WO2023016147A1 - 多目标预测方法、装置、设备、存储介质及程序产品 - Google Patents

多目标预测方法、装置、设备、存储介质及程序产品 Download PDF

Info

Publication number
WO2023016147A1
WO2023016147A1 PCT/CN2022/104024 CN2022104024W WO2023016147A1 WO 2023016147 A1 WO2023016147 A1 WO 2023016147A1 CN 2022104024 W CN2022104024 W CN 2022104024W WO 2023016147 A1 WO2023016147 A1 WO 2023016147A1
Authority
WO
WIPO (PCT)
Prior art keywords
historical behavior
historical
predicted
target
sequence
Prior art date
Legal status (The legal status 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 status listed.)
Ceased
Application number
PCT/CN2022/104024
Other languages
English (en)
French (fr)
Inventor
付煜文
陈亮
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Tencent Technology Shenzhen Co Ltd
Original Assignee
Tencent Technology Shenzhen Co Ltd
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 Tencent Technology Shenzhen Co Ltd filed Critical Tencent Technology Shenzhen Co Ltd
Publication of WO2023016147A1 publication Critical patent/WO2023016147A1/zh
Priority to US18/137,993 priority Critical patent/US20230259959A1/en
Anticipated expiration legal-status Critical
Ceased legal-status Critical Current

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; 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 OR CALCULATING; COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/20—Information retrieval; Database structures therefor; File system structures therefor of structured data, e.g. relational data
    • G06F16/24—Querying
    • G06F16/245—Query processing
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/20—Information retrieval; Database structures therefor; File system structures therefor of structured data, e.g. relational data
    • G06F16/27—Replication, distribution or synchronisation of data between databases or within a distributed database system; Distributed database system architectures therefor
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; 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
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; 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

Definitions

  • the embodiments of the present application relate to the field of computer and Internet technologies, and in particular to multi-objective prediction.
  • merchants will actively promote relevant content of the product to users when selling products.
  • the historical behavior data sequence corresponding to the product is obtained, and combined with the user's own characteristics, the user's business effect on the product is predicted.
  • Embodiments of the present application provide a multi-objective prediction method, device, device, storage medium, and program product.
  • differences between different prediction objects are considered to improve the accuracy of prediction results. Described technical scheme is as follows:
  • a multi-objective prediction method comprising:
  • the prediction results of the target objects corresponding to the events to be predicted under each of the predicted targets are obtained.
  • a multi-target prediction device includes:
  • a coding sequence acquisition module configured to obtain a historical behavior coding sequence according to the historical behavior data sequence of the target object; wherein, a historical behavior coding in the historical behavior coding sequence is a historical behavior coding sequence in the historical behavior data sequence The code representation corresponding to the behavior data;
  • a characteristic data generating module for each of the plurality of predicted targets, according to the correlation between each historical behavior code in the historical behavior code sequence and the predicted target, generate the target object relative to Historical feature data of the predicted target;
  • a prediction result acquisition module configured to obtain the prediction results of the target objects corresponding to the events to be predicted under each of the prediction targets according to the event information of the event to be predicted and the historical feature data corresponding to each of the prediction targets .
  • a server includes a processor and a memory, the memory stores at least one instruction, at least one program, a code set or an instruction set, and the at least one instruction, The at least one program, the code set or the instruction set is loaded and executed by the processor to implement the above multi-objective prediction method.
  • a computer-readable storage medium stores at least one instruction, at least one program, a code set or an instruction set, and the at least one instruction, all The at least one program, the code set or the instruction set is loaded and executed by the processor to realize the above multi-objective prediction method.
  • a computer program product or computer program includes computer instructions, and the computer instructions are stored in a computer-readable storage medium.
  • the processor of the server reads the computer instructions from the computer-readable storage medium, and the processor executes the computer instructions, so that the server executes the above multi-objective prediction method.
  • the corresponding historical characteristic data of the target object relative to different predicted targets are generated. Since the correlation between the target object and different predicted targets is different, the different predicted targets The historical feature data will also be different. Furthermore, on the basis of the event information of the event to be predicted, combined with the historical feature data corresponding to each predicted target, from the dimensions of each predicted target, the prediction results related to the target object and the event to be predicted are respectively obtained. , taking into account the difference in historical feature data of the target object between different prediction targets, so the obtained prediction results can reflect the difference of the target object under different prediction targets for the event to be predicted, thus improving the accuracy and comprehensiveness of the prediction results .
  • FIG. 1 is a schematic diagram of a multi-objective prediction system provided by an embodiment of the present application
  • Fig. 2 exemplarily shows a schematic diagram of a multi-target prediction system
  • FIG. 3 is a flowchart of a multi-objective prediction method provided by an embodiment of the present application.
  • Fig. 4 exemplarily shows a schematic diagram of the manner of acquiring event characteristic data corresponding to each prediction target
  • Fig. 5 exemplarily shows a schematic diagram of an acquisition method of historical behavior embedding
  • Fig. 6 exemplarily shows a schematic diagram of an acquisition method of a historical behavior code sequence
  • FIG. 7 exemplarily shows a schematic diagram of a flow of a multi-target prediction manner
  • Fig. 8 is a block diagram of a multi-target prediction device provided by an embodiment of the present application.
  • FIG. 9 is a block diagram of a multi-target prediction device provided by another embodiment of the present application.
  • FIG. 1 shows a schematic diagram of a multi-object prediction system provided by an embodiment of the present application.
  • the multi-object prediction system may include: a terminal 10 and a server 20 .
  • the terminal 10 may be an electronic device such as a mobile phone, a tablet computer, a game console, an e-book reader, a multimedia playback device, a wearable device, and a PC (Personal Computer, personal computer).
  • the terminal 10 may include a client of an application program, for example, the application program may be a shopping application program, a social application program, a game application program, a video application program, and the like.
  • the application program may be an application program that needs to be downloaded and installed, or a click-to-run application program, which is not limited in this embodiment of the present application.
  • the server 20 is used to provide background services for the terminal 10 .
  • the server 20 may be one server, or a server cluster composed of multiple servers, or a cloud computing service center.
  • the server 20 may be a background server of the above-mentioned client.
  • the server 20 provides background services for a plurality of terminals 10 .
  • the terminal 10 and the server 20 communicate with each other through the network 30 .
  • the above-mentioned application program is an application program with an item push function.
  • the terminal 10 sends to the server 20 a request for obtaining a prediction result of the target item.
  • the prediction result acquisition request includes the identification information of the target item.
  • the server 20 determines the target item according to the identification information of the target item, determines the event information of the event to be predicted according to the target item, and acquires the historical behavior code sequence of the target object.
  • the event information includes related item information, related user information and related scene information;
  • a historical behavior code in the historical behavior code sequence refers to a code representation corresponding to a historical behavior data in the historical behavior data sequence.
  • the server 20 obtains the historical feature data corresponding to each predicted target according to the correlation between each historical behavior code in the historical behavior code sequence and the predicted target, and combines the historical feature data corresponding to each predicted target according to the above event information
  • the prediction results of the target objects corresponding to the events to be predicted under each prediction target are obtained respectively, and then the server 20 sends the prediction results corresponding to each prediction target to the terminal 10 .
  • the terminal 10 determines to push the above-mentioned target item to the user.
  • the target item is the associated item corresponding to the above event information.
  • FIG. 3 shows a flowchart of a multi-object prediction method provided by an embodiment of the present application.
  • This method can be applied to the multi-objective prediction system shown in FIG. 1 , for example, the execution body of each step can be the server 20 .
  • the method may include the following steps (301-303):
  • Step 301 according to the historical behavior data sequence of the target object, obtain the historical behavior coding sequence.
  • the historical behavior data sequence is used to reflect the operation behavior of the target object on the item.
  • the historical behavior data sequence includes a plurality of historical behavior data, and one historical behavior data corresponds to one operation behavior.
  • the above-mentioned operation behavior refers to any one of click, conversion, search, bookmark, etc., which is not limited in this embodiment of the present application.
  • the aforementioned target objects may be objects corresponding to any one or more user accounts, and the aforementioned items may be any one or more items.
  • the historical behavior code sequence refers to the code representation corresponding to the historical behavior data sequence.
  • a historical behavior code in the historical behavior code sequence is a code representation corresponding to a historical behavior data in the historical behavior data sequence.
  • the server before obtaining the prediction result, obtains the historical behavior data sequence of the target object, and then obtains the historical behavior coding sequence according to the historical behavior data sequence.
  • the server after the server acquires the above-mentioned historical behavior data sequence, it performs coding processing on each historical behavior data in the historical behavior data sequence to obtain the above-mentioned historical behavior coding sequence.
  • each item of historical behavior data includes multiple types of historical feature information
  • the server when encoding the historical behavior data, the server encodes different types of historical feature information respectively to obtain various types of historical feature information.
  • the feature vectors corresponding to each type and then splicing the feature vectors corresponding to each type to obtain the historical behavior code corresponding to the historical behavior data.
  • Step 302 for each of the multiple predicted targets, according to the correlation between each historical behavior code in the historical behavior code sequence and the predicted target, generate historical feature data of the target object relative to the predicted target.
  • the prediction target refers to the index parameter used to measure the execution effect of the event to be predicted.
  • the event to be predicted may be any event, such as a product sale event, an advertisement conversion event, a video exposure event, etc., which is not limited in this embodiment of the present application.
  • one event to be predicted corresponds to one or more prediction targets.
  • the prediction targets corresponding to different events to be predicted may be the same or different.
  • the prediction target includes the click rate of the advertisement, the conversion rate of the advertisement, and the purchase quantity of the commodities in the advertisement.
  • the server obtains the historical characteristic data through weighting processing.
  • the above-mentioned correlation is used to indicate the degree of correlation between the historical behavior code and the predicted target.
  • the server obtains the above-mentioned historical behavior coding sequence, for the first predicted target among the multiple predicted targets, according to the correlation between the first predicted target and each historical behavior code, respectively determine the relationship between each historical behavior code and the first predicted target.
  • a degree of correlation between predicted targets, and then according to the degree of correlation the weight parameters of each historical behavior code for the first predicted target are respectively determined.
  • the weight parameter is positively correlated with the above-mentioned correlation degree.
  • the server may also obtain the historical characteristic data through code filtering and weighting processing.
  • the above association information is used to indicate the degree of association between the historical behavior code and the predicted target.
  • the server acquires the above-mentioned historical behavior codes, for the first predicted target among the multiple predicted targets, according to the degree of association between each historical behavior code and the first predicted target, obtain the historical behavior associated with the first predicted target. Behavior codes, and weight parameters corresponding to the historical behavior codes associated with the first predicted target are obtained, and historical feature data corresponding to the first predicted target is generated through weighting processing.
  • the historical feature data corresponding to the first predicted target only includes the historical behavior codes associated with the predicted target, and different historical behavior codes correspond to different weight parameters.
  • Step 303 according to the event information of the event to be predicted and the historical feature data corresponding to each prediction target, the prediction results of the target object corresponding to the event to be predicted under each prediction target are obtained.
  • the event information is used to indicate characteristic data of an event to be predicted.
  • the server after obtaining the historical characteristic data corresponding to each predicted target, the server respectively obtains the corresponding target object under each predicted target according to the event information of the event to be predicted and the historical characteristic data corresponding to each predicted target. The predicted outcome of a predicted event.
  • the above forecast results include click rate, conversion rate and purchase quantity.
  • the following sub-steps are also included:
  • the embodiment of the present application can sink down to specific application scenarios to perform prediction results for multiple specific prediction target dimensions, which effectively improves the recommendation of events to be predicted in terms of click-through rate, conversion rate, and purchase quantity. efficiency.
  • pCTCVR pCTR*pCVR
  • the purchase situation prediction result P is:
  • the above event information includes associated item information, associated user information, and associated scene information.
  • the server obtains related item information according to the related item corresponding to the event information; obtains related user information according to the user account corresponding to the event information; and obtains related scene information.
  • the server obtains the event information according to the associated item information, associated user information, and associated scene information, and further encodes the event information to obtain an encoded representation corresponding to the event information.
  • the above-mentioned associated item information includes historical purchase conditions of associated items.
  • the historical purchase situation may be the historical purchase situation of the user account for the associated item, or the historical purchase situation of the target object for the associated item, which is not limited in this embodiment of the present application.
  • the above-mentioned purchase information includes but not limited to at least one of the following: whether it has been purchased, the time of purchase, the number of purchases, the scene information at the time of purchase, and the like.
  • the above-mentioned associated item information may also include but not limited to at least one of the following: identification information of the associated item, type of the associated item, display platform of the associated item, display location of the associated item, and the like.
  • the above associated user information is used to reflect account tag information corresponding to the user account.
  • the account label information may be drawn by the server according to various user data corresponding to the user account, and used to reflect the characteristics of the user account.
  • the user data includes but is not limited to at least one of the following: user age corresponding to the user account, user gender corresponding to the user account, balance corresponding to the user account, and the like.
  • the above associated scene information is used to indicate the environment in which the operation behavior occurs.
  • the environment in which the operation behavior occurs is the current scene environment.
  • the related scene information includes but is not limited to at least one of the following: recent (today, within the last week, within the last month, etc.) traffic of people on the display platform of related items, and recent transactions of items on the exhibit platform of related items The quantity, the recent market competitiveness of the display platform of related items, etc.
  • the event information of the event to be predicted can be comprehensively represented from multiple dimensions, so that the coded representation of event information can represent the event to be predicted more accurately, which is helpful To improve the accuracy of subsequent prediction results.
  • the technical solution provided by the embodiment of the present application through the correlation between each historical behavior code and the prediction target, generates the historical feature data corresponding to the target object with respect to different prediction targets, because the target object and different prediction targets The correlation between them is different, so the historical feature data for different prediction targets will also be different. Furthermore, on the basis of the event information of the event to be predicted, combined with the historical feature data corresponding to each predicted target, from the dimensions of each predicted target, the prediction results related to the target object and the event to be predicted are respectively obtained. , taking into account the difference in historical feature data of the target object between different prediction targets, so the obtained prediction results can reflect the difference of the target object under different prediction targets for the event to be predicted, thus improving the accuracy and comprehensiveness of the prediction results .
  • step 302 includes the following steps:
  • the server when the server acquires the above-mentioned historical feature data, it acquires the operation behaviors corresponding to the respective historical behavior codes in the mountain score historical behavior code sequence.
  • the server after acquiring the operation behavior corresponding to each of the above historical behavior codes, acquires the correlation between each operation behavior and the prediction target for each of the multiple prediction targets.
  • the correlation degree is used to quantify and characterize the above correlation.
  • the above correlation degree is used to indicate whether the operation behavior is related to the predicted target. For example, if the operation behavior is related to the predicted target, the correlation between the historical behavior code and the predicted target is "1"; if the operational behavior is not related to the predicted target, the correlation between the historical behavior code and the predicted target is "0".
  • whether the operation behavior is related to the predicted target may be preset information. For example, if the prediction target is the conversion rate, the relevant operation behaviors may be preset to include purchase and collection; if the prediction target is the click rate, the relevant operation behaviors may be preset to include click and search.
  • the above-mentioned degree of correlation is used to indicate a degree of correlation between the operation behavior and the predicted target. For example, if the operational behavior is strongly correlated with the predicted target, the correlation between the historical behavior code and the predicted target is "1"; if the operational behavior is correlated with the predicted target, the correlation between the historical behavior code and the predicted target The degree is "0.5"; if the operational behavior is weakly correlated with the predicted target, then the correlation between the historical behavior code and the predicted target is "0.1".
  • the degree of correlation between the operation behavior and the predicted target may be preset information.
  • the prediction target is the conversion rate
  • the forecast target is click rate
  • each historical behavior code is weighted to obtain the historical characteristic data of the target object relative to the predicted target.
  • the server after obtaining the above correlation, performs weighting processing on each historical behavior code according to the correlation between each operation behavior and the predicted target, and obtains historical characteristic data corresponding to the predicted target.
  • the server when performing weighting processing, obtains the similarity between each historical behavior code and the code representation corresponding to the event information; further, according to the correlation and similarity corresponding to each historical behavior code, determine each historical behavior The weight parameters corresponding to the behavior codes; then, according to the weight parameters corresponding to the historical behavior codes, each historical behavior code is weighted to obtain the historical characteristic data corresponding to the prediction target.
  • both the above-mentioned degree of correlation and the above-mentioned degree of similarity are expressed in the form of specific numerical values, and the magnitude of the degree of correlation is positively correlated with the magnitude of the numerical value, and the magnitude of the degree of similarity is also positively correlated with the magnitude of the numerical value.
  • the server obtains the above weight parameter by multiplying the correlation degree and the similarity degree.
  • the server directly uses the product as the above weight parameter; in another possible implementation manner, after obtaining the above product, the server performs normalization processing on the product to obtain the above weight parameter.
  • the above-mentioned correlation degree and the above-mentioned similarity degree may also have other representation forms, which are not limited in this embodiment of the present application.
  • the representation forms of correlation degree include extremely high correlation degree, high correlation degree, general correlation degree, weak correlation degree, extremely weak correlation degree, no correlation degree, etc.
  • the similarity degree representation forms include extremely high similarity degree, high similarity degree, General similarity, weak similarity, very weak similarity, no similarity, etc.
  • the server determines the importance of each historical behavior code for the prediction target, and then determines the corresponding weight parameter.
  • the weight parameter is positively correlated with the above-mentioned degree of importance.
  • the product between the correlation degree and the similarity degree may be used to represent the above-mentioned association situation.
  • the above-mentioned product is directly used to represent the above-mentioned degree of association; in another possible implementation, after obtaining the above-mentioned product, the historical behavior code whose product is greater than a certain value is determined to be related to the predicted target.
  • the associated historical behavior code the value may be any value, which is not limited in this embodiment of the present application.
  • step 303 includes the following steps:
  • the server when acquiring the above-mentioned prediction results, the server extracts event feature data corresponding to each prediction target based on the above-mentioned event information.
  • the prediction target includes click-through rate, conversion rate and purchase quantity.
  • FIG. Indicates that the feature extraction process is performed, and the feature extraction results from different expert networks are obtained.
  • the first weighted parameters corresponding to each feature extraction result are determined by the click-through rate weighted gate based on the above coded representation
  • the second weighted parameters corresponding to each feature extraction result are determined by the conversion rate weighted gate based on the above coded representation, weighted by the number of purchases
  • the gate determines the third weighting parameters corresponding to the respective feature extraction results based on the above encoding representation.
  • weighted summation processing is performed on each feature extraction result respectively, and event feature data corresponding to each predicted target is obtained.
  • the feature extraction result set composed of the feature extraction results given by different expert networks for event information can fully characterize the essence of event information, thereby improving the accuracy of event feature data corresponding to event information and each prediction target.
  • the server after obtaining the event feature data corresponding to each of the forecast targets, acquires the forecast result corresponding to the forecast target according to the historical feature data and event feature data corresponding to the forecast target.
  • the server performs fusion processing on the historical feature data and event feature data corresponding to the predicted target to obtain the fused feature data corresponding to the predicted target; data to generate prediction results corresponding to the prediction target.
  • the above prediction network is a Tower network.
  • the server when acquiring the above-mentioned historical behavior encoding sequence, the server first acquires the historical behavior data sequence, and then obtains the historical behavior embedding vector according to the historical behavior data sequence.
  • the above-mentioned historical behavior data sequence is a data sequence related to the above-mentioned event information, and when acquiring the historical behavior data sequence, the server acquires the historical behavior data sequence based on the above-mentioned event information.
  • the server acquires the historical behavior data sequence based on the user account corresponding to the event information.
  • the server acquires the user account corresponding to the above event information, and then determines the target object according to the user account.
  • the target object includes user accounts and/or similar user accounts.
  • the similar user account has similar user characteristics between the user accounts.
  • the similar user characteristics include but are not limited to at least one of the following: the age difference between users is smaller than the first target value, the users have the same gender, the users have purchased similar items, and the number of similar items purchased by the users The difference is less than the second target value, the user has browsed similar videos, and so on.
  • the server acquires historical behavior data based on the target item corresponding to the event information.
  • the server acquires the target item corresponding to the above event information, and then determines the item targeted by the historical behavior data according to the target item.
  • the item includes the target item and/or similar items.
  • the similar item has similar item features with the target item.
  • the characteristics of the similar items include but are not limited to at least one of the following: the same item type, the same display platform of the items, the same user group targeted by the items, and the like.
  • the server may also acquire the above historical behavior data sequence by taking the user account and the target item corresponding to the event information as constraints.
  • the server determines the above-mentioned target object according to the user account, and then obtains the historical behavior data corresponding to the item indicated by the target item within the scope of the entire amount of historical behavior data corresponding to the target object, and generates the above-mentioned historical behavior data sequence.
  • a historical behavior embedding in the above historical behavior embedding sequence refers to an embedding representation corresponding to a piece of historical behavior data in the historical behavior data sequence.
  • each item of historical behavior data includes historical item information, historical behavior information and historical scene information.
  • the server acquires the first feature vector, the second feature vector and the third feature vector for each item of historical behavior data in the historical behavior data sequence.
  • the above-mentioned first feature vector refers to a feature vector corresponding to historical object information
  • the above-mentioned second feature vector refers to a feature vector corresponding to historical behavior information
  • the above-mentioned third feature vector refers to a feature vector corresponding to historical scene information.
  • the server splices the first eigenvector, the second eigenvector, and the third eigenvector corresponding to each historical behavior data to obtain the above-mentioned historical behavior embedding sequence.
  • the historical behavior embedding obtained by splicing the feature vectors representing the aforementioned different dimensions has a more comprehensive representation ability.
  • the historical behavior encoding obtained by encoding the historical behavior embedding can effectively reduce the amount of data while ensuring the representation accuracy, thereby saving the resource occupation of subsequent processing.
  • the server acquires at least one previous historical behavior embedding according to the occurrence time of the operation behavior corresponding to the historical behavior embedding.
  • the occurrence time of the operation behavior corresponding to the pre-order historical behavior embedding is before the occurrence time of the operation behavior corresponding to the historical behavior embedding.
  • the server encodes each historical behavior embedding and at least one preorder historical behavior embedding corresponding to the historical behavior embedding to obtain a historical behavior encoding sequence.
  • the historical behavior embedding sequence includes historical behavior embedding E 1 , historical behavior embedding E 2 ... historical behavior embedding E n , and the occurrence time corresponding to historical behavior embedding E 1 is the earliest, and the historical behavior Embedding E n corresponds to the latest occurrence time.
  • the server obtains the historical behavior embedding sequence, for each historical behavior embedding, according to the occurrence time corresponding to the historical behavior embedding, at least one preorder historical behavior embedding is obtained, and each historical behavior embedding and historical At least one preorder historical behavior embedding corresponding to the behavior embedding is encoded to obtain a historical behavior encoding sequence.
  • the historical behavior coding sequence includes historical behavior coding T 1 , historical behavior coding T 2 ... historical behavior coding T n .
  • the pre-sequence historical behavior embedding before the historical behavior embedding can be obtained through the occurrence time of the operation behavior, and the pre-sequence historical behavior embedding is used to encode the historical behavior embedding, which is equivalent to incorporating the previous operation behavior in the historical behavior coding , which is equivalent to incorporating the above information, thus adding time-series dimension information to the historical behavior coding and expanding the observation horizon corresponding to the historical behavior coding.
  • the complete multi-target forecast method of this application will be introduced.
  • the historical behavior encoding sequence is obtained by the second feature processing block shown in Figure 6, wherein the second feature processing block includes a self-attention mechanism; the first feature processing block shown in Figure 4 passes the event information
  • the corresponding code indicates that the event feature data corresponding to the click-through rate, the event feature data corresponding to the conversion rate, and the event feature data corresponding to the purchase quantity are obtained.
  • the server obtains the prediction result corresponding to the click rate through the prediction network corresponding to the click rate, according to the historical feature data corresponding to the click rate and the event feature data corresponding to the click rate; through the prediction network corresponding to the conversion rate, according to the historical feature data corresponding to the conversion rate Feature data and event feature data corresponding to the conversion rate to obtain the prediction result corresponding to the conversion rate; through the forecast network corresponding to the purchase quantity, according to the historical feature data corresponding to the purchase quantity and the event characteristic data corresponding to the purchase quantity, obtain the forecast corresponding to the purchase quantity result.
  • the server can obtain the prediction result corresponding to the click conversion rate according to the prediction result corresponding to the click rate and the prediction result corresponding to the conversion rate; and obtain the purchase situation prediction result according to the prediction result corresponding to the click conversion rate and the prediction result corresponding to the purchase quantity .
  • FIG. 8 shows a block diagram of a multi-object prediction device provided by an embodiment of the present application.
  • the device has the function of realizing the above-mentioned multi-objective prediction method, and the function can be realized by hardware, and can also be realized by hardware executing corresponding software.
  • the device may be a server, or be set in the server.
  • the apparatus 800 may include: an encoding sequence acquisition module 810 , a feature data generation module 820 and a prediction result acquisition module 830 .
  • a coding sequence acquisition module 810 configured to obtain a historical behavior coding sequence according to the historical behavior data sequence of the target object; wherein, a historical behavior coding in the historical behavior coding sequence is an item in the historical behavior data sequence The coded representation corresponding to historical behavior data.
  • the feature data generation module 820 is configured to, for each of the multiple predicted targets, generate the relative Based on the historical feature data of the predicted target.
  • the prediction result acquisition module 830 is configured to obtain the predictions of the target objects corresponding to the events to be predicted under each of the prediction targets according to the event information of the event to be predicted and the historical feature data corresponding to each of the prediction targets result.
  • the feature data generation module 820 includes: an operation behavior acquisition unit 821 , a correlation acquisition unit 822 and a feature data acquisition unit 823 .
  • the operation behavior acquiring unit 821 is configured to acquire the operation behavior corresponding to each historical behavior code in the historical behavior code sequence.
  • the correlation acquisition unit 822 is configured to obtain, for each of the plurality of predicted targets, the correlation between each of the operation behaviors and the predicted target, and the correlation is used to quantify and characterize the Correlation.
  • the feature data acquisition unit 823 is configured to perform weighting processing on each of the historical behavior codes according to the correlation between each of the operation behaviors and the predicted target, and obtain the historical characteristics of the target object relative to the predicted target data.
  • the feature data acquiring unit 823 is configured to respectively acquire the similarity between each of the historical behavior codes and the code representations corresponding to the event information; Relevance and similarity, determine the weight parameters corresponding to each of the historical behavior codes respectively; according to the weight parameters corresponding to each of the historical behavior codes, carry out weighting processing on each of the historical behavior codes to obtain the corresponding prediction target historical feature data.
  • the prediction result acquisition module 830 includes: an event feature acquisition unit 831 and a prediction result acquisition unit 832.
  • the event feature acquisition unit 831 is configured to extract event feature data corresponding to each of the predicted targets based on the event information.
  • the prediction result acquisition unit 832 is configured to, for each prediction target, obtain the prediction result of the target object corresponding to the event to be predicted at the prediction target according to the historical characteristic data and event characteristic data corresponding to the prediction target.
  • the event feature acquisition unit 831 is configured to acquire a coded representation corresponding to the event information; use different expert networks to perform feature extraction processing on the coded representation to obtain a feature extraction result set; wherein , the feature extraction result set includes feature extraction results from different expert networks; different weight gates are used respectively to obtain multiple sets of weighted parameters for the feature extraction result set; wherein, each set of weighted parameters includes the feature
  • Each feature extraction result in the extraction result set corresponds to the weighting parameters respectively, and there is a one-to-one correspondence between the weighted gate and the prediction target; for each prediction target, based on the multiple groups of weighting parameters and the described A set of weighted parameters corresponding to the predicted target, performing a weighted summation process on each feature extraction result in the feature extraction result set, to obtain event feature data corresponding to the predicted target.
  • the prediction result acquisition unit 832 is configured to, for each prediction target, perform fusion processing on the historical feature data and event feature data corresponding to the prediction target to obtain the fusion feature corresponding to the prediction target data; through the prediction network corresponding to the prediction target, according to the fusion feature data, the prediction results of the target object corresponding to the event to be predicted under each of the prediction targets are generated.
  • the coding sequence obtaining module 810 includes: an embedded sequence obtaining unit 811 and a coding sequence obtaining unit 812 .
  • An embedding sequence acquisition unit 811 configured to acquire a historical behavior embedding sequence according to the historical behavior data sequence; wherein, a historical behavior embedding in the historical behavior embedding sequence is a historical behavior embedding in the historical behavior data sequence Embedded representations corresponding to behavioral data.
  • the encoding sequence acquisition unit 812 is configured to perform encoding processing on each of the historical behavior embeddings in the historical behavior embedding sequence to obtain the historical behavior encoding sequence.
  • the encoding sequence acquisition unit 812 is configured to, for each historical behavior embedding in the historical behavior embedding sequence, acquire at least one Item preorder historical behavior embedding; wherein, the preorder historical behavior embedding corresponds to the occurrence moment of the operation behavior, before the occurrence moment of the historical behavior embedding corresponding operation behavior; respectively embedding each of the historical behaviors, and the The at least one preorder historical behavior embedding corresponding to the historical behavior embedding is encoded to obtain the historical behavior encoding sequence.
  • each item of historical behavior data includes historical item information, historical behavior information, and historical scene information;
  • the embedded sequence acquisition unit 811 is configured to Behavioral data, obtaining a first feature vector, a second feature vector, and a third feature vector; wherein, the first feature vector refers to the feature vector corresponding to the historical item information, and the second feature vector refers to the historical item information
  • the feature vector corresponding to the behavior information, the third feature vector refers to the feature vector corresponding to the historical scene information; respectively carry out the first feature vector, the second feature vector and the third feature vector corresponding to each of the historical behavior data Splicing processing to obtain the historical behavior embedding sequence.
  • the event information includes associated item information, associated user information, and associated scene information; as shown in FIG. 9 , the apparatus 800 further includes: an event code acquiring module 840 .
  • the event code acquiring module 840 is configured to acquire the related item information according to the related item corresponding to the event information, the related item information includes the historical purchase status of the related item; according to the user account corresponding to the event information, Obtain the associated user information, the associated user information is used to reflect the account label information corresponding to the user account; obtain associated scene information, the associated scene information is used to indicate the environment in which the operation behavior occurs; according to the associated item information, The associating user information and the associating scene information obtain the event information; encode the event information to obtain a coded representation corresponding to the event information.
  • the predicted result includes a click rate, a conversion rate, and a purchase quantity; as shown in FIG. 9 , the apparatus 800 further includes: a predicted result processing module 850 .
  • the prediction result processing module 850 is configured to obtain, according to the click rate and the conversion rate, the click conversion rate of the user account corresponding to the event information for the associated item; and/or, according to the click conversion rate and the purchase Quantity, to obtain the prediction result of the purchase situation of the associated item by the user account.
  • the technical solution provided by the embodiment of the present application generates the historical feature data corresponding to the predicted target through the correlation between each historical behavior code and the predicted target, and then based on the event information of the event to be predicted, respectively Combining the historical feature data corresponding to each forecast target, the forecast results corresponding to each of the forecast targets are respectively obtained.
  • the differences between different forecast targets are considered.
  • according to different historical features Data to obtain prediction results and improve the accuracy of prediction results.
  • FIG. 10 shows a structural block diagram of a server provided by an embodiment of the present application.
  • the server can be used to realize the functions of the above-mentioned multi-objective prediction method. Specifically:
  • the server 1000 includes a central processing unit (Central Processing Unit, CPU) 1001, a system memory 1004 including a random access memory (Random Access Memory, RAM) 1002 and a read only memory (Read Only Memory, ROM) 1003, and a connection system memory 1004 and the system bus 1005 of the central processing unit 1001.
  • the server 1000 also includes a basic input/output (Input/Output, I/O) system 1006 to help transfer information between various devices in the computer, and a large capacity for storing the operating system 1013, application programs 1014 and other program modules 1015 Storage device 1007.
  • I/O input/output
  • the basic input/output system 1006 includes a display 1008 for displaying information and an input device 1009 such as a mouse and a keyboard for a user to input information. Both the display 1008 and the input device 1009 are connected to the central processing unit 1001 through the input and output controller 1010 connected to the system bus 1005 .
  • the basic input/output system 1006 may also include an input output controller 1010 for receiving and processing input from a number of other devices such as a keyboard, mouse, or electronic stylus. Similarly, input output controller 1010 also provides output to a display screen, printer, or other type of output device.
  • Computer-readable media may comprise computer storage media and communication media.
  • Computer storage media includes volatile and nonvolatile, removable and non-removable media implemented in any method or technology for storage of information such as computer readable instructions, data structures, program modules or other data.
  • Computer storage media include RAM, ROM, EPROM (Erasable Programmable Read Only Memory, Erasable Programmable Read Only Memory), EEPROM (Electrically Erasable Programmable Read Only Memory, Electrically Erasable Programmable Read Only Memory), flash memory or other solid-state storage Its technology, CD-ROM, DVD (Digital Video Disc, high-density digital video disc) or other optical storage, tape cassette, tape, magnetic disk storage or other magnetic storage devices.
  • the computer storage medium is not limited to the above-mentioned ones.
  • the above-mentioned system memory 1004 and mass storage device 1007 may be collectively referred to as memory.
  • the server 1000 can also run on a remote computer connected to the network through a network such as the Internet. That is to say, the server 1000 can be connected to the network 1012 through the network interface unit 1011 connected to the system bus 1005, or can use the network interface unit 1011 to connect to other types of networks or remote computer systems (not shown).
  • the memory also includes a computer program stored in the memory and configured to be executed by one or more processors to implement the above-mentioned multi-objective prediction method.
  • a computer-readable storage medium is also provided. At least one instruction, at least one program, code set or instruction set are stored in the storage medium, and the at least one instruction, the at least one program , the code set or the instruction set is executed by a processor to implement the above multi-objective prediction method.
  • the computer-readable storage medium may include: ROM (Read Only Memory, read-only memory), RAM (Random Access Memory, random access memory), SSD (Solid State Drives, solid state drive) or an optical disc, etc.
  • the random access memory may include ReRAM (Resistance Random Access Memory, resistive random access memory) and DRAM (Dynamic Random Access Memory, dynamic random access memory).
  • a computer program product or computer program comprising computer instructions stored on a computer readable storage medium.
  • the processor of the server reads the computer instructions from the computer-readable storage medium, and the processor executes the computer instructions, so that the server executes the above multi-objective prediction method.

Landscapes

  • Engineering & Computer Science (AREA)
  • Business, Economics & Management (AREA)
  • Strategic Management (AREA)
  • Finance (AREA)
  • Accounting & Taxation (AREA)
  • Development Economics (AREA)
  • Theoretical Computer Science (AREA)
  • Entrepreneurship & Innovation (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • Economics (AREA)
  • Game Theory and Decision Science (AREA)
  • Marketing (AREA)
  • General Business, Economics & Management (AREA)
  • Databases & Information Systems (AREA)
  • General Engineering & Computer Science (AREA)
  • Computational Linguistics (AREA)
  • Computing Systems (AREA)
  • Management, Administration, Business Operations System, And Electronic Commerce (AREA)

Abstract

一种多目标预测方法，包括：根据目标对象的历史行为数据序列，获取历史行为编码序列，对于多个预测目标中的每一个预测目标，根据历史行为编码序列中的各个历史行为编码与所述预测目标之间的相关性，生成目标对象相对于预测目标的历史特征数据；根据待预测事件的事件信息，以及各个预测目标分别对应的历史特征数据，获取目标对象在各个预测目标下分别对应待预测事件的预测结果。还公开了一种多目标预测装置、设备、存储介质及程序产品。

Description

多目标预测方法、装置、设备、存储介质及程序产品
本申请要求于2021年08月09日提交中国专利局、申请号为202110907940.9、申请名称为“多目标预测方法、装置、设备及存储介质”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
本申请实施例涉及计算机和互联网技术领域，特别涉及多目标预测。
背景技术
目前，商家在售卖产品时会主动向用户推动该产品的相关内容。在相关技术中，在向用户推送产品的相关内容之前，获取该产品对应的历史行为数据序列，结合用户自身的特征，预测用户针对该产品的业务效果。
然而，在上述相关技术中，直接将历史行为数据序列作为整体来预测业务效果，预测结果不准确。
发明内容
本申请实施例提供了一种多目标预测方法、装置、设备、存储介质及程序产品，在获取预测结果时，考虑了不同预测目标之间的差异，提高预测结果的准确性。所述技术方案如下：
根据本申请实施例的一个方面，提供了一种多目标预测方法，所述方法包括：
根据目标对象的历史行为数据序列，获取历史行为编码序列；其中，所述历史行为编码序列中的一项历史行为编码，是所述历史行为数据序列中的一项历史行为数据对应的编码表示；
对于多个预测目标中的每一个预测目标，根据所述历史行为编码序列中的各个历史行为编码与所述预测目标之间的相关性，生成所述目标对象相对于所述预测目标的历史特征数据；
根据待预测事件的事件信息，以及各个所述预测目标分别对应的历史特征数据，获取所述目标对象在各个所述预测目标下分别对应所述待预测事件的预测结果。
根据本申请实施例的一个方面，提供了一种多目标预测装置，所述装置包括：
编码序列获取模块，用于根据目标对象的历史行为数据序列，获取历史行为编码序列；其中，所述历史行为编码序列中的一项历史行为编码，是所述历史行为数据序列中的一项历史行为数据对应的编码表示；
特征数据生成模块，用于对于多个预测目标中的每一个预测目标，根据所述历史行为编码序列中的各个历史行为编码与所述预测目标之间的相关性，生成所述目标对象相对于所述预测目标的历史特征数据；
预测结果获取模块，用于根据待预测事件的事件信息，以及各个所述预测目标分别对应的历史特征数据，获取所述目标对象在各个所述预测目标下分别对应所述待预测事件的预测结果。
根据本申请实施例的一个方面，提供了一种服务器，所述服务器包括处理器和存储器，所述存储器中存储有至少一条指令、至少一段程序、代码集或指令集，所述至少一条指令、 所述至少一段程序、所述代码集或指令集由所述处理器加载并执行以实现上述多目标预测方法。
根据本申请实施例的一个方面，提供了一种计算机可读存储介质，所述计算机可读存储介质中存储有至少一条指令、至少一段程序、代码集或指令集，所述至少一条指令、所述至少一段程序、所述代码集或指令集由处理器加载并执行以实现上述多目标预测方法。
根据本申请实施例的一个方面，提供了一种计算机程序产品或计算机程序，该计算机程序产品或计算机程序包括计算机指令，该计算机指令存储在计算机可读存储介质中。服务器的处理器从计算机可读存储介质读取该计算机指令，处理器执行该计算机指令，使得该服务器执行上述多目标预测方法。
本申请实施例提供的技术方案可以包括如下有益效果：
通过各个历史行为编码与预测目标之间的相关性，生成目标对象相对于不同预测目标分别对应的历史特征数据，由于目标对象和不同预测目标之间的相关性有所区别，故针对不同预测目标的历史特征数据也会不同。进而在待预测事件的事件信息的基础上，结合各个预测目标分别对应的历史特征数据，从各个预测目标的维度下，分别获取目标对象与待预测事件相关的预测结果，由于在获取预测结果时，考虑了目标对象在不同预测目标之间历史特征数据的差异，故获取的预测结果可以体现出目标对象针对待预测事件在不同预测目标下的区别，从而提高了预测结果的准确性和全面性。
附图说明
图1是本申请一个实施例提供的多目标预测系统的示意图；
图2示例性示出了一种多目标预测系统的示意图；
图3是本申请一个实施例提供的多目标预测方法的流程图；
图4示例性示出了各个预测目标分别对应的事件特征数据的获取方式的示意图；
图5示例性示出了一种历史行为嵌入的获取方式的示意图；
图6示例性示出了一种历史行为编码序列的获取方式的示意图；
图7示例性示出了一种多目标预测方式的流程的示意图；
图8是本申请一个实施例提供的多目标预测装置的框图；
图9是本申请另一个实施例提供的多目标预测装置的框图；
图10是本申请一个实施例提供的服务器结构的示意图。
具体实施方式
为使本申请的目的、技术方案和优点更加清楚，下面将结合附图对本申请实施方式作进一步地详细描述。
请参考图1，其示出了本申请一个实施例提供的多目标预测系统的示意图。该多目标预测系统可以包括：终端10和服务器20。
终端10可以是诸如手机、平板电脑、游戏主机、电子书阅读器、多媒体播放设备、可穿戴设备、PC(Personal Computer，个人计算机)等电子设备。终端10中可以包括应用程序的客户端，如该应用程序可以是购物应用程序、社交应用程序、游戏应用程序、视频应用程序等。可选地，该应用程序可以是需要下载安装的应用程序，也可以是即点即用的应 用程序，本申请实施例对此不作限定。
服务器20用于为终端10提供后台服务。服务器20可以是一台服务器，也可以是由多台服务器组成的服务器集群，或者是一个云计算服务中心。可选地，服务器20可以是上述客户端的后台服务器。在示例性实施例中，服务器20为多个终端10提供后台服务。
上述终端10和上述服务器20通过网络30进行互相通信。
可选地，上述应用程序为具有物品推送功能的应用程序。示例性地，由终端10向服务器20发送针对目标物品的预测结果获取请求。其中，该预测结果获取请求中包括目标物品的标识信息。进一步地，服务器20在根据目标物品的标识信息确定目标物品，依据该目标物品确定待预测事件的事件信息，并获取目标对象的历史行为编码序列。其中，事件信息中包括关联物品信息、关联用户信息和关联场景信息；历史行为编码序列中的一项历史行为编码，是指历史行为数据序列中的一项历史行为数据对应的编码表示。之后，服务器20根据历史行为编码序列中的各个历史行为编码与预测目标之间的相关性，获取各个预测目标分别对应的历史特征数据，并根据上述事件信息，结合各个预测目标分别对应的历史特征数据，分别获取目标对象在各个预测目标下分别对应待预测事件的预测结果，进而由服务器20向终端10发送各个预测目标分别对应的预测结果。之后，终端10在预测结果满足条件的情况下，确定向用户推送上述目标物品。需要说明的一点是，目标物品即为上述事件信息对应的关联物品。
可选地，在本申请实施例中，上述服务器20为多个服务器中的任一服务器，其中，多个服务器可组成为一区块链，而服务器为区块链上的节点，即上述服务器20为区块链上的节点。可选地，本申请中，通过服务器中的程序指令来执行实现上述多目标预测方法。作为示例，程序指令可被部署在一个服务器上执行，或者在位于一个地点的多个服务器上执行，又或者，在分布在多个地点且通过通信网络互连的多个服务器上执行，分布在多个地点且通过通信网络互连的多个服务器可以组成区块链系统。
可以理解的是，在本申请的具体实施方式中，涉及到历史行为数据等相关的数据，当本申请以上实施例运用到具体产品或技术中时，任意一项的运用均需要单独获得用户许可或者同意，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。
请参考图3，其示出了本申请一个实施例提供的多目标预测方法的流程图。该方法可应用于图1所示的多目标预测系统中，如各步骤的执行主体可以是服务器20。该方法可以包括以下几个步骤(301～303)：
步骤301，根据目标对象的历史行为数据序列，获取历史行为编码序列。
历史行为数据序列用于反映目标对象针对物品的操作行为。其中，该历史行为数据序列中包括多个历史行为数据，一个历史行为数据对应一个操作行为。可选地，上述操作行为是指点击、转化、搜索、收藏等其中的任意一种，本申请实施例对此不作限定。需要说明的是，上述目标对象可以是任意一个或多个用户账号对应地对象，而且，上述物品可以是任意一个或多个物品。
历史行为编码序列是指历史行为数据序列对应的编码表示。其中，历史行为编码序列 中的一项历史行为编码，是历史行为数据序列中的一项历史行为数据对应的编码表示。
在本申请实施例中，服务器在获取预测结果之前，获取目标对象的历史行为数据序列，进而根据该历史行为数据序列，获取历史行为编码序列。可选地，服务器在获取上述历史行为数据序列之后，对该历史行为数据序列中的各个历史行为数据分别进行编码处理，得到上述历史行为编码序列。
可选地，在本申请实施例中，每项历史行为数据中包括多个类型的历史特征信息，服务器在对历史行为数据进行编码时，分别对不同类型的历史特征信息进行编码，得到各个类型分别对应的特征向量，进而将各个类型分别对应的特征向量进行拼接，以获取该历史行为数据对应的历史行为编码。
步骤302，对于多个预测目标中的每一个预测目标，根据历史行为编码序列中的各个历史行为编码与预测目标之间的相关性，生成目标对象相对于预测目标的历史特征数据。
预测目标是指用于衡量待预测事件的执行效果的指标参数。待预测事件可以为任意事件，如商品的出售事件、广告的转化事件、视频的曝光事件等，本申请实施例对此不作限定。可选地，一个待预测事件对应一个或多个预测目标。而且，不同的待预测事件对应的预测目标可以相同，也可以不同。示例性地，若待预测事件为广告的转化事件，则预测目标包括广告的点击率、广告的转化率、广告中商品的购买数量。
在本申请实施例中，服务器在获取上述历史行为编码序列之后，对于多个预测目标中的每一个预测目标，根据历史行为编码序列中的各个历史行为编码与预测目标之间的相关性，生成目标对象对应预测目标的历史特征数据。其中，上述相关性用于指示历史行为编码与预测目标之间的关联情况，目标对象对应不同预测目标的历史特征数据可以不同。
在一种可能的实施方式中，服务器通过编码过滤的方式来获取上述历史特征数据，在这种情况下，上述关联情况用于指示历史行为编码与预测目标是否关联。可选地，服务器在获取上述历史行为编码序列之后，对于多个预测目标中的第一预测目标，根据第一预测目标与各个历史行为编码之间的相关性，从多个历史行为编码中选择与第一预测目标关联的历史行为编码，生成该第一预测目标对应的历史特征数据。此时，第一预测目标对应的历史特征数据中仅仅包括与该预测目标关联的历史行为编码，减少后续数据处理的计算量。
在另一种可能的实时方式中，服务器通过加权处理的方式来获取上述历史特征数据，在这种情况下，上述关联情况用于指示历史行为编码与预测目标之间的关联程度。可选地，服务器在获取上述历史行为编码序列之后，对于多个预测目标中的第一预测目标，根据第一预测目标与各个历史行为编码之间的相关性，分别确定各个历史行为编码与第一预测目标之间的关联程度，进而根据该关联程度，分别确定各个历史行为编码针对第一预测目标的权重参数。其中，该权重参数与上述关联程度呈正相关关系。之后，基于各个历史行为编码分别对应的权重参数，对各个历史行为编码分别进行加权处理，生成该第一预测目标对应的历史特征数据。此时，第一预测目标对应的历史特征数据中包括全量历史行为编码，且不同历史行为编码对应有不同的权重参数，提高后续数据处理的准确性。
当然，在其它可能的实施方式中，服务器也可以通过编码过滤和加权处理来获取上述历史特征数据，在这种情况下，上述关联情况用于指示历史行为编码与预测目标之间的关 联程度。可选地，服务器在获取上述历史行为编码之后，对于多个预测目标中的第一预测目标，根据各个历史行为编码与第一预测目标之间的关联程度，获取与第一预测目标关联的历史行为编码，并获取各个与第一预测目标关联的历史行为编码分别对应的权重参数，通过加权处理生成该第一预测目标对应的历史特征数据。此时，第一预测目标对应的历史特征数据中仅仅包括与该预测目标关联的历史行为编码，且不同历史行为编码对应有不同的权重参数。
步骤303，根据待预测事件的事件信息，以及各个预测目标分别对应的历史特征数据，获取目标对象在各个预测目标下分别对应待预测事件的预测结果。
事件信息用于指示待预测事件的特征数据。在本申请实施例中，服务器在获取各个预测目标分别对应的历史特征数据之后，根据待预测事件的事件信息，以及各个预测目标分别对应的历史特征数据，分别获取各个预测目标下目标对象对应待预测事件的预测结果。
可选地，上述预测结果包括点击率、转化率和购买数量。在示例性实施例中，上述步骤303之后还包括以下几个子步骤：
1、根据点击率和转化率，获取事件信息对应的用户账号针对关联物品的点击转化率；其中，关联物品即为上文中的目标物品；
和/或，
2、根据点击转化率和购买数量，获取用户账号针对关联物品的购买情况预测结果。
通过具体化预测目标，使得本申请实施例可以下沉到具体应用场景进行针对多个具体预测目标维度的预测结果，有效的提升了针对点击率、转化率、购买数量上的待预测事件的推荐效率。
示例性地，假设点击率为pCTR，转化率为pCVR，购买数量为n，则点击转化率pCTCVR为：
pCTCVR＝pCTR*pCVR；
购买情况预测结果P为：
P＝pCTR*Pcvr*n。
可选地，上述事件信息包括关联物品信息、关联用户信息和关联场景信息。服务器在获取上述事件信息之后，根据该事件信息对应的关联物品，获取关联物品信息；根据事件信息对应的用户账号，获取关联用户信息；获取关联场景信息。之后，服务器根据关联物品信息、关联用户信息和关联场景信息，获取上述事件信息，进一步地，对该事件信息进行编码，得到该事件信息对应的编码表示。
上述关联物品信息包括关联物品的历史购买情况。可选地，该历史购买情况可以是用户账号针对关联物品的历史购买情况，也可以是目标对象针对关联物品的历史购买情况，本申请实施例对此不作限定。其中，上述购买情况包括但不限于以下至少一项：是否购买过、购买时刻、购买次数、购买时的场景信息等。当然，在示例性实施例中，上述关联物品信息中还可以包括但不限于以下至少一项：关联物品的标识信息、关联物品的类型、关联物品的展示平台、关联物品的展示位置等。
上述关联用户信息用于反映用户账号对应的账号标签信息。其中，该账号标签信息可 以为服务器根据用户账号对应的各种用户数据绘制的，且用于反映用户账号的特征。可选地，上述用户数据包括但不限于以下至少一项：用户账号对应的用户年龄、用户账号对应的用户性别、用户账号对应的余额等。
上述关联场景信息用于指示操作行为发生环境。其中，操作行为发生环境即为当前的场景环境。可选地，该关联场景信息中包括但不限于以下至少一项：关联物品的展示平台的近期(今日、近期一周内、近期一个月内等)人流量、关联物品的展品平台的近期物品成交量、关联物品的展示平台的近期市场竞争力等。
通过对关联物品信息、关联用户信息和关联场景信息的获取，从而可以从多个维度全面的表征待预测事件的事件信息，从而事件信息的编码表示可以更为准确的表征待预测事件，有助于提高后续的预测结果的准确性。
综上所述，本申请实施例提供的技术方案，通过各个历史行为编码与预测目标之间的相关性，生成目标对象相对于不同预测目标分别对应的历史特征数据，由于目标对象和不同预测目标之间的相关性有所区别，故针对不同预测目标的历史特征数据也会不同。进而在待预测事件的事件信息的基础上，结合各个预测目标分别对应的历史特征数据，从各个预测目标的维度下，分别获取目标对象与待预测事件相关的预测结果，由于在获取预测结果时，考虑了目标对象在不同预测目标之间历史特征数据的差异，故获取的预测结果可以体现出目标对象针对待预测事件在不同预测目标下的区别，从而提高了预测结果的准确性和全面性。
下面，对历史特征数据的获取方式进行介绍。
在示例性实施例中，上述步骤302包括以下几个步骤：
1、获取历史行为编码序列中的各个历史行为编码分别对应的操作行为。
在本申请实施例中，服务器在获取上述历史特征数据时，获取山分数历史行为编码序列中的各个历史行为编码分别对应的操作行为。
2、对于多个预测目标中的每一个预测目标，分别获取各个操作行为与预测目标之间的相关度。
在本申请实施例中，服务器在获取上述各个历史行为编码分别对应的操作行为之后，对于多个预测目标中的每一个预测目标，分别获取各个操作行为与预测目标之间的相关度。可选地，该相关度用于量化表征上述相关性。
在一种可能的实施方式中，上述相关度用于指示操作行为与预测目标是否相关。示例性地，若操作行为与预测目标相关，则历史行为编码与预测目标之间的相关度为“1”；若操作行为与预测目标不相关，则历史行为编码与预测目标之间的相关度为“0”。其中，操作行为是否与预测目标相关，可以为预先设置的信息。例如，若预测目标为转化率，则可以预先设置相关的操作行为包括购买、收藏；若预测目标为点击率，则可以预先设置相关的操作行为包括点击、搜索。
在另一种可能的实施方式中，上述相关度用于指示操作行为与预测目标之间的相关程度。示例性地，若操作行为与预测目标强相关，则历史行为编码与预测目标之间的相关度为“1”；若操作行为与预测目标中相关，则历史行为编码与预测目标之间的相关度为“0.5”； 若操作行为与预测目标弱相关，则历史行为编码与预测目标之间的相关度为“0.1”。其中，操作行为与预测目标之间的相关程度，可以为预先设置的信息。例如，若预测目标为转化率，则可以预先设置相关的操作行为包括购买(相关度1)、收藏(相关度0.5)、点击(相关度0.3)、搜索(相关度0.1)；若预测目标为点击率，则可以预先设置相关的操作行为包括点击(相关度1)、搜索(相关度0.3)。
3、根据各个操作行为与预测目标之间的相关度，对各个历史行为编码进行加权处理，得到所述目标对象相对于预测目标的历史特征数据。
在本申请实施例中，服务器在获取上述相关度之后，根据各个操作行为与所述预测目标之间的相关度，对各个历史行为编码进行加权处理，得到预测目标对应的历史特征数据。通过细化到操作行为这一维度来确定历史行为编码分别对应预测目标间的相关度，可以有效的提高历史特征数据表征操作行为的能力，且能够着重体现出与操作行为的相关程度，提高了历史特征数据的准确性。
可选地，服务器在进行加权处理时，分别获取各个历史行为编码和事件信息对应的编码表示之间的相似度；进一步地，根据各个历史行为编码分别对应的相关度和相似度，确定各个历史行为编码分别对应的权重参数；之后，根据各个历史行为编码分别对应的权重参数，对各个历史行为编码进行加权处理，得到预测目标对应的历史特征数据。
由于不同的操作行为和用于表征待预测事件的事件信息间的关联程度有所区别，故为了提高后续确定预测结果的准确性，可以在确定目标对象对应不同预测目标的历史特征数据时，预先确定历史行为编码与事件信息间的相似度，从而得到更为准确的历史特征数据。
需要说明的一点是，上述权重参数与上述相关度呈正相关关系，即相关度越大，权重参数越大；而且。上述权重参数与上述相似度也呈正相关关系，即相似度越大，权重参数越大。
可选地，上述相关度与上述相似度均以具体数值的形式进行表示，而且，相关度的大小与数值大小呈正相关关系，相似度的大小与数值大小也呈正相关关系。可选地，服务器通过该相关度和该相似度之间的乘积来获取上述权重参数。在一种可能的实施方式中，服务器直接将该乘积作为上述权重参数；在另一种可能的实施方式中，服务器在获取上述乘积之后，对该乘积进行归一化处理以获取上述权重参数。
当然，在示例性实施例中，上述相关度与上述相似度也可以有其它表示形式，本申请实施例对此不作限定。例如，相关度的表示形式有极高相关度、高相关度、一般相关度、弱相关度、极弱相关度、无相关度等，相似度的表示形式有极高相似度、高相似度、一般相似度、弱相似度、极弱相似度、无相似度等。可选地，服务器在获取上述相关度和上述相似度之后，基于该相关度和该相似度，确定各个历史行为编码针对预测目标的重要程度，进而根据该重要程度确定各个历史行为编码分别对应的权重参数。其中，权重参数与上述重要程度呈正相关关系。
可选地，在上述相关度与上述相似度均以具体数值的形式进行表示的情况下，该相关度和该相似度之间的乘积可用于表示上述关联情况。在一种可能的实施方式中，直接将上述乘积来表示上述关联程度；在另一种可能的实施方式中，在获取上述乘积之后，将乘积 大于某个数值的历史行为编码确定为与预测目标关联的历史行为编码，该数值可以是任意数值，本申请实施例对此不作限定。
下面，对预测结果的获取方式进行介绍。
在示例性实施例中，上述步骤303包括以下几个步骤：
1、基于事件信息，提取与各个预测目标分别对应的事件特征数据。
在本申请实施例中，服务器在获取上述预测结果时，基于上述事件信息，提取与各个预测目标分别对应的事件特征数据。
可选地，服务器获取该事件信息对应的编码表示，并采用不同的专家网络分别对编码表示进行特征提取处理，得到特征提取结果集合。其中，该特征提取结果集合中包括来自不同专家网络的特征提取结果。进一步地，服务器分别采用不同的加权门，获取针对特征提取结果集合的多组加权参数。其中，每组加权参数中包括特征提取结果集合中的各个特征提取结果分别对应的加权参数，且加权门与上述预测目标之间具有一一对应的关系。之后，服务器对于每一个预测目标，基于多组加权参数中与预测目标对应的一组加权参数，对特征提取结果集合中的各个特征提取结果进行加权求和处理，得到预测目标对应的事件特征数据。
示例性地，假设预测目标包括点击率、转化率和购买数量，如图4所示，服务器在获取事件信息对应的编码表示之后，采用专家网络1、专家网络2和专家网络3分别对该编码表示进行特征特征提取处理，得到来自不同专家网络的特征提取结果。而且，由点击率加权门基于上述编码表示确定各个特征提取结果分别对应的第一加权参数，由转化率加权门基于上述编码表示确定各个特征提取结果分别对应的第二加权参数，由购买次数加权门基于上述编码表示确定各个特征提取结果分别对应的第三加权参数。之后，根据第一加权参数组、第二加权参数组、第三加权参数组分别对各个特征提取结果进行加权求和处理，得到各个预测目标分别对应的事件特征数据。
通过不同专家网络针对事件信息分别给出的特征提取结果所构成的特征提取结果集合，可以全面的表征事件信息的实质，从而提高了事件信息与各个预测目标分别对应的事件特征数据的准确性。
2、对于每一个预测目标，根据预测目标对应的历史特征数据和事件特征数据，获取目标对象在预测目标下对应待预测事件的预测结果。
在本申请实施例中，服务器在获取上述各个预测目标分别对应的事件特征数据之后，对于每一个预测目标，根据预测目标对应的历史特征数据和事件特征数据，获取预测目标对应的预测结果。
也就是说，在确定目标对象与待预测事件间的预测结果时，不仅要考虑目标对象与待预测事件间的关系，还可以进一步参考待预测事件与预测目标间的关系，通过引入事件特征数据，可以有效的表征上述参考待预测事件与预测目标间的关系，从而起到提高预测结果的准确性的作用。
可选地，对于每一个预测目标，服务器对预测目标对应的历史特征数据和事件特征数据进行融合处理，得到预测目标对应的融合特征数据；进一步地，通过该预测目标对应的 预测网络根据融合特征数据，生成预测目标对应的预测结果。示例性地，上述预测网络为Tower网络。
通过对历史特征数据和事件特征数据进行融合处理，可以在基于融合特征数据确定预测结果时更全面的共享历史特征数据和事件特征数据间的信息，提高预测结果的准确性。
下面，对历史行为编码序列的获取方式进行介绍。
1、根据历史行为数据序列，获取历史行为嵌入序列。
在本申请实施例中，服务器在获取上述历史行为编码序列时，先获取历史行为数据序列，进而根据该历史行为数据序列，获取历史行为嵌入向量。可选地，上述历史行为数据序列是与上述事件信息相关的数据序列，服务器在获取历史行为数据序列时，基于上述事件信息获取历史行为数据序列。
在一种可能的实施方式中，服务器基于事件信息所对应的用户账号获取历史行为数据序列。可选地，服务器在获取历史行为数据序列时，获取上述事件信息对应的用户账号，进而根据该用户账号确定目标对象。可选地，该目标对象包括用户账号和/或相似用户账号。其中，该相似用户账号与用户账号之间具有相似的用户特征。可选地，该相似的用户特征包括但不限于以下至少一项：用户之间的年龄差小于第一目标值、用户性别相同、用户购买过相似物品、用户购买过的相似物品之间的数量差小于第二目标值、用户浏览过相似视频等。
在另一种可能的实施方式中，服务器基于事件信息对应的目标物品获取历史行为数据。可选地，服务器在获取历史行为数据序列时，获取上述事件信息对应的目标物品，进而根据该目标物品确定历史行为数据所针对的物品。可选地，该物品包括目标物品和/或相似物品。其中，该相似物品与目标物品之间具有相似的物品特征。可选地，该相似的物品特征包括但不限于以下至少一项：物品类型相同、物品的展示平台相同、物品所针对的用户群体相同等。
当然，在其它可能的实施方式中，服务器也可以通过将事件信息所对应的用户账号和目标物品作为约束条件，来获取上述历史行为数据序列。示例性地，服务器根据该用户账号确定上述目标对象，进而以该目标对象对应的全量历史行为数据为范围，获取目标物品所指示的物品对应的历史行为数据，生成上述历史行为数据序列。
上述历史行为嵌入序列中的一项历史行为嵌入，是指历史行为数据序列中的一项历史行为数据对应的嵌入表示。可选地，每一项历史行为数据中包括历史物品信息、历史行为信息和历史场景信息。服务器在获取历史行为嵌入序列时，对于所述历史行为数据序列中的每一项历史行为数据，获取第一特征向量、第二特征向量和第三特征向量。其中，上述第一特征向量是指历史物品信息对应的特征向量，上述第二特征向量是指历史行为信息对应的特征向量，上述第三特征向量是指历史场景信息对应的特征向量。之后，服务器分别对各个历史行为数据对应的第一特征向量、第二特征向量和第三特征向量进行拼接处理，得到上述历史行为嵌入序列。
示例性地，如图5所示，历史行为数据中包括历史物品信息、历史行为信息和历史场景信息。其中，历史物品信息包括但不限于以下至少一项：物品标识、物品类型、物品热度 等，历史行为信息包括但不限于以下至少一项：行为类型、停留时长、申购金额等。历史场景信息包括但不限于以下至少一项：当前页面、当前时间、时长行情等。服务器在获取上述历史行为数据之后，对历史物品信息、历史行为信息和历史场景信息分别进行编码，以获取第一特征向量、第二特征向量和第三特征向量。之后，对第一特征向量、第二特征向量和第三特征向量进行拼接，以获取上述历史行为数据对应的历史行为嵌入。
通过从物品、场景和行为多个维度确定对应历史行为数据的特征向量，使得通过表征前述不同维度的特征向量所拼接得到历史行为嵌入具有更为全面的表征能力。
2、对历史行为嵌入序列中的各个历史行为嵌入分别进行编码处理，得到历史行为编码序列。
在本申请实施例中，服务器在获取上述历史行为嵌入序列之后，对历史行为嵌入序列中的各个历史行为嵌入分别进行编码处理，以获取上述历史行为编码序列。
通过对历史行为嵌入进行编码得到的历史行为编码，可以在保证表征精度的前提下有效降低数据量，从而节约了后续处理的资源占用。
可选地，服务器对于历史行为嵌入序列中的每一项历史行为嵌入，根据历史行为嵌入对应的操作行为的发生时刻，获取至少一项前序历史行为嵌入。其中，前序历史行为嵌入对应的操作行为的发生时刻，在历史行为嵌入对应的操作行为的发生时刻之前。之后，服务器分别对各个历史行为嵌入，以及历史行为嵌入对应的至少一项前序历史行为嵌入进行编码处理，获取历史行为编码序列。
示例性地，如图6所示，在历史行为嵌入序列中包括历史行为嵌入E 1、历史行为嵌入E 2……历史行为嵌入E n，且历史行为嵌入E 1对应的发生时刻最早，历史行为嵌入E n对应的发生时刻最晚。服务器在获取该历史行为嵌入序列之后，对于每一项历史行为嵌入，根据历史行为嵌入对应的发生时刻，获取至少一项前序历史行为嵌入，并通过神经网络分别对各个历史行为嵌入，以及历史行为嵌入对应的至少一项前序历史行为嵌入进行编码处理，获取历史行为编码序列。其中，该历史行为编码序列中包括历史行为编码T 1、历史行为编码T 2……历史行为编码T n。
通过操作行为的发生时刻可以获取历史行为嵌入之前的前序历史行为嵌入，并将前序历史行为嵌入用于对历史行为嵌入的编码中，从而相当于在历史行为编码中融入了之前的操作行为，相当于融入了上文信息，由此在历史行为编码增加了时序维度的信息，扩展了历史行为编码对应的观察视野。
另外，假设上述预测目标包括点击率、转化率和购买数量，结合参考图7，对本申请完整的多目标预测方法进行介绍。在服务器中，由图6所示的第二特征处理版块获取历史行为编码序列，其中，该第二特征处理版块中包括自注意力机制；由图4所示的第一特征处理版块通过事件信息对应的编码表示，获取点击率对应的事件特征数据、转化率对应的事件特征数据、购买数量对应的事件特征数据。而且，在获取历史行为编码序列之后，通过点击率注意力机制，结合事件信息对应的编码表示，获取点击率对应的历史特征数据；通过转化率注意力机制，结合事件信息对应的编码表示，获取转化率对应的历史特征数据；通过购买数量注意力机制，结合事件信息对应的编码表示，获取购买数量对应的历史特征数据。 之后，服务器通过点击率对应的预测网络，根据点击率对应的历史特征数据和点击率对应的事件特征数据，获取点击率对应的预测结果；通过转化率对应的预测网络，根据转化率对应的历史特征数据和转化率对应的事件特征数据，获取转化率对应的预测结果；通过购买数量对应的预测网络，根据购买数量对应的历史特征数据和购买数量对应的事件特征数据，获取购买数量对应的预测结果。最终，服务器可以根据点击率对应的预测结果和转化率对应的预测结果，获取点击转化率对应的预测结果；并且，根据点击转化率对应的预测结果和购买数量对应的预测结果获取购买情况预测结果。
下述为本申请装置实施例，可以用于执行本申请方法实施例。对于本申请装置实施例中未披露的细节，请参照本申请方法实施例。
请参考图8，其示出了本申请一个实施例提供的多目标预测装置的框图。该装置具有实现上述多目标预测方法的功能，所述功能可以由硬件实现，也可以由硬件执行相应的软件实现。该装置可以是服务器，也可以设置在服务器中。该装置800可以包括：编码序列获取模块810、特征数据生成模块820和预测结果获取模块830。
编码序列获取模块810，用于根据目标对象的历史行为数据序列，获取历史行为编码序列；其中，所述历史行为编码序列中的一项历史行为编码，是所述历史行为数据序列中的一项历史行为数据对应的编码表示。
特征数据生成模块820，用于对于多个预测目标中的每一个预测目标，根据所述历史行为编码序列中的各个历史行为编码与所述预测目标之间的相关性，生成所述目标对象相对于所述预测目标的历史特征数据。
预测结果获取模块830，用于根据待预测事件的事件信息，以及各个所述预测目标分别对应的历史特征数据，获取所述目标对象在各个所述预测目标下分别对应所述待预测事件的预测结果。
在示例性实施例中，如图9所示，所述特征数据生成模块820，包括：操作行为获取单元821、相关性获取单元822和特征数据获取单元823。
操作行为获取单元821，用于获取所述历史行为编码序列中的各个历史行为编码分别对应的操作行为。
相关性获取单元822，用于对于所述多个预测目标中的每一个预测目标，分别获取各个所述操作行为与所述预测目标之间的相关度，所述相关度用于量化表征所述相关性。
特征数据获取单元823，用于根据各个所述操作行为与所述预测目标之间的相关度，对各个所述历史行为编码进行加权处理，得到所述目标对象相对于所述预测目标的历史特征数据。
在示例性实施例中，所述特征数据获取单元823，用于分别获取各个所述历史行为编码和所述事件信息对应的编码表示之间的相似度；根据各个所述历史行为编码分别对应的相关度和相似度，确定各个所述历史行为编码分别对应的权重参数；根据各个所述历史行为编码分别对应的权重参数，对各个所述历史行为编码进行加权处理，得到所述预测目标对应的历史特征数据。
在示例性实施例中，如图9所示，所述预测结果获取模块830，包括：事件特征获取单 元831和预测结果获取单元832。
事件特征获取单元831，用于基于所述事件信息，提取与各个所述预测目标分别对应的事件特征数据。
预测结果获取单元832，用于对于每一个预测目标，根据所述预测目标对应的历史特征数据和事件特征数据，获取所述目标对象在所述预测目标对应所述待预测事件的预测结果。
在示例性实施例中，所述事件特征获取单元831，用于获取所述事件信息对应的编码表示；采用不同的专家网络分别对所述编码表示进行特征提取处理，得到特征提取结果集合；其中，所述特征提取结果集合中包括来自不同专家网络的特征提取结果；分别采用不同的加权门，获取针对所述特征提取结果集合的多组加权参数；其中，每组加权参数中包括所述特征提取结果集合中的各个特征提取结果分别对应的加权参数，所述加权门与所述预测目标之间具有一一对应的关系；对于每一个预测目标，基于所述多组加权参数中与所述预测目标对应的一组加权参数，对所述特征提取结果集合中的各个特征提取结果进行加权求和处理，得到所述预测目标对应的事件特征数据。
在示例性实施例中，所述预测结果获取单元832，用于对于每一个预测目标，对所述预测目标对应的历史特征数据和事件特征数据进行融合处理，得到所述预测目标对应的融合特征数据；通过所述预测目标对应的预测网络根据所述融合特征数据，生成所述目标对象在各个所述预测目标下分别对应所述待预测事件的预测结果。
在示例性实施例中，如图9所示，所述编码序列获取模块810，包括：嵌入序列获取单元811和编码序列获取单元812。
嵌入序列获取单元811，用于根据所述历史行为数据序列，获取历史行为嵌入序列；其中，所述历史行为嵌入序列中的一项历史行为嵌入，是所述历史行为数据序列中的一项历史行为数据对应的嵌入表示。
编码序列获取单元812，用于对所述历史行为嵌入序列中的各个所述历史行为嵌入分别进行编码处理，得到所述历史行为编码序列。
在示例性实施例中，所述编码序列获取单元812，用于对于所述历史行为嵌入序列中的每一项历史行为嵌入，根据所述历史行为嵌入对应的操作行为的发生时刻，获取至少一项前序历史行为嵌入；其中，所述前序历史行为嵌入对应的操作行为的发生时刻，在所述历史行为嵌入对应的操作行为的发生时刻之前；分别对各个所述历史行为嵌入，以及所述历史行为嵌入对应的所述至少一项前序历史行为嵌入进行编码处理，获取所述历史行为编码序列。
在示例性实施例中，每一项历史行为数据中包括历史物品信息、历史行为信息和历史场景信息；所述嵌入序列获取单元811，用于对于所述历史行为数据序列中的每一项历史行为数据，获取第一特征向量、第二特征向量和第三特征向量；其中，所述第一特征向量是指所述历史物品信息对应的特征向量，所述第二特征向量是指所述历史行为信息对应的特征向量，所述第三特征向量是指所述历史场景信息对应的特征向量；分别对各个所述历史行为数据对应的第一特征向量、第二特征向量和第三特征向量进行拼接处理，得到所述历史行为嵌入序列。
在示例性实施例中，所述事件信息包括关联物品信息、关联用户信息和关联场景信息；如图9所示，所述装置800还包括：事件编码获取模块840。
事件编码获取模块840，用于根据所述事件信息对应的关联物品，获取所述关联物品信息，所述关联物品信息包括所述关联物品的历史购买情况；根据所述事件信息对应的用户账号，获取所述关联用户信息，所述关联用户信息用于反映所述用户账号对应的账号标签信息；获取关联场景信息，所述关联场景信息用于指示操作行为发生环境；根据所述关联物品信息、所述关联用户信息和所述关联场景信息，获取所述事件信息；对所述事件信息进行编码，得到所述事件信息对应的编码表示。
在示例性实施例中，所述预测结果包括点击率、转化率和购买数量；如图9所示，所述装置800还包括：预测结果处理模块850。
预测结果处理模块850，用于根据所述点击率和所述转化率，获取所述事件信息对应的用户账号针对关联物品的点击转化率；和/或，根据所述点击转化率和所述购买数量，获取所述用户账号针对所述关联物品的购买情况预测结果。
综上所述，本申请实施例提供的技术方案，通过各个历史行为编码与预测目标之间的相关性，生成预测目标对应的历史特征数据，进而在待预测事件的事件信息的基础上，分别结合各个预测目标分别对应的历史特征数据，分别取各个所述预测目标分别对应的预测结果，在获取预测结果时，考虑了不同预测目标之间的差异，对于不同的预测目标根据不同的历史特征数据来获取预测结果，提高预测结果的准确性。
请参考图10，其示出了本申请一个实施例提供的服务器的结构框图。该服务器可用于实现上述多目标预测方法的功能。具体来讲：
服务器1000包括中央处理单元(Central Processing Unit，CPU)1001、包括随机存取存储器(Random Access Memory，RAM)1002和只读存储器(Read Only Memory，ROM)1003的系统存储器1004，以及连接系统存储器1004和中央处理单元1001的系统总线1005。服务器1000还包括帮助计算机内的各个器件之间传输信息的基本输入/输出(Input/Output，I/O)系统1006，和用于存储操作系统1013、应用程序1014和其他程序模块1015的大容量存储设备1007。
基本输入/输出系统1006包括有用于显示信息的显示器1008和用于用户输入信息的诸如鼠标、键盘之类的输入设备1009。其中显示器1008和输入设备1009都通过连接到系统总线1005的输入输出控制器1010连接到中央处理单元1001。基本输入/输出系统1006还可以包括输入输出控制器1010以用于接收和处理来自键盘、鼠标、或电子触控笔等多个其他设备的输入。类似地，输入输出控制器1010还提供输出到显示屏、打印机或其他类型的输出设备。
大容量存储设备1007通过连接到系统总线1005的大容量存储控制器(未示出)连接到中央处理单元1001。大容量存储设备1007及其相关联的计算机可读介质为服务器1000提供非易失性存储。
不失一般性，计算机可读介质可以包括计算机存储介质和通信介质。计算机存储介质包括以用于存储诸如计算机可读指令、数据结构、程序模块或其他数据等信息的任何方法 或技术实现的易失性和非易失性、可移动和不可移动介质。计算机存储介质包括RAM、ROM、EPROM(Erasable Programmable Read Only Memory，可擦除可编程只读存储器)、EEPROM(Electrically Erasable Programmable Read Only Memory，电可擦可编程只读存储器)、闪存或其他固态存储其技术，CD-ROM、DVD(Digital Video Disc，高密度数字视频光盘)或其他光学存储、磁带盒、磁带、磁盘存储或其他磁性存储设备。当然，本领域技术人员可知计算机存储介质不局限于上述几种。上述的系统存储器1004和大容量存储设备1007可以统称为存储器。
根据本申请的各种实施例，服务器1000还可以通过诸如因特网等网络连接到网络上的远程计算机运行。也即服务器1000可以通过连接在系统总线1005上的网络接口单元1011连接到网络1012，或者说，也可以使用网络接口单元1011来连接到其他类型的网络或远程计算机系统(未示出)。
所述存储器还包括计算机程序，该计算机程序存储于存储器中，且经配置以由一个或者一个以上处理器执行，以实现上述多目标预测方法。
在示例性实施例中，还提供了一种计算机可读存储介质，所述存储介质中存储有至少一条指令、至少一段程序、代码集或指令集，所述至少一条指令、所述至少一段程序、所述代码集或所述指令集在被处理器执行时以实现上述多目标预测方法。
可选地，该计算机可读存储介质可以包括：ROM(Read Only Memory，只读存储器)、RAM(Random Access Memory，随机存取存储器)、SSD(Solid State Drives，固态硬盘)或光盘等。其中，随机存取记忆体可以包括ReRAM(Resistance Random Access Memory，电阻式随机存取记忆体)和DRAM(Dynamic Random Access Memory，动态随机存取存储器)。
在示例性实施例中，提供了一种计算机程序产品或计算机程序，该计算机程序产品或计算机程序包括计算机指令，该计算机指令存储在计算机可读存储介质中。服务器的处理器从计算机可读存储介质读取该计算机指令，处理器执行该计算机指令，使得该服务器执行上述多目标预测方法。
应当理解的是，在本文中提及的“多个”是指两个或两个以上。“和/或”，描述关联对象的关联关系，表示可以存在三种关系，例如，A和/或B，可以表示：单独存在A，同时存在A和B，单独存在B这三种情况。字符“/”一般表示前后关联对象是一种“或”的关系。另外，本文中描述的步骤编号，仅示例性示出了步骤间的一种可能的执行先后顺序，在一些其它实施例中，上述步骤也可以不按照编号顺序来执行，如两个不同编号的步骤同时执行，或者两个不同编号的步骤按照与图示相反的顺序执行，本申请实施例对此不作限制。
以上所述仅为本申请的示例性实施例，并不用以限制本申请，凡在本申请的精神和原则之内，所作的任何修改、等同替换、改进等，均应包含在本申请的保护范围之内。

Claims (15)

  1. 一种多目标预测方法，所述方法由服务器执行，所述方法包括：
    根据目标对象的历史行为数据序列，获取历史行为编码序列；其中，所述历史行为编码序列中的一项历史行为编码，是所述历史行为数据序列中的一项历史行为数据对应的编码表示；
    对于多个预测目标中的每一个预测目标，根据所述历史行为编码序列中的各个历史行为编码与所述预测目标之间的相关性，生成所述目标对象相对于所述预测目标的历史特征数据；
    根据待预测事件的事件信息，以及各个所述预测目标分别对应的历史特征数据，获取所述目标对象在各个所述预测目标下分别对应所述待预测事件的预测结果。
  2. 根据权利要求1所述的方法，所述对于多个预测目标中的每一个预测目标，根据所述历史行为编码序列中的各个历史行为编码与所述预测目标之间的相关性，生成所述目标对象相对于所述预测目标的历史特征数据，包括：
    获取所述历史行为编码序列中的各个历史行为编码分别对应的操作行为；
    对于所述多个预测目标中的每一个预测目标，分别获取各个所述操作行为与所述预测目标之间的相关度，所述相关度用于量化表征所述相关性；
    根据各个所述操作行为与所述预测目标之间的相关度，对各个所述历史行为编码进行加权处理，得到所述目标对象相对于所述预测目标的历史特征数据。
  3. 根据权利要求2所述的方法，所述根据各个所述操作行为与所述预测目标之间的相关度，对各个所述历史行为编码进行加权处理，得到所述目标对象相对于所述预测目标的历史特征数据，包括：
    分别获取各个所述历史行为编码和所述事件信息对应的编码表示之间的相似度；
    根据各个所述历史行为编码分别对应的相关度和相似度，确定各个所述历史行为编码分别对应的权重参数；
    根据各个所述历史行为编码分别对应的权重参数，对各个所述历史行为编码进行加权处理，得到所述预测目标对应的历史特征数据。
  4. 根据权利要求1所述的方法，所述根据待预测事件的事件信息，以及各个所述预测目标分别对应的历史特征数据，获取所述目标对象在各个所述预测目标下分别对应所述待预测事件的预测结果，包括：
    基于所述事件信息，提取与各个所述预测目标分别对应的事件特征数据；
    对于每一个预测目标，根据所述预测目标对应的历史特征数据和事件特征数据，获取所述目标对象在所述预测目标下对应所述待预测事件的预测结果。
  5. 根据权利要求4所述的方法，所述基于所述事件信息，提取与各个所述预测目标分别对应的事件特征数据，包括：
    获取所述事件信息对应的编码表示；
    采用不同的专家网络分别对所述编码表示进行特征提取处理，得到特征提取结果集合；其中，所述特征提取结果集合中包括来自不同专家网络的特征提取结果；
    分别采用不同的加权门，获取针对所述特征提取结果集合的多组加权参数；其中，每组加权参数中包括所述特征提取结果集合中的各个特征提取结果分别对应的加权参数，所述加权门与所述预测目标之间具有一一对应的关系；
    对于每一个预测目标，基于所述多组加权参数中与所述预测目标对应的一组加权参数，对所述特征提取结果集合中的各个特征提取结果进行加权求和处理，得到所述预测目标对应的事件特征数据。
  6. 根据权利要求4所述的方法，所述对于每一个预测目标，根据所述预测目标对应的历史特征数据和事件特征数据，获取所述目标对象在各个所述预测目标下分别对应所述待预测事件的预测结果，包括：
    对于每一个预测目标，对所述预测目标对应的历史特征数据和事件特征数据进行融合处理，得到所述预测目标对应的融合特征数据；
    通过所述预测目标对应的预测网络根据所述融合特征数据，生成所述目标对象在各个所述预测目标下分别对应所述待预测事件的预测结果。
  7. 根据权利要求1所述的方法，所述根据目标对象的历史行为数据序列，获取历史行为编码序列，包括：
    根据所述历史行为数据序列，获取历史行为嵌入序列；其中，所述历史行为嵌入序列中的一项历史行为嵌入，是所述历史行为数据序列中的一项历史行为数据对应的嵌入表示；
    对所述历史行为嵌入序列中的各个所述历史行为嵌入分别进行编码处理，得到所述历史行为编码序列。
  8. 根据权利要求7所述的方法，所述对所述历史行为嵌入序列中的各个所述历史行为嵌入分别进行编码处理，得到所述历史行为编码序列，包括：
    对于所述历史行为嵌入序列中的每一项历史行为嵌入，根据所述历史行为嵌入对应的操作行为的发生时刻，获取至少一项前序历史行为嵌入；其中，所述前序历史行为嵌入对应的操作行为的发生时刻，在所述历史行为嵌入对应的操作行为的发生时刻之前；
    分别对各个所述历史行为嵌入，以及所述历史行为嵌入对应的所述至少一项前序历史行为嵌入进行编码处理，获取所述历史行为编码序列。
  9. 根据权利要求7所述的方法，每一项历史行为数据中包括历史物品信息、历史行为信息和历史场景信息；
    所述根据所述历史行为数据序列，获取历史行为嵌入序列，包括：
    对于所述历史行为数据序列中的每一项历史行为数据，获取第一特征向量、第二特征向量和第三特征向量；其中，所述第一特征向量是指所述历史物品信息对应的特征向量，所述第二特征向量是指所述历史行为信息对应的特征向量，所述第三特征向量是指所述历史场景信息对应的特征向量；
    分别对各个所述历史行为数据对应的第一特征向量、第二特征向量和第三特征向量进行拼接处理，得到所述历史行为嵌入序列。
  10. 根据权利要求1所述的方法，所述事件信息包括关联物品信息、关联用户信息和关联场景信息；所述方法还包括：
    根据所述事件信息对应的关联物品，获取所述关联物品信息，所述关联物品信息包括所述关联物品的历史购买情况；
    根据所述事件信息对应的用户账号，获取所述关联用户信息，所述关联用户信息用于反映所述用户账号对应的账号标签信息；
    获取关联场景信息，所述关联场景信息用于指示操作行为发生环境；
    根据所述关联物品信息、所述关联用户信息和所述关联场景信息，获取所述事件信息；
    对所述事件信息进行编码，得到所述事件信息对应的编码表示。
  11. 根据权利要求1至10任一项所述的方法，所述预测结果包括点击率、转化率和购买数量；
    所述根据待预测事件的事件信息，以及各个所述预测目标分别对应的历史特征数据，获取所述目标对象在各个所述预测目标下分别对应所述待预测事件的预测结果之后，还包括：
    根据所述点击率和所述转化率，获取所述事件信息对应的用户账号针对关联物品的点击转化率；
    和/或，
    根据所述点击转化率和所述购买数量，获取所述用户账号针对所述关联物品的购买情况预测结果。
  12. 一种多目标预测装置，所述装置包括：
    编码序列获取模块，用于根据目标对象的历史行为数据序列，获取历史行为编码序列；其中，所述历史行为编码序列中的一项历史行为编码，是所述历史行为数据序列中的一项历史行为数据对应的编码表示；
    特征数据生成模块，用于对于多个预测目标中的每一个预测目标，根据所述历史行为编码序列中的各个历史行为编码与所述预测目标之间的相关性，生成所述目标对象相对于所述预测目标的历史特征数据；
    预测结果获取模块，用于根据待预测事件的事件信息，以及各个所述预测目标分别对应的历史特征数据，获取所述目标对象在各个所述预测目标下分别对应所述待预测事件的预测结果。
  13. 一种服务器，所述服务器包括处理器和存储器，所述存储器中存储有至少一条指令、至少一段程序、代码集或指令集，所述至少一条指令、所述至少一段程序、所述代码集或指令集由所述处理器加载并执行以实现如权利要求1至11任一项所述的多目标预测方法。
  14. 一种计算机可读存储介质，所述计算机可读存储介质中存储有至少一条指令、至少一段程序、代码集或指令集，所述至少一条指令、所述至少一段程序、所述代码集或指令集由处理器加载并执行以实现如权利要求1至11任一项所述的多目标预测方法。
  15. 一种包括指令的计算机程序产品，当其在计算机上运行时，使得所述计算机执行权利要求1至11任一项所述的多目标预测方法。
PCT/CN2022/104024 2021-08-09 2022-07-06 多目标预测方法、装置、设备、存储介质及程序产品 Ceased WO2023016147A1 (zh)

Priority Applications (1)

Application Number Priority Date Filing Date Title
US18/137,993 US20230259959A1 (en) 2021-08-09 2023-04-21 Multi-target prediction method and apparatus, device, storage medium and program product

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN202110907940.9A CN115705583A (zh) 2021-08-09 2021-08-09 多目标预测方法、装置、设备及存储介质
CN202110907940.9 2021-08-09

Related Child Applications (1)

Application Number Title Priority Date Filing Date
US18/137,993 Continuation US20230259959A1 (en) 2021-08-09 2023-04-21 Multi-target prediction method and apparatus, device, storage medium and program product

Publications (1)

Publication Number Publication Date
WO2023016147A1 true WO2023016147A1 (zh) 2023-02-16

Family

ID=85179271

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2022/104024 Ceased WO2023016147A1 (zh) 2021-08-09 2022-07-06 多目标预测方法、装置、设备、存储介质及程序产品

Country Status (3)

Country Link
US (1) US20230259959A1 (zh)
CN (1) CN115705583A (zh)
WO (1) WO2023016147A1 (zh)

Cited By (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN116383014A (zh) * 2023-06-02 2023-07-04 北京集度科技有限公司 监控计算机、软件调试方法、软件调试系统、介质及产品
CN117077789A (zh) * 2023-08-15 2023-11-17 合肥工业大学 推理个体与群体关联度的方法、装置、电子设备及介质
CN117290599A (zh) * 2023-09-25 2023-12-26 中国联合网络通信集团有限公司 预测模型的生成方法、物品推荐方法、装置、设备和介质
CN117392396A (zh) * 2023-12-08 2024-01-12 安徽蔚来智驾科技有限公司 跨模态目标状态的检测方法、设备、智能设备和介质
CN118627674A (zh) * 2024-06-11 2024-09-10 东南大学 类别和时间敏感的事件序列预测方法、系统、介质及设备

Families Citing this family (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN117540136B (zh) * 2024-01-10 2025-03-28 腾讯科技（深圳）有限公司 异常信号检测方法、装置、设备及存储介质
CN117875522A (zh) * 2024-03-12 2024-04-12 之江实验室 一种事件数量预测的方法、装置、存储介质、设备
CN119254632B (zh) * 2024-12-03 2025-03-28 长沙瑞腾信息技术有限公司 一种数据分布式存储方法和计算机设备
CN120263282B (zh) * 2025-06-03 2025-08-15 新华三人工智能科技有限公司 一种数据预测方法、装置及设备

Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20110099136A1 (en) * 2009-10-23 2011-04-28 Gm Global Technology Operations, Inc. Method and system for concurrent event forecasting
CN111723910A (zh) * 2020-06-17 2020-09-29 腾讯科技（北京）有限公司 构建多任务学习模型的方法、装置、电子设备及存储介质
CN112328646A (zh) * 2021-01-04 2021-02-05 平安科技（深圳）有限公司 多任务课程推荐方法、装置、计算机设备及存储介质
US20210064998A1 (en) * 2019-08-27 2021-03-04 Nec Laboratories America, Inc. Tensorized lstm with adaptive shared memory for learning trends in multivariate time series
CN112883257A (zh) * 2021-01-11 2021-06-01 北京达佳互联信息技术有限公司 行为序列数据处理方法、装置、电子设备及存储介质
CN112905648A (zh) * 2021-02-04 2021-06-04 北京邮电大学 一种基于多任务学习的多目标推荐方法及系统

Family Cites Families (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110008974A (zh) * 2018-11-23 2019-07-12 阿里巴巴集团控股有限公司 行为数据预测方法、装置、电子设备及计算机存储介质
CN112905876B (zh) * 2020-03-16 2024-03-15 腾讯科技（深圳）有限公司 基于深度学习的信息推送方法、装置和计算机设备
CN112381607B (zh) * 2020-11-12 2023-11-24 杭州时趣信息技术有限公司 一种网络商品排序方法、装置、设备及介质
CN112232510A (zh) * 2020-12-14 2021-01-15 蚂蚁智信(杭州)信息技术有限公司 多目标推荐模型的训练及信息推荐方法以及装置

Patent Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20110099136A1 (en) * 2009-10-23 2011-04-28 Gm Global Technology Operations, Inc. Method and system for concurrent event forecasting
US20210064998A1 (en) * 2019-08-27 2021-03-04 Nec Laboratories America, Inc. Tensorized lstm with adaptive shared memory for learning trends in multivariate time series
CN111723910A (zh) * 2020-06-17 2020-09-29 腾讯科技（北京）有限公司 构建多任务学习模型的方法、装置、电子设备及存储介质
CN112328646A (zh) * 2021-01-04 2021-02-05 平安科技（深圳）有限公司 多任务课程推荐方法、装置、计算机设备及存储介质
CN112883257A (zh) * 2021-01-11 2021-06-01 北京达佳互联信息技术有限公司 行为序列数据处理方法、装置、电子设备及存储介质
CN112905648A (zh) * 2021-02-04 2021-06-04 北京邮电大学 一种基于多任务学习的多目标推荐方法及系统

Cited By (8)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN116383014A (zh) * 2023-06-02 2023-07-04 北京集度科技有限公司 监控计算机、软件调试方法、软件调试系统、介质及产品
CN116383014B (zh) * 2023-06-02 2023-08-01 北京集度科技有限公司 监控计算机、软件调试方法、软件调试系统、介质及产品
CN117077789A (zh) * 2023-08-15 2023-11-17 合肥工业大学 推理个体与群体关联度的方法、装置、电子设备及介质
CN117290599A (zh) * 2023-09-25 2023-12-26 中国联合网络通信集团有限公司 预测模型的生成方法、物品推荐方法、装置、设备和介质
CN117392396A (zh) * 2023-12-08 2024-01-12 安徽蔚来智驾科技有限公司 跨模态目标状态的检测方法、设备、智能设备和介质
CN117392396B (zh) * 2023-12-08 2024-03-05 安徽蔚来智驾科技有限公司 跨模态目标状态的检测方法、设备、智能设备和介质
CN118627674A (zh) * 2024-06-11 2024-09-10 东南大学 类别和时间敏感的事件序列预测方法、系统、介质及设备
CN118627674B (zh) * 2024-06-11 2025-07-18 东南大学 类别和时间敏感的事件序列预测方法、系统、介质及设备

Also Published As

Publication number Publication date
CN115705583A (zh) 2023-02-17
US20230259959A1 (en) 2023-08-17

Similar Documents

Publication Publication Date Title
WO2023016147A1 (zh) 多目标预测方法、装置、设备、存储介质及程序产品
CN110413877B (zh) 一种资源推荐方法、装置及电子设备
CN110334289B (zh) 出行目的地的确定方法和目标用户的确定方法
CN106649681B (zh) 一种数据处理方法、装置及设备
CN110909182A (zh) 多媒体资源搜索方法、装置、计算机设备及存储介质
CN114219664B (zh) 产品推荐方法、装置、计算机设备及存储介质
CN113836391B (zh) 资源推荐方法、装置、计算机设备及存储介质
CN111369293A (zh) 一种广告出价方法、装置及电子设备
CN109426894A (zh) 用户信息共享、竞价方法、装置、系统及电子设备
CN111369080A (zh) 一种智能客服解决率预测方法和系统以及多业务预测模型
CN113297462A (zh) 数据处理方法、装置、设备和存储介质
CN119646316B (zh) 一种用户偏好分析方法及系统
CN116703555A (zh) 一种预警方法、装置、电子设备及计算机可读介质
CN114943279A (zh) 招投标合作关系的预测方法、设备及系统
CN113836390B (zh) 资源推荐方法、装置、计算机设备及存储介质
CN115017362A (zh) 数据处理方法、电子设备及存储介质
CN115271769A (zh) 投放效果数据的预估方法、装置及设备
CN115329183A (zh) 数据处理方法、装置、存储介质及设备
CN113919866A (zh) 模型训练方法、广告投放方法、装置、设备及存储介质
CN111582901B (zh) 一种展示链接效果的量化方法、装置、系统及存储介质
CN117853247A (zh) 基于人工智能的产品推荐方法、装置、设备及存储介质
CN110807313A (zh) 预估文本阅读时间的方法、装置、电子设备及存储介质
HK40080489A (zh) 多目标预测方法、装置、设备及存储介质
CN111666250B (zh) 一种用于处理阅读应用中书籍催更请求信息的方法与设备
CN115130756A (zh) 一种在线服务的管理方法、装置、电子设备及存储介质

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: 22855131

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

32PN Ep: public notification in the ep bulletin as address of the adressee cannot be established

Free format text: NOTING OF LOSS OF RIGHTS PURSUANT TO RULE 112(1) EPC (EPO FORM 1205A DATED 19.06.2024)

122 Ep: pct application non-entry in european phase

Ref document number: 22855131

Country of ref document: EP

Kind code of ref document: A1