaining the weights corresponding to each attention status feature, use the weight to Weighted calculation is performed on each attention status feature to obtain the attention degree feature of the user object for the target resource object at the current time.
  • the feature generation model may also include a weight prediction feature generation network, which is used to generate the weight prediction feature, and the server may input the second concatenation feature into the weight prediction feature generation network to obtain the weight prediction Features
  • Figure 4 shows the weight prediction feature generation network in the feature generation model, the time status feature at the current time is tq, the resource status feature at the current time is mq, the object encoding feature is X1, and the second concatenation
  • the feature is [X1, tq, mq], input [X1, tq, mq] to the weight prediction feature generation network, and get the weight prediction feature q (q in the figure is the weight prediction feature), assuming the weight of the weight prediction feature generation network
  • the parameter is Wq
  • the bias parameter is bq
  • the activation function is ⁇
  • an attention mechanism may be used to calculate weights corresponding to each attention status feature.
  • the following formulas can be used to calculate the respective weights corresponding to each status feature of concern.
  • W a is the model of the network used by the attention mechanism
  • ⁇ is the activation function of the network
  • e i is the result calculated by the attention mechanism
  • e i is the comparison between the i-th attention status feature and the weight prediction feature
  • the results obtained by the attention calculation, the results of each attention calculation are normalized to obtain the weight
  • ⁇ i is the weight corresponding to the i-th attention status feature.
  • the weighted calculation is performed on each state of attention feature to obtain the degree of attention feature.
  • the degree of attention feature can be expressed as:
  • act embedding is the feature of attention degree.
  • the weight corresponding to the concerned situation characteristics is determined, so that the calculated weights can conform to the characteristics of the object and the dynamic influencing factors at the current time.
  • the situation under time makes the weight more real and reliable.
  • predicting the conversion possibility of the user object for the target resource object at the current time based on the conversion prediction features includes: obtaining the conversion link corresponding to the target resource object; the conversion link includes the process of the user object converting the target resource object The interactive behavior that needs to occur in the conversion link; for each interactive behavior in the conversion link, based on the conversion prediction feature, the possibility of the user object’s interactive behavior for the target resource object is predicted, and the corresponding behavior occurrence probability of the interactive behavior is obtained; based on each behavior The possibility of occurrence obtains the conversion possibility of the user object for the target resource object at the current time; the conversion possibility is positively correlated with the behavior occurrence possibility.
  • the corresponding behavior probability of the interactive behavior is used to represent the probability that the user object has the interactive behavior for the target resource object.
  • the probability of occurrence of the behavior corresponding to the interactive behavior is used to represent the probability that the user object has the interactive behavior on the target resource object under the condition that the user object has already performed the forward behavior on the target resource object.
  • the greater the probability of behavior occurrence the greater the probability of the interaction behavior occurring.
  • the forward behavior of the interactive behavior refers to the interactive behavior that is arranged before the interactive behavior in the conversion link.
  • the forward behavior includes “visit” and "click”.
  • the behavior occurrence probability corresponding to the interactive behavior may be used to represent the possibility of the user object occurring the interactive behavior for the target resource object at the current time.
  • the server may acquire a trained object conversion prediction model, and the object conversion prediction model is used to predict the conversion possibility.
  • the object conversion prediction model may include a behavior prediction network corresponding to each interaction behavior in the conversion link, and the behavior prediction network corresponding to the interaction behavior is used to predict the occurrence probability of the behavior corresponding to the interaction behavior.
  • the server can input the transformation prediction features into each behavior prediction network to obtain the corresponding behavior occurrence probability of each interaction behavior. Taking the conversion link as "visit ⁇ click ⁇ purchase" as an example, as shown in Figure 4, the object conversion prediction model is shown.
  • the object conversion prediction model includes 3 behavior prediction networks, and the first behavior prediction network is the “visit” corresponding The behavior prediction network predicts the probability of "visiting”, the second behavior prediction network is the behavior prediction network corresponding to "click”, and predicts the probability of "clicking” when "visiting” has occurred, and the third behavior The prediction network is the behavior prediction network corresponding to "subscription”, which predicts the probability of "subscription” when "visit” and "click” have occurred.
  • the conversion possibility is positively correlated with the behavior occurrence possibility.
  • the server may perform a multiplication operation on the behavior occurrence probability corresponding to each interaction behavior, and use the result of the multiplication operation as the conversion probability of the user object for the target resource object at the current time.
  • the first behavior prediction network outputs the first probability
  • the second behavior prediction network outputs the second probability
  • the third behavior prediction network outputs the third probability
  • the first probability is multiplied by the second probability to obtain the fourth probability
  • the fourth probability is multiplied by the third probability to obtain the fifth probability
  • the fifth probability is taken as the conversion possibility.
  • the first probability is the probability that the user object "visits” and the second probability is the probability that the user object "visits” occurs when the user object has already “visited”.
  • the third probability is the probability of "subscription” occurring when the user object has "visited” and "clicked”.
  • the positive correlation refers to: when other conditions remain unchanged, the two variables change in the same direction, and when one variable changes from large to small, the other variable also changes from large to small. It is understandable that the positive correlation here means that the direction of change is the same, but it does not require that when one variable changes a little, the other variable must also change. For example, when variable a is 10 to 20, variable b is 100, and when variable a is 20 to 30, variable b is 120. In this way, the direction of change of a and b is that when a becomes larger, b also becomes larger. But when a is in the range of 10 to 20, b can be unchanged.
  • the conversion possibility of the user object for the target resource object at the current time is obtained based on each behavior occurrence possibility. Since the conversion possibility is positively correlated with the behavior occurrence probability, the accuracy of the conversion possibility is improved.
  • predicting the possibility of the interaction behavior of the user object with respect to the target resource object, and obtaining the behavior occurrence probability corresponding to the interaction behavior includes: obtaining the forward behavior of the interaction behavior from the conversion link; Based on the conversion prediction feature, predict the possibility of the user object interacting with the target resource object when the user object has already taken the forward behavior, and obtain the behavior occurrence probability corresponding to the interaction behavior.
  • the possibility of the user object interacting with the target resource object is predicted when the user object has already taken the forward behavior, and the corresponding behavior occurrence probability of the interaction behavior is obtained, which improves the behavior occurrence possibility. degree of efficiency and accuracy.
  • predicting the possibility of the interaction behavior of the user object with respect to the target resource object, and obtaining the behavior occurrence probability corresponding to the interaction behavior includes: obtaining a trained object conversion prediction model; the object conversion prediction model includes The behavior prediction network corresponding to each interactive behavior in the conversion link; the behavior prediction network corresponding to the interactive behavior is used to predict the occurrence probability of the behavior corresponding to the interactive behavior; the conversion prediction features are input into the behavior prediction network corresponding to each interactive behavior , to predict the probability of occurrence of behaviors corresponding to each interaction behavior.
  • the process of obtaining the trained object conversion prediction model may include: obtaining a sample user object set, the sample user object set includes a plurality of sample user objects, and the sample user object is the user used for training the object conversion prediction model object.
  • the preset duration may be, for example, a period of time after the marketing is delivered to the sample user objects in the sample user object set, for example, 1 month or 3 months after the delivery.
  • the sample user object may have behaviors in the conversion link. For example, during the conversion process, the user may have three behaviors: visit, click, and conversion.
  • the sample space of the whole scene is shown in Figure 6.
  • the conversion link is "Visit ⁇ Click ⁇ Purchase (that is, conversion)" is explained, including 3 sample labels, such as F1, F2 and F3, F1 represents the probability of "visiting" of the sample user object, F2 represents the “visiting” of the sample user object and The probability of "click”, F3 represents the probability of "visit", “click” and "purchase” of the sample user object.
  • V ⁇ 0,1 ⁇ represents "visit”
  • Y ⁇ 0,1 ⁇ represents "click”
  • Z ⁇ 0,1 ⁇ represents "conversion (purchase)”
  • the sample user object A has only "visited”
  • the sample user object B has "visited” and "clicked”
  • the sample user object B is determined
  • the conversion prediction features corresponding to the sample user objects are obtained, and the conversion prediction features may be generated by using a feature generation model.
  • Input the conversion prediction features of the sample user object into each behavior prediction network for example, into the behavior prediction network corresponding to "visit”, the behavior prediction network corresponding to "click”, and the behavior prediction network corresponding to "subscription”, and get
  • V 1, X)
  • the fourth predicted probability represents the probability of “clicking” and “visiting” the sample user object
  • model loss value (loss) corresponding to the object conversion prediction model based on the first prediction probability, the fourth prediction probability, the fifth prediction probability and the sample label, and use the model loss value to adjust each behavior prediction network in the object conversion prediction model until the model Converge to get the trained object transformation prediction model.
  • the sample label corresponding to the first predicted probability is F1
  • the sample label corresponding to the fourth predicted probability is F2
  • the sample label corresponding to the fifth predicted probability is F3.
  • the model loss value L can be expressed by the following formula:
  • V in the formula indicates that the sample label is F1
  • V&Y indicates that the sample label is F2
  • V&Y&Z indicates that the sample label is F3
  • n indicates the number of sample user objects used for one training.
  • pvisit represents the visit rate, for example, it represents the visit rate after marketing
  • the visit rate is used to represent the ratio of the number of visitors to the number of people placed
  • pctr represents the click rate
  • the click rate is used to represent the ratio of the number of clicks to the number of visitors
  • pcvr stands for conversion rate, which is used to characterize the ratio of the number of people who convert to the number of people who click.
  • the sample user object can be a user at any conversion stage, for example, it can be a user without “visiting”, a user with only “visiting”, a user of "visiting” and “clicking", “visiting” ", “click” and “subscribe” users, that is, the full number of users are used for training.
  • This training method has lower requirements for training samples, and the richness of samples is higher, which improves the accuracy of model training, and It can be applied to scenarios with a small number of feature samples, for example, it can be applied to financial scenarios to improve the accuracy of predicting conversion rates in financial scenarios.
  • the behavior occurrence probability corresponding to each interaction behavior is obtained through each behavior prediction network in the object conversion prediction model, which improves the efficiency and accuracy of obtaining the behavior occurrence probability.
  • the present application also provides an application scenario, where the above object processing method is applied.
  • the application scenario is a financial scenario
  • the resource object is a resource object in the financial scenario, for example, a fund
  • the application of the object processing method in this application scenario is as follows:
  • Step 702 receiving a push request for the target resource object sent by the terminal, where the push request carries the identifier of the target resource object, and obtaining a set of user objects in response to the push request;
  • Step 702 for each user object in the user object set, obtain the historical interaction feature sequence of the user object within the preset time range, each historical interaction feature in the historical interaction feature sequence corresponds to an interaction time, and the historical interaction feature sequence in the historical interaction feature sequence Each historical interaction feature is arranged according to the interaction time, and the interaction time is a time within the preset time range;
  • each historical interaction feature corresponds to a resource object, and the resource objects corresponding to different historical interaction features may be the same or different.
  • Step 704 for each interaction moment, obtain the time status feature at the interaction moment, and obtain the resource status feature at the interaction moment, and the resource status feature is used to represent the resource factor of the resource object corresponding to the historical interaction feature at the interaction moment the situation at the moment of the interaction;
  • the resource objects are different, and the resource factors of the resources may be the same or different.
  • the time status features at each interaction moment are respectively T 1 , T 2 , and T 3 corresponding to
  • the characteristics of the time, the characteristics of the resource status at each interaction moment are the characteristics of the resource factors corresponding to the time T 1 , T 2 , and T 3 , such as the duration of the market.
  • resource status features such as market conditions can be appropriately time-shifted according to actual conditions.
  • the resource status feature used is the change in closing price on day T, and If the forecast day has not yet closed and the closing price of T-day cannot be obtained, the closing price of T-1 day can be used during training (that is, shifted forward by one day), so as to ensure that the prediction can have the same logic data (T-1 day closing price).
  • Step 706 obtain the trained feature generation model, which includes multiple feature processing networks, determine the feature processing networks corresponding to each interaction moment, and for each interaction moment, the historical interaction features, time The status features and resource status features are input into the feature processing network corresponding to the interaction moment, and the attention status features at each interaction moment are obtained;
  • the feature generation network also includes a weight prediction feature generation network, which acquires the object coding feature of the user object, obtains the time status feature at the current time and the resource status feature at the current time, and converts the object coding feature, the time at the current time
  • the status features and the resource status features at the current time are input into the weight prediction feature generation network, and the weight prediction features are predicted.
  • the weights corresponding to the attention status features at each interaction moment are determined, and based on the obtained weights.
  • the weighted calculation is performed on each attention status feature to obtain the attention degree feature at the current time.
  • the resource status feature at the current time is used to represent the status of the resource factor of the target resource object at the current time
  • the feature generation network may also include an object feature extraction network, input the object encoding feature into the object feature extraction network to obtain the object extraction feature, and splicing the object extraction feature and the degree of attention feature to obtain the conversion prediction at the current time feature.
  • Step 712 obtain the trained object conversion prediction model, the object conversion prediction model includes the behavior prediction network corresponding to each interaction behavior in the conversion link corresponding to the target resource object, and input the conversion prediction features into the corresponding interaction behavior
  • the behavior prediction network the behavior occurrence probability corresponding to each interactive behavior is predicted, and the occurrence probability of each behavior is multiplied to obtain the conversion probability of the user object for the target resource object at the current time.
  • Step 714 Based on the conversion possibilities corresponding to each user object in the user object set, filter the target user object from the user object set, and push the target resource object or the content associated with the target resource object to the target user object.
  • the server may compare the conversion possibility with the possibility threshold, and when it is determined that the conversion possibility is greater than the possibility threshold, use the user object as the target user object.
  • the server can arrange the user objects in the user object collection according to the conversion possibility. For example, arrange the user objects in descending order of the conversion possibility to obtain the sequence of user objects. The greater the conversion possibility, the user object in The higher the ranking in the user object sequence is, the server can acquire the user object that is ranked before the sorting threshold from the user object sequence as the target user object.
  • the possibility threshold and the sorting threshold can be preset or set as required.
  • the content associated with the target resource object may be content used to motivate users to purchase the target resource object, for example, when the target resource object is a fund, it may be a fund coupon or the like.
  • the conversion prediction features are obtained by using time status features and resource status features, which improves the authenticity and reliability of the conversion prediction features. Efficiency and accuracy of conversion possibilities.
  • the object conversion prediction model is used in combination with the feature generation model. Since the object conversion prediction model can be obtained by using the full amount of samples for training, and the object conversion prediction model includes multiple behavior prediction networks, it is realized. Entire Space Multi-Task Model (ESMM, Entire Space Multi-Task Model), and since the feature generation model can be an improved long-term short-term memory neural network, the combination of the two models realizes a fusion of improved long-term short-term memory and full-space Neural Networks for Multi-Task Learning.
  • ESMM Space Multi-Task Model
  • the object processing method provided by this application can achieve better results when applied to financial scenarios.
  • Table 2 shows the effect of the object processing method provided by this application in the financial scenario.
  • the historical interaction feature sequence is constructed from 16 fund products that have been clicked, purchased, and searched in the last 30 days, and the characteristics of each fund product (such as yield, closing period, maximum drawdown, etc.) as the representation of the fund, if the length of the user's historical interaction feature sequence is shorter than 16, it will be complemented with all 0 features, and if it exceeds 16, it will be truncated to 16 in chronological order.
  • the DNN model is an ordinary fully-connected neural network that only uses whether it is transformed into a modeling target after delivery
  • the ESMM model is an ESMM structure neural network that introduces two goals of click and conversion Network
  • LSTM represents adding LSTM unit modeling user historical interest on the basis of ESMM model
  • FLSTM model refers to replacing LSTM unit with FLSTM unit
  • MFLSTM multi-task Financial Long short-term memory
  • the feature generation model of this application can also be based on the Transformer model.
  • a long-term user behavior sequence (such as nearly half a year) can be introduced, and its structure can be optimized according to the particularity of the financial scene, such as the time, market Factors such as market conditions are represented as embeddings, which are directly spliced with sequence element representation features, or as a position embedding-like structure fused with sequence element representations and then added to the Transformer structure to predict user conversion prediction features.
  • steps in the flow charts of the above embodiments are shown sequentially as indicated by the arrows, these steps are not necessarily executed sequentially in the order indicated by the arrows. Unless otherwise specified herein, there is no strict order restriction on the execution of these steps, and these steps can be executed in other orders. Moreover, at least some of the steps in the above-mentioned embodiments may include multiple steps or stages. These steps or stages are not necessarily executed at the same time, but may be executed at different times. The execution of these steps or stages The sequence is not necessarily performed sequentially, but may be performed alternately or alternately with other steps or at least a part of steps or stages in other steps.
  • an object processing device may adopt a software module or a hardware module, or a combination of the two becomes part of a computer device.
  • the device specifically includes: interactive feature acquisition Module 802, condition feature acquisition module 804, prediction feature determination module 806 and possibility degree prediction module 808, wherein:
  • An interaction feature acquisition module 802 configured to acquire historical interaction features of the user object with respect to the historical resource object.
  • the status feature acquisition module 804 is used to acquire the historical status feature of the dynamic influencing factors of the historical resource object; the dynamic influencing factor is used to dynamically affect the change of the resource attribute of the historical resource object; the historical status feature is based on the historical status of the dynamic influencing factor Information is determined.
  • a prediction feature determining module 806, configured to determine the conversion prediction feature of the user object for the target resource object at the current time based on the historical interaction feature and the historical status feature.
  • the possibility prediction module 808 is configured to predict the conversion possibility of the user object for the target resource object at the current time based on the conversion prediction feature, so as to determine the processing method for the target resource object for the user object based on the conversion possibility.
  • the dynamic influencing factors include at least one of resource factors or time factors; resource factors are resource factors that change dynamically in resource scenarios; the status feature acquisition module is also used to: determine the interaction moment when historical interaction features are generated , based on the time information of the time factor at the interaction moment, determine the time status feature corresponding to the time factor; obtain the resource information of the resource factor at the interaction time, and determine the resource status feature corresponding to the resource factor based on the resource information; and, based on the time status feature or At least one of the resource condition characteristics determines a historical condition characteristic.
  • the prediction feature determination module is also used to: determine the feature of the degree of attention of the user object to the target resource object at the current time based on the feature of historical interaction and the feature of the historical situation; and, based on the feature of the degree of attention of the user object, determine The time-targeted conversion prediction feature for the resource object.
  • the prediction feature determination module is also used to: for each interaction moment corresponding to the historical interaction feature, determine the previous moment of the interaction moment; the interaction moment corresponding to the historical interaction feature is generated for obtaining the historical interaction feature The moment of historical interaction data; obtain the attention status feature of the user object at the previous moment, and obtain the prior attention status feature; the prior attention status feature is used to represent the user object’s attention to the target resource object at the previous moment; based on the First pay attention to the status features and the historical interaction features at the interaction moment, and get the incremental features at the interaction moment; the incremental feature is the feature added by the historical interaction feature compared with the previous attention status feature; based on the historical interaction feature at the interaction moment and the historical status feature, processing the incremental feature to obtain the attention status feature at the interaction moment; the historical status feature at the interaction moment is the status feature of the dynamic influencing factors of the historical resource object at the interaction moment; and, based on each interaction The feature of the attention status at the moment determines the feature of the degree of attention of the user object to the target resource object
  • the predictive feature determination module is also used to: obtain the aggregated features of the user object at the previous moment to obtain the previous aggregated features; based on the historical interaction features and historical status features, determine the incremental weight corresponding to the incremental feature ; Determine the aggregation weight corresponding to the previous aggregation feature, based on the incremental weight and the aggregation weight, perform weighted calculations on the incremental feature and the previous aggregation feature, and obtain the aggregation feature at the interaction time; and, determine based on the aggregation feature at the interaction time Attention status characteristics at the moment of interaction.
  • the historical status feature includes at least one of the time status feature at the interaction moment or the resource status feature at the interaction moment
  • the predictive feature determination module is also used for: based on the historical interaction feature at the interaction moment and the interaction moment
  • the first weight corresponding to the incremental feature is obtained based on the time status feature under the interaction time
  • the second weight corresponding to the incremental feature is obtained based on the historical interaction feature at the interaction time and the resource status feature at the interaction time
  • based on the first weight or At least one of the second weights is to determine an incremental weight corresponding to the incremental feature.
  • the attention status feature is generated by inputting the historical interaction feature and the historical status feature into the feature processing network corresponding to the interaction moment; the feature processing network includes an incremental weight prediction network; and, the prediction feature determination module also uses Yu: Input the historical interaction features and historical status features into the incremental weight prediction network, and predict the incremental weights corresponding to the incremental features.
  • the feature processing network also includes an aggregation weight prediction network
  • the prediction feature determination module is also used to: input the prior attention status features and the historical status features at the interaction moment into the aggregation weight prediction network, and predict to obtain Aggregation weights corresponding to previously aggregated features.
  • the prediction feature determination module is also used to: obtain the object features of the user object and the current status features of the dynamic influencing factors of the target resource object at the current time; based on the object features and the current status features, determine the attention at each interaction moment The weights corresponding to the status features; and, using the weights corresponding to the status features of attention, perform weighted calculations on each status feature of concern, and determine the feature of the degree of attention of the user object for the target resource object at the current time.
  • the predictive feature determination module is also used to: obtain the object information of the user object; encode the object information to obtain the object encoding feature of the user object; and, based on the object encoding feature and attention degree feature of the user object, obtain The conversion prediction characteristics of the user object for the target resource object at the current time.
  • the possibility prediction module is also used to: obtain the conversion link corresponding to the target resource object; the conversion link includes the interaction behavior that needs to occur during the conversion process of the user object for the target resource object; for the conversion link For each interaction behavior in , based on the conversion prediction features, predict the possibility of user object interacting with the target resource object, and obtain the behavior occurrence probability corresponding to the interaction behavior; and, based on the occurrence probability of each behavior, get the user object in the current Time is the conversion possibility of the target resource object; the conversion possibility is positively correlated with the behavior occurrence probability.
  • the possibility prediction module is also used to: obtain the forward behavior of the interactive behavior from the conversion link; based on the conversion prediction feature, predict that the user object will target the target resource object when the forward behavior has occurred.
  • the probability of occurrence of interactive behavior the probability of occurrence of the behavior corresponding to the interactive behavior is obtained.
  • the possibility degree prediction module is also used to: obtain the trained object conversion prediction model; the object conversion prediction model includes the behavior prediction network corresponding to each interaction behavior in the conversion link; the behavior prediction network corresponding to the interaction behavior is used To predict the occurrence probability of the behavior corresponding to the interactive behavior; and input the conversion prediction features into the behavior prediction network corresponding to each interactive behavior, and predict the corresponding behavior occurrence probability of each interactive behavior.
  • Each module in the above-mentioned object processing apparatus may be fully or partially realized by software, hardware or a combination thereof.
  • the above-mentioned modules can be embedded in or independent of the processor in the computer device in the form of hardware, and can also be stored in the memory of the computer device in the form of software, so that the processor can invoke and execute the corresponding operations of the above-mentioned modules.
  • a computer device is provided.
  • the computer device may be a terminal, and its internal structure may be as shown in FIG. 9 .
  • the computer device includes a processor, a memory, a communication interface, a display screen and an input device connected through a system bus. Wherein, the processor of the computer device is used to provide calculation and control capabilities.
  • the memory of the computer device includes a non-volatile storage medium and an internal memory.
  • the non-volatile storage medium stores an operating system and computer readable instructions.
  • the internal memory provides an environment for the execution of the operating system and computer readable instructions in the non-volatile storage medium.
  • the communication interface of the computer device is used to communicate with an external terminal in a wired or wireless manner, and the wireless manner can be realized through WIFI, an operator network, NFC (Near Field Communication) or other technologies.
  • WIFI Wireless Fidelity
  • NFC Near Field Communication
  • an object processing method is realized.
  • the display screen of the computer device may be a liquid crystal display screen or an electronic ink display screen
  • the input device of the computer device may be a touch layer covered on the display screen, or a button, a trackball or a touch pad provided on the casing of the computer device , and can also be an external keyboard, touchpad, or mouse.
  • a computer device is provided.
  • the computer device may be a server, and its internal structure may be as shown in FIG. 10 .
  • the computer device includes a processor, memory and a network interface connected by a system bus. Wherein, the processor of the computer device is used to provide calculation and control capabilities.
  • the memory of the computer device includes a non-volatile storage medium and an internal memory.
  • the non-volatile storage medium stores an operating system, computer readable instructions and a database.
  • the internal memory provides an environment for the execution of the operating system and computer readable instructions in the non-volatile storage medium.
  • the database of the computer device is used to store the data involved in the object processing method.
  • the network interface of the computer device is used to communicate with an external terminal via a network connection. When the computer readable instructions are executed by the processor, an object processing method is realized.
  • Figure 9 and Figure 10 are only block diagrams of partial structures related to the solution of this application, and do not constitute a limitation on the computer equipment on which the solution of this application is applied, specifically
  • the computer device may include more or fewer components than shown in the figures, or combine certain components, or have a different arrangement of components.
  • a computer device including a memory and one or more processors, the memory stores computer readable instructions, and when executed by the processor, the computer readable instructions cause one or more processing The device executes the steps in the above method embodiments.
  • non-volatile readable storage media storing computer readable instructions that, when executed by one or more processors, cause the one or more processors to The steps in the foregoing method embodiments are implemented.
  • a computer program product including computer-readable instructions, when the computer-readable instructions are executed by a processor, the steps of the above object processing method are implemented.
  • Non-volatile memory may include read-only memory (Read-Only Memory, ROM), magnetic tape, floppy disk, flash memory or optical memory, etc.
  • Volatile memory can include Random Access Memory (RAM) or external cache memory.
  • RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM).

Landscapes

  • Engineering & Computer Science (AREA)
  • Business, Economics & Management (AREA)
  • Strategic Management (AREA)
  • Development Economics (AREA)
  • Accounting & Taxation (AREA)
  • Finance (AREA)
  • Entrepreneurship & Innovation (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Economics (AREA)
  • Marketing (AREA)
  • General Business, Economics & Management (AREA)
  • Game Theory and Decision Science (AREA)
  • Data Mining & Analysis (AREA)
  • Databases & Information Systems (AREA)
  • Human Resources & Organizations (AREA)
  • General Engineering & Computer Science (AREA)
  • Operations Research (AREA)
  • Quality & Reliability (AREA)
  • Tourism & Hospitality (AREA)
  • Management, Administration, Business Operations System, And Electronic Commerce (AREA)

Abstract

本申请涉及一种对象处理方法、装置、计算机设备和存储介质。所述方法包括：获取用户对象针对历史资源对象的历史交互特征(202)；获取历史资源对象的动态影响因素的历史状况特征；动态影响因素，用于动态影响目标资源对象的资源属性的变化；历史状况特征，是基于动态影响因素的历史状况信息确定的(204)；基于历史交互特征以及历史状况特征，确定用户对象在当前时间针对目标资源对象的转化预测特征(206)；基于转化预测特征预测用户对象针对目标资源对象的转化可能度，以基于转化可能度确定针对用户对象的处理方式(208)。本申请提供的对象处理方法可以应用于金融领域中。

