WO2019001463A1 - 数据处理方法及装置 - Google Patents

数据处理方法及装置 Download PDF

Info

Publication number
WO2019001463A1
WO2019001463A1 PCT/CN2018/093093 CN2018093093W WO2019001463A1 WO 2019001463 A1 WO2019001463 A1 WO 2019001463A1 CN 2018093093 W CN2018093093 W CN 2018093093W WO 2019001463 A1 WO2019001463 A1 WO 2019001463A1
Authority
WO
WIPO (PCT)
Prior art keywords
data
processing
feature data
order
feature
Prior art date
Application number
PCT/CN2018/093093
Other languages
English (en)
French (fr)
Inventor
唐睿明
郭慧丰
李震国
何秀强
Original Assignee
华为技术有限公司
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 华为技术有限公司 filed Critical 华为技术有限公司
Publication of WO2019001463A1 publication Critical patent/WO2019001463A1/zh
Priority to US16/729,043 priority Critical patent/US11334758B2/en
Priority to US17/661,448 priority patent/US11748452B2/en

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/23—Clustering techniques
    • G06F18/232—Non-hierarchical techniques
    • G06F18/2321—Non-hierarchical techniques using statistics or function optimisation, e.g. modelling of probability density functions
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/213—Feature extraction, e.g. by transforming the feature space; Summarisation; Mappings, e.g. subspace methods
    • G06F18/2133—Feature extraction, e.g. by transforming the feature space; Summarisation; Mappings, e.g. subspace methods based on naturality criteria, e.g. with non-negative factorisation or negative correlation
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/24—Classification techniques
    • G06F18/245—Classification techniques relating to the decision surface
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/24—Classification techniques
    • G06F18/245—Classification techniques relating to the decision surface
    • G06F18/2451—Classification techniques relating to the decision surface linear, e.g. hyperplane
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/24—Classification techniques
    • G06F18/245—Classification techniques relating to the decision surface
    • G06F18/2453—Classification techniques relating to the decision surface non-linear, e.g. polynomial classifier
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F3/00—Input arrangements for transferring data to be processed into a form capable of being handled by the computer; Output arrangements for transferring data from processing unit to output unit, e.g. interface arrangements
    • G06F3/01—Input arrangements or combined input and output arrangements for interaction between user and computer
    • G06F3/048—Interaction techniques based on graphical user interfaces [GUI]
    • G06F3/0481—Interaction techniques based on graphical user interfaces [GUI] based on specific properties of the displayed interaction object or a metaphor-based environment, e.g. interaction with desktop elements like windows or icons, or assisted by a cursor's changing behaviour or appearance
    • G06F3/04817—Interaction techniques based on graphical user interfaces [GUI] based on specific properties of the displayed interaction object or a metaphor-based environment, e.g. interaction with desktop elements like windows or icons, or assisted by a cursor's changing behaviour or appearance using icons
    • 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
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
    • G06Q10/00—Administration; Management
    • G06Q10/04—Forecasting or optimisation specially adapted for administrative or management purposes, e.g. linear programming or "cutting stock problem"

Definitions

  • the present application relates to the field of computer technologies, and in particular, to a data processing method and apparatus.
  • the order data determines prediction data, wherein the high order data is a non-linear combination of at least three of the plurality of feature data.
  • the data processing apparatus may perform high-order nonlinear combination processing on the plurality of feature data of an application in the application mall to obtain prediction data for characterizing the click rate of the icon of the application, and according to the prediction data, the application may be determined. Recommended priority.
  • the prediction data in the related art is obtained by the data processing apparatus based on a plurality of high-order data, and the accurate prediction needs to be implemented based on a large amount of data, the data volume of the plurality of high-order data still cannot satisfy the accurate prediction requirement, and therefore, The accuracy of the prediction data determined by the related art is low.
  • the present application provides a data processing method and apparatus.
  • the technical solution is as follows:
  • a data processing method includes: acquiring a plurality of feature data; automatically performing two different non-linear combination processes on the plurality of feature data to obtain two sets of processed data, wherein the The two sets of processing data include a set of high-order data and a set of low-order data, the high-order data being associated with a nonlinear combination of m feature data of the plurality of feature data, the low-order data and the plurality of The nonlinear combination of n feature data in the feature data is related, m ⁇ 3, m>n ⁇ 2; the prediction data is determined according to the plurality of target data, the plurality of target data including the two sets of processed data.
  • a set of processing data (including at least one high-order data) that can be obtained by performing a nonlinear combination processing (high-order nonlinear combination processing) on a plurality of feature data; and the present application can automatically perform multiple feature data automatically.
  • Two different nonlinear combination processes (high-order nonlinear combination processing and low-order nonlinear combination processing) and two sets of processing data are obtained. That is, the data amount of the two sets of processing data obtained in the present application is larger than the data amount of the set of processing data obtained in the related art, and the accurate prediction needs to be implemented based on a large amount of data, and therefore, the present application is based on two groups.
  • the accuracy of the predicted data obtained by processing the data is higher than the accuracy of the predicted data obtained based on a set of processed data in the related art.
  • the automatically performing two different non-linear combination processing on the plurality of feature data to obtain two sets of processing data including: automatically performing high-order nonlinear combination processing on the plurality of feature data, to obtain the a set of high-order data; automatically performing low-order nonlinear combination processing on the plurality of feature data to obtain the set of low-order data.
  • the number of high-order data obtained by performing high-order nonlinear combination processing on a part of the feature data is the first number
  • the other part of the feature data is subjected to low-order nonlinear combination processing, and the number of low-order data is the second number
  • the number of target data on which the prediction data is based is equal to the first number and the first number. The sum of two numbers. If the high-order nonlinear combination processing is performed on the plurality of data, the obtained high-order data is larger than the first number, and the low-order nonlinear combination processing is performed on the plurality of data, and the obtained low-order data is obtained.
  • the number of the data is greater than the second number, and the number of target data on which the prediction data is finally determined is also greater than the sum of the first number and the second number. That is, when performing high-order nonlinear combination processing and low-order nonlinear combination processing on a plurality of feature data, the number of target data on which the prediction data is finally determined is more, and the accuracy of the prediction data is The higher it will be.
  • the method further includes: automatically performing linear combination processing on the plurality of feature data to obtain a set of first auxiliary data, the first auxiliary data and the A linear combination of at least two of the plurality of feature data, the plurality of target data further comprising: the at least one first assistance data.
  • the present application can perform not only high-order nonlinear combination processing and low-order nonlinear combination processing on a plurality of feature data, but also linear combination processing on a plurality of feature data to obtain a set of first auxiliary data, thereby Further, the number of target data on which the prediction data is finalized is further increased, and the accuracy of the prediction data is further improved.
  • the automatically performing low-order nonlinear combination processing on the plurality of feature data to obtain the set of low-order data including: automatically, by using a processing module based on a factorization machine FM, the plurality of feature data Performing low-order nonlinear combination processing to obtain the set of low-order data; and automatically performing linear combination processing on the plurality of feature data to obtain a set of first auxiliary data, including: automatically adopting an FM-based processing module Performing linear combination processing on the plurality of feature data to obtain the set of first auxiliary data; or automatically performing low-order nonlinear combination processing on the plurality of feature data to obtain the set of low-order data
  • the method includes: performing, by using a processing module of the domain-aware decomposition machine FFM, a low-order nonlinear combination process on the plurality of feature data to obtain the set of low-order data; and automatically performing the multiple feature data
  • the linear combination processing obtains a set of first auxiliary data, including: automatically performing linear combination processing on the plurality of feature data by using an FFM
  • the method further includes: combining the expert experience and the plurality of feature data, determining x empirical feature data, x ⁇ 1, the x experiences
  • the present application is capable of determining a set of second auxiliary data obtained by processing x empirical feature data having a large influence on the predicted data and processing the x empirical feature data according to the guidance of the expert experience. And finally determining the target data on which the prediction data is based further includes the second auxiliary data, thereby further increasing the number of target data on which the prediction data is finalized, and further improving the accuracy of the prediction data.
  • the multiple feature data is a plurality of sparse feature data
  • the method includes: converting the plurality of sparse feature data into multiple dense feature data; Automatically performing two different non-linear combination processes on the plurality of feature data to obtain two sets of processed data, including: automatically performing two different non-linear combination processes on the plurality of dense feature data to obtain the two sets of processing data.
  • the present application can also convert the plurality of sparse feature data to obtain a plurality of dense feature data, and the data amount of the dense feature data is much smaller than the data amount of the sparse feature data.
  • the efficiency of processing multiple dense feature data is higher than the efficiency of processing multiple sparse feature data.
  • a data processing apparatus comprising: an acquiring unit, configured to acquire a plurality of feature data; and a nonlinear processing unit, configured to automatically perform two different types of the plurality of feature data
  • the nonlinear combination processing obtains two sets of processing data, wherein the two sets of processing data includes a set of high-order data and a set of low-order data, and the high-order data and the non-m characteristic data of the plurality of feature data are non- a linear combination correlation, the low-order data is related to a nonlinear combination of n feature data of the plurality of feature data, m ⁇ 3, m>n ⁇ 2; and a prediction unit configured to determine prediction data according to the plurality of target data
  • the plurality of target data includes the two sets of processing data.
  • the non-linear processing unit includes: a high-order processing sub-unit, configured to automatically perform high-order nonlinear combination processing on the plurality of feature data to obtain the set of high-order data; and the low-order processing sub-unit And performing automatic low-order nonlinear combination processing on the plurality of feature data to obtain the set of low-order data.
  • a high-order processing sub-unit configured to automatically perform high-order nonlinear combination processing on the plurality of feature data to obtain the set of high-order data
  • the low-order processing sub-unit And performing automatic low-order nonlinear combination processing on the plurality of feature data to obtain the set of low-order data.
  • the data processing apparatus further includes: a linear processing unit, configured to automatically perform linear combination processing on the plurality of feature data to obtain a set of first auxiliary data, the first auxiliary data and the plurality of A linear combination of at least two feature data in the feature data, the plurality of target data further comprising: the at least one first assistance data.
  • a linear processing unit configured to automatically perform linear combination processing on the plurality of feature data to obtain a set of first auxiliary data, the first auxiliary data and the plurality of A linear combination of at least two feature data in the feature data
  • the plurality of target data further comprising: the at least one first assistance data.
  • the data processing apparatus further includes: a processing module based on a factorization machine FM or a processing module based on a domain aware decomposition machine FFM, the FM-based processing module includes: the low-order processing sub-unit and the a linear processing unit; the FFM-based processing module includes: the low-order processing sub-unit and the linear processing unit.
  • a processing module based on a factorization machine FM or a processing module based on a domain aware decomposition machine FFM the FM-based processing module includes: the low-order processing sub-unit and the a linear processing unit
  • the FFM-based processing module includes: the low-order processing sub-unit and the linear processing unit.
  • a determining unit combining the expert experience and the plurality of feature data, determining x empirical feature data, x ⁇ 1, wherein the x empirical feature data includes:
  • the plurality of feature data is a plurality of sparse feature data
  • the data processing device further includes: a converting unit, configured to convert the plurality of sparse feature data into multiple dense feature data;
  • the linear processing unit is further configured to: automatically perform two different non-linear combination processes on the plurality of dense feature data to obtain the two sets of processing data.
  • a computer readable storage medium having stored thereon a computer program that, when executed by a processor, implements the data processing method of the first aspect.
  • a computer program product which, when run on a computer, causes the computer to perform the data processing method of the first aspect.
  • a fifth aspect provides a data processing apparatus, the data processing apparatus comprising: at least one processor, at least one network interface, a memory, and at least one bus, wherein the memory and the network interface are respectively connected to the processor through a bus; the processor is The instructions are configured to execute the instructions stored in the memory; the processor implements the data processing method provided by any of the possible implementations of the first aspect or the first aspect by executing the instructions.
  • a set of processing data (including at least one high-order data) that can be obtained by performing a nonlinear combination processing (high-order nonlinear combination processing) on a plurality of feature data; and the present application can automatically perform multiple feature data automatically.
  • Two different nonlinear combination processes (high-order nonlinear combination processing and low-order nonlinear combination processing) and two sets of processing data are obtained. That is, the data amount of the two sets of processing data obtained in the present application is larger than the data amount of the set of processing data obtained in the related art, and the accurate prediction needs to be implemented based on a large amount of data, and therefore, the present application is based on two groups.
  • the accuracy of the predicted data obtained by processing the data is higher than the accuracy of the predicted data obtained based on a set of processed data in the related art.
  • FIG. 1 is a schematic diagram of an application scenario of a data processing method according to an embodiment of the present disclosure
  • FIG. 2 is a schematic structural diagram of a data processing apparatus according to an embodiment of the present invention.
  • FIG. 3 is a schematic structural diagram of a program module according to an embodiment of the present disclosure.
  • FIG. 4 is a flowchart of a method for processing a data according to an embodiment of the present invention.
  • FIG. 5 is a schematic structural diagram of another data processing apparatus according to an embodiment of the present disclosure.
  • FIG. 6 is a schematic structural diagram of still another data processing apparatus according to an embodiment of the present disclosure.
  • FIG. 7 is a schematic structural diagram of still another data processing apparatus according to an embodiment of the present disclosure.
  • FIG. 8 is a schematic structural diagram of a data processing apparatus according to another embodiment of the present disclosure.
  • the server 1 can establish a communication connection with multiple clients 2, and the server 1 can acquire data related to the client 2.
  • the client 2 can also acquire data stored on the server from the server 1.
  • the data processing device 0 is connected to the server 1, and the data processing device 0 can also be provided on the server 1, and the data processing device 0 can process the data acquired by the server.
  • both data processing device 0 and server 1, as well as server 1 and client 2 can be connected by a wired network or a wireless network.
  • a plurality of feature data may be combined according to a preset equation.
  • the preset equation is a linear equation
  • the combined processing of the plurality of feature data is also referred to as “linear combination processing”
  • the data obtained by the combination processing is related to the "linear combination” of the plurality of feature data.
  • the preset equation is a nonlinear equation
  • the “combination processing” of the plurality of feature data is also referred to as “non-linear combination processing”, and at this time, the “non-linear combination of the data obtained by the combination processing and the plurality of feature data” Related”.
  • the linear equation is a multivariate one-time equation.
  • the relationship between the independent variable and the dependent variable in the linear equation is linear, and the linear equation is represented as a straight line on the Cartesian coordinate system.
  • the nonlinear equation is a multivariate multiple equation.
  • the relationship between the independent variable and the dependent variable in the nonlinear equation is not linear, and the nonlinear equation is represented as a curve on the Cartesian coordinate system.
  • d2 is the weight of the feature data a2
  • y1 is related to the linear combination of the feature data a1 and the feature data a2;
  • the data obtained by the linear combination processing may not be d1*a1+d2*a2, and the data obtained by the non-linear combination processing may not be a1*a2, which is not limited in the embodiment of the present invention.
  • FIG. 2 is a schematic structural diagram of a data processing apparatus according to an embodiment of the present invention.
  • the data processing apparatus may be the data processing apparatus shown in FIG. 1.
  • the data processing apparatus can include: at least one processor 10221 (eg, a central processing unit), at least one network interface 10222, a memory 10223, and at least one bus 10224, which can be used to implement a processor, The connection communication between the network interface and the memory; the memory 10223 and the network interface 10222 are connected to the processor 10221 via the bus 10224, respectively.
  • the processor 10221 is configured to execute an executable module, such as a computer program, stored in the memory 10223.
  • the memory 10223 may include a high speed random access memory (RAM), and may also include a non-volatile memory such as at least one disk memory.
  • RAM random access memory
  • non-volatile memory such as at least one disk memory.
  • the communication connection between the data processing device and at least one other device is achieved by at least one network interface 10222 (wired or wireless).
  • the memory 10223 stores a program module 10225 that can be executed by the processor 10221 to implement the data processing method as shown in FIG.
  • the program module may include:
  • the obtaining module 01 is configured to acquire a plurality of feature data, where the plurality of feature data is a plurality of sparse feature data.
  • the conversion module 02 is configured to convert the plurality of sparse feature data into a plurality of dense feature data
  • the nonlinear processing module (not shown in FIG. 3) is configured to automatically perform two different non-linear combination processing on multiple feature data to obtain two sets of processing data, wherein the two sets of processing data include a set of high-order data and one Group lower-order data, high-order data is related to nonlinear combination of m feature data in multiple feature data, and low-order data is related to nonlinear combination of n feature data in multiple feature data, m ⁇ 3, m>n ⁇ 2; the nonlinear processing module is further configured to: automatically perform two different nonlinear combination processing on multiple dense feature data to obtain two sets of processed data.
  • the linear processing module 04 is configured to automatically perform linear combination processing on the plurality of feature data to obtain a set of first auxiliary data, and the first auxiliary data is related to a linear combination of at least two of the plurality of feature data, and the plurality of targets
  • the data also includes: at least one first auxiliary data.
  • the prediction module 05 is configured to determine prediction data according to the plurality of target data, where the plurality of target data includes two sets of processing data;
  • the nonlinear processing module may include: a high-order processing sub-module 031 and a low-order processing sub-module 032, and the high-order processing sub-module 031 may be configured to automatically perform high-order nonlinear combination processing on multiple feature data to obtain A set of high-order data; the low-order processing sub-module 032 can be used to automatically perform low-order nonlinear combination processing on a plurality of feature data to obtain a set of low-order data.
  • the program module may further include: a Factorization Machine (FM)-based processing module or a Field-aware Factorization Machine (FFM)-based processing module, where the FM-based processing module includes: The low-order processing sub-module 032 and the linear processing module 04; the FFM-based processing module includes a low-order processing sub-module 032 and a linear processing module 04.
  • FM Factorization Machine
  • FFM Field-aware Factorization Machine
  • the program module may further include: a determining module and an experience processing module (neither shown in FIG. 3), wherein the determining module may be used to determine x empirical feature data by combining expert experience and multiple feature data, x ⁇
  • the data includes: data acquired in a plurality of feature data in combination with expert experience, and/or combined with expert experience to perform nonlinear combination processing on at least two of the plurality of feature data to obtain data;
  • FIG. 4 is a flowchart of a method for processing a data according to an embodiment of the present invention.
  • the data processing method can be used in the data processing apparatus shown in FIG. 2. As shown in FIG. 4, the data processing method can include:
  • Step 401 Acquire a plurality of sparse feature data.
  • an icon of multiple applications is displayed in the client of the application mall.
  • the icon of the application to be downloaded may be clicked in the multiple applications, thereby entering the download interface of the application, and determining whether Download the app. Therefore, whether the user downloads the application is related to whether the user clicks on the icon of the application, and the server of the application mall can determine the recommendation priority of the application based on the click through rate (CTR) of the icon of the application, and the click rate is high.
  • CTR click through rate
  • the application assigns a higher priority, and correspondingly displays the icons of the applications in the application mall to recommend the application that meets the user's requirements to the user in time and accurately, thereby improving user stickiness. In this scenario, the application is applied.
  • the plurality of sparse feature data acquired by the data processing device in step 401 may be sparse feature data of the application, wherein the sparse feature data of the application includes: attribute data of the user and The attribute data of the application.
  • the attribute data of the user may include attribute data such as the identifier of the user, the gender of the user, the age of the user, the occupation of the user, the income of the user, the hobby of the user, and the education status of the user.
  • the attribute data of the application may include: an identifier of the application, a name of the application, a developer of the application, a type of the application, an installation package size of the application, a label of the application, a score of the application, a comment of the application, and the like.
  • the identifier of the user is Xiao Ming
  • the gender of the user is male
  • the age of the user is 20 years old
  • the name of the application is F1
  • the type of the application is a chat application
  • Feature data can be arranged according to certain rules, such as "Xiaoming, Male, 20, F1, Chat”.
  • the website can control the terminal to display the entries of the plurality of files related to the keyword according to the keyword, and the user can The requirement is to click on the entry of a file to download the file.
  • a file that is downloaded multiple times usually meets the downloading requirements of multiple users. Therefore, the website can determine the recommendation priority of the file according to the download amount of the file, and assign a higher priority to the file with a higher download amount.
  • the corresponding control terminal preferentially displays these file entries.
  • the plurality of sparse feature data of the file may include: a name of the file, a size of the file, and a label of the file.
  • the file name is F2
  • the file size is 2 megabytes
  • the file label is technology
  • the sparse feature data obtained in step 401 can be arranged according to certain rules, for example, “F2. 2 trillion, technology.”
  • the program module in the data processing apparatus may include an acquisition module.
  • the data processing apparatus may acquire a plurality of sparse feature data through the acquisition module.
  • Step 402 Convert a plurality of sparse feature data to obtain a plurality of dense feature data.
  • the data processing device may convert the plurality of sparse feature data to obtain a plurality of dense feature data with less data amount, and the data amount of the dense feature data is much smaller than The data amount of the sparse feature data, the efficiency of processing the plurality of dense feature data by the subsequent data processing device is higher than the efficiency of processing the plurality of sparse feature data by the data processing device.
  • the program module in the data processing device may further include a conversion module.
  • the data processing device may convert the plurality of sparse feature data acquired by the acquisition module by using the conversion module to obtain multiple dense data. Feature data.
  • Step 403 Perform high-order nonlinear combination processing on the plurality of dense feature data automatically to obtain a set of high-order data.
  • the data processing device can perform high-order nonlinear combination processing on the input plurality of dense feature data to obtain a set of high-order data, and each high-order data and the plurality of dense feature data are The nonlinear combination of m characteristic data is related.
  • m is any integer greater than or equal to 3, that is, a set of higher order data may include: a plurality of higher order data related to a nonlinear combination of at least three feature data.
  • m is 5, 6, or 7.
  • the set of high-order data includes: high-order data related to nonlinear combination of five feature data, high-order data related to nonlinear combination of six feature data. And high-order data related to the nonlinear combination of the seven feature data.
  • the program module in the data processing device may further include a nonlinear processing module, where the nonlinear processing module includes: a high-order processing sub-module, and the data processing device may automatically and automatically use the high-order processing sub-module in step 403.
  • a dense feature data is processed by high-order nonlinear combination to obtain a set of high-order data.
  • the high-order processing sub-module may be: a Deep Neural Network (DNN)-based processing sub-module, a Factorization-machine Supported Neural Network (FNN)-based processing sub-module, or based on Product-based Neural Network (PNN) processing sub-module.
  • DNN Deep Neural Network
  • FNN Factorization-machine Supported Neural Network
  • PNN Product-based Neural Network
  • the high-order processing sub-module may be a DNN-based processing sub-module, and the DNN-based processing sub-module may include a plurality of concatenated first processing elements 0311.
  • the DNN-based processing sub-module includes a first processing element of the H+1 layer (the two-layer first processing element 0311 is schematically shown in FIG.
  • the first processing element of the first layer and the conversion module 02, and The first processing elements of the non-first layer are connected; the first processing element of the i-th layer is connected to the first processing element of the i-1th layer and the first processing element of the i+1th layer; the H+1th layer
  • the first processing element is connected to the first processing element of the Hth level and the prediction module 05 finally used to determine the prediction data in the data processing apparatus, i is an integer greater than 1 and less than H+1, and H is greater than or equal to An integer of 1.
  • the conversion module 02 can input the plurality of dense feature data to the first processing element of the first layer, and the first processing element 021 of each layer can perform nonlinear combination processing on the input data, and the processed data is obtained.
  • each first processing element of the H+1th level can output a high-order data to the prediction module 05 after performing nonlinear combination processing on the input data, each high order
  • the data is all related to a non-linear combination of m feature data among the plurality of dense feature data acquired by the data processing device.
  • any two connected first processing elements have a connected edge for connecting the two first processing elements, and the connected edges have preset weights.
  • the first processing element of the (H+1)th layer and the connected side of the prediction module and the first processing unit of the first layer and the connecting side of the conversion module also have preset weights.
  • the data processed by each first processing unit is: data obtained by inputting the data of the first processing unit by the side and the weight calculated on the side.
  • the high-order processing sub-module may be a DNN-based processing sub-module, and the DNN-based processing sub-module may include a plurality of cascaded second processing elements. It is assumed that the DNN-based processing sub-module comprises a second processing element of the H+1 layer, and the high-order processing sub-module automatically performs high-order nonlinear combination processing on the plurality of dense feature data to obtain a set of high-order data.
  • ⁇ (x) is the activation function
  • a H is the vector whose data is processed by the second processing element of the Hth level
  • b H+1 is the offset of the second processing element of the H+1th level
  • W H+1 is a matrix whose weights on the side of the second processing element of the Hth level and the second processing element of the H+1th level are elements.
  • the activation function may be an activation function such as a sigmoid function (also called a sigmoid function), a rectified linear unit (relu) function, or a tangential hyperbolic (tanh) function.
  • the second processing unit of the Hth layer includes m H second processing elements, and the data obtained after processing each second processing element of the m H second processing elements are real numbers, and the second processing element of the H+1th layer A total of m H+1 second processing elements are included, and data obtained by processing each second processing element in the m H+1 second processing elements are real numbers, and the second processing element and the H+1th layer of the Hth layer
  • the weights on the sides of the second processing element of the layer include m H ⁇ m H+1 real numbers.
  • l is an integer greater than or equal to 1, and less than or equal to H
  • a l is a vector with the data processed by the second processing element of each layer 1 as an element
  • b l+1 is the first +1 layer second
  • W l+1 is a vector whose weight on each side of the second processing element of the first layer and the second processing element of the l+1th layer is an element
  • Step 404 Perform low-order nonlinear combination processing on the plurality of dense feature data to obtain a set of low-order data.
  • the data processing device is further capable of performing low-order nonlinear combination processing on the input plurality of dense feature data to obtain a set of low-order data, each low-order data and the plurality of dense feature data.
  • the nonlinear combination of n feature data is related. Where n is any integer greater than or equal to 2 and less than m. For example, m is 5, 6, or 7. In this case, n can be 2, 3, or 4.
  • the set of low-order data includes: low-order data related to nonlinear combination of two feature data, and three feature data. The low-order data associated with the nonlinear combination and the low-order data associated with the nonlinear combination of the four eigen data.
  • Step 405 Perform linear combination processing on the plurality of dense feature data to obtain a set of first auxiliary data.
  • the data processing device is further capable of linearly combining the input plurality of dense feature data to obtain a set of first auxiliary data, each of the first auxiliary data and the plurality of dense feature data.
  • a linear combination of at least two characteristic data is associated.
  • a set of first auxiliary data may include: first auxiliary data related to linear combination of two feature data, first auxiliary data related to linear combination of three feature data, and linear combination with four feature data Related first auxiliary data.
  • the data processing apparatus can automatically perform low-order nonlinear combination processing on the plurality of dense feature data through the low-order processing sub-module in the data processing device to obtain a set of low-order data.
  • the data processing apparatus may automatically perform linear combination processing on the plurality of dense feature data through a linear processing module in the data processing device to obtain a set of first auxiliary data.
  • the program module in the data processing apparatus shown in FIG. 4 may further include an FM-based processing module or an FFM-based processing module, and the FM-based processing module includes: a low-order processing sub-module 032 and a linear processing module 04.
  • the FFM-based processing module includes a low-order processing sub-module 032 and a linear processing module 04. That is, when performing step 404 and step 405, the data processing apparatus may perform low-order nonlinear combination processing and linear combination on the plurality of dense feature data directly through the FM-based processing module or the FFM-based processing module, respectively. A set of low order data and a set of first auxiliary data are obtained.
  • the data processing apparatus can perform low-order nonlinear combination processing and linear combination on the plurality of dense feature data directly through the FM-based processing module to obtain a set of low-order data and a set of first
  • the auxiliary data which is processed by the FM-based processing module, can be expressed as:
  • the A set of low order data and a set of first auxiliary data are included.
  • the set of low-order data is
  • the first set of auxiliary data is
  • each element in V a is a real number
  • V b is an implicit vector of the b-th feature data of the plurality of dense feature data, a+1 ⁇ b ⁇ n
  • V b has a specified dimension
  • Each element in b is a real number
  • X i, a is the a-th feature data in the plurality of dense feature data
  • X i, b is the b-th feature data in the plurality of dense feature data
  • X i , j is the jth feature data of the plurality of dense feature data, 1 ⁇ j ⁇ n
  • w j is the weight of the jth feature data.
  • the data processing apparatus can perform low-order nonlinear combination processing and linear combination on the plurality of dense feature data directly through the FFM-based processing module to obtain a set of low-order data and a set of first Auxiliary data, which is processed by the FFM-based processing module, can be expressed as: A set of low order data and a set of first auxiliary data are included.
  • the set of low-order data is The first set of auxiliary data is
  • the function f a represents a domain in which the a-th feature data of the plurality of dense feature data is located
  • the function f b represents a domain in which the b-th feature data of the plurality of dense feature data is input
  • An implicit vector of the a-th feature data pair f b of the plurality of dense feature data and To have a specified dimension, Each element in the real number is a real number.
  • Step 406 Determine prediction data according to the plurality of target data, the plurality of target data including a set of low order data, a set of high order data, and a set of first auxiliary data.
  • the data processing apparatus can be based on the set of low-order data, a set of high-order data, and a set of first auxiliary data.
  • Multiple target data to determine forecast data may be provided.
  • the program module in the data processing apparatus may further include a prediction module 05.
  • the data processing apparatus may determine prediction data through the prediction module 05, optionally, the prediction module determines
  • the forecast data can be:
  • the prediction data obtained by the data processing device in step 406 can be used to represent the The clickthrough rate of the app's icon.
  • the server 1 the server of the application mall
  • the server 1 can analyze each user in the application and the application mall through the data processing device 0 to determine the user to each application in the application mall.
  • the click rate of the icon, and sorting the multiple applications according to the click rate of the icon of each application, and the information of the corresponding application mall interface is sent to the client 2 (the client of the application mall), and the client 2 displays based on the information.
  • the application mall interface, the icons of the applications displayed in the interface are arranged in the order determined above.
  • the server 1 reorders the icons of all the applications for each user, so that the order of all the applications in the client of the application mall used by each user is related to the attribute characteristics of the user, so that the application The sorting is more in line with the needs of each user. For example, if the application mall includes four applications, namely, application Q1, application Q2, application Q3, and application Q4, the data processing device determines that the user has the highest click rate for the icon of the application Q1 after processing the sparse feature data of the application. It is an icon for applying Q2, an icon for applying Q4, and an icon for applying Q3.
  • the server may arrange the icons of the four applications according to the icons of the application Q1, the icon of the application Q2, the icon of the application Q3, and the icon of the application Q4, so that the four applications in the application mall interface displayed by the client are displayed.
  • the icons are arranged in the order of the icon of the application Q1, the icon of the application Q2, the icon of the application Q3, and the icon of the application Q4.
  • the server may acquire a plurality of pieces of training data, and may train the data processing device according to the plurality of pieces of data through the training device.
  • each piece of training data may include: a set of sparse feature data and click data corresponding to the set of sparse feature data.
  • a set of sparse feature data in a piece of training data may include: attribute data of the user P and attribute data of the application Q
  • the click data in the piece of training data may include: an identifier of the user P
  • the click data in the piece of training data may further include: a time for collecting the training data, a weather condition when the training data is counted, a traffic condition when the training data is counted, and a statistical data when the training data is collected.
  • the identifier of the user P may be the identification number (ID) of the user P
  • the identifier of the application Q may also be the ID of the application Q.
  • the training device may first input multiple sets of sparse feature data in the plurality of pieces of training data to the acquiring module in the data processing device, and the data processing device can sequentially perform the input of each set of sparse feature data. Processing, a prediction data is obtained, which is used to indicate whether the user clicks on the icon of the application. The training device is further configured to compare the prediction data obtained by the data processing device according to the set of sparse feature data with the click data corresponding to the set of sparse feature data, and determine whether the predicted data and the click data are both indicated: the user clicks on the icon of the application. , or the user won’t click the app’s icon.
  • the training device needs to adjust the weight used in the process of processing the data by the data processing device. After the data processing device is trained by the training device, the weight used by the data processing device in processing the data has been adjusted a plurality of times, and the data processing device can process the plurality of sparse feature data to obtain a more accurate prediction. data.
  • the plurality of feature data After acquiring a plurality of feature data, if the number of high-order data obtained by performing high-order nonlinear combination processing on a part of the feature data is the first number, the plurality of feature data The other part of the feature data is subjected to low-order nonlinear combination processing, and the number of low-order data is the second number, and finally the number of target data on which the prediction data is based is equal to the first number and the second number. Sum. If the high-order nonlinear combination processing is performed on the plurality of data, the obtained high-order data is larger than the first number, and the low-order nonlinear combination processing is performed on the plurality of data, and the obtained low-order data is obtained.
  • the number of the data is greater than the second number, and the number of target data on which the prediction data is finally determined is also greater than the sum of the first number and the second number. That is, when performing high-order nonlinear combination processing and low-order nonlinear combination processing on a plurality of feature data, the number of target data on which the prediction data is finally determined is more, and the accuracy of the prediction data is The higher it will be. Therefore, in the embodiment of the present invention, all the dense feature data obtained in step 402 are processed in steps 403, 404, and 405, so that the target data according to the prediction data is determined in step 406, and the steps are performed. The accuracy of the prediction data obtained in 406 is higher.
  • the program module in the data processing device shown in FIG. 3 may further include a determining module and an experience processing module.
  • the data processing device may further combine the expert experience and the plurality of dense feature data by using the determining module. Determining x empirical feature data, and automatically processing the x empirical feature data by the empirical processing module to obtain a set of second auxiliary data.
  • the plurality of target data by which the prediction module in the program module in the data processing device determines the prediction data in step 406 may further include: the set of second auxiliary data. That is, the plurality of target data may include: a set of high-order data, a set of low-order data, a set of first auxiliary data, and a set of second auxiliary data.
  • x is an integer greater than or equal to 1
  • x empirical feature data includes: data acquired in multiple dense feature data in combination with expert experience, and/or combined with expert experience on at least one of multiple dense feature data
  • the two feature data are subjected to nonlinear combination processing to obtain data.
  • the second auxiliary data is related to an empirical feature data
  • the second auxiliary data is related to a linear combination of at least two empirical feature data of the x empirical feature data
  • the second auxiliary data is associated with a non-linear combination of at least two empirical feature data of the x empirical feature data.
  • the “expert” mentioned in the embodiments of the present invention refers to a technician or a plurality of technicians in the field of computer technology, and the “expert” knows all common technical knowledge in the field of computer technology, and can learn the field of computer technology.
  • the "expert experience” mentioned in the embodiment of the present invention is the "preset formula” determined by the "expert".
  • the determining module combines the "expert experience” and the plurality of dense feature data to determine x empirical feature data, including: the determining module inputs the plurality of dense feature data as the independent variable of the "preset formula” into the "preset formula", and the preset formula The results obtained are x empirical characterization data.
  • the plurality of dense feature data are: data a1, data a2, data a3, data a4, data a5, data a6, data a7, data a8, data a9, and data a10, a total of 10 data.
  • the determining module can combine the expert experience to obtain the data a1 in the plurality of dense feature data as an empirical feature data; or the determining module can perform nonlinear combination processing on the data a9 and the data a10 in the plurality of dense feature data in combination with the expert experience.
  • the determining module acquires the data a1 and uses the data a1 as an empirical feature data, and performs nonlinear combination processing on the data a9 and the data a10 in combination with expert experience to obtain another empirical data.
  • the empirical processing unit can also automatically process the x empirical feature data to obtain a set of second auxiliary data.
  • the second auxiliary data is related to an empirical feature data
  • the linear combination of the second auxiliary data and at least two empirical feature data of the x empirical feature data Associated with a combination of nonlinear and/or nonlinear combinations.
  • the second auxiliary data is related to a linear combination of at least two empirical feature data of the x empirical feature data; or at least two empirical features of the second auxiliary data and the x empirical feature data.
  • the experience processing module may be a Logistic Regression (LR) based processing module, an FM based processing module, an FFM based processing module, or a DNN based processing module.
  • LR Logistic Regression
  • the experience processing module needs to combine expert experience when processing multiple dense feature data, but the nonlinear processing module and the linear processing module do not need to combine expert experience when processing multiple dense feature data. , the feature data can be processed automatically.
  • the embodiment of the present invention uses a plurality of pieces of training data and a plurality of sets of detection data, respectively, and the data processing apparatus provided by the embodiment of the present invention and related
  • the data processing device in the technology has performed several experiments separately. And using the area under the ROC curve (AUC) and the loss function (Logloss) to perform data processing in the data processing apparatus and related technologies in the embodiments of the present invention. The device was compared. It is known that the accuracy of the prediction data obtained by the data processing apparatus provided by the embodiment of the present invention is higher than the prediction data obtained by the data processing apparatus in the related art.
  • the data processing device may acquire a plurality of sparse feature data of the application Q, and convert the plurality of sparse feature data to obtain Apply multiple dense feature data of Q. Then, the data processing device automatically performs multiple dense feature data of the application Q: high-order nonlinear combination processing, low-order nonlinear combination processing, and linear combination processing to obtain a set of high-order data and a set of low-order data. And a set of first auxiliary data. Further, the data processing device may further determine x empirical feature data by combining expert experience and multiple dense feature data of the application Q; and automatically processing the x empirical feature data to obtain a set of second auxiliary data.
  • the data processing apparatus determines the prediction data based on the plurality of target data including the set of high-order data, the set of low-order data, the set of first auxiliary data, and the set of second auxiliary data.
  • the predicted data can be used to characterize the click rate of the icon of the application Q.
  • the embodiment of the present invention provides a data processing method that can automatically perform two different nonlinear combination processing (high-order nonlinear combination processing and low-order nonlinear combination processing) on feature data and obtain two The group processes the data.
  • a set of processing data including at least one high-order data
  • the data volume of the two sets of processing data obtained by the embodiment of the present invention is greater than the data volume of the set of processing data obtained in the related art, and the accurate prediction needs to be implemented based on a large amount of data. Therefore, in the embodiment of the present invention, The accuracy of the prediction data obtained based on the two sets of processing data is higher than the accuracy of the prediction data obtained based on a set of processing data in the related art.
  • the target data on which the prediction data is determined in the embodiment of the present invention may further include: a set of first auxiliary data and a set of second auxiliary data, that is, the target data in the embodiment of the present invention may include: The group of high-order data, a set of low-order data, a set of first auxiliary data, and a set of second auxiliary data have four sets of data, thereby further increasing the number of target data on which the finalized prediction data is based, and further improving Predict the accuracy of the data.
  • FIG. 5 is a schematic structural diagram of another data processing apparatus according to an embodiment of the present invention.
  • the data processing apparatus may be the data processing apparatus shown in FIG. 1.
  • the data processing apparatus 50 may include:
  • the obtaining unit 501 is configured to acquire multiple feature data.
  • the non-linear processing unit 502 is configured to automatically perform two different non-linear combination processing on the plurality of feature data to obtain two sets of processing data, wherein the two sets of processing data include a set of high-order data and a set of low-order data, high order
  • the data is related to the nonlinear combination of m feature data among the plurality of feature data
  • the low-order data is related to the nonlinear combination of the n feature data of the plurality of feature data, m ⁇ 3, m>n ⁇ 2;
  • the prediction unit 503 is configured to determine prediction data according to the plurality of target data, where the plurality of target data includes two sets of processing data.
  • the embodiment of the present invention provides a data processing apparatus, and the nonlinear processing unit can automatically perform two different nonlinear combination processing on the feature data (high-order nonlinear combination processing and low-order nonlinear combination processing). And get two sets of processing data.
  • a set of processing data (including at least one high-order data) that can be obtained by performing a nonlinear combination processing (high-order nonlinear combination processing) on the feature data. That is, the data volume of the two sets of processing data obtained by the embodiment of the present invention is greater than the data volume of the set of processing data obtained in the related art, and the accurate prediction needs to be implemented based on a large amount of data. Therefore, in the embodiment of the present invention, The accuracy of the prediction data obtained based on the two sets of processing data is higher than the accuracy of the prediction data obtained based on a set of processing data in the related art.
  • the nonlinear processing unit 502 can include:
  • the high-order processing sub-unit 5021 is configured to automatically perform high-order nonlinear combination processing on the plurality of feature data to obtain a set of high-order data;
  • the low-order processing sub-unit 5022 is configured to automatically perform low-order nonlinear combination processing on the plurality of feature data to obtain a set of low-order data.
  • FIG. 6 is a schematic structural diagram of another data processing apparatus according to an embodiment of the present disclosure.
  • the data processing apparatus 50 may further include:
  • the linear processing unit 504 is configured to automatically perform linear combination processing on the plurality of feature data to obtain a set of first auxiliary data, where the first auxiliary data is related to a linear combination of at least two of the plurality of feature data, and the plurality of targets
  • the data also includes: at least one first auxiliary data.
  • the data processing apparatus further includes: a processing module based on a factorization machine FM or a processing module based on a domain aware decomposition machine FFM, wherein the FM-based processing module comprises: a low-order processing sub-unit and a linear processing unit, based on the FFM
  • the processing module includes: a low-order processing sub-unit and a linear processing unit.
  • FIG. 7 is a schematic structural diagram of another data processing apparatus according to an embodiment of the present invention.
  • the data processing apparatus 50 may further include:
  • the determining unit 505 combines the expert experience and the plurality of feature data to determine x pieces of empirical feature data, x ⁇ 1, and the x pieces of empirical feature data include: data acquired in the plurality of feature data in combination with expert experience, and/or combined
  • the expert experience performs nonlinear combination processing on at least two of the plurality of feature data to obtain data;
  • FIG. 8 is a schematic structural diagram of a data processing device according to another embodiment of the present application. As shown in FIG. 8, on the basis of FIG. 5, the data processing device 50 further Can include:
  • the converting unit 507 is configured to convert the plurality of sparse feature data into a plurality of dense feature data
  • the non-linear processing unit 502 is further configured to: automatically perform two different non-linear combination processing on the plurality of dense feature data to obtain two sets of processed data.
  • the embodiment of the present invention provides a data processing apparatus, and the nonlinear processing unit can automatically perform two different nonlinear combination processing on the feature data (high-order nonlinear combination processing and low-order nonlinear combination processing). And get two sets of processing data.
  • a set of processing data (including at least one high-order data) that can be obtained by performing a nonlinear combination processing (high-order nonlinear combination processing) on the feature data. That is, the data volume of the two sets of processing data obtained by the embodiment of the present invention is greater than the data volume of the set of processing data obtained in the related art, and the accurate prediction needs to be implemented based on a large amount of data. Therefore, in the embodiment of the present invention, The accuracy of the prediction data obtained based on the two sets of processing data is higher than the accuracy of the prediction data obtained based on a set of processing data in the related art.
  • the target data by which the prediction data is determined in the embodiment of the present invention may further include: a set of first auxiliary data obtained by the linear processing unit and a set of second auxiliary data obtained by the empirical processing unit, that is, the present invention
  • the target data in the embodiment may include: a set of high-order data, a set of low-order data, a set of first auxiliary data, and a set of second auxiliary data, and a total of four sets of data, thereby further increasing the basis for finalizing the predicted data.
  • the number of target data further improves the accuracy of the predicted data.
  • the above embodiments it can be implemented in whole or in part by software, hardware, firmware, or any combination thereof.
  • software it can be implemented in whole or in part in the form of a computer program product comprising one or more computer instructions.
  • the computer program instructions When the computer program instructions are loaded and executed on a computer, the processes or functions described in accordance with embodiments of the present invention are generated in whole or in part.
  • the computer can be a general purpose computer, a computer network, or other programmable device.
  • the computer instructions can be stored in a computer readable storage medium or transferred from one computer readable storage medium to another computer readable storage medium, for example, the computer instructions can be from a website site, computer, server or data
  • the center transmits to another website site, computer, server, or data center by wire (eg, coaxial cable, fiber optic, digital subscriber line) or wireless (eg, infrared, wireless, microwave, etc.).
  • the computer readable storage medium can be any available media that can be accessed by a computer or a data storage device such as a server, data center, or the like that includes one or more available media.
  • the usable medium can be a magnetic medium (eg, a floppy disk, a hard disk, a magnetic tape), an optical medium, or a semiconductor medium (eg, a solid state hard disk) or the like.
  • the embodiment of the method provided by the embodiment of the present invention can refer to the corresponding device embodiment, which is not limited by the embodiment of the present invention.
  • the sequence of the steps of the method embodiments of the present invention can be appropriately adjusted, and the steps can be correspondingly increased or decreased according to the situation. Any person skilled in the art can easily think of the technical scope disclosed in the embodiments of the present invention. The method of the change should be included in the scope of protection of the embodiments of the present invention, and therefore will not be described again.
  • the serial numbers of the embodiments of the present invention are merely for the description, and do not represent the advantages and disadvantages of the embodiments.

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Data Mining & Analysis (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • Evolutionary Computation (AREA)
  • Artificial Intelligence (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Evolutionary Biology (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Mathematical Physics (AREA)
  • Business, Economics & Management (AREA)
  • Probability & Statistics with Applications (AREA)
  • Biophysics (AREA)
  • Economics (AREA)
  • Health & Medical Sciences (AREA)
  • Computational Linguistics (AREA)
  • General Health & Medical Sciences (AREA)
  • Molecular Biology (AREA)
  • Computing Systems (AREA)
  • Software Systems (AREA)
  • Strategic Management (AREA)
  • Human Resources & Organizations (AREA)
  • Biomedical Technology (AREA)
  • Nonlinear Science (AREA)
  • Development Economics (AREA)
  • Game Theory and Decision Science (AREA)
  • Human Computer Interaction (AREA)
  • Entrepreneurship & Innovation (AREA)
  • Marketing (AREA)
  • Operations Research (AREA)
  • Quality & Reliability (AREA)
  • Tourism & Hospitality (AREA)
  • General Business, Economics & Management (AREA)
  • Management, Administration, Business Operations System, And Electronic Commerce (AREA)
  • Information Retrieval, Db Structures And Fs Structures Therefor (AREA)

Abstract

本申请公开了一种数据处理方法及装置，属于计算机技术领域。所述方法包括：获取多个特征数据；自动对多个特征数据进行两种不同的非线性组合处理得到两组处理数据，其中，两组处理数据包括一组高阶数据和一组低阶数据，高阶数据与多个特征数据中m个特征数据的非线性组合相关，低阶数据与多个特征数据中n个特征数据的非线性组合相关，m≥3，m＞n≥2；根据多个目标数据确定预测数据，多个目标数据包括两组处理数据。本申请解决了相关技术确定的预测数据的准确度较低的问题，提高了预测数据的准确度，本申请用于数据处理。

