WO2020044098A2 - 一种信息流中的排序方法、装置和设备/终端/服务器 - Google Patents

一种信息流中的排序方法、装置和设备/终端/服务器 Download PDF

Info

Publication number
WO2020044098A2
WO2020044098A2 PCT/IB2018/057166 IB2018057166W WO2020044098A2 WO 2020044098 A2 WO2020044098 A2 WO 2020044098A2 IB 2018057166 W IB2018057166 W IB 2018057166W WO 2020044098 A2 WO2020044098 A2 WO 2020044098A2
Authority
WO
WIPO (PCT)
Prior art keywords
user
recommended object
model
information flow
recommended
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/IB2018/057166
Other languages
English (en)
French (fr)
Other versions
WO2020044098A3 (zh
Inventor
马泽锋
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.)
Ucweb Singapore Pte Ltd
Original Assignee
Ucweb Singapore Pte 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 Ucweb Singapore Pte Ltd filed Critical Ucweb Singapore Pte Ltd
Publication of WO2020044098A2 publication Critical patent/WO2020044098A2/zh
Publication of WO2020044098A3 publication Critical patent/WO2020044098A3/zh
Anticipated expiration legal-status Critical
Ceased legal-status Critical Current

Links

Classifications

    • GPHYSICS
    • G06COMPUTING OR CALCULATING; COUNTING
    • G06FELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00Information retrieval; Database structures therefor; File system structures therefor
    • GPHYSICS
    • G06COMPUTING OR CALCULATING; COUNTING
    • G06FELECTRIC DIGITAL DATA PROCESSING
    • G06F17/00Digital computing or data processing equipment or methods, specially adapted for specific functions
    • G06F17/10Complex mathematical operations
    • G06F17/16Matrix or vector computation, e.g. matrix-matrix or matrix-vector multiplication, matrix factorization
    • GPHYSICS
    • G06COMPUTING OR CALCULATING; COUNTING
    • G06NCOMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N20/00Machine learning