Description

对象处理方法、装置、计算机设备和存储介质
本申请要求于2021年12月08日提交中国专利局，申请号为202111492149.2，申请名称为“对象处理方法、装置、计算机设备和存储介质”的中国专利申请的优先权，其全部内容通过引用结合在本申请中。
技术领域
本申请涉及计算机技术领域，特别是涉及一种对象处理方法、装置、计算机设备和存储介质。
背景技术
随着计算机技术和互联网技术的发展，越来越多的通过互联网向用户推送内容，例如，向用户投放广告或者向用户发放优惠券。由于不同的用户群体，对同样的推送内容可能产生不同的反应，故在推送内容之前需要先从用户中筛选出目标用户，然后将内容推送给目标用户。
传统技术中，通常是根据人工经验从用户群中筛选出目标用户，向目标用户推送内容，并且不向用户群中的非目标用户推送内容，然而由于人工筛选的误差较大，因而筛选出的目标用户并不一定是推送的内容的受众，而未被推送内容的用户有可能是该内容的受众，从而导致对用户的处理方式的准确度较低。
发明内容
根据本申请提供的各种实施例，提供一种对象处理方法、装置、计算机设备、存储介质和计算机程序产品。
一种对象处理方法，由计算机设备执行，所述方法包括：获取用户对象针对历史资源对象的历史交互特征；获取所述历史资源对象的动态影响因素的历史状况特征；所述动态影响因素，用于动态影响所述历史资源对象的资源属性的变化；历史状况特征，是基于所述动态影响因素的历史状况信息确定的；基于所述历史交互特征以及所述历史状况特征，确定所述用户对象在当前时间针对目标资源对象的转化预测特征；及，基于所述转化预测特征，预测所述用户对象在当前时间针对所述目标资源对象的转化可能度，以基于所述转化可能度确定针对所述目标资源对象对所述用户对象的处理方式。
一种对象处理装置，所述装置包括：交互特征获取模块，用于获取用户对象针对历史资源对象的历史交互特征；状况特征获取模块，用于获取所述历史资源对象的动态影响因素的历史状况特征；所述动态影响因素，用于动态影响所述历史资源对象的资源属性的变化；历史状况特征，是基于所述动态影响因素的历史状况信息确定的；预测特征确定模块，用于基于所述历史交互特征以及所述历史状况特征，确定所述用户对象在当前时间针对目标资源对象的转化预测特征；及，可能度预测模块，用于基于所述转化预测特征，预测所述用户对象在当前时间针对所述目标资源对象的转化可能度，以基于所述转化可能度确定针对所述目标资源对象对所述用户对象的处理方式。
一种计算机设备，包括存储器和一个或多个处理器，所述存储器存储有计算机可读指令，所述计算机可读指令被所述处理器执行时，使得所述一个或多个处理器执行上述对象处理方法中的步骤。
一个或多个非易失性可读存储介质，其上存储有计算机可读指令，所述计算机可读指令 被一个或多个处理器执行时，使得所述一个或多个处理器实现上述对象处理方法中的步骤。
一种计算机程序产品，包括计算机可读指令，所述计算机可读指令被处理器执行时实现上述对象处理方法的步骤。
本申请的一个或多个实施例的细节在下面的附图和描述中提出。本申请的其他特征、目的和优点将从说明书、附图以及权利要求书变得明显。
附图说明
为了更清楚地说明本申请实施例中的技术方案，下面将对实施例描述中所需要使用的附图作简单地介绍，显而易见地，下面描述中的附图仅仅是本申请的一些实施例，对于本领域普通技术人员来讲，在不付出创造性劳动的前提下，还可以根据这些附图获得其他的附图。
图1为一些实施例中对象处理方法的应用环境图；
图2为一些实施例中对象处理方法的流程示意图；
图3为一些实施例中上证指数与投放转化率的对比图；
图4为一些实施例中的特征生成模型以及对象转化预测模型的结构图；
图5为一些实施例中特征处理网络的结构图；
图6为一些实施例全场景的样本空间的示意图；
图7为一些实施例中对象处理方法的流程示意图；
图8为一些实施例中对象处理装置的结构框图；
图9为一些实施例中计算机设备的内部结构图；
图10为一些实施例中计算机设备的内部结构图。
具体实施方式
为了使本申请的目的、技术方案及优点更加清楚明白，以下结合附图及实施例，对本申请进行进一步详细说明。应当理解，此处描述的具体实施例仅仅用以解释本申请，并不用于限定本申请。
本申请提供的对象处理方法，可以应用于如图1所示的应用环境中。其中，终端102通过网络与服务器104进行通信。
具体地，服务器104可以获取用户对象针对历史资源对象的历史交互特征，获取历史资源对象的动态影响因素的历史状况特征，动态影响因素用于动态影响目标资源对象的资源属性的变化，历史状况特征是基于动态影响因素的历史状况信息确定的，基于历史交互特征以及历史状况特征，确定用户对象在当前时间针对目标资源对象的转化预测特征，基于转化预测特征预测用户对象针对目标资源对象的转化可能度，以基于转化可能度确定针对用户对象的处理方式。
本申请提供的对象处理方法，可以应用于金融领域中。例如目标资源对象可以为基金，用户对象可以是购买或关注基金的用户，转化可能度为用户购买基金的概率，利用本申请提供的对象处理方法，可以确定用户购买基金的概率，当用户购买基金的概率大于概率阈值时，向用户推送一些内容例如优惠券以激励用于购买基金，当用户购买基金的概率小于概率阈值时，不对用户进行处理。概率阈值可以预设或根据需要设置，例如为0.6。
需要说明的是，本申请所涉及的用户信息(包括但不限于用户设备信息、用户个人信息等)和数据(包括但不限于用于分析的数据、存储的数据、展示的数据等)，均为经用户授权或者经过各方充分授权的信息和数据，且相关数据的收集、使用和处理需要遵守相关国家 和地区的相关法律法规和标准。例如，本申请中涉及到的用户对象、用户资源对象、交互特征以及状况特征等信息都是在充分授权的情况下获取的。
其中，终端102可以但不限于是各种个人计算机、笔记本电脑、智能手机、平板电脑和便携式可穿戴设备，服务器104可以用独立的服务器或者是多个服务器组成的服务器集群来实现。
本申请提供的对象处理方法可以是基于人工智能的，例如，本申请中，可以利用特征生成模型对历史交互特征以及历史状况特征进行处理，从而确定用户对象在当前时间针对目标资源对象的转化预测特征。特征生成模型是基于人工智能的模型，例如为训练好的神经网络模型，用于生成转化预测特征。再例如，本申请中，可以利用对象转化预测模型对转化预测特征进行处理，得到用户对象针对目标资源对象的转化可能度。对象转化预测模型是基于人工智能的模型，例如为训练好的神经网络模型，用于预测转化可能度。
可以理解，上述应用场景仅是一种示例，并不构成对本申请实施例提供的对象处理方法的限定，本申请实施例提供的方法还可以应用在其他应用场景中，例如本申请提供的对象处理方法可以是由终端102执行的，终端102可以将得到的用户对象对应的转化可能度上传至服务器104，服务器104可以存储用户对象对应的转化可能度，也可以将用户对象对应的转化可能度转发至其他终端设备。
在一些实施例中，如图2所示，提供了一种对象处理方法，该方法可以由终端执行，也可以由服务器执行，还可以由终端和服务器共同执行，以该方法应用于图1中的服务器104为例进行说明，包括以下步骤：
步骤202，获取用户对象针对历史资源对象的历史交互特征。
其中，用户对象可以是任意的自然人，例如可以是使用应用程序的用户。应用程序包括但不限于是购物类的软件或金融类的软件，例如可以是理财类的软件。资源对象可以是虚拟的资源，包括但不限于是游戏装备、游戏宠物、电子优惠券或电子红包等。资源对象也可以是真实的资源，包括但不限于是现金或实物礼品。历史资源对象是指历史时间中与用户对象产生交互行为的资源对象，历史资源对象可以包括当前时间之前的所有时间中与用户对象产生交互行为的资源对象，也可以是当前时间之前的一段时间中与用户对象产生交互行为的资源对象。用户对象与资源对象之间的交互行为可以包括购买、关注、点击或申购，例如购买是指用户对象购买资源对象，可以是线下购买也可以是线上购买。关注可以是指用户对象通过互联网对资源对象的关注操作，例如，可以是对理财类的应用软件中展示的理财产品的关注操作。点击是指用户对象通过互联网对资源对象的点击操作。申购是指用户对象通过互联网申请购买资源对象的操作。例如，当用户对象在历史时间中购买了资源对象A，则资源对象A为历史资源对象。
历史交互特征是历史交互数据对应的特征，历史交互数据可以对应一个交互时刻，该交互时刻是产生该历史交互数据时的时刻，也即用户对象与资源对象发生交互时的时刻，例如用户对象访问基金A的时刻。由于历史交互特征是基于历史交互数据得到的，故历史交互数据对应的交互时刻也是产生历史交互特征的时刻，因此，历史交互数据对应的交互时刻也是历史交互特征对应的交互时刻。例如，当历史交互数据是用户对象在时刻A产生的交互数据时，则历史交互特征对应的时刻为时刻A。历史交互数据可以是服务器中存储的，也可以是服务器从其他设备获取的。历史交互数据中包括历史资源对象的信息或交互行为类型。历史资源对象的信息包括但不限于是历史资源对象的标识或历史资源对象的价格。
交互行为类型可以是转化链路中的任意的行为，转化链路包括用户对象针对资源对象发生转化的过程中所需要发生的交互行为，转换链路中的交互行为包括但不限于是访问(即来访)、点击或申购。每个资源对象可以对应有转化链路，不同的资源对象所对应的转化链路可以相同也可以不同。转换链路中的交互行为按照行为发生顺序排列，行为发生顺序越靠前，则交互行为在转化链路中的排序越靠前。靠后的行为发生顺序对应的交互行为是在靠前的行为发生顺序对应的交互行为发生的情况下才会发生，例如由于只有“来访”来可能“点击”，故“来访”的行为发生顺序是靠前的，而“点击”的行为发生顺序是靠后的，故在转化链路中“来访”是排列在“点击”之前的，例如转化链路可以为“来访→点击→申购”。发生转化是指用户对象对资源对象产生了转化链路中的所有的交互行为，也即产生了转化链路中排列在最后的交互行为，例如若资源对象A的转化链路为“来访→点击→申购”，则当用户对象与资源对象之间产生了申购行为，则确定用户对象针对资源对象A发生了转化。
具体地，服务器可以获取用户对象的历史交互数据，对历史交互数据进行编码得到历史交互特征，当历史交互数据中包括多种维度的数据时，可以对各个维度的数据分别进行编码，得到各个维度的数据分别对应的编码特征，将各个编码特征进行组合例如进行拼接，得到历史交互特征。例如，历史交互数据为“基金A，基金A的价格，购买”，则分别对“基金A”、“基金A的价格”以及“购买”进行编码，得到这3者分别对应的编码特征。其中，编码所采用的方法可以是任意的编码算法，包括但不限于是独热编码(One-Hot Encoding)。当然，也可以通过将数据输入到嵌入层(embedding层)得到编码特征。
在一些实施例中，服务器可以获取用户在预设时间范围内的历史交互数据序列，历史交互数据序列中包括多个历史交互数据，各个历史交互数据对应不同的交互时刻，即各个历史交互数据是用户对象在不同时刻所产生的交互数据，历史交互数据序列中的历史交互数据按照交互时刻排列，交互时间越早，则历史交互数据在交互数据序列中的排序越靠前。服务器可以将历史交互数据序列中的各个历史交互数据进行编码，得到各个历史交互数据分别对应的历史交互特征。其中，预设时间范围为当前时间之前的时间范围，用户对象在不同时刻产生的历史交互数据所对应的资源对象可以是不同的也可以是相同的，例如，历史交互数据序列中，T 1时刻对应的历史交互数据是用户对象访问基金的数据，T 2时刻对应的历史交互数据是用户对象购买电脑的数据，T 1时刻的资源对象为基金，而T 2时刻的资源对象为电脑。服务器可以按照交互时刻，对各个历史交互特征进行排列，得到历史交互特征序列，交互时刻越靠前，则历史交互特征在历史交互特征序列中的排列越靠前。历史交互特征序列中包括了用户对象在多个时刻产生的历史交互特征，从而可以反映用户的兴趣偏好。
步骤204，获取历史资源对象的动态影响因素的历史状况特征；动态影响因素，用于动态影响历史资源对象的资源属性的变化；历史状况特征，是基于动态影响因素的历史状况信息确定的。
其中，动态影响因素是指动态影响历史资源对象的资源属性的变化的因素，包括但不限于资源因素或时间因素。资源因素是指与资源相关的因素，资源因素对应有资源信息，例如，资源信息可以包括市场行情信息，市场行情信息包括但不限于是上证指数、道琼斯指数、美元指数或新能源行业指数，资源信息还可以包括市场行情信息的变化信息，包括但不限于是上证指数、道琼斯指数、美元指数或新能源行业指数的指数变化量。时间因素是指与时间相关的因素，时间因素对应有时间信息，时间信息包括但不限于是周、月、日、第一交易标识或第二交易标识。第一交易标识为交易日标识或非交易日标识中的任意一个。第二交易标识 为交易时刻标识或非交易时刻标识中的任意一个。交易日标识用于表示是交易日，非交易日标识用于表示不是交易日。交易时刻标识用于表示是交易时刻，非交易时刻标识用于表示不是交易时刻。例如用1作为交易日标识，用0作为非交易标识。
动态影响因素的历史状况信息对应有历史时刻，不同的历史状况信息对应不同的历史时刻。动态影响因素的历史状况信息用于表征动态影响因素在历史时刻或历史时刻前的一段时间内的状况，例如，T 1时刻对应的历史状况信息用于表征动态影响因素在T 1时刻的状况或在T 1时刻前的一段时间内的状况。当动态影响因素为时间因素时，历史状况信息为历史时刻对应的时间信息，可以包括历史时刻对应的周/月/第几天、第几个小时、交易日标识或交易时段标识中的至少一个，当动态影响因素为资源因素时，历史状况信息可以为历史时刻的资源信息，例如，历史时刻的上证指数，也可以为历史时刻前的一段时间内的资源信息的变化量，例如可以是历史时刻对应的近1天/7天/30天上证指数、道琼斯指数、美元指数或新能源行业指数等全球主要市场和行业指数变化量。动态影响因素的历史状况信息可以是存储在服务器中的，也可以是服务器从其他设备获取的。
动态影响因素的历史状况特征是通过对动态影响因素的历史状况信息进行编码所得到的特征。动态影响因素的历史状况特征对应的历史时刻与动态影响因素的历史状况信息对应的历史时刻一致。
具体地，历史状况特征对应的历史时刻可以为历史交互特征对应的交互时间。服务器可以确定历史交互特征对应的交互时间，确定该历史交互特征对应的历史资源对象，获取该历史资源对象的动态影响因素在历史交互特征对应的交互时刻时的历史状况特征。例如，历史交互特征是用户对象在T 1时刻购买基金的数据产生的特征，则T 1时刻为交互时刻，服务器可以获取在T 1时刻基金的动态影响因素的历史状况特征。从而历史交互特征与历史状况特征对应的时刻是一致的。服务器可以对历史状况特征进行编码，得到历史状况特征。
在一些实施例中，历史交互特征有多个，不同的历史交互特征所对应的历史资源对象有可能是不同的，不同的历史交互特征所对应的交互时间也有可能是不同，对于每一个历史交互特征，服务器可以确定该历史交互特征对应的历史资源对象以及交互时刻，获取该历史资源对象的动态影响因素在该交互时刻时的历史状况特征，例如获取该历史资源对象的动态影响因素在该交互时刻的历史状况信息，对历史状况信息进行编码，得到历史状况特征。从而一个历史交互特征可以获取一个历史状况特征，并且根据历史交互特征得到历史状况特征对应的时刻为该历史交互特征的交互时刻。
在一些实施例中，时间信息以及资源信息如表1所示。
表1时间信息以及资源信息
Figure PCTCN2022125251-appb-000001
步骤206，基于历史交互特征以及历史状况特征，确定用户对象在当前时间针对目标资源对象的转化预测特征。
其中，目标资源对象可以是任意的资源对象，目标资源对象可以与历史资源对象相同也可以不同，例如可以是基金。资源对象的资源属性是与资源相关的属性，例如可以是资源对象的价格，例如可以是基金的价格。历史交互特征与历史状况特征对应的时刻一致，均为该历史交互特征的交互时刻。
转化预测特征是用于预测用户对象针对目标资源对象的转化可能度的特征，转化可能度是指发生转化的概率。
具体地，服务器可以基于历史交互特征以及历史状况特征进行特征融合，得到用户对象在当前时间针对目标资源对象的转化预测特征。其中，特征融合可以为特征拼接、特征相加或特征相乘中的至少一个。
在一些实施例中，服务器可以基于历史交互特征进行特征运算，得到增量特征，服务器可以将历史交互特征与历史状况特征进行拼接，得到历史拼接特征，对历史拼接特征进行特征运算，得到增量特征对应的增量过滤特征，利用增量过滤特征对增量特征进行过滤处理，得到用户对象在当前时间针对目标资源对象的转化预测特征。其中，特征运算包括线性运算或非线性运算中的至少一种，线性运算包括但不限于是乘法运算或加法运算，非线性运算包括但不限于是指数运算、对数运算或双曲正切(tanh函数)运算。过滤处理可以是通过特征相乘实现的，例如，当增量过滤特征与增量特征的维度相同时，服务器可以将增量过滤特征与增量特征中对应位置的数值进行相乘，得到用户对象在当前时间针对目标资源对象的转化预测特征，当增量过滤特征与增量特征的维度不相同时，服务器可以先统一增量过滤特征与增量特征的维度，然后将统一维度后的增量过滤特征与增量特征进行特征相乘，得到用户对象在当前时间针对目标资源对应的转化预测特征。
在一些实施例中，服务器可以获取已训练的特征生成模型，特征生成模型用于生成转化预测特征，服务器可以将历史交互特征以及历史状况特征输入特征生成模型中，预测得到转化预测特征。
在一些实施例中，目标资源对象可以是待推送的资源对象。终端可以向服务器发送针对目标资源对象的资源对象推送请求，资源对象推送请求中可以携带目标资源对象的标识。服务器可以响应于资源对象推送请求，获取用户对象集合，对于用户对象集合中的每个用户对象，获取用户对象针对历史资源对象的历史交互特征，获取历史资源对象的动态影响因素的历史状况特征，以根据获取的数据确定用户对象是否为目标资源对象的受众，当确定用户对象为目标资源对象的受众时，则向用户对象推送目标资源对象，当确定用户对象不是目标资源对象的受众时，则不向用户对象推送目标资源对象。
步骤208，基于转化预测特征预测用户对象针对目标资源对象的转化可能度，以基于转化可能度确定针对目标资源对象对用户对象的处理方式。
其中，处理方式包括但不限于是激励或忽视。激励是指对用户对象执行激励操作，以促使用户对象针对目标资源对象发生转化。忽视是指针对目标资源对象不对用户对象执行激励操作。激励操作包括但不限于是推送目标资源对象、推送用于购买目标资源对象的优惠券等。
具体地，服务器可以将转化可能度与可能度阈值进行对比，当确定转化可能度大于可能度阈值时，确定对用户对象的处理方式为激励，反之，确定对用户对象的处理方式为忽视。当处理方式为激励时，确定用户对象为目标资源对象的受众，当处理方式为忽视时，确定用户对象不是目标资源对象的受众。其中，可能度阈值可以是预设或根据需要设置的，例如可以为60％。
在一些实施例中，服务器可以获取已训练的对象转化预测模型，对象转化预测模型用于基于转化预测特征预测转化可能度。对象转化预测模型与特征生成模型可以是独立训练得到的，也可以是联合训练得到的，例如，服务器可以获取训练样本，将训练样本输入到特征生成模型中，将特征生成模型的输出作为样本转化预测特征，将样本转化预测特征输入到对象转化预测模型中，将对象转化预测模型的输出作为样本转化可能度，基于样本转化可能度确定模型损失函数，利用模型损失值调整特征生成模型以及对象转化预测模型的模型参数，迭代训练，直到特征生成模型以及对象转化预测模型均收敛，得到已训练的特征生成模型以及已训练的对象转化预测模型。
在一些实施例中，目标资源对象为待推送的资源对象。服务器可以获取用户对象集合，用户对象集合中包括多个用户对象，用户对象集合中的用户对象可以是存储在服务器中的也可以是服务器从其他设备获取的。对于用户对象集合中的每个用户对象，服务器可以采用步骤202-步骤208的方法确定对用户对象的处理方式，从用户对象集合中获取处理方式为激励的用户对象，作为目标用户对象，向目标用户对象推送目标资源对象。目标用户对象为目标资源对象的受众用户。
上述对象处理方法中，获取用户对象针对历史资源对象的历史交互特征，获取历史资源对象的动态影响因素的历史状况特征，基于历史交互特征以及历史状况特征，确定用户对象在当前时间针对目标资源对象的转化预测特征，基于转化预测特征预测用户对象针对目标资源对象的转化可能度，以基于转化可能度确定针对用户对象的处理方式。由于动态影响因素用于动态影响历史资源对象的资源属性的变化，历史状况特征是基于动态影响因素的历史状况信息确定的，因此历史状况特征能够反映目标资源对象的动态影响因素在历史时候的状况，由于历史交互特征能够反映用户对象针对历史资源对象产生的交互情况，从而基于历史交互特征以及历史状况特征，确定用户对象在当前时间针对目标资源对象的转化预测特征，使得在得到转化预测特征时，既考虑了用户对象与历史时间中与资源对象的交互情况，又考虑了资源对象的动态影响因素在历史时候的状况，从而提高了转化预测特征的准确度，进一步的提高了根据转化预测特征得到的转化可能度的准确度。从而当根据转化可能度确定用户对象的处理方式时，可以得到与用户对象相符的处理方式，提高了对用户对象的处理方式的准确度。可以理解，由于传统方法确定的对用户对象的处理方式不够准确，会导致计算机进行一些无效的处理，从而造成计算机资源的浪费，本申请的处理方式的准确度得到提高，因而能够一定程度上减少无效处理，从而节省了计算机资源。
经过研究发现金融营销场景中用户决策受市场行情、时间等因素的影响，用户在金融场景下的决策(如理财产品的点击、申购)受到市场、时间等因素的影响，如在某基金营销短信投放场景下，申购转化率随着上证指数波动的变化情况如图3所示，从图中可以看出用户申购转化率与上证指数变化趋势具有较强的一致性。产生这种现象的原因是，不同用户对时长行情、时间等因素波动的敏感程度不同(如市场下跌时有的用户会选择抄底，有的用户会选择止损，周末和节假日无法确定基金的份额，有的用户倾向于在交易日申购等)，这也会导致其历史行为序列中不同行为在预估时的重要性产生差异(如某用户喜欢在市场下跌时抄底，那么该用户历史行为序列中，市场下跌时的对应的行为重要性应该更高)，其中，用户行为序列中是用户在一段时间中的行为信息按照时间先后顺序排列成的序列，对应历史交互特征序列。由此可知，金融场景中市场行情以及时间对用户的转化具有较大的关系。本申请提供的对象处理方法，可以应用于金融场景下的精准营销领域中，可以提高转化率。例如， 可以将金融场景中的用户作为用户对象，将基金作为资源对象，例如，历史资源对象以及目标资源对象可以为基金，将市场行情以及时间作为基金的动态影响因素，基于本申请提供的对象处理方法，得到金融场景中的用户针对基金的转化概率，向转化概率较大的用户投放关于基金的信息，例如投放基金营销短信，从而提高了转化率。
在一些实施例中，动态影响因素包括资源因素或时间因素中的至少一个；资源因素是在资源场景下动态变化的资源因素；获取历史资源对象的动态影响因素的历史状况特征包括：确定产生历史交互特征时的交互时刻，基于时间因素在交互时刻时的时间信息确定时间因素对应的时间状况特征；获取资源因素在交互时刻的资源信息，基于资源信息确定资源因素对应的资源状况特征；基于时间状况特征或资源状况特征中的至少一个，确定历史状况特征。
其中，资源因素是与资源相关的因素，资源因素的资源价值随着时间动态变化。资源价值是指资源所具有的价值，例如为基金的价格。资源场景是指资源所处的场景，资源场景例如为市场，例如可以是国内市场或全球市场等中的至少一个。
产生历史交互特征时的交互时刻即上述的历史交互特征对应的交互时刻。状况特征是用于反映状况的特征。时间状况特征是指时间因素对应的状况特征，用于反映时间因素的状况。资源状况特征是指资源因素对应的状况特征，用于反映资源因素的状况。历史状况特征可以包括时间状况特征或资源状况特征中的至少一个。
具体地，服务器可以获取时间因素在历史交互特征对应的交互时刻下的时间信息，对该时间信息进行编码，得到时间因素在该交互时刻下的时间状况特征。服务器可以获取资源因素在历史交互特征对应的交互时刻下的资源信息，对该资源信息进行编码，得到资源因素在该交互时刻下的资源状况特征。
在一些实施例中，服务器可以将时间状况特征或资源状况特征中的至少一个作为历史状况特征。例如，服务器可以将时间状况特征作为历史状况特征，或将资源状况特征作为历史状况特征，或将时间状况特征以及资源状况特征作为历史状况特征。
本实施例中，由于时间因素以及资源因素的状况对资源对象的资源属性的变化具有较大的关系，故基于时间因素对应的时间状况特征或资源因素对应的资源状况特征，确定历史状况特征，可以使得历史状况特征更加符合实际情况，从而提高了历史状况特征的准确度。
在一些实施例中，基于历史交互特征以及历史状况特征，确定用户对象在当前时间针对目标资源对象的转化预测特征包括：基于历史交互特征以及历史状况特征，确定用户对象在当前时间针对目标资源对象的关注程度特征；基于关注程度特征，确定用户对象在当前时间针对目标资源对象的转化预测特征。
其中，关注程度特征用于反映用户对象对目标资源对象的关注情况。由于用户对象对目标资源对象的关注情况，与用户对象针对目标资源对象发生转化具有较大的影响，故基于用户对象的关注程度特征确定用户对象的转化预测特征，可以提高转化预测特征的准确度。
具体地，服务器可以将历史交互特征与历史状况特征进行特征融合，得到用户对象在当前时间针对目标资源对象的关注程度特征。当有多个历史交互特征时，对于每一个历史交互特征，服务器可以确定该历史交互特征的交互时刻，获取该历史交互特征对应的历史资源对象的动态影响因素在该交互时刻时的历史状况特征，基于该历史交互特征以及该历史状况特征，确定用户对象在该交互时刻下针对目标资源对象的关注状况特征。由于历史交互特征对应的交互时刻不同，故可以得到用户对象在多个不同的交互时刻下分别对目标资源对象的关注状况特征。服务器可以对各个交互时刻下的关注状况特征进行特征融合，得到关注程度特 征，例如，服务器可以确定各个关注状况特征分别对应的权重，基于确定的权重，对各个关注状况特征进行加权计算，将加权计算的结果作为关注程度特征。其中，交互时刻下的关注状况特征用于反映用户对象在该交互时刻对目标资源对象的关注情况。
在一些实施例中，特征生成模型中可以包括多个特征处理网络，特征生成模型中的各个特征处理网络可以是有连接关系的，例如，一个特征处理网络的输出数据输入到另一个特征处理网络中，各个特征处理网络对应有连接顺序，靠前的连接顺序对应的特征处理网络的输出数据输入到靠后的连接顺序对应的特征处理网络中。如图4所示，展示了一个特征生成模型，该特征生成模型中包括n个特征处理网络，这n个特征处理网络具有连接关系，并且第1个特征处理网络的输出数据输入到第2个特征处理网络中，故第1个特征处理网络的连接顺序靠前，第2个特征处理网络的连接顺序相对于第1个特征处理网络的连接顺序是靠后的。特征处理网络用于生成关注状况特征，每个特征处理网络可以用于生成一个交互时刻下的关注状况特征，例如，对于每个交互时刻，服务器可以确定该交互时刻对应的特征处理网络，将该交互时刻对应的历史交互特征以及该交互时刻对应的历史状况特征，输入到该交互时刻对应的特征处理网络中，利用该特征处理网络得到该交互时刻下的关注状况特征。交互时刻对应的特征处理网络可以是基于交互时刻确定的，例如，交互时刻越靠前，则交互时刻对应的特征处理网络的连接顺序越靠前。如图4所示，T 1-T n为n个交互时刻，T j-1时刻是T j时刻之前的时刻，则T j时刻对应的特征处理网络是第j个特征处理网络。x 1-x n是与T 1-T n中各个交互时刻分别对应的历史交互特征，例如，x j是与T j对应的历史交互特征。t 1-t n为T 1-T n中各个交互时刻分别对应的时间状况特征，例如，t j为T j时刻对应的时间状况特征，m 1-m n是与T 1-T n中各个交互时刻分别对应的资源状况特征，例如，m j为T j时刻对应的资源状况特征，1≤j≤n。其中，特征提取网络为自定义的结构，可以是对已有网络结构改进后所得到的网络，例如可以是基于长短期记忆神经网络(LSTM，Long short-term memory)进行改进得到的网络。当特征提取网络是基于长短期记忆神经网络改进得到的时，特征提取网络也可以称为改进长短期记忆神经网络(FLSTM，Financial Long short-term memory)。当然还可以是基于Transformer模型的。
在一些实施例中，服务器可以将用户对象在当前时间针对目标资源对象的关注程度特征，作为用户对象在当前时间针对目标资源对象的转化预测特征。
在一些实施例中，服务器可以获取用户对象的对象信息，对对象信息进行编码，得到用户对象的对象编码特征，基于用户对象的对象编码特征以及关注程度特征，得到转化预测特征。例如，服务器可以将对象编码特征与关注程度特征进行特征融合例如特征拼接处理，将处理的结果作为转化预测特征。其中，对象信息可以包括用户对象的属性信息，还可以包括用户对象的资源交互信息，用户对象的属性信息包括但不限于是用户对象的年龄、性别、职业或地区等，资源交互信息中可以包括用户对象在当前时间之前与资源对象产生的交互信息，例如可以包括当前时间之前的指定的时间范围内用户对象产生的交互信息。本实施例中，结合用户对象的对象信息和关注程度特征，得到转化预测特征，由于用户对象的对象信息能够反映用户对象的特征，故使得转化预测特征中融合用户对象的特征，提高了转化预测特征的表达能力。
在一些实施例中，转化预测特征中可以包括关注程度特征，转化预测特征中还可以包括对象的特征。例如，服务器还可以对对象编码特征进行特征提取，得到对象提取特征，将对象提取特征与关注程度特征进行拼接，得到转化预测特征，从而转化预测特征中包括了对象 的特征。例如，转化预测特征可以表示为share embedding＝[act embedding,feature embedding]。其中，share embedding为转化预测特征，act embedding为关注程度特征，feature embedding为对象提取特征。
在一些实施例中，特征生成模型中还可以包括对象特征提取网络，对象特征提取网络用于提取得到对象提取特征，对象特征提取网络可以使用任意层数的全连接神经网络，以两层全连接网络为了，如图4所示，展示了特征生成模型中的对象特征提取网络。其中，Y 1指的是对象信息。对象特征提取网络中包括特征编码层，第一特征提取层以及第二特征提取层，第一特征提取层以及第二特征提取层分别为一个全连接神经网络，特征编码层用于对对象信息进行编码，得到对象编码特征，第一特征提取层用于对对象编码特征进行特征提取，得到第一提取特征，第二特征提取层用于对第一提取特征进行特征提取，得到对象提取特征。例如，对象提取特征feature embedding可以表示为feature embedding＝σ(W 2O 1+b 2)，O 1＝σ(W 1Y 2+b 1)，其中，Y 2是特征编码层对对象信息Y 1进行编码所输出的对象编码特征，W 1以及b 1为第一特征提取层的参数，W 2以及b 2为第二特征提取层的参数，O 1为第一特征提取层得到的第一提取特征。
本实施例中，由于用户对象对目标资源对象的关注情况，与用户对象针对目标资源对象发生转化具有较大的影响，故基于用户对象的关注程度特征确定用户对象的转化预测特征，提高了转化预测特征的准确度。
在一些实施例中，历史交互特征为多个，各个历史交互特征对应有交互时刻，交互时刻是产生历史交互特征的时刻，历史状况特征是历史资源对象的动态影响因素在交互时刻下的状况特征；交互时刻是当前时间之前的预设时间范围内的时刻；基于历史交互特征以及历史状况特征，确定用户对象在当前时间针对目标资源对象的关注程度特征包括：对于每一个历史交互特征对应的交互时刻，确定交互时刻的在先时刻，获取用户对象在在先时刻下的关注状况特征，得到在先关注状况特征；在先关注状况特征用于表征用户对象在在先时刻对目标资源对象的关注情况；基于在先关注状况特征以及交互时刻下的历史交互特征，对增量特征进行处理，得到交互时刻下的增量特征；增量特征为历史交互特征相较于在先关注状况特征所增加的特征；基于交互时刻下的历史交互特征以及历史状况特征，得到交互时刻下的关注状况特征；基于各个交互时刻下的关注状况特征，确定用户对象在当前时间针对目标资源对象的关注程度特征。
其中，交互时刻是产生用于得到历史交互特征的历史交互数据的时刻。预设时间范围可以根据需要预先设置，例如可以是最近3个月或最近半年。交互时刻为预设时间范围中的时刻。交互时刻的在先时刻包括该交互时刻之前的交互时刻中的至少一个，例如，交互时刻的在先时刻是该交互时刻之前与该交互时刻最近的交互时刻。例如，历史交互特征序列中包括多个历史交互特征，各个历史交互特征按照交互时刻排列，则排列在前的历史交互特征对应的交互时刻，为排列在后的历史交互特征对应的交互时刻的在先时刻，例如，历史交互特征序列为“T 1时刻的历史交互特征，T 2时刻的历史交互特征，T 3时刻的历史交互特征”，虽然，T 1时刻以及T 2时刻都在T 3时刻之前，但由于T 2时刻与T 3时刻最近，故T 2时刻为T 3时刻的在先时刻。关注状况特征用于表征用户对象对目标资源对象的关注情况。增量特征可以反映交互时刻的历史交互特征相对于在先时刻的关注状况特征所带来的新的特征。
具体地，服务器可以确定预设时间范围，获取预设时间范围内用户对象的资源交互信息，资源交互信息中包括用户对象在该预设时间范围中的多个时刻产生的历史交互数据，产生历 史交互数据的时刻确定为交互时刻，对历史交互数据进行编码，得到交互时刻下的历史交互特征，按照交互时刻对各个历史交互特征进行排列，得到历史交互特征序列。历史交互特征序列例如为“x 1,x 2,x 3,…,x n”，x j对应的交互时刻为T j，1≤j≤n。
在一些实施例中，对于历史交互特征序列中的每个交互时刻分别对应的历史交互特征，服务器可以确定各个交互时刻分别对应的特征处理网络，当交互时刻不存在在先时刻时，例如交互时间为历史交互特征序列中排列在第一位的历史交互特征对应的交互时间时，将该交互时刻下的历史交互特征以及历史状况特征，输入到该交互时刻对应的特征处理网络中，得到该交互时刻对应的关注状况特征，例如可以先得到该交互时刻对应的聚合特征，根据该交互时刻对应的聚合特征生成该交互时刻下的关注状况特征。例如，图4中，历史交互特征序列为“x 1,x 2,x 3,…,x n”，c j为T j时刻下的聚合特征，h j为T j时刻下的关注状况特征，x j为T j时刻的历史交互特征，t j为T j时刻的时间状况特征，m j为T j时刻的资源状况特征，对于T 1时刻，x 1为T 1时刻的历史交互特征，t 1为T 1时刻的时间状况特征，m 1为T 1时刻的资源状况特征，将x 1、t 1以及m 1输入到第一个特征处理网络中，得到T 1时刻的聚合特征c 1，并基于聚合特征c 1得到T 1时刻的关注状况特征h 1。
在一些实施例中，交互时刻具有在先时刻，服务器可以获取该交互时刻的在先时刻，获取在先时刻下的关注状况特征，得到在先关注状况特征，将在先关注状况特征与交互时刻下的历史交互特征进行拼接，对拼接后得到的特征进行特征运算，得到增量特征。基于交互时刻下的历史交互特征以及历史状况特征，对增量特征进行处理，得到交互时刻下的关注状况特征。可以利用特征处理网络得到关注状况特征，如图4中，T1时刻为T 2时刻的在先时刻，x 2为T 2时刻的历史交互特征，t 2为T 2时刻的时间状况特征，m 2为T 2时刻的资源状况特征，第一个特征处理网络得到的T 1时刻的关注状况特征h 1输入到第二个特征处理网络中，并且将x 2、t 2以及m 2输入到第二个特征处理网络中，第二个特征处理网络基于T 1时刻的关注状况特征h 1以及x 2得到T 2时刻下的增量特征，并基于x 2、t 2以及m 2对增量特征进行处理，得到T 2时刻的关注状况特征h 2。
在一些实施例中，特征处理网络中还可以包括增量特征生成网络，增量特征生成网络用于生成增量特征，服务器可以将交互时刻下的历史状况特征与在先时刻下的关注状况特征进行拼接，将拼接后的特征输入到增量特征生成网络中，利用增量特征生产网络的参数以及激活函数对拼接后的特征进行特征运算，得到交互时刻下的增量特征。如图5所示，展示了一个T时刻对应的特征处理网络，该特征处理网络中包括增量特征生成网络，增量特征生成网络的输入数据包括T时刻的历史交互特征x t以及T-1时刻的关注状况特征h t-1，增量特征生成网络的输出为T时刻下的增量特征CS t。例如，增量特征CS t＝tanh(W c[h t-1,x t]+b c)，其中，W c以及b c为增量特征生成网络的参数，tanh为双曲正切函数，为增量特征生成网络的激活函数。
本实施例中，由于在先关注状况特征用于表征用户对象在在先时刻对目标资源对象的关注情况，故基于在先关注状况特征以及交互时刻下的历史交互特征，得到交互时刻下的增量特征，基于交互时刻下的历史交互特征以及历史状况特征，对增量特征进行处理，得到交互时刻下的关注状况特征，基于各个交互时刻下的关注状况特征，确定用户对象在当前时间针对目标资源对象的关注程度特征，故确定关注程度特征的过程中涉及到用户在在先时刻对目标资源对象的关注情况，从而进一步的提高了关注程度特征的准确度。
在一些实施例中，基于交互时刻下的历史交互特征以及历史状况特征，对增量特征进行 处理，得到交互时刻下的关注状况特征包括：获取用户对象在在先时刻下的聚合特征，得到在先聚合特征；基于历史交互特征以及历史状况特征，确定增量特征对应的增量权重；确定在先聚合特征对应的聚合权重，基于增量权重以及聚合权重，对增量特征以及在先聚合特征进行加权计算，得到交互时刻下的聚合特征；基于交互时刻下的聚合特征确定交互时刻下的关注状况特征。
其中，当交互时刻在历史交互特征序列中具有在先时刻时，交互时刻对应的聚合特征可以采用本实施例的方式计算得到。当交互时刻在历史交互特征序列中不具有在先时刻时，即交互时刻为历史交互特征序列中排列在第一位的历史交互特征对应的交互时刻时，则交互时刻对应的聚合特征基于该交互时刻的历史交互特征以及历史状况特征得到。
具体地，服务器可以将交互时刻下的历史交互特征与交互时刻下的历史状况特征进行拼接，得到历史拼接特征，基于历史拼接特征确定增量特征对应的增量权重，基于增量权重以及聚合权重，对增量特征以及在先聚合特征进行加权计算，得到交互时刻下的聚合特征。其中，通过加权计算得到交互时刻下的聚合特征，可以通过权重确定交互时刻下的聚合特征中的特征有多大程度是来自于在先聚合特征的，以及有多大程度是来自于增量特征的，聚合权重越大，则来自与在先聚合特征的程度越大，增量权重越大，则来自与增量特征的程度越大。从而可以使得交互时刻下的聚合特征更加的符合真实情况。
在一些实施例中，服务器可以将交互时刻的历史交互特征与在先时刻的关注状况特征进行特征拼接，得到第一拼接特征，基于第一拼接特征以及历史拼接特征确定增量特征对应的增量特征。
在一些实施例中，服务器可以基于在先时刻的关注状况特征以及交互时刻的历史交互特征，确定在先聚合特征对应的聚合权重。例如，可以将在先时刻的关注状况特征以及交互时刻的历史交互特征进行拼接，得到第一拼接特征，对第一拼接特征进行特征运算，得到在先聚合特征对应的聚合权重。
在一些实施例中，服务器可以在先时刻的关注状况特征以及交互时刻的历史交互特征，对交互时刻下的聚合特征进行处理，得到交互时刻下的关注状况特征，例如，服务器可以将在先时刻的关注状况特征与交互时刻的历史交互特征进行拼接，得到第一拼接特征，基于第一拼接特征对交互时刻下的聚合特征进行处理，得到交互时刻下的关注状况特征，例如，可以对第一拼接特征进行特征运算，得到特征运算后的第一拼接特征，对交互时刻下的聚合特征进行非线性运算，得到非线性运算后的聚合特征，将特征运算后的第一拼接特征与非线性运算后的聚合特征进行乘积运算，得到交互时刻下的关注状况特征。
在一些实施例中，交互时刻下的聚合特征可以是利用特征处理网络得到的，如图4中，以获取T 2时刻的聚合特征为例，第一个特征处理网络将得到的T 1时刻的聚合特征c 1以及T 1时刻的关注状况特征h 1输入到第二个特征处理网络中，第二个特征处理网络可以基于h 1(即在先时刻的关注状况特征)以及x 2确定c 1(即在先聚合特征)对应的聚合权重，基于x 2、t 2以及m 2确定增量特征对应的增量权重，从而加权计算得到T 2时刻的聚合特征c 2。
在一些实施例中，交互时刻下的关注状况特征也可以是利用特征处理网络得到的，例如，特征处理网络可以对交互时刻下的聚合特征进行特征运算，得到交互时刻下的关注状况特征，例如特征处理网络中可以包括调整值生成网络，调整值生成网络用于生成用于对聚合特征进行调整的聚合调整值，聚合调整值可以是基于在先时刻的关注状况特征以及交互时刻的历史交互特征生成的，例如，可以将在先时刻的关注状况特征以及交互时刻的历史交互特征输入 到调整值生成网络中，得到交互时刻下的聚合特征对应的聚合调整值。如图5所示，展示了一个T时刻对应的特征处理网络，特征处理网络中包括调整值生成网络，调整值生成网络的输入包括T时刻下的历史交互特征x t以及T-1时刻下的关注状况特征，调整值生成网络的输出为聚合调整值O t。聚合调整值O t＝σ(W o[h t-1,x t]+b o)，其中，W o以及b o为调整值生成网络的参数。[h t-1,x t]表示将h t-1与x t进行拼接。服务器可以利用得到的聚合调整值对交互时刻下的聚合特征进行调整得到交互时刻下的关注状况特征。服务器还可以对交互时刻下的聚合特征进行非线性运算，利用聚合调整值对非线性运算后的聚合特征进行调整，得到交互时刻下的关注状况特征。如图5所示，特征处理网络中还包括非线性运算层，非线性运算层的输入为T时刻下的聚合特征c t，将非线性运算层的输出结果即非线性运算后的聚合特征c t以及聚合调整值O t输入到乘法运算模块中，图5中一个圆圈中有一个“×”的图形代表乘法运算模块，乘法运算模块用于进行特征相乘，利用乘法运算模块对聚合特征以及聚合调整值进行特征相乘，得到交互时刻下的关注状况特征h t。其中，聚合调整值的维度可以与聚合特征的维度相同。当特征生成网络中的特征处理网络为改进长短期记忆神经网络时，其中，O t长短期记忆神经网络中的遗忘门输出门。
本实施例中，由于增量权重是基于交互时刻下的历史交互特征以及历史状况特征确定，因此增量权重更加符合真实情况，而聚合特征反映了交互时刻之前遗留下来的特征，基于增量权重对增量特征以及在先聚合特征进行加权计算，得到交互时刻下的聚合特征，使得交互时刻下的聚合特征是基于交互时刻下的特征以及交互时刻之间的特征所生成的特征，从而可以继承交互时刻之前的一部分特征又包括了交互时刻增加的特征，提高了交互时刻下的聚合特征的准确度。
在一些实施例中，历史状况特征包括交互时刻下的时间状况特征、或交互时刻下的资源状况特征中的至少一个，基于历史交互特征以及历史状况特征，确定增量特征对应的增量权重包括：基于交互时刻下的历史交互特征以及交互时刻下的时间状况特征，得到增量特征对应的第一权重；基于交互时刻下的历史交互特征以及交互时刻下的资源状况特征，得到增量特征对应的第二权重；基于第一权重或第二权重中的至少一个，确定增量特征对应的增量权重。
具体地，服务器可以将交互时刻下的历史交互特征与交互时刻下的时间状况特征进行拼接，得到第一历史拼接特征，将交互时刻下的历史交互特征与交互时刻下的资源状况特征进行拼接，得到第二历史拼接特征。服务器可以基于第一历史拼接特征或第二历史拼接特征中的至少一个确定增量特征对应的增量权重。例如，服务器可以基于第一历史拼接特征确定增量特征对应的第一权重，基于第二历史拼接特征确定增量特征对应的第二权重，基于第一权重或第二权重中的至少一个得到增量特征对应的增量权重，例如，可以将第一权重作为增量权重，或者，将第二权重作为增量权重，或者将第一权重与第二权重进行加和运算，将加和运算的结果作为增量权重。
在一些实施例中，服务器可以基于第一拼接特征确定增量特征对应的第三权重，将第一权重或第二权重中的至少一个与第三权重进行加和运算，得到增量特征对应的增量权重。例如，可以将第一权重、第二权重以及第三权重进行加和运算，将加和运算的结果作为增量特征对应的增量权重，利用该增量权重加权计算得到交互时刻下的聚合特征。例如，交互时刻下的聚合特征c t＝f t*c t-1+(i t+M t+T t)*CS t，其中，c t为交互时刻即T时刻下的聚合特征，c t-1为在先时刻即T-1时刻下的聚合特征，f t为在先聚合特征对应的聚合权重，i t为第三权重，M t为第 二权重，T t为第一权重，CS t为T时刻下的增量特征。如图5中，一个圆圈中有一个“十”的模块为加和模块，加和模型用于进行加和运算即求和运算的模型，将第一权重T t、第二权重M t以及第三权重i t，输入到加和模型中进行求和，得到(i t+M t+T t)。将(i t+M t+T t)以及T时刻下的增量特征CS t输入到乘法运算模块中，得到(i t+M t+T t)*CS t。将T-1时刻下的聚合特征c t-1对应的聚合权重f t以及T-1时刻下的聚合特征c t-1输入到乘法运算模块中，得到f t*c t-1。将(i t+M t+T t)*CS t以及f t*c t-1输入加和模块中得到f t*c t-1+(i t+M t+T t)*CS t，即得到c t。
在一些实施例中，特征生成网络中的特征处理网络可以是改进长短期记忆神经网络。其中，f t还可以为长短期记忆神经网络中的遗忘门，用于决定上一个状态c t-1以多大程度往后传递，i t还可以为长短期记忆神经网络中的输入门，决定t步的更新信息以多大程度引入。当资源影响因素为市场时，M t可以称为是在长短期记忆神经网络的基础上增加的市场偏置门，T t可以称为是在长短期记忆神经网络的基础上增加的时间偏置门。
本实施例中，由于第一权重基于交互时刻下的历史交互特征以及交互时刻下的时间状况特征确定，第二权重基于交互时刻下的历史交互特征以及交互时刻下的资源状况特征确定，因此，第一权重符合时间状况，第二权重符合资源的状况，即第一权重以及第二权重符合真实情况，从而基于第一权重或第二权重中的至少一个，确定增量特征对应的增量权重，使得增量权重符合实际的情况，提高了增量权重的精准度。
在一些实施例中，关注状况特征是将历史交互特征以及历史状况特征输入至交互时刻对应的特征处理网络中生成的；特征处理网络中包括增量权重预测网络；基于历史交互特征以及历史状况特征，确定增量特征对应的增量权重包括：将历史交互特征以及历史状况特征输入到增量权重预测网络中，预测得到增量特征对应的增量权重。
具体地，服务器可以将历史交互特征以及历史状况特征，输入到增量权重预测网络中，预测到增量特征对应的增量权重，当有多个历史状况特征时，增量权重预测网络也可以有多个，例如每个历史状况特征可以对应有一个增量权重预测网络，对于每一个历史状况特征，服务器可以将该历史状况特征以及历史交互特征，输入到该历史状况特征对应的增量权重预测网络中，得到该历史状况特征预测出的权重。以历史状况特征包括时间状况特征以及资源状况特征为例，如图5所示，展示了一个T时刻对应的特征处理网络，该特征处理网络中包括第一增量权重预测网络以及第二增量权重预测网络，第一增量权重预测网络为时间状况特征对应的增量权重预测网络，第二增量权重预测网络为资源状况特征对应的增量权重预测网络。T时刻对应的特征处理网络的输入数据包括T时刻的历史交互特征x t、T时刻的时间状况特征Time t、T时刻的资源状况特征Market t、T-1时刻的聚合特征c t-1以及T-1时刻的关注状况特征h t-1，T时刻对应的特征处理网络的输出数据包括T时刻的聚合特征c t以及T时刻的关注状况特征h t。第一增量权重预测网络的输入包括历史交互特征x t以及时间状况特征Time t，第一增量权重预测网络的输出结果为T时刻下的增量特征对应的第一权重T t，第二增量权重预测网络的输出结果为T时刻下的增量特征对应的第二权重M t。例如，第一权重T t＝σ(W t[Time t,x t]+b t),第二权重M t＝σ(W m[Market t,x t]+b m)，其中，W t以及b t为第一增量权重预测网络的参数，W m以及b m为第二增量权重预测网络的参数。[Time t,x t]表示将Time t与x t拼接。σ为网络的激活函数。其中，得到M t以及T t时采用的激活函数可以为tanh，由于tanh激活函数输出结果的取值范围为[-1，+1]，从而可以更好的反映出时间T和时长因素对时间T加入的信息(即历史交互特征)是正向还是负向的效果。在一些实施例中，特征处理网络中还包括第三增量权重预测网络，第三增量权重预测网络用于基于在先时刻的关注状况特征以 及交互时刻的历史交互特征，预测得到交互时刻下的增量特征对应的第三权重。如图5所示，T时刻对应的特征处理网络中包括第三增量权重预测网络，第三增量权重预测网络的输入数据包括T时刻的历史交互特征x t以及T-1时刻的关注状况特征h t-1，第三增量权重预测网络的输出数据为T时刻下的增量特征对应的第三权重i t。例如，第三权重i t＝σ(W i[h t-1,x t]+b i)，其中，W i以及b i为第三增量权重预测网络的参数。σ为网络的激活函数。
本实施例中，将历史交互特征以及历史状况特征输入增量权重预测网络中，预测得到增量特征对应的增量权重，从而可以准确且快速的预测出增量权重，提高了增量权重的精准度以及预测效率。
在一些实施例中，特征处理网络中还包括聚合权重预测网络，确定在先聚合特征对应的聚合权重包括：将在先关注状况特征以及交互时刻下的历史状况特征，输入到聚合权重预测网络中，预测得到在先聚合特征对应的聚合权重。
具体地，服务器可以将在先关注状况特征与交互时刻下的历史状况特征进行拼接，将拼接后的特征输入到聚合权重预测网络中，利用聚合权重预测网络的网络参数以及激活函数对拼接后的特征进行特征运算，得到在先聚合特征对应的聚合权重。如图5所示，T时刻的特征处理网络中包括聚合权重预测网络，聚合权重预测网络的输入包括T-1时刻的关注状况特征h t-1以及T时刻的历史交互特征x t，输出结果为T-1时刻的聚合特征c t-1对应的聚合权重f t。例如，聚合权重f t＝σ(W f[h t-1,x t]+b f),其中，W f以及b f为聚合权重预测网络的参数。σ代表激活函数。
图5中的各个特征处理网络构成的整体可以称为User behavior part，即是进行用户历史行为兴趣抽取和表示的模型。权重预测特征生成网络可以称为是Query part，即是用于生成attention机制中的查询(query)向量的模块。对象特征提取网络可以称为DNN part，即用于抽取和表示用户的特征的模块。对象转化预测模型也可以称为全空间多目标模块(Multi-task part)。其中，DNN为Deep Neural Networks的简称，中文意思为深度神经网络。
本实施例中，将在先关注状况特征以及交互时刻下的历史状况特征，输入到聚合权重预测网络中，得到在先聚合特征对应的聚合权重，提高了预测聚合权重的效率以及精准度。
在一些实施例中，基于各个交互时刻下的关注状况特征，确定用户对象在当前时间针对目标资源对象的关注程度特征包括：获取用户对象的对象特征以及当前时间目标资源对象的动态影响因素的当前状况特征；基于对象特征以及当前状况特征，确定各个交互时刻下的关注状况特征分别对应的权重；利用各个关注状况特征分别对应的权重，对各个关注状况特征进行加权计算，确定用户对象在当前时间针对目标资源对象的关注程度特征。
其中，当前状况特征用于表征目标资源对象的动态影响因素在当前时间下的状况。当动态影响因素包括时间因素时，当前状况特征包括时间因素在当前时间下的时间状况特征，当动态影响因素包括资源因素时，当前状况特征包括资源因素在当前时间下的资源状况特征。
具体地，服务器可以将对象特征与当前状况特征进行拼接，得到第二拼接特征，基于第二拼接特征得到权重预测特征，权重预测特征用于预测各个交互时刻下的关注状况特征分别对应的权重。服务器可以将当前时间下的时间状况特征、当前时间下的资源状况特征以及对象特征进行拼接，得到第二拼接特征。服务器可以将第二拼接特征作为权重预测特征，或者对权重预测特征进行特征运算，得到权重预测特征。对于每个交互时刻下的关注状况特征，服务器可以基于权重预测特征以及该关注状况特征进行权重预测，得到该关注状况特征对应的权重，获取到各个关注状况特征分别对应的权重后，利用权重对各个关注状况特征进行加 权计算，得到用户对象在当前时间针对目标资源对象的关注程度特征。
在一些实施例中，特征生成模型中还可以包括权重预测特征生成网络，权重预测特征生成网络用于生成权重预测特征，服务器可以将第二拼接特征输入到权重预测特征生成网络中，得到权重预测特征，如图4所示，展示了特征生成模型中的权重预测特征生成网络，当前时间下的时间状况特征为tq，当前时间下的资源状况特征为mq，对象编码特征为X1，第二拼接特征为[X1,tq,mq]，将[X1,tq,mq]输入到权重预测特征生成网络，得到权重预测特征q(图中的q为权重预测特征)，假设权重预测特征生成网络的权重参数为Wq，偏置参数为bq，激活函数为σ，σ包括但不限于是Sigmoid、relu或tanh，则权重预测特征q＝σ(Wq[tq,mq,X1]+bq)。
在一些实施例中，可以利用注意力(attention)机制计算得到各个关注状况特征分别对应的权重。例如，可以利用下面的公式计算得到各个关注状况特征分别对应的权重。其中，W a是注意力机制所采用的网络的模型，σ为该网络的激活函数，e i是通过注意力机制计算出的结果，e i为将第i个关注状况特征与权重预测特征进行注意力计算所得到的结果，将各个注意力计算的结果进行归一化处理得到权重，α i为第i个关注状况特征对应的权重。
Figure PCTCN2022125251-appb-000002
在一些实施例中，得到各个关注状况特征对应的权重后，对各个关注状况特征进行加权计算，得到关注程度特征。例如，关注程度特征可以表示为:
Figure PCTCN2022125251-appb-000003
其中，act embedding为关注程度特征。
本实施例中，基于对象特征以及目标资源对象的动态影响因素在当前时间下的当前状况特征，确定关注状况特征对应的权重，可以使得计算出的权重符合对象的特征并且符合动态影响因素在当前时间下的状况，使得权重更加的真实可靠。
在一些实施例中，基于转化预测特征预测用户对象在当前时间针对目标资源对象的转化可能度包括：获取目标资源对象对应的转化链路；转化链路包括用户对象针对目标资源对象发生转化的过程中所需要发生的交互行为；对于转化链路中的每个交互行为，基于转化预测特征预测用户对象针对目标资源对象发生交互行为的可能度，得到交互行为对应的行为发生可能度；基于各个行为发生可能度得到用户对象在当前时间针对目标资源对象的转化可能度；转化可能度与行为发生可能度成正相关关系。
其中，当交互行为在转化链路中不存在前向行为时，交互行为对应的行为发生可能度，用于表征用户对象针对目标资源对象发生该交互行为的概率，当交互行为在转化链路中存在前向行为时，交互行为对应的行为发生可能度，用于表征在用户对象已对目标资源对象发生了前向行为的情况下，用户对象对目标资源对象发生该交互行为的概率。行为发生可能度越大，则发生该交互行为的概率也越大。交互行为的前向行为是指转化链路中排列在该交互行为之前的交互行为。例如，转化链路为“来访→点击→申购”，则对于“申购”，前向行为包括“来访”以及“点击”。交互行为对应的行为发生可能度，可以用于表征用户对象在当前时间下针对目标资源对象发生该交互行为的可能度。
具体地，服务器可以获取已训练的对象转化预测模型，对象转化预测模型用于预测转化可能度。对象转化预测模型中可以包括转化链路中各个交互行为分别对应的行为预测网络，交互行为对应的行为预测网络，用于预测该交互行为对应的行为发生可能度。服务器可以将 转化预测特征分别输入到每个行为预测网络中，得到各个交互行为分别对应的行为发生可能度。以转化链路为“来访→点击→申购”为例，如图4所示，展示了对象转化预测模型，对象转化预测模型中包括3个行为预测网络，第一行为预测网络是“来访”对应的行为预测网络，预测得到发生“来访”的概率，第二行为预测网络是“点击”对应的行为预测网络，预测得到在已发生“来访”的情况下发生“点击”的概率，第三行为预测网络是“申购”对应的行为预测网络，预测得到在已发生“来访”以及“点击”的情况下发生“申购”的概率。
在一些实施例中，转化可能度与行为发生可能度成正相关关系。服务器可以将各个交互行为分别对应的行为发生可能度进行相乘运算，将相乘运算的结果作为用户对象在当前时间针对目标资源对象的转化可能度。如图4所示，第一行为预测网络输出第一概率，第二行为预测网络输出第二概率，第三行为预测网络输出第三概率，将第一概率与第二概率相乘得到第四概率，将第四概率与第三概率相乘得到第五概率，将第五概率作为转化可能度。以转化链路为“来访→点击→申购”为例，则第一概率为用户对象发生“来访”的概率，第二概率为用户对象已发生“来访”的情况下发生“点击”的概率，第三概率为用户对象已发生“来访”以及“点击”的情况下发生“申购”的概率。将这3个概率相乘即可以得到用户发生转化的概率即用户发生申购的概率。
其中，正相关关系指的是：在其他条件不变的情况下，两个变量变动方向相同，一个变量由大到小变化时，另一个变量也由大到小变化。可以理解的是，这里的正相关关系是指变化的方向是一致的，但并不是要求当一个变量有一点变化，另一个变量就必须也变化。例如，可以设置当变量a为10至20时，变量b为100，当变量a为20至30时，变量b为120。这样，a与b的变化方向都是当a变大时，b也变大。但在a为10至20的范围内时，b可以是没有变化的。
本实施例中，基于各个行为发生可能度，得到用户对象在当前时间针对目标资源对象的转化可能度，由于转化可能度与行为发生可能度成正相关关系，提高了转化可能度的准确度。
在一些实施例中，基于转化预测特征，预测用户对象针对目标资源对象发生交互行为的可能度，得到交互行为对应的行为发生可能度包括：从转化链路中，获取交互行为的前向行为；基于转化预测特征，预测用户对象在已发生前向行为时，用户对象针对目标资源对象发生交互行为的可能度，得到交互行为对应的行为发生可能度。
其中，“在已发生前向行为时”即上述的“在用户对象已对目标资源对象发生了前向行为的情况下”。
本实施例中，基于转化预测特征，预测用户对象在已发生前向行为时，用户对象针对目标资源对象发生交互行为的可能度，得到交互行为对应的行为发生可能度，提高了得到行为发生可能度的效率以及准确度。
在一些实施例中，基于转化预测特征，预测用户对象针对目标资源对象发生交互行为的可能度，得到交互行为对应的行为发生可能度包括：获取已训练的对象转化预测模型；对象转化预测模型包括转化链路中各个交互行为分别对应的行为预测网络；交互行为对应的行为预测网络用于预测交互行为对应的行为发生可能度；将转化预测特征分别输入到每个交互行为对应的行为预测网络中，预测得到各个交互行为分别对应的行为发生可能度。
具体地，得到已训练的对象转化预测模型的过程可以包括：获取样本用户对象集合，样本用户对象集合中包括多个样本用户对象，样本用户对象是用于训练对象转化预测模型时所使用的用户对象。确定样本用户对象集合中各个样本用户对象在预设时长内针对目标资源对 象所产生的交互行为。预设时长例如可以是向样本用户对象集合中的样本用户对象进行投放营销后的一段时间，例如投放后的1个月或3个月等。投放后样本用户对象有可能会产生转化链路中的行为，例如用户在发生转化的过程中，有可能发生来访、点击以及转化三种行为，全场景的样本空间如图6所示。
根据已产生的交互行为确定各个样本对象分别对应的样本标签，每个样本用户对象的样本标签有多个，例如样本标签的数量与转化链路中的交互行为的数量相同，以转化链路为“来访→点击→申购(即转化)”进行说明，包括3个样本标签，例如为F1、F2以及F3，F1表示样本用户对象发生“来访”的概率，F2表示样本用户对象发生“来访”以及“点击”的概率，F3表示样本用户对象发生“来访”、“点击”以及“申购”的概率。在确定样本标签时，根据样本用户对象在该预设时长内的交互行为确定，以V∈{0,1}代表“来访”，V＝1代表用户来访，V＝0代表用户没有来访，以Y∈{0,1}代表“点击”，Y＝1代表用户点击，Y＝0代表用户没有点击，以Z∈{0,1}代表“转化(申购)”，Z＝1代表用户转化，Z＝0代表用户没有转化。若样本用户对象A为“V＝1，Y＝0，Z＝0”时，则说明样本用户对象A只发生了“访问”，则确定样本用户对象A的样本标签为“F1＝1，F2＝0，F3＝0”，若样本用户对象B为“V＝1，Y＝1，Z＝0”，则说明样本用户对象B发生了“访问”以及“点击”，则确定样本用户对象B的样本标签为“F1＝1，F2＝1，F3＝0”，若样本用户对象C为“V＝1，Y＝1，Z＝1”，则说明样本用户对象C发生了“访问”、“点击”以及“申购”，则确定样本用户对象C的样本标签为“F1＝1，F2＝1，F3＝1”。
当确定了样本用户对象集合中的各个样本用户对象分别对应的样本标签后，获取样本用户对象对应的转化预测特征，转化预测特征可以是利用特征生成模型生成的。将样本用户对象的转化预测特征分别输入到各个行为预测网络的，例如，输入到“访问”对应的行为预测网络、“点击”对应的行为预测网络、“申购”对应的行为预测网络中，得到样本用户对象发生“访问”的第一预测概率即pvisit＝p(V＝1|X)、在发生“访问”的前提下发生“点击”的第二预测概率即pctr＝p(Y＝1|V＝1，X)，在发生“访问”以及“点击”的前提下，发生“申购”的第三预测概率即pcvr＝p(Z＝1|Y＝1，V＝1，X)，其中，X用于代表一个样本用户对象，例如可以是样本用户对象的对象特征。将第一预测概率与第二预测概率相乘，得到第四预测概率pvisit-ctr＝p(Y&V＝1|X)＝p(Y＝1|V＝1，X)*p(V＝1|X)，第四预测概率代表样本用户对象发生“点击”以及“访问”的概率，将第三预测概率与第四预测概率相乘，得到第五预测概率pvisit-ctcvr＝p(Z&Y&V＝1|X)＝p(Z＝1|Y＝1,V＝1,X)*p(Y&V＝1|X)。基于第一预测概率、第四预测概率、第五预测概率以及样本标签生成对象转化预测模型对应的模型损失值(loss)，利用模型损失值调整对象转化预测模型中的各个行为预测网络，直到模型收敛，得到已训练的对象转化预测模型。第一预测概率对应的样本标签为F1，第四预测概率对应的样本标签为F2，第五预测概率对应的样本标签为F3。模型损失值L可以用下面的公式表示：
Figure PCTCN2022125251-appb-000004
其中，公式中的V表示样本标签为F1，V&Y表示样本标签为F2，V&Y&Z表示样本标签为F3，n表示一次训练所使用的样本用户对象的数量。其中，pvisit代表来访率，例如代表营销投放后的来访率，来访率用于表征来访的人数与投放的人数的比值，pctr代表点击率，点击率用于表征点击的人数与来访的人数的比值，pcvr代表转化率，用于表征转化的人数与点击的人数的比值。
由于在对对象转化预测模型时，样本用户对象可以是处于任意转化阶段的用户，例如可以是没有“访问”的用户、仅仅“访问”的用户，“访问”以及“点击”的用户、“访问”、“点击” 以及“申购”的用户，即训练时采用的是全量的用户，该训练方法对于训练样本的要求较低，且样本的丰富程度较高，提高了模型训练的准确度，并且能够适用于特征的样本的数量较少的场景中，例如可以应用于金融场景中，提高金融场景中预测转化率的准确度。由于金融场景下用户决策成本高，行为相比广告、推荐等场景更加低频，深度转化样本更加稀少，且不同行为往往存在关联性和递进性，如对于理财产品站内广告投放，用户首先需要有来访行为，来访后如果对投放内容感兴趣则会发生点击行为，如果点击后对产品认可则会发生申购行为，投放、来访、点击、申购这些行为之间层层递进，但最末端的申购行为往往非常稀疏，对建模产生较大的挑战。如果仅仅采用申购了的用户进行样本训练，由于样本稀疏会导致模型训练的精度较低。而本申请实施例中的训练方法以及对象转化预测模型应用于金融场景中，可以将金融场景中的任何的用户都作为样本，提高了样本数量以及丰富程度，从而提高了模型训练的精确度。
本实施例中，通过对象转化预测模型中的各个行为预测网络，得到各个交互行为分别对应的行为发生可能度，提高了得到行为发生可能度的效率以及准确度。
本申请还提供一种应用场景，该应用场景应用上述的对象处理方法。具体地，该应用场景为金融场景，资源对象为金融场景中资源对象例如可以是基金，如图7所示，该对象处理方法在该应用场景的应用如下：
步骤702，接收终端发送的针对目标资源对象的推送请求，推送请求中携带目标资源对象的标识，响应于推送请求，获取用户对象集合；
步骤702，对于用户对象集合中的每个用户对象，获取预设时间范围内，用户对象的历史交互特征序列，历史交互特征序列中的各个历史交互特征对应有交互时刻，历史交互特征序列中的各个历史交互特征按照交互时刻排列，交互时刻为预设时间范围内的时刻；
其中，交互时刻越靠前，则历史交互特征在历史交互特征序列中的排序越靠前。各个历史交互特征对应有资源对象，不同的历史交互特征对应的资源对象可以相同也可以不同。
步骤704，对于每个交互时刻，获取该交互时刻下的时间状况特征，获取该交互时刻下的资源状况特征，该资源状况特征用于表征该交互时刻的历史交互特征对应的资源对象的资源因素在该交互时刻下的状况；
其中，资源对象不同，资源的资源因素可以相同也可以不同。例如，一个用户的历史交互特征为item1,item2,item3，对应的交互时间为T 1,T 2,T 3，则各个交互时刻下的时间状况特征分别为T 1,T 2,T 3时刻对应的时间的特征，各个交互时刻下的资源状况特征分别为T 1,T 2,T 3时刻对应的资源因素例如时长行情的特征。
为了确定数据不泄露以及训练和预测时取值逻辑一致，资源状况特征例如市场行情的特征可以根据实际情况进行适当的时间偏移，例如，使用的资源状况特征是T日收盘价变化量，而预测当日还未收盘，无法获取T日收盘价，则训练时可以使用T-1日的收盘价(即向前偏移一天)，这样保证预测时可以有同样逻辑的数据(T-1日收盘价)。
步骤706，获取已训练的特征生成模型，特征生成模型中包括多个特征处理网络，确定各个交互时刻分别对应的特征处理网络，对于每个交互时刻，将该交互时刻下的历史交互特征、时间状况特征以及资源状况特征，输入到该交互时刻对应的特征处理网络中，得到各个交互时刻下的关注状况特征；
步骤708，特征生成网络中还包括权重预测特征生成网络，获取用户对象的对象编码特征，获取当前时间下的时间状况特征以及当前时间下的资源状况特征，将对象编码特征、当 前时间下的时间状况特征、以及当前时间下的资源状况特征，输入到权重预测特征生成网络中，预测得到权重预测特征，基于权重预测特征确定各个交互时刻下的关注状况特征分别对应的权重，基于得到的权重对各个关注状况特征进行加权计算，得到当前时间下的关注程度特征。
其中，当前时间下的资源状况特征用于表征目标资源对象的资源因素在当前时间下的状况；
步骤710，特征生成网络中还可以包括对象特征提取网络，将对象编码特征输入到对象特征提取网络中，得到对象提取特征，将对象提取特征与关注程度特征进行拼接，得到当前时间下的转化预测特征。
步骤712，获取已训练的对象转化预测模型，对象转化预测模型中包括目标资源对象对应的转化链路中的各个交互行为分别对应的行为预测网络，将转化预测特征分别输入到各个交互行为对应的行为预测网络中，预测得到各个交互行为分别对应的行为发生可能度，将各个行为发生可能度进行相乘，得到用户对象在当前时间下针对目标资源对象的转化可能度。
步骤714，基于用户对象集合中各个用户对象分别对应的转化可能度，从用户对象集合中筛选得到目标用户对象，向目标用户对象推送目标资源对象或推送与目标资源对象相关联的内容。
其中，服务器可以将转化可能度与可能度阈值进行对比，当确定转化可能度大于可能度阈值时，将用户对象作为目标用户对象。服务器可以按照转化可能度对用户对象集合中的各个用户对象进行排列，例如，按照转化可能度从大到小的顺序对用户对象进行排列，得到用户对象序列，转化可能度越大，用户对象在用户对象序列中的排序越靠前，服务器可以从用户对象序列中获取排序在排序阈值之前的用户对象，作为目标用户对象。可能度阈值以及排序阈值可以预设或根据需要设置。与目标资源对象相关联的内容可以是用于激励用户购买目标资源对象的内容，例如，当目标资源对象为基金时，可以是基金的优惠券等。
本实施例中，利用时间状况特征以及资源状况特征，预测得到转化预测特征，提高了转化预测特征的真实可靠性，利用对象转化预测模型中的各个行为预测网络，确定转化可能度，提高了计算转化可能度的效率以及准确度。
本申请实施例中，将对象转化预测模型与特征生成模型结合使用，由于对象转化预测模型可以是利用全量的样本进行训练得到的，并且对象转化预测模型中包括多个行为预测网络，故实现了全空间多任务学习模型(ESMM，Entire Space Multi-Task Model)，而由于特征生成模型可以是改进长短期记忆神经网络，因此，两个模型的结合实现了一种融合改进长短期记忆和全空间多任务学习的神经网络。
通过实验证明了本申请提供的对象处理方法，应用于金融场景中可以取得较好的效果。如表2所示，展示了本申请提供的对象处理方法在金融场景下所取得的效果。表2中，本实验中，经过统计分布分析，历史交互特征序列由最近30天的点击、申购、搜索的16只基金产品构建，并将每个基金产品的特征(如收益率、封闭期、最大回撤等)作为该基金的表示，如果用户的历史交互特征序列的长度短于16，则使用全0特征补足，如果超过16，则按照时间先后顺序截断为16。
表2线上投放效果
模型名称 转化AUC 相对提升 千次投放转化人数 相对提升
DNN 0.9011 - 10.57 -
ESMM 0.9149 1.53％ 10.82 2.37％
LSTM 0.9252 2.67％ 11.57 9.46％
FLSTM 0.9301 3.22％ 13.45 27.2％
MFLSTM 0.9382 4.12％ 15.53 46.9％
表2中，千次投放转化人数＝pvisit-ctcvr*1000，DNN模型为只使用投放后是否转化为建模目标的普通全连接神经网络，ESMM模型为引入点击、转化两个目标的ESMM结构神经网络，LSTM代表在ESMM模型基础上增加LSTM单元建模用户历史兴趣，FLSTM模型指的是将LSTM单元更换为FLSTM单元，MFLSTM(multi-task Financial Long short-term memory)为考虑了来访-点击-转化三目标，并引入FLSTM提取用户历史兴趣的模型。可以看出，FLSTM相较原始的LSTM结构，可以提升16.2％，继续引入来访目标全空间建模后，相对FLSTM可以提升15.5％，最终相对基线DNN模型取得了46.9％的线上提升效果。
本申请的特征生成模型还可以是基于Transformer模型的，例如可以引入长时间的用户行为序列(如近半年)，并针对金融场景的特殊性对其结构进行优化，如将金融场景下时间、市场行情等因素表征为embedding，将其直接与序列元素表征特征进行拼接，或作为类似position embedding的结构与序列元素表征融合后加入Transformer结构中从而预测出用户的转化预测特征。
应该理解的是，虽然上述各实施例的流程图中的各个步骤按照箭头的指示依次显示，但是这些步骤并不是必然按照箭头指示的顺序依次执行。除非本文中有明确的说明，这些步骤的执行并没有严格的顺序限制，这些步骤可以以其它的顺序执行。而且，上述各实施例中的至少一部分步骤可以包括多个步骤或者多个阶段，这些步骤或者阶段并不必然是在同一时间执行完成，而是可以在不同的时间执行，这些步骤或者阶段的执行顺序也不必然是依次进行，而是可以与其它步骤或者其它步骤中的步骤或者阶段的至少一部分轮流或者交替地执行。
在一些实施例中，如图8所示，提供了一种对象处理装置，该装置可以采用软件模块或硬件模块，或者是二者的结合成为计算机设备的一部分，该装置具体包括：交互特征获取模块802、状况特征获取模块804、预测特征确定模块806和可能度预测模块808，其中：
交互特征获取模块802，用于获取用户对象针对历史资源对象的历史交互特征。
状况特征获取模块804，用于获取历史资源对象的动态影响因素的历史状况特征；动态影响因素，用于动态影响历史资源对象的资源属性的变化；历史状况特征，是基于动态影响因素的历史状况信息确定的。
预测特征确定模块806，用于基于历史交互特征以及历史状况特征，确定用户对象在当前时间针对目标资源对象的转化预测特征。
可能度预测模块808，用于基于转化预测特征，预测用户对象在当前时间针对目标资源对象的转化可能度，以基于转化可能度确定针对目标资源对象对用户对象的处理方式。
在一些实施例中，动态影响因素包括资源因素或时间因素中的至少一个；资源因素是在资源场景下动态变化的资源因素；状况特征获取模块还用于：确定产生历史交互特征时的交互时刻，基于时间因素在交互时刻时的时间信息确定时间因素对应的时间状况特征；获取资源因素在交互时刻时的资源信息，基于资源信息确定资源因素对应的资源状况特征；及，基 于时间状况特征或资源状况特征中的至少一个，确定历史状况特征。
在一些实施例中，预测特征确定模块还用于：基于历史交互特征以及历史状况特征，确定用户对象在当前时间针对目标资源对象的关注程度特征；及，基于关注程度特征，确定用户对象在当前时间针对目标资源对象的转化预测特征。
在一些实施例中，预测特征确定模块还用于：对于每一个历史交互特征对应的交互时刻，确定交互时刻的在先时刻；历史交互特征对应的交互时刻，是产生用于得到历史交互特征的历史交互数据的时刻；获取用户对象在在先时刻下的关注状况特征，得到在先关注状况特征；在先关注状况特征用于表征用户对象在在先时刻对目标资源对象的关注情况；基于在先关注状况特征以及交互时刻下的历史交互特征，得到交互时刻下的增量特征；增量特征为历史交互特征相较于在先关注状况特征所增加的特征；基于交互时刻下的历史交互特征以及历史状况特征，对增量特征进行处理，得到交互时刻下的关注状况特征；交互时刻下的历史状况特征，是历史资源对象的动态影响因素在交互时刻下的状况特征；及，基于各个交互时刻下的关注状况特征，确定用户对象在当前时间针对目标资源对象的关注程度特征。
在一些实施例中，预测特征确定模块还用于：获取用户对象在在先时刻下的聚合特征，得到在先聚合特征；基于历史交互特征以及历史状况特征，确定增量特征对应的增量权重；确定在先聚合特征对应的聚合权重，基于增量权重以及聚合权重，对增量特征以及在先聚合特征进行加权计算，得到交互时刻下的聚合特征；及，基于交互时刻下的聚合特征确定交互时刻下的关注状况特征。
在一些实施例中，历史状况特征包括交互时刻下的时间状况特征、或交互时刻下的资源状况特征中的至少一个，预测特征确定模块还用于：基于交互时刻下的历史交互特征以及交互时刻下的时间状况特征，得到增量特征对应的第一权重；基于交互时刻下的历史交互特征以及交互时刻下的资源状况特征，得到增量特征对应的第二权重；及，基于第一权重或第二权重中的至少一个，确定增量特征对应的增量权重。
在一些实施例中，关注状况特征是将历史交互特征以及历史状况特征输入至交互时刻对应的特征处理网络中生成的；特征处理网络中包括增量权重预测网络；及，预测特征确定模块还用于：将历史交互特征以及历史状况特征输入到增量权重预测网络中，预测得到增量特征对应的增量权重。
在一些实施例中，特征处理网络中还包括聚合权重预测网络，预测特征确定模块还用于：将在先关注状况特征以及交互时刻下的历史状况特征，输入到聚合权重预测网络中，预测得到在先聚合特征对应的聚合权重。
在一些实施例中，预测特征确定模块还用于：获取用户对象的对象特征以及当前时间目标资源对象的动态影响因素的当前状况特征；基于对象特征以及当前状况特征，确定各个交互时刻下的关注状况特征分别对应的权重；及，利用各个关注状况特征分别对应的权重，对各个关注状况特征进行加权计算，确定用户对象在当前时间针对目标资源对象的关注程度特征。
在一些实施例中，预测特征确定模块还用于：获取用户对象的对象信息；对对象信息进行编码，得到用户对象的对象编码特征；及，基于用户对象的对象编码特征以及关注程度特征，得到用户对象在当前时间针对目标资源对象的转化预测特征。
在一些实施例中，可能度预测模块还用于：获取目标资源对象对应的转化链路；转化链路包括用户对象针对目标资源对象发生转化的过程中所需要发生的交互行为；对于转化链路 中的每个交互行为，基于转化预测特征，预测用户对象针对目标资源对象发生交互行为的可能度，得到交互行为对应的行为发生可能度；及，基于各个行为发生可能度，得到用户对象在当前时间针对目标资源对象的转化可能度；转化可能度与行为发生可能度成正相关关系。
在一些实施例中，可能度预测模块还用于：从转化链路中，获取交互行为的前向行为；基于转化预测特征，预测用户对象在已发生前向行为时，用户对象针对目标资源对象发生交互行为的可能度，得到交互行为对应的行为发生可能度。
在一些实施例中，可能度预测模块还用于：获取已训练的对象转化预测模型；对象转化预测模型包括转化链路中各个交互行为分别对应的行为预测网络；交互行为对应的行为预测网络用于预测交互行为对应的行为发生可能度；及，将转化预测特征分别输入到每个交互行为对应的行为预测网络中，预测得到各个交互行为分别对应的行为发生可能度。
关于对象处理装置的具体限定可以参见上文中对于对象处理方法的限定，在此不再赘述。上述对象处理装置中的各个模块可全部或部分通过软件、硬件及其组合来实现。上述各模块可以硬件形式内嵌于或独立于计算机设备中的处理器中，也可以以软件形式存储于计算机设备中的存储器中，以便于处理器调用执行以上各个模块对应的操作。
在一些实施例中，提供了一种计算机设备，该计算机设备可以是终端，其内部结构图可以如图9所示。该计算机设备包括通过系统总线连接的处理器、存储器、通信接口、显示屏和输入装置。其中，该计算机设备的处理器用于提供计算和控制能力。该计算机设备的存储器包括非易失性存储介质、内存储器。该非易失性存储介质存储有操作系统和计算机可读指令。该内存储器为非易失性存储介质中的操作系统和计算机可读指令的运行提供环境。该计算机设备的通信接口用于与外部的终端进行有线或无线方式的通信，无线方式可通过WIFI、运营商网络、NFC(近场通信)或其他技术实现。该计算机可读指令被处理器执行时以实现一种对象处理方法。该计算机设备的显示屏可以是液晶显示屏或者电子墨水显示屏，该计算机设备的输入装置可以是显示屏上覆盖的触摸层，也可以是计算机设备外壳上设置的按键、轨迹球或触控板，还可以是外接的键盘、触控板或鼠标等。
在一些实施例中，提供了一种计算机设备，该计算机设备可以是服务器，其内部结构图可以如图10所示。该计算机设备包括通过系统总线连接的处理器、存储器和网络接口。其中，该计算机设备的处理器用于提供计算和控制能力。该计算机设备的存储器包括非易失性存储介质、内存储器。该非易失性存储介质存储有操作系统、计算机可读指令和数据库。该内存储器为非易失性存储介质中的操作系统和计算机可读指令的运行提供环境。该计算机设备的数据库用于存储对象处理方法中涉及到的数据。该计算机设备的网络接口用于与外部的终端通过网络连接通信。该计算机可读指令被处理器执行时以实现一种对象处理方法。
本领域技术人员可以理解，图9和图10中示出的结构，仅仅是与本申请方案相关的部分结构的框图，并不构成对本申请方案所应用于其上的计算机设备的限定，具体的计算机设备可以包括比图中所示更多或更少的部件，或者组合某些部件，或者具有不同的部件布置。
在一些实施例中，还提供了一种计算机设备，包括存储器和一个或多个处理器，存储器中存储有计算机可读指令，计算机可读指令被该处理器执行时，使得一个或多个处理器执行上述各方法实施例中的步骤。
在一些实施例中，提供了一个或多个非易失性可读存储介质，存储有计算机可读指令，该计算机可读指令被一个或多个处理器执行时，使得一个或多个处理器实现上述各方法实施例中的步骤。
一种计算机程序产品，包括计算机可读指令，所述计算机可读指令被处理器执行时实现上述对象处理方法的步骤。
本领域普通技术人员可以理解实现上述实施例方法中的全部或部分流程，是可以通过计算机可读指令来指令相关的硬件来完成，所述的计算机可读指令可存储于一非易失性计算机可读取存储介质中，该计算机可读指令在执行时，可包括如上述各方法的实施例的流程。其中，本申请所提供的各实施例中所使用的对存储器、存储、数据库或其它介质的任何引用，均可包括非易失性和易失性存储器中的至少一种。非易失性存储器可包括只读存储器(Read-Only Memory，ROM)、磁带、软盘、闪存或光存储器等。易失性存储器可包括随机存取存储器(Random Access Memory，RAM)或外部高速缓冲存储器。作为说明而非局限，RAM可以是多种形式，比如静态随机存取存储器(Static Random Access Memory，SRAM)或动态随机存取存储器(Dynamic Random Access Memory，DRAM)等。
以上实施例的各技术特征可以进行任意的组合，为使描述简洁，未对上述实施例中的各个技术特征所有可能的组合都进行描述，然而，只要这些技术特征的组合不存在矛盾，都应当认为是本说明书记载的范围。
以上所述实施例仅表达了本申请的几种实施方式，其描述较为具体和详细，但并不能因此而理解为对发明专利范围的限制。应当指出的是，对于本领域的普通技术人员来说，在不脱离本申请构思的前提下，还可以做出若干变形和改进，这些都属于本申请的保护范围。因此，本申请专利的保护范围应以所附权利要求为准。

Claims (17)

  1. 一种对象处理方法，其特征在于，由计算机设备执行，所述方法包括：
    获取用户对象针对历史资源对象的历史交互特征；
    获取所述历史资源对象的动态影响因素的历史状况特征；所述动态影响因素，用于动态影响所述历史资源对象的资源属性的变化；历史状况特征，是基于所述动态影响因素的历史状况信息确定的；
    基于所述历史交互特征以及所述历史状况特征，确定所述用户对象在当前时间针对目标资源对象的转化预测特征；及
    基于所述转化预测特征，预测所述用户对象在当前时间针对所述目标资源对象的转化可能度，以基于所述转化可能度确定针对所述目标资源对象对所述用户对象的处理方式。
  2. 根据权利要求1所述的方法，其特征在于，所述动态影响因素包括资源因素或时间因素中的至少一个；所述资源因素是在资源场景下动态变化的资源因素；
    所述获取所述历史资源对象的动态影响因素的历史状况特征包括：
    确定产生所述历史交互特征时的交互时刻，基于所述时间因素在所述交互时刻时的时间信息确定所述时间因素对应的时间状况特征；
    获取所述资源因素在所述交互时刻时的资源信息，基于所述资源信息确定所述资源因素对应的资源状况特征；及
    基于所述时间状况特征或所述资源状况特征中的至少一个，确定所述历史状况特征。
  3. 根据权利要求1所述的方法，其特征在于，所述基于所述历史交互特征以及所述历史状况特征，确定所述用户对象在当前时间针对目标资源对象的转化预测特征包括：
    基于所述历史交互特征以及所述历史状况特征，确定所述用户对象在当前时间针对目标资源对象的关注程度特征；及
    基于所述关注程度特征，确定所述用户对象在当前时间针对所述目标资源对象的转化预测特征。
  4. 根据权利要求3所述的方法，其特征在于，所述历史交互特征为多个，所述基于所述历史交互特征以及所述历史状况特征，确定所述用户对象在当前时间针对目标资源对象的关注程度特征包括：
    对于每一个所述历史交互特征对应的交互时刻，确定所述交互时刻的在先时刻；所述历史交互特征对应的交互时刻，是产生用于得到所述历史交互特征的历史交互数据的时刻；
    获取所述用户对象在所述在先时刻下的关注状况特征，得到在先关注状况特征；所述在先关注状况特征用于表征所述用户对象在所述在先时刻对所述目标资源对象的关注情况；
    基于所述在先关注状况特征以及所述交互时刻下的历史交互特征，得到所述交互时刻下的增量特征；所述增量特征为所述历史交互特征相较于所述在先关注状况特征所增加的特征；
    基于所述交互时刻下的历史交互特征以及历史状况特征，对所述增量特征进行处理，得到所述交互时刻下的关注状况特征；所述交互时刻下的历史状况特征，是所述历史资源对象的动态影响因素在所述交互时刻下的状况特征；及
    基于各个所述交互时刻下的关注状况特征，确定所述用户对象在当前时间针对目标资源对象的关注程度特征。
  5. 根据权利要求4所述的方法，其特征在于，所述基于所述交互时刻下的历史交互特征以及历史状况特征，对所述增量特征进行处理，得到所述交互时刻下的关注状况特征包括：
    获取所述用户对象在所述在先时刻下的聚合特征，得到在先聚合特征；基于所述历史交互特征以及所述历史状况特征，确定所述增量特征对应的增量权重；
    确定所述在先聚合特征对应的聚合权重，基于所述增量权重以及所述聚合权重，对所述 增量特征以及所述在先聚合特征进行加权计算，得到所述交互时刻下的聚合特征；及
    基于所述交互时刻下的聚合特征确定所述交互时刻下的关注状况特征。
  6. 根据权利要求5所述的方法，其特征在于，所述历史状况特征包括所述交互时刻下的时间状况特征、或所述交互时刻下的资源状况特征中的至少一个，所述基于所述历史交互特征以及所述历史状况特征，确定所述增量特征对应的增量权重包括：
    基于所述交互时刻下的历史交互特征以及所述交互时刻下的时间状况特征，得到所述增量特征对应的第一权重；
    基于所述交互时刻下的历史交互特征以及所述交互时刻下的资源状况特征，得到所述增量特征对应的第二权重；及
    基于所述第一权重或所述第二权重中的至少一个，确定所述增量特征对应的增量权重。
  7. 根据权利要求5所述的方法，其特征在于，所述关注状况特征是将所述历史交互特征以及所述历史状况特征输入至所述交互时刻对应的特征处理网络中生成的；所述特征处理网络中包括增量权重预测网络；及
    所述基于所述历史交互特征以及所述历史状况特征，确定所述增量特征对应的增量权重包括：
    将所述历史交互特征以及所述历史状况特征输入到所述增量权重预测网络中，预测得到所述增量特征对应的增量权重。
  8. 根据权利要求7所述的方法，其特征在于，所述特征处理网络中还包括聚合权重预测网络，所述确定所述在先聚合特征对应的聚合权重包括：
    将所述在先关注状况特征以及所述交互时刻下的历史状况特征，输入到所述聚合权重预测网络中，预测得到所述在先聚合特征对应的聚合权重。
  9. 根据权利要求4所述的方法，其特征在于，所述基于各个所述交互时刻下的关注状况特征，确定所述用户对象在当前时间针对目标资源对象的关注程度特征包括：
    获取所述用户对象的对象特征以及当前时间所述目标资源对象的动态影响因素的当前状况特征；
    基于所述对象特征以及所述当前状况特征，确定各个所述交互时刻下的关注状况特征分别对应的权重；及
    利用各个所述关注状况特征分别对应的权重，对各个所述关注状况特征进行加权计算，确定所述用户对象在当前时间针对所述目标资源对象的关注程度特征。
  10. 根据权利要求3所述的方法，其特征在于，所述基于所述关注程度特征，确定所述用户对象在当前时间针对所述目标资源对象的转化预测特征包括：
    获取所述用户对象的对象信息；
    对所述对象信息进行编码，得到所述用户对象的对象编码特征；及
    基于所述用户对象的对象编码特征以及所述关注程度特征，得到所述用户对象在当前时间针对所述目标资源对象的转化预测特征。
  11. 根据权利要求1至10任意一项所述的方法，其特征在于，所述基于所述转化预测特征，预测所述用户对象在当前时间针对所述目标资源对象的转化可能度包括：
    获取所述目标资源对象对应的转化链路；所述转化链路包括所述用户对象针对所述目标资源对象发生转化的过程中所需要发生的交互行为；
    对于所述转化链路中的每个交互行为，基于所述转化预测特征，预测所述用户对象针对所述目标资源对象发生所述交互行为的可能度，得到所述交互行为对应的行为发生可能度；及
    基于各个所述行为发生可能度，得到所述用户对象在当前时间针对所述目标资源对象的转化可能度；所述转化可能度与所述行为发生可能度成正相关关系。
  12. 根据权利要求11所述的方法，其特征在于，所述基于所述转化预测特征，预测所 述用户对象针对所述目标资源对象发生所述交互行为的可能度，得到所述交互行为对应的行为发生可能度包括：
    从所述转化链路中，获取所述交互行为的前向行为；及
    基于所述转化预测特征，预测所述用户对象在已发生所述前向行为时，所述用户对象针对所述目标资源对象发生所述交互行为的可能度，得到所述交互行为对应的行为发生可能度。
  13. 根据权利要求11所述的方法，其特征在于，所述基于所述转化预测特征，预测所述用户对象针对所述目标资源对象发生所述交互行为的可能度，得到所述交互行为对应的行为发生可能度包括：
    获取已训练的对象转化预测模型；所述对象转化预测模型包括所述转化链路中各个交互行为分别对应的行为预测网络；交互行为对应的行为预测网络用于预测所述交互行为对应的行为发生可能度；及
    将所述转化预测特征分别输入到每个所述交互行为对应的行为预测网络中，预测得到各个所述交互行为分别对应的行为发生可能度。
  14. 一种对象处理装置，其特征在于，所述装置包括：
    交互特征获取模块，用于获取用户对象针对历史资源对象的历史交互特征；
    状况特征获取模块，用于获取所述历史资源对象的动态影响因素的历史状况特征；所述动态影响因素，用于动态影响所述历史资源对象的资源属性的变化；历史状况特征，是基于所述动态影响因素的历史状况信息确定的；
    预测特征确定模块，用于基于所述历史交互特征以及所述历史状况特征，确定所述用户对象在当前时间针对目标资源对象的转化预测特征；及
    可能度预测模块，用于基于所述转化预测特征，预测所述用户对象在当前时间针对所述目标资源对象的转化可能度，以基于所述转化可能度确定针对所述目标资源对象对所述用户对象的处理方式。
  15. 一种计算机设备，包括存储器和一个或多个处理器，所述存储器存储有计算机可读指令，其特征在于，所述计算机可读指令被所述处理器执行时，使得所述一个或多个处理器执行权利要求1至13中任一项所述的方法。
  16. 一个或多个非易失性可读存储介质，存储有计算机可读指令，其特征在于，所述计算机可读指令被一个或多个处理器执行时，使得所述一个或多个处理器实现如权利要求1至13中任一项所述的方法。
  17. 一种计算机程序产品，包括计算机可读指令，其特征在于，所述计算机可读指令被处理器执行时实现权利要求1至13中任一项所述的方法。
PCT/CN2022/125251 2021-12-08 2022-10-14 对象处理方法、装置、计算机设备和存储介质 WO2023103584A1 (zh)

Priority Applications (1)

Application Number Priority Date Filing Date Title
US18/215,303 US20230342797A1 (en) 2021-12-08 2023-06-28 Object processing method based on time and value factors

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN202111492149.2 2021-12-08
CN202111492149.2A CN114330837A (zh) 2021-12-08 2021-12-08 对象处理方法、装置、计算机设备和存储介质

Related Child Applications (1)

Application Number Title Priority Date Filing Date
US18/215,303 Continuation US20230342797A1 (en) 2021-12-08 2023-06-28 Object processing method based on time and value factors

Publications (1)

Publication Number Publication Date
WO2023103584A1 true WO2023103584A1 (zh) 2023-06-15

Family

ID=81051267

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2022/125251 WO2023103584A1 (zh) 2021-12-08 2022-10-14 对象处理方法、装置、计算机设备和存储介质

Country Status (3)

Country Link
US (1) US20230342797A1 (zh)
CN (1) CN114330837A (zh)
WO (1) WO2023103584A1 (zh)

Families Citing this family (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114330837A (zh) * 2021-12-08 2022-04-12 腾讯科技（深圳）有限公司 对象处理方法、装置、计算机设备和存储介质
CN114580794B (zh) * 2022-05-05 2022-07-22 腾讯科技（深圳）有限公司 数据处理方法、装置、程序产品、计算机设备和介质

Citations (8)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US8468110B1 (en) * 2010-07-22 2013-06-18 Intuit Inc. Real-time user behavior prediction
US20190043115A1 (en) * 2017-08-02 2019-02-07 Thomas Purves Machine learning tool
CN109447687A (zh) * 2018-09-27 2019-03-08 中国平安人寿保险股份有限公司 基于大数据分析的用户挖掘方法及装置
CN110110203A (zh) * 2018-01-11 2019-08-09 腾讯科技（深圳）有限公司 资源信息推送方法及服务器、资源信息展示方法及终端
US20190340684A1 (en) * 2017-03-10 2019-11-07 Cerebri AI Inc. Monitoring and controlling continuous stochastic processes based on events in time series data
CN110476176A (zh) * 2017-03-20 2019-11-19 微软技术许可有限责任公司 用户目标辅助技术
CN113222732A (zh) * 2021-05-17 2021-08-06 中国农业银行股份有限公司 信息处理方法、装置、设备及存储介质
CN114330837A (zh) * 2021-12-08 2022-04-12 腾讯科技（深圳）有限公司 对象处理方法、装置、计算机设备和存储介质

Patent Citations (8)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US8468110B1 (en) * 2010-07-22 2013-06-18 Intuit Inc. Real-time user behavior prediction
US20190340684A1 (en) * 2017-03-10 2019-11-07 Cerebri AI Inc. Monitoring and controlling continuous stochastic processes based on events in time series data
CN110476176A (zh) * 2017-03-20 2019-11-19 微软技术许可有限责任公司 用户目标辅助技术
US20190043115A1 (en) * 2017-08-02 2019-02-07 Thomas Purves Machine learning tool
CN110110203A (zh) * 2018-01-11 2019-08-09 腾讯科技（深圳）有限公司 资源信息推送方法及服务器、资源信息展示方法及终端
CN109447687A (zh) * 2018-09-27 2019-03-08 中国平安人寿保险股份有限公司 基于大数据分析的用户挖掘方法及装置
CN113222732A (zh) * 2021-05-17 2021-08-06 中国农业银行股份有限公司 信息处理方法、装置、设备及存储介质
CN114330837A (zh) * 2021-12-08 2022-04-12 腾讯科技（深圳）有限公司 对象处理方法、装置、计算机设备和存储介质

Also Published As

Publication number Publication date
US20230342797A1 (en) 2023-10-26
CN114330837A (zh) 2022-04-12

Similar Documents

Publication Publication Date Title
US20230101487A1 (en) Customer journey management engine
US11694257B2 (en) Utilizing artificial intelligence to make a prediction about an entity based on user sentiment and transaction history
US20210264272A1 (en) Training method and system of neural network model and prediction method and system
WO2023103584A1 (zh) 对象处理方法、装置、计算机设备和存储介质
CN110751286B (zh) 神经网络模型的训练方法和训练系统
CN109783730A (zh) 产品推荐方法、装置、计算机设备和存储介质
CN109493199A (zh) 产品推荐方法、装置、计算机设备和存储介质
US8285622B1 (en) Method and system for providing budgeting recommendations based on financial data from similarly situated individuals
JP2023531100A (ja) エンティティがターゲットパラメータを充足しない確率を計算するための機械学習モデルアンサンブル
US11334941B2 (en) Systems and computer-implemented processes for model-based underwriting
Ye et al. Cold start to improve market thickness on online advertising platforms: Data-driven algorithms and field experiments
CN112348592A (zh) 广告推荐方法、装置、电子设备及介质
Shang et al. Need for speed, but how much does it cost? Unpacking the fee‐speed relationship in Bitcoin transactions
CN113034168A (zh) 内容项投放方法、装置、计算机设备及存储介质
Liu et al. Cost-Restricted Feature Selection for Data Acquisition
Singhal et al. An E‐commerce prediction system for product allocation to bridge the gap between cultural analytics and data science
Droftina et al. A diffusion model for churn prediction based on sociometric theory
Li et al. Learning-based ad auction design with externalities: the framework and a matching-based approach
Cheng’ai et al. PC-IEN: a click-through rate prediction method based on dynamic collaborative personalized interest extraction
CN115774813A (zh) 产品推荐方法、装置、计算机设备和存储介质
Qi et al. Recommendations based on social relationships in mobile services
CN113743906A (zh) 一种确定业务处理策略的方法和装置
CN111178535A (zh) 实现自动机器学习的方法和装置
Vora et al. Forecasting User Payment Behavior Using Machine Learning
US20240028935A1 (en) Context-aware prediction and recommendation

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 22903008

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE