CN108921300A - The method and apparatus for executing automaton study - Google Patents
The method and apparatus for executing automaton study Download PDFInfo
- Publication number
- CN108921300A CN108921300A CN201810641700.7A CN201810641700A CN108921300A CN 108921300 A CN108921300 A CN 108921300A CN 201810641700 A CN201810641700 A CN 201810641700A CN 108921300 A CN108921300 A CN 108921300A
- Authority
- CN
- China
- Prior art keywords
- learning model
- machine learning
- model group
- prediction data
- prediction
- 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
Landscapes
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
Abstract
A kind of method and apparatus of execution automaton study are provided, the method includes：The initial machine learning model group including at least one machine learning model is obtained by way of automaton study；Persistently obtain prediction data；It monitors and whether occurs the distribution variation beyond preset threshold in the prediction data persistently obtained；In the case where there is the distribution variation beyond threshold value, initial machine learning model group is automatically updated.
Description
Technical field
The application all things considered is related to artificial intelligence field, more particularly, be related to it is a kind of execution automaton study
Method and apparatus.
Background technique
With the appearance of mass data, artificial intelligence technology is rapidly developed, and machine learning is Artificial Intelligence Development to one
The inevitable outcome for determining the stage, is dedicated to the means by calculating, and valuable potential information is excavated from mass data.
In machine learning field, carry out training machine learning model often through empirical data is supplied to machine learning algorithm
To determine the ideal parameters for being constituted machine learning model, and trained machine learning model can be applied in face of newly pre-
Corresponding prediction result is provided when measured data.However, many work are (for example, feature is located in advance involved in machine-learning process
Reason and selection, model algorithm selection, hyper parameter adjustment etc.) it had often both needed to have computer (especially machine learning) profession and had known
Know, it is also desirable to have specific business experience relevant to prediction scene, therefore, it is necessary to expend a large amount of human cost.In order to mention
High machine learning efficiency proposes automaton study (AutoML) technology in recent years, realizes from data prediction to ginseng
Therefore a series of automation of processes of number selection greatly reduces machine learning threshold and reduces for machine learning
Human cost.
However, AutoML technology is assumed always on carrying out automaton study current events reality for machine learning model training
Data be it is independent identically distributed, in other words, it is assumed that the distribution of data for machine learning model training is constant
(that is, static data), and the case where do not consider data distribution state change.But in fact, in the reality using machine learning
In the scene of border, over time, the distribution of data can usually change, and if still according to previously training
Machine learning model have occurred and that the new data of variation executes prediction for data distribution state, then be difficult to obtain accurate
Prediction result.In consideration of it, needing more perfect automaton study technology.
Summary of the invention
According to the application exemplary embodiment, a kind of method of execution automaton study is provided, the method can wrap
It includes：The initial machine learning model group including at least one machine learning model is obtained by way of automaton study；Continue
Obtain prediction data；It monitors and whether occurs the distribution variation beyond preset threshold in the prediction data persistently obtained；Going out
Now in the case where the distribution variation beyond threshold value, initial machine learning model group is automatically updated.
Optionally, in the initial machine learning model group, each machine learning model is provided for prediction data
Prediction result can be weighted prediction result of the summation to be directed to prediction data as the initial machine learning model group, and
And monitoring step may include：The distribution and/or initial machine learning model group for monitoring the prediction data persistently obtained are directed to
The prediction effect of the prediction data, to determine whether that the distribution beyond preset threshold changes.
Optionally, it in the case where there is the distribution variation beyond preset threshold, fixedly uses or adaptively
Initial machine learning model group is automatically updated using one of scheduled a variety of update modes.
Optionally, scheduled a variety of update modes may include following update mode：First update mode, re -training
New machine learning model group is to substitute initial machine learning model group；Second update mode adjusts initial machine learning model
The weight of each machine learning model in group；Or third update mode, the new machine learning model group of training simultaneously will be new
Machine learning model group combines to constitute updated machine learning model group with initial machine learning model group.
Optionally, it in the first update mode, is formed using based at least part prediction data and its legitimate reading
Training data, the new machine learning model group of re -training；It, can be according to initial machine learning model in the second update mode
Group is adjusting initial machine based on the prediction effect at least part prediction data and its observation data of legitimate reading formation
The weight of each machine learning model in device learning model group；In third update mode, using based at least part
The new machine learning model group of the training data training that prediction data and its legitimate reading are formed, also, by by new machine
Learning model group is combined with initial machine learning model group, and is being based at least one according to the machine learning model group after combination
The prediction effect in observation data for dividing prediction data and its legitimate reading to be formed determines in the machine learning model group after combination
The weight of each machine learning model constitute updated machine learning model group.
Optionally, in the first update mode, at least part prediction data may include all pre- of lasting acquisition
Measured data；In the second update mode and third update mode, at least part prediction data may include that distribution becomes
The whole prediction data changing the prediction data beyond preset threshold or persistently obtaining.
Optionally, initial machine learning model group is adaptively automatically updated using one of scheduled a variety of update modes
The step of may include：According at least to information self-adapting related with prediction data among scheduled a variety of update modes
Corresponding update mode is selected to automatically update initial machine learning model group.
Optionally, information related with prediction data may include the prediction that can be used in updating initial machine learning model group
The data volume of data and/or information relevant to the variation of the distribution of prediction data.Optionally, with the distribution shape of prediction data
State changes the week of the distribution variation of the distribution change rate that relevant information may include prediction data and/or prediction data
Phase property.
Optionally, the step of corresponding update mode is adaptive selected may include：Data volume be less than first threshold and
In the case that distribution change rate is greater than second threshold, third update mode is selected；It is less than first threshold in data volume and divides
In the case that cloth status variation rate is less than second threshold, the first update mode is selected；It is greater than first threshold and distribution in data volume
In the case that status variation rate is less than second threshold, the second update mode is selected；It is greater than first threshold and distribution shape in data volume
State variation selects the second update mode in periodic situation；It is greater than first threshold in data volume, distribution change rate is big
In second threshold, and distribution variation selects third update mode not in periodic situation.
Optionally, monitoring step may include：Obtain the prediction knot that initial machine learning model group is directed to the prediction data
Fruit；It is compared by the prediction result that will acquire with the legitimate reading of the prediction data, determines initial machine learning model
Group is directed to the prediction effect of the prediction data；According to determining prediction effect, it is determined whether occur exceeding the distribution shape of threshold value
State variation.
According to the application another exemplary embodiment, provide a kind of for executing the computer-readable of automaton study
Medium, wherein record has the computer program for executing method as described above on the computer-readable medium.
According to the application another exemplary embodiment, a kind of computing device of execution automaton study is provided, including
Storage unit and processor, wherein set of computer-executable instructions conjunction is stored in storage unit, when the computer is executable
When instruction set is executed by the processor, processor is promoted to execute method as described above.
According to the application another exemplary embodiment, a kind of device of execution automaton study, described device are provided
May include：Initial machine learning model group acquiring unit is configured as obtaining by way of automaton study including at least one
The initial machine learning model group of a machine learning model；Prediction data acquiring unit is configured as continuing to obtain prediction data；
Monitoring unit is configured as the distribution variation for whether occurring beyond preset threshold in the prediction data for monitoring lasting acquisition；
Updating unit is configured as automatically updating initial machine learning model in the case where there is the distribution variation beyond threshold value
Group.
