CN110033098A - Online GBDT model learning method and device - Google Patents
Online GBDT model learning method and device Download PDFInfo
- Publication number
- CN110033098A CN110033098A CN201910243086.3A CN201910243086A CN110033098A CN 110033098 A CN110033098 A CN 110033098A CN 201910243086 A CN201910243086 A CN 201910243086A CN 110033098 A CN110033098 A CN 110033098A
- Authority
- CN
- China
- Prior art keywords
- decision tree
- sample data
- model
- prediction model
- learning
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N20/00—Machine learning
Abstract
The disclosure provides a kind of online GBDT prediction model learning method, comprising: obtains the sample data set for learning GBDT prediction model；And model learning is carried out based on the model parameter of at least one decision tree in GBDT prediction model using the sample data set, to create new decision tree and update the model parameter of at least one decision tree.Further, it is also possible to remove partial decision tree from least one decision tree, and treated that GBDT prediction model to create new decision tree and updates the model parameter of decision tree based on removing.Using this method, it can efficiently realize that GBDT prediction model learns.Further, it is also possible to which each decision tree for GBDT prediction model assigns weight factor, model prediction accuracy is thus improved.
Description
Technical field
The disclosure is usually directed to field of computer technology, more particularly, to the method for on-line study GBDT model
And device.
Background technique
With the development of artificial intelligence, in internet, big data data mining and machine learning work become more next
It is more important, many models for prediction are also therefore produced, for handling data to be predicted.It is calculated in a variety of machine learning
In method, GBDT (Gradient Boosting Deision Tree, gradient promote decision tree) algorithm is due to its excellent study
Performance has been more and more widely used.GBDT algorithm is a kind of machine learning skill for the tasks such as return, classify, sort
Art, by obtaining strong prediction model in conjunction with multiple weak learners (usually decision tree).
Fig. 1 shows the example of the GBDT prediction model an of routine, and the GBDT prediction model is by T decision tree g1(x)~gT
(x) it constitutes, and the GBDT prediction model can indicate are as follows:New sample data set is being used every time
When training GBDT prediction model, to need to all decision tree g in the 1st to the T decision tree of numberi(x) it is all trained,
So that it is computationally intensive, and it is more to consume resource.
Summary of the invention
Using a problem brought by traditional GBDT model, not by the obtained GBDT model of same sample training
It can reflect the variation of application scenarios, condition and data in time.For example, there is commodity to update for operation system,
The possibility of undercarriage, even and same user, due to the variation in time (such as season), place, for user different
Time or place and the identical purchase inquiry request issued, practical desired end article also can be different, but if platform begins
Inquiry response is provided as a result, the prediction result then provided is then difficult to embody difference using same prediction model eventually.
The sample data of update is generallyd use in existing scheme to relearn whole decision trees of GBDT prediction model.But
It can be found that it is very big to normally result in calculation amount in this way, system burden is increased, and still cannot accurately reflect that the time becomes
The feature of change.
The present invention proposes a kind of improved GBDT prediction model learning method, by obtaining the sample changed over time online
Data, and GBDT prediction mould is carried out using the new decision tree of the sample data training obtained online GBDT prediction model is added
Type updates, and GBDT prediction model can be made to update rapider, and feature can be followed to change with time, so that
The prediction result of GBDT prediction model is more accurate.
According to one aspect of the disclosure, a kind of method for on-line study GBDT prediction model is provided, it is described
GBDT prediction model includes at least one decision tree, which comprises obtains the sample number for learning GBDT prediction model
According to collection；And model learning is carried out based on the model parameter of at least one decision tree using the sample data set, with
It creates new decision tree and updates the model parameter of at least one decision tree.
Optionally, in an example of above-mentioned aspect, the method can also include: at least one described decision tree
It is removed processing, and model is carried out based on the model parameter of at least one decision tree using the sample data set
Study includes: to use the sample data to create new decision tree and update the model parameter of at least one decision tree
Collection to carry out model learning based on the model parameter by removal treated at least one decision tree, to create new decision
It sets and updates the model parameter by removal treated at least one decision tree.
Optionally, in an example of above-mentioned aspect, being removed processing at least one described decision tree be can wrap
It includes: removing the earliest predetermined number decision tree of creation time from least one described decision tree.
Optionally, in an example of above-mentioned aspect, each decision tree in the GBDT prediction model is endowed power
Repeated factor, the value of the weight factor of each decision tree and the creation time of the decision tree or the sample for creating the decision tree
The generation time of notebook data is inversely.
Optionally, in an example of above-mentioned aspect, the weight factor of each decision tree is the wound with the decision tree
Build the time decay factor for generating time decaying of time or the sample data for creating the decision tree.
Optionally, in an example of above-mentioned aspect, each decision tree in the GBDT prediction model is endowed power
The weight factor of repeated factor and each decision tree be arranged to the decision tree decision tree number increase and list
Adjust decline, wherein the decision tree number of each decision tree is the creation time serial number based on the decision tree.
Optionally, in an example of above-mentioned aspect, at least one described decision tree be removed processing include: from
The decision tree that weight factor is less than predetermined threshold is removed at least one described decision tree.
Optionally, in an example of above-mentioned aspect, acquisition can for learning the sample data set of GBDT prediction model
To include: the sample data obtained in predetermined time interval, as the sample data set for learning GBDT prediction model；Or
The sample data for obtaining predetermined amount of data, as the sample data set for learning GBDT prediction model.
According to another aspect of the present disclosure, a kind of device for on-line study GBDT prediction model, the GBDT are provided
Prediction model includes at least one decision tree, and described device includes: sample data acquiring unit, is configured as obtaining for learning
The sample data set of GBDT prediction model；And model learning unit, the sample data set is configured with to be based on
The model parameter for stating at least one decision tree carries out model learning, to create new decision tree and update at least one described decision
The model parameter of tree.
Optionally, in an example of above-mentioned aspect, described device can also include: decision tree removal unit, be matched
Be set to at least one described decision tree be removed processing and the model learning unit be configured as: use the sample
Notebook data collection to carry out model learning based on the model parameter by removal treated at least one decision tree, with creation
New decision tree and update the model parameter by removal treated at least one decision tree.
Optionally, in an example of above-mentioned aspect, the decision tree removal unit is configured as: from described at least one
The earliest predetermined number decision tree of creation time is removed in a decision tree.
Optionally, in an example of above-mentioned aspect, each decision tree in the GBDT prediction model is endowed power
Repeated factor, the value of the weight factor of each decision tree and the creation time of the decision tree or the sample for creating the decision tree
The generation time of notebook data is inversely.
Optionally, in an example of above-mentioned aspect, the weight factor of each decision tree is the wound with the decision tree
Build the time decay factor for generating time decaying of time or the sample data for creating the decision tree.
Optionally, in an example of above-mentioned aspect, each decision tree in the GBDT prediction model is endowed power
The weight factor of repeated factor and each decision tree be arranged to the decision tree decision tree number increase and list
Adjust decline, wherein the decision tree number of each decision tree is the creation time serial number based on the decision tree.
Optionally, in an example of above-mentioned aspect, the decision tree removal unit is configured as: from described at least one
The decision tree that weight factor is less than predetermined threshold is removed in a decision tree.
Optionally, in an example of above-mentioned aspect, the sample data acquiring unit is configured as: obtaining pre- timing
Between interval in sample data, as the sample data set for learning GBDT prediction model；Or obtain predetermined amount of data
Sample data, as the sample data set for learning GBDT prediction model.
According to another aspect of the present disclosure, a kind of calculating equipment is provided, comprising: at least one processor, and with it is described
The memory of at least one processor coupling, the memory store instruction, when described instruction is by least one described processor
When execution, so that at least one described processor executes the method for being used for on-line study GBDT prediction model as described above.
According to another aspect of the present disclosure, a kind of non-transitory machinable medium is provided, is stored with executable
Instruction, described instruction make the machine execute the side for being used for on-line study GBDT prediction model as described above upon being performed
Method.
Detailed description of the invention
By referring to following attached drawing, may be implemented to further understand the nature and advantages of present disclosure.?
In attached drawing, similar assembly or feature can have identical appended drawing reference.
Fig. 1 shows the GBDT modular concept schematic diagram of the prior art；
Fig. 2 shows the block diagrams of the system according to an embodiment of the present disclosure for on-line study GBDT prediction model；
Fig. 3 shows the flow chart of the method according to an embodiment of the present disclosure for on-line study GBDT prediction model；
Fig. 4 shows on-line study GBDT modular concept schematic diagram according to an embodiment of the present disclosure；
Fig. 5 shows the block diagram of the device according to an embodiment of the present disclosure for on-line study GBDT prediction model；
Fig. 6 shows the box of the calculating equipment according to an embodiment of the present disclosure for on-line study GBDT prediction model
Figure.
Specific embodiment
Theme described herein is discussed referring now to example embodiment.It should be understood that discussing these embodiments only
It is in order to enable those skilled in the art can better understand that being not to claim to realize theme described herein
Protection scope, applicability or the exemplary limitation illustrated in book.It can be in the protection scope for not departing from present disclosure
In the case of, the function and arrangement of the element discussed are changed.Each example can according to need, omit, substitute or
Add various processes or component.For example, described method can be executed according to described order in a different order, with
And each step can be added, omits or combine.In addition, feature described in relatively some examples is in other examples
It can be combined.
As used in this article, term " includes " and its modification indicate open term, are meant that " including but not limited to ".
Term "based" indicates " being based at least partially on ".Term " one embodiment " and " embodiment " expression " at least one implementation
Example ".Term " another embodiment " expression " at least one other embodiment ".Term " first ", " second " etc. may refer to not
Same or identical object.Here may include other definition, either specific or implicit.Unless bright in context
It really indicates, otherwise the definition of a term is consistent throughout the specification.
It is described in detail below in conjunction with attached drawing according to an embodiment of the present disclosure for on-line study GBDT prediction model
Method and device.
Fig. 2 shows the systems according to an embodiment of the present disclosure for on-line study GBDT prediction model (hereinafter referred to as
For on-time model learning system) 10 block diagram.
As shown in Fig. 2, on-time model learning system 10 includes at least one sample data occurrence of equipment 100, on-time model
Learning device 200 and GBDT prediction model 300.
At least one sample data occurrence of equipment 100 is configurable to generate sample data, and sample data set is consequently formed.For
Explanation is simple, illustrate only a sample data occurrence of equipment 100 in Fig. 2.Sample data occurrence of equipment 100 for example may be used
To be operation system equipment (hereinafter referred to as operation system).Business is run in operation system, thus generates business datum.
In addition, operation system 100 can also include sample database 110, for storing sample data generated.In the disclosure,
The business datum for example may include user data, commodity data, business transaction data etc..Correspondingly, sample database 110
It may include at least one merchandising database, at least one customer data base and at least one business transaction database.In quotient
The entire service data that can be traded in the operation system equipment are stored in product database, these data include related every quotient
The information such as one or more attributes of product, mark.It can be appreciated that with various factors such as the variation of time in season and user preferences
Influence, be available in the commodity traded in operation system and often change, and these information changes are stored in commodity data
In library.
The user information for accessing operation system is stored in customer data base, these information both include to be registered to business
Commodity transaction is realized in system or not yet carries out the information of the user of commodity transaction, is also realized comprising temporary visit operation system
Commodity transaction or not yet carry out commodity transaction user information.For chartered user, then the user information packet
Terminal iidentification (the example that the identity and its attribute information (such as age, gender, occupation etc.) and user for including the user use
Such as MAC Address, IP address) etc. characteristic informations.For the user of temporary visit, then the user information includes that the user uses
The characteristic informations such as terminal iidentification.Relevant to the commodity transaction actually occurred the Transaction Information of business transaction database purchase,
User, user query request, the end article of real trade, time of origin including transaction etc..
On-time model learning device 200 is configured as communicating at least one sample data occurrence of equipment 100, with from sample
Data transmitting equipment 100 obtains sample data set.After obtaining sample data set, on-time model learning device 200 is using being obtained
The sample data set taken based on the model parameter of each decision tree in GBDT prediction model 300 carries out model learning, with
Obtain updated GBDT prediction model.Operations and functions about on-time model learning device 200 will be below in reference to attached drawing
It is described.
GBDT prediction model 300 includes at least one decision tree.In one example, GBDT prediction model 300 for example may be used
To be stored in GBDT prediction model database.When carrying out on-time model study, on-time model learning device 200 is pre- from GBDT
It surveys in model database and obtains current GBDT prediction model, be then based on the model of each decision tree of current GBDT prediction model
Parameter carries out model learning.After on-time model learning device 200 completes model learning, updated GBDT can be predicted into mould
Type is stored into GBDT prediction model database.
Fig. 3 shows the flow chart of the method according to an embodiment of the present disclosure for on-line study GBDT prediction model.
As shown in figure 3, obtaining the sample data set for learning GBDT prediction model in block 310.The sample data set
In each sample data include the characteristic for indicating sample characteristics and the flag data for identifying sample attribute (such as
Labeled as positive sample or negative sample).For example, can be obtained by way of obtaining online for learning GBDT prediction model
Sample data set.
In an example of the disclosure, the sample data set obtained for learning GBDT prediction model may include: to obtain
The sample data in predetermined time interval is taken, as the sample data set for learning GBDT prediction model.Here, show at one
Example in, predetermined time interval for example can be rule of thumb or concrete application scene setting time interval, for example, 1 is small
When, 10 minutes etc..In another example, the predetermined time interval can be set dynamically according to service traffics size.For example,
Predetermined time interval can be set by a timer.For example, can start to obtain data when timer starts, and
Stop obtaining data in timer expiration.
In another example of the disclosure, the sample data set obtained for learning GBDT prediction model may include: to obtain
The sample data for taking predetermined amount of data, as the sample data set for learning GBDT prediction model.Equally, predetermined amount of data example
It such as can be the data volume size set according to the processing capacity or concrete application scene of on-time model learning device, for example,
100M etc..
After as above obtaining sample data set, in block 320, at least one decision tree in GBDT prediction model is gone
Except processing.Here, at least one decision tree refers to all decision trees for including in GBDT prediction model.One in the disclosure is shown
In example, being removed processing at least one described decision tree may include: to remove to create from least one described decision tree
Time earliest predetermined number decision tree.Due to each decision tree in GBDT prediction model be sequence train, and
When carrying out the training of each decision tree, need using all model parameters in preceding decision tree, so that GBDT prediction model
In the creation time of each decision tree there are sequencings.When carrying out decision tree removal processing, from earliest created decision
Tree starts, and removes predetermined number decision tree.The predetermined number for example can be 1 or other suitable numerical value.
