CN107239532A - Data digging method and device - Google Patents
Data digging method and device Download PDFInfo
- Publication number
- CN107239532A CN107239532A CN201710397745.XA CN201710397745A CN107239532A CN 107239532 A CN107239532 A CN 107239532A CN 201710397745 A CN201710397745 A CN 201710397745A CN 107239532 A CN107239532 A CN 107239532A
- Authority
- CN
- China
- Prior art keywords
- target
- cnn
- content
- statistics
- vector
- 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.)
- Granted
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/20—Information retrieval; Database structures therefor; File system structures therefor of structured data, e.g. relational data
- G06F16/24—Querying
- G06F16/245—Query processing
- G06F16/2458—Special types of queries, e.g. statistical queries, fuzzy queries or distributed queries
- G06F16/2465—Query processing support for facilitating data mining operations in structured databases
-
- 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
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- Databases & Information Systems (AREA)
- Software Systems (AREA)
- Mathematical Physics (AREA)
- Computational Linguistics (AREA)
- Data Mining & Analysis (AREA)
- General Physics & Mathematics (AREA)
- Biophysics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Computing Systems (AREA)
- General Health & Medical Sciences (AREA)
- Evolutionary Computation (AREA)
- Biomedical Technology (AREA)
- Artificial Intelligence (AREA)
- Molecular Biology (AREA)
- Health & Medical Sciences (AREA)
- Fuzzy Systems (AREA)
- Probability & Statistics with Applications (AREA)
- Image Analysis (AREA)
- Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
Abstract
The invention discloses a kind of data digging method and device, it is related to data mining technology field.Method therein includes：By the content-data input content convolutional neural networks CNN of target, the content feature vector of target is obtained；The content feature vector of target is inputted into deep neural network DNN, the estimated data of target is obtained.It is efficient to obtain the more accurate estimated data of target so as in the case where the historical data of target need not be provided.
Description
Technical field
The present invention relates to data mining technology field, more particularly to a kind of data digging method and device.
Background technology
Logistic Scheduling is the important step of electric business website operation, and efficient Logistic Scheduling is pre- dependent on accurate Sales Volume of Commodity
Estimate.Therefore, optimize sales volume predictor method, to reduction electric business website operation cost, improve user's purchase experiences for it is most important.
, can be in advance to end article in each region if sales volume of the end article under object time and regional condition can be estimated accurately
Stock be scheduled and get ready the goods, or guiding consumer reasonable arrangement consumption in advance, so as to save cargo transport and stock
Cost, improves the purchase experiences of user.
Traditional sales volume predictor method, it usually needs the history sales volume data of accumulation end article, and pass through history sales volume number
Actual sales volume data of end article according to estimates.History sales volume data of this method excessively dependent on accumulation end article, no
It can draw inferences about other cases from one instance, therefore the accuracy estimated to new product sales volume is poor, easily causes new product goods and overstocks or get ready the goods not enough show
As in turn resulting in loss.
The sales volume predictor method also existed in the prior art is to analyze the attribute of end article, and utilizes end article
Attribute is associated with other commodity by end article.For example, the attribute of books includes author, classifying content etc..Then, never
Same attribute dimensions statistics end article, dependent merchandise or attribute polymerize the statistics sales volume data of commodity.Finally, according to obtaining
Each dimension counts sales volume data, and the statistics sales volume data of such as author and the classifying content, prediction obtain end article
Sales volume.The problem of presence of this method is, associated articles it is tactful relatively simple, and associated articles quantity limitation too
It is few.For the various electric business commodity of classification, brand, this method excessively simplifies problem, and subjectivity is strong so that estimated data is accurate
True property is relatively low.Simultaneously as needing the work of a large amount of manual features engineerings during estimating, such as extract and unite according to a variety of dimensions
Count, design various Rule Extraction dependent merchandises etc., these work are highly dependent on the experience of engineer, easily by subjectivity
Influence, the characteristic mass of extraction is uneven, there are a large amount of invalid or even reactive manual features, and to optimize therein
Feature extraction rule is very time-consuming, less efficient.
The content of the invention
The technical problem that the present invention is solved is, how in the case where that need not provide the historical data of target, high
The more accurate estimated data of acquisition target of effect.
One side according to embodiments of the present invention there is provided a kind of data digging method, including：By the content number of target
According to input content convolutional neural networks CNN, the content feature vector of target is obtained；The content feature vector of target is inputted into depth
Neutral net DNN, obtains the estimated data of target.
In certain embodiments, this method also includes：By the historical data input statistics CNN of target, the system of target is obtained
Count characteristic vector；Deep neural network DNN will be inputted after the content feature vector of target and the vector cascade of the statistical nature of target,
Obtain the estimated data of target.
In certain embodiments, by the content-data input content convolutional neural networks CNN of target, the content of target is obtained
Characteristic vector includes：Using text depth representing model, the text description of target is encoded to two-dimensional matrix；Two-dimensional matrix is defeated
Enter content CNN, obtain the Text eigenvector of target, and using Text eigenvector as target content feature vector.
In certain embodiments, by the content-data input content convolutional neural networks CNN of target, the content of target is obtained
Characteristic vector includes：It is multi-dimensional matrix by the coding of graphics of target according to the pixel value of the picture of target；Multi-dimensional matrix is inputted
Content CNN, obtain target picture feature vector, and using picture feature vector as target content feature vector.
In certain embodiments, content CNN includes first content CNN and the 2nd CNN；By in the content-data input of target
Hold convolutional neural networks CNN, obtaining the content feature vector of target includes：Using text depth representing model, by the text of target
This description is encoded to two-dimensional matrix；Two-dimensional matrix is inputted into first content CNN, the Text eigenvector of target is obtained；According to mesh
The pixel value of target picture, is multi-dimensional matrix by the coding of graphics of target；Multi-dimensional matrix is inputted into the second content CNN, mesh is obtained
Target picture feature vector；Using Text eigenvector and the vectorial content feature vector as target of picture feature.
In certain embodiments, by the historical data input statistics CNN of target, the statistical nature vector bag of target is obtained
Include：Target is cascaded as Time-domain Statistics sequence in the historical data of different periods；Time-domain Statistics sequence inputting is counted into CNN, obtained
The Time-domain Statistics characteristic vector of target is taken, and the statistical nature of Time-domain Statistics characteristic vector as target is vectorial.
In certain embodiments, by the historical data input statistics CNN of target, the statistical nature vector bag of target is obtained
Include：Multiple subregions are turned to by the statistical regions of target are discrete；By historical data of the target in different subregions of identical period
It is encoded to region statistical matrix；By region statistical matrix input statistics CNN, the region statistical nature vector of target is obtained, and will
Region statistical nature vector is vectorial as the statistical nature of target.
In certain embodiments, statistics CNN includes the first statistics CNN and the second statistics CNN；The historical data of target is defeated
Enter and count CNN, obtaining the statistical nature vector of target includes：History sales volume by target in different periods is cascaded as Time-domain Statistics
Sequence；Time-domain Statistics sequence inputting first is counted into CNN, the Time-domain Statistics characteristic vector of target is obtained；By the Statistical Area of target
Domain is discrete to turn to multiple subregions；Historical data of the target in different subregions of identical period is encoded to region statistical moment
Battle array；Region statistical matrix input second is counted into CNN, the region statistical nature vector of target is obtained；By Time-domain Statistics feature to
Amount and statistical nature vector of the statistical nature vector in region as target.
In certain embodiments, content CNN is made up of convolutional layer, pond layer, full articulamentum and output layer.
In certain embodiments, statistics CNN is made up of convolutional layer, pond layer and output layer.
In certain embodiments, DNN is made up of full articulamentum and output layer, and each neuron in full articulamentum performs non-
Linear activation primitive computing.
In certain embodiments, this method also includes：Calculate the real data of target and the mean square of estimated data is missed
Difference；Using back-propagation algorithm, the combination neural net that Mean Square Error is constituted from content CNN, statistics CNN and DNN
Output layer transmitted to input layer, with iteration update combination neural net in each layer parameter.
In certain embodiments, this method also includes：Using the content-data and the historical data of sample of sample, to content
The combination neural net of CNN, statistics CNN and DNN compositions is trained.
Other side according to embodiments of the present invention there is provided a kind of data mining device, including：Content feature vector
Acquisition module, is configured to the content-data input content convolutional neural networks CNN of target, obtain the content characteristic of target to
Amount；Estimated data acquisition module, is configured to the content feature vector of target inputting deep neural network DNN, obtains target
Estimated data.
In certain embodiments, the device also includes the vectorial acquisition module of statistical nature, is configured to the history number of target
CNN is counted according to input, the statistical nature vector of target is obtained；Estimated data acquisition module is additionally configured to：The content of target is special
Deep neural network DNN is inputted after the statistical nature vector cascade for levying vector sum target, the estimated data of target is obtained.
In certain embodiments, content feature vector acquisition module is configured to：Using text depth representing model, by target
Text description be encoded to two-dimensional matrix；By two-dimensional matrix input content CNN, the Text eigenvector of target is obtained, and by text
The vectorial content feature vector as target of eigen.
In certain embodiments, content feature vector acquisition module is configured to：According to the pixel value of the picture of target, by mesh
Target coding of graphics is multi-dimensional matrix；By multi-dimensional matrix input content CNN, the picture feature vector of target is obtained, and by picture
Characteristic vector as target content feature vector.
In certain embodiments, content CNN includes first content CNN and the second content CNN；Content feature vector obtains mould
Block is configured to：Using text depth representing model, the text description of target is encoded to two-dimensional matrix；By two-dimensional matrix input the
One content CNN, obtains the Text eigenvector of target；It is many by the coding of graphics of target according to the pixel value of the picture of target
Tie up matrix；Multi-dimensional matrix is inputted into the second content CNN, the picture feature vector of target is obtained；By Text eigenvector and picture
Characteristic vector as target content feature vector.
In certain embodiments, the vectorial acquisition module of statistical nature is configured to：By target different periods historical data
It is cascaded as Time-domain Statistics sequence；Time-domain Statistics sequence inputting is counted into CNN, the Time-domain Statistics characteristic vector of target is obtained, and will
Time-domain Statistics characteristic vector is vectorial as the statistical nature of target.
In certain embodiments, the vectorial acquisition module of statistical nature is configured to：By the statistical regions of target it is discrete turn to it is many
Sub-regions；Historical data of the target in different subregions of identical period is encoded to region statistical matrix；Region is counted
Input matrix counts CNN, obtains the region statistical nature vector of target, and using region statistical nature vector as target statistics
Characteristic vector.
In certain embodiments, statistics CNN includes the first statistics CNN and the second statistics CNN；Statistical nature vector obtains mould
Block is configured to：History sales volume by target in different periods is cascaded as Time-domain Statistics sequence；By Time-domain Statistics sequence inputting first
CNN is counted, the Time-domain Statistics characteristic vector of target is obtained；Multiple subregions are turned to by the statistical regions of target are discrete；By target
Historical data in different subregions of identical period is encoded to region statistical matrix；Region statistical matrix input second is counted
CNN, obtains the region statistical nature vector of target；It regard Time-domain Statistics characteristic vector and region statistical nature vector as target
Statistical nature vector.
In certain embodiments, content CNN is made up of convolutional layer, pond layer, full articulamentum and output layer.
In certain embodiments, statistics CNN is made up of convolutional layer, pond layer and output layer.
In certain embodiments, DNN is made up of full articulamentum and output layer, and each neuron in full articulamentum performs non-
Linear activation primitive computing, obtains the estimated data of target.
In certain embodiments, the device also includes parameter update module, is configured to：Calculate the real data of target with it is pre-
Estimate the Mean Square Error of data；Using back-propagation algorithm, by Mean Square Error from content CNN, statistics CNN and DNN
The output layer of the combination neural net of composition is transmitted to input layer, and the parameter of each layer in combination neural net is updated with iteration.
In certain embodiments, the device also includes training module, is configured to content-data and sample using sample
Historical data, is trained to content CNN, statistics CNN and DNN compositions combination neural net.
Another aspect according to embodiments of the present invention there is provided a kind of data mining device, including：Memory；And
The processor of memory is coupled to, processor is configured as the instruction based on storage in memory, performs foregoing data and dig
Pick method.
Another aspect according to embodiments of the present invention is computer-readable there is provided a kind of computer-readable recording medium
Storage medium is stored with computer instruction, and instruction realizes foregoing data digging method when being executed by processor.
The data digging method that the present invention is provided, can by count convolutional neural networks obtain target content characteristic to
Amount, and then pass through the estimated data of deep neural network acquisition target.So as in the feelings for the historical data that need not provide target
It is efficient to obtain the more accurate estimated data of target under condition.
By referring to the drawings to the detailed description of the exemplary embodiment of the present invention, further feature of the invention and its
Advantage will be made apparent from.
Brief description of the drawings
In order to illustrate more clearly about the embodiment of the present invention or technical scheme of the prior art, below will be to embodiment or existing
There is the accompanying drawing used required in technology description to be briefly described, it should be apparent that, drawings in the following description are only this
Some embodiments of invention, for those of ordinary skill in the art, without having to pay creative labor, may be used also
To obtain other accompanying drawings according to these accompanying drawings.
Fig. 1 shows the schematic diagram of the combination neural net of content CNN, statistics CNN and DNN compositions.
Fig. 2 shows content CNN schematic network structure.
Fig. 3 shows statistics CNN schematic network structure.
Fig. 4 shows DNN schematic network structure.
Fig. 5 shows the schematic flow sheet of one embodiment of data digging method of the present invention.
Fig. 6 shows the schematic flow sheet of another embodiment of data digging method of the present invention.
Fig. 7 shows the structural representation of one embodiment of data mining device of the present invention.
Fig. 8 shows the structure chart of another embodiment of data mining device of the present invention.
Fig. 9 shows the structure chart of another embodiment of data mining device of the present invention.
Embodiment
Below in conjunction with the accompanying drawing in the embodiment of the present invention, the technical scheme in the embodiment of the present invention is carried out clear, complete
Site preparation is described, it is clear that described embodiment is only a part of embodiment of the invention, rather than whole embodiments.Below
Description only actually at least one exemplary embodiment is illustrative, is never used as to the present invention and its application or makes
Any limitation.Based on the embodiment in the present invention, those of ordinary skill in the art are not making creative work premise
Lower obtained all other embodiment, belongs to the scope of protection of the invention.
It is related to CNN (Convolutional Neural Network, convolutional Neural net in the technical scheme that the present invention is provided
Network) and DNN (Deep Neural Network, deep neural network).Fig. 1 shows content CNN, statistics CNN and DNN groups
Into combination neural net schematic diagram.
By taking commodity as an example, content CNN input is the content-data of end article, and content CNN can be from content angle point
End article is analysed, and extracts the content feature vector of end article.Count sales volume data of the CNN input for end article, system
Count CNN and pass through multilayer convolutional neural networks, the slip convolution window of many sizes, the automatic different dimensions statistics for extracting end article
Characteristic vector.DNN input is spliced into by the content feature vector of end article and the statistical nature vector of end article
Characteristic vector, DNN can learn the assemblage characteristic of high-order, and last linear combination obtains the estimated data of end article.
Before the estimated data of target is obtained using content CNN, statistics CNN and DNN compositions combination neural net,
Need content CNN, statistics CNN and DNN are constituted using the content-data of sample commodity and the historical data of sample commodity
Combination neural net is trained.To cause the training data of input to meet the data entry requirement of neutral net, it is necessary to sample
The initial data of this commodity is pre-processed.
For the content-data of commodity, text description or commodity picture that can be by commodity be encoded to content number
According to.
For example, the text of commodity is described as " iPadMini tablet personal computers 32GB ", then text description can be divided into 3
Word " iPadMini ", " tablet personal computer ", " 32GB ", then using text depth representing model Word2Vec, by the text of target
This description is encoded to two-dimensional matrix as shown in table 1.Two-dimensional matrix shown in table 1 is the matrix that 50 rows 3 are arranged, each row of matrix
The term vector corresponding to a word is represented, these term vectors can be generated coding two-dimensional matrix according to occurring being linked in sequence.
Every a line of matrix represents some dimension (totally 50 dimension) of each word in term vector space.
Table 1
For another example, the commodity picture of RGB triple channels can be scaled to the picture of 100 × 100 pixels, it is then whole with 8
Type represents the pixel value of every passage, and pixel value range is 0~255.Because picture has tri- passages of RGB, and have 100 altogether
Coding of graphics can be 100 × 100 by × 100 pixels, therefore pixel value with pixel in every passage encoded ×
3 encoder matrix.Certainly, in order to reduce the computational loads of follow-up CNN networks, the pixel value of each pixel can also be carried out
Normalize computing.
For the historical data of commodity, history sales volume that can be by commodity in different periods is encoded to historical data,
History sales volume of the commodity in different subregions of identical period can also be encoded to historical data.
For example, certain commodity is as shown in table 2 in the history sales volume data in past 5 days.The data can be so cascaded as
Time-domain Statistics sequence 10,15,5,4,3.
| 10 | 15 | 5 | 4 | 3 |
Table 2
For another example, map can be divided into the square of equal sizes according to space, region statistical matrix, ground are set up with this
Each element difference corresponding goods in the statistical matrix of domain are in the regional sales volume data of each segmentation square, such as shown in table 3.Its
In, X represents that the sales volume of commodity in the region is unknown, it is necessary to be estimated.
| 1 | 0 | 1 | 0 | 0 |
| 0 | 0 | X | 0 | 0 |
| 0 | 3 | 5 | 0 | 0 |
| 0 | 0 | 0 | 6 | 0 |
| 0 | 0 | 0 | 0 | 0 |
Table 3
After the pretreatment for completing initial data, pretreated region statistical matrix can be inputted combination neural net,
To be trained to combination neural net.
Fig. 2 shows content CNN schematic network structure.Figure it is seen that content CNN is by convolutional layer, Chi Hua
Layer, full articulamentum and output layer composition.Wherein, convolutional layer and pond layer learn the local message of commodity, then will be local special
Reference breath converges to full articulamentum, and full articulamentum then learns to include the global information of commodity, final output content feature vector, example
Such as the vector of one 50 dimension, DNN input is used as.
It will be understood by those skilled in the art that when extracting content feature vector from the description of the text of commodity, can also make
Content CNN is replaced with RNN (Recurrent Neural Networks, Recognition with Recurrent Neural Network).
Fig. 3 shows statistics CNN schematic network structure.From figure 3, it can be seen that statistics CNN is by convolutional layer, Chi Hua
Layer and output layer composition.Statistics CNN input can be the chronological sales volume data of commodity (with the depth generation of color
The sales volume height of the table period).In training process, if time domain sequences shown in input table 2, also need to input sample business simultaneously
The actual sales volume of product, such as Time-domain Statistics sequence 10,15,5,4,3,6；If region sequence shown in table 3 is inputted, while also needing
Input sample commodity are wanted in the actual sales volume (X value) of corresponding time specific region.
