CN107301221A - A kind of data digging method of multiple features dimension heap fusion - Google Patents
A kind of data digging method of multiple features dimension heap fusion Download PDFInfo
- Publication number
- CN107301221A CN107301221A CN201710458628.XA CN201710458628A CN107301221A CN 107301221 A CN107301221 A CN 107301221A CN 201710458628 A CN201710458628 A CN 201710458628A CN 107301221 A CN107301221 A CN 107301221A
- Authority
- CN
- China
- Prior art keywords
- model
- training
- dimension
- submodel
- feature
- 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
- 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
- G06F16/22—Indexing; Data structures therefor; Storage structures
- G06F16/2228—Indexing structures
- G06F16/2264—Multidimensional index structures
-
- 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
- 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
- G06F16/28—Databases characterised by their database models, e.g. relational or object models
- G06F16/284—Relational databases
- G06F16/285—Clustering or classification
Landscapes
- Engineering & Computer Science (AREA)
- Databases & Information Systems (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- Software Systems (AREA)
- Fuzzy Systems (AREA)
- Mathematical Physics (AREA)
- Probability & Statistics with Applications (AREA)
- Computational Linguistics (AREA)
- Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
Abstract
The present invention discloses a kind of data digging method of multiple features dimension heap fusion.The data digging method is divided to feature architecture according to business dimension first, then using different character subset training patterns and tuning, then submodel output result constitutes new characteristic set with initial characteristicses, finally carries out heap Fusion training and tuning to new characteristic set.Multiple features dimension heap blending algorithm of the present invention is mainly used in solving the problems, such as that single model is unstable, easy over-fitting, passes through and combines different learning models and strengthen model stability and predictive ability.Because single dimension model portrays limited in one's ability, each dimensional model of the invention combines, then can the whole dimension of comprehensive understanding feature.
Description
Technical field
The present invention relates to a kind of Ensemble Learning Algorithms, more particularly to a kind of data mining side of multiple features dimension heap fusion
Method, belongs to field of artificial intelligence.
Background technology
Heap model (Stack) basic thought is the multiple submodels of training independent to initial data, then exports submodel
It is input in built-up pattern and trains as feature, built-up pattern predicts the outcome as final result.Heap model is main by two parts
It is separate between composition, Part I training submodel, submodel, it is ensured that preferably to accomplish difference while submodel effect
Property；Part II combination submodel is predicted the outcome, and wherein built-up pattern can be linear, or nonlinear.
Depth stacks network model (Deep Stack Network, abbreviation DSN) and is mainly used in accelerating deep neural network
Study optimization process.Current deep learning is taken seriously in every profession and trade, such as pattern-recognition, natural language processing, bioengineering
Deng.Being defeated mankind top expert by the artificial intelligence go program AlphaGo that Google releases simultaneously is influenceed, deep learning and people
Work has intelligently more been pulled to the teeth of the storm.But deep neural network DNN learning process is not easy in fact, is included in DNN
Connection before many intermediate layers and neurode, neuron is even more that number is not very counted.Usually, DNN parameter optimization passes through nothing
Supervised learning carries out training in advance, and the global parameter for then have supervision further according to label data is adjusted, even if so, with
The lifting of network complexity, the calculating time is difficult to control to.
Therefore depth stacks network DSN and is proposed out as a kind of new multi-level network structure, its main concept
It is that model stacks thought, the first layer model is trained first with initial data, then further according to single model output and initial data
The model of the second layer is trained, third layer model then is trained to first and second layer model and initial data, it is reciprocal with this, until most
After export.So, the extraction of level from low to high just can be effectively carried out to raw information, can also accelerate network iteration convergence speed
Degree.
What single hidden layer was done in each layer models of DSN is a kind of nonlinear calculating conversion work, and more is
Tanh functions, also there is a lot of other transfer functions certainly.From input vector, calculated by hidden layer is nonlinear, obtain defeated
Outgoing vector.From output is input in fact it is also seen that being the process of Automatic Feature Extraction, without the feature that manual construction is complicated.
Such as in the application of image recognition, start input be pixel value, by extraction in layer, be constantly transitioned into edge,
The process that shape, topography to whole image recognize.
Comprehensive described, DSN is constantly learnt by the structure of level, the model on upper strata be combined with all models of lower floor and
Original input information, then learns the feature of higher level.Each straton model structure is identical, mainly the input by enriching constantly
Information carrys out lift scheme and portrays ability.The number of plies that DSN can be stacked by Controlling model is come the complexity of Controlling model, so that right
Models fitting ability and generalization ability make equilibrium.
The content of the invention
The present invention combines basic heap model and deep neural network, proposes the data mining side of multiple features dimension heap fusion
Method.
The object of the invention is achieved through the following technical solutions：
A kind of data digging method of multiple features dimension heap fusion, including step：
(1) multiple characteristic dimensions are divided：Feature architecture is classified according to business dimension, between each characteristic dimension mutually
Supplement, design feature pond featurepool={ fc1, fc2..., fcn, fciRepresent single features dimension；
(2) submodel training and tