CN110503186A - Commodity sequence neural network model training method, device, electronic equipment - Google Patents
Commodity sequence neural network model training method, device, electronic equipment Download PDFInfo
- Publication number
- CN110503186A CN110503186A CN201910656545.0A CN201910656545A CN110503186A CN 110503186 A CN110503186 A CN 110503186A CN 201910656545 A CN201910656545 A CN 201910656545A CN 110503186 A CN110503186 A CN 110503186A
- Authority
- CN
- China
- Prior art keywords
- commodity
- neural network
- network model
- hidden layer
- businessman
- 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
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
- G06N3/084—Backpropagation, e.g. using gradient descent
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
- G06Q30/00—Commerce
- G06Q30/02—Marketing; Price estimation or determination; Fundraising
- G06Q30/0201—Market modelling; Market analysis; Collecting market data
- G06Q30/0202—Market predictions or forecasting for commercial activities
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Business, Economics & Management (AREA)
- Strategic Management (AREA)
- Data Mining & Analysis (AREA)
- Finance (AREA)
- Development Economics (AREA)
- Accounting & Taxation (AREA)
- General Physics & Mathematics (AREA)
- Artificial Intelligence (AREA)
- Biophysics (AREA)
- Software Systems (AREA)
- Mathematical Physics (AREA)
- General Engineering & Computer Science (AREA)
- Health & Medical Sciences (AREA)
- Life Sciences & Earth Sciences (AREA)
- Entrepreneurship & Innovation (AREA)
- Biomedical Technology (AREA)
- Computing Systems (AREA)
- Computational Linguistics (AREA)
- Evolutionary Computation (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Marketing (AREA)
- Game Theory and Decision Science (AREA)
- Economics (AREA)
- General Business, Economics & Management (AREA)
- Image Analysis (AREA)
Abstract
The embodiment of the present application discloses a kind of commodity sequence neural network model training method, belongs to field of computer technology, facilitates lift scheme training effectiveness.The described method includes: migrating the department pattern parameter of businessman's sequence neural network model trained in advance to commodity sequence neural network model, with the parameter of part hidden layer in the first network for the neural network model that sorted according to the department pattern parameter initialization commodity, and the parameter of each hidden layer in the sequence neural network model of commodity described in random initializtion in addition to the part hidden layer；The businessman's dimensional characteristics and commodity dimensional characteristics of each commodity training sample are obtained respectively, construct the sample data of corresponding commodity training sample, the training commodity sequence neural network model.
Description
Technical field
This application involves field of computer technology, more particularly to a kind of commodity sequence neural network model training method,
Device, electronic equipment and computer readable storage medium.
Background technique
In commodity sequence scene, it is common practice to collect commodity original, then, be based on preset product features dimension
The product features for the commodity original collected are extracted, later, the product features training commodity sequence neural network model based on extraction,
The index score to be sorted with the lower single probability etc. for exporting commodity for commodity.Commodity sequence neural network is trained in this way
Model needs to obtain a large amount of training sample.Since commodity original haves the characteristics that sample is sparse, cause in model training process
In, model is not easy to restrain, and needs repetition learning and adjustment model parameter, and training effectiveness is relatively low.Also, by this method
The accuracy of the index score of the commodity sequence of trained neural network model output is directly influenced by the quantity of commodity original,
In the case where commodity original is sparse, pass through the index score accuracy rate meeting for the commodity sequence that trained neural network model exports
It reduces.
Therefore, in order to promote the training effectiveness of commodity sequence neural network model, a kind of commodity sequence neural network is needed
The training method of model can train the commodity sequence of the accurate sequence index score of output in the case where commodity original is sparse
Neural network model.
Summary of the invention
The embodiment of the present application provides a kind of commodity sequence neural network model training method, facilitates training for promotion and obtains quotient
The training effectiveness of product sequence neural network model.
To solve the above-mentioned problems, in a first aspect, the embodiment of the present application provides a kind of commodity sequence neural network model
Training method, comprising:
The commodity sequence neural network model includes parallel arrangement of first network and the second network, the first network
Network structure and the businessman neural network model that sorts match, which comprises
The department pattern parameter of businessman's sequence neural network model trained in advance is migrated to commodity sequence neural network
Model, with part hidden layer in the first network of the sequence neural network model of the commodity according to the department pattern parameter initialization
Parameter, and, commodity described in random initializtion sort each hidden layer in addition to the part hidden layer in neural network model
Parameter；
The businessman's dimensional characteristics and commodity dimensional characteristics of each commodity training sample, construction corresponding commodity training sample are obtained respectively
This sample data；
For each commodity training sample, using businessman's dimensional characteristics of the commodity training sample as described first
The input of network, and using the commodity dimensional characteristics of the commodity training sample as the input of second network, training
The commodity sequence neural network model.
Second aspect, the embodiment of the present application provide a kind of commodity sequence neural network model training device, comprising:
The commodity sequence neural network model includes parallel arrangement of first network and the second network, the first network
Network structure and the businessman neural network model that sorts match, described device includes:
Commodity order models initialization module, the department pattern of businessman's sequence neural network model for that will train in advance
Parameter is migrated to commodity sequence neural network model, with the sequence nerve net of the commodity according to the department pattern parameter initialization
The parameter of part hidden layer in the first network of network model, and, it is removed in the sequence neural network model of commodity described in random initializtion
The parameter of each hidden layer except the part hidden layer；
Commodity training sample constructs module, for obtaining the businessman's dimensional characteristics and commodity dimension of each commodity training sample respectively
Feature is spent, the sample data of corresponding commodity training sample is constructed；
Commodity order models training module is used for for each commodity training sample, by the commodity training sample
Input of businessman's dimensional characteristics as the first network, and the commodity dimensional characteristics of the commodity training sample are made
For the input of second network, the training commodity sequence neural network model.
The third aspect, the embodiment of the present application also disclose a kind of electronic equipment, including memory, processor and are stored in institute
The computer program that can be run on memory and on a processor is stated, the processor realizes this when executing the computer program
Apply for the sequence neural network model training method of commodity described in embodiment.
Fourth aspect, the embodiment of the present application provide a kind of computer readable storage medium, are stored thereon with computer journey
Sequence, when which is executed by processor the step of the sequence of commodity disclosed in the embodiment of the present application neural network model training method.
The sequence neural network model training method of commodity disclosed in the embodiment of the present application passes through businessman row that will be trained in advance
The department pattern parameter of sequence neural network model is migrated to commodity sequence neural network model, according to the department pattern parameter
The parameter of part hidden layer in the first network of commodity sequence neural network model is initialized, and, commodity described in random initializtion
The parameter of each hidden layer in sequence neural network model in addition to the part hidden layer；The quotient of each commodity training sample is obtained respectively
Family's dimensional characteristics and commodity dimensional characteristics construct the sample data of corresponding commodity training sample；For each commodity training
Sample, using businessman's dimensional characteristics of the commodity training sample as the input of the first network, and by the commodity training
Input of the commodity dimensional characteristics of sample as second network, the training commodity sequence neural network model, has
Help training for promotion and obtains the training effectiveness of commodity sequence neural network model.
Detailed description of the invention
Technical solution in ord to more clearly illustrate embodiments of the present application, below will be in embodiment or description of the prior art
Required attached drawing is briefly described, it should be apparent that, the accompanying drawings in the following description is only some realities of the application
Example is applied, it for those of ordinary skill in the art, without any creative labor, can also be attached according to these
Figure obtains other attached drawings.
Fig. 1 is the commodity sequence neural network model training method flow chart of the embodiment of the present application one；
Fig. 2 is businessman's sequence Artificial Neural Network Structures schematic diagram in the embodiment of the present application one；
Fig. 3 is the commodity sequence Artificial Neural Network Structures schematic diagram in the embodiment of the present application one；
Fig. 4 is the commodity sequence neural network model training device structural schematic diagram of the embodiment of the present application two.
Specific embodiment
Below in conjunction with the attached drawing in the embodiment of the present application, technical solutions in the embodiments of the present application carries out clear, complete
Site preparation description, it is clear that described embodiment is some embodiments of the present application, instead of all the embodiments.Based on this Shen
Please in embodiment, every other implementation obtained by those of ordinary skill in the art without making creative efforts
Example, shall fall in the protection scope of this application.
Embodiment one
A kind of commodity sequence neural network model training method disclosed in the embodiment of the present application, as shown in Figure 1, this method packet
It includes: step 110 to step 130.
Step 110, the department pattern parameter of businessman's sequence neural network model trained in advance is migrated to commodity and is sorted
Neural network model, in the first network of the sequence neural network model of the commodity according to the department pattern parameter initialization
The parameter of part hidden layer, and, in the sequence neural network model of commodity described in random initializtion in addition to the part hidden layer
The parameter of each hidden layer.
The sequence neural network model of businessman described in the embodiment of the present application is that preparatory training is completed, on-line operation
, the training process of businessman's sequence neural network model includes: to obtain businessman's dimension of several businessman's training samples respectively
Feature constructs the sample data of corresponding businessman's training sample；Based on the sample data of businessman's training sample, the training quotient
Family's sequence neural network model.
In some embodiments of the present application, as shown in Fig. 2, the businessman sorts, the network structure of neural network model is by spy
Assemble-publish code layer 210 and feature coding layer 220, feature input layer 230 and at least two hidden layers 240 and output layer 250 are constituted,
In, feature coding layer 210 and feature coding layer 220 are met the requirements for encoding to different types of input feature vector
Feature vector, and be input to feature input layer 230；The feature input layer 230 is for spelling the feature vector of input
It connects, obtains spliced feature vector, and spliced feature vector is input to hidden layer 240；At least two hidden layer 240
For carrying out eigentransformation processing to the feature vector of input, and the training objective based on model carries out Feature Mapping；Finally, defeated
The vector that layer 250 is exported based on hidden layer 240 out, obtains the output of model, if user is to lower single probability of specified businessman.
In some embodiments of the present application, businessman's sequence neural network model may include 3 hidden layers, wherein
Two hidden layers connecting with the feature input layer 230 are used to learn to be input to the feature of businessman's sequence neural network model
Public characteristic, the hidden layer being connect with output layer 250 be used for according to the businessman sort neural network model output target into
Row further study, obtains the depth characteristic with output object matching.
In some embodiments of the present application, before the training businessman sorts neural network model, first according to businessman
Data construct businessman's training sample.For example, obtaining lower forms data, browsing data etc. of the user in businessman；Then, according to acquisition
Lower forms data constructs positive sample, businessman's dimensional characteristics is extracted, as the sample data of businessman's training sample, according to the browsing of acquisition
Data construct negative sample, extract businessman's dimensional characteristics, as the sample data of businessman's training sample, sample label is used to indicate down
Single probability can be set to 1 or 0.Wherein, businessman's dimensional characteristics include: user characteristics, businessman feature, scene characteristic, use
It is any one or more in family and businessman's interaction feature.
In some embodiments of the present application, when the training businessman sorts neural network model, by every businessman's training
Sample is separately input into the businessman and sorts neural network model, by the businessman sort neural network model to training sample into
Row study, training pattern parameter.
Wherein, businessman's dimensional characteristics include discrete features.For example, the user characteristics may include: gender, it is whether white
The information such as neck, income rank；Businessman feature may include: merchant identification, whether brand marketers, whether high-quality businessman, place city
The information such as city；Scene characteristic may include: the information such as lower single period, user client type, commodity category；User and businessman
Interaction feature may include: whether place an order in businessman, the information such as quantity of order, browsing time.
In some embodiments of the present application, businessman's dimensional characteristics further include continuous feature.For example, the user characteristics
It may include: 7 days/30 days/90 days Order average prices；Businessman feature may include: 7 days/30 days/90 days sales volumes, moon sale
Volume, businessman's order average price, independent visitor's number of businessman, website browsing amount, clicking rate, conversion ratio, businessman's comment point etc.；User and
Businessman's interaction feature may include: distance of the user to businessman.
Several businessman's training samples can be constructed by above-mentioned side.
In other embodiments of the application, businessman's training sample can also be constructed using other methods, or obtain
The training sample of other businessman's clicking rate prediction models of training, as businessman's training sample.Alternatively, selecting other businessman's dimensions special
Sign is used as sample data, and it is not the restriction to specific implementation that the application provided, which is only some preferable examples,.
During businessman's sequence neural network model described based on businessman's training sample training, the businessman row
Sequence neural network model first encodes the sample data of input.For example, leading to for discrete features in input sample data
It crosses feature coding layer 210 to be encoded, obtains the feature vector of designated length, the coding parameter of feature coding layer 210 can lead to
Cross training determination；It for feature continuous in sample data, is normalized by feature coding layer 220, by different dimensions
Characteristic value be normalized between [0,1].In some embodiments of the present application, MinMax normalization mode can be used, by turning
Exchange the letters number x‘=(x-min)/(max-min) is normalized, wherein min indicates the minimum value of this feature, and max is indicated
The maximum value of this feature, x indicate the current value of this feature, x‘Feature vector value after indicating feature normalization.
Later, the feature vector that the feature input layer 230 inputs feature coding layer 220 and 210 is spliced, and is obtained
Spliced feature vector, and spliced feature vector is input to hidden layer 240.With businessman sequence neural network model
Illustrate including at least three hidden layers, close to 240 pairs of at least two hidden layers of the feature input layer 230 inputs feature vectors into
Row eigentransformation processing, and the training objective based on model carries out Feature Mapping, a hidden layer close to output layer 250 is according to institute
The output target for stating businessman's sequence neural network model is further learnt, and the depth characteristic with output object matching is obtained.
Finally, the depth characteristic that the output layer 250 exports hidden layer 240 is input to loss function, computation model it is pre-
The penalty values (i.e. prediction error) between result and the sample label of training sample are surveyed, and to predict the minimum target of error, it is excellent
Change model parameter, completes the training of businessman's sequence neural network model.
Based on trained businessman's sequence neural network model, is migrated by model parameter, further train commodity
Sort neural network model.Referring to Fig. 3, commodity described in the embodiment of the present application sort neural network model including being arranged parallel
First network 310, the second network 320, merging features layer 330 and output layer 340.Wherein, the first network 310 for pair
Business Information carries out characteristic processing；Second network 320 is used to carry out characteristic processing to merchandise news；The merging features layer
330 are used for first network 3 to output layer 340；Output layer 340 is for calculating characteristic loss and output model prediction result.Institute
The network structure and businessman's sequence neural network model for stating first network 310 match.
In the present embodiment, with the base of the network structure of first network 310 and businessman shown in Fig. 2 sequence neural network model
For plinth network is identical, illustrate the structure and training process of commodity sequence neural network model.
Referring to Fig. 3, the first network 310 includes: second hidden layer 3101 and at least one first hidden layer 3102, institute
It states at least one first hidden layer 3102 and is connected to the commodity sequence neural network mould in turn (as included 2 the first hidden layers 3102)
The input side of type, second hidden layer 3101 are connect with the outlet side of commodity sequence neural network model, and described second is hidden
Layer 3101 exports user to the preference of businessman after carrying out conversion process for the output at least one first hidden layer 3102
Vector.The first network 310 further includes feature input network structure 3104, discrete features vectorization network structure 3105 and connects
Continuous feature normalization network structure 3106, wherein discrete features vectorization network structure 3105 and continuous feature normalization network
Feature coding layer 210 and 220 of the specific structure and the course of work of structure 3106 referring to businessman's sequence neural network model, feature
Input network structure 3104 specific structure and the course of work referring to businessman sort neural network model feature input layer 230,
The specific structure and the course of work of first hidden layer 3102,3103 and the second hidden layer 3101 are referring to businessman's sequence neural network model
Corresponding hidden layer.
Second network 320 includes: at least two hidden layers, for example, second network 320 includes third hidden layer
3202,3203 and the 4th hidden layer 3201, second network 320 further include: feature inputs network structure 3204 and discrete spy
Levy vectorization network structure 3205 and/or continuous feature normalization network structure 3206.Wherein, the discrete features vectorization net
Network structure 3205 is used to the discrete features in the commodity dimensional characteristics carrying out vector expression, obtains the first commodity dimensional characteristics
Vector；The continuous feature normalization network structure 3206 is used to the continuous feature in the commodity dimensional characteristics carrying out normalizing
Change processing, obtains the second commodity dimensional characteristics vector；Feature inputs network structure 3204 and is used for the first commodity dimension spy
The second commodity dimensional characteristics vector described in vector sum is levied after splicing, is input to the hidden layer of second network 320, such as third
Hidden layer 3203；The third hidden layer 3203 and 3202 is used to carry out eigentransformation processing to the feature vector obtained after splicing, and
Training objective based on model carries out Feature Mapping, and the 4th hidden layer 3201 is used for the neural network mould that sorts according to the commodity
The output target of type is further learnt, and the depth characteristic with output object matching is obtained.
In some embodiments of the present application, the sequence of the commodity according to the department pattern parameter initialization nerve
The parameter of part hidden layer in the first network of network model, comprising: according to the portion of businessman sequence neural network model
Sub-model parameter initializes the parameter of the part hidden layer of close input side in first network in commodity sequence neural network model.
For example, when the businessman sort neural network model include N number of hidden layer, commodity sequence neural network model first network in
When including corresponding N number of hidden layer, neural network model can be sorted by the businessman in N-1 hidden layer of input side
Corresponding hidden layer in the first network of the model parameter initialization commodity sequence neural network model of any one or more hidden layers
Model parameter.For example, the businessman sort neural network model in N-1 hidden layer of input side any one or it is more
The value of the model parameter of a hidden layer copies in the first network of commodity sequence neural network model close to the corresponding of outlet side
Hidden layer to the parameter assignment of the corresponding hidden layer, and in the first network of commodity sequence neural network model close to outlet side
Hidden layer parameter carry out random initializtion.
In some embodiments of the present application, as shown in figure 3, the first network 310 includes second hidden layer 3101
With at least one the first hidden layer 3102 and 3103, at least one described first hidden layer 3103 and 3102 is connected to the commodity in turn
The input side of sequence neural network model, second hidden layer 3101 and the outlet side of commodity sequence neural network model connect
It connects, second hidden layer 3101 is used to export user after the output at least one first hidden layer 3102 carries out conversion process
To the preference vector of businessman.The neural network mould correspondingly, the commodity according to the department pattern parameter initialization sort
The parameter of part hidden layer in the first network of type, comprising: according to the department pattern of businessman sequence neural network model
Parameter initializes the parameter of at least one first hidden layer, and, the parameter of second hidden layer is carried out random
Initialization.
The model parameter of neural network model for example, the businessman that can be obtained according to abovementioned steps training sorts, just
Close to the portion of feature input network structure 3104 in the first network 310 of beginningization commodity sequence neural network model shown in Fig. 3
Divide the parameter of hidden layer, such as model parameter of the first hidden layer 3102 and/or the first hidden layer 3103.And for institute in first network 310
The parameter of the hidden layer except the hidden layer of part is stated, then carries out random initializtion.And it is then complete for the model parameter in the second network 320
Portion carries out random initializtion.
Because the input feature vector learnt close to the hidden layer of input side belongs to public feature, use can be shared, still
It is connected to close to the hidden layer and output layer of output layer, the depth characteristic learnt close to the hidden layer of output layer can determine final
Training objective, since the target of businessman's sequence and commodity sequence is different, the first of commodity sequence neural network model
Network carries out random initializtion effect close to the hidden layer of output layer can be more preferable, and whether the depth characteristic of output will be related to user right
Commodity place an order, and the network parameter by learning public characteristic in businessman's sequence neural network model initializes commodity sequence nerve
The parameter of corresponding network in network model, commodity sequence neural network model can be restrained than faster, lift scheme training effectiveness.
Step 120, the businessman's dimensional characteristics and commodity dimensional characteristics for obtaining each commodity training sample respectively, construct corresponding quotient
The sample data of product training sample.
Before training commodity sequence neural network model, commodity training sample is constructed according to commodity data first.For example,
User is obtained to the lower forms datas of commodity, browsing data etc.；Then, positive sample is constructed according to the lower forms data of acquisition, extracts quotient
Family's dimensional characteristics and commodity dimensional characteristics construct negative as the sample data of commodity training sample according to the browsing data of acquisition
Sample extracts businessman's dimensional characteristics and commodity dimensional characteristics, and as the sample data of commodity training sample, sample label is for referring to
Show lower single probability, can be set to 1 or 0.
In some embodiments of the present application, businessman's dimensional characteristics include: user characteristics, businessman feature, scene characteristic,
It is any one or more in user and businessman's interaction feature.Based on obtaining user to the lower forms datas of commodity, browsing data etc.,
Specific embodiment of the specific embodiment of businessman's dimensional characteristics referring to building businessman's sequence neural network model when is extracted, this
Place repeats no more.
In some embodiments of the present application, the commodity dimensional characteristics include: user characteristics, standard product unit spy
It is any one or more in sign, scene characteristic, user and standard product unit interaction feature.Wherein, commodity dimensional characteristics
It further comprise discrete features.For example, the user characteristics may include: gender, whether white collar, the income information such as rank；Mark
Standardization product unit feature may include: the information such as the price of commodity, category；Scene characteristic may include: lower single period, use
Family client type, etc. information；User and standard product unit interaction feature may include: whether place an order to commodity, order
The information such as quantity, browsing time.
In some embodiments of the present application, the commodity dimensional characteristics further include continuous feature.For example, the user characteristics
It may include: 7 days/30 days/90 days Order average price standard product element characteristics can also include: 7 days/30 days/90 days
Sales volume, independent visitor's number of commodity, website browsing amount, clicking rate, conversion ratio；User and standard product unit interaction feature can
To include: clicking rate, lower single quantity etc. of the user to commodity.
Several commodity training samples can be constructed by above-mentioned side.
Step 130, for each commodity training sample, using businessman's dimensional characteristics of the commodity training sample as
The input of the first network, and using the commodity dimensional characteristics of the commodity training sample as the defeated of second network
Enter, the training commodity sequence neural network model.
Next, the commodity training sample based on building, the training commodity sequence neural network model.
By abovementioned steps it is found that the sample data of commodity training sample includes businessman's dimensional characteristics and commodity dimensional characteristics two
Part is respectively used to study user couple correspondingly, commodity sequence neural network model includes first network and the second network
The feature of the feature of businessman and user to commodity.Therefore, when carrying out model training, the quotient by the commodity training sample
Input of family's dimensional characteristics as the first network, and using the commodity dimensional characteristics of the commodity training sample as institute
The step of stating the input of the second network, training the commodity sequence neural network model, comprising: by the first network to defeated
The businessman's dimensional characteristics entered carry out eigentransformation, obtain user to the preference vector of businessman；And pass through second network pair
The commodity dimensional characteristics of input carry out eigentransformation, obtain user to the preference vector of commodity；To the user to the inclined of businessman
User described in good vector sum splices the preference vector of commodity；The vector that splicing is obtained sorts neural as the commodity
The input of the activation primitive of network model, with the output of activation primitive sample label corresponding with the commodity training sample
The minimum target of difference, adjust the parameter of each hidden layer in the first network and second network until reach the mesh
Mark, then the commodity sequence neural network model training is completed, wherein the sample label is used to indicate the sample data pair
Whether the commodity answered place an order.
By taking commodity shown in Fig. 3 sort neural network model as an example, for a commodity training sample, by its sample data
In businessman's dimensional characteristics in discrete features be input to first network 310 discrete features vectorization network structure 3105 carry out
Feature coding obtains vector F1；Continuous feature in businessman's dimensional characteristics in its sample data is input to first network 310
Continuous feature normalization network structure 3106 be normalized, obtain vector F2；Then, feature inputs network structure
3104 couples of vectors F1 and F2 splice, and obtain vector F3；First hidden layer 3103 and the first hidden layer 3102 by vector F3 into
Row study, obtains vector F4, later, the second hidden layer 3101 obtains user to businessman to vector F4 further progress deep learning
Preference vector.
Meanwhile the discrete features in the commodity dimensional characteristics in its sample data are input to the discrete of the second network 320
Feature vector network structure 3205 carries out feature coding, obtains vector F6；It will be in the commodity dimensional characteristics in its sample data
Continuous feature be input to the continuous feature normalization network structure 3206 of the second network 320 and be normalized, obtain to
Measure F7；Then, feature input network structure 3204 splices vector F6 and F7, obtains vector F8；3203 He of third hidden layer
Third hidden layer 3202 obtains vector F9 by learning to vector F8, later, the 4th hidden layer 3201 to vector F9 further into
Row deep learning obtains user to the preference vector of commodity.For example, by user to the preference vector of businessman [a1, a2, a3, a4,
A5] and user two data of preference vector [b1, b2, b3, b4, b5] of commodity is spliced, obtain vector [a1, a2, a3,
a4,a5,b1,b2,b3,b4,b5]。
The concrete mode encoded to discrete features sorts the coding mode in neural network model referring to businessman, or existing
There is technology；The concrete mode that continuous feature is normalized sorts the normalization mode in neural network model referring to businessman,
Or the prior art, details are not described herein again.
Next, 330 couples of user of merging features layer carry out the preference vector of commodity and user to the preference vector of businessman
Splicing obtains the feature vector comprising user to commodity and businessman's preference information, and the vector that splicing is obtained is as the commodity
The input of the activation primitive of sequence neural network model, the parameter of training pattern, the output of activation primitive are then the commodity row
The output of sequence neural network model, for example, output is the numerical value between one 0 to 1, for indicating click of the user to commodity
Rate, clicking rate then can be used as commodity sequence index score.
During model training, with the output of activation primitive sample label corresponding with the commodity training sample
The minimum target of difference, the first network and second network are adjusted by Adam optimization algorithm based on backpropagation
In each network structure parameter, until reach the target, then the commodity sequence neural network model training is completed.Wherein,
The sample label is used to indicate whether the corresponding commodity of the sample data place an order, for example, sample label is 1 representative sample number
According to the feature extracted for user to forms data under commodity；Sample label is that 0 representative sample data are user to goods browse data
The feature of extraction.
The sequence neural network model training method of commodity disclosed in the embodiment of the present application passes through businessman row that will be trained in advance
The department pattern parameter of sequence neural network model is migrated to commodity sequence neural network model, according to the department pattern parameter
The parameter of part hidden layer in the first network of the commodity sequence neural network model is initialized, and, described in random initializtion
The parameter of each hidden layer in commodity sequence neural network model in addition to the part hidden layer；Each commodity training sample is obtained respectively
Businessman's dimensional characteristics and commodity dimensional characteristics, construct corresponding commodity training sample sample data；For each commodity
Training sample, using businessman's dimensional characteristics of the commodity training sample as the input of the first network, and by the commodity
Input of the commodity dimensional characteristics of training sample as second network, the training commodity sequence neural network mould
Type facilitates training for promotion and obtains the training effectiveness of commodity sequence neural network model.
On the other hand, since commodity training sample haves the characteristics that sample is sparse, merely according to commodity training sample
Training commodity order models will lead to the problem of the sequence index score inaccuracy of output.But the sample size of businessman's training sample
It is very big, accuracy height more abundant by parameter learning in businessman's sequence neural network model of businessman's training sample training.Cause
This, carries out businessman using Fine-tune technology and sorts neural network model to the migration of commodity sequence neural network model, realize
Department pattern parameter sharing, can be sorted index score with the commodity for the commodity sequence neural network model output that training for promotion obtains
Accuracy, effectively make up the few deficiency of commodity original amount.
Embodiment two
A kind of commodity sequence neural network model training device disclosed in the present embodiment, as shown in figure 3, the commodity sort
Neural network model includes parallel arrangement of first network and the second network, and the network structure of the first network and businessman sort
Neural network model matching, as shown in figure 4, described device includes:
Commodity order models initialization module 410, the part of businessman's sequence neural network model for that will train in advance
Model parameter is migrated to commodity sequence neural network model, is sorted with the commodity according to the department pattern parameter initialization refreshing
The parameter of part hidden layer in first network through network model, and, the sequence neural network model of commodity described in random initializtion
In each hidden layer in addition to the part hidden layer parameter；
Commodity training sample constructs module 420, for obtaining businessman's dimensional characteristics and the quotient of each commodity training sample respectively
Product dimensional characteristics construct the sample data of corresponding commodity training sample；
Commodity order models training module 430 is used for for each commodity training sample, by the commodity training sample
The input of this businessman's dimensional characteristics as the first network, and by the commodity dimensional characteristics of the commodity training sample
As the input of second network, the training commodity sequence neural network model.
Wherein, the training process of businessman's sequence neural network model includes: to obtain several businessman's training samples respectively
Businessman's dimensional characteristics, construct the sample data of corresponding businessman's training sample；
Based on the sample data of businessman's training sample, training businessman's sequence neural network model.
The training process of businessman's sequence neural network model is referring to embodiment one, and this embodiment is not repeated.
In some embodiments of the present application, the commodity order models training module 430 is further used for:
Eigentransformation is carried out by businessman dimensional characteristics of the first network to input, obtains user to the preference of businessman
Vector；And eigentransformation is carried out by commodity dimensional characteristics of second network to input, user is obtained to the inclined of commodity
Good vector；
The user splices the preference vector of businessman and the user to the preference vector of commodity；
The vector that splicing is obtained is swashed as the input of the activation primitive of commodity sequence neural network model with described
The minimum target of difference of the output sample label corresponding with the commodity training sample of function living, adjusts the first network
Parameter with each hidden layer in second network is until reach the target, then the commodity sequence neural network model has been trained
At, wherein the sample label is used to indicate whether the corresponding commodity of the sample data place an order.
In some embodiments of the present application, second network further includes discrete features vectorization network structure, continuous
Feature normalization network structure, the discrete features vectorization network structure are used for the discrete spy in the commodity dimensional characteristics
Sign carries out vector expression, obtains the first commodity dimensional characteristics vector, the continuous feature normalization network structure is used for will be described
Continuous feature in commodity dimensional characteristics is normalized, and obtains the second commodity dimensional characteristics vector；First commodity
Second commodity dimensional characteristics vector described in dimensional characteristics vector sum is input to the hidden layer of second network after splicing.
In some embodiments of the present application, the sequence of the commodity according to the department pattern parameter initialization nerve
The parameter of part hidden layer in the first network of network model, comprising:
According to the department pattern parameter of businessman sequence neural network model, initialization commodity sequence neural network
Close to the parameter of the part hidden layer of input side in first network in model.
In some embodiments of the present application, the first network includes second hidden layer and at least one is first hidden
Layer, at least one described first hidden layer are connected to the input side of the commodity sequence neural network model in turn, and described second is hidden
Layer is connect with the outlet side of commodity sequence neural network model, and second hidden layer is at least one to be first hidden to described
The output of layer carries out after conversion process output user to the preference vector of businessman, and the commodity order models initialization module is into one
Step is used for:
According to the department pattern parameter of businessman sequence neural network model at least one described first hidden layer
Parameter initialized, and, random initializtion is carried out to the parameter of second hidden layer；And to second network
Parameter carry out random initializtion.
Because the input feature vector learnt close to the hidden layer of input side belongs to public feature, use can be shared, still
It is connected to close to the hidden layer and output layer of output layer, the depth characteristic learnt close to the hidden layer of output layer can determine final
Training objective, since the target of businessman's sequence and commodity sequence is different, the first of commodity sequence neural network model
Network carries out random initializtion effect close to the hidden layer of output layer can be more preferable, and whether the depth characteristic of output will be related to user right
Commodity place an order, and the network parameter by learning public characteristic in businessman's sequence neural network model initializes commodity sequence nerve
The parameter of corresponding network in network model, commodity sequence neural network model can be restrained than faster, lift scheme training effectiveness.
In some embodiments of the present application, businessman's dimensional characteristics include: user characteristics, businessman feature, scene spy
Any one or more in sign, user and businessman's interaction feature, the commodity dimensional characteristics include user characteristics, standardization production
It is any one or more in article unit feature, scene characteristic, user and standard product unit interaction feature.
The sequence neural network model training device of commodity disclosed in the embodiment of the present application, for realizing the embodiment of the present application one
Described in commodity sequence neural network model training method each step, the specific embodiment of each module of device is referring to phase
Step is answered, details are not described herein again.
The sequence neural network model training device of commodity disclosed in the embodiment of the present application passes through businessman row that will be trained in advance
The department pattern parameter of sequence neural network model is migrated to commodity sequence neural network model, according to the department pattern parameter
The parameter of part hidden layer in the first network of commodity sequence neural network model is initialized, and, commodity described in random initializtion
The parameter of each hidden layer in sequence neural network model in addition to the part hidden layer；The quotient of each commodity training sample is obtained respectively
Family's dimensional characteristics and commodity dimensional characteristics construct the sample data of corresponding commodity training sample；For each commodity training
Sample, using businessman's dimensional characteristics of the commodity training sample as the input of the first network, and by the commodity training
Input of the commodity dimensional characteristics of sample as second network, the training commodity sequence neural network model, has
Help training for promotion and obtains the efficiency of commodity sequence neural network model.
On the other hand, since commodity training sample haves the characteristics that sample is sparse, merely according to commodity training sample
Training commodity order models will lead to the problem of the sequence index score inaccuracy of output.But the sample size of businessman's training sample
It is very big, accuracy height more abundant by parameter learning in businessman's sequence neural network model of businessman's training sample training.Cause
This, carries out businessman using Fine-tune technology and sorts neural network model to the migration of commodity sequence neural network model, realize
Department pattern parameter sharing, can be sorted index score with the commodity for the commodity sequence neural network model output that training for promotion obtains
Accuracy, effectively make up the few deficiency of commodity original amount.
Correspondingly, disclosed herein as well is a kind of electronic equipment, including memory, processor and it is stored in the memory
Computer program that is upper and can running on a processor, the processor are realized when executing the computer program as the application is real
Apply the sequence neural network model training method of commodity described in example one.The electronic equipment can be PC machine, mobile terminal, individual
Digital assistants, tablet computer etc..
Disclosed herein as well is a kind of computer readable storage mediums, are stored thereon with computer program, which is located
Manage the step of commodity sequence neural network model training method as described in the embodiment of the present application one is realized when device executes.
All the embodiments in this specification are described in a progressive manner, the highlights of each of the examples are with
The difference of other embodiments, the same or similar parts between the embodiments can be referred to each other.For Installation practice
For, since it is basically similar to the method embodiment, so being described relatively simple, referring to the portion of embodiment of the method in place of correlation
It defends oneself bright.
Detailed Jie has been carried out to a kind of commodity sequence neural network model training method provided by the present application and device above
It continues, specific examples are used herein to illustrate the principle and implementation manner of the present application, and the explanation of above embodiments is only
It is to be used to help understand the method for this application and its core ideas；At the same time, for those skilled in the art, according to this Shen
Thought please, there will be changes in the specific implementation manner and application range, in conclusion the content of the present specification should not manage
Solution is the limitation to the application.
Through the above description of the embodiments, those skilled in the art can be understood that each embodiment can
It realizes by means of software and necessary general hardware platform, naturally it is also possible to pass through hardware realization.Based on such reason
Solution, substantially the part that contributes to existing technology can embody above-mentioned technical proposal in the form of software products in other words
Come, which may be stored in a computer readable storage medium, such as ROM/RAM, magnetic disk, CD, including
Some instructions are used so that a computer equipment (can be personal computer, server or the network equipment etc.) executes respectively
Method described in certain parts of a embodiment or embodiment.
Claims (13)
- A kind of neural network model training method 1. commodity sort, which is characterized in that the commodity sequence neural network model packet Include parallel arrangement of first network and the second network, the network structure of the first network and businessman's sequence neural network model Match, which comprisesThe department pattern parameter of businessman's sequence neural network model trained in advance is migrated to commodity sequence neural network model, With the ginseng of part hidden layer in the first network of the sequence neural network model of the commodity according to the department pattern parameter initialization Number, and, the parameter of each hidden layer in the sequence neural network model of commodity described in random initializtion in addition to the part hidden layer；The businessman's dimensional characteristics and commodity dimensional characteristics of each commodity training sample are obtained respectively, construction corresponding commodity training sample Sample data；For each commodity training sample, using businessman's dimensional characteristics of the commodity training sample as the first network Input, and using the commodity dimensional characteristics of the commodity training sample as the input of second network, described in training Commodity sequence neural network model.
- 2. the method according to claim 1, wherein businessman's dimensional characteristics by the commodity training sample As the input of the first network, and using the commodity dimensional characteristics of the commodity training sample as second network Input, the step of training commodity sequence neural network model, comprising:Eigentransformation is carried out to businessman's dimensional characteristics of input by the first network, obtain user to the preference of businessman to Amount；And eigentransformation is carried out by commodity dimensional characteristics of second network to input, user is obtained to the preference of commodity Vector；The user splices the preference vector of businessman and the user to the preference vector of commodity；The vector that splicing is obtained is as the input of the activation primitive of commodity sequence neural network model, with the activation letter The minimum target of difference of several output sample labels corresponding with the commodity training sample, adjusts the first network and institute The parameter of each hidden layer in the second network is stated until reaching the target, then the commodity sequence neural network model training is completed, Wherein, the sample label is used to indicate whether the corresponding commodity of the sample data place an order.
- 3. according to the method described in claim 2, it is characterized in that, second network further includes discrete features vectorization network Structure, continuous feature normalization network structure, the discrete features vectorization network structure are used for the commodity dimensional characteristics In discrete features carry out vector expression, obtain the first commodity dimensional characteristics vector, the continuous feature normalization network structure For the continuous feature in the commodity dimensional characteristics to be normalized, the second commodity dimensional characteristics vector is obtained；Institute The second commodity dimensional characteristics vector described in the first commodity dimensional characteristics vector sum is stated after splicing, is input to second network Hidden layer.
- 4. method according to any one of claims 1 to 3, which is characterized in that described according at the beginning of the department pattern parameter In the first network of the beginningization commodity sequence neural network model the step of parameter of part hidden layer, comprising:According to the department pattern parameter of businessman sequence neural network model, initialization commodity sequence neural network model Close to the parameter of the part hidden layer of input side in middle first network.
- 5. method according to any one of claims 1 to 3, which is characterized in that the first network includes one second hidden Layer and at least one first hidden layer, at least one described first hidden layer are connected to the commodity sequence neural network model in turn The outlet side of input side, second hidden layer and the commodity sequence neural network model connect, and second hidden layer is used for pair Output user is described according to the preference vector of businessman after the output of at least one first hidden layer carries out conversion process The parameter of part hidden layer in the first network of the sequence neural network model of commodity described in department pattern parameter initialization, and, with Machine initializes the step of parameter of each hidden layer in the commodity sequence neural network model in addition to the part hidden layer, packet It includes:According to the department pattern parameter of businessman sequence neural network model to the ginseng of at least one first hidden layer Number is initialized, and, random initializtion is carried out to the parameter of second hidden layer；And the ginseng to second network Number carries out random initializtion.
- 6. method according to any one of claims 1 to 3, which is characterized in that businessman's dimensional characteristics include: user spy It is any one or more in sign, businessman feature, scene characteristic, user and businessman's interaction feature, the commodity dimensional characteristics packet It includes any one in user characteristics, standard product element characteristic, scene characteristic, user and standard product unit interaction feature Item is multinomial.
- A kind of neural network model training device 7. commodity sort, which is characterized in that the commodity sequence neural network model packet Include parallel arrangement of first network and the second network, the network structure of the first network and businessman's sequence neural network model Match, described device includes:Commodity order models initialization module, the department pattern parameter of businessman's sequence neural network model for that will train in advance It migrates to commodity sequence neural network model, with the sequence neural network mould of the commodity according to the department pattern parameter initialization The parameter of part hidden layer in the first network of type, and, except described in the sequence neural network model of commodity described in random initializtion The parameter of each hidden layer except the hidden layer of part；Commodity training sample constructs module, and the businessman's dimensional characteristics and commodity dimension for obtaining each commodity training sample respectively are special Sign constructs the sample data of corresponding commodity training sample；Commodity order models training module is used for for each commodity training sample, by the quotient of the commodity training sample Input of family's dimensional characteristics as the first network, and using the commodity dimensional characteristics of the commodity training sample as institute State the input of the second network, the training commodity sequence neural network model.
- 8. device according to claim 7, which is characterized in that the commodity order models training module is further used for:Eigentransformation is carried out to businessman's dimensional characteristics of input by the first network, obtain user to the preference of businessman to Amount；And eigentransformation is carried out by commodity dimensional characteristics of second network to input, user is obtained to the preference of commodity Vector；The user splices the preference vector of businessman and the user to the preference vector of commodity；The vector that splicing is obtained is as the input of the activation primitive of commodity sequence neural network model, with the activation letter The minimum target of difference of several output sample labels corresponding with the commodity training sample, adjusts the first network and institute The parameter of each hidden layer in the second network is stated until reaching the target, then the commodity sequence neural network model training is completed, Wherein, the sample label is used to indicate whether the corresponding commodity of the sample data place an order.
- 9. device according to claim 8, which is characterized in that second network further includes discrete features vectorization network Structure, continuous feature normalization network structure, the discrete features vectorization network structure are used for the commodity dimensional characteristics In discrete features carry out vector expression, obtain the first commodity dimensional characteristics vector, the continuous feature normalization network structure For the continuous feature in the commodity dimensional characteristics to be normalized, the second commodity dimensional characteristics vector is obtained；Institute The second commodity dimensional characteristics vector described in the first commodity dimensional characteristics vector sum is stated after splicing, is input to second network Hidden layer.
- 10. device according to any one of claims 7 to 9, which is characterized in that described according at the beginning of the department pattern parameter The parameter of part hidden layer in the first network of the beginningization commodity sequence neural network model, comprising:According to the department pattern parameter of businessman sequence neural network model, initialization commodity sequence neural network model Close to the parameter of the part hidden layer of input side in middle first network.
- 11. device according to any one of claims 7 to 9, which is characterized in that the first network includes one second hidden Layer and at least one first hidden layer, at least one described first hidden layer are connected to the commodity sequence neural network model in turn The outlet side of input side, second hidden layer and the commodity sequence neural network model connect, and second hidden layer is used for pair The output of at least one first hidden layer carries out preference vector of the output user to businessman after conversion process, the commodity sequence Model initialization module is further used for:According to the department pattern parameter of businessman sequence neural network model to the ginseng of at least one first hidden layer Number is initialized, and, random initializtion is carried out to the parameter of second hidden layer；And the ginseng to second network Number carries out random initializtion.
- 12. a kind of electronic equipment, including memory, processor and it is stored on the memory and can runs on a processor Computer program, which is characterized in that the processor realizes claim 1 to 6 any one when executing the computer program The commodity sequence neural network model training method.
- 13. a kind of computer readable storage medium, is stored thereon with computer program, which is characterized in that the program is by processor The step of sequence neural network model training method of commodity described in claim 1 to 6 any one is realized when execution.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910656545.0A CN110503186A (en) | 2019-07-19 | 2019-07-19 | Commodity sequence neural network model training method, device, electronic equipment |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910656545.0A CN110503186A (en) | 2019-07-19 | 2019-07-19 | Commodity sequence neural network model training method, device, electronic equipment |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN110503186A true CN110503186A (en) | 2019-11-26 |
Family
ID=68586724
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910656545.0A Pending CN110503186A (en) | 2019-07-19 | 2019-07-19 | Commodity sequence neural network model training method, device, electronic equipment |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN110503186A (en) |
Cited By (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN112132281A (en) * | 2020-09-29 | 2020-12-25 | 腾讯科技（深圳）有限公司 | Model training method, device, server and medium based on artificial intelligence |
| CN112465042A (en) * | 2020-12-02 | 2021-03-09 | 中国联合网络通信集团有限公司 | Generation method and device of classification network model |
| CN113065635A (en) * | 2021-02-27 | 2021-07-02 | 华为技术有限公司 | Model training method, image enhancement method and device |
| CN113344127A (en) * | 2021-06-29 | 2021-09-03 | 中国平安人寿保险股份有限公司 | Data prediction method, device, equipment and storage medium |
-
2019
- 2019-07-19 CN CN201910656545.0A patent/CN110503186A/en active Pending
Cited By (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN112132281A (en) * | 2020-09-29 | 2020-12-25 | 腾讯科技（深圳）有限公司 | Model training method, device, server and medium based on artificial intelligence |
| CN112132281B (en) * | 2020-09-29 | 2024-04-26 | 腾讯科技（深圳）有限公司 | Model training method, device, server and medium based on artificial intelligence |
| CN112465042A (en) * | 2020-12-02 | 2021-03-09 | 中国联合网络通信集团有限公司 | Generation method and device of classification network model |
| CN112465042B (en) * | 2020-12-02 | 2023-10-24 | 中国联合网络通信集团有限公司 | Method and device for generating classified network model |
| CN113065635A (en) * | 2021-02-27 | 2021-07-02 | 华为技术有限公司 | Model training method, image enhancement method and device |
| CN113344127A (en) * | 2021-06-29 | 2021-09-03 | 中国平安人寿保险股份有限公司 | Data prediction method, device, equipment and storage medium |
| CN113344127B (en) * | 2021-06-29 | 2024-04-26 | 中国平安人寿保险股份有限公司 | Data prediction method, device, equipment and storage medium |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110503186A (en) | Commodity sequence neural network model training method, device, electronic equipment | |
| CN111178624B (en) | New product demand prediction method | |
| CN109190044A (en) | Personalized recommendation method, device, server and medium | |
| CN111784455A (en) | Article recommendation method and recommendation equipment | |
| CN108564414A (en) | Method of Commodity Recommendation based on behavior under line and system | |
| EP4198864A1 (en) | User feature-based page display method and apparatus, medium, and electronic device | |
| WO2023011382A1 (en) | Recommendation method, recommendation model training method, and related product | |
| CN104718547A (en) | Customer data analysis system | |
| CN110415002A (en) | Customer behavior prediction method and system | |
| CN112287238B (en) | User characteristic determination method and device, storage medium and electronic equipment | |
| CN108053322A (en) | The customer investment return evaluation method and system of vehicle | |
| CN106600039A (en) | E-commerce commodity information pushing optimization method | |
| CN106897776A (en) | A kind of continuous type latent structure method based on nominal attribute | |
| CN107688865A (en) | Identify the method and apparatus of potential high consumption user in online game | |
| CN112288471A (en) | Advertisement click rate prediction method based on user historical behavior sequence | |
| Kang | Combining grey relationship analysis and neural network to develop attractive automobile booth design | |
| CN112487109A (en) | Entity relationship extraction method, terminal and computer readable storage medium | |
| CN109034853A (en) | Similar users method, apparatus, medium and electronic equipment are found based on seed user | |
| CN113781139A (en) | Item recommendation method, item recommendation device, equipment and medium | |
| CN116737922A (en) | Tourist online comment fine granularity emotion analysis method and system | |
| CN112529415B (en) | Article scoring method based on combined multiple receptive field graph neural network | |
| Gupta et al. | House prices prediction using machine learning regression models | |
| CN108876422A (en) | For the method, apparatus of information popularization, electronic equipment and computer-readable medium | |
| CN105741143A (en) | Load characteristic and cluster analysis based electric power commodity pricing model establishment method | |
| CN111768218A (en) | Method and device for processing user interaction information |
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 |