Distinguished with the content CNN shown in Fig. 2, statistics CNN does not have full articulamentum, because full articulamentum can be destroyed
The locality of statistics.Here locality refers to that, for the statistical nature extracted here, each feature is simply with certain
The sales volume sequence of section time is relevant, not comprising global sales volume information.If inserting full articulamentum in statistics CNN, quite
In just having carried out statistical nature integration in CNN ends, and the integration processing of statistical nature is arranged in follow-up DNN and more closed
Reason.For example, statistics CNN output can be the vector of 10 dimensions, DNN input is used as.
Fig. 4 shows DNN schematic network structure.From fig. 4, it can be seen that DNN is made up of full articulamentum and output layer.
The input of DNN networks is the one-dimensional characteristic vector by the above characteristic vector and statistical nature vector connection, for example, can be
The characteristic vector for 60 dimensions being cascaded into by the content feature vector of 50 dimensions and the statistical nature vector of 10 dimensions.
Each neuron in full articulamentum is to the content characteristic after the content feature vector of input or the cascade of input
Vector sum statistical nature vector carries out the nonlinear activation functional operation such as relu or sigmoid, and complicated composite function is arrived in study,
Final output commodity estimate sales volume, and this, which estimates sales volume, to be a floating number.Typically, relu algorithms are better than
Sigmoid algorithms, because relu algorithms do not have the larger saturation region such as sigmoid algorithms.The larger saturation of sigmoid algorithms
Area causes gradient back-propagation algorithm often to cause gradient disperse problem, so that DNN networks cannot be trained effectively.
The one embodiment for the data digging method that the present invention is provided is described with reference to Fig. 5.
Fig. 5 shows the schematic flow sheet of one embodiment of data digging method of the present invention.As shown in figure 5, the embodiment
Data digging method include：
Step S502, by the content-data input content CNN of target, obtains the content feature vector of target.
Step S502 three kinds of implementations are set forth below.
(1) text depth representing model is utilized, the text description of target is encoded to two-dimensional matrix.Two-dimensional matrix is inputted
First content CNN, obtain target Text eigenvector, and using Text eigenvector as target content feature vector.
(2) it is multi-dimensional matrix by the coding of graphics of target according to the pixel value of the picture of target.By multi-dimensional matrix input the
Two content CNN, obtain target picture feature vector, and using picture feature vector as target content feature vector.
(3) text depth representing model is utilized, the text description of target is encoded to two-dimensional matrix.Two-dimensional matrix is inputted
First content CNN, obtains the Text eigenvector of target.According to the pixel value of the picture of target, it is by the coding of graphics of target
Multi-dimensional matrix.Multi-dimensional matrix is inputted into the second content CNN, the picture feature vector of target is obtained.By Text eigenvector and figure
Piece characteristic vector as target content feature vector.
Above-mentioned first content CNN and the second content CNN can be obtained by foregoing content CNN training methods respectively.By
It is similar in the content-data preprocessing process of target and the content-data preprocessing process of sample, content number not to target herein
Data preprocess process is repeated.
Step S504, deep neural network DNN is inputted by the content feature vector of target, obtains the estimated data of target.
For new product, the method normal extraction that the content feature vector of new product can be according to the present embodiment, but system
Meter feature can not obtain therefore could be arranged to 0, then DNN networks will only be through content feature vector and obtain pin to estimate
Amount.Due to manually calculate end article and other commodity the degree of association, directly given a forecast with DNN networks.If by
Great amount of samples data are trained up, DNN networks may learn other commodity similar to new product sales volume estimation (estimation
Reference feature contains content statisticses feature).Therefore for new product, can also reasonably it be sold according to its content characteristic
Amount is estimated.
The data digging method based on CNN and DNN that above-described embodiment is proposed, can be contour from text, the image of commodity
Content feature vector is extracted in dimension content-data, and it is pre- based on content feature vector progress dependent merchandise analysis and new product sales volume
Survey, compared to traditional rule-based method, generalization ability is stronger, extended capability more preferably, and can quantify, not by subjective factor
Influence, therefore, it is possible to obtain more accurately new product Method for Sales Forecast effect.Meanwhile, CNN than artificial statistical rules method more
Flexibly, can be more efficient than manually adjusting statistical rules with intelligence learning adjusting parameter；DNN can be with intelligence learning characteristic dimension
The parameters such as conversion, combinations of features, it is higher than manual features treatment effeciency, and estimate it is more accurate.
Inventor also found that sales volume method of estimation of the prior art does not account for the relevance of region to Sales Volume of Commodity
Influence, such as A with B ground distance with A with C compare closer to, humanity custom it is more like, then end article is on A ground
Sales volume should to B it is more related.Therefore, solving this present invention also offers another of data digging method embodiment
Technical problem.
Fig. 6 shows the schematic flow sheet of another embodiment of data digging method of the present invention.As shown in fig. 6, the implementation
The data digging method of example includes：
Step S602, by the content-data input content convolutional neural networks CNN of target, obtain the content characteristic of target to
Amount.Specific implementation is referred to step S502.
Step S604, by the historical data input statistics CNN of target, obtains the statistical nature vector of target.
Step S604 three kinds of implementations are set forth below.
(1) target is cascaded as Time-domain Statistics sequence in the historical data of different periods；By Time-domain Statistics sequence inputting
One statistics CNN, obtain target Time-domain Statistics characteristic vector, and using Time-domain Statistics characteristic vector as target statistical nature to
Amount.
For example, data input first shown in table 2 is counted after CNN, it is 3 that the first statistics CNN, which is adopted for example with window size,
Convolution kernel it is handled, wherein convolution kernel weight is first initialized as 1/3,1/3,1/3.Sales volume data sequence so to 5 days
Carry out after convolution operation, resulting result is as shown in table 4.The example for sliding window, has been counted 3 days equivalent to 3 days
Sales volume average value.For practical application, convolution kernel weight can automatically be learnt by algorithm, and convolution kernel size can be set
Multiple, different scale is put, and convolution operation can form depth network with stacked multilayer.
| 10 | 15 | 5 |
Table 4
(2) multiple subregions are turned to by the statistical regions of target are discrete；By target in different subregions of identical period
Historical data is encoded to region statistical matrix；Region statistical matrix input second is counted into CNN, the region statistics for obtaining target is special
Vector is levied, and region statistical nature vector is vectorial as the statistical nature of target.
The statistics of objective area, can do convolution operation by the statistics of adjacent area and estimate obtaining.For example, will
The sales volume in " X " region in computational chart 3, can calculate the convolution statistical value of 3 × 3 window around target area.Likewise, convolution
The weighted value of core can automatically be learnt by the method for machine learning, and number, the size of convolution kernel can pass through experimental configuration.
If output region statistical nature be it is two-dimentional, can by its by row be launched into one-dimensional vector.
(3) the history sales volume by target in different periods is cascaded as Time-domain Statistics sequence；By Time-domain Statistics sequence inputting
One statistics CNN, obtains the Time-domain Statistics characteristic vector of target；Multiple subregions are turned to by the statistical regions of target are discrete；By mesh
The historical data being marked in different subregions of identical period is encoded to region statistical matrix；Region statistical matrix input second is united
CNN is counted, the region statistical nature vector of target is obtained；It regard Time-domain Statistics characteristic vector and region statistical nature vector as target
Statistical nature vector.
Above-mentioned first statistics CNN can be obtained by foregoing statistics CNN training methods respectively with the second statistics CNN.By
It is similar in the historical data preprocessing process of target and the historical data preprocessing process of sample, history number not to target herein
Data preprocess process is repeated.
Step S606, will input depth nerve net after the content feature vector of target and the vector cascade of the statistical nature of target
Network DNN, obtains the estimated data of target.
In above-described embodiment, for asking for " estimating actual sales volume of the end article under object time and regional condition "
Topic, has carried out more comprehensively neural net model establishing.It is related by establishing the three classes content correlated characteristic related to target, time domain
Feature, region correlated characteristic, and extract, handle and encode in terms of contents attribute association, time domain association, regional interrelation three
It is adapted to the input data of Processing with Neural Network, eventually through content CNN, statistics CNN and DNN combination neural net structure,
Effectively three category features of processing and combination, save the work of a large amount of manual features processing, are efficiently completed the standard of end article sales volume
Really estimate.
In certain embodiments, after the estimated data for obtaining target, the real data of target can also be calculated and number is estimated
According to Mean Square Error.Then utilize back-propagation algorithm, by Mean Square Error from content CNN, statistics CNN and DNN
The output layer of the combination neural net of composition is transmitted to input layer, and the parameter of each layer in combination neural net is updated with iteration.From
And by the combination neural net after renewal, more accurately estimate the data of succeeding target.
The data mining device of one embodiment of the invention is described with reference to Fig. 7.
Fig. 7 shows the structural representation of one embodiment of data mining device of the present invention.As shown in fig. 7, the embodiment
Data mining device 70 include：
Content feature vector acquisition module 702, is configured to the content-data input content convolutional neural networks of target
CNN, obtains the content feature vector of target.
Estimated data acquisition module 704, is configured to the content feature vector of target inputting deep neural network DNN, obtains
Take the estimated data of target.
In certain embodiments, content feature vector acquisition module 702 is configured to：, will using text depth representing model
The text description of target is encoded to two-dimensional matrix；By two-dimensional matrix input first content CNN, obtain target text feature to
Amount, and using Text eigenvector as target content feature vector.
In certain embodiments, content feature vector acquisition module 702 is configured to：According to the pixel value of the picture of target,
It is multi-dimensional matrix by the coding of graphics of target；Multi-dimensional matrix is inputted into the second content CNN, the picture feature vector of target is obtained,
And using picture feature vector as target content feature vector.
In certain embodiments, content feature vector acquisition module 702 is configured to：, will using text depth representing model
The text description of target is encoded to two-dimensional matrix；By two-dimensional matrix input first content CNN, obtain target text feature to
Amount；It is multi-dimensional matrix by the coding of graphics of target according to the pixel value of the picture of target；Multi-dimensional matrix is inputted into the second content
CNN, obtains the picture feature vector of target；Using Text eigenvector and picture feature vector as target content characteristic to
Amount.
The data digging method based on CNN and DNN that above-described embodiment is proposed, can be contour from text, the image of commodity
Content feature vector is extracted in dimension content-data, and it is pre- based on content feature vector progress dependent merchandise analysis and new product sales volume
Survey, compared to traditional rule-based method, generalization ability is stronger, extended capability more preferably, and can quantify, not by subjective factor
Influence, therefore, it is possible to obtain more accurately new product Method for Sales Forecast effect.Meanwhile, CNN than artificial statistical rules method more
Flexibly, can be more efficient than manually adjusting statistical rules with intelligence learning adjusting parameter；DNN can be with intelligence learning characteristic dimension
The parameters such as conversion, combinations of features, it is higher than manual features treatment effeciency, and estimate it is more accurate.
In certain embodiments, the device 70 also includes the vectorial acquisition module 703 of statistical nature, is configured to going through target
History data input counts CNN, obtains the statistical nature vector of target.Estimated data acquisition module 704 is additionally configured to：By target
Deep neural network DNN is inputted after content feature vector and the vector cascade of the statistical nature of target, the estimated data of target is obtained.
In certain embodiments, the vectorial acquisition module 703 of statistical nature is configured to：By target different periods history number
According to being cascaded as Time-domain Statistics sequence；Time-domain Statistics sequence inputting first is counted into CNN, obtain the Time-domain Statistics feature of target to
Amount, and the statistical nature of Time-domain Statistics characteristic vector as target is vectorial.
In certain embodiments, the vectorial acquisition module 703 of statistical nature is configured to：Turn to the statistical regions of target are discrete
Many sub-regions；Historical data of the target in different subregions of identical period is encoded to region statistical matrix；Region is united
The statistics CNN of Input matrix second is counted, the region statistical nature vector of target is obtained, and regard region statistical nature vector as target
Statistical nature vector.
In certain embodiments, the vectorial acquisition module 703 of statistical nature is configured to：By target different periods history pin
Amount is cascaded as Time-domain Statistics sequence；Time-domain Statistics sequence inputting first is counted into CNN, obtain the Time-domain Statistics feature of target to
Amount；Multiple subregions are turned to by the statistical regions of target are discrete；By historical data of the target in different subregions of identical period
It is encoded to region statistical matrix；Region statistical matrix input second is counted into CNN, the region statistical nature vector of target is obtained；
Time-domain Statistics characteristic vector and region statistical nature vector is vectorial as the statistical nature of target.
In above-described embodiment, for asking for " estimating actual sales volume of the end article under object time and regional condition "
Topic, has carried out more comprehensively neural net model establishing.It is related by establishing the three classes content correlated characteristic related to target, time domain
Feature, region correlated characteristic, and extract, handle and encode in terms of contents attribute association, time domain association, regional interrelation three
It is adapted to the input data of Processing with Neural Network, eventually through content CNN, statistics CNN and DNN combination neural net structure,
Effectively three category features of processing and combination, save the work of a large amount of manual features processing, are efficiently completed the standard of end article sales volume
Really estimate.
In certain embodiments, content CNN is made up of convolutional layer, pond layer, full articulamentum and output layer.
In certain embodiments, statistics CNN is made up of convolutional layer, pond layer and output layer.
In certain embodiments, DNN is made up of full articulamentum and output layer, and each neuron in full articulamentum performs non-
Linear activation primitive computing, obtains the estimated data of target.
In certain embodiments, the device 70 also includes parameter update module 705, is configured to：Calculate the actual number of target
According to the Mean Square Error with estimated data；Using back-propagation algorithm, by Mean Square Error from content CNN, statistics CNN with
And the output layer of the combination neural net of DNN compositions is transmitted to input layer, and the ginseng of each layer in combination neural net is updated with iteration
Number.
In certain embodiments, the device also includes training module 701, is configured to content-data and sample using sample
Historical data, to content CNN, statistics CNN and DNN composition combination neural net be trained.
Fig. 8 shows the structure chart of another embodiment of data mining device of the present invention.As shown in figure 8, the embodiment
Data mining device 80 include：Memory 810 and the processor 820 for being coupled to the memory 810, processor 820 by with
It is set to based on the instruction being stored in memory 810, performs the data digging method in any one foregoing embodiment.
Wherein, memory 810 is such as can include system storage, fixed non-volatile memory medium.System is stored
Device is such as the operating system that is stored with, application program, Boot loader (Boot Loader) and other programs.
Fig. 9 shows the structure chart of another embodiment of data mining device of the present invention.As shown in figure 9, the embodiment
Device 90 include：Memory 810 and processor 820, can also include input/output interface 930, network interface 940, deposit
Store up interface 950 etc..It can for example pass through bus between these interfaces 930,940,950 and memory 810 and processor 820
950 connections.Wherein, input/output interface 930 is display, the input-output equipment such as mouse, keyboard, touch-screen provides connection and connect
Mouthful.Network interface 940 provides connecting interface for various networked devices.Memory interface 950 is the external storages such as SD card, USB flash disk
Connecting interface is provided.
Present invention additionally comprises a kind of computer-readable recording medium, computer instruction is stored thereon with, the instruction is processed
Device realizes the data digging method in any one foregoing embodiment when performing.
It should be understood by those skilled in the art that, embodiments of the invention can be provided as method, system or computer program
Product.Therefore, the present invention can be using the reality in terms of complete hardware embodiment, complete software embodiment or combination software and hardware
Apply the form of example.Moreover, the present invention can be used in one or more computers for wherein including computer usable program code
The calculating implemented on non-transient storage medium (including but is not limited to magnetic disk storage, CD-ROM, optical memory etc.) can be used
The form of machine program product.
The present invention is the flow with reference to method according to embodiments of the present invention, equipment (system) and computer program product
Figure and/or block diagram are described.It should be understood that can be by every first-class in computer program instructions implementation process figure and/or block diagram
Journey and/or the flow in square frame and flow chart and/or block diagram and/or the combination of square frame.These computer programs can be provided
The processor of all-purpose computer, special-purpose computer, Embedded Processor or other programmable data processing devices is instructed to produce
A raw machine so that produced by the instruction of computer or the computing device of other programmable data processing devices for real
The device for the function of being specified in present one flow of flow chart or one square frame of multiple flows and/or block diagram or multiple square frames.
These computer program instructions, which may be alternatively stored in, can guide computer or other programmable data processing devices with spy
Determine in the computer-readable memory that mode works so that the instruction being stored in the computer-readable memory, which is produced, to be included referring to
Make the manufacture of device, the command device realize in one flow of flow chart or multiple flows and/or one square frame of block diagram or
The function of being specified in multiple square frames.
These computer program instructions can be also loaded into computer or other programmable data processing devices so that in meter
Series of operation steps is performed on calculation machine or other programmable devices to produce computer implemented processing, thus in computer or
The instruction performed on other programmable devices is provided for realizing in one flow of flow chart or multiple flows and/or block diagram one
The step of function of being specified in individual square frame or multiple square frames.
The foregoing is only presently preferred embodiments of the present invention, be not intended to limit the invention, it is all the present invention spirit and
Within principle, any modification, equivalent substitution and improvements made etc. should be included in the scope of the protection.
Claims (28)
1. a kind of data digging method, it is characterised in that including：
By the content-data input content convolutional neural networks CNN of target, the content feature vector of target is obtained；
The content feature vector of target is inputted into deep neural network DNN, the estimated data of target is obtained.
2. the method as described in claim 1, it is characterised in that methods described also includes：
By the historical data input statistics CNN of target, the statistical nature vector of target is obtained；
Deep neural network DNN will be inputted after the content feature vector of target and the vector cascade of the statistical nature of target, obtain mesh
Target estimated data.
3. the method as described in claim 1, it is characterised in that the content-data input content convolutional Neural net by target
Network CNN, obtaining the content feature vector of target includes：
Using text depth representing model, the text description of target is encoded to two-dimensional matrix；
The two-dimensional matrix is inputted into the content CNN, the Text eigenvector of target is obtained, and by the Text eigenvector
It is used as the content feature vector of target.
4. the method as described in claim 1, it is characterised in that the content-data input content convolutional Neural net by target
Network CNN, obtaining the content feature vector of target includes：
It is multi-dimensional matrix by the coding of graphics of target according to the pixel value of the picture of target；
The multi-dimensional matrix is inputted into the content CNN, the picture feature vector of target is obtained, and the picture feature is vectorial
It is used as the content feature vector of target.
5. the method as described in claim 1, it is characterised in that the content CNN includes first content CNN and the second content
CNN；
The content-data input content convolutional neural networks CNN by target, obtaining the content feature vector of target includes：
Using text depth representing model, the text description of target is encoded to two-dimensional matrix；
The two-dimensional matrix is inputted into first content CNN, the Text eigenvector of target is obtained；
It is multi-dimensional matrix by the coding of graphics of target according to the pixel value of the picture of target；
The multi-dimensional matrix is inputted into the second content CNN, the picture feature vector of target is obtained；
Using the Text eigenvector and the vectorial content feature vector as target of the picture feature.
6. method as claimed in claim 2, it is characterised in that the historical data input statistics CNN by target, obtains mesh
Target statistical nature vector includes：
Target is cascaded as Time-domain Statistics sequence in the historical data of different periods；
It will count CNN described in the Time-domain Statistics sequence inputting, obtain the Time-domain Statistics characteristic vector of target, and by the time domain
Statistical nature vector is vectorial as the statistical nature of target.
7. method as claimed in claim 2, it is characterised in that the historical data input statistics CNN by target, obtains mesh
Target statistical nature vector includes：
Multiple subregions are turned to by the statistical regions of target are discrete；
Historical data of the target in different subregions of identical period is encoded to region statistical matrix；
The region statistical matrix is inputted into the statistics CNN, the region statistical nature vector of target is obtained, and by the region
Statistical nature vector is vectorial as the statistical nature of target.
8. method as claimed in claim 2, it is characterised in that the statistics CNN includes the first statistics CNN and the second statistics
CNN；
The historical data input statistics CNN by target, obtaining the statistical nature vector of target includes：
History sales volume by target in different periods is cascaded as Time-domain Statistics sequence；
The Time-domain Statistics sequence inputting first is counted into CNN, the Time-domain Statistics characteristic vector of target is obtained；
Multiple subregions are turned to by the statistical regions of target are discrete；
Historical data of the target in different subregions of identical period is encoded to region statistical matrix；
Region statistical matrix input second is counted into CNN, the region statistical nature vector of target is obtained；
The Time-domain Statistics characteristic vector and region statistical nature vector is vectorial as the statistical nature of the target.
9. the method as any one of claim 1 to 5, it is characterised in that the content CNN by convolutional layer, pond layer,
Full articulamentum and output layer composition.
10. the method as any one of claim 2,6 to 8, it is characterised in that the statistics CNN is by convolutional layer, Chi Hua
Layer and output layer composition.
11. method as claimed in claim 1 or 2, it is characterised in that the DNN is made up of full articulamentum and output layer, described
Each neuron in full articulamentum performs nonlinear activation functional operation.
12. method as claimed in claim 2, it is characterised in that methods described also includes：
Calculate the real data of target and the Mean Square Error of estimated data；
Using back-propagation algorithm, by the Mean Square Error from the content CNN, the statistics CNN and the DNN groups
Into the output layer of combination neural net transmitted to input layer, the parameter of each layer in the combination neural net is updated with iteration.
13. method as claimed in claim 2, it is characterised in that methods described also includes：
Using the content-data and the historical data of sample of sample, to the content CNN, the statistics CNN and the DNN groups
Into combination neural net be trained.
14. a kind of data mining device, it is characterised in that including：
Content feature vector acquisition module, is configured to, by the content-data input content convolutional neural networks CNN of target, obtain mesh
Target content feature vector；
Estimated data acquisition module, for the content feature vector of target to be inputted into deep neural network DNN, obtains the pre- of target
Estimate data.
15. device as claimed in claim 14, it is characterised in that described device also includes the vectorial acquisition module of statistical nature,
It is configured to, by the historical data input statistics CNN of target, obtain the statistical nature vector of target；
The estimated data acquisition module is additionally configured to：By the content feature vector of target and the vector cascade of the statistical nature of target
Deep neural network DNN is inputted afterwards, obtains the estimated data of target.
16. device as claimed in claim 14, it is characterised in that the content feature vector acquisition module is configured to：
Using text depth representing model, the text description of target is encoded to two-dimensional matrix；
The two-dimensional matrix is inputted into the content CNN, the Text eigenvector of target is obtained, and by the Text eigenvector
It is used as the content feature vector of target.
17. device as claimed in claim 14, it is characterised in that the content feature vector acquisition module is configured to：
It is multi-dimensional matrix by the coding of graphics of target according to the pixel value of the picture of target；
The multi-dimensional matrix is inputted into the content CNN, the picture feature vector of target is obtained, and the picture feature is vectorial
It is used as the content feature vector of target.
18. device as claimed in claim 14, it is characterised in that the content CNN includes first content CNN and the second content
CNN；
The content feature vector acquisition module is configured to：
Using text depth representing model, the text description of target is encoded to two-dimensional matrix；
The two-dimensional matrix is inputted into first content CNN, the Text eigenvector of target is obtained；
It is multi-dimensional matrix by the coding of graphics of target according to the pixel value of the picture of target；
The multi-dimensional matrix is inputted into the second content CNN, the picture feature vector of target is obtained；
Using the Text eigenvector and the vectorial content feature vector as target of the picture feature.
19. device as claimed in claim 15, it is characterised in that the vectorial acquisition module of the statistical nature is configured to：
Target is cascaded as Time-domain Statistics sequence in the historical data of different periods；
It will count CNN described in the Time-domain Statistics sequence inputting, obtain the Time-domain Statistics characteristic vector of target, and by the time domain
Statistical nature vector is vectorial as the statistical nature of target.
20. device as claimed in claim 15, it is characterised in that the vectorial acquisition module of the statistical nature is configured to：
Multiple subregions are turned to by the statistical regions of target are discrete；
Historical data of the target in different subregions of identical period is encoded to region statistical matrix；
The region statistical matrix is inputted into the statistics CNN, the region statistical nature vector of target is obtained, and by the region
Statistical nature vector is vectorial as the statistical nature of target.
21. device as claimed in claim 15, it is characterised in that the statistics CNN includes the first statistics CNN and the second statistics
CNN；
The vectorial acquisition module of the statistical nature is configured to：
History sales volume by target in different periods is cascaded as Time-domain Statistics sequence；
The Time-domain Statistics sequence inputting first is counted into CNN, the Time-domain Statistics characteristic vector of target is obtained；
Multiple subregions are turned to by the statistical regions of target are discrete；
Historical data of the target in different subregions of identical period is encoded to region statistical matrix；
Region statistical matrix input second is counted into CNN, the region statistical nature vector of target is obtained；
The Time-domain Statistics characteristic vector and region statistical nature vector is vectorial as the statistical nature of the target.
22. the device as any one of claim 14 to 18, it is characterised in that the content CNN is by convolutional layer, Chi Hua
Layer, full articulamentum and output layer composition.
23. the device as any one of claim 15,19 to 21, it is characterised in that the statistics CNN by convolutional layer,
Pond layer and output layer composition.
24. device as claimed in claim 14, it is characterised in that the DNN is made up of full articulamentum and output layer, described complete
Each neuron in articulamentum carries out nonlinear activation functional operation.
25. device as claimed in claim 15, it is characterised in that described device also includes parameter update module, is configured to：
Calculate the real data of target and the Mean Square Error of estimated data；
Using back-propagation algorithm, by the Mean Square Error from the content CNN, the statistics CNN and the DNN groups
Into the output layer of combination neural net transmitted to input layer, the parameter of each layer in the combination neural net is updated with iteration.
26. device as claimed in claim 15, it is characterised in that described device also includes training module, is configured to utilize sample
This content-data and the historical data of sample, the combination god constituted to the content CNN, the statistics CNN and the DNN
It is trained through network.