Description

数据处理方法及装置
本申请要求于2017年6月30日提交中国国家知识产权局、申请号为201710524501.3、发明名称为“数据处理方法及装置”的中国专利申请的优先权，其全部内容通过引用结合在本申请中。
技术领域
本申请涉及计算机技术领域，特别涉及一种数据处理方法及装置。
背景技术
随着计算机技术的发展，人们在对可动态变化的事物(例如某一网站的访问次数、某一文件的下载量、应用商城中某一应用的图标的点击率或视频网站中某一视频的点击率)进行分析时，通常可以获取事物的多个特征数据(特征数据用于表征事物的属性)，并对该多个特征数据进行处理，得到用于表征该事物的发展方向的预测数据。
相关技术中，人们可以向数据处理装置输入事物的多个特征数据，数据处理装置能够对事物的多个特征数据进行高阶非线性组合处理，得到多个高阶数据，并根据该多个高阶数据确定预测数据，其中，高阶数据为该多个特征数据中至少三个特征数据的非线性组合。例如，数据处理装置可以对应用商城中某一应用的多个特征数据进行高阶非线性组合处理，得到用于表征该应用的图标的点击率的预测数据，根据该预测数据可以确定该应用的推荐优先级。
由于相关技术中的预测数据是数据处理装置基于多个高阶数据得到的，而准确的预测需要基于大量的数据来实现，上述多个高阶数据的数据量依然无法满足准确预测需求，因此，相关技术确定的预测数据的准确度较低。
发明内容
为了解决相关技术确定的预测数据的准确度较低的问题，本申请提供了一种数据处理方法及装置。所述技术方案如下：
第一方面，提供了一种数据处理方法，所述方法包括：获取多个特征数据；自动对所述多个特征数据进行两种不同的非线性组合处理得到两组处理数据，其中，所述两组处理数据包括一组高阶数据和一组低阶数据，所述高阶数据与所述多个特征数据中m个特征数据的非线性组合相关，所述低阶数据与所述多个特征数据中n个特征数据的非线性组合相关，m≥3，m＞n≥2；根据多个目标数据确定预测数据，所述多个目标数据包括所述两组处理数据。
相关技术中对多个特征数据进行一种非线性组合处理(高阶非线性组合处理)能够得到的一组处理数据(包括至少一个高阶数据)；而本申请能够自动对多个特征数据进行两种不同的非线性组合处理(高阶非线性组合处理和低阶非线性组合处理)并得到两组处理数据。也即是，本申请中得到的两组处理数据的数据量大于相关技术中得到的一组处理数据的数据量，且准确的预测需要基于大量的数据来实现，因此，本申请中基于两组处理数据得到的预测数据的准确度高于相关技术中基于一组处理数据得到的预测数据的准确度。
可选的，所述自动对所述多个特征数据进行两种不同的非线性组合处理得到两组处理数据，包括：自动对所述多个特征数据进行高阶非线性组合处理，得到所述一组高阶数据；自动对所述多个特征数据 进行低阶非线性组合处理，得到所述一组低阶数据。
需要说明的是，在获取到多个特征数据后，如果对该多个特征数据中的一部分特征数据进行高阶非线性组合处理得到的高阶数据的个数为第一个数，对该多个特征数据中的另一部分特征数据进行低阶非线性组合处理得到的低阶数据的个数为第二个数，则最终确定预测数据所基于的目标数据的个数等于第一个数与第二个数之和。如果对该多个数据进行高阶非线性组合处理，则得到的高阶数据的个数大于第一个数，且对该多个数据也进行低阶非线性组合处理，则得到的低阶数据的个数大于第二个数，且最终确定预测数据所基于的目标数据的个数也大于第一个数与第二个数之和。也即是，在对多个特征数据分别进行高阶非线性组合处理以及低阶非线性组合处理时，最终确定预测数据所基于的目标数据的个数就会较多，预测数据的准确度就会越高。
可选的，在所述获取多个特征数据之后，所述方法还包括：自动对所述多个特征数据进行线性组合处理，得到一组第一辅助数据，所述第一辅助数据与所述多个特征数据中的至少两个特征数据的线性组合相关，所述多个目标数据还包括：所述至少一个第一辅助数据。
也即是，本申请不仅能够对多个特征数据进行高阶非线性组合处理以及低阶非线性组合处理，并且还能够对多个特征数据进行线性组合处理，得到一组第一辅助数据，从而进一步的增多了最终确定预测数据所基于的目标数据的个数，进一步提高了预测数据的准确度。
可选的，所述自动对所述多个特征数据进行低阶非线性组合处理，得到所述一组低阶数据，包括：通过基于因子分解机FM的处理模块自动对所述多个特征数据进行低阶非线性组合处理，得到所述一组低阶数据；所述自动对所述多个特征数据进行线性组合处理，得到一组第一辅助数据，包括：通过基于FM的处理模块，自动对所述多个特征数据进行线性组合处理，得到所述一组第一辅助数据；或者，所述自动对所述多个特征数据进行低阶非线性组合处理，得到所述一组低阶数据，包括：通过基于域感知分解机FFM的处理模块，自动对所述多个特征数据进行低阶非线性组合处理，得到所述一组低阶数据；所述自动对所述多个特征数据进行线性组合处理，得到一组第一辅助数据，包括：通过基于FFM的处理模块，自动对所述多个特征数据进行线性组合处理，得到所述一组第一辅助数据。
可选的，在所述根据多个目标数据确定预测数据之前，所述方法还包括：结合专家经验以及所述多个特征数据，确定x个经验特征数据，x≥1，所述x个经验特征数据包括：结合专家经验在所述多个特征数据中获取到的数据，和/或，结合专家经验对所述多个特征数据中的至少两个特征数据进行非线性组合处理得到数据；自动对所述x个经验特征数据进行处理，得到一组第二辅助数据，所述多个目标数据还包括：所述一组第二辅助数据，其中，当x＝1时，所述第二辅助数据与一个经验特征数据相关，当x＞1时，所述第二辅助数据与所述x个经验特征数据中的至少两个经验特征数据的线性组合相关，和/或，所述第二辅助数据与所述x个经验特征数据中的至少两个经验特征数据的非线性组合相关。
也即是，本申请能够根据专家经验的指导，确定对预测数据影响较大的x个经验特征数据，并对x个经验特征数据进行处理得到的一组第二辅助数据。且最终确定预测数据所依据的目标数据还包括该第二辅助数据，从而进一步的增多了最终确定预测数据所基于的目标数据的个数，进一步提高了预测数据的准确度。
可选的，所述多个特征数据为多个稀疏特征数据，在所述获取多个特征数据之后，所述方法包括：将所述多个稀疏特征数据，转换为多个稠密特征数据；所述自动对所述多个特征数据进行两种不同的非线性组合处理得到两组处理数据，包括：自动对所述多个稠密特征数据进行两种不同的非线性组合处理得到所述两组处理数据。
也即是，本申请在获取到多个稀疏特征数据后，还能够对该多个稀疏特征数据进行转换，得到多个 稠密特征数据，且稠密特征数据的数据量远小于稀疏特征数据的数据量，对多个稠密特征数据进行处理的效率高于对多个稀疏特征数据进行处理的效率。
第二方面，提供了一种数据处理装置，所述数据处理装置包括：获取单元，用于获取多个特征数据；非线性处理单元，用于自动对所述多个特征数据进行两种不同的非线性组合处理得到两组处理数据，其中，所述两组处理数据包括一组高阶数据和一组低阶数据，所述高阶数据与所述多个特征数据中m个特征数据的非线性组合相关，所述低阶数据与所述多个特征数据中n个特征数据的非线性组合相关，m≥3，m＞n≥2；预测单元，用于根据多个目标数据确定预测数据，所述多个目标数据包括所述两组处理数据。
可选的，所述非线性处理单元包括：高阶处理子单元，用于自动对所述多个特征数据进行高阶非线性组合处理，得到所述一组高阶数据；低阶处理子单元，用于自动对所述多个特征数据进行低阶非线性组合处理，得到所述一组低阶数据。
可选的，所述数据处理装置还包括：线性处理单元，用于自动对所述多个特征数据进行线性组合处理，得到一组第一辅助数据，所述第一辅助数据与所述多个特征数据中的至少两个特征数据的线性组合相关，所述多个目标数据还包括：所述至少一个第一辅助数据。
可选的，所述数据处理装置还包括：基于因子分解机FM的处理模块或基于域感知分解机FFM的处理模块，所述基于FM的处理模块包括：所述低阶处理子单元和所述线性处理单元；所述基于FFM的处理模块包括：所述低阶处理子单元和所述线性处理单元。
可选的，所述数据处理装置还包括：确定单元，结合专家经验以及所述多个特征数据，确定x个经验特征数据，x≥1，所述x个经验特征数据包括：结合专家经验在所述多个特征数据中获取到的数据，和/或，结合专家经验对所述多个特征数据中的至少两个特征数据进行非线性组合处理得到数据；经验处理单元，用于自动对所述x个经验特征数据进行处理，得到一组第二辅助数据，所述多个目标数据还包括：所述一组第二辅助数据，其中，当x＝1时，所述第二辅助数据与一个经验特征数据相关，当x＞1时，所述第二辅助数据与所述x个经验特征数据中的至少两个经验特征数据的线性组合相关，和/或，所述第二辅助数据与所述x个经验特征数据中的至少两个经验特征数据的非线性组合相关。
可选的，所述多个特征数据为多个稀疏特征数据，所述数据处理装置还包括：转换单元，用于将所述多个稀疏特征数据，转换为多个稠密特征数据；所述非线性处理单元还用于：自动对所述多个稠密特征数据进行两种不同的非线性组合处理得到所述两组处理数据。
第三方面，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第一方面述的数据处理方法。
第四方面，提供了一种计算机程序产品，当其在计算机上运行时，使得计算机执行第一方面所述的数据处理方法。
第五方面，提供了一种数据处理装置，所述数据处理装置包括：至少一个处理器、至少一个网络接口、存储器以及至少一个总线，存储器与网络接口分别通过总线与处理器相连；处理器被配置为执行存储器中存储的指令；处理器通过执行指令来实现上述第一方面或第一方面中任意一种可能的实现方式所提供的数据处理方法。
上述第二方面至第五方面所获得的技术效果与上述第一方面中对应的技术手段获得的技术效果近似，本申请在此不再赘述。
本申请提供的技术方案带来的有益效果是：
相关技术中对多个特征数据进行一种非线性组合处理(高阶非线性组合处理)能够得到的一组处理数据(包括至少一个高阶数据)；而本申请能够自动对多个特征数据进行两种不同的非线性组合处理(高 阶非线性组合处理和低阶非线性组合处理)并得到两组处理数据。也即是，本申请中得到的两组处理数据的数据量大于相关技术中得到的一组处理数据的数据量，且准确的预测需要基于大量的数据来实现，因此，本申请中基于两组处理数据得到的预测数据的准确度高于相关技术中基于一组处理数据得到的预测数据的准确度。
附图说明
图1为本发明实施例提供的一种数据处理方法的应用场景示意图；
图2为本发明实施例提供的一种数据处理装置的结构示意图；
图3为本发明实施例提供的一种程序模块的结构示意图；
图4为本发明实施例提供的一种数据处理方法的方法流程图；
图5为本发明实施例提供的另一种数据处理装置的结构示意图；
图6为本发明实施例提供的又一种数据处理装置的结构示意图；
图7为本发明实施例提供的再一种数据处理装置的结构示意图；
图8为本申请另一实施例提供的一种数据处理装置的结构示意图。
具体实施方式
为使本申请的目的、技术方案和优点更加清楚，下面将结合附图对本申请实施方式作进一步地详细描述。
图1为本发明实施例提供的一种数据处理方法的应用场景示意图，如图1所示，服务器1可以与多个客户端2建立通信连接，且服务器1能够获取与客户端2相关的数据；客户端2也能够从服务器1上获取服务器上存储的数据。数据处理装置0与服务器1相连接，数据处理装置0还能够设置在服务器1上，数据处理装置0能够对服务器获取到的数据进行处理。示例的，数据处理装置0和服务器1，以及服务器1和客户端2均可以通过有线网络或者无线网络连接。
需要说明的是，本发明实施例会涉及到“线性组合相关”、“非线性组合相关”、“线性组合处理”以及“非线性组合处理”这几个名词，本发明实施例在此对这几个名词进行解释说明：
示例的，在对多个特征数据进行组合处理(线性组合处理或者非线性组合处理)时，可以根据预设方程对多个特征数据进行组合处理。当该预设方程为线性方程时，对多个特征数据的组合处理也称为“线性组合处理”，此时，组合处理得到的数据与该多个特征数据的“线性组合相关”。当该预设方程为非线性方程时，对多个特征数据的“组合处理”也称为“非线性组合处理”，此时，组合处理得到的数据与该多个特征数据的“非线性组合相关”。其中，线性方程为多元一次方程，线性方程中的自变量与因变量的关系是线性关系，线性方程在笛卡尔坐标系上表示为一条直线。非线性方程为多元多次方程，非线性方程中的自变量与因变量的关系并不是线性关系，且非线性方程在笛卡尔坐标系上表示为曲线。
例如，若该多个特征数据包括：特征数据a1和特征数据a2，则对多个特征数据进行线性组合处理能够得到数据y1＝d1*a1+d2*a2，其中，d1为特征数据a1的权重，d2为特征数据a2的权重，y1与特征数据a1和特征数据a2的线性组合相关；对多个特征数据进行非线性组合处理能够得到数据y2＝a1*a2，y2与特征数据a1和特征数据a2的非线性组合相关。需要说明的是，本发明实施例仅仅以y1＝d1*a1+d2*a2与y2＝a1*a2为例，对多个特征数据的线性组合处理以及非线性组合处理进行举例说明，实际应用中，线性组合处理得到的数据可以不为d1*a1+d2*a2，非线性组合处理得到的数据可以不为a1*a2，本发明实施例对此不作限定。
图2为本发明实施例提供的一种数据处理装置的结构示意图，可选的，该数据处理装置可以为图1所示的数据处理装置。如图2所示，该数据处理装置可以为包括：至少一个处理器10221(例如中央处理器)，至少一个网络接口10222，存储器10223，和至少一个总线10224，总线10224可以用于实现处理器、网络接口和存储器之间的连接通信；存储器10223与网络接口10222分别通过总线10224与处理器10221相连。处理器10221用于执行存储器10223中存储的可执行模块，例如计算机程序。存储器10223可能包含高速随机存取存储器(Random Access Memory，RAM)，也可能还包括非不稳定的存储器(non-volatile memory)，例如至少一个磁盘存储器。通过至少一个网络接口10222(有线或者无线)实现该数据处理装置与至少一个其他装置之间的通信连接。在一些实施方式中，存储器10223存储了程序模块10225，程序模块10225能够被处理器10221执行，以实现如图4所示的数据处理方法。
示例的，如图3所示，该程序模块可以包括：
获取模块01，用于获取多个特征数据，该多个特征数据为多个稀疏特征数据。
转换模块02，用于将多个稀疏特征数据，转换为多个稠密特征数据；
非线性处理模块(图3中未标出)，用于自动对多个特征数据进行两种不同的非线性组合处理得到两组处理数据，其中，两组处理数据包括一组高阶数据和一组低阶数据，高阶数据与多个特征数据中m个特征数据的非线性组合相关，低阶数据与多个特征数据中n个特征数据的非线性组合相关，m≥3，m＞n≥2；非线性处理模块还用于：自动对多个稠密特征数据进行两种不同的非线性组合处理得到两组处理数据。
线性处理模块04，用于自动对多个特征数据进行线性组合处理，得到一组第一辅助数据，第一辅助数据与多个特征数据中的至少两个特征数据的线性组合相关，多个目标数据还包括：至少一个第一辅助数据。
预测模块05，用于根据多个目标数据确定预测数据，多个目标数据包括两组处理数据；
可选的，该非线性处理模块可以包括：高阶处理子模块031和低阶处理子模块032，高阶处理子模块031可以用于自动对多个特征数据进行高阶非线性组合处理，得到一组高阶数据；低阶处理子模块032可以用于自动对多个特征数据进行低阶非线性组合处理，得到一组低阶数据。
可选的，程序模块还可以包括：基于因子分解机(Factorization Machine，FM)的处理模块或基于域感知分解机(Field-aware Factorization Machine，FFM)的处理模块，该基于FM的处理模块包括：低阶处理子模块032和线性处理模块04；基于FFM的处理模块包括：低阶处理子模块032和线性处理模块04。
可选的，该程序模块还可以包括：确定模块和经验处理模块(图3中均未示出)，确定模块可以用于结合专家经验以及多个特征数据，确定x个经验特征数据，x≥1，x个经验特征数据包括：结合专家经验在多个特征数据中获取到的数据，和/或，结合专家经验对多个特征数据中的至少两个特征数据进行非线性组合处理得到数据；经验处理模块可以用于自动对x个经验特征数据进行处理，得到一组第二辅助数据，多个目标数据还包括：一组第二辅助数据，其中，当x＝1时，第二辅助数据与一个经验特征数据相关，当x＞1时，第二辅助数据与所述x个经验特征数据中的至少两个经验特征数据的线性组合相关，和/或，第二辅助数据与所述x个经验特征数据中的至少两个经验特征数据的非线性组合相关。
图4为本发明实施例提供的一种数据处理方法的方法流程图，该数据处理方法可以用于图2所示的数据处理装置，如图4所示，该数据处理方法可以包括：
步骤401、获取多个稀疏特征数据。
人们在对可动态变化的事物(例如某一网站的访问次数、某一文件的下载量、应用商城中某一应用的图标的点击率或视频网站中某一视频的点击率)进行分析时，通常可以获取事物的多个稀疏特征数据。
例如，应用商城的客户端中显示有多个应用的图标，用户需要下载某一应用时，可以在该多个应用中点击需要下载的应用的图标，从而进入该应用的下载界面，并决定是否下载该应用。因此，用户是否会下载应用与用户是否会点击应用的图标相关，应用商城的服务器可以基于应用的图标的点击率(Click Through Rate，CTR)来决定该应用的推荐优先级，为点击率高的应用的分配较高的优先级，相应的在应用商城中优先显示这些应用的图标，以将符合用户需求的应用及时准确得推荐给用户，提高用户黏性，在这种场景中，对应用的图标的点击率的预测就显得尤为重要。在需要对应用的图标的点击率进行预测时，在步骤401中数据处理装置获取的多个稀疏特征数据可以为该应用的稀疏特征数据，其中，应用的稀疏特征数据包括：用户的属性数据和应用的属性数据。用户的属性数据可以包括：用户的标识、用户的性别、用户的年龄、用户的职业、用户的收入、用户的爱好、用户的教育情况等属性数据。应用的属性数据可以包括：应用的标识、应用的名称、应用的开发者、应用的类型、应用的安装包大小、应用的标签、应用的打分、应用的评论等属性数据。例如，某一应用的稀疏特征数据中：用户的标识为小明、用户的性别为男、用户的年龄为20岁、应用的名称为F1以及应用的类型为聊天应用，步骤401中获取的上述稀疏特征数据可以按照一定规则排布，例如“小明、男、20、F1、聊天”。
又例如，当用户通过终端登录网站，并在能够下载文件的网站中搜索某一关键词后，网站可以根据该关键词，控制终端显示与该关键词相关的多个文件的入口，用户可以根据需求点击某一文件的入口进行文件的下载。示例的，被多次下载的文件通常会较符合多个用户的下载需求，因此，网站可以根据文件的下载量来决定文件的推荐优先级，为下载量高的文件分配较高的优先级，相应的控制终端优先显示这些文件入口。当人们需要对某一文件的下载量进行分析时，人们可以获取该文件的多个稀疏特征数据。示例的，文件的多个稀疏特征数据可以包括：文件的名称、文件的大小和文件的标签。例如，某一文件的稀疏特征数据中：文件的名称为F2、文件的大小为2兆和文件的标签为科技，步骤401中获取的上述稀疏特征数据可以按照一定规则排布，例如“F2、2兆、科技”。
如图3所示，数据处理装置中的程序模块可以包括获取模块，在步骤401中数据处理装置可以通过该获取模块获取多个稀疏特征数据。
步骤402、对多个稀疏特征数据进行转换，得到多个稠密特征数据。
数据处理装置在获取到多个数据量较多的稀疏特征数据后，可以对多个稀疏特征数据进行转换，以得到数据量较少的多个稠密特征数据，且稠密特征数据的数据量远小于稀疏特征数据的数据量，后续数据处理装置对该多个稠密特征数据进行处理的效率高于数据处理装置对多个稀疏特征数据进行处理的效率。
如图3所示，数据处理装置中的程序模块还可以包括转换模块，在步骤402中数据处理装置可以通过该转换模块对获取模块获取到的多个稀疏特征数据进行转换，以得到多个稠密特征数据。
步骤403、自动对多个稠密特征数据进行高阶非线性组合处理，得到一组高阶数据。
在获取到多个稠密特征数据后，数据处理装置能够对输入的多个稠密特征数据进行高阶非线性组合处理，得到一组高阶数据，每个高阶数据与该多个稠密特征数据中的m个特征数据的非线性组合相关。其中，m为大于或等于3的任意一个整数，也即是，一组高阶数据可以包括：多个与至少三个特征数据的非线性组合相关的高阶数据。例如，m为5、6或7，此时，该一组高阶数据包括：与五个特征数据的非线性组合相关的高阶数据、与六个特征数据的非线性组合相关的高阶数据以及与七个特征数据的非线 性组合相关的高阶数据。
可选的，数据处理装置中的程序模块还可以包括非线性处理模块，该非线性处理模块包括：高阶处理子模块，在步骤403中数据处理装置可以通过该高阶处理子模块自动对多个稠密特征数据进行高阶非线性组合处理，得到一组高阶数据。示例的，该高阶处理子模块可以为：基于深度神经网络(Deep Neural Network，DNN)的处理子模块、基于分解机支持神经网络(Factorization-machine supported Neural Network，FNN)的处理子模块或基于产品基神经网络(Product-based Neural Network，PNN)的处理子模块。
在第一种可实现方式中，如图3所示，高阶处理子模块可以为基于DNN的处理子模块，基于DNN的处理子模块可以包括多个级联的第一处理元0311。该基于DNN的处理子模块共包括H+1层的第一处理元(图3中示意性的示出了两层第一处理元0311)，第一层的第一处理元与转换模块02以及非第一层的第一处理元相连接；第i层的第一处理元与第i-1层的第一处理元以及第i+1层的第一处理元相连接；第H+1层的第一处理元与第H层级的第一处理元以及该数据处理装置中最终用于确定预测数据的预测模块05相连接，i为大于1且小于H+1的整数，H为大于或等于1的整数。该转换模块02可以向第一层的第一处理元输入该多个稠密特征数据，每一层的第一处理元021均可以对输入的数据进行非线性组合处理，并将处理后得到的数据输入下一层的第一处理元021，第H+1层级的每个第一处理元能够在对输入的数据进行非线性组合处理后，向预测模块05输出一个高阶数据，每个高阶数据均与数据处理装置获取到的多个稠密特征数据中的m个特征数据的非线性组合相关。
需要说明的是，多个级连的第一处理元中，任意两个相连的第一处理元具有一个用于连接这两个第一处理元的连边，且该连边具有预设的权重；第H+1层的第一处理元与预测模块的连边以及第一层的第一处理元与转换模块的连边也具有预设的权重。每个第一处理元处理后的数据均为：由通过连边输入该第一处理元的数据以及该连边上的权重计算得到的数据。
在第二种可实现方式中，高阶处理子模块可以为基于DNN的处理子模块，基于DNN的处理子模块可以包括多个级联的第二处理元。假设该基于DNN的处理子模块共包括H+1层的第二处理元，该高阶处理子模块自动对多个稠密特征数据进行高阶非线性组合处理得到的一组高阶数据为
Figure PCTCN2018093093-appb-000001
其中，σ(x)是为激活函数，a H为以第H层级的第二处理元处理后数据为元素的向量，b H+1为第H+1层级的第二处理元的偏移项，W H+1为以第H层级的第二处理元与第H+1层级的第二处理元的连边上的权重为元素的矩阵。示例的，激活函数可以为sigmoid函数(也称S形函数)、修正线性单元(rectifiedlinearunits，relu)函数或双曲正切(tangentialhyperbolic，tanh)函数等激活函数。
第H层第二处理元共包括m H个第二处理元，且m H个第二处理元中每个第二处理元处理后得到的数据均为实数，第H+1层第二处理元共包括m H+1个第二处理元，且m H+1个第二处理元中每个第二处理元处理后得到的数据均为实数，第H层第二处理元与第H+1层第二处理元的连边上的所有权重共包括m H×m H+1个实数。l为大于或等于1，且小于或等于H的整数，a l为以每个第l层的第二处理元处理后的数据为元素的向量，b l+1为第l+1层第二处理元的偏移项，W l+1为以第l层的第二处理元与第l+1层第二处理元的每个连边上的权重为元素的向量，a l+1为以每个第l+1层的第二处理元处理后的数据为元素的向量，a l+1＝σ(W l+1a l+b l+1)。
步骤404、自动对多个稠密特征数据进行低阶非线性组合处理，得到一组低阶数据。
在获取到多个稠密特征数据后，数据处理装置还能够对输入的多个稠密特征数据进行低阶非线性组合处理，得到一组低阶数据，每个低阶数据与该多个稠密特征数据中的n个特征数据的非线性组合相关。其中，n为大于或等于2，且小于m的任意一个整数。例如，m为5、6或7，此时，n可以为2、3或4，该一组低阶数据包括：与两个特征数据的非线性组合相关的低阶数据、与三个特征数据的非线性组合相关的低阶数据以及与四个特征数据的非线性组合相关的低阶数据。
步骤405、自动对多个稠密特征数据进行线性组合处理，得到一组第一辅助数据。
在获取到多个稠密特征数据后，数据处理装置还能够对输入的多个稠密特征数据进行线性组合处理，得到一组第一辅助数据，每个第一辅助数据与该多个稠密特征数据中的至少两个特征数据的线性组合相关。示例的，一组第一辅助数据可以包括：与两个特征数据的线性组合相关的第一辅助数据、与三个特征数据的线性组合相关的第一辅助数据以及与四个特征数据的线性组合相关的第一辅助数据。
需要说明的是，在步骤404中数据处理装置可以通过数据处理装置中的低阶处理子模块自动对多个稠密特征数据进行低阶非线性组合处理，得到一组低阶数据。在步骤405中数据处理装置可以通过数据处理装置中的线性处理模块自动对多个稠密特征数据进行线性组合处理，得到一组第一辅助数据。
进一步的，图4所示的数据处理装置中的程序模块还可以包括基于FM的处理模块或基于FFM的处理模块，该基于FM的处理模块包括：低阶处理子模块032和线性处理模块04，基于FFM的处理模块包括：低阶处理子模块032和线性处理模块04。也即是，在执行步骤404和步骤405时，数据处理装置可以直接通过基于FM的处理模块或基于FFM的处理模块对多个稠密特征数据分别进行低阶非线性组合处理以及线性组合出来，以得到一组低阶数据和一组第一辅助数据。
在第一种可实现方式中，数据处理装置可以直接通过基于FM的处理模块对多个稠密特征数据分别进行低阶非线性组合处理以及线性组合，以得到一组低阶数据和一组第一辅助数据，该基于FM的处理模块处理得到的可以表示为：
Figure PCTCN2018093093-appb-000002
该
Figure PCTCN2018093093-appb-000003
包括一组低阶数据和一组第一辅助数据。该一组低阶数据为
Figure PCTCN2018093093-appb-000004
该一组第一辅助数据为
Figure PCTCN2018093093-appb-000005
Figure PCTCN2018093093-appb-000006
其中，n为获取模块获取到的多个稠密特征数据的数量，V a为该多个稠密特征数据中第a个特征数据的隐式向量，1≤a≤n-1，且V a为具有指定维度，V a中的每个元素均为实数，V b为该多个稠密特征数据中第b个特征数据的隐式向量，a+1≤b≤n，且V b具有指定维度，V b中的每个元素均为实数，X i,a为该多个稠密特征数据中的第a个特征数据，X i,b为该多个稠密特征数据中的第b个特征数据，X i,j为该多个稠密特征数据中的第j个特征数据，1≤j≤n，w j为该第j个特征数据的权重。
在第二种可实现方式中，数据处理装置可以直接通过基于FFM的处理模块对多个稠密特征数据分别进行低阶非线性组合处理以及线性组合，以得到一组低阶数据和一组第一辅助数据，该基于FFM的处理模块处理得到的可以表示为：
Figure PCTCN2018093093-appb-000007
包括一组低阶数据和一组第一辅助数据。该一组低阶数据为
Figure PCTCN2018093093-appb-000008
该一组第一辅助数据为
Figure PCTCN2018093093-appb-000009
Figure PCTCN2018093093-appb-000010
其中，函数f a表示该多个稠密特征数据中的第a个特征数据所处的域，函数f b表示输入该多个稠密特征数据中的第b个特征数据所处的域，
Figure PCTCN2018093093-appb-000011
为该多个稠密特征数据中第a个特征数据对f b的隐式向量，且
Figure PCTCN2018093093-appb-000012
为具有指定维度，
Figure PCTCN2018093093-appb-000013
中的每个元素均为实数，
Figure PCTCN2018093093-appb-000014
为该多个稠密特征数据中第b个特征数据对f a的隐式向量，且
Figure PCTCN2018093093-appb-000015
具有指定维度，
Figure PCTCN2018093093-appb-000016
中的每个元素均为实数。
步骤406、根据多个目标数据确定预测数据，该多个目标数据包括一组低阶数据、一组高阶数据和一组第一辅助数据。
在获取到一组低阶数据、一组高阶数据以及一组第一辅助数据后，数据处理装置就可以根据包括该一组低阶数据、一组高阶数据以及一组第一辅助数据的多个目标数据，确定预测数据。示例的，请继续参考图3，该数据处理装置中的程序模块还可以包括预测模块05，在步骤406中，数据处理装置可以通 过该预测模块05确定预测数据，可选的，预测模块确定的预测数据可以为：
Figure PCTCN2018093093-appb-000017
Figure PCTCN2018093093-appb-000018
示例的，当步骤401中获取到的多个稀疏特征数据为应用的稀疏特征数据(包括用户的属性数据和应用的属性数据)时，步骤406中数据处理装置得到的预测数据可以用于表征该应用的图标的点击率。需要说明的是，如图1所示，服务器1(应用商城的服务器)可以通过数据处理装置0对用户以及应用商城中的每个应用进行分析，以确定该用户对应用商城中每个应用的图标的点击率，并根据每个应用的图标的点击率对多个应用进行排序，并相应的应用商城界面的信息发送给客户端2(应用商城的客户端)，客户端2基于该信息显示应用商城界面，该界面中显示的应用的图标按照上述确定的顺序排布。
也即是，服务器1会针对每个用户，重新对所有的应用的图标进行排序，从而使得每个用户使用的应用商城的客户端中所有应用的排序均与该用户的属性特征相关，使应用的排序更能够符合每个用户的需求。例如，若应用商城包括四个应用，分别为应用Q1、应用Q2、应用Q3和应用Q4，数据处理装置在应用的稀疏特征数据进行处理后，确定用户对应用Q1的图标的点击率最大，其次是应用Q2的图标、应用Q4的图标和应用Q3的图标。则服务器可以将这四个应用的图标按照应用Q1的图标、应用Q2的图标、应用Q3的图标和应用Q4的图标的顺序进行排列，进而使得客户端显示的应用商城界面中的这四个应用的图标按照应用Q1的图标、应用Q2的图标、应用Q3的图标和应用Q4的图标的顺序排列。
进一步的，为了提高数据处理装置得到的预测数据的准确性，在数据处理装置执行步骤401之前，服务器可以获取多条训练数据，并可以通过训练装置根据该多条数据对数据处理装置进行训练，以调整数据处理装置在处理数据过程中使用到的权重。示例的，每条训练数据可以包括：一组稀疏特征数据和与该组稀疏特征数据对应的点击数据。
以用户P和应用Q为例，一条训练数据中的一组稀疏特征数据可以包括：用户P的属性数据和应用Q的属性数据，该条训练数据中的点击数据可以包括：用户P的标识、应用Q的标识以及用于指示用户P是否对应用Q的图标进行点击的指示数据。可选的，该条训练数据中的点击数据还可以包括：统计该条训练数据的时间、统计该条训练数据时的天气情况、统计该条训练数据时的交通情况、统计该条训练数据时的网络连接情况等统计该条训练数据时的上下文信息。用户P的标识可以为用户P的身份标识号(identification，ID)，应用Q的标识也可以为应用Q的ID。
训练装置在对该数据处理装置进行训练时，可以首先向该数据处理装置中的获取模块输入多条训练数据中的多组稀疏特征数据，数据处理装置能够依次对输入的每组稀疏特征数据进行处理，得到一个预测数据，该预测数据用于指示用户是否会点击应用的图标。训练装置还能够将数据处理装置根据一组稀疏特征数据得到的预测数据与该组稀疏特征数据对应的点击数据进行比较，并判断该预测数据与该点击数据是否均指示：用户会点击应用的图标，或用户不会点击应用的图标。如果该预测数据与该点击数据不是同时指示：用户会点击应用的图标，或用户不会点击应用的图标，则该训练装置需要对数据处理装置处理数据的过程中使用到的权重进行调整。在通过训练装置对数据处理装置训练完毕后，该数据处理装置在处理数据的过程中使用到的权重已经经过了多次调整，数据处理装置能够对多个稀疏特征数据进行处理得到较准确的预测数据。
另外，在获取到多个特征数据后，如果对该多个特征数据中的一部分特征数据进行高阶非线性组合处理得到的高阶数据的个数为第一个数，对该多个特征数据中的另一部分特征数据进行低阶非线性组合处理得到的低阶数据的个数为第二个数，则最终确定预测数据所基于的目标数据的个数等于第一个数与第二个数之和。如果对该多个数据进行高阶非线性组合处理，则得到的高阶数据的个数大于第一个数， 且对该多个数据也进行低阶非线性组合处理，则得到的低阶数据的个数大于第二个数，且最终确定预测数据所基于的目标数据的个数也大于第一个数与第二个数之和。也即是，在对多个特征数据分别进行高阶非线性组合处理以及低阶非线性组合处理时，最终确定预测数据所基于的目标数据的个数就会较多，预测数据的准确度就会越高。因此，本发明实施例中，在步骤403、步骤404和步骤405中均是对步骤402中得到的所有稠密特征数据进行处理，从而使得步骤406中确定预测数据所依据的目标数据较多，步骤406中得到的预测数据的准确度较高。
可选的，图3所示的数据处理装置中的程序模块还可以包括确定模块和经验处理模块，在步骤406之前，数据处理装置还可以通过该确定模块结合专家经验以及该多个稠密特征数据，确定x个经验特征数据，以及通过该经验处理模块自动对x个经验特征数据进行处理，以得到一组第二辅助数据。在步骤406中数据处理装置中的程序模块中的预测模块确定预测数据所依据的多个目标数据还可以包括：该一组第二辅助数据。也即是，该多个目标数据可以包括：一组高阶数据、一组低阶数据、一组第一辅助数据以及一组第二辅助数据。
示例的，x为大于或等于1的整数，x个经验特征数据包括：结合专家经验在多个稠密特征数据中获取到的数据，和/或，结合专家经验对多个稠密特征数据中的至少两个特征数据进行非线性组合处理得到数据。其中，当x＝1时，第二辅助数据与一个经验特征数据相关，当x＞1时，第二辅助数据与x个经验特征数据中的至少两个经验特征数据的线性组合相关，和/或，第二辅助数据与x个经验特征数据中的至少两个经验特征数据的非线性组合相关。
可选的，本发明实施例中提到的“专家”是指计算机技术领域的一个技术人员或多个技术人员，且“专家”知晓计算机技术领域所有的普通技术知识，并且能够获知计算机技术领域的现有技术，并且具有常规的实验手段能力。本发明实施例中提到的“专家经验”为“专家”确定出的“预设公式”。确定模块结合“专家经验”以及多个稠密特征数据，确定x个经验特征数据，包括：确定模块将多个稠密特征数据作为“预设公式”的自变量输入“预设公式”，预设公式得到的结果为x个经验特征数据。示例的，假设该多个稠密特征数据分别为：数据a1、数据a2、数据a3、数据a4、数据a5、数据a6、数据a7、数据a8、数据a9以及数据a10共10个数据。则确定模块可以结合专家经验在多个稠密特征数据中获取数据a1，作为一个经验特征数据；或者，确定模块可以结合专家经验对多个稠密特征数据中的数据a9和数据a10进行非线性组合处理得到一个经验数据；或者，确定模块获取数据a1并将数据a1作为一个经验特征数据，并且结合专家经验对数据a9和数据a10进行非线性组合处理得到另一个经验数据。
确定模块在得到x个经验特征数据后，经验处理单元还能够自动对x个经验特征数据进行处理，得到一组第二辅助数据。需要说明的是，当x＝1时，该第二辅助数据与一个经验特征数据相关，当x＞1时，第二辅助数据与x个经验特征数据中的至少两个经验特征数据的线性组合和/或非线性组合组合相关。示例的，当x＞1时，该第二辅助数据与x个经验特征数据中至少两个经验特征数据的线性组合相关；或者该第二辅助数据与x个经验特征数据中至少两个经验特征数据的非线性组合相关；或者第二辅助数据同时与该至少两个经验特征数据的线性组合和非线性组合相关。可选的，该经验处理模块可以为基于对数几率回归(Logistic Regression，LR)的处理模块、基于FM的处理模块、基于FFM的处理模块或基于DNN的处理模块。
需要说明的是，经验处理模块在对多个稠密特征数据进行处理时，需要结合专家经验，但是非线性处理模块和线性处理模块在对多个稠密特征数据进行处理时，均不需要结合专家经验，就能够自动的对特征数据进行处理。
另外，为了对本发明实施例提供的数据处理装置得到的预测数据的准确度进行验证，本发明实施例使用了多条训练数据和多组检测数据，分别对本发明实施例提供的数据处理装置以及相关技术中的数据处理装置分别进行了多次实验。并采用受试者工作特征(receiver operating characteristic curve，ROC)曲线下的面积(Area under the ROC curve，AUC)和损失函数(Logloss)对本发明实施例中的数据处理装置与相关技术中的数据处理装置进行了比较。经过比较得知：本发明实施例提供的数据处理装置得到的预测数据的准确度高于相关技术中的数据处理装置得到的预测数据。
示例的，当需要对应用Q的图标的点击率进行预测时，基于上述数据处理方法可知，数据处理装置可以获取应用Q的多个稀疏特征数据，并对该多个稀疏特征数据进行转换，得到应用Q的多个稠密特征数据。然后，数据处理装置再分别自动对应用Q的多个稠密特征数据分别进行：高阶非线性组合处理、低阶非线性组合处理以及线性组合处理，得到一组高阶数据、一组低阶数据以及一组第一辅助数据。进一步的，数据处理装置还可以结合专家经验以及应用Q的多个稠密特征数据，确定x个经验特征数据；以及自动对该x个经验特征数据进行处理，得到一组第二辅助数据。最后，数据处理装置根据包括该一组高阶数据、一组低阶数据、一组第一辅助数据以及一组第二辅助数据的多个目标数据，确定预测数据。该预测数据可以用于表征应用Q的图标的点击率。
综上所述，本发明实施例提供了一种数据处理方法中，能够自动对特征数据进行两种不同的非线性组合处理(高阶非线性组合处理和低阶非线性组合处理)并得到两组处理数据。而相关技术中对特征数据进行一种非线性组合处理(高阶非线性组合处理)能够得到的一组处理数据(包括至少一个高阶数据)。也即是，本发明实施例得到的两组处理数据的数据量大于相关技术中得到的一组处理数据的数据量，且准确的预测需要基于大量的数据来实现，因此，本发明实施例中基于两组处理数据得到的预测数据的准确度高于相关技术中基于一组处理数据得到的预测数据的准确度。
进一步的，本发明实施例中确定预测数据所依据的目标数据还可以包括：一组第一辅助数据和一组第二辅助数据，也即是，本发明实施例中的目标数据可以包括：一组高阶数据、一组低阶数据、一组第一辅助数据和一组第二辅助数据共四组数据，从而进一步的增多了最终确定预测数据所基于的目标数据的个数，进一步提高了预测数据的准确度。
图5为本发明实施例提供的另一种数据处理装置的结构示意图，该数据处理装置可以为图1所示的数据处理装置，如图5所示，该数据处理装置50可以包括：
获取单元501，用于获取多个特征数据；
非线性处理单元502，用于自动对多个特征数据进行两种不同的非线性组合处理得到两组处理数据，其中，两组处理数据包括一组高阶数据和一组低阶数据，高阶数据与多个特征数据中m个特征数据的非线性组合相关，低阶数据与多个特征数据中n个特征数据的非线性组合相关，m≥3，m＞n≥2；
预测单元503，用于根据多个目标数据确定预测数据，多个目标数据包括两组处理数据。
综上所述，本发明实施例提供了一种数据处理装置，非线性处理单元能够自动对特征数据进行两种不同的非线性组合处理(高阶非线性组合处理和低阶非线性组合处理)并得到两组处理数据。而相关技术中对特征数据进行一种非线性组合处理(高阶非线性组合处理)能够得到的一组处理数据(包括至少一个高阶数据)。也即是，本发明实施例得到的两组处理数据的数据量大于相关技术中得到的一组处理数据的数据量，且准确的预测需要基于大量的数据来实现，因此，本发明实施例中基于两组处理数据得到的预测数据的准确度高于相关技术中基于一组处理数据得到的预测数据的准确度。
可选的，请继续参考图5，该非线性处理单元502可以包括：
高阶处理子单元5021，用于自动对多个特征数据进行高阶非线性组合处理，得到一组高阶数据；
低阶处理子单元5022，用于自动对多个特征数据进行低阶非线性组合处理，得到一组低阶数据。
可选的，图6为本发明实施例提供的又一种数据处理装置的结构示意图，如图6所示，在图5的基础上，该数据处理装置50还可以包括：
线性处理单元504，用于自动对多个特征数据进行线性组合处理，得到一组第一辅助数据，第一辅助数据与多个特征数据中的至少两个特征数据的线性组合相关，多个目标数据还包括：至少一个第一辅助数据。
可选的，数据处理装置还包括：基于因子分解机FM的处理模块或基于域感知分解机FFM的处理模块，其中，基于FM的处理模块包括：低阶处理子单元和线性处理单元，基于FFM的处理模块包括：低阶处理子单元和线性处理单元。
可选的，图7为本发明实施例提供的再一种数据处理装置的结构示意图，如图7所示，在图5的基础上，该数据处理装置50还可以包括：
确定单元505，结合专家经验以及多个特征数据，确定x个经验特征数据，x≥1，x个经验特征数据包括：结合专家经验在多个特征数据中获取到的数据，和/或，结合专家经验对多个特征数据中的至少两个特征数据进行非线性组合处理得到数据；
经验处理单元506，用于自动对x个经验特征数据进行处理，得到一组第二辅助数据，多个目标数据还包括：一组第二辅助数据，其中，当x＝1时，第二辅助数据与一个经验特征数据相关，当x＞1时，第二辅助数据与所述x个经验特征数据中的至少两个经验特征数据的线性组合相关，和/或，第二辅助数据与所述x个经验特征数据中的至少两个经验特征数据的非线性组合相关。
多个特征数据为多个稀疏特征数据，图8为本申请另一实施例提供的一种数据处理装置的结构示意图，如图8所示，在图5的基础上，该数据处理装置50还可以包括：
转换单元507，用于将多个稀疏特征数据，转换为多个稠密特征数据；
该非线性处理单元502还可以用于：自动对多个稠密特征数据进行两种不同的非线性组合处理得到两组处理数据。
综上所述，本发明实施例提供了一种数据处理装置，非线性处理单元能够自动对特征数据进行两种不同的非线性组合处理(高阶非线性组合处理和低阶非线性组合处理)并得到两组处理数据。而相关技术中对特征数据进行一种非线性组合处理(高阶非线性组合处理)能够得到的一组处理数据(包括至少一个高阶数据)。也即是，本发明实施例得到的两组处理数据的数据量大于相关技术中得到的一组处理数据的数据量，且准确的预测需要基于大量的数据来实现，因此，本发明实施例中基于两组处理数据得到的预测数据的准确度高于相关技术中基于一组处理数据得到的预测数据的准确度。
进一步的，本发明实施例中确定预测数据所依据的目标数据还可以包括：线性处理单元得到的一组第一辅助数据和经验处理单元得到的一组第二辅助数据，也即是，本发明实施例中的目标数据可以包括：一组高阶数据、一组低阶数据、一组第一辅助数据和一组第二辅助数据共四组数据，从而进一步的增多了最终确定预测数据所基于的目标数据的个数，进一步提高了预测数据的准确度。
在上述实施例中，能够全部或部分地通过软件、硬件、固件或者其任意组合来实现。当使用软件实现时，能够全部或部分地以计算机程序产品的形式实现，所述计算机程序产品包括一个或多个计算机指令。在计算机上加载和执行所述计算机程序指令时，全部或部分地产生按照本发明实施例所述的流程或功能。所述计算机能够是通用计算机、计算机网络、或者其他可编程装置。所述计算机指令能够存储在 计算机的可读存储介质中，或者从一个计算机可读存储介质向另一个计算机可读存储介质传输，例如，所述计算机指令能够从一个网站站点、计算机、服务器或数据中心通过有线(例如同轴电缆、光纤、数字用户线)或无线(例如红外、无线、微波等)方式向另一个网站站点、计算机、服务器或数据中心传输。所述计算机可读存储介质能够是计算机能够存取的任何可用介质或者包含一个或多个可用介质集成的服务器、数据中心等数据存储设备。所述可用介质能够是磁性介质(例如，软盘、硬盘、磁带)、光介质，或者半导体介质(例如固态硬盘)等。
本发明实施例中术语“和/或”，仅仅是一种描述关联对象的关联关系，表示可以存在三种关系，例如，A和/或B，可以表示：单独存在A，同时存在A和B，单独存在B这三种情况。另外，本文中字符“/”，一般表示前后关联对象是一种“或”的关系。
需要说明的是，本发明实施例提供的方法实施例能够与相应的装置实施例相互参考，本发明实施例对此不做限定。本发明实施例提供的方法实施例步骤的先后顺序能够进行适当调整，步骤也能够根据情况进行相应增减，任何熟悉本技术领域的技术人员在本发明实施例揭露的技术范围内，可轻易想到变化的方法，都应涵盖在本发明实施例的保护范围之内，因此不再赘述。上述本发明实施例序号仅仅为了描述，不代表实施例的优劣。
以上所述仅为本申请的可选实施例，并不用以限制本申请，凡在本申请的精神和原则之内，所作的任何修改、等同替换、改进等，均应包含在本申请的保护范围之内。

Claims (13)

  1. 一种数据处理方法，其特征在于，所述方法包括：
    获取多个特征数据；
    自动对所述多个特征数据进行两种不同的非线性组合处理得到两组处理数据，其中，所述两组处理数据包括一组高阶数据和一组低阶数据，所述高阶数据与所述多个特征数据中m个特征数据的非线性组合相关，所述低阶数据与所述多个特征数据中n个特征数据的非线性组合相关，m≥3，m＞n≥2；
    根据多个目标数据确定预测数据，所述多个目标数据包括所述两组处理数据。
  2. 根据权利要求1所述的方法，其特征在于，所述自动对所述多个特征数据进行两种不同的非线性组合处理得到两组处理数据，包括：
    自动对所述多个特征数据进行高阶非线性组合处理，得到所述一组高阶数据；
    自动对所述多个特征数据进行低阶非线性组合处理，得到所述一组低阶数据。
  3. 根据权利要求1或2所述的方法，其特征在于，在所述获取多个特征数据之后，所述方法还包括：
    自动对所述多个特征数据进行线性组合处理，得到一组第一辅助数据，所述第一辅助数据与所述多个特征数据中的至少两个特征数据的线性组合相关，所述多个目标数据还包括：所述至少一个第一辅助数据。
  4. 根据权利要求3所述的方法，其特征在于，
    所述自动对所述多个特征数据进行低阶非线性组合处理，得到所述一组低阶数据，包括：通过基于因子分解机FM的处理模块自动对所述多个特征数据进行低阶非线性组合处理，得到所述一组低阶数据；所述自动对所述多个特征数据进行线性组合处理，得到一组第一辅助数据，包括：通过基于FM的处理模块，自动对所述多个特征数据进行线性组合处理，得到所述一组第一辅助数据；
    或者，所述自动对所述多个特征数据进行低阶非线性组合处理，得到所述一组低阶数据，包括：通过基于域感知分解机FFM的处理模块，自动对所述多个特征数据进行低阶非线性组合处理，得到所述一组低阶数据；所述自动对所述多个特征数据进行线性组合处理，得到一组第一辅助数据，包括：通过基于FFM的处理模块，自动对所述多个特征数据进行线性组合处理，得到所述一组第一辅助数据。
  5. 根据权利要求1或2所述的方法，其特征在于，在所述根据多个目标数据确定预测数据之前，所述方法还包括：
    结合专家经验以及所述多个特征数据，确定x个经验特征数据，x≥1，所述x个经验特征数据包括：结合专家经验在所述多个特征数据中获取到的数据，和/或，结合专家经验对所述多个特征数据中的至少两个特征数据进行非线性组合处理得到数据；
    自动对所述x个经验特征数据进行处理，得到一组第二辅助数据，所述多个目标数据还包括：所述一组第二辅助数据；其中，当x＝1时，所述第二辅助数据与一个经验特征数据相关；当x＞1时，所述第二辅助数据与所述x个经验特征数据中的至少两个经验特征数据的线性组合相关，和/或，所述第二辅助数据与所述x个经验特征数据中的至少两个经验特征数据的非线性组合相关。
  6. 根据权利要求1或2所述的方法，其特征在于，所述多个特征数据为多个稀疏特征数据，
    在所述获取多个特征数据之后，所述方法包括：
    将所述多个稀疏特征数据，转换为多个稠密特征数据；
    所述自动对所述多个特征数据进行两种不同的非线性组合处理得到两组处理数据，包括：
    自动对所述多个稠密特征数据进行两种不同的非线性组合处理得到所述两组处理数据。
  7. 一种数据处理装置，其特征在于，所述数据处理装置包括：
    获取单元，用于获取多个特征数据；
    非线性处理单元，用于自动对所述多个特征数据进行两种不同的非线性组合处理得到两组处理数据，其中，所述两组处理数据包括一组高阶数据和一组低阶数据，所述高阶数据与所述多个特征数据中m个特征数据的非线性组合相关，所述低阶数据与所述多个特征数据中n个特征数据的非线性组合相关，m≥3，m＞n≥2；
    预测单元，用于根据多个目标数据确定预测数据，所述多个目标数据包括所述两组处理数据。
  8. 根据权利要求7所述的数据处理装置，其特征在于，所述非线性处理单元包括：
    高阶处理子单元，用于自动对所述多个特征数据进行高阶非线性组合处理，得到所述一组高阶数据；
    低阶处理子单元，用于自动对所述多个特征数据进行低阶非线性组合处理，得到所述一组低阶数据。
  9. 根据权利要求7或8所述的数据处理装置，其特征在于，所述数据处理装置还包括：
    线性处理单元，用于自动对所述多个特征数据进行线性组合处理，得到一组第一辅助数据，所述第一辅助数据与所述多个特征数据中的至少两个特征数据的线性组合相关，所述多个目标数据还包括：所述至少一个第一辅助数据。
  10. 根据权利要求9所述的数据处理装置，其特征在于，所述数据处理装置还包括：基于因子分解机FM的处理模块或基于域感知分解机FFM的处理模块，
    所述基于FM的处理模块包括：所述低阶处理子单元和所述线性处理单元；
    所述基于FFM的处理模块包括：所述低阶处理子单元和所述线性处理单元。
  11. 根据权利要求7或8所述的数据处理装置，其特征在于，所述数据处理装置还包括：
    确定单元，结合专家经验以及所述多个特征数据，确定x个经验特征数据，x≥1，所述x个经验特征数据包括：结合专家经验在所述多个特征数据中获取到的数据，和/或，结合专家经验对所述多个特征数据中的至少两个特征数据进行非线性组合处理得到数据；
    经验处理单元，用于自动对所述x个经验特征数据进行处理，得到一组第二辅助数据，所述多个目标数据还包括：所述一组第二辅助数据；其中，当x＝1时，所述第二辅助数据与一个经验特征数据相关；当x＞1时，所述第二辅助数据与所述x个经验特征数据中的至少两个经验特征数据的线性组合相关，和/或，所述第二辅助数据与所述x个经验特征数据中的至少两个经验特征数据的非线性组合相关。
  12. 根据权利要求7或8所述的数据处理装置，其特征在于，所述多个特征数据为多个稀疏特征数据，所述数据处理装置还包括：
    转换单元，用于将所述多个稀疏特征数据，转换为多个稠密特征数据；
    所述非线性处理单元还用于：
    自动对所述多个稠密特征数据进行两种不同的非线性组合处理得到所述两组处理数据。
  13. 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权利要求1至6任一所述的数据处理方法。
PCT/CN2018/093093 2017-06-30 2018-06-27 数据处理方法及装置 WO2019001463A1 (zh)

Priority Applications (2)

Application Number Priority Date Filing Date Title
US16/729,043 US11334758B2 (en) 2017-06-30 2019-12-27 Method and apparatus of data processing using multiple types of non-linear combination processing
US17/661,448 US11748452B2 (en) 2017-06-30 2022-04-29 Method for data processing by performing different non-linear combination processing

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN201710524501.3 2017-06-30
CN201710524501.3A CN109214543B (zh) 2017-06-30 2017-06-30 数据处理方法及装置

Related Child Applications (1)

Application Number Title Priority Date Filing Date
US16/729,043 Continuation US11334758B2 (en) 2017-06-30 2019-12-27 Method and apparatus of data processing using multiple types of non-linear combination processing

Publications (1)

Publication Number Publication Date
WO2019001463A1 true WO2019001463A1 (zh) 2019-01-03

Family

ID=64742771

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2018/093093 WO2019001463A1 (zh) 2017-06-30 2018-06-27 数据处理方法及装置

Country Status (3)

Country Link
US (2) US11334758B2 (zh)
CN (2) CN109214543B (zh)
WO (1) WO2019001463A1 (zh)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2021104132A1 (zh) * 2019-11-29 2021-06-03 华为技术有限公司 一种基于云虚拟机的数据访问方法及设备

Families Citing this family (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109214543B (zh) * 2017-06-30 2021-03-30 华为技术有限公司 数据处理方法及装置
WO2021022441A1 (zh) * 2019-08-05 2021-02-11 华为技术有限公司 数据传输方法、装置、电子设备及可读存储介质
CN112243021A (zh) * 2020-05-25 2021-01-19 北京沃东天骏信息技术有限公司 信息推送方法、装置、设备及计算机可读存储介质
CN111737578B (zh) * 2020-06-22 2024-04-02 陕西师范大学 一种推荐方法及系统

Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN103222243A (zh) * 2012-12-05 2013-07-24 华为技术有限公司 一种数据处理的方法及装置
US20140310301A1 (en) * 2013-04-15 2014-10-16 Qifang Li Hierarchical Data Server
CN105844330A (zh) * 2016-03-22 2016-08-10 华为技术有限公司 神经网络处理器的数据处理方法及神经网络处理器

Family Cites Families (10)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US5784402A (en) * 1995-01-09 1998-07-21 Kamilo Feher FMOD transceivers including continuous and burst operated TDMA, FDMA, spread spectrum CDMA, WCDMA and CSMA
US7233932B2 (en) * 2005-05-31 2007-06-19 Honeywell International, Inc. Fault detection system and method using approximate null space base fault signature classification
CN1897015A (zh) * 2006-05-18 2007-01-17 王海燕 基于机器视觉的车辆检测和跟踪方法及系统
JP2008148055A (ja) * 2006-12-11 2008-06-26 Sony Corp 画像処理装置、画像処理方法、表示装置、および投射型表示装置
US9443211B2 (en) * 2010-10-13 2016-09-13 International Business Machines Corporation Describing a paradigmatic member of a task directed community in a complex heterogeneous environment based on non-linear attributes
CN102867516B (zh) * 2012-09-10 2014-08-27 大连理工大学 一种采用高阶线性预测系数分组矢量量化的语音编解方法
CN106096877A (zh) * 2016-05-31 2016-11-09 乐视控股（北京）有限公司 一种数据生成方法及生成系统
WO2018219839A1 (en) * 2017-05-29 2018-12-06 Universiteit Gent Training of photonic reservoir computing systems
CN109214543B (zh) * 2017-06-30 2021-03-30 华为技术有限公司 数据处理方法及装置
CN109902222B (zh) * 2018-11-30 2022-05-13 华为技术有限公司 一种推荐方法及装置

Patent Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN103222243A (zh) * 2012-12-05 2013-07-24 华为技术有限公司 一种数据处理的方法及装置
US20140310301A1 (en) * 2013-04-15 2014-10-16 Qifang Li Hierarchical Data Server
CN105844330A (zh) * 2016-03-22 2016-08-10 华为技术有限公司 神经网络处理器的数据处理方法及神经网络处理器

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2021104132A1 (zh) * 2019-11-29 2021-06-03 华为技术有限公司 一种基于云虚拟机的数据访问方法及设备

Also Published As

Publication number Publication date
US11748452B2 (en) 2023-09-05
CN109214543A (zh) 2019-01-15
US20220261591A1 (en) 2022-08-18
CN109214543B (zh) 2021-03-30
US20200134361A1 (en) 2020-04-30
US11334758B2 (en) 2022-05-17
CN113076981A (zh) 2021-07-06

Similar Documents

Publication Publication Date Title
US10958748B2 (en) Resource push method and apparatus
WO2020182122A1 (zh) 用于生成文本匹配模型的方法和装置
WO2019001463A1 (zh) 数据处理方法及装置
CN109460514B (zh) 用于推送信息的方法和装置
CN106873799B (zh) 输入方法和装置
CN108197652B (zh) 用于生成信息的方法和装置
WO2020156389A1 (zh) 信息推送方法和装置
CN108520470B (zh) 用于生成用户属性信息的方法和装置
US20140095308A1 (en) Advertisement distribution apparatus and advertisement distribution method
WO2018149337A1 (zh) 一种信息投放方法、装置及服务器
US11275994B2 (en) Unstructured key definitions for optimal performance
EP3352121A1 (en) Content delivery method and device
JP7504192B2 (ja) 画像を検索するための方法及び装置
CN112182370A (zh) 物品类目信息推送方法、装置、电子设备和介质
CN112783468A (zh) 目标对象的排序方法和排序装置
US10474688B2 (en) System and method to recommend a bundle of items based on item/user tagging and co-install graph
CN109886300A (zh) 一种用户聚类方法、装置及设备
CN113793160B (zh) 投放数据处理方法、装置、设备及存储介质
CN110852078A (zh) 生成标题的方法和装置
CN113362097B (zh) 一种用户确定方法和装置
CN113780827A (zh) 一种物品筛选方法、装置、电子设备及计算机可读介质
CN109408716B (zh) 用于推送信息的方法和设备
CN112884538A (zh) 物品推荐的方法和装置
JP6570501B2 (ja) 情報提供装置、情報提供方法、および情報提供プログラム
CN111897910A (zh) 信息推送方法和装置

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: 18824754

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 18824754

Country of ref document: EP

Kind code of ref document: A1