Definitions

  • the present application relates to the field of Internet technologies, and in particular, to a method, an apparatus, and a device / terminal / server for ordering information flows.
  • the algorithm model cannot accurately describe the relationship between the information flow object and the user's interest in the ranking process of the information flow recommendation, which makes it impossible to recommend the information that the user is actually interested in, thereby affecting the user's experience.
  • the embodiments of the present application provide a sorting method, an apparatus, and a device / terminal / server in an information flow, which completely or partially solve a problem in the prior art.
  • a ranking method in an information flow includes: using a user behavior data to train a factoring machine model to obtain hidden vectors of multiple dimensions of a user and a recommended object; Describe the hidden vectors of the user and the recommended object in multiple dimensions, extract the matching features of the user and the recommended object in each dimension; process the matching features of the user and the recommended object in each dimension to form a sample, and train a gradient boosting tree model, and Ranking scores are provided by the gradient boosted tree model.
  • a sorting device in an information flow is also provided.
  • the device includes: a vector obtaining module configured to train a factorization machine model using user behavior data to obtain users and recommended objects.
  • Hidden vectors of multiple dimensions a feature extraction module configured to extract matching features of each dimension of the user and the recommended object according to the hidden vectors of the user and the recommended object in multiple dimensions; a model training module configured to convert the The matching feature of each dimension of the user and the recommended object is processed to form a sample, a gradient boosting tree model is trained, and the gradient boosting tree model provides ranking scores.
  • a device / terminal / server including: one or more processors; a storage device, configured to store one or more programs, and when the one or more programs are Being executed by the one or more processors, so that the one or more processors implement operations corresponding to the sorting method in the information flow as described above.
  • a computer-readable storage medium in which a computer program is stored, and when the program is executed by a processor, the operations corresponding to the sorting method in the information flow as described above are implemented.
  • the embodiment of the present application uses a user behavior data training factor decomposition machine model to obtain hidden vectors of multiple dimensions of the user and the recommended object. Based on the hidden vectors of multiple dimensions of the user and the recommended object, the matching features of each dimension of the user and the recommended object are extracted, and the matched features of each dimension of the user and the recommended object are processed to form a sample to train a gradient boosting tree. Model, and ranking scores are provided by the gradient boosting tree model.
  • a combination of a factorization machine model and a gradient boosting tree model is used to sort the recommended objects in the information flow recommendation, and the information flow ranking model is optimized, so that the recommended objects can more satisfy the user's interest points and improve the user experience.
  • FIG. 1 is a flowchart of steps in a sorting method in an information flow according to the first embodiment of the present application
  • FIG. 2 is a structural block diagram of a sorting device in an information flow according to Embodiment 3 of the present application
  • FIG. 3 is a structural block diagram of a device / terminal / server according to Embodiment 5 of the present application. detailed description
  • FIG. 1 there is shown a flowchart of steps in a sorting method in an information flow according to the first embodiment of the present application.
  • steps S 101 to S 103 described in this application do not represent the order of execution.
  • Step S 101 Use the user behavior data to train a factoring machine model to obtain hidden vectors of multiple dimensions of the user and the recommended object.
  • FM Factorization Machines
  • LR logistic regression model
  • the logistic regression model (LR) adds the combined features as new features to the model by means of feature crossing.
  • the model complexity is 0 (N A 2) (N is the number of interactive features), and the memory is strong and general. Weakness.
  • FM subtly improves the generalization ability of the model by characterizing features as latent vectors and representing feature associations through the similarity (inner product) of latent vectors.
  • the complexity of the FM model is 0 (N * k) (k is the hyperparameter of the hidden vector dimension).
  • FM is essentially a linear model, and different terms affect the output of the model in a linear combination.
  • a nonlinear model is introduced to optimize the FM model.
  • the Gradient Boosted Tree Algorithm (GBM) is used as a tree model to optimize the FM nonlinear model.
  • the embodiment of the present application obtains user behavior data, and uses the user behavior data to train an FM model, thereby obtaining hidden vectors of multiple dimensions of the user and the recommended object according to the user behavior data.
  • the FM model can be considered as a neural network structure that can more closely characterize this degree of matching.
  • This embodiment of the present application uses the FM model to decompose the user behavior data (that is, the user's click and presentation data on the recommended object) to obtain the hidden vectors of each dimension of the recommended object in the user and the information stream. These hidden vectors can be used to calculate the matching degree between users and information flow objects.
  • Step S102 According to the hidden vectors of the user and the recommended object in multiple dimensions, Matching characteristics of households and recommended objects in each dimension.
  • the matching characteristics of the user and the recommended object in each dimension described in the embodiments of this application include:
  • the characteristics of the recommended object side in the information flow that is, the signals of the recommended object side in the information flow.
  • Such characteristics describe the characteristics of the recommended object in the information flow from various dimensions: heat, consumption time, quality, and so on. Such characteristics help to screen premium content and improve the quality benchmark of recommendations.
  • the correlation feature between the recommended object and the user in the information flow is used to characterize the degree of association between the user and the recommended object in the information flow.
  • the correlation degree can be characterized by the user's action of clicking the recommended object, or can be played by the user. The length of time the recommended object is portrayed, thereby increasing the personalization of the recommendation.
  • the correlation characteristics between the recommended object and the user in the information stream are obtained by calculating the matching degree between the hidden object vector and the user hidden vector in the information stream.
  • the structured information of each dimension of the recommended object in the user and the information stream is embedded into a low-dimensional hidden vector, and the matching degree between the user and the recommended object in the information stream is calculated.
  • the embodiment of the present application uses the FM model to divide all users and the features of the recommended object into hidden vectors in the same space, so all vectors are comparable. Not only the matching degree between users and the recommended objects in the information flow, but also between users and between the recommended objects in the information flow, can be obtained through simple vector operations.
  • the recommended object hidden vector and the user hidden vector in the information stream are aligned using a version alignment mechanism.
  • FM requires routine training, there is a time difference between the user's latent vector filling library and the latent vector loading of the recommended objects in the information flow, and the latent vectors of different versions of the model are not comparable.
  • the time window for routine training is usually 4-6 hours, retaining two versions of the data is sufficient to ensure that most of the hidden vectors can be aligned.
  • the version alignment mechanism retains the data of the last two versions for all hidden vectors.
  • the logic of version alignment is implemented, and the hidden vector of the latest version after alignment is used to calculate the matching degree.
  • the number of versions can be increased to ensure model alignment.
  • Step S103 After processing the matching features of the user and the recommended object in each dimension to form a sample, a gradient boosted tree model is trained, and the gradient boosted tree model provides ranking scores.
  • Gradient Boosting Tree Model (GBM), which is called Gradient Boosting Machine, is a kind of Boosting algorithm.
  • the main idea of GBM is to build the next new base learner based on the gradient descent direction of the previously established base learner's loss function. The goal is to integrate these base learners to make the overall loss function of the model continuously decrease and the model continuously improve.
  • This application combines the FM model with the GBM model to estimate the click-through rate through the GBM.
  • the FM model combined with the GBM model is a tree model, which can handle the highly nonlinear complex relationship between the signal and the target, and also has better interpretability.
  • the embodiment of the present application uses user behavior data to train a factoring machine model to obtain hidden vectors of multiple dimensions of the user and the recommended object. Based on the hidden vectors of multiple dimensions of the user and the recommended object, the matching features of each dimension of the user and the recommended object are extracted, and the matched features of each dimension of the user and the recommended object are processed to form a sample to train a gradient boosting tree. Model, and ranking scores are provided by the gradient boosting tree model.
  • the embodiment of the present application uses a combination of a factorization machine model and a gradient boosting tree model to rank recommended objects in information flow recommendation, and optimizes the information flow ranking model, so that the recommended objects can more satisfy the user's points of interest and improve the user experience.
  • the sorting method in the information flow in this embodiment may be executed by any appropriate device having a sorting capability in the information flow, including but not limited to: various device terminals or servers, including but not limited to a PC, a tablet, and a mobile Terminal, etc.
  • This embodiment includes the above steps S101 to S103.
  • the step S102 further includes: performing a normalization process on the extracted features.
  • the extracted special features are normalized and returned to the distribution server, and the distribution server returns the flow to the log server and places the order. Click logs are also placed on the log server at the same time.
  • the version alignment mechanism performs alignment, and after cleaning, filtering, and anti-cheating processing, the reflux features are extracted for model training.
  • the embodiment of the present application uses user behavior data to train a factoring machine model to obtain hidden vectors of multiple dimensions of the user and the recommended object. And based on the user and recommended Like the hidden vectors of multiple dimensions, the matching features of each dimension of the user and the recommended object are extracted, and the matched features of the user and the recommended object are processed to form a sample, and a gradient boosting tree model is trained, and the gradient boosting The tree model provides ranking scores.
  • a combination of a factorization machine model and a gradient boosting tree model is used to sort the recommended objects in the information flow recommendation, and the information flow ranking model is optimized, so that the recommended objects can more satisfy the user's interest points and improve the user experience.
  • the sorting method in the information flow in this embodiment may be executed by any appropriate device having a sorting capability in the information flow, including but not limited to: various device terminals or servers, including but not limited to a PC, a tablet, and a mobile Terminal, etc.
  • FIG. 2 a structural block diagram of a sorting device in an information flow according to Embodiment 3 of the present application is shown.
  • the vector obtaining module 201 is configured to train a factorization machine model using user behavior data to obtain hidden vectors of multiple dimensions of a user and a recommended object.
  • the feature extraction module 202 is configured to extract matching features of each dimension of the user and the recommended object according to the hidden vectors of the user and the recommended object in multiple dimensions.
  • the model training module 203 is configured to process a matching feature of the user and the recommended object in each dimension to form a sample, train a gradient boosting tree model, and provide a ranking score by the gradient boosting tree model.
  • FM Factorization Machines
  • LR logistic regression model
  • the logistic regression model (LR) adds the combined features as new features to the model by way of feature crossing.
  • the model complexity is 0 (N A 2) (N is the number of interactive features), and the memory is strong and general. Weakness.
  • FM subtly improves the generalization ability of the model by characterizing features as latent vectors and representing feature associations through the similarity (inner product) of latent vectors.
  • the complexity of the FM model is 0 (N * k) (k is the hyperparameter of the hidden vector dimension).
  • FM is essentially a linear model, and different terms affect the output of the model in a linear combination.
  • a nonlinear model is introduced to optimize the FM model.
  • the Gradient Boosted Tree Algorithm (GBM) is used as a tree model to optimize the FM nonlinear model.
  • GBM Gradient Boosted Tree Algorithm
  • the embodiment of the present application obtains user behavior data, and uses the user behavior data to train an FM model, so as to obtain hidden vectors of multiple dimensions of a user and a recommended object according to the user behavior data.
  • the FM model can be considered as a neural network structure that can more closely characterize this degree of matching.
  • This embodiment of the present application uses the FM model to decompose the user behavior data (that is, the user's click and presentation data on the recommended object) to obtain the hidden vectors of each dimension of the recommended object in the user and the information stream. These hidden vectors can be used to calculate the matching degree between users and information flow objects.
  • the matching characteristics of the user and the recommended object in each dimension described in the embodiments of this application include:
  • the characteristics of the recommended object side in the information flow that is, the signals of the recommended object side in the information flow.
  • Such characteristics describe the characteristics of the recommended object in the information flow from various dimensions: heat, consumption time, quality, and so on. Such characteristics help to screen premium content and improve the quality benchmark of recommendations.
  • the correlation feature between the recommended object and the user in the information flow is used to characterize the degree of association between the user and the recommended object in the information flow.
  • the correlation degree can be characterized by the user's action of clicking the recommended object, or can be played by the user. The length of time the recommended object is portrayed, thereby increasing the personalization of the recommendation.
  • the correlation characteristics between the recommended object and the user in the information stream are obtained by calculating the matching degree between the hidden object vector and the user hidden vector in the information stream.
  • the structured information of each dimension of the recommended object in the user and the information stream is embedded into a low-dimensional hidden vector, and the matching degree between the user and the recommended object in the information stream is calculated.
  • the embodiment of the present application uses the FM model to divide all users and the features of the recommended object into hidden vectors in the same space, so all vectors are comparable. Not only the matching degree between users and the recommended objects in the information flow, but also between users and between the recommended objects in the information flow, can be obtained through simple vector operations.
  • the recommended object hidden vector and the user hidden vector in the information stream are aligned using a version alignment mechanism.
  • FM requires routine training, there is a time difference between the user's hidden vector database and the hidden object load of the recommended object in the information stream, and the hidden vectors of different versions of the model are not comparable. Specifically, since the time window for routine training is usually 4-6 hours, retaining two versions of the data is sufficient to ensure that most of the hidden vectors can be aligned.
  • the version alignment mechanism retains the data of the last two versions for all hidden vectors.
  • the logic of version alignment is implemented, and the matching degree is calculated using the hidden vector of the latest version after alignment.
  • the number of versions can be increased to ensure model alignment.
  • GBM Gradient Boosting Tree Model
  • the main idea of GBM is to build the next new base learner based on the gradient descent direction of the loss function of the previously built base learner. The purpose is to hope that by integrating these base learners, the overall loss function of the model will continue to decline, and the model will continue to improve.
  • This application combines the FM model with the GBM model to estimate the click-through rate through the GBM.
  • the FM model combined with the GBM model is a tree model, which can handle the highly nonlinear complex relationship between the signal and the target, and also has better interpretability.
  • the embodiment of the present application uses user behavior data to train a factoring machine model to obtain hidden vectors of multiple dimensions of the user and the recommended object. Based on the hidden vectors of multiple dimensions of the user and the recommended object, the matching features of each dimension of the user and the recommended object are extracted, and the matched features of each dimension of the user and the recommended object are processed to form a sample to train a gradient boosting tree. Model, and ranking scores are provided by the gradient boosting tree model.
  • the embodiment of the present application uses a combination of a factorization machine model and a gradient boosting tree model to rank recommended objects in information flow recommendation, and optimizes the information flow ranking model, so that the recommended objects can more satisfy the user's points of interest and improve the user experience.
  • the sorting method in the information flow in this embodiment may be executed by any appropriate device having a sorting capability in the information flow, including but not limited to: various device terminals or servers, including but not limited to a PC, a tablet, and a mobile Terminal, etc.
  • This embodiment includes the above-mentioned vector obtaining module, feature extraction module, and model training module.
  • the feature extraction module further includes:
  • the extracted special features are normalized and returned to the distribution server, and the distribution server returns the flow to the log server and places the order. Click logs are also placed through the log server at the same time.
  • the embodiment of the present application uses user behavior data to train a factoring machine model to obtain hidden vectors of multiple dimensions of the user and the recommended object. Based on the hidden vectors of multiple dimensions of the user and the recommended object, the matching features of each dimension of the user and the recommended object are extracted, and the matched features of each dimension of the user and the recommended object are processed to form a sample to train a gradient boosting tree. Model, and ranking scores are provided by the gradient boosting tree model.
  • the embodiment of the present application uses a combination of a factorization machine model and a gradient boosting tree model to rank recommended objects in information flow recommendation, and optimizes the information flow ranking model, so that the recommended objects can more satisfy the user's points of interest and improve the user experience.
  • the sorting method in the information flow in this embodiment may be executed by any appropriate device having a sorting capability in the information flow, including but not limited to: various device terminals or servers, including but not limited to a PC, a tablet, and a mobile Terminal, etc.
  • FIG. 3 a structural block diagram of a device / terminal / server according to Embodiment 5 of the present application is shown.
  • the specific embodiment of the present application does not limit the specific implementation of the device / terminal / server.
  • the device / terminal / server may include: one or more processors
  • processor storage device
  • the processor 302 is configured to execute a program 306. Specifically, the processor 302 may execute related steps in the foregoing method embodiment of the information flow.
  • the program 306 may include program code, where the program code includes a computer operation instruction.
  • the processor 302 may be a central processing unit CPU, or an application specific integrated circuit (ASIC), or one or more integrated circuits configured to implement the embodiments of the present application.
  • the device / terminal / server includes one or more processors, which can be processors of the same type, such as one or more CPUs; or different types of processors, such as one or more CPUs and one or more ASICs .
  • the storage device 304 is configured to store one or more programs 306.
  • the storage device 304 may include a high-speed RAM memory, and may also include a non-volatile memory (non-volatile memory), for example, at least one magnetic disk memory.
  • the program 306 may be specifically configured to cause the processor 302 to perform the following operations: use the user behavior data to train a factorization machine model to obtain hidden vectors of multiple dimensions of the user and the recommended object; and according to the hidden dimensions of the user and the recommended object of multiple dimensions Vector to extract the matching features of the user and the recommended object in each dimension; process the matching features of the user and the recommended object in each dimension to form a sample, train a gradient boosting tree model, and provide a ranking score by the gradient boosting tree model.
  • the program 306 is further configured to perform normalization processing on the matching features of each dimension of the user and the recommended object.
  • the matching characteristics of the user and the recommended object in each dimension include: information flow object side characteristics and the recommended object and user correlation characteristics in the information flow.
  • the feature of the correlation between the recommended object and the user in the information stream is obtained by performing a matching degree calculation using the hidden vector of the recommended object and the user hidden vector in the information stream.
  • the recommended object hidden vector and the user hidden vector in the information stream are aligned using a version alignment mechanism.
  • the embodiment of the present application uses user behavior data to train a factoring machine model to obtain hidden vectors of multiple dimensions of the user and the recommended object. Based on the hidden vectors of multiple dimensions of the user and the recommended object, the matching features of each dimension of the user and the recommended object are extracted, and the matched features of each dimension of the user and the recommended object are processed to form a sample to train a gradient boosting tree. Model, and ranking scores are provided by the gradient boosting tree model.
  • the embodiment of the present application uses a combination of a factorization machine model and a gradient boosting tree model to rank recommended objects in information flow recommendation, and optimizes the information flow ranking model, so that the recommended objects can more satisfy the user's points of interest and improve the user experience.
  • each component / step described in the embodiment of the present application may be split into more components / steps, or two or more components / steps or partial operations of components / steps may be combined into New components / steps to achieve the purpose of the embodiments of the present application.
  • the process described above with reference to the flowchart may be implemented as a computer software program.
  • embodiments of the present disclosure include a computer program product including a computer program carried on a computer-readable medium, the computer program containing program code for performing a method shown in a flowchart.
  • the computer The program may be downloaded and installed from a network through a communication section, and / or installed from a removable medium.
  • CPU central processing unit
  • the computer-readable medium described in this application may be a computer-readable signal medium or a computer-readable storage medium or any combination of the foregoing.
  • the computer-readable storage medium may be, for example, but not limited to, an electronic, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of computer-readable storage media may include, but are not limited to: electrical connections with one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable memories Programming read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination of the foregoing.
  • a computer-readable storage medium may be any tangible medium containing or storing a program, and the program may be used by or in combination with an instruction execution system, apparatus, or device.
  • a computer-readable signal medium may include a data signal that is included in baseband or propagated as part of a carrier wave, and which carries computer-readable program code. Such a propagated data signal may take many forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination of the foregoing.
  • the computer-readable signal medium may also be any computer-readable medium other than a computer-readable storage medium, and the computer-readable medium may send, propagate, or transmit a program for use by or in connection with an instruction execution system, apparatus, or device .
  • Program code embodied on a computer-readable medium may be transmitted using any appropriate medium, including but not limited to: wireless, wire, optical fiber cable, RF, etc., or any suitable combination of the foregoing.
  • Computer program code for performing the operations of the present application may be written in one or more programming languages or a combination thereof, the programming languages including an object-oriented programming language such as Java, Smalltalk, C ++, and also conventional A procedural programming language such as "C" or a similar programming language.
  • the program code can be executed entirely on the user's computer, partly on the user's computer, as an independent software package, partly on the user's computer, partly on a remote computer, or entirely on the remote computer or server.
  • the remote computer can be connected to the user's computer through any kind of network, including a local area network (LAN) or a wide area network (WAN), or it can be connected to an external computer (such as through an Internet service provider using the Internet Connection).
  • LAN local area network
  • WAN wide area network
  • each block in the flowchart or block diagram may represent a module, a program segment, or a portion of a code, which module, program segment, or part of the code contains one or more functions for implementing a specified logical function Executable instructions.
  • the functions marked in the blocks may also occur in a different order than those marked in the drawings. For example, two successively represented boxes may actually be executed substantially in parallel, and they may sometimes be executed in the reverse order, depending on the functions involved.
  • each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts may be implemented in a dedicated hardware-based system that performs the specified function or operation. Or, it can be implemented by a combination of dedicated hardware and computer instructions.
  • a processor includes a receiving unit, a parsing unit, an information selecting unit, and a generating unit. Among them, the names of these units do not in any way constitute a limitation on the unit itself.
  • the receiving unit may also be described as a "unit for receiving a user's web browsing request".
  • the present application also provides a computer-readable medium having stored thereon a computer program that, when executed by a processor, implements a method as described in any one of the above embodiments.
  • the present application further provides a computer-readable medium, which may be included in the device described in the foregoing embodiments; or may exist alone without being assembled into the device.
  • the computer-readable medium carries one or more programs, and when the one or more programs are executed by the device, the device is caused to: use user behavior data to train a factoring machine model to obtain multiple dimensions of the user and the recommended object Hidden vector; extracting matching features of the user and the recommended object in each dimension according to the hidden vectors of the user and the recommended object in multiple dimensions; processing the matching features of the user and the recommended object in each dimension to form a sample to train the gradient A boosted tree model, and ranking scores are provided by the gradient boosted tree model.

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • Mathematical Physics (AREA)
  • General Engineering & Computer Science (AREA)
  • Software Systems (AREA)
  • Mathematical Optimization (AREA)
  • Mathematical Analysis (AREA)
  • Databases & Information Systems (AREA)
  • Pure & Applied Mathematics (AREA)
  • Computing Systems (AREA)
  • Computational Mathematics (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Artificial Intelligence (AREA)
  • Evolutionary Computation (AREA)
  • Algebra (AREA)
  • Medical Informatics (AREA)
  • User Interface Of Digital Computer (AREA)
  • Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
  • Two-Way Televisions, Distribution Of Moving Picture Or The Like (AREA)

Abstract

本申请实施例提供了一种信息流中的排序方法、装置和设备 /终端 / 服务器,所述方法包括:使用用户行为数据训练因子分解机模型,获得用户与被推荐对象多个维度的隐向量;根据所述用户与被推荐对象多个维度的隐向量,提取用户与被推荐对象各个维度的匹配特征;将所述用户与被推荐对象各个维度的匹配特征处理后构成样本,训练梯度提升树模型,并由所述梯度提升树模型提供排序打分。本申请实施例采用了因子分解机和梯度提升树模型结合进行信息流中被推荐对象的排序,优化了信息流排序模型,令推荐的对象能够更加满足用户的兴趣点,改善用户体验。

Description

一种信息流中的排序方法、 装置和设备 /终端 /服务器
申请要求在 2018 年 08 月 30 日提交中国专利局、 申请号为 201811004222.5、 发明名称为“一种信息流中的排序方法、 装置和设备 / 终端 /服务器”的中国专利申请的优先权, 其全部内容通过引用结合在本 申请中。 技术领域
本申请涉及互联网技术领域, 尤其涉及一种信息流中的排序方法、 装置和设备 /终端 /服务器。
背景技术
随着互联网技术的发展,信息流成为用户日常生活娱乐的主要方式, 通过排序算法提供用户感兴趣的信息 (文章 /视频等) 成为各信息流应用 程序吸引用户的重要手段。 如何向用户推荐符合其兴趣的对象, 通常通 过算法模型的计算来实现。
但是通常算法模型在信息流推荐的排序过程中无法准确表述信息流 对象和用户兴趣之间的关系,造成无法向用户推荐其实际感兴趣的信息, 从而影响用户的使用感受。
因此, 如何更佳的实现信息流中的排序成为现有技术中亟待解决的 技术问题。
发明内容
本申请实施例提供了一种信息流中的排序方法、 装置和设备 /终端 / 服务器, 全部或者部分解决现有技术中存在的问题。
根据本申请实施例的一个方面, 提供了一种信息流中的排序方法, 所述方法包括: 使用用户行为数据训练因子分解机模型, 获得用户与被 推荐对象多个维度的隐向量; 根据所述用户与被推荐对象多个维度的隐 向量, 提取用户与被推荐对象各个维度的匹配特征; 将所述用户与被推 荐对象各个维度的匹配特征处理后构成样本, 训练梯度提升树模型, 并 由所述梯度提升树模型提供排序打分。 根据本申请实施例的另一个方面, 还提供了一种信息流中的排序装 置, 所述装置包括: 向量获得模块, 设置用于使用用户行为数据训练因 子分解机模型, 获得用户与被推荐对象多个维度的隐向量; 特征提取模 块, 设置用于根据所述用户与被推荐对象多个维度的隐向量, 提取用户 与被推荐对象各个维度的匹配特征; 模型训练模块, 设置用于将所述用 户与被推荐对象各个维度的匹配特征处理后构成样本, 训练梯度提升树 模型, 并由所述梯度提升树模型提供排序打分
根据本申请实施例的又一个方面, 还提供了一种设备 /终端 /服务器, 包括: 一个或多个处理器; 存储装置, 用于存储一个或多个程序, 当所 述一个或多个程序被所述一个或多个处理器执行, 使得所述一个或多个 处理器实现如上所述的信息流中的排序方法对应的操作。
根据本申请实施例的又一个方面, 还提供了一种计算机可读存储介 质, 其上存储有计算机程序, 该程序被处理器执行时实现如上所述的信 息流中的排序方法对应的操作。
根据本申请实施例提供的技术方案, 本申请实施例使用用户行为数 据训练因子分解机模型, 获得用户与被推荐对象多个维度的隐向量。 并 根据所述用户与被推荐对象多个维度的隐向量, 提取用户与被推荐对象 各个维度的匹配特征, 将所述用户与被推荐对象各个维度的匹配特征处 理后构成样本, 训练梯度提升树模型, 并由所述梯度提升树模型提供排 序打分。 本申请实施例采用了因子分解机模型和梯度提升树模型结合进 行信息流推荐中被推荐对象的排序, 优化了信息流排序模型, 令推荐的 对象能够更加满足用户的兴趣点, 改善用户体验。
附图说明
图 1是根据本申请实施例一的一种信息流中的排序方法的步骤流程 图;
图 2是根据本申请实施例三的一种信息流中的排序装置的结构框图; 图 3是根据本申请实施例五的一种设备 /终端 /服务器的结构框图。 具体实施方式
下面结合附图(若干附图中相同的标号表示相同的元素)和实施例, 对本申请实施例的具体实施方式作进一步详细说明。 以下实施例用于说 明本申请, 但不用来限制本申请的范围。 本领域技术人员可以理解, 本申请实施例中的“第一”、 “第二”等术 语仅用于区别不同步骤、 设备或模块等, 既不代表任何特定技术含义, 也不表示它们之间的必然逻辑顺序。
实施例一
参照图 1, 示出了根据本申请实施例一的一种信息流中的排序方法 的步骤流程图。
值得说明的是, 本申请所述步骤 S 101至 S 103并不代表其执行的先 后顺序。
本实施例的信息流中的排序方法包括以下步骤:
步骤 S 101 : 使用用户行为数据训练因子分解机模型, 获得用户与被 推荐对象多个维度的隐向量。
因子分解机模型 (FM, Factorization Machines) 是对传统逻辑回归 模型(LR)在处理高阶交互特征问题上的优化和改进。逻辑回归模型(LR) 通过特征交叉的方式, 将组合后的特征作为新特征加入到模型中, 模型 复杂度为 0(NA2) (N 为交互特征的数量) , 记忆性较强而泛化性偏弱。 FM通过将特征表征为隐向量, 通过隐向量的相似度(内积)来表示特征 关联这种方式来巧妙地提升模型的泛化能力。 FM模型的复杂度为 0(N*k) (k为隐向量维度超参) 。
FM本质上是一个线性模型,不同项之间以线性组合的方式影响模型 的输出。 引入非线性的模型来优化 FM模型, 梯度提升树算法 (GBM) 作为树模型可以优化 FM非线性模型。
本申请实施例获得用户行为数据, 并利用所述用户行为数据来训练 FM模型,从而根据所述用户行为数据得到用户与被推荐对象多个维度的 隐向量。
从模型结构看, FM模型可以认为是能够更加紧密刻画这种匹配度的 神经网络结构。 本申请实施例利用 FM模型分解用户行为数据 (即, 用 户对被推荐对象的点击和展现数据) , 得到用户和信息流中被推荐对象 各个维度的隐向量。 进而可以通过这些隐向量计算用户和信息流对象的 匹配度。
步骤 S 102 : 根据所述用户与被推荐对象多个维度的隐向量, 提取用 户与被推荐对象各个维度的匹配特征。
在本申请一具体实现中, 本申请实施例所述用户与被推荐对象各个 维度的匹配特征包括:
信息流中被推荐对象侧特征和信息流中被推荐对象与用户相关性特 征。
所述信息流中被推荐对象侧特征, 即信息流中被推荐对象侧信号, 这类特征从各个维度刻画了信息流中被推荐对象的特性: 热度、 消费时 长、 质量等。 这类特征有助于筛选精品内容, 提升推荐的质量基准。
所述信息流中被推荐对象与用户相关性特征用于刻画用户和信息流 中被推荐对象的关联度, 所述关联度可以通过用户的点击被推荐对象的 动作进行刻画, 也可以通过用户播放被推荐对象的时长刻画, 从而提升 推荐的个性化。
所述信息流中被推荐对象与用户相关性特征采用信息流中被推荐对 象隐向量和用户隐向量进行匹配度计算获得。
具体地, 将用户和信息流中被推荐对象各个维度的结构化信息嵌入 到一个低维隐向量, 计算用户和信息流中被推荐对象的匹配度。
本申请实施例是通过 FM模型, 将所有用户与被推荐对象特征都分 成在同一个空间内的隐向量, 因而所有的向量都是可比的。 不仅用户与 信息流中被推荐对象各个维度的匹配度, 甚至用户之间、 信息流中被推 荐对象之间, 都可以通过简单的向量运算得到匹配度。
在本申请另一具体实现中, 所述信息流中被推荐对象隐向量和用户 隐向量采用版本对齐机制令所述其能够对齐。
由于 FM需要例行训练, 用户隐向量灌库和信息流中被推荐对象隐 向量加载之间存在时间差, 而不同版本模型的隐向量之间是不可比的。
具体地, 由于例行训练的时间窗口通常为 4~6个小时, 保留两个版 本的数据是足以保证绝大部分隐向量能够对齐。 所述版本对齐机制将所 有隐向量都会保留最近两个版本的数据。 在 FM在线计算模块中, 实现 版本对齐的逻辑, 用对齐后最新版本的隐向量计算匹配度。 此外, 在更 加高频的模型训练中, 可以增加版本的数量来确保模型对齐。
步骤 S 103 : 将所述用户与被推荐对象各个维度的匹配特征处理后构 成样本,训练梯度提升树模型,并由所述梯度提升树模型提供排序打分。 梯度提升树模型 (GBM) , 全称为 Gradient Boosting Machine, 是 Boosting算法的一种。 GBM主要思想是基于之前建立的基学习器的损失 函数的梯度下降方向来建立下一个新的基学习器, 目的就是希望通过集 成这些基学习器使得模型总体的损失函数不断下降, 模型不断改进。
本申请将 FM模型结合 GBM模型,通过 GBM进行点击率预估。 FM 模型结合 GBM 模型是树模型, 能够处理信号与目标间高度非线性的复 杂关系, 也具备更好的可解释性。
由此可知, 本申请实施例使用用户行为数据训练因子分解机模型, 获得用户与被推荐对象多个维度的隐向量。 并根据所述用户与被推荐对 象多个维度的隐向量, 提取用户与被推荐对象各个维度的匹配特征, 将 所述用户与被推荐对象各个维度的匹配特征处理后构成样本, 训练梯度 提升树模型, 并由所述梯度提升树模型提供排序打分。 本申请实施例采 用了因子分解机模型和梯度提升树模型结合进行信息流推荐中被推荐对 象的排序, 优化了信息流排序模型, 令推荐的对象能够更加满足用户的 兴趣点, 改善用户体验。
本实施例的信息流中的排序方法可以由任意适当的具有信息流中的 排序能力的设备执行, 包括但不限于: 各种设备终端或者服务端, 包括 但不限于 PC机、 平板电脑、 移动终端等。
实施例二
本实施例包括上述步骤 S 101至步骤 S 103。 所述步骤 S 102还包括: 对提取的所述特征进行归一化处理。
由于本申请实施例提取的特征如未经过归一化处理, 则不同用户间 的特征进行打分会造成差异显著,令 FM模型获得的特征利用 GBM模型 训练过程很难收敛。
因此本申请实施例将提取的所述特正经归一化后返回给分发服务器, 由分发服务器回流至日志服务器落盘。 点击日志也同时经由日志服务器 落盘。经过所述版本对齐机制进行对齐,经清洗、过滤和反作弊处理后, 提取回流特征用于模型训练。
由此可知, 本申请实施例使用用户行为数据训练因子分解机模型, 获得用户与被推荐对象多个维度的隐向量。 并根据所述用户与被推荐对 象多个维度的隐向量, 提取用户与被推荐对象各个维度的匹配特征, 将 所述用户与被推荐对象各个维度的匹配特征处理后构成样本, 训练梯度 提升树模型, 并由所述梯度提升树模型提供排序打分。 本申请实施例采 用了因子分解机模型和梯度提升树模型结合进行信息流推荐中被推荐对 象的排序, 优化了信息流排序模型, 令推荐的对象能够更加满足用户的 兴趣点, 改善用户体验。
本实施例的信息流中的排序方法可以由任意适当的具有信息流中的 排序能力的设备执行, 包括但不限于: 各种设备终端或者服务端, 包括 但不限于 PC机、 平板电脑、 移动终端等。
实施例三
参照图 2 , 示出了根据本申请实施例三的一种信息流中的排序装置 的结构框图。
本实施例的信息流中的排序装置包括:
向量获得模块 201, 设置用于使用用户行为数据训练因子分解机模 型, 获得用户与被推荐对象多个维度的隐向量。
特征提取模块 202, 设置用于根据所述用户与被推荐对象多个维度 的隐向量, 提取用户与被推荐对象各个维度的匹配特征。
模型训练模块 203, 设置用于将所述用户与被推荐对象各个维度的 匹配特征处理后构成样本, 训练梯度提升树模型, 并由所述梯度提升树 模型提供排序打分。
因子分解机模型 (FM, Factorization Machines) 是对传统逻辑回归 模型(LR)在处理高阶交互特征问题上的优化和改进。逻辑回归模型(LR) 通过特征交叉的方式, 将组合后的特征作为新特征加入到模型中, 模型 复杂度为 0(NA2) (N 为交互特征的数量) , 记忆性较强而泛化性偏弱。 FM通过将特征表征为隐向量, 通过隐向量的相似度(内积)来表示特征 关联这种方式来巧妙地提升模型的泛化能力。 FM模型的复杂度为 0(N*k) (k为隐向量维度超参) 。
FM本质上是一个线性模型,不同项之间以线性组合的方式影响模型 的输出。 引入非线性的模型来优化 FM模型, 梯度提升树算法 (GBM) 作为树模型可以优化 FM非线性模型。 本申请实施例获得用户行为数据, 并利用所述用户行为数据来训练 FM模型,从而根据所述用户行为数据得到用户与被推荐对象多个维度的 隐向量。
从模型结构看, FM模型可以认为是能够更加紧密刻画这种匹配度的 神经网络结构。 本申请实施例利用 FM模型分解用户行为数据 (即, 用 户对被推荐对象的点击和展现数据) , 得到用户和信息流中被推荐对象 各个维度的隐向量。 进而可以通过这些隐向量计算用户和信息流对象的 匹配度。
在本申请一具体实现中, 本申请实施例所述用户与被推荐对象各个 维度的匹配特征包括:
信息流中被推荐对象侧特征和信息流中被推荐对象与用户相关性特 征。
所述信息流中被推荐对象侧特征, 即信息流中被推荐对象侧信号, 这类特征从各个维度刻画了信息流中被推荐对象的特性: 热度、 消费时 长、 质量等。 这类特征有助于筛选精品内容, 提升推荐的质量基准。
所述信息流中被推荐对象与用户相关性特征用于刻画用户和信息流 中被推荐对象的关联度, 所述关联度可以通过用户的点击被推荐对象的 动作进行刻画, 也可以通过用户播放被推荐对象的时长刻画, 从而提升 推荐的个性化。
所述信息流中被推荐对象与用户相关性特征采用信息流中被推荐对 象隐向量和用户隐向量进行匹配度计算获得。
具体地, 将用户和信息流中被推荐对象各个维度的结构化信息嵌入 到一个低维隐向量, 计算用户和信息流中被推荐对象的匹配度。
本申请实施例是通过 FM模型, 将所有用户与被推荐对象特征都分 成在同一个空间内的隐向量, 因而所有的向量都是可比的。 不仅用户与 信息流中被推荐对象各个维度的匹配度, 甚至用户之间、 信息流中被推 荐对象之间, 都可以通过简单的向量运算得到匹配度。
在本申请另一具体实现中, 所述信息流中被推荐对象隐向量和用户 隐向量采用版本对齐机制令所述其能够对齐。
由于 FM需要例行训练, 用户隐向量灌库和信息流中被推荐对象隐 向量加载之间存在时间差, 而不同版本模型的隐向量之间是不可比的。 具体地, 由于例行训练的时间窗口通常为 4〜 6个小时, 保留两个版 本的数据是足以保证绝大部分隐向量能够对齐。 所述版本对齐机制将所 有隐向量都会保留最近两个版本的数据。 在 FM在线计算模块中, 实现 版本对齐的逻辑, 用对齐后最新版本的隐向量计算匹配度。 此外, 在更 加高频的模型训练中, 可以增加版本的数量来确保模型对齐。
梯度提升树模型 (GBM) , 全称为 Gradient Boosting Machine, 是 Boosting算法的一种。 GBM主要思想是基于之前建立的基学习器的损失 函数的梯度下降方向来建立下一个新的基学习器, 目的就是希望通过集 成这些基学习器使得模型总体的损失函数不断下降, 模型不断改进。
本申请将 FM模型结合 GBM模型,通过 GBM进行点击率预估。 FM 模型结合 GBM 模型是树模型, 能够处理信号与目标间高度非线性的复 杂关系, 也具备更好的可解释性。
由此可知, 本申请实施例使用用户行为数据训练因子分解机模型, 获得用户与被推荐对象多个维度的隐向量。 并根据所述用户与被推荐对 象多个维度的隐向量, 提取用户与被推荐对象各个维度的匹配特征, 将 所述用户与被推荐对象各个维度的匹配特征处理后构成样本, 训练梯度 提升树模型, 并由所述梯度提升树模型提供排序打分。 本申请实施例采 用了因子分解机模型和梯度提升树模型结合进行信息流推荐中被推荐对 象的排序, 优化了信息流排序模型, 令推荐的对象能够更加满足用户的 兴趣点, 改善用户体验。
本实施例的信息流中的排序方法可以由任意适当的具有信息流中的 排序能力的设备执行, 包括但不限于: 各种设备终端或者服务端, 包括 但不限于 PC机、 平板电脑、 移动终端等。
实施例四
本实施例包括上述向量获得模块、 特征提取模块、 模型训练模块。 所述特征提取模块还包括:
对提取的所述特征进行归一化处理。
由于本申请实施例提取的特征如未经过归一化处理, 则不同用户间 的特征进行打分会造成差异显著,令 FM模型获得的特征利用 GBM模型 训练过程很难收敛。 因此本申请实施例将提取的所述特正经归一化后返回给分发服务器, 由分发服务器回流至日志服务器落盘。 点击日志也同时经由日志服务器 落盘。经过所述版本对齐机制进行对齐,经清洗、过滤和反作弊处理后, 提取回流特征用于模型训练。
由此可知, 本申请实施例使用用户行为数据训练因子分解机模型, 获得用户与被推荐对象多个维度的隐向量。 并根据所述用户与被推荐对 象多个维度的隐向量, 提取用户与被推荐对象各个维度的匹配特征, 将 所述用户与被推荐对象各个维度的匹配特征处理后构成样本, 训练梯度 提升树模型, 并由所述梯度提升树模型提供排序打分。 本申请实施例采 用了因子分解机模型和梯度提升树模型结合进行信息流推荐中被推荐对 象的排序, 优化了信息流排序模型, 令推荐的对象能够更加满足用户的 兴趣点, 改善用户体验。
本实施例的信息流中的排序方法可以由任意适当的具有信息流中的 排序能力的设备执行, 包括但不限于: 各种设备终端或者服务端, 包括 但不限于 PC机、 平板电脑、 移动终端等。
实施例五
参照图 3 , 示出了根据本申请实施例五的一种设备 /终端 /服务器的结 构框图, 本申请具体实施例并不对设备 /终端 /服务器的具体实现做限定。
如图 3 所示, 该设备 /终端 /服务器可以包括: 一个或者多个处理器
(processor)302、 存储装置 (memory )304。
其中:
处理器 302 , 用于执行程序 306 , 具体可以执行上述信息流中的排序 方法实施例中的相关步骤。
具体地, 程序 306可以包括程序代码, 该程序代码包括计算机操作 指令。
处理器 302 可能是中央处理器 CPU, 或者是特定集成电路 ASIC ( Application Specific Integrated Circuit ) , 或者是被配置成实施本申请 实施例的一个或多个集成电路。设备 /终端 /服务器包括的一个或多个处理 器, 可以是同一类型的处理器, 如一个或多个 CPU; 也可以是不同类型 的处理器, 如一个或多个 CPU以及一个或多个 ASIC。 存储装置 304, 用于存放一个或多个程序 306。 存储装置 304可能包 含高速 RAM 存储器, 也可能还包括非易失性存储器 ( non-volatile memory ) , 例如至少一个磁盘存储器。
程序 306具体可以用于使得处理器 302执行以下操作: 使用用户行 为数据训练因子分解机模型,获得用户与被推荐对象多个维度的隐向量; 根据所述用户与被推荐对象多个维度的隐向量, 提取用户与被推荐对象 各个维度的匹配特征; 将所述用户与被推荐对象各个维度的匹配特征处 理后构成样本, 训练梯度提升树模型, 并由所述梯度提升树模型提供排 序打分。
在一种可选的实施方式中, 程序 306还用于将所述用户与被推荐对 象各个维度的匹配特征进行归一化处理。
在一种可选的实施方式中, 所述用户与被推荐对象各个维度的匹配 特征包括:信息流对象侧特征和信息流中被推荐对象与用户相关性特征。
在一种可选的实施方式中, 所述信息流中被推荐对象与用户相关性 特征采用信息流中被推荐对象隐向量和用户隐向量进行匹配度计算获得。
在一种可选的实施方式中, 所述信息流中被推荐对象隐向量和用户 隐向量采用版本对齐机制令所述其能够对齐。
由此可知, 本申请实施例使用用户行为数据训练因子分解机模型, 获得用户与被推荐对象多个维度的隐向量。 并根据所述用户与被推荐对 象多个维度的隐向量, 提取用户与被推荐对象各个维度的匹配特征, 将 所述用户与被推荐对象各个维度的匹配特征处理后构成样本, 训练梯度 提升树模型, 并由所述梯度提升树模型提供排序打分。 本申请实施例采 用了因子分解机模型和梯度提升树模型结合进行信息流推荐中被推荐对 象的排序, 优化了信息流排序模型, 令推荐的对象能够更加满足用户的 兴趣点, 改善用户体验。
需要指出, 根据实施的需要, 可将本申请实施例中描述的各个部件 / 步骤拆分为更多部件 /步骤, 也可将两个或多个部件 /步骤或者部件 /步骤 的部分操作组合成新的部件 /步骤, 以实现本申请实施例的目的。
特别地, 根据本公开的实施例, 上文参考流程图描述的过程可以被 实现为计算机软件程序。 例如, 本公开的实施例包括一种计算机程序产 品, 其包括承载在计算机可读介质上的计算机程序, 该计算机程序包含 用于执行流程图所示的方法的程序代码。 在这样的实施例中, 该计算机 程序可以通过通信部分从网络上被下载和安装, 和 /或从可拆卸介质被安 装。 在该计算机程序被中央处理单元 (CPU) 执行时, 执行本申请的方 法中限定的上述功能。 需要说明的是, 本申请所述的计算机可读介质可 以是计算机可读信号介质或者计算机可读存储介质或者是上述两者的任 意组合。计算机可读存储介质例如可以是一一但不限于一一电、磁、光、 电磁、 红外线、 或半导体的系统、 装置或器件, 或者任意以上的组合。 计算机可读存储介质的更具体的例子可以包括但不限于: 具有一个或多 个导线的电连接、便携式计算机磁盘、硬盘、 随机访问存储器(RAM)、 只读存储器 (ROM) 、 可擦式可编程只读存储器 (EPROM 或闪存) 、 光纤、 便携式紧凑磁盘只读存储器 (CD-ROM) 、 光存储器件、 磁存储 器件、 或者上述的任意合适的组合。 在本申请中, 计算机可读存储介质 可以是任何包含或存储程序的有形介质, 该程序可以被指令执行系统、 装置或者器件使用或者与其结合使用。 而在本申请中, 计算机可读的信 号介质可以包括在基带中或者作为载波一部分传播的数据信号, 其中承 载了计算机可读的程序代码。 这种传播的数据信号可以采用多种形式, 包括但不限于电磁信号、 光信号或上述的任意合适的组合。 计算机可读 的信号介质还可以是计算机可读存储介质以外的任何计算机可读介质, 该计算机可读介质可以发送、 传播或者传输用于由指令执行系统、 装置 或者器件使用或者与其结合使用的程序。 计算机可读介质上包含的程序 代码可以用任何适当的介质传输, 包括但不限于: 无线、 电线、 光缆、 RF等等, 或者上述的任意合适的组合。
可以以一种或多种程序设计语言或其组合来编写用于执行本申请的 操作的计算机程序代码, 所述程序设计语言包括面向对象的程序设计语 言一诸如 Java、 Smalltalk、 C++, 还包括常规的过程式程序设计语言一 诸如” C”语言或类似的程序设计语言。 程序代码可以完全地在用户计算 机上执行、 部分地在用户计算机上执行、 作为一个独立的软件包执行、 部分在用户计算机上部分在远程计算机上执行、 或者完全在远程计算机 或服务器上执行。 在涉及远程计算机的情形中, 远程计算机可以通过任 意种类的网络—包括局域网(LAN)或广域网(WAN)—连接到用户计算 机, 或者, 可以连接到外部计算机 (例如利用因特网服务提供商来通过 因特网连接) 。
附图中的流程图和框图, 图示了按照本申请各种实施例的系统、 方 法和计算机程序产品的可能实现的体系架构、 功能和操作。 在这点上, 流程图或框图中的每个方框可以代表一个模块、 程序段、 或代码的一部 分, 该模块、 程序段、 或代码的一部分包含一个或多个用于实现规定的 逻辑功能的可执行指令。 也应当注意, 在有些作为替换的实现中, 方框 中所标注的功能也可以以不同于附图中所标注的顺序发生。 例如, 两个 接连地表示的方框实际上可以基本并行地执行, 它们有时也可以按相反 的顺序执行, 这依所涉及的功能而定。 也要注意的是, 框图和 /或流程图 中的每个方框、 以及框图和 /或流程图中的方框的组合, 可以用执行规定 的功能或操作的专用的基于硬件的系统来实现, 或者可以用专用硬件与 计算机指令的组合来实现。
描述于本申请实施例中所涉及到的单元可以通过软件的方式实现, 也可以通过硬件的方式来实现。 所描述的单元也可以设置在处理器中, 例如, 可以描述为: 一种处理器包括接收单元、 解析单元、 信息选取单 元和生成单元。 其中, 这些单元的名称在某种情况下并不构成对该单元 本身的限定, 例如, 接收单元还可以被描述为 “接收用户的网页浏览请 求的单元” 。
作为另一方面, 本申请还提供了一种计算机可读介质, 其上存储有 计算机程序, 该程序被处理器执行时实现如上述任一实施例中所描述的 方法。
作为另一方面, 本申请还提供了一种计算机可读介质, 该计算机可 读介质可以是上述实施例中描述的装置中所包含的;也可以是单独存在, 而未装配入该装置中。 上述计算机可读介质承载有一个或者多个程序, 当上述一个或者多个程序被该装置执行时, 使得该装置: 使用用户行为 数据训练因子分解机模型, 获得用户与被推荐对象多个维度的隐向量; 根据所述用户与被推荐对象多个维度的隐向量, 提取用户与被推荐对象 各个维度的匹配特征; 将所述用户与被推荐对象各个维度的匹配特征处 理后构成样本, 训练梯度提升树模型, 并由所述梯度提升树模型提供排 序打分。
以上描述仅为本申请的较佳实施例以及对所运用技术原理的说明。 本领域技术人员应当理解, 本申请中所涉及的发明范围, 并不限于上述 技术特征的特定组合而成的技术方案, 同时也应涵盖在不脱离上述发明 构思的情况下, 由上述技术特征或其等同特征进行任意组合而形成的其 它技术方案。 例如上述特征与本申请中公开的 (但不限于) 具有类似功 能的技术特征进行互相替换而形成的技术方案。

Claims

权 利 要 求 书
1、 一种信息流中的排序方法, 其特征在于, 所述方法包括: 使用用户行为数据训练因子分解机模型, 获得用户与被推荐对象多 个维度的隐向量;
根据所述用户与被推荐对象多个维度的隐向量, 提取用户与被推荐 对象各个维度的匹配特征;
将所述用户与被推荐对象各个维度的匹配特征处理后构成样本, 训 练梯度提升树模型, 并由所述梯度提升树模型提供排序打分。
2、 根据权利要求 1所述的方法, 其特征在于, 所述将所述用户与被 推荐对象各个维度的匹配特征处理后构成样本, 训练梯度提升树模型, 并由所述梯度提升树模型提供排序打分还包括:
将所述用户与被推荐对象各个维度的匹配特征进行归一化处理。
3、 根据权利要求 1所述的方法, 其特征在于, 所述用户与被推荐对 象各个维度的匹配特征包括:
信息流中被推荐对象侧特征和信息流中被推荐对象与用户相关性特 征。
4、 根据权利要求 3所述的方法, 其特征在于, 所述信息流中被推荐 对象与用户相关性特征采用信息流对象隐向量和用户隐向量进行匹配度 计算获得。
5、 根据权利要求 4所述的方法, 其特征在于, 所述信息流对象隐向 量和用户隐向量采用版本对齐机制令所述其能够对齐。
6、 一种信息流中的排序装置, 其特征在于, 所述装置包括: 向量获得模块, 设置用于使用用户行为数据训练因子分解机模型, 获得用户与被推荐对象多个维度的隐向量;
特征提取模块, 设置用于根据所述用户与被推荐对象多个维度的隐 向量, 提取用户与被推荐对象各个维度的匹配特征;
模型训练模块, 设置用于将所述用户与被推荐对象各个维度的匹配 特征处理后构成样本, 训练梯度提升树模型, 并由所述梯度提升树模型 提供排序打分。
7、 根据权利要求 6所述的装置, 其特征在于, 所述模型训练模块还 设置用于: 将所述用户与被推荐对象各个维度的匹配特征进行归一化处理。
8、 根据权利要求 6所述的装置, 其特征在于, 所述用户与被推荐对 象各个维度的匹配特征包括:
信息流中被推荐对象侧特征和信息流中被推荐对象与用户相关性特 征。
9、 根据权利要求 8所述的装置, 其特征在于, 所述信息流中被推荐 对象与用户相关性特征采用信息流对象隐向量和用户隐向量进行匹配度 计算获得。
10、 根据权利要求 9所述的装置, 其特征在于, 所述信息流对象隐 向量和用户隐向量采用版本对齐机制令所述其能够对齐。
1 1、 一种设备 /终端 /服务器, 包括:
一个或多个处理器;
存储装置, 用于存储一个或多个程序,
当所述一个或多个程序被所述一个或多个处理器执行, 使得所述一 个或多个处理器实现如权利要求 1-5中任一所述的方法。
12、 一种计算机可读存储介质, 其上存储有计算机程序, 其特征在 于, 该程序被处理器执行时实现如权利要求 1-5中任一所述的方法。
PCT/IB2018/057166 2018-08-30 2018-09-18 一种信息流中的排序方法、装置和设备/终端/服务器 Ceased WO2020044098A2 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN201811004222.5 2018-08-30
CN201811004222.5A CN109033460A (zh) 2018-08-30 2018-08-30 一种信息流中的排序方法、装置和设备/终端/服务器

Publications (2)

Publication Number Publication Date
WO2020044098A2 true WO2020044098A2 (zh) 2020-03-05
WO2020044098A3 WO2020044098A3 (zh) 2020-08-13

Family

ID=64625867

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/IB2018/057166 Ceased WO2020044098A2 (zh) 2018-08-30 2018-09-18 一种信息流中的排序方法、装置和设备/终端/服务器

Country Status (2)

Country Link
CN (1) CN109033460A (zh)
WO (1) WO2020044098A2 (zh)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112801760A (zh) * 2021-03-30 2021-05-14 南京蓝鲸人网络科技有限公司 一种内容个性化推荐系统的排序优化方法及系统

Families Citing this family (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111695960A (zh) * 2019-03-12 2020-09-22 阿里巴巴集团控股有限公司 一种对象推荐系统、方法、电子设备及存储介质
CN111950733B (zh) * 2019-05-15 2024-06-11 阿里巴巴集团控股有限公司 信息流的排序方法、装置及计算机存储介质
CN113836406A (zh) * 2021-09-10 2021-12-24 北京小米移动软件有限公司 信息流推荐方法及装置

Family Cites Families (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN107562758B (zh) * 2016-06-30 2020-12-01 北京金山安全软件有限公司 一种推送信息的方法、装置及电子设备
CN106250461A (zh) * 2016-07-28 2016-12-21 北京北信源软件股份有限公司 一种基于Spark框架利用梯度提升决策树进行数据挖掘的算法
CN107515909B (zh) * 2017-08-11 2020-05-19 深圳市云网拜特科技有限公司 一种视频推荐方法及系统
CN107729488A (zh) * 2017-10-17 2018-02-23 北京搜狐新媒体信息技术有限公司 一种信息推荐方法及装置
CN108427708B (zh) * 2018-01-25 2021-06-25 腾讯科技(深圳)有限公司 数据处理方法、装置、存储介质和电子装置

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112801760A (zh) * 2021-03-30 2021-05-14 南京蓝鲸人网络科技有限公司 一种内容个性化推荐系统的排序优化方法及系统

Also Published As

Publication number Publication date
CN109033460A (zh) 2018-12-18
WO2020044098A3 (zh) 2020-08-13

Similar Documents

Publication Publication Date Title
JP6745384B2 (ja) 情報をプッシュするための方法及び装置
CN109492772B (zh) 生成信息的方法和装置
US9384233B2 (en) Product synthesis from multiple sources
JP7150090B2 (ja) ショッピング検索のための商品属性抽出方法
CN105677931B (zh) 信息搜索方法和装置
CN107273436A (zh) 一种推荐模型的训练方法和训练装置
WO2018149115A1 (zh) 用于提供搜索结果的方法和装置
CN111159578B (zh) 一种推荐对象的方法和系统
CN107506402A (zh) 搜索结果的排序方法、装置、设备及计算机可读存储介质
CN105144164A (zh) 使用深度网络对概念术语评分
CN109241403B (zh) 项目推荐方法、装置、机器设备和计算机可读存储介质
CN106030569A (zh) 重构搜索查询的技术
US11599822B1 (en) Generation and use of literary work signatures reflective of entity relationships
CN106484766B (zh) 基于人工智能的搜索方法和装置
WO2020044098A2 (zh) 一种信息流中的排序方法、装置和设备/终端/服务器
WO2020238502A1 (zh) 物品推荐方法及装置、电子设备及存储介质
CN113934851A (zh) 用于文本分类的数据增强方法、装置及电子设备
CN112650942A (zh) 产品推荐方法、装置、计算机系统和计算机可读存储介质
US20250316062A1 (en) Self-supervised audio-visual learning for correlating music and video
JP7849447B2 (ja) インタラクション情報の特定方法、装置、電子機器、記憶媒体及びコンピュータプログラム
WO2024230757A1 (zh) 一种数据处理方法及相关装置
CN111046285B (zh) 一种推荐排序确定方法、装置、服务器及存储介质
CN108563648A (zh) 数据显示方法和装置、存储介质及电子装置
CN115146157A (zh) 一种推荐列表生成方法、匹配模型训练方法、装置及设备
US20240386048A1 (en) Natural language-guided music audio recommendation for video using machine learning

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: 18931398

Country of ref document: EP

Kind code of ref document: A2

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 18931398

Country of ref document: EP

Kind code of ref document: A2