27. a kind of data mining device, it is characterised in that including：
Memory；And
The processor of the memory is coupled to, the processor is configured as based on the instruction being stored in the memory,
Perform the data digging method as any one of claim 1 to 13.
28. a kind of computer-readable recording medium, it is characterised in that the computer-readable recording medium storage has computer to refer to
The data digging method as any one of claim 1 to 13 is realized in order, the instruction when being executed by processor.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710397745.XA CN107239532B (en) | 2017-05-31 | 2017-05-31 | Data mining method and device |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710397745.XA CN107239532B (en) | 2017-05-31 | 2017-05-31 | Data mining method and device |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN107239532A true CN107239532A (en) | 2017-10-10 |
| CN107239532B CN107239532B (en) | 2020-07-31 |
Family
ID=59985214
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201710397745.XA Active CN107239532B (en) | 2017-05-31 | 2017-05-31 | Data mining method and device |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN107239532B (en) |
Cited By (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107944488A (en) * | 2017-11-21 | 2018-04-20 | 清华大学 | Long time series data processing method based on stratification depth network |
| CN108564326A (en) * | 2018-04-19 | 2018-09-21 | 安吉汽车物流股份有限公司 | Prediction technique and device, computer-readable medium, the logistics system of order |
| CN109447716A (en) * | 2018-11-09 | 2019-03-08 | 四川长虹电器股份有限公司 | Method for Sales Forecast method and server based on Recognition with Recurrent Neural Network |
| CN113762481A (en) * | 2021-09-13 | 2021-12-07 | 中国地震局地球物理研究所 | Tomography method and system based on deep learning |
| CN113822689A (en) * | 2020-07-01 | 2021-12-21 | 北京沃东天骏信息技术有限公司 | Advertisement conversion rate estimation method and device, storage medium and electronic equipment |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN104077613A (en) * | 2014-07-16 | 2014-10-01 | 电子科技大学 | Crowd density estimation method based on cascaded multilevel convolution neural network |
| CN105095833A (en) * | 2014-05-08 | 2015-11-25 | 中国科学院声学研究所 | Network constructing method for human face identification, identification method and system |
| CN106021364A (en) * | 2016-05-10 | 2016-10-12 | 百度在线网络技术（北京）有限公司 | Method and device for establishing picture search correlation prediction model, and picture search method and device |
-
2017
- 2017-05-31 CN CN201710397745.XA patent/CN107239532B/en active Active
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN105095833A (en) * | 2014-05-08 | 2015-11-25 | 中国科学院声学研究所 | Network constructing method for human face identification, identification method and system |
| CN104077613A (en) * | 2014-07-16 | 2014-10-01 | 电子科技大学 | Crowd density estimation method based on cascaded multilevel convolution neural network |
| CN106021364A (en) * | 2016-05-10 | 2016-10-12 | 百度在线网络技术（北京）有限公司 | Method and device for establishing picture search correlation prediction model, and picture search method and device |
Cited By (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107944488A (en) * | 2017-11-21 | 2018-04-20 | 清华大学 | Long time series data processing method based on stratification depth network |
| CN107944488B (en) * | 2017-11-21 | 2018-12-11 | 清华大学 | Long time series data processing method based on stratification depth network |
| CN108564326A (en) * | 2018-04-19 | 2018-09-21 | 安吉汽车物流股份有限公司 | Prediction technique and device, computer-readable medium, the logistics system of order |
| CN108564326B (en) * | 2018-04-19 | 2021-12-21 | 安吉汽车物流股份有限公司 | Order prediction method and device, computer readable medium and logistics system |
| CN109447716A (en) * | 2018-11-09 | 2019-03-08 | 四川长虹电器股份有限公司 | Method for Sales Forecast method and server based on Recognition with Recurrent Neural Network |
| CN113822689A (en) * | 2020-07-01 | 2021-12-21 | 北京沃东天骏信息技术有限公司 | Advertisement conversion rate estimation method and device, storage medium and electronic equipment |
| CN113762481A (en) * | 2021-09-13 | 2021-12-07 | 中国地震局地球物理研究所 | Tomography method and system based on deep learning |
| CN113762481B (en) * | 2021-09-13 | 2024-03-12 | 中国地震局地球物理研究所 | Tomographic imaging method and system based on deep learning |
Also Published As
| Publication number | Publication date |
|---|---|
| CN107239532B (en) | 2020-07-31 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN108510012B (en) | Target rapid detection method based on multi-scale feature map | |
| CN107273784B (en) | Image pattern recognition apparatus and method | |
| CN107239532A (en) | Data digging method and device | |
| WO2021188354A1 (en) | Automated and adaptive design and training of neural networks | |
| CN107423442A (en) | Method and system, storage medium and computer equipment are recommended in application based on user's portrait behavioural analysis | |
| CN106709461A (en) | Video based behavior recognition method and device | |
| Fang et al. | DOG: A new background removal for object recognition from images | |
| AU2017101803A4 (en) | Deep learning based image classification of dangerous goods of gun type | |
| CN104008538A (en) | Super-resolution method based on single image | |
| Silver et al. | In vino veritas: Estimating vineyard grape yield from images using deep learning | |
| CN109726725A (en) | The oil painting writer identification method of heterogeneite Multiple Kernel Learning between a kind of class based on large-spacing | |
| Jiang et al. | Nonlinear CNN: improving CNNs with quadratic convolutions | |
| CN107506792A (en) | A kind of semi-supervised notable method for checking object | |
| CN113128478A (en) | Model training method, pedestrian analysis method, device, equipment and storage medium | |
| CN112560967A (en) | Multi-source remote sensing image classification method, storage medium and computing device | |
| CN109255382A (en) | For the nerve network system of picture match positioning, method and device | |
| Cao et al. | Ancient mural classification method based on improved AlexNet network | |
| Verdhan et al. | Image classification using LeNet | |
| CN103440651A (en) | Multi-label image annotation result fusion method based on rank minimization | |
| CN105160354A (en) | Apple disease identifying method based on sparse representation | |
| Xu et al. | Dilated convolution capsule network for apple leaf disease identification | |
| WO2024078112A1 (en) | Method for intelligent recognition of ship outfitting items, and computer device | |
| CN113449103A (en) | Bank transaction flow classification method and system integrating label and text interaction mechanism | |
| CN115982654B (en) | Node classification method and device based on self-supervision graph neural network | |
| CN116029760A (en) | Message pushing method, device, computer equipment and storage medium |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |