WO2024104241A1 - 一种基于模型隐式多目标融合的消息推送方法及装置 - Google Patents

一种基于模型隐式多目标融合的消息推送方法及装置 Download PDF

Info

Publication number
WO2024104241A1
WO2024104241A1 PCT/CN2023/130616 CN2023130616W WO2024104241A1 WO 2024104241 A1 WO2024104241 A1 WO 2024104241A1 CN 2023130616 W CN2023130616 W CN 2023130616W WO 2024104241 A1 WO2024104241 A1 WO 2024104241A1
Authority
WO
WIPO (PCT)
Prior art keywords
model
sample data
target sub
label
module
Prior art date
Application number
PCT/CN2023/130616
Other languages
English (en)
French (fr)
Inventor
张金鹏
王耀宣
费浩峻
Original Assignee
上海淇玥信息技术有限公司
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 上海淇玥信息技术有限公司 filed Critical 上海淇玥信息技术有限公司
Publication of WO2024104241A1 publication Critical patent/WO2024104241A1/zh

Links

Classifications

    • 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
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/25—Fusion techniques
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
    • Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
    • Y02D—CLIMATE CHANGE MITIGATION TECHNOLOGIES IN INFORMATION AND COMMUNICATION TECHNOLOGIES [ICT], I.E. INFORMATION AND COMMUNICATION TECHNOLOGIES AIMING AT THE REDUCTION OF THEIR OWN ENERGY USE
    • Y02D30/00—Reducing energy consumption in communication networks
    • Y02D30/70—Reducing energy consumption in communication networks in wireless communication networks

Definitions

  • the present invention relates to the technical field of intelligent message push, and in particular to a message push method, device, electronic device and computer-readable medium based on model implicit multi-target fusion.
  • the message push will integrate the information of each target sub-process to determine the push strategy. For example, if the message push is expected to have both high clicks and high conversions, it is necessary to combine and optimize the business goals of each target sub-process.
  • a model is first built for the business goal of a single target sub-process, and then a linear combination is made for the prediction results of multiple business goals, and the push strategy is optimized through the combination results; or, prediction and optimization are performed based on the data of each target sub-process through multi-target training methods and models.
  • these methods do not take into account: 1.
  • the sample size corresponding to some business indicators may be relatively sparse, and there will be sample imbalance problems during joint training; 2. It is impossible to directly end-to-end (End2End) model the scores that integrate multiple business goals; 3. It is impossible to reflect the post-business goals as a priori in the modeling. Therefore, it will lead to poor message push effects, and the push system is complex and has large computational requirements.
  • the main purpose of the present invention is to propose a message push method, device, electronic device and computer-readable medium based on model implicit multi-target fusion, in order to at least partially solve at least one of the above-mentioned technical problems.
  • the first aspect of the present invention proposes a message push method based on model implicit multi-target fusion, the method comprising:
  • each target sub-process is arranged in chronological order, and each target sub-process is a post-process of the previous target sub-process and/or a pre-process of the next target sub-process;
  • the first model is trained according to the sample data of the post-process, and the second model is trained according to the sample data of the pre-process, and the sample information of the first model is implicitly integrated into the second model;
  • the sample data of the pre-process includes: hard labels and soft labels output by the first model;
  • obtaining sample data of a single target sub-process includes:
  • the sample data is labeled according to the triggering moment
  • the sample data is labeled with a hard label and a soft label according to the triggering moment.
  • the hard label of the sample data corresponding to the triggering moment is marked as 1, and the soft label softlabel is:
  • T is the model parameter
  • x is the predicted hard label corresponding to the sample data
  • the hard label of the sample data corresponding to the non-trigger moment is marked as 0, and the soft label is 0.
  • training the second model according to the sample data of the previous process includes:
  • the parameters of the second model are updated based on the first loss function and the second loss function.
  • the second aspect of the present invention provides a message push device based on model implicit multi-target fusion, the device comprising:
  • the acquisition module is used to acquire sample data of each target sub-process of the message push; each target sub-process is arranged in chronological order, and each target sub-process is a post-process of the previous target sub-process and/or a pre-process of the next target sub-process;
  • a training module used to train a first model according to sample data of a post-process, train a second model according to sample data of a pre-process, and implicitly integrate sample information of the first model into the second model;
  • the sample data of the pre-process includes: hard labels and soft labels output by the first model;
  • a prediction module used for inputting the prediction data of the previous process into the second model to obtain a prediction result
  • the push module is used to push messages according to the prediction results.
  • the acquisition module includes:
  • a sub-acquisition module used to acquire sample data within the sliding window
  • a detection module used for detecting the triggering moment of the target sub-process event in the sliding window
  • a labeling module used to label the sample data according to the location of the target sub-process and the triggering time
  • the update module is used to update the sliding window.
  • the labeling module labels the sample data according to the triggering moment
  • the labeling module labels the sample data with a hard label and a soft label according to the triggering moment.
  • the labeling module labels the hard label of the sample data corresponding to the triggering moment as 1, and the soft label softlabel is:
  • T is the model parameter
  • x is the predicted hard label corresponding to the sample data
  • the hard label of the sample data corresponding to the non-trigger moment is marked as 0, and the soft label is 0.
  • the training module includes:
  • a first configuration module used to configure the model parameters of the first model and the second model to be the same
  • a first input module used to input sample data of the previous process into the second model to obtain a predicted soft label
  • a first calculation module used to calculate a first loss function according to the marked soft label and the predicted soft label
  • a second configuration module used to configure a model parameter of the second model to be 1;
  • a second input module used to input the sample data of the previous process into the second model to obtain a predicted hard label
  • a second calculation module used to calculate a second loss function according to the marked hard label and the predicted hard label
  • An updating module is used to update the parameters of the second model based on the first loss function and the second loss function.
  • the third aspect of the present invention provides an electronic device, including:
  • a memory storing computer executable instructions, which when executed cause the processor to perform any of the methods described above.
  • the fourth aspect of the present invention provides a computer-readable storage medium, wherein the computer-readable storage medium stores one or more programs, and when the one or more programs are executed by a processor, the above method is implemented.
  • the present invention obtains sample data of each target sub-process for message push; wherein: each target sub-process is a post-process of a previous target sub-process and/or a pre-process of a subsequent target sub-process; a first model is trained according to sample data of the post-process, and a second model is trained according to sample data of the pre-process; wherein: the sample data of the pre-process includes: hard labels and soft labels output by the first model, so that the sample information of the first model is implicitly integrated into the second model, and through this training
  • the training method can avoid the problem of sample imbalance among multiple business objectives to a certain extent, and realize End2End prediction at the same time, reducing the complexity of the push system; and the prediction value of the second model obtained by training directly feeds back the comprehensive value after weighing multiple business objectives, which can guide message push from multiple business objectives, thereby improving the effect of message push.
  • FIG1 is a flow chart of a message push method based on implicit multi-target fusion of a model according to an embodiment of the present invention
  • FIG2 is a schematic diagram of obtaining sample data according to an embodiment of the present invention.
  • FIG3 is a schematic diagram of training a second model according to an embodiment of the present invention.
  • FIG. 4 is a schematic diagram of a message push process based on model implicit multi-target fusion according to an embodiment of the present invention
  • FIG. 5 is a schematic diagram of the structural framework of a message push device based on implicit multi-target fusion of a model according to an embodiment of the present invention
  • FIG6 is a block diagram of an exemplary embodiment of an electronic device according to the present invention.
  • FIG. 7 is a schematic diagram of an embodiment of a computer readable medium of the present invention.
  • FIG. 1 is a message push method based on model implicit multi-target fusion provided by the present invention. As shown in FIG. 1 , the method includes:
  • each target sub-process is arranged in chronological order, and each target sub-process is a post-process of one or more previous target sub-processes, and is also a pre-process of one or more subsequent target sub-processes; for example: exposure is a pre-process of clicks, favorites, etc., clicks are a post-process of exposure, favorites are a post-process of exposure and clicks, and so on.
  • the sample data collected during the time period of each target sub-process is the sample data of the target sub-process;
  • the sample data of each target sub-process can be obtained by window sliding.
  • a sliding window [T1, T2] can be pre-set, and the sample data of each target sub-process can be obtained in the sliding window, and then the sliding window can be slid to the next position, and the cycle is repeated until the sliding window slides to the end time.
  • obtaining sample data of a single target sub-process in each sliding window includes:
  • the sample data includes: the device data that receives the push message in the sliding window; for example: when a message is pushed on a website, the sample data includes the device data that logs into the website in the sliding window.
  • the device data may be the data disclosed by the user who uses the device on the Internet service platform.
  • the data may include one or more items: user name, user age, user occupation, user income, user place of origin, user authorization information on the Internet platform, user product application behavior information, user security credibility and other public information, but not limited to this.
  • the data processing of this scheme can also be performed only through user information that cannot identify the user's identity, such as age, education, household registration, etc., to protect user privacy; the protection of user privacy can be achieved by deleting or anonymizing the information that can identify the user in the user information, and the anonymization processing can be performed by encrypting the data.
  • each target sub-process corresponds to a target sub-process event, for example, exposure corresponds to a display event, click corresponds to a click event, and so on. If the device triggers the target sub-process event, it indicates that the device has received the push of the message in the target sub-process, and the push target of the target sub-process has been completed; otherwise, it indicates that the device has not received the push of the message in the target sub-process, and the push target of the target sub-process has not been completed.
  • the triggering time of the target sub-process event can be detected by parsing the message push log.
  • the position of the target sub-process refers to whether the target sub-process is a pre-process or a post-process during model training. If the target sub-process is in a post-process, the sample data of the post-process is used to train the first model, and the sample data can be labeled directly according to the triggering moment; for example, the label of the sample data corresponding to the triggering moment is labeled as 1 as a positive sample, and the label of the sample data corresponding to the non-triggering moment is labeled as 0 as a negative sample.
  • the sample data of the front process is used to train the second model. Therefore, the sample data of the front process includes both the true label of the sample data (i.e., hard label) and the probability distribution of the output of the first model (i.e., soft label); the hard label and soft label can be labeled for the sample data according to the triggering moment.
  • the hard label of the sample data corresponding to the triggering moment is labeled as 1, as a positive sample
  • the hard label of the sample data corresponding to the non-triggering moment is labeled as 0, as a negative sample
  • the sample data is input into the trained first model to obtain the soft label corresponding to each sample data.
  • the soft label is determined according to the business characteristics. For example: the hard label of the sample data corresponding to the triggering moment is labeled as 1, and the soft label softlabel is:
  • T is the model parameter, specifically T is the parameter of the model distillation, and x is the predicted hard label corresponding to the sample data;
  • the hard label of the sample data corresponding to the non-trigger moment is marked as 0, and the soft label is 0.
  • the pre-process and post-process are determined according to the optimization index in the message push process. For example, when optimizing clicks, it is also desired to optimize payments simultaneously, so clicks are the pre-process and payments are the post-process. When the optimization index is greater than two, the target sub-process that is ranked last when the message is pushed is used as the post-process, and the other processes are used as the pre-process. For example, when optimizing clicks, it is also desired to optimize collections and payments simultaneously, so clicks and collections are the pre-process and payments are the post-process.
  • the sample data of the post-process is input into the first model to obtain a trained first model.
  • the sample data of the front-process is input into the second model for training. Since the sample data of the front-process includes the hard labels corresponding to the sample data of the front-process and the soft labels output by the first model predicting the sample data of the front-process, the sample information of the first model is implicitly integrated into the second model.
  • the training of the second model according to the sample data of the front-process includes:
  • the model parameter T of the first model and the second model is m.
  • the first model may adopt a k-layer neural network model
  • the second model may adopt an L-layer neural network model.
  • this step trains the first model with payment sample data and trains the second model with click sample data.
  • This training method implicitly incorporates the payment information of the first model into the second model, which can avoid the problem of sample imbalance among multiple targets to a certain extent and achieve End2End.
  • the predicted value of the trained second model directly feeds back the comprehensive value after weighing multiple business targets, making message push more efficient.
  • the front-end process may be one or more.
  • the front-end process may be one or more. For example, taking the RTA (Realtime API) advertising push mode as an example, as shown in Figure 4, after receiving the RTA request, the historical sample data of each front-end process is collected as the prediction data and input into the second model to obtain the prediction result.
  • RTA Realtime API
  • the model distillation technology for the number of exposures in message push (taking advertising as an example) to solve the end-to-end prediction problem, and to transform the multi-objective prediction modeling of customer acquisition and return on investment into an independent prediction problem of a single model. That is, the prediction result of a single model takes into account both types of objectives, which greatly reduces the storage scale and complexity of the model.
  • FIG5 is a message push device based on model implicit multi-target fusion of the present invention. As shown in FIG5 , the device includes:
  • the acquisition module 51 is used to acquire sample data of each target sub-process of the message push; each target sub-process is arranged in chronological order, and each target sub-process is a post-process of the previous target sub-process and/or a pre-process of the next target sub-process;
  • a training module 52 is used to train a first model according to sample data of a post-process, train a second model according to sample data of a pre-process, and implicitly integrate sample information of the first model into the second model;
  • the sample data of the pre-process includes: hard labels and soft labels output by the first model;
  • a prediction module 53 used for inputting the prediction data of the previous process into the second model to obtain a prediction result
  • the push module 54 is used to push messages according to the prediction results.
  • the acquisition module 51 includes:
  • a sub-acquisition module used to acquire sample data within the sliding window
  • a detection module used for detecting the triggering moment of the target sub-process event in the sliding window
  • a labeling module used to label the sample data according to the location of the target sub-process and the triggering time
  • the update module is used to update the sliding window.
  • the labeling module labels the sample data according to the triggering moment
  • the labeling module labels the sample data with a hard label and a soft label according to the triggering time. Specifically, the labeling module labels the hard label of the sample data corresponding to the triggering time as 1, and the soft label softlabel as:
  • T is the parameter of the model
  • x is the predicted hard label corresponding to the sample data
  • the hard label of the sample data corresponding to the non-trigger moment is marked as 0, and the soft label is 0.
  • the training module 52 includes:
  • a first configuration module used to configure the model parameters of the first model and the second model to be the same
  • a first input module used to input sample data of the previous process into the second model to obtain a predicted soft label
  • a first calculation module used to calculate a first loss function according to the marked soft label and the predicted soft label
  • a second configuration module used to configure a model parameter of the second model to be 1;
  • a second input module used to input the sample data of the previous process into the second model to obtain a predicted hard label
  • a second calculation module used to calculate a second loss function according to the marked hard label and the predicted hard label
  • An updating module is used to update the parameters of the second model based on the first loss function and the second loss function.
  • modules in the above device embodiments may be distributed in the device as described, or may be changed accordingly and distributed in one or more devices different from the above embodiments.
  • the modules in the above embodiments may be combined into one module, or may be further split into multiple sub-modules.
  • the electronic device embodiment of the present invention is described below, and the electronic device can be regarded as an implementation of the method and device embodiments of the present invention in a physical form.
  • the details described in the electronic device embodiment of the present invention should be regarded as a supplement to the above method or device embodiments; details not disclosed in the electronic device embodiment of the present invention can be implemented with reference to the above method or device embodiments.
  • Fig. 6 is a structural block diagram of an exemplary embodiment of an electronic device according to the present invention.
  • the electronic device shown in Fig. 6 is only an example and should not bring any limitation to the functions and application scope of the embodiments of the present invention.
  • the electronic device 600 of this exemplary embodiment is in the form of a general data processing device.
  • the components of the electronic device 600 may include, but are not limited to: at least one processing unit 610, at least one storage unit 620, a bus 630 connecting different electronic device components (including the storage unit 620 and the processing unit 610), a display unit 640, etc.
  • the storage unit 620 stores a computer-readable program, which may be a source program or a code of a read-only program.
  • the program may be executed by the processing unit 610, so that the processing unit 610 performs the steps of various embodiments of the present invention. For example, the processing unit 610 may perform the steps shown in FIG. 1 .
  • the storage unit 620 may include a readable medium in the form of a volatile storage unit, such as a random access storage unit (RAM) 6201 and/or a cache storage unit 6202, and may further include a read-only storage unit (ROM) 6203.
  • the storage unit 620 may also include a program/utility 6204 having a set (at least one) of program modules 6205, such program modules 6205 including but not limited to: operating the electronic device, one or more application programs, other program modules, and program data, each of which or some combination may include the implementation of a network environment.
  • Bus 630 may represent one or more of several types of bus structures, including a memory unit bus or memory unit controller, a peripheral bus, an accelerated graphics port, a processing unit, or a local bus using any of a variety of bus architectures.
  • the electronic device 600 may also communicate with one or more external devices 100 (e.g., keyboards, displays, network devices, Bluetooth devices, etc.) so that a user can interact with the electronic device 600 via these external devices 100, and/or so that the electronic device 600 can communicate with one or more other data processing devices (e.g., routers, modems, etc.). Such communication may be performed via an input/output (I/O) interface 650, and may also be performed with one or more via a network adapter 660.
  • the network adapter 660 may communicate with other modules of the electronic device 600 via a bus 630. It should be understood that, although not shown in FIG. 6 , other hardware and/or software modules may be used in the electronic device 600, including but not limited to: microcode, device drivers, redundant processing units, external disk drive arrays, RAID electronics, tape drives, and data backup storage electronics, etc.
  • FIG. 7 is a schematic diagram of a computer readable medium embodiment of the present invention.
  • the computer program may be stored on one or more computer readable media.
  • the computer readable medium may be a readable signal medium or a readable storage medium.
  • the readable storage medium may be, for example, but not limited to, an electronic device, device or device of electricity, magnetism, light, electromagnetic, infrared, or semiconductor, or any combination thereof.
  • the computer readable medium is enabled to implement the method as shown in FIG. 1 .
  • the technical solution according to the implementation mode of the present invention can be embodied in the form of a software product, which can be stored in a computer-readable storage medium (which can be a CD-ROM, a USB flash drive, a mobile hard disk, etc.) or on a network, and includes several instructions to enable a data processing device (which can be a personal computer, a server, or a network device, etc.) to execute the above method according to the present invention.
  • a computer-readable storage medium which can be a CD-ROM, a USB flash drive, a mobile hard disk, etc.
  • a data processing device which can be a personal computer, a server, or a network device, etc.
  • the present invention can be implemented by a method, apparatus, electronic device or computer-readable medium that executes a computer program.
  • a general data processing device such as a microprocessor or a digital signal processor (DSP) can be used to implement some or all functions of the present invention.
  • DSP digital signal processor

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • General Engineering & Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Artificial Intelligence (AREA)
  • Databases & Information Systems (AREA)
  • Evolutionary Computation (AREA)
  • Molecular Biology (AREA)
  • General Health & Medical Sciences (AREA)
  • Software Systems (AREA)
  • Health & Medical Sciences (AREA)
  • Biomedical Technology (AREA)
  • Biophysics (AREA)
  • Computational Linguistics (AREA)
  • Mathematical Physics (AREA)
  • Computing Systems (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Evolutionary Biology (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Management, Administration, Business Operations System, And Electronic Commerce (AREA)

Abstract

本发明公开了一种基于模型隐式多目标融合的消息推送方法及装置，所述方法包括：获取消息推送各个目标子流程的样本数据；各个目标子流程按照时间顺序排列，每个目标子流程为前面目标子流程的后置流程和/或为后面目标子流程的前置流程；根据后置流程的样本数据训练第一模型，根据前置流程的样本数据训练第二模型，将第一模型的样本信息隐式融入第二模型中；将前置流程的预测数据输入第二模型中得到预测结果；根据预测结果进行消息推送。本发明能规避多个业务目标间样本不均衡的问题，实现端到端，降低推送系统的复杂性；并且训练得到第二模型的预测值直接反馈了多个业务目标权衡后的综合值，从多个业务目标指导消息推送，从而提高消息推送效果。

Description

一种基于模型隐式多目标融合的消息推送方法及装置 技术领域
本发明涉及智能消息推送技术领域，具体而言，涉及一种基于模型隐式多目标融合的消息推送方法、装置、电子设备及计算机可读介质。
背景技术
目前，由于大数据驱动，使得效果指标的量化评估成为了可能。为了实现好的消息推送效果，消息推送中会融合各个目标子流程的信息来确定推送策略。比如：期望消息推送既有高点击也存在高转化，这就需要综合各个目标子流程的业务目标进行组合优化。
现有技术中，先对单个目标子流程的业务目标构建模型，再对多个业务目标的预测结果做线性组合，通过组合结果对推送策略进行优化；或者，基于各个目标子流程数据通过多目标训练方法和模型进行预测和优化。而这些方法并未考虑到：1、部分业务指标对应样本量可能比较稀疏，联合训练时会存在样本不均衡问题；2、无法直接端到端(End2End)建模融合多个业务目标的score；3、无法将后置业务目标作为先验在建模中体现。因此会导致消息推送效果差，并且推送系统复杂、计算量大的技术问题。
发明内容
有鉴于此，本发明主要目的在于提出一种基于模型隐式多目标融合的消息推送方法、装置、电子设备及计算机可读介质，以期至少部分地解决上述技术问题中的至少之一。
为了解决上述技术问题，本发明第一方面提出一种基于模型隐式多目标融合的消息推送方法，所述方法包括：
获取消息推送各个目标子流程的样本数据；各个目标子流程按照时间顺序排列，每个目标子流程为前面目标子流程的后置流程和/或为后面目标子流程的前置流程；
根据后置流程的样本数据训练第一模型，根据前置流程的样本数据训练第二模型，将第一模型的样本信息隐式融入第二模型中；所述前置流程的样本数据中包括：硬标签和第一模型输出的软标签；
将前置流程的预测数据输入所述第二模型中得到预测结果；
根据所述预测结果进行消息推送。
根据本发明一种优选实施方式，获取单个目标子流程的样本数据包括：
获取滑动窗口内的样本数据；
在滑动窗口内检测触发目标子流程事件的触发时刻；
根据目标子流程所处的位置及所述触发时刻为样本数据标注标签；
更新滑动窗口。
根据本发明一种优选实施方式，若目标子流程处于后置流程，根据所述触发时刻为样本数据标注标签；
若目标子流程处于前置流程，根据所述触发时刻为样本数据标注硬标签和软标签。
根据本发明一种优选实施方式，触发时刻对应样本数据的硬标签标注为1，软标签softlabel为：
其中：T为模型参数，x为样本数据对应的预测硬标签；
非触发时刻对应样本数据的硬标签标注为0，软标签为0。
根据本发明一种优选实施方式，消息推送的时间序列为：{T0，T1，…，Tn-k，…，Tn}；若终端在Ti时刻触发前置流程事件并在Tn时刻触发后置流程事件，且Ti～Tn时刻共推送消息m次，则模型参数T＝m。
根据本发明一种优选实施方式，所述根据前置流程的样本数据训练第二模型包括：
配置第一模型和第二模型的模型参数相同；
将前置流程的样本数据输入第二模型得到预测软标签；
根据标注的软标签和预测软标签计算第一损失函数；
配置第二模型的模型参数为1；
将前置流程的样本数据输入第二模型得到预测硬标签；
根据标注的硬标签和预测硬标签计算第二损失函数；
基于所述第一损失函数和第二损失函数更新第二模型的参数。
为了解决上述技术问题，本发明第二方面提供一种基于模型隐式多目标融合的消息推送装置，所述装置包括：
获取模块，用于获取消息推送各个目标子流程的样本数据；各个目标子流程按照时间顺序排列，每个目标子流程为前面目标子流程的后置流程和/或为后面目标子流程的前置流程；
训练模块，用于根据后置流程的样本数据训练第一模型，根据前置流程的样本数据训练第二模型，将第一模型的样本信息隐式融入第二模型中；所述前置流程的样本数据中包括：硬标签和所述第一模型输出的软标签；
预测模块，用于将前置流程的预测数据输入所述第二模型中得到预测结果；
推送模块，用于根据所述预测结果进行消息推送。
根据本发明一种优选实施方式，所述获取模块包括：
子获取模块，用于获取滑动窗口内的样本数据；
检测模块，用于在滑动窗口内检测触发目标子流程事件的触发时刻；
标注模块，用于根据目标子流程所处的位置及所述触发时刻为样本数据标注标签；
更新模块，用于更新滑动窗口。
根据本发明一种优选实施方式，若目标子流程处于后置流程，所述标注模块根据所述触发时刻为样本数据标注标签；
若目标子流程处于前置流程，所述标注模块根据所述触发时刻为样本数据标注硬标签和软标签。
根据本发明一种优选实施方式，若所述目标子流程处于前置流程，所述标注模块将触发时刻对应样本数据的硬标签标注为1，软标签softlabel为：
其中：T为模型参数，x为样本数据对应的预测硬标签；
将非触发时刻对应样本数据的硬标签标注为0，软标签为0。
根据本发明一种优选实施方式，消息推送的时间序列为：{T0，T1，…，Tn-k，…，Tn}；若终端在Ti时刻触发前置流程事件并在Tn时刻触发后置流程事件，且Ti～Tn时刻共推送消息m次，则模型参数T＝m。
根据本发明一种优选实施方式，所述训练模块包括：
第一配置模块，用于配置第一模型和第二模型的模型参数相同；
第一输入模块，用于将前置流程的样本数据输入第二模型得到预测软标签；
第一计算模块，用于根据标注的软标签和预测软标签计算第一损失函数；
第二配置模块，用于配置第二模型的模型参数为1；
第二输入模块，用于将前置流程的样本数据输入第二模型得到预测硬标签；
第二计算模块，用于根据标注的硬标签和预测硬标签计算第二损失函数；
更新模块，用于基于所述第一损失函数和第二损失函数更新第二模型的参数。
为解决上述技术问题，本发明第三方面提供一种电子设备，包括：
处理器；以及
存储计算机可执行指令的存储器，所述计算机可执行指令在被执行时使所述处理器执行上述任一项所述的方法。
为解决上述技术问题，本发明第四方面提供一种计算机可读存储介质，其中，所述计算机可读存储介质存储一个或多个程序，当所述一个或多个程序被处理器执行时，实现上述方法。
本发明获取消息推送各个目标子流程的样本数据；其中：每个目标子流程为前面目标子流程的后置流程和/或为后面目标子流程的前置流程；根据后置流程的样本数据训练第一模型，根据前置流程的样本数据训练第二模型；其中：前置流程的样本数据中包括：硬标签和所述第一模型输出的软标签，从而将第二模型中隐式融入了第一模型的样本信息，通过这种训 练方式可以在一定程度上规避多个业务目标间样本不均衡的问题，同时实现End2End预测，降低推送系统的复杂性；并且训练得到第二模型的预测值直接反馈了多个业务目标权衡后的综合值，可以从多个业务目标指导消息推送，从而提高消息推送效果。
附图说明
为了使本发明所解决的技术问题、采用的技术手段及取得的技术效果更加清楚，下面将参照附图详细描述本发明的具体实施例。但需声明的是，下面描述的附图仅仅是本发明的示例性实施例的附图，对于本领域的技术人员来讲，在不付出创造性劳动的前提下，可以根据这些附图获得其他实施例的附图。
图1是本发明实施例一种基于模型隐式多目标融合的消息推送方法的流程示意图；
图2是本发明实施例获取样本数据的示意图；
图3是本发明实施例训练第二模型的示意图；
图4是本发明实施例一种基于模型隐式多目标融合的消息推送过程的示意图；
图5是本发明实施例一种基于模型隐式多目标融合的消息推送装置的结构框架示意图；
图6是根据本发明的电子设备的示例性实施例的结构框图；
图7是本发明一种计算机可读介质实施例的示意图。
具体实施方式
在符合本发明的技术构思的前提下，在某个特定的实施例中描述的结构、性能、效果或者其他特征可以以任何合适的方式结合到一个或更多其他的实施例中。
在对于具体实施例的介绍过程中，对结构、性能、效果或者其他特征的细节描述是为了使本领域的技术人员对实施例能够充分理解。但是，并不排除本领域技术人员可以在特定情况下，以不含有上述结构、性能、效果或者其他特征的技术方案来实施本发明。附图中的流程图仅是一种示例 性的流程演示，不代表本发明的方案中必须包括流程图中的所有的内容、操作和步骤，也不代表必须按照图中所显示的的顺序执行。
请参阅图1，图1是本发明提供的一种基于模型隐式多目标融合的消息推送方法，如图1所示，所述方法包括：
S1、获取消息推送各个目标子流程的样本数据；
产品在进行目标消息推送时会经历多个目标子流程，比如：曝光、点击、收藏、付款等，且各个目标子流程之间存在时序关系，即点击目标子流程必然发生在曝光目标子流程后，以此类推。因此，本实施例中，各个目标子流程按照时间顺序排列，每个目标子流程为前面一个或多个目标子流程的后置流程，同时也是后面一个或多个目标子流程的前置流程；比如：曝光为点击、收藏等的前置流程，点击为曝光的后置流程，收藏为曝光、点击的后置流程，等等。
本实施例中，每个目标子流程所在时间段内所采集的样本数据为该目标子流程的样本数据；示例性的，可以通过窗口滑动的方式获取各个目标子流程的样本数据。比如：可以预先设定滑动窗口[T1,T2]，在滑动窗口内分别获取各个目标子流程的样本数据，然后再滑动所述滑动窗口至下一位置，如此循环，直至滑动窗口滑动到终止时间为止。其中：在每个滑动窗口内获取单个目标子流程的样本数据包括：
S11、获取滑动窗口内的样本数据；
其中：所述样本数据包括：滑动窗口内接收到推送消息的设备数据；比如：在某网站推送消息，则样本数据包括在滑动窗口内登陆该网站的设备数据。设备数据可以是使用该设备的用户在互联网服务平台公开的数据。该数据可以包括一项或多项：用户名称、用户年龄、用户职业、用户收入、用户籍贯、用户在互联网平台上的授权信息、用户申请产品行为信息、用户安全可信度等公开信息，但不限于此，还可以仅通过无法识别用户身份的用户信息进行本方案的数据处理，比如，年龄、学历、户籍等，以实现对于保护用户隐私；可以采用对用户信息中可以识别出用户身份的信息删除或者匿名化处理的方式来实现对于用户隐私的保护，匿名化处理可以是通过加密手段对数据进行处理。
S12、在滑动窗口内检测触发目标子流程事件的触发时刻；
本实施例中，每个目标子流程对应一个目标子流程事件，比如：曝光对应展示事件，点击对应点击事件；等等。设备若触发该目标子流程事件表明设备接收到消息在该目标子流程的推送，该目标子流程的推送目标已完成；反之，则表明设备未接收到消息在该目标子流程的推送，该目标子流程的推送目标未完成。示例性的，可以通过对消息推送日志的解析来检测触发目标子流程事件的触发时刻。
S13、根据目标子流程所处的位置及所述触发时刻为样本数据标注标签；
所述目标子流程所处的位置指在模型训练时目标子流程为前置流程还是后置流程。若目标子流程处于后置流程，后置流程的样本数据用于训练第一模型，可以直接根据所述触发时刻为样本数据标注标签；比如：将触发时刻对应样本数据的标签标注为1，作为正样本，将非触发时刻对应样本数据的标签标注为0，作为负样本。
若所述目标子流程处于前置流程，前置流程的样本数据用于训练第二模型，因此，前置流程的样本数据既包含样本数据的真实标签(即：硬标签)，也包含第一模型输出的概率分布(即软标签)；可以根据所述触发时刻为样本数据标注硬标签和软标签。比如：触发时刻对应样本数据的硬标签标注为1，作为正样本；非触发时刻对应样本数据的硬标签标注为0，作为负样本；将样本数据输入训练好的第一模型后得到各个样本数据对应的软标签。或者：在一种简化方式中，根据业务特点确定软标签，比如：触发时刻对应样本数据的硬标签标注为1，软标签softlabel为：
其中：T为模型参数，具体T为模型蒸馏的参数，x为样本数据对应的预测硬标签；
非触发时刻对应样本数据的硬标签标注为0，软标签为0。
本实施例中，消息推送的时间序列为：{T0，T1，…，Tn-k，…，Tn}；若终端在Ti时刻触发前置流程事件并在Tn时刻触发后置流程事件，且Ti～Tn时刻共推送消息m次，则模型蒸馏参数T＝m。
S14、更新滑动窗口。
比如：滑动所述滑动窗口至下一位置。
以点击作为前置流程为例，图2所示：若在滑动窗口若检测到设备i点击时刻为：Ti+1，Ti+k，Ti+s，且在最后一次点击后推送2次消息(如图2中曝光两次)，则m＝2，三个点击时刻对应三条样本正样本，对应硬标签为1，软标签为：其他时刻对应负样本；再通过窗口滑动方式获取其他样本。
S2、根据后置流程的样本数据训练第一模型，根据前置流程的样本数据训练第二模型，将第一模型的样本信息隐式融入第二模型中；
本实施例中，前置流程和后置流程根据消息推送过程中的优化指标来确定，比如：优化点击时也希望同步优化付款，则点击为前置流程、付款为后置流程。当优化指标大于两个时，将消息推送时排在最后的目标子流程作为后置流程，其他流程作为前置流程。比如：优化点击时也希望同步优化收藏和付款，则点击和收藏为前置流程、付款为后置流程。
示例性的，如图3，将后置流程的样本数据输入第一模型得到训练好的第一模型。训练好第一模型后，再将前置流程的样本数据输入第二模型进行训练。由于前置流程的样本数据中包括了前置流程的样本数据对应的硬标签和第一模型预测前置流程的样本数据输出的软标签，从而将第一模型的样本信息隐式融入第二模型中。其中：所述根据前置流程的样本数据训练第二模型包括：
S21、配置第一模型和第二模型的模型参数相同；
本实施例中，第一模型和第二模型的模型参数T＝m。第一模型可以采用k层神经网络模型，第二模型可以采用L层神经网络模型。
S22、将前置流程的样本数据输入第二模型得到预测软标签；
S23、根据标注的软标签和预测软标签计算第一损失函数F1；
S24、配置第二模型的模型参数为1；
S25、将前置流程的样本数据输入第二模型得到预测硬标签；
S26、根据标注的硬标签和预测硬标签计算第二损失函数F2；
S27、基于所述第一损失函数和第二损失函数更新第二模型的参数。
以点击和付款为例，则本步骤通过付款的样本数据训练第一模型，通过点击的样本数据训练第二模型，通过这种训练方式使第二模型中隐式融入了第一模型的付款信息，可以一定程度上规避多个目标间样本不均衡的问题，同时实现End2End，同时训练后的第二模型的预测值直接反馈了多个业务目标权衡后的综合值，使消息推送效率更高。
S3、将前置流程的预测数据输入所述第二模型中得到预测结果；
其中：前置流程可能是一个，也可能是多个。示例性的，以RTA(Realtime API)广告推送模式为例，如图4，当接收到RTA请求后，采集各个前置流程的历史样本数据作为预测数据输入所述第二模型中得到预测结果。
S4、根据所述预测结果进行消息推送。
比如：根据预测结果调整消息推送时间、推送方式、推送价格等等，生成新的消息策略，并基于新的消息策略推送消息。
本公开实施例中首次提出在消息推送(可以以广告为例)中针对曝光次数融合模型蒸馏技术来解决端到端预测问题，将获客和投资回报率的的多目标预测建模转化为单模型的独立预测问题，即单模型预测结果兼顾此两类目标，极大地降低了模型的存储规模和复杂性。
图5是本发明一种基于模型隐式多目标融合的消息推送装置，如图5所示，所述装置包括：
获取模块51，用于获取消息推送各个目标子流程的样本数据；各个目标子流程按照时间顺序排列，每个目标子流程为前面目标子流程的后置流程和/或为后面目标子流程的前置流程；
训练模块52，用于根据后置流程的样本数据训练第一模型，根据前置流程的样本数据训练第二模型，将第一模型的样本信息隐式融入第二模型中；所述前置流程的样本数据中包括：硬标签和所述第一模型输出的软标签；
预测模块53，用于将前置流程的预测数据输入所述第二模型中得到预测结果；
推送模块54，用于根据所述预测结果进行消息推送。
在一种实施方式中，所述获取模块51包括：
子获取模块，用于获取滑动窗口内的样本数据；
检测模块，用于在滑动窗口内检测触发目标子流程事件的触发时刻；
标注模块，用于根据目标子流程所处的位置及所述触发时刻为样本数据标注标签；
更新模块，用于更新滑动窗口。
可选的，若目标子流程处于后置流程，所述标注模块根据所述触发时刻为样本数据标注标签；
若目标子流程处于前置流程，所述标注模块根据所述触发时刻为样本数据标注硬标签和软标签。具体的，所述标注模块将触发时刻对应样本数据的硬标签标注为1，软标签softlabel为：
其中：T为模型的参数，x为样本数据对应的预测硬标签；
将非触发时刻对应样本数据的硬标签标注为0，软标签为0。
根据权利要求10所述的装置，其特征在于，消息推送的时间序列为：{T0，T1，…，Tn-k，…，Tn}；若终端在Ti时刻触发前置流程事件并在Tn时刻触发后置流程事件，且Ti～Tn时刻共推送消息m次，则模型参数T＝m。
所述训练模块52包括：
第一配置模块，用于配置第一模型和第二模型的模型参数相同；
第一输入模块，用于将前置流程的样本数据输入第二模型得到预测软标签；
第一计算模块，用于根据标注的软标签和预测软标签计算第一损失函数；
第二配置模块，用于配置第二模型的模型参数为1；
第二输入模块，用于将前置流程的样本数据输入第二模型得到预测硬标签；
第二计算模块，用于根据标注的硬标签和预测硬标签计算第二损失函数；
更新模块，用于基于所述第一损失函数和第二损失函数更新第二模型的参数。
本领域技术人员可以理解，上述装置实施例中的各模块可以按照描述分布于装置中，也可以进行相应变化，分布于不同于上述实施例的一个或多个装置中。上述实施例的模块可以合并为一个模块，也可以进一步拆分成多个子模块。
下面描述本发明的电子设备实施例，该电子设备可以视为对于上述本发明的方法和装置实施例的实体形式的实施方式。对于本发明电子设备实施例中描述的细节，应视为对于上述方法或装置实施例的补充；对于在本发明电子设备实施例中未披露的细节，可以参照上述方法或装置实施例来实现。
图6是根据本发明的一种电子设备的示例性实施例的结构框图。图6显示的电子设备仅仅是一个示例，不应对本发明实施例的功能和使用范围带来任何限制。
如图6所示，该示例性实施例的电子设备600以通用数据处理设备的形式表现。电子设备600的组件可以包括但不限于：至少一个处理单元610、至少一个存储单元620、连接不同电子设备组件(包括存储单元620和处理单元610)的总线630、显示单元640等。
其中，所述存储单元620存储有计算机可读程序，其可以是源程序或都只读程序的代码。所述程序可以被处理单元610执行，使得所述处理单元610执行本发明各种实施方式的步骤。例如，所述处理单元610可以执行如图1所示的步骤。
所述存储单元620可以包括易失性存储单元形式的可读介质，例如随机存取存储单元(RAM)6201和/或高速缓存存储单元6202，还可以进一步包括只读存储单元(ROM)6203。所述存储单元620还可以包括具有一组(至少一个)程序模块6205的程序/实用工具6204，这样的程序模块6205包括但不限于：操作电子设备、一个或者多个应用程序、其它程序模块以及程序数据，这些示例中的每一个或某种组合中可能包括网络环境的实现。
总线630可以为表示几类总线结构中的一种或多种，包括存储单元总线或者存储单元控制器、外围总线、图形加速端口、处理单元或者使用多种总线结构中的任意总线结构的局域总线。
电子设备600也可以与一个或多个外部设备100(例如键盘、显示器、网络设备、蓝牙设备等)通信，使得用户能经由这些外部设备100与该电子设备600交互，和/或使得该电子设备600能与一个或多个其它数据处理设备(例如路由器、调制解调器等等)进行通信。这种通信可以通过输入/输出(I/O)接口650进行，还可以通过网络适配器660与一个或者多个进行。网络适配器660可以通过总线630与电子设备600的其它模块通信。应当明白，尽管图6中未示出，电子设备600中可使用其它硬件和/或软件模块，包括但不限于：微代码、设备驱动器、冗余处理单元、外部磁盘驱动阵列、RAID电子设备、磁带驱动器以及数据备份存储电子设备等。
图7是本发明的一个计算机可读介质实施例的示意图。如图7所示，所述计算机程序可以存储于一个或多个计算机可读介质上。计算机可读介质可以是可读信号介质或者可读存储介质。可读存储介质例如可以为但不限于电、磁、光、电磁、红外线、或半导体的电子设备、装置或器件，或者任意以上的组合。当所述计算机程序被一个或多个数据处理设备执行时，使得该计算机可读介质能够实现如图1的方法。
通过以上的实施方式的描述，本领域的技术人员易于理解，本发明描述的示例性实施例可以通过软件实现，也可以通过软件结合必要的硬件的方式来实现。因此，根据本发明实施方式的技术方案可以以软件产品的形式体现出来，该软件产品可以存储在一个计算机可读的存储介质(可以是CD-ROM，U盘，移动硬盘等)中或网络上，包括若干指令以使得一台数据处理设备(可以是个人计算机、服务器、或者网络设备等)执行根据本发明的上述方法。
综上所述，本发明可以执行计算机程序的方法、装置、电子设备或计算机可读介质来实现。可以在实践中使用微处理器或者数字信号处理器(DSP)等通用数据处理设备来实现本发明的一些或者全部功能。
以上所述的具体实施例，对本发明的目的、技术方案和有益效果进行了进一步详细说明，应理解的是，本发明不与任何特定计算机、虚拟装置 或者电子设备固有相关，各种通用装置也可以实现本发明。以上所述仅为本发明的具体实施例而已，并不用于限制本发明，凡在本发明的精神和原则之内，所做的任何修改、等同替换、改进等，均应包含在本发明的保护范围之内。

Claims (10)

  1. 一种基于模型隐式多目标融合的消息推送方法，其特征在于，所述方法包括：
    获取消息推送各个目标子流程的样本数据；各个目标子流程按照时间顺序排列，每个目标子流程为前面目标子流程的后置流程和/或为后面目标子流程的前置流程；
    根据后置流程的样本数据训练第一模型，根据前置流程的样本数据训练第二模型，所述第一模型的样本信息隐式融入于第二模型中；所述前置流程的样本数据中包括：硬标签和所述第一模型输出的软标签；
    将前置流程的预测数据输入所述第二模型中得到预测结果；
    根据所述预测结果进行消息推送。
  2. 根据权利要求1所述的方法，其特征在于，获取单个目标子流程的样本数据包括：
    获取滑动窗口内的样本数据；
    在滑动窗口内检测触发目标子流程事件的触发时刻；
    根据目标子流程所处的位置及所述触发时刻为样本数据标注标签；
    更新滑动窗口。
  3. 根据权利要求2所述的方法，其特征在于，
    若目标子流程处于后置流程，根据所述触发时刻为样本数据标注标签；
    若目标子流程处于前置流程，根据所述触发时刻为样本数据标注硬标签和软标签。
  4. 根据权利要求3所述的方法，其特征在于，
    触发时刻对应样本数据的硬标签标注为1，软标签softlabel为：
    其中：T为模型参数，x为样本数据对应的预测硬标签；
    非触发时刻对应样本数据的硬标签标注为0，软标签为0。
  5. 根据权利要求4所述的方法，其特征在于，消息推送的时间序列为：{T0，T1，…，Tn-k，…，Tn}；若终端在Ti时刻触发前置流程事件并在Tn时刻触发后置流程事件，且Ti～Tn时刻共推送消息m次，则模型参数T＝m。
  6. 根据权利要求2所述的方法，其特征在于，所述根据前置流程的样本数据训练第二模型包括：
    配置第一模型和第二模型的模型参数相同；
    将前置流程的样本数据输入第二模型得到预测软标签；
    根据标注的软标签和预测软标签计算第一损失函数；
    配置第二模型的模型参数为1；
    将前置流程的样本数据输入第二模型得到预测硬标签；
    根据标注的硬标签和预测硬标签计算第二损失函数；
    基于所述第一损失函数和第二损失函数更新第二模型的参数。
  7. 一种基于模型隐式多目标融合消息推送装置，其特征在于，所述装置包括：
    获取模块，用于获取消息推送各个目标子流程的样本数据；各个目标子流程按照时间顺序排列，每个目标子流程为前面目标子流程的后置流程和/或为后面目标子流程的前置流程；
    训练模块，用于根据后置流程的样本数据训练第一模型，根据前置流程的样本数据训练第二模型，将第一模型的样本信息隐式融入第二模型中；所述前置流程的样本数据中包括：硬标签和所述第一模型输出的软标签；
    预测模块，用于将前置流程的预测数据输入所述第二模型中得到预测结果；
    推送模块，用于根据所述预测结果进行消息推送。
  8. 根据权利要求7所述的装置，其特征在于，所述获取模块包括：
    子获取模块，用于获取滑动窗口内的样本数据；
    检测模块，用于在滑动窗口内检测触发目标子流程事件的触发时刻；
    标注模块，用于根据目标子流程所处的位置及所述触发时刻为样本数据标注标签；
    更新模块，用于更新滑动窗口。
  9. 一种电子设备，包括：
    处理器；以及
    存储计算机可执行指令的存储器，所述计算机可执行指令在被执行时使所述处理器执行根据权利要求1至6中任一项所述的方法。
  10. 一种计算机可读存储介质，其中，所述计算机可读存储介质存储一个或多个程序，当所述一个或多个程序被处理器执行时，实现权利要求1至6中任一项所述的方法。
PCT/CN2023/130616 2022-11-14 2023-11-09 一种基于模型隐式多目标融合的消息推送方法及装置 WO2024104241A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN202211417236.6A CN115878888A (zh) 2022-11-14 2022-11-14 一种基于模型隐式多目标融合的消息推送方法及装置
CN202211417236.6 2022-11-14

Publications (1)

Publication Number Publication Date
WO2024104241A1 true WO2024104241A1 (zh) 2024-05-23

Family

ID=85759742

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2023/130616 WO2024104241A1 (zh) 2022-11-14 2023-11-09 一种基于模型隐式多目标融合的消息推送方法及装置

Country Status (2)

Country Link
CN (1) CN115878888A (zh)
WO (1) WO2024104241A1 (zh)

Families Citing this family (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN115878888A (zh) * 2022-11-14 2023-03-31 上海淇玥信息技术有限公司 一种基于模型隐式多目标融合的消息推送方法及装置

Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111461566A (zh) * 2020-04-10 2020-07-28 武汉大学 一种基于消息流划分与合并的跨界服务流程融合方法及系统
CN113159314A (zh) * 2021-04-02 2021-07-23 支付宝(杭州)信息技术有限公司 业务模型的训练方法、装置及系统
CN113449821A (zh) * 2021-08-31 2021-09-28 浙江宇视科技有限公司 融合语义和图像特征的智能训练方法、装置、设备及介质
CN113761375A (zh) * 2021-09-10 2021-12-07 未鲲(上海)科技服务有限公司 基于神经网络的消息推荐方法、装置、设备及存储介质
WO2022227400A1 (zh) * 2021-04-27 2022-11-03 商汤集团有限公司 神经网络训练方法和装置、设备，及计算机存储介质
CN115878888A (zh) * 2022-11-14 2023-03-31 上海淇玥信息技术有限公司 一种基于模型隐式多目标融合的消息推送方法及装置

Patent Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111461566A (zh) * 2020-04-10 2020-07-28 武汉大学 一种基于消息流划分与合并的跨界服务流程融合方法及系统
CN113159314A (zh) * 2021-04-02 2021-07-23 支付宝(杭州)信息技术有限公司 业务模型的训练方法、装置及系统
WO2022227400A1 (zh) * 2021-04-27 2022-11-03 商汤集团有限公司 神经网络训练方法和装置、设备，及计算机存储介质
CN113449821A (zh) * 2021-08-31 2021-09-28 浙江宇视科技有限公司 融合语义和图像特征的智能训练方法、装置、设备及介质
CN113761375A (zh) * 2021-09-10 2021-12-07 未鲲(上海)科技服务有限公司 基于神经网络的消息推荐方法、装置、设备及存储介质
CN115878888A (zh) * 2022-11-14 2023-03-31 上海淇玥信息技术有限公司 一种基于模型隐式多目标融合的消息推送方法及装置

Also Published As

Publication number Publication date
CN115878888A (zh) 2023-03-31

Similar Documents

Publication Publication Date Title
CN112445854B (zh) 多源业务数据实时处理方法、装置、终端及存储介质
WO2024104241A1 (zh) 一种基于模型隐式多目标融合的消息推送方法及装置
US20200357063A1 (en) Systems and computer-implemented processes for model-based underwriting
CN111586126A (zh) 小程序预下载方法、装置、设备及存储介质
CN112686717B (zh) 一种广告召回的数据处理方法及系统
CN108830715A (zh) 批量文件部分回盘处理方法和系统
CN110795697A (zh) 逻辑表达式的获取方法、装置、存储介质以及电子装置
CN110717597A (zh) 利用机器学习模型获取时序特征的方法和装置
CN111651524B (zh) 利用机器学习模型进行线上预测的辅助实现方法及装置
CN112070564A (zh) 广告拉取方法、装置、系统与电子设备
US20050165596A1 (en) Method and apparatus for determining expected values in the presence of uncertainty
CN113705698A (zh) 基于点击行为预测的信息推送方法及装置
CN112102099A (zh) 保单数据处理方法、装置、电子设备及存储介质
CN116226173A (zh) 数据查询方法、装置、存储介质和电子设备
CN113780675B (zh) 一种消耗预测方法、装置、存储介质及电子设备
CN109785118A (zh) 贷后数据处理的方法、装置、介质及电子设备
CN115269085A (zh) 手机银行页面展示方法及装置
CN114066507A (zh) 推广信息的分析方法、装置、电子设备及存储介质
CN114372892A (zh) 支付数据监控方法、装置、设备及介质
CN114298825A (zh) 还款积极度评估方法及装置
CN114282940A (zh) 用于意图识别的方法及装置、存储介质及电子设备
CN113987351A (zh) 基于人工智能的智能推荐方法、装置、电子设备及介质
CN113590484A (zh) 算法模型服务测试方法、系统、设备及存储介质
CN113191844A (zh) 基于匿名用户在线操作的产品推荐方法、装置和电子设备
CN111079991A (zh) 一种业务指标预测方法、装置、设备和存储介质

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: 23890677

Country of ref document: EP

Kind code of ref document: A1