After being removed processing at least one decision tree, in block 330, acquired sample data set and base are used
Carry out model learning in the model parameter by removal treated at least one decision tree, with create new decision tree and
Update the model parameter by removal treated at least one decision tree.
For example, can use acquired sample data set [y (i), x (i)] to learn new decision tree gT+1(x), i.e.,Wherein, L is loss function, and F is the space of all trees, and Gi-1(x)
It is by current the 2nd to i-1 decision tree g having determined2(x)~gT(x) model constituted is (assuming that eliminate decision tree g1
(x)), 1≤i≤T.As indicated in the formula, GBDT model algorithm is that new decision is determined by empirical loss minimization
Set gT+1(x).Specific to the selection of loss function L itself, various schemes well known in the prior art can be used, such as square
Loss function, 0-1 loss function, logarithm loss function etc..In addition, carrying out model using sample data set [y (i), x (i)]
Study, decision tree g2(x)~gT(x) model parameter also can be adjusted correspondingly, to obtain updated decision tree g'2(x)~
g'T(x).Each decision tree in GBDT model substantially envisages the change that with the time and bring sample may occur as a result,
Change.
GBDT prediction model and before study according to an embodiment of the present disclosure are illustrated with an example below
GBDT prediction model after habit.
For example, it is assumed that GBDT prediction model is by T decision tree g before carrying out model learning1(x)~gT(x) it constitutes, then
Before carrying out model learning, which can be indicated are as follows:
When before carrying out model learning, if removing 1 decision tree in decision tree removal processing, that is, removal creation
Time earliest decision tree g1(x), then by decision tree removal, treated that GBDT prediction model includes T-1 decision tree g2(x)
~gT(x).When carrying out model learning, using acquired sample data set, it is based on T-1 decision tree g2(x)~gT(x)
Model parameter creates new decision tree gT+1(x), T-1 decision tree g while is adaptively adjusted2(x)~gT(x) model ginseng
Number, that is, T-1 decision tree g' after being adjusted2(x)~g'T(x), thus the obtained GBDT after model learning is pre-
Surveying model includes T decision tree g'2(x)~g'T(x) and gT+1(x), that is,
In addition, it is contemplated that the influence that decision tree corresponding to the sample data set that different time obtains exports actual prediction
Be it is different, time closer sample data set and corresponding decision tree more have reference significance.In an example of the disclosure
In, each decision tree in the GBDT prediction model can also be endowed weight factor, for example, being directed to each decision tree gi
(x), corresponding weight factor Y is assignedi(x).Here, each decision tree gi(x) weight factor Yi(x) value and the decision
The generation time of the creation time of tree or the sample data for creating the decision tree is inversely.Correspondingly, acquired
The GBDT prediction model after model learningFig. 4 is shown according to this
The on-line study GBDT modular concept schematic diagram of disclosed embodiment.
For example, can be by each decision tree gi(x) weight factor Yi(x) it is set as the creation time with the decision tree
Or the time decay factor for generating time decaying of the sample data for creating the decision tree.In one example, for example,
It can be by each decision tree gi(x) weight factor Yi(x) it is set as Yi(T-ti), wherein T is current time, tiIt is the decision
The generation time or acquisition time of the creation time of tree or the sample data for creating the decision tree, and Yi(T-ti) with
T-tiIncrease and reduce.For example, can be by Yi(x) it is set asWherein, γ is the positive number less than 1.
In an example of the disclosure, each decision tree can have decision tree number, and each decision tree
Decision tree number be creation time serial number based on the decision tree.For example, the decision tree of earliest created decision tree
Number is 1, and the decision tree number of the decision tree then created successively increases.In this case, the weight of each decision tree because
Son can be set to the decision tree decision tree number increase and monotonic decreasing.
In an example of the disclosure, each decision tree have weight factor in the case where, to it is described at least one
It may include: to remove weight factor from least one decision tree of GBDT prediction model to be less than in advance that decision tree, which is removed processing,
Determine the decision tree of threshold value.For example, each decision tree in GBDT prediction model respective weight under current time can be calculated
Then the calculated each weight factor of institute is compared by the factor with predetermined threshold, and weight factor is less than predetermined threshold
Decision tree removed from GBDT prediction model, be achieved in decision tree removal process.For example, being determined as kth decision tree
Weight factor when being less than preset threshold, since this means that influence very little of the 1~k decision tree to GBDT decision model,
It is possible thereby to remove the 1~k decision tree.
It is described above with reference to Fig. 3 to according to the method for on-line study GBDT prediction model of the disclosure.Here
It is noted that shown in Fig. 3 being only a reality according to the method for on-line study GBDT prediction model of the disclosure
Example is applied, it, can also be to the method for being used for on-line study GBDT prediction model shown in Fig. 3 in the other embodiments of the disclosure
It modifies.
For example, the method for on-line study GBDT prediction model can not include decision in another embodiment of the present disclosure
Set removal process, that is, do not include the operation of block 320.Correspondingly, in block 330, using acquired sample data set and based on described
The model parameter of at least one decision tree in GBDT prediction model carries out model learning, to create new decision tree and update institute
State the model parameter of at least one decision tree.Obtained GBDT prediction model isOr
Person
Using the method for on-line study GBDT prediction model according to the disclosure, by obtaining subsequent update online
Sample data set, and new decision tree is created based on original GBDT prediction model using acquired sample data set, together
When the model parameter of each decision tree of original GBDT prediction model is adaptively adjusted, thus use created it is new certainly
Plan tree and each decision tree after being adjusted construct the GBDT prediction model after model learning, so that online
The GBDT prediction model for learning to obtain can reflect that in real time data sample changes with time, simultaneously because being only to create newly
Decision tree, rather than entire GBDT model is trained, so as to reduce the calculation amount of GBDT prediction model study, into
And GBDT prediction model learning efficiency is improved, and reduce resource needed for model learning.
In addition, using the method for on-line study GBDT prediction model according to the disclosure, by for by model
Each decision tree in GBDT prediction model after habit assigns weight factor, which is arranged to the wound with the decision tree
Generation time of time or the sample data for creating the decision tree is built inversely, it is possible thereby to make based on relatively early
The influence that the decision tree that the sample data set that time obtains is created exports actual prediction is smaller, to improve model prediction
Accuracy rate.
In addition, using the method for on-line study GBDT prediction model according to the disclosure, by pre- from original GBDT
Surveying removal creation time in each decision tree of model, decision tree or weight factor are lower than the decision tree of predetermined threshold earlier,
Quantity is handled so as to reduce decision tree when model learning, thus reduces the calculation amount of GBDT prediction model study, in turn
Improve GBDT prediction model learning efficiency.Further, since decision tree or weight factor are lower than predetermined threshold to creation time earlier
Influence of the decision tree of value for model prediction is smaller, and thereby may be ensured that will not be pre- to the GBDT obtained after above-mentioned processing
The predictablity rate for surveying model generates large effect.
Fig. 5 shows the device according to an embodiment of the present disclosure for on-line study GBDT prediction model (hereinafter referred to as
For on-time model learning device) 200 block diagram.As shown in figure 5, on-time model learning device 200 is obtained including sample data set
Take unit 210, decision tree removal unit 220 and on-time model unit 230.
Sample data acquiring unit 210 is configured as obtaining the sample data set for learning GBDT prediction model.For example,
In one example, sample data acquiring unit 210 is configured as: obtain predetermined time interval in sample data, as with
In the sample data set of study GBDT prediction model.Alternatively, in another example, sample data acquiring unit 210 is configured as:
The sample data for obtaining predetermined amount of data, as the sample data set for learning GBDT prediction model.Sample data obtains single
The operation of member 210 can be with reference to the operation above with reference to Fig. 3 block 310 described.
Decision tree removal unit 220 is configured as being removed at least one decision tree processing.Decision tree removal unit
220 operation can be with reference to the operation above with reference to Fig. 3 block 320 described.
Model learning unit 230 be configured with the sample data set be based on it is described by removal treated extremely
The model parameter of a few decision tree carries out model learning, to create new decision tree and update that described treated by removal
The model parameter of at least one decision tree.The operation of model learning unit 230 can be with reference to the block 330 described above with reference to Fig. 3
Operation.
In an example of the disclosure, decision tree removal unit 220 be can be configured as: from least one decision tree
Remove the earliest predetermined number decision tree of creation time.
In another example of the disclosure, each decision tree in GBDT prediction model is endowed weight factor, each to determine
The generation of the value of the weight factor of plan tree and the creation time of the decision tree or the sample data for creating the decision tree
Time is inversely.For example, the weight factor of each decision tree is with the creation time of the decision tree or for creating
The time decay factor for generating time decaying of the sample data of the decision tree.Alternatively, each decision in GBDT prediction model
Tree can have decision tree number, and the decision tree number of each decision tree is that the creation time sequence based on the decision tree is compiled
Number.The weight factor of each decision tree be arranged to the decision tree decision tree number increase and monotonic decreasing.?
In this case, decision tree removal unit 220 can be additionally configured to: removing weight factor from least one decision tree and is less than
The decision tree of predetermined threshold.
In addition, on-time model learning device 200 may include that timer unit (does not show in another example of the disclosure
Out).The timer unit is configured as setting predetermined time interval.
In another example of the disclosure, on-time model learning device 200 may include that weight factor determination unit (is not shown
Out).The weight factor determination unit is configured to determine that each decision tree in the weight factor at current time.Then, decision
Weight factor of the removal unit 220 based on each decision tree at least one decision tree determined is set, from least one
The decision tree that weight factor is less than predetermined threshold is removed in decision tree.
In another embodiment of the present disclosure, on-time model learning device 200 can not also include decision tree removal unit
220.Correspondingly, model learning unit 230 is configured with the sample data set and comes based on the GBDT prediction model
The model parameter of at least one decision tree carries out model learning, to create new decision tree and update at least one described decision tree
Model parameter.
Above with reference to Fig. 1 to Fig. 5, to according to the method and device for on-line study GBDT prediction model of the disclosure
Embodiment is described.The device for on-line study GBDT prediction model above can use hardware realization, can also be with
It is realized using the combination of software or hardware and software.
Fig. 6 shows the calculating equipment 600 according to an embodiment of the present disclosure for on-line study GBDT prediction model
Hardware structure diagram.As shown in fig. 6, calculating equipment 600 may include at least one processor 610, memory 620,630 and of memory
Communication interface 640, and at least one processor 610, memory 620, memory 630 and communication interface 640 connect via bus 660
It is connected together.At least one processor 610 executes at least one computer-readable instruction for storing or encoding in memory
(that is, above-mentioned element realized in a software form).
In one embodiment, computer executable instructions are stored in memory, make at least one when implemented
Processor 610: the sample data set for learning GBDT prediction model is obtained；And using the sample data set to be based on
The model parameter for stating at least one decision tree carries out model learning, to create new decision tree and update at least one described decision
The model parameter of tree.
It should be understood that the computer executable instructions stored in memory make at least one processor when implemented
610 carry out the above various operations and functions described in conjunction with Fig. 1-5 in each embodiment of the disclosure.
In the disclosure, calculating equipment 600 can include but is not limited to: personal computer, server computer, work
It stands, desktop computer, laptop computer, notebook computer, mobile computing device, smart phone, tablet computer, bee
Cellular telephone, personal digital assistant (PDA), hand-held device, messaging devices, wearable calculating equipment, consumer-elcetronics devices etc.
Deng.
According to one embodiment, a kind of program product of such as non-transitory machine readable media is provided.Non-transitory
Machine readable media can have instruction (that is, above-mentioned element realized in a software form), which when executed by a machine, makes
It obtains machine and executes the above various operations and functions described in conjunction with Fig. 1-5 in each embodiment of the disclosure.Specifically, Ke Yiti
For being furnished with the system or device of readable storage medium storing program for executing, store on the readable storage medium storing program for executing any in realization above-described embodiment
The software program code of the function of embodiment, and read and execute the computer of the system or device or processor and be stored in
Instruction in the readable storage medium storing program for executing.
According to one embodiment, a kind of program product of such as non-transitory machine readable media is provided.Non-transitory
Machine readable media can have instruction (that is, above-mentioned element realized in a software form), which when executed by a machine, makes
It obtains machine and executes the above various operations and functions described in conjunction with Fig. 1-5 in each embodiment of the disclosure.Specifically, Ke Yiti
For being furnished with the system or device of readable storage medium storing program for executing, store on the readable storage medium storing program for executing any in realization above-described embodiment
The software program code of the function of embodiment, and read and execute the computer of the system or device or processor and be stored in
Instruction in the readable storage medium storing program for executing.
In this case, it is real that any one of above-described embodiment can be achieved in the program code itself read from readable medium
The function of example is applied, therefore the readable storage medium storing program for executing of machine readable code and storage machine readable code constitutes of the invention one
Point.
The embodiment of readable storage medium storing program for executing include floppy disk, hard disk, magneto-optic disk, CD (such as CD-ROM, CD-R, CD-RW,
DVD-ROM, DVD-RAM, DVD-RW, DVD-RW), tape, non-volatile memory card and ROM.It selectively, can be by communication network
Network download program code from server computer or on cloud.
It will be appreciated by those skilled in the art that each embodiment disclosed above can be in the situation without departing from invention essence
Under make various changes and modifications.Therefore, protection scope of the present invention should be defined by the appended claims.
It should be noted that step and unit not all in above-mentioned each process and each system construction drawing is all necessary
, certain step or units can be ignored according to the actual needs.Each step execution sequence be not it is fixed, can be according to need
It is determined.Apparatus structure described in the various embodiments described above can be physical structure, be also possible to logical construction, that is, have
A little units may be realized by same physical entity, be realized alternatively, some units may divide by multiple physical entities, alternatively, can be with
It is realized jointly by certain components in multiple autonomous devices.
In the above various embodiments, hardware cell or module mechanically or can be realized electrically.For example, one
Hardware cell, module or processor may include permanent dedicated circuit or logic (such as special processor, FPGA or
ASIC) corresponding operating is completed.Hardware cell or processor can also include programmable logic or circuit (such as general processor or
Other programmable processors), interim setting can be carried out by software to complete corresponding operating.Concrete implementation mode is (mechanical
Mode or dedicated permanent circuit or the circuit being temporarily arranged) it can be determined based on cost and temporal consideration.
The specific embodiment illustrated above in conjunction with attached drawing describes exemplary embodiment, it is not intended that may be implemented
Or fall into all embodiments of the protection scope of claims." exemplary " meaning of the term used in entire this specification
Taste " be used as example, example or illustration ", be not meant to than other embodiments " preferably " or " there is advantage ".For offer pair
The purpose of the understanding of described technology, specific embodiment include detail.However, it is possible in these no details
In the case of implement these technologies.In some instances, public in order to avoid the concept to described embodiment causes indigestion
The construction and device known is shown in block diagram form.
The foregoing description of present disclosure is provided so that any those of ordinary skill in this field can be realized or make
Use present disclosure.To those skilled in the art, the various modifications carried out to present disclosure are apparent
, also, can also answer generic principles defined herein in the case where not departing from the protection scope of present disclosure
For other modifications.Therefore, present disclosure is not limited to examples described herein and design, but disclosed herein with meeting
Principle and novel features widest scope it is consistent.
Claims (18)
1. a kind of method for on-line study GBDT prediction model, the GBDT prediction model includes at least one decision tree,
The described method includes:
Obtain the sample data set for learning GBDT prediction model；And
Model learning is carried out based on the model parameter of at least one decision tree using the sample data set, it is new to create
Decision tree and the model parameter for updating at least one decision tree.
2. the method as described in claim 1, further includes:
Processing is removed at least one described decision tree, and
Model learning is carried out based on the model parameter of at least one decision tree using the sample data set, it is new to create
The decision tree and model parameter for updating at least one decision tree includes:
It is carried out using the sample data set based on the model parameter by removal treated at least one decision tree
Model learning, to create new decision tree and update the model parameter by removal treated at least one decision tree.
3. method according to claim 2, wherein being removed processing at least one described decision tree includes:
The earliest predetermined number decision tree of creation time is removed from least one described decision tree.
4. method according to claim 2, wherein each decision tree in the GBDT prediction model be endowed weight because
Son, the value of the weight factor of each decision tree and the creation time of the decision tree or the sample number for creating the decision tree
According to the generation time inversely.
5. method as claimed in claim 4, wherein the weight factor of each decision tree is the creation time with the decision tree
Or the time decay factor for generating time decaying of the sample data for creating the decision tree.
6. method according to claim 2, wherein each decision tree in the GBDT prediction model be endowed weight because
The weight factor of son and each decision tree is arranged to the increase of the decision tree number of the decision tree and under dullness
Drop,
Wherein, the decision tree number of each decision tree is the creation time serial number based on the decision tree.
7. the method as described in any in claim 4 to 6, wherein be removed processing packet at least one described decision tree
It includes:
The decision tree that weight factor is less than predetermined threshold is removed from least one described decision tree.
8. the method for claim 1, wherein obtaining the sample data set for learning GBDT prediction model includes:
The sample data in predetermined time interval is obtained, as the sample data set for learning GBDT prediction model；Or
The sample data for obtaining predetermined amount of data, as the sample data set for learning GBDT prediction model.
9. a kind of device for on-line study GBDT prediction model, the GBDT prediction model includes at least one decision tree,
Described device includes:
Sample data acquiring unit is configured as obtaining the sample data set for learning GBDT prediction model；And
Model learning unit is configured with the sample data set and carrys out the model parameter based at least one decision tree
Model learning is carried out, to create new decision tree and update the model parameter of at least one decision tree.
10. device as claimed in claim 9, further includes:
Decision tree removal unit is configured as being removed processing at least one described decision tree, and
The model learning unit is configured as: be based on using the sample data set it is described by removal treated at least
The model parameter of one decision tree carries out model learning, with create new decision tree and update it is described by removal treated extremely
The model parameter of a few decision tree.
11. device as claimed in claim 10, wherein the decision tree removal unit is configured as:
The earliest predetermined number decision tree of creation time is removed from least one described decision tree.
12. device as claimed in claim 10, wherein each decision tree in the GBDT prediction model be endowed weight because
Son, the value of the weight factor of each decision tree and the creation time of the decision tree or the sample number for creating the decision tree
According to the generation time inversely.
13. device as claimed in claim 12, wherein when the weight factor of each decision tree is the creation with the decision tree
Between or sample data for creating the decision tree the time decay factor for generating time decaying.
14. device as claimed in claim 10, wherein each decision tree in the GBDT prediction model be endowed weight because
The weight factor of son and each decision tree is arranged to the increase of the decision tree number of the decision tree and under dullness
Drop,
Wherein, the decision tree number of each decision tree is the creation time serial number based on the decision tree.
15. the device as described in any in claim 12 to 14, wherein the decision tree removal unit is configured as:
The decision tree that weight factor is less than predetermined threshold is removed from least one described decision tree.
16. device as claimed in claim 9, wherein the sample data acquiring unit is configured as:
The sample data in predetermined time interval is obtained, as the sample data set for learning GBDT prediction model；Or
The sample data for obtaining predetermined amount of data, as the sample data set for learning GBDT prediction model.
17. a kind of calculating equipment, comprising:
At least one processor, and
The memory coupled at least one described processor, the memory store instruction, when described instruction by it is described at least
When one processor executes, so that at least one described processor executes the method as described in any in claims 1 to 8.
18. a kind of non-transitory machinable medium, is stored with executable instruction, described instruction makes upon being performed
The machine executes the method as described in any in claims 1 to 8.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910243086.3A CN110033098A (en) | 2019-03-28 | 2019-03-28 | Online GBDT model learning method and device |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910243086.3A CN110033098A (en) | 2019-03-28 | 2019-03-28 | Online GBDT model learning method and device |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN110033098A true CN110033098A (en) | 2019-07-19 |
Family
ID=67236817
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910243086.3A Pending CN110033098A (en) | 2019-03-28 | 2019-03-28 | Online GBDT model learning method and device |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN110033098A (en) |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106156809A (en) * | 2015-04-24 | 2016-11-23 | 阿里巴巴集团控股有限公司 | For updating the method and device of disaggregated model |
| US20170213280A1 (en) * | 2016-01-27 | 2017-07-27 | Huawei Technologies Co., Ltd. | System and method for prediction using synthetic features and gradient boosted decision tree |
| CN107330464A (en) * | 2017-06-30 | 2017-11-07 | 众安信息技术服务有限公司 | Data processing method and device |
| CN109472296A (en) * | 2018-10-17 | 2019-03-15 | 阿里巴巴集团控股有限公司 | A kind of model training method and device promoting decision tree based on gradient |
| US20190318202A1 (en) * | 2016-10-31 | 2019-10-17 | Tencent Technology (Shenzhen) Company Limited | Machine learning model training method and apparatus, server, and storage medium |
-
2019
- 2019-03-28 CN CN201910243086.3A patent/CN110033098A/en active Pending
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106156809A (en) * | 2015-04-24 | 2016-11-23 | 阿里巴巴集团控股有限公司 | For updating the method and device of disaggregated model |
| US20170213280A1 (en) * | 2016-01-27 | 2017-07-27 | Huawei Technologies Co., Ltd. | System and method for prediction using synthetic features and gradient boosted decision tree |
| US20190318202A1 (en) * | 2016-10-31 | 2019-10-17 | Tencent Technology (Shenzhen) Company Limited | Machine learning model training method and apparatus, server, and storage medium |
| CN107330464A (en) * | 2017-06-30 | 2017-11-07 | 众安信息技术服务有限公司 | Data processing method and device |
| CN109472296A (en) * | 2018-10-17 | 2019-03-15 | 阿里巴巴集团控股有限公司 | A kind of model training method and device promoting decision tree based on gradient |
Non-Patent Citations (1)
| Title |
|---|
| ZHIYUAN CHEN AND BING LIU: "Lifelong Machine Learning", MORGAN & CLAYPOOL PUBLISHERS, pages: 146 - 149 * |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110765117B (en) | Fraud identification method, fraud identification device, electronic equipment and computer readable storage medium | |
| CN106980623B (en) | Data model determination method and device | |
| CN108108821A (en) | Model training method and device | |
| CN108563710B (en) | Knowledge graph construction method and device and storage medium | |
| CN108399564B (en) | Credit scoring method and device | |
| CN106651213B (en) | Service order processing method and device | |
| CN113254549B (en) | Character relation mining model training method, character relation mining method and device | |
| US11423307B2 (en) | Taxonomy construction via graph-based cross-domain knowledge transfer | |
| CN112214499B (en) | Graph data processing method and device, computer equipment and storage medium | |
| CN111967971A (en) | Bank client data processing method and device | |
| CN109685805A (en) | A kind of image partition method and device | |
| CN114462582A (en) | Data processing method, device and equipment based on convolutional neural network model | |
| CN109992679A (en) | A kind of classification method and device of multi-medium data | |
| CN111667018B (en) | Object clustering method and device, computer readable medium and electronic equipment | |
| CN111984842B (en) | Bank customer data processing method and device | |
| CN110033098A (en) | Online GBDT model learning method and device | |
| CN115238194A (en) | Book recommendation method, computing device and computer storage medium | |
| CN110727705B (en) | Information recommendation method and device, electronic equipment and computer-readable storage medium | |
| CN114418120A (en) | Data processing method, device, equipment and storage medium of federal tree model | |
| CN113159926A (en) | Loan transaction repayment date determination method and device | |
| CN109840535B (en) | Method and device for realizing terrain classification | |
| CN111400413A (en) | Method and system for determining category of knowledge points in knowledge base | |
| CN111931035B (en) | Service recommendation method, device and equipment | |
| CN109712012B (en) | Social network partitioning method, device, equipment and storage medium | |
| CN114579735A (en) | Information point state determining method and device, electronic equipment and readable storage medium |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| TA01 | Transfer of patent application right |
Effective date of registration: 20200929 Address after: English genus Applicant after: Innovative advanced technology Co.,Ltd. Address before: English genus Applicant before: Advanced innovation technology Co.,Ltd. Effective date of registration: 20200929 Address after: English genus Applicant after: Advanced innovation technology Co.,Ltd. Address before: A four-storey 847 mailbox in Grand Cayman Capital Building, British Cayman Islands Applicant before: Alibaba Group Holding Ltd. |
|
| TA01 | Transfer of patent application right |