Optionally, in the initial machine learning model group, each machine learning model is provided for prediction data
Prediction result can be weighted prediction result of the summation to be directed to prediction data as the initial machine learning model group, and
And the monitoring unit distribution that can monitor the prediction data of lasting acquisition and/or initial machine learning model group are for described
The prediction effect of prediction data, to determine whether that the distribution beyond preset threshold changes.
Optionally, in the case where there is the distribution variation beyond preset threshold, updating unit can be used fixedly
Or initial machine learning model group is adaptively automatically updated using one of scheduled a variety of update modes.
Optionally, scheduled a variety of update modes may include following update mode：First update mode, re -training
New machine learning model group is to substitute initial machine learning model group；Second update mode adjusts initial machine learning model
The weight of each machine learning model in group；Alternatively, third update mode, the new machine learning model group of training simultaneously will be new
Machine learning model group combines to constitute updated machine learning model group with initial machine learning model group.
Optionally, in the first update mode, updating unit is using based at least part prediction data and its really
As a result the training data formed, the new machine learning model group of re -training；In the second update mode, updating unit can basis
Prediction of the initial machine learning model group in the observation data based at least part prediction data and its legitimate reading formation
Effect adjusts the weight of each machine learning model in initial machine learning model group；In third update mode, update
The unit machine learning mould new using the training data training formed based at least part prediction data and its legitimate reading
Type group, also, by combining new machine learning model group with initial machine learning model group, and according to the machine after combination
Learning model group is determined based on the prediction effect at least part prediction data and its observation data of legitimate reading formation
The weight of each machine learning model in machine learning model group after combination constitutes updated machine learning model group.
Optionally, in the first update mode, at least part prediction data may include all pre- of lasting acquisition
Measured data；In the second update mode and third update mode, at least part prediction data may include that distribution becomes
The whole prediction data changing the prediction data beyond preset threshold or persistently obtaining.
Optionally, updating unit can be according at least to information self-adapting related with prediction data described scheduled a variety of
Corresponding update mode is selected among update mode to automatically update initial machine learning model group.
Optionally, information related with prediction data may include the prediction that can be used in updating initial machine learning model group
The data volume of data and/or information relevant to the variation of the distribution of prediction data.Optionally, with the distribution shape of prediction data
State changes the week of the distribution variation of the distribution change rate that relevant information may include prediction data and/or prediction data
Phase property.
Optionally, it in the case where data volume is less than first threshold and distribution change rate is greater than second threshold, updates
Third update mode may be selected in unit；The case where data volume is less than first threshold and distribution change rate is less than second threshold
Under, the first update mode may be selected in updating unit；It is greater than first threshold and distribution change rate less than the second threshold in data volume
In the case where value, the second update mode is may be selected in updating unit；It is greater than first threshold and distribution variation in week in data volume
In the case where phase property, the second update mode is may be selected in updating unit；It is greater than first threshold in data volume, distribution change rate is big
In second threshold, and distribution variation, not in periodic situation, third update mode may be selected in updating unit.
Optionally, monitoring unit can obtain the prediction result that initial machine learning model group is directed to the prediction data, lead to
The legitimate reading for crossing the prediction result and prediction data that will acquire is compared to determine initial machine learning model group needle
To the prediction effect of the prediction data, and according to determining prediction effect, it is determined whether occur exceeding the distribution shape of threshold value
State variation.
According to the method and apparatus for executing automaton study of the application exemplary embodiment, by robotics
The distribution variation of monitoring and forecasting data during habit, and the distribution beyond preset threshold occur in prediction data and become
Automatically updating initial machine learning model in the case where change, further perfect automaton at present learns technology, also, by
It, therefore, can in can constantly learn by automatically updating machine learning model group to the distribution of prediction data to change
More accurate prediction result is provided for new prediction data using updated machine learning model group.
Detailed description of the invention
From the detailed description with reference to the accompanying drawing to the embodiment of the present application, these and or other aspects of the application and
Advantage will become clearer and be easier to understand, wherein：
Fig. 1 is the block diagram for showing the device for executing automaton study according to the application exemplary embodiment；
Fig. 2 is the flow chart for showing the method for executing automaton study according to the application exemplary embodiment；
Fig. 3 is the schematic diagram for showing the method for executing automaton study according to the application exemplary embodiment.
Specific embodiment
In order to make those skilled in the art more fully understand the application, with reference to the accompanying drawings and detailed description to this Shen
Exemplary embodiment please is described in further detail.
Fig. 1 be show according to the application exemplary embodiment execute automaton study device 100 (hereinafter,
Be called for short automaton study device 100) block diagram.It may include initial machine learning model that automaton, which learns device 100,
Group acquiring unit 110, prediction data acquiring unit 120, monitoring unit 130 and updating unit 140.
Particularly, initial machine learning model group acquiring unit 110 can be obtained by way of automaton study including
The initial machine learning model group of at least one machine learning model.Here, automaton study mode can be such as
The combination of any automaton study mode of AutoML and different automaton study modes.In addition, it should be noted that,
Initial machine learning model group not only may include one or more machine learning models, but also work as initial machine learning model group packet
When including multiple machine learning models, the type of the multiple machine learning model can be identical or different (for example, can be according to identical
Or different algorithms apply identical or different characteristic processing etc.), also, each machine learning model can be it is any kind of
Machine learning model, for example, logistic regression (LR) model, support vector machines (SVM), gradient promote decision tree or depth nerve net
Network etc., but not limited to this.
Accoding to exemplary embodiment, initial machine learning model group acquiring unit 110 can by way of automaton study,
Initial machine learning model group is trained based on the training data obtained in advance.Here, training data can be learnt by initial machine
Other data capture unit (not shown) or other numbers included by model group acquiring unit 110, automaton study device 100
It obtains, or can be input by input unit certainly from external data source (for example, server, database etc.) according to acquisition device
Movement machine learning device 100.For example, the above-mentioned training data obtained in advance can be continuously defeated according to defeated mode is spread
Enter, or can input disposably or in batch, initial machine learning model group acquiring unit 110 is enabled to utilize these training
Data train initial machine learning model group.As an example, initial machine learning model group acquiring unit 110 can be used and be based on
The AutoML mode of Ensemble Learning (integrated study), trains initial machine to learn based on the training data of acquisition
Model group.Here, initial machine learning model group acquiring unit 110 can be based on greedy algorithm, each by continuous iteration adjustment
The weight of machine learning model is come final true with maximizing prediction effect of the initial machine learning model group on training dataset
Determine the weight of each machine learning model in initial machine learning model group.Accoding to exemplary embodiment, pass through robotics
The initial machine learning model group that habit mode obtains can be used for executing prediction for prediction data, and including multiple machine learning
The initial machine learning model group of model usually can provide better prediction effect than individual machine learning model.
The sustainable acquisition prediction data of prediction data acquiring unit 120.As an example, prediction data can be defeated according to spreading
Mode continually enters, and here, prediction data can be inside the entity that expectation obtains model prediction result (for example, deriving from
It is expected that obtaining bank, enterprise, the school etc. of prediction result), alternatively, prediction data can also derive from other than above-mentioned entity, for example,
From metadata provider, internet (for example, social network sites), mobile operator, APP operator, express company, credit institution
Deng.In addition, prediction data is either the data obtained online, are also possible to the data obtained offline, also, prediction data can
It is pre-processed before being obtained by prediction data acquiring unit 120 to be handled in predetermined format with being used for subsequent prediction, or
Person can be treated as predetermined format for subsequent pre- by characteristic processing after being obtained by prediction data acquiring unit 120
Survey processing.In addition, prediction data can also be input to prediction data acquiring unit 120 by input unit, or can be by prediction number
It is automatically generated according to acquiring unit 120 according to the data obtained, or can be by prediction data acquiring unit 120 from network
(for example, storage medium (for example, data warehouse) on network) obtains.In addition, the intermediate data switch of such as server
It can help to prediction data acquiring unit 120 and obtain corresponding data from external data source.It should be noted that the application is to this
In producing method, existence form, type, source and the acquisition modes of prediction data etc. do not do any restrictions, if its
Suitable for initial machine learning model group.
The prediction data of acquisition can be used to form forecast sample corresponding with prediction data, to be mentioned by machine learning model group
For predicting process accordingly.Accoding to exemplary embodiment, in above-mentioned initial machine learning model group, each machine learning model
Summation is weighted to be directed to prediction as the initial machine learning model group for the prediction result that prediction data provides
The prediction result of data.
Monitoring unit 130 can monitor the distribution change whether occurred in the prediction data of lasting acquisition beyond preset threshold
Change.Here, distribution variation may include the joint probability distribution variation of prediction data.In addition, preset threshold here can be by
User presets according to the practical application scene of machine learning model.Accoding to exemplary embodiment, monitoring unit 130 can monitor
The distribution and/or initial machine learning model group of the prediction data persistently obtained are imitated for the prediction of the prediction data
Fruit, to determine whether that the distribution beyond preset threshold changes.Here, the distribution of prediction data can refer to prediction number
According to and its legitimate reading probability distribution state, variation size can characterize spy included in prediction data to a certain extent
Whether therefore the variation degree of sign is compared by being changed size with preset threshold, it may be determined that occur in prediction data
Distribution variation beyond preset threshold.
As an example, initial machine learning model group may include that prediction is accurate for the prediction effect of the prediction data
Rate, accurate rate, recall rate etc., still, it will be clear to someone skilled in the art that the index for evaluation and foreca effect is not limited to
This.In general, if the distribution of prediction data changes, the initial machine trained using previous training data
The prediction effect that model group is practised for the prediction data newly to arrive can also decline, and therefore, optionally, monitoring unit 130 can also lead to
Cross in prediction data of the monitoring initial machine learning model group for the prediction effect of prediction data to determine lasting acquisition whether
There is the distribution beyond preset threshold to change.Specifically, for example, monitoring unit 130 can obtain initial machine learning model
Group passes through the true knot of the prediction result and the prediction data that will acquire for the prediction result of the prediction data persistently obtained
Fruit is compared to determine the prediction effect that initial machine learning model group is directed to the prediction data, and according to determining pre-
Survey effect, it is determined whether the distribution beyond threshold value occur and change.As an example, monitoring unit can be by the variation of prediction effect
Size (for example, variation size of predictablity rate) is compared to determine in prediction data whether occur with the preset threshold
Distribution variation beyond the preset threshold.As an example, monitoring unit 130 can learn the prediction of device from automaton
Unit (not shown) obtains the prediction result that initial machine learning model group is directed to the prediction data persistently obtained.Specifically, in advance
Prediction can be executed for the prediction data persistently obtained to generate to be directed to and persistently obtain using initial machine learning model group by surveying unit
The prediction result of the prediction data taken, and the prediction result of generation is supplied to monitoring unit 130.
Alternatively, optionally, monitoring unit 130 can also monitor the distribution and initial machine of the prediction data of lasting acquisition
Learning model group is directed to both prediction effects of the prediction data, to determine whether the distribution beyond preset threshold
Whether variation changes more accurately default about occurring exceeding in prediction data to obtain so as to comprehensive analysis the two
The definitive result of the distribution variation of threshold value.
In the case where not occurring the distribution variation beyond preset threshold, initial machine learning model group is not necessarily to by more
Newly, and initial machine learning model group is provided to expectation for the prediction result of prediction data and obtains model prediction result
Entity.However, updating unit 140 is automatically updated just in the case where there is the distribution variation beyond preset threshold
Beginning machine learning model group.Accoding to exemplary embodiment, in the case where there is the distribution variation beyond preset threshold, more
New unit 140 can fixedly using or adaptively automatically update initial machine using one of scheduled a variety of update modes
Practise model group." fixedly using " can refer to：In the case where there is the distribution variation beyond threshold value, updating unit 140 can
Select a kind of update mode fixed in scheduled a variety of update modes to automatically update initial machine learning model group always.
" adaptively using " can refer to：In the case where there is the distribution variation beyond threshold value, updating unit 140 can be adaptively
Corresponding update mode is selected among scheduled a variety of update modes to automatically update initial machine learning model group.Root
According to exemplary embodiment, scheduled a variety of update modes here may include following update mode：First update mode, is instructed again
Practice new machine learning model group to substitute initial machine learning model group；Second update mode, adjustment initial machine learn mould
The weight of each machine learning model in type group；Or third update mode, the new machine learning model group of training simultaneously will be new
Machine learning model group combined with initial machine learning model group to constitute updated machine learning model group.However, more
New paragon is without being limited thereto, for example, following update mode can also be used：Prediction effect in initial machine learning model group is deleted to be lower than
The part machine learning model of threshold value and only retain the part machine that prediction effect is relatively high in initial machine learning model group
Then the machine learning model group of the part machine learning model of reservation and re -training is combined to constitute by learning model
Updated machine learning model group.
Accoding to exemplary embodiment, in above first update mode, updating unit 140 is using based at least part
The training data that prediction data and its legitimate reading are formed, the new machine learning model group of re -training.Here, described at least one
Fractional prediction data can be whole prediction data that prediction data acquiring unit 120 persistently obtains, but not limited to this, it is optional
Ground, at least part prediction data are also possible to the distribution that includes at least in the whole prediction data for continuing acquisition and send out
A part of prediction data of prediction data of the changing beyond preset threshold, and whole prediction data of non-continuous acquisition.Here,
The legitimate reading of prediction data may include authentic signature (label) information about prediction data.As an example, authentic signature is believed
Breath may include the true feedback information about the prediction data obtained from external data source.In addition, as an example, with previously obtaining
Take the mode of initial machine learning model group identical, it is new that automaton study mode re -training can also be used in updating unit 140
Machine learning model group, and can be according to prediction effect of the new machine learning model group on above-mentioned training data, based on greedy
Center algorithm determines the weight of each machine learning model in new machine learning model group.About greedy algorithm, in view of this field skill
Therefore its known related content of art personnel here no longer repeats the detail of greedy algorithm.
In the second update mode, updating unit 140 can be according to initial machine learning model group based at least part
The prediction effect in observation data that prediction data and its legitimate reading are formed is each in initial machine learning model group to adjust
The weight of a machine learning model.Accoding to exemplary embodiment, at least part prediction data may include that distribution becomes
The whole prediction data changing the prediction data beyond preset threshold or persistently obtaining.It should be noted that being different from first more
New paragon, the second update does not need the new machine learning model group of re -training, and is only the previous initial machine of adjustment
The weight of each machine learning model in model group is practised to obtain updated machine learning model group.As an example, updating
Unit 140 can pass through prediction of each machine learning model in assessment initial machine learning model group in above-mentioned observation data
Effect performance is come machine each when calculating the prediction effect maximization when initial machine learning model group in above-mentioned observation data
The percentage contribution of learning model, with the weight of each machine learning model of determination.
In the third update mode, updating unit 140 is using based at least part prediction data and its true knot
Fruit shape at the new machine learning model group of training data training, also, by by new machine learning model group and initial machine
The combination of device learning model group, and according to the machine learning model group after combination based at least part prediction data and its really
As a result the prediction effect in the observation data formed determines each machine learning model in the machine learning model group after combination
Weight constitute updated machine learning model group.That is, updating unit 140 can be by learning mould in initial machine
New machine learning model group is introduced in type group to constitute updated machine learning model group.Specifically, updating unit 140 is first
First with the new machine learning model group of the training data training based at least part prediction data and its legitimate reading formation.
Here, at least part prediction data may include that distribution variation exceeds the prediction data of preset threshold or persistently obtains
The whole prediction data taken, and in the third update mode in the way of the new machine learning model group of training data training
Can it is identical as the new mode of machine learning model group of re -training in the first above-mentioned update mode (for example,
In the way of automaton study), therefore, which is not described herein again.Then, updating unit 140 can be by new machine learning model
Group is combined with initial machine learning model group, and according to the machine learning model group after combination based at least part prediction number
According to and its legitimate reading formed observation data on prediction effect determine combine after machine learning model group in each machine
The weight of device learning model.Here, the observation data formed based at least part prediction data and its legitimate reading can with
The training data of upper description formed based at least part prediction data and its legitimate reading is identical or part is identical.Separately
Outside, similar with other update modes, in the third update mode, updating unit 140 may be based on greedy algorithm and determine combination
The weight of each machine learning model in machine learning model group afterwards, which is not described herein again.
As described above, updating unit 140 can be automatically updated adaptively using one of scheduled a variety of update modes just
Beginning machine learning model group.Specifically, accoding to exemplary embodiment, updating unit 140 can have according at least to prediction data
It is initial to automatically update to select corresponding update mode among scheduled a variety of update modes to the information self-adapting of pass
Machine learning model group.Here, as an example, information related with prediction data, which may include, can be used in update initial machine
Practise the data volume and/or information relevant to the variation of the distribution of prediction data of the prediction data of model group.As described above, such as
Fruit can update initial machine learning model group using the whole prediction data persistently obtained, then can be used in updating initial machine
The data volume of the prediction data of device learning model group can be the data volume of whole prediction data of lasting acquisition.If being capable of benefit
Initial machine learning model group is updated with the changed prediction data of distribution, then can be used in updating initial machine
The data volume for practising the prediction data of model group can be the data volume of the changed prediction data of distribution.As an example,
Change the distribution change rate and/or prediction data that relevant information may include prediction data to the distribution of prediction data
Distribution variation periodicity.Here, the distribution that the distribution change rate of prediction data can refer to prediction data becomes
Change speed, and it can be obtained for example, by directly calculating the distribution of prediction data variation speed in the given time
, alternatively, it can also be obtained by computing machine learning model group for the variation speed of the prediction effect of prediction data.In advance
Whether the distribution that the periodicity of the distribution variation of measured data can refer to prediction data is in periodically variable.
Accoding to exemplary embodiment, it is assumed that scheduled a variety of update modes include three of the above update mode (that is,
One update mode, the second update mode and third update mode), then corresponding update is adaptive selected in updating unit 140
When mode, in the case where data volume is less than first threshold and distribution change rate is greater than second threshold, updating unit 140 can
Select third update mode；In the case where data volume is less than first threshold and distribution change rate is less than second threshold, more
The first update mode may be selected in new unit 140；It is greater than first threshold in data volume and distribution change rate is less than second threshold
In the case where, the second update mode may be selected in updating unit 140；It is greater than first threshold and distribution variation in week in data volume
In the case where phase property, the second update mode is may be selected in updating unit 140；It is greater than first threshold, distribution variation in data volume
Rate is greater than second threshold, and distribution variation, not in periodic situation, third update side may be selected in updating unit 140
Formula.This is because the first update mode is due to machine usually new come re -training using whole prediction data of lasting acquisition
Learning model group, thus expend computing resource and the time it is often larger, therefore if data volume it is smaller and respectively state change compared with
Slowly, then the first update mode is conveniently selected.Second update mode is not due to needing the new machine learning model of re -training
Group and only adjust the weight of initial machine learning model group, therefore, the computing resource of consuming and time are relatively small, therefore if
Data volume is larger and the slower or data volume of distribution variation is larger and distribution variation is in periodical, then relatively adaptation choosing
Select the second update mode.Third update mode due to introducing new machine learning model group in initial machine learning model group,
Therefore conveniently processing distribution changes very fast situation.Here, first threshold and second threshold can be according to machine learning
Practical application scene is configured, and there is no restriction to this by the application.
It should be noted that although being referred to the above adaptively selected corresponding update mode in the exemplary embodiment
Several situations, but it will be clear to someone skilled in the art that factor based on either adaptively selected update mode, still
The specific choice strategy of adaptively selected update mode is not limited to above example.
Updated machine learning model group can be used for the prediction data obtained continuing with prediction data acquiring unit 120
Prediction is executed, and updated machine learning model group is provided to monitoring unit for the prediction result of prediction data
130, prediction data is determined in order to which monitoring unit 130 changes according to the distribution of the prediction result and/or prediction data
In whether occur beyond preset threshold distribution change.Determine in prediction data occur beyond default in monitoring unit 130
In the case where the distribution variation of threshold value, updating unit 140 fixedly can be used or adaptively be used predetermined as described above
One of a variety of update modes automatically update current machine learning model group.
In addition, accoding to exemplary embodiment, automaton study device 100 can be disposed beyond the clouds or local.Automaton
Learning device 100 can be deployed in such as public cloud, private clound or mixed cloud, and the reality of corresponding prediction result can be obtained to expectation
Body (for example, content operator) provides machine learning service related with prediction data.Optionally, automaton learns device
100 can also be deployed in local, for example, the local system of content operator.
It should be noted that although above describe automaton study device 100 when, for convenience of description, by automatic machine
Device learning device 100 is divided into for executing respective treated unit respectively (for example, initial machine learning model group acquiring unit
110, prediction data acquiring unit 120, monitoring unit 130 and updating unit 140), however, those skilled in the art are clear
It is that the processing that above-mentioned each unit executes can also learn device 100 itself in automaton and draw without any specific unit
Point or each unit between have no clearly demarcate in the case where execute.In addition, learning dress above by reference to Fig. 1 automaton described
It sets 100 and is not limited to include unit described above, but some units can be increased as needed, and the above unit can also quilt
Combination.
Learning device according to the automaton of the application exemplary embodiment can persistently supervise during automaton learns
The distribution variation of the prediction data of input (for example, inflow) is surveyed, and constantly automatically updates initial machine according to monitoring result
Device learning model, so that the initial machine learning model group of continuous updating is capable of providing accurate prediction result always, from
And further perfect automaton at present learns technology.Further, since can be in the case where there is distribution variation not
Initial machine learning model group is automatically updated disconnectedly, therefore, the distribution of prediction data can be made to change constantly automatic
Learn to arrive, consequently facilitating providing more accurately prediction for new prediction data using updated initial machine learning model group
As a result.
More than, it has described referring to Fig.1 and device 100 is learnt according to the automaton of the application exemplary embodiment.?
Hereinafter, the method learnt according to the execution automaton of the application exemplary embodiment will be retouched referring to Fig. 2 and Fig. 3
It states.
Fig. 2 shows the flow charts for executing the method that automaton learns according to the application exemplary embodiment.
Here, as an example, automaton learning method shown in Fig. 2 can automaton as shown in Figure 1 learn device
100 execute, and can also be realized completely by computer program with software mode, can also be held by the computing device of specific configuration
Row.For convenience, it is assumed that the automaton of method shown in Fig. 2 as shown in Figure 1 learns device 100 to execute, and assumes
Automaton study device 100 can have configuration shown in FIG. 1.
Referring to Fig. 2, in step S210, initial machine learning model group acquiring unit 110 can pass through automaton study side
Formula obtains the initial machine learning model group including at least one machine learning model.More than, referring to Fig.1 to automaton
Mode of learning and initial machine learning model group are described, and therefore, which is not described herein again.As an example, initial machine
Learning model group acquiring unit 110 can be trained initial by way of automaton study based on the training data obtained in advance
Machine learning model group.The mode of initial machine learning model group about training data and is specifically trained, it is above also to have existed
It describes to carry out description during initial machine learning model group acquiring unit 110 referring to Fig.1, which is not described herein again.
In step S220, the sustainable acquisition prediction data of prediction data acquiring unit 120.As an example, prediction data obtains
Take unit 120 that can constantly obtain prediction data by manual, semi or fully automatic mode.The prediction data of acquisition can quilt
Pre-processing is predetermined format or form, alternatively, pre- fix can be treated as after being predicted data capture unit 120 and obtaining
Formula or form.For example, prediction data acquiring unit 120 can by input unit (for example, work station) receive operator (for example,
The operator of the entity of prediction result is obtained from expectation) prediction data that persistently inputs manually.Alternatively, prediction data obtains list
Member 120 from data source systems can persistently obtain prediction data by full automatic mode, for example, by with software, firmware, hard
The timer mechanism that part or combinations thereof is realized obtains requested prediction data come systematically request data source and from response.Institute
Stating data source may include one or more databases or server.Prediction data acquiring unit 120 can via internal network and/or
External network realizes the acquisition of prediction data, and prediction data can be the data by encryption.Server, database,
Network etc. is configured as in the case where communicating with one another, and obtaining for prediction data can be carried out automatically in the case where no manual intervention
It takes, it should be noted that certain user still may be present in this manner inputs operation.For example, receiving specific user
In the case where input, the request for obtaining prediction data is generated.In addition, will can be obtained as needed when getting prediction data every time
The prediction data taken is stored in corresponding storage medium.As an example, availability data warehouse stores the pre- of lasting acquisition
Measured data.
Accoding to exemplary embodiment, the prediction data of acquisition can be used to form forecast sample corresponding with prediction data, with
Participate in the prediction process of initial machine learning model group.Accoding to exemplary embodiment, in above-mentioned initial machine learning model group,
The prediction result that each machine learning model is provided for prediction data is weighted summation using as the initial machine
Practise the prediction result that model group is directed to prediction data.
In step S230, monitoring unit 130 can monitor in the prediction data of lasting acquisition whether occur beyond preset threshold
Distribution variation.Accoding to exemplary embodiment, in step S230, monitoring unit 230 can monitor the prediction number of lasting acquisition
According to distribution and/or initial machine learning model group be directed to the prediction data prediction effect, to determine whether
Distribution variation beyond preset threshold.Above during the description as described in monitoring unit referring to Fig.1, to distribution shape
State transformation, preset threshold and prediction effect etc. are described, and therefore, which is not described herein again.As an example, in step
S230, monitoring unit 130 can monitoring and forecasting data joint probability distribution state variation size, and by the transform size and pre-
If threshold value is compared to determine that whether occurring the distribution beyond preset threshold in prediction data changes.Show as another
Example, in step S230, monitoring unit 130 can obtain the prediction that initial machine learning model group is directed to the prediction data persistently obtained
As a result, being compared to determine that initial machine learns mould by the legitimate reading of the prediction result and the prediction data that will acquire
Type group is directed to the prediction effect of the prediction data, and according to determining prediction effect, it is determined whether occurs beyond threshold value
Distribution variation.Optionally, in step S230, monitoring unit 130 can monitor the distribution of the prediction data of lasting acquisition
It is directed to both prediction effects of the prediction data with initial machine learning model group, and passes through the change of comprehensive analysis both of the above
Change to determine that whether occurring the distribution beyond preset threshold in prediction data changes.
In the case where not occurring the distribution variation beyond preset threshold, initial machine learning model group is not necessarily to by more
Newly, prediction data acquiring unit 120 can continue to obtain new prediction data.In addition, initial machine learning model group is for prediction
The prediction result of data is provided to the entity that expectation obtains model prediction result.However, in step S230 monitoring unit 130
Determine that, in step S240, updating unit 140 can in the case where occurring the distribution variation beyond preset threshold in prediction data
Automatically update initial machine learning model group.
Specifically, in step S240, in the case where there is the distribution variation beyond preset threshold, updating unit
140 can fixedly using or adaptively automatically update initial machine learning model using one of scheduled a variety of update modes
Group.Accoding to exemplary embodiment, scheduled a variety of update modes here may include following update mode：First update mode,
The new machine learning model group of re -training is to substitute initial machine learning model group；Second update mode adjusts initial machine
The weight of each machine learning model in learning model group；Or third update mode, the new machine learning model group of training
And it combines new machine learning model group to constitute updated machine learning model group with initial machine learning model group.So
And update mode is without being limited thereto.
Accoding to exemplary embodiment, in the first update mode, updating unit 140 is predicted using based at least part
The training data that data and its legitimate reading are formed, the new machine learning model group of re -training.As an example, described at least one
Fractional prediction data can be the whole prediction data persistently obtained in step S120.In the second update mode, updating unit
140 can be according to initial machine learning model group in the observation data based at least part prediction data and its legitimate reading formation
On prediction effect adjust the weight of each machine learning model in initial machine learning model group.As an example, described
At least part prediction data may include the whole that distribution variation exceeds the prediction data of preset threshold or persistently obtains
Prediction data.In the third update mode, updating unit 140 is using based at least part prediction data and its true knot
Fruit shape at the new machine learning model group of training data training, also, by by new machine learning model group and initial machine
The combination of device learning model group, and according to the machine learning model group after combination based at least part prediction data and its really
As a result the prediction effect in the observation data formed determines each machine learning model in the machine learning model group after combination
Weight constitute updated machine learning model group.As an example, in the third update mode, described at least part
Prediction data may include whole prediction data that distribution variation exceeds the prediction data of preset threshold or persistently obtains.
As described above, can adaptively be come using one of scheduled a variety of update modes in step S240 updating unit 140
Automatically update initial machine learning model group.Specifically, accoding to exemplary embodiment, updating unit 140 can according at least to
Corresponding update mode is selected to come among scheduled a variety of update modes to the related information self-adapting of prediction data
It is dynamic to update initial machine learning model group.Here, information related with prediction data may include that can be used in updating initial machine
The data volume of the prediction data of learning model group and/or information relevant to the variation of the distribution of prediction data.As an example,
Change the distribution change rate and/or prediction data that relevant information may include prediction data to the distribution of prediction data
Distribution variation periodicity.
Accoding to exemplary embodiment, it is assumed that scheduled a variety of update modes include three of the above update mode (that is,
One update mode, the second update mode and third update mode), then in step S240, data volume be less than first threshold and point
In the case that cloth status variation rate is greater than second threshold, third update mode is may be selected in updating unit 140；In data volume less than
In the case that one threshold value and distribution change rate are less than second threshold, the first update mode is may be selected in updating unit 140；In number
It is greater than first threshold according to amount and distribution change rate is less than in the case where second threshold, updating unit 140 may be selected second more
New paragon；It is greater than first threshold in data volume and distribution variation is in periodic situation, updating unit 140 may be selected the
Two update modes；It is greater than first threshold in data volume, distribution change rate is greater than second threshold, and distribution variation is not in
In periodic situation, third update mode is may be selected in updating unit 140.
Above during describing updating unit referring to Fig.1, to three of the above update mode and its phase being related to
Hold inside the Pass and carried out introduction, therefore, which is not described herein again.In addition, it should be noted that, being adaptive selected more in step S240
Factor based on new paragon is not limited in can be used in updating the data volume of the prediction data of initial machine learning model group
And change related information with the distribution of prediction data, but it is also possible to consider other information related with prediction data,
And the specific choice strategy of adaptively selected update mode can also change according to other Considerations.
After step S240 is updated initial machine learning model group, updated machine learning model group is available
In executing prediction, and updated machine in the prediction data that step S220 is obtained continuing with prediction data acquiring unit 120
Device learning model group is provided to monitoring unit 130 for the prediction result of prediction data, so that monitoring unit 130 can continue
Step S230 is executed to determine that whether occurring the distribution beyond preset threshold in prediction data changes, and is being occurred beyond pre-
If can be automatically updated as described above just in the case where the distribution variation of threshold value in step S240 updating unit 140
The mode of beginning machine learning model group automatically updates current machine learning model group.Process as described above it is sustainable into
Row, so that can consider the variation of prediction data distribution always in automaton study and by automatically updating current machine
Device learning model group constantly learns this variation, so that machine learning model group be enable to provide for new prediction data
More accurate prediction result.
It should be noted that although the step in Fig. 2 is described in order when describing Fig. 2 above, still,
It will be clear to someone skilled in the art that each step in the above method not necessarily executes in order, but can also concurrently hold
Row, for example, process described above S220 and step S230 can be executed parallel, that is to say, that monitored in monitoring unit 230
While whether occurring the distribution variation beyond preset threshold in prediction data through obtaining, prediction data acquiring unit
The new prediction data of 220 still sustainable acquisitions.In addition, during step S240 automatically updates initial machine learning model, step
Rapid S220 can also be performed simultaneously with lasting acquisition prediction data.
It can be monitored during automaton learns according to the automaton learning method of the application exemplary embodiment pre-
The distribution of measured data changes, and occurs in the case that the distribution beyond preset threshold changes automatically in prediction data
Initial machine learning model is updated, so that further perfect automaton at present learns technology.Further, since more than passing through
Monitoring operation and operation is automatically updated, machine learning model group is enable constantly to learn the distribution to the prediction data of appearance automatically
State change, therefore, convenient for providing more accurate prediction result for new prediction data.
For convenient for being more clearly understood that the automaton learning method according to the application exemplary embodiment, referring to figure
3 are more intuitively briefly described the example process of automaton learning method.
Fig. 3 is the schematic diagram for showing the method for executing automaton study according to the application exemplary embodiment.
In the exemplary embodiment shown in Fig. 3, it is assumed that only by monitoring initial machine learning model group for prediction number
According to prediction effect come determine whether occur in prediction data beyond preset threshold distribution change.But, it should be clear that
It is that as described above with reference to FIG. 2, in this application, the distribution of the prediction data of lasting acquisition can also be monitored directly to determine
Whether occur the distribution beyond preset threshold in prediction data to change, alternatively, the prediction data of lasting acquisition can be monitored
Distribution and initial machine learning model group are determined whether for both prediction effects of prediction data beyond default
The distribution of threshold value changes.
As shown in figure 3, firstly, initial machine learning model group acquiring unit 110 can obtain training data, and pass through
AutoML mode trains initial machine learning model group based on the training data of acquisition.Then, prediction data acquiring unit 120
Constantly obtain prediction data.Here, the prediction data of acquisition can be input into initial machine learning model at predetermined time intervals
Group, to execute prediction for prediction data using initial machine learning model group.Next, initial machine learning model group is directed to
The prediction result of prediction data is provided to monitoring unit 130.Then, monitoring unit 130 can learn mould according to initial machine
Type group determines whether occur distribution variation in prediction data for the prediction effect (for example, predictablity rate) of prediction data
Distribution variation beyond preset threshold.Specifically, for example, monitoring unit 130 can pass through the prediction result that will acquire and institute
The legitimate reading for stating prediction data is compared to determine that initial machine learning model group is imitated for the prediction of the prediction data
Fruit, and according to determining prediction effect, it is determined whether there is the distribution beyond threshold value and changes.Do not occurring beyond default
In the case where the distribution variation of threshold value, prediction data acquiring unit 120 continues to obtain prediction data.Occurring beyond default
In the case where the distribution variation of threshold value, the automatically updated initial machine learning model group of updating unit 140.Updated machine
The prediction data that device learning model group can be obtained continuing with prediction data acquiring unit 120 executes prediction.Such process can
It is repeatedly executed, so that change whenever occurring the distribution beyond preset threshold in the prediction data for monitoring to continue acquisition,
Machine learning model group can be automatically updated, thus the constantly distribution variation of study prediction data automatically, to make machine
Device learning model group can provide more accurate prediction result for new prediction data.
The dress for executing automaton study according to the application exemplary embodiment is described with reference to Fig. 1 to Fig. 3 above
It sets and method.It is to be understood, however, that：Device and its unit shown in figure 1 can be individually configured to execute specific function
Any combination of software, hardware, firmware or above-mentioned item.For example, these devices or unit can correspond to dedicated integrated circuit,
It can correspond to pure software code, also correspond to the module that software is combined with hardware.In addition, these devices or unit institute
The one or more functions of realization can also be by the component in physical entity equipment (for example, processor, client or server etc.)
To seek unity of action.
In addition, the above method can be realized by the program being recorded in computer-readable media, for example, being shown according to the application
Example property embodiment, it is possible to provide it is a kind of for execute automaton study computer-readable medium, wherein the computer can
It reads to record the computer program having for executing following methods step on medium：It is obtained by way of automaton study including extremely
The initial machine learning model group of a few machine learning model；Persistently obtain prediction data；Monitor the prediction number persistently obtained
Whether occur the distribution beyond preset threshold in change；In the case where there is the distribution variation beyond threshold value,
Automatically update initial machine learning model group.
Computer program in above-mentioned computer-readable medium can be in client, host, agent apparatus, server etc.
Run in the environment disposed in computer equipment, it should be noted that the computer program can also be used in execute in addition to above-mentioned steps with
Outer additional step or execute when executing above-mentioned steps more specifically handles, these additional steps and is further processed
Content referring to Fig. 2 carry out correlation technique description during refer to, therefore here in order to avoid repeat will no longer carry out
It repeats.
It should be noted that computer program can be completely dependent on by learning device according to the automaton of the application exemplary embodiment
Operation is to realize corresponding function, that is, each unit is corresponding to each step in the function structure of computer program, so that entirely
Device is called by special software package (for example, the library lib), to realize corresponding function.
On the other hand, device or unit shown in FIG. 1 can also by hardware, software, firmware, middleware, microcode or
Any combination thereof is realized.When with the realization of software, firmware, middleware or microcode, for executing the program generation of corresponding operating
Code or code segment can store in the computer-readable medium of such as storage medium, so that processor can be by reading and transporting
The corresponding program code of row or code segment execute corresponding operation.
For example, the exemplary embodiment of the application is also implemented as computing device, which includes storage unit
And processor, set of computer-executable instructions conjunction is stored in storage unit, when the set of computer-executable instructions is closed by institute
When stating processor execution, method comprising the following steps are executed：It is obtained by way of automaton study including at least one machine
The initial machine learning model group of device learning model；Persistently obtain prediction data；Monitor in the prediction data that persistently obtains whether
There is the distribution beyond preset threshold to change；In the case where there is the distribution variation beyond threshold value, automatically update
Initial machine learning model group.
Particularly, the computing device can be deployed in server or client, can also be deployed in distributed network
On node apparatus in network environment.In addition, the computing device can be PC computer, board device, personal digital assistant, intelligence
Energy mobile phone, web are applied or other are able to carry out the device of above-metioned instruction set.
Here, the computing device is not necessarily single computing device, can also be it is any can be alone or in combination
Execute the device of above-metioned instruction (or instruction set) or the aggregate of circuit.Computing device can also be integrated control system or system
A part of manager, or can be configured to Local or Remote (for example, via wireless transmission) with the portable of interface inter-link
Formula electronic device.
In the computing device, processor may include central processing unit (CPU), graphics processor (GPU), may be programmed and patrol
Collect device, dedicated processor systems, microcontroller or microprocessor.As an example, not a limit, processor may also include simulation
Processor, digital processing unit, microprocessor, multi-core processor, processor array, network processing unit etc..
Certain operations described in the method learnt according to the execution automaton of the application exemplary embodiment can lead to
Software mode is crossed to realize, certain operations can be realized by hardware mode, in addition, can also by way of software and hardware combining come
Realize these operations.
Processor can run the instruction being stored in one of storage unit or code, wherein the storage unit can be with
Storing data.Instruction and data can be also sent and received via Network Interface Unit and by network, wherein the network connects
Any of transport protocol can be used in mouth device.
Storage unit can be integral to the processor and be integrated, for example, RAM or flash memory are arranged in integrated circuit microprocessor etc.
Within.In addition, storage unit may include independent device, such as, external dish driving, storage array or any Database Systems can
Other storage devices used.Storage unit and processor can be coupled operationally, or can for example by the port I/O,
Network connection etc. communicates with each other, and enables a processor to read the file being stored in storage unit.
In addition, the computing device may also include video display (such as, liquid crystal display) and user's interactive interface is (all
Such as, keyboard, mouse, touch input device etc.).The all components of computing device can be connected to each other via bus and/or network.
It operates and can be described as involved in the method learnt according to the execution automaton of the application exemplary embodiment
The functional block or function diagram of various interconnections or coupling.However, these functional blocks or function diagram can be equably integrated into list
A logic device is operated according to non-exact boundary.
The method and dress of the execution automaton study that exemplary embodiment is described according to the application are had been combined above
It sets, can be widely applied to data dependent with any machine learning scene of distribution, for example, these machine learning scenes can be with
It is such as online content (such as, news, advertisement, music etc.) recommendation, credit card fraud detection, unusual checking, intelligence battalion
The machine learning scene of pin, intellectual investment consultant, network traffic analysis etc..
Although the foregoing describe the exemplary embodiments of the application, it should be understood that：Foregoing description is merely exemplary, and
Non-exclusive.The application is not limited to disclosed each exemplary embodiment, and without departing from the scope and spirit of the present application
In the case where, many modifications and changes are obvious for those skilled in the art.Therefore, originally
The protection scope of application should be subject to the scope of the claims.
Claims (10)
1. a kind of method for executing automaton study, including：
The initial machine learning model group including at least one machine learning model is obtained by way of automaton study；
Persistently obtain prediction data；
It monitors and whether occurs the distribution variation beyond preset threshold in the prediction data persistently obtained；
In the case where there is the distribution variation beyond threshold value, initial machine learning model group is automatically updated.
2. the method for claim 1, wherein in the initial machine learning model group, each machine learning model
Summation is weighted to be directed to prediction as the initial machine learning model group for the prediction result that prediction data provides
The prediction result of data,
Also, monitoring step includes：Monitor the distribution and/or initial machine learning model group of the prediction data persistently obtained
For the prediction effect of the prediction data, to determine whether that the distribution beyond preset threshold changes.
3. method according to claim 2, wherein in the case where there is the distribution variation beyond preset threshold, Gu
Surely it uses or uses one of scheduled a variety of update modes adaptively to automatically update initial machine learning model group.
4. method as claimed in claim 3, wherein scheduled a variety of update modes include following update mode：
First update mode, the new machine learning model group of re -training is to substitute initial machine learning model group；
Second update mode adjusts the weight of each machine learning model in initial machine learning model group；Or
New machine learning model group and initial machine are simultaneously learnt mould by third update mode, the new machine learning model group of training
Type group combines to constitute updated machine learning model group.
5. method as claimed in claim 4, wherein
In the first update mode, the training data based at least part prediction data and its legitimate reading formation, weight are utilized
New machine learning model group is newly trained,
In the second update mode, according to initial machine learning model group based at least part prediction data and its true knot
Fruit shape at observation data on prediction effect adjust the power of each machine learning model in initial machine learning model group
Weight；
In third update mode, the training data training based at least part prediction data and its legitimate reading formation is utilized
New machine learning model group, also, by combining new machine learning model group with initial machine learning model group, and root
According to the machine learning model group after combination in the observation data based at least part prediction data and its legitimate reading formation
Prediction effect determine combination after machine learning model group in each machine learning model weight it is updated to constitute
Machine learning model group.
6. method as claimed in claim 5, wherein in the first update mode, at least part prediction data includes
The whole prediction data persistently obtained, in the second update mode and third update mode, at least part prediction data
The whole prediction data for changing the prediction data for exceeding preset threshold including distribution or persistently obtaining.
7. method as claimed in claim 4, wherein adaptively automatically updated using one of scheduled a variety of update modes
The step of initial machine learning model group includes：According at least to information self-adapting related with prediction data described scheduled
Corresponding update mode is selected among a variety of update modes to automatically update initial machine learning model group.
8. a kind of for executing the computer-readable medium of automaton study, wherein remember on the computer-readable medium
Record has the computer program for executing the method as described in any claim in claim 1 to 7.
9. a kind of computing device for executing automaton study, including storage unit and processor, wherein stored in storage unit
There is set of computer-executable instructions conjunction, when the set of computer-executable instructions, which is closed, to be executed by the processor, promotes to handle
Device executes the method as described in any claim in claim 1 to 7.
10. a kind of device for executing automaton study, including：
Initial machine learning model group acquiring unit is configured as obtaining by way of automaton study including at least one machine
The initial machine learning model group of device learning model；
Prediction data acquiring unit is configured as continuing to obtain prediction data；
Monitoring unit is configured as the distribution change for whether occurring beyond preset threshold in the prediction data for monitoring lasting acquisition
Change；
Updating unit is configured as automatically updating initial machine study in the case where there is the distribution variation beyond threshold value
Model group.
Priority Applications (2)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810641700.7A CN108921300A (en) | 2018-06-21 | 2018-06-21 | The method and apparatus for executing automaton study |
| CN201910540759.1A CN110705719B (en) | 2018-06-21 | 2019-06-21 | Method and apparatus for performing automatic machine learning |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810641700.7A CN108921300A (en) | 2018-06-21 | 2018-06-21 | The method and apparatus for executing automaton study |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN108921300A true CN108921300A (en) | 2018-11-30 |
Family
ID=64421672
Family Applications (2)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201810641700.7A Pending CN108921300A (en) | 2018-06-21 | 2018-06-21 | The method and apparatus for executing automaton study |
| CN201910540759.1A Active CN110705719B (en) | 2018-06-21 | 2019-06-21 | Method and apparatus for performing automatic machine learning |
Family Applications After (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910540759.1A Active CN110705719B (en) | 2018-06-21 | 2019-06-21 | Method and apparatus for performing automatic machine learning |
Country Status (1)
| Country | Link |
|---|---|
| CN (2) | CN108921300A (en) |
Cited By (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111245805A (en) * | 2020-01-06 | 2020-06-05 | 北京元心科技有限公司 | EMM (emergency management message) based method, terminal device, server and system for generating management and control strategy |
| CN111382874A (en) * | 2018-12-28 | 2020-07-07 | 第四范式（北京）技术有限公司 | Method and device for realizing update iteration of online machine learning model |
| WO2020253775A1 (en) * | 2019-06-18 | 2020-12-24 | 第四范式（北京）技术有限公司 | Method and system for realizing machine learning modeling process |
| CN112149836A (en) * | 2019-06-28 | 2020-12-29 | 杭州海康威视数字技术股份有限公司 | Machine learning program updating method, device and equipment |
| CN113196313A (en) * | 2019-01-18 | 2021-07-30 | 欧姆龙株式会社 | Model integration device, model integration method, model integration program, estimation system, inspection system, and control system |
| CN111382346B (en) * | 2018-12-28 | 2023-09-01 | 第四范式（北京）技术有限公司 | Method and system for recommending content |
Families Citing this family (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109242105B (en) * | 2018-08-17 | 2024-03-15 | 第四范式（北京）技术有限公司 | Code optimization method, device, equipment and medium |
| CN111752703B (en) * | 2020-04-09 | 2024-03-19 | 杭州海康威视数字技术股份有限公司 | Processing resource configuration method and device for neural network training and intelligent analysis |
| CN111860629A (en) * | 2020-06-30 | 2020-10-30 | 北京滴普科技有限公司 | Jewelry classification system, method, device and storage medium |
| CN114005016A (en) * | 2020-07-28 | 2022-02-01 | 华为技术有限公司 | Image processing method, electronic equipment, image processing system and chip system |
| CN112446048A (en) * | 2020-11-26 | 2021-03-05 | 平安科技（深圳）有限公司 | Data sharing method, system, terminal and storage medium based on block chain |
| CN113516128A (en) * | 2021-03-04 | 2021-10-19 | 贵州电网有限责任公司 | OCR-based method and system capable of accepting picture input |
Family Cites Families (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US10068186B2 (en) * | 2015-03-20 | 2018-09-04 | Sap Se | Model vector generation for machine learning algorithms |
| US10163061B2 (en) * | 2015-06-18 | 2018-12-25 | International Business Machines Corporation | Quality-directed adaptive analytic retraining |
| CN107103187B (en) * | 2017-04-10 | 2020-12-29 | 四川省肿瘤医院 | Lung nodule detection grading and management method and system based on deep learning |
| CN107169573A (en) * | 2017-05-05 | 2017-09-15 | 第四范式（北京）技术有限公司 | Using composite machine learning model come the method and system of perform prediction |
| CN107358317A (en) * | 2017-06-28 | 2017-11-17 | 北京优特捷信息技术有限公司 | The method and device of time series forecasting is carried out by machine learning |
| CN108023876B (en) * | 2017-11-20 | 2021-07-30 | 西安电子科技大学 | Intrusion detection method and intrusion detection system based on sustainability ensemble learning |
| CN108090429B (en) * | 2017-12-08 | 2020-07-24 | 浙江捷尚视觉科技股份有限公司 | Vehicle type recognition method for graded front face bayonet |
-
2018
- 2018-06-21 CN CN201810641700.7A patent/CN108921300A/en active Pending
-
2019
- 2019-06-21 CN CN201910540759.1A patent/CN110705719B/en active Active
Cited By (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111382874A (en) * | 2018-12-28 | 2020-07-07 | 第四范式（北京）技术有限公司 | Method and device for realizing update iteration of online machine learning model |
| CN111382346B (en) * | 2018-12-28 | 2023-09-01 | 第四范式（北京）技术有限公司 | Method and system for recommending content |
| CN111382874B (en) * | 2018-12-28 | 2024-04-12 | 第四范式（北京）技术有限公司 | Method and device for realizing update iteration of online machine learning model |
| CN113196313A (en) * | 2019-01-18 | 2021-07-30 | 欧姆龙株式会社 | Model integration device, model integration method, model integration program, estimation system, inspection system, and control system |
| WO2020253775A1 (en) * | 2019-06-18 | 2020-12-24 | 第四范式（北京）技术有限公司 | Method and system for realizing machine learning modeling process |
| CN112149836A (en) * | 2019-06-28 | 2020-12-29 | 杭州海康威视数字技术股份有限公司 | Machine learning program updating method, device and equipment |
| CN112149836B (en) * | 2019-06-28 | 2024-05-24 | 杭州海康威视数字技术股份有限公司 | Machine learning program updating method, device and equipment |
| CN111245805A (en) * | 2020-01-06 | 2020-06-05 | 北京元心科技有限公司 | EMM (emergency management message) based method, terminal device, server and system for generating management and control strategy |
Also Published As
| Publication number | Publication date |
|---|---|
| CN110705719B (en) | 2024-09-03 |
| CN110705719A (en) | 2020-01-17 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN108921300A (en) | The method and apparatus for executing automaton study | |
| Xu et al. | Unleashing the power of edge-cloud generative ai in mobile networks: A survey of aigc services | |
| US10824478B2 (en) | Methods and systems for quantum ready and quantum enabled computations | |
| US11544494B2 (en) | Algorithm-specific neural network architectures for automatic machine learning model selection | |
| CN103502899B (en) | Dynamic prediction Modeling Platform | |
| CN110188910A (en) | The method and system of on-line prediction service are provided using machine learning model | |
| CN107392319A (en) | Generate the method and system of the assemblage characteristic of machine learning sample | |
| CN107077385B (en) | For reducing system, method and the storage medium of calculated examples starting time | |
| US20140358828A1 (en) | Machine learning generated action plan | |
| CN107844837A (en) | The method and system of algorithm parameter tuning are carried out for machine learning algorithm | |
| CN107169573A (en) | Using composite machine learning model come the method and system of perform prediction | |
| TWI785346B (en) | Dual machine learning pipelines for transforming data and optimizing data transformation | |
| CN108090570A (en) | For selecting the method and system of the feature of machine learning sample | |
| CN110383298A (en) | Data efficient intensified learning for continuous control task | |
| CN108665072A (en) | A kind of machine learning algorithm overall process training method and system based on cloud framework | |
| US10474926B1 (en) | Generating artificial intelligence image processing services | |
| US20230259829A1 (en) | Warm starting an online bandit learner model utilizing relevant offline models | |
| CN110447041A (en) | Noise neural net layer | |
| CN108108820A (en) | For selecting the method and system of the feature of machine learning sample | |
| US20240111739A1 (en) | Tuning large data infrastructures | |
| Almadhor et al. | A new offloading method in the green mobile cloud computing based on a hybrid meta-heuristic algorithm | |
| Horng et al. | Merging crow search into ordinal optimization for solving equality constrained simulation optimization problems | |
| KR102531299B1 (en) | A learning model recommendation device based on similarity and a cloud integrated operating system comprising the same | |
| US20140156334A1 (en) | Setting constraints in project portfolio optimization | |
| Mohd et al. | Rapid modelling of machine learning in predicting office rental price |
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 | ||
| WD01 | Invention patent application deemed withdrawn after publication | ||
| WD01 | Invention patent application deemed withdrawn after publication |
Application publication date: 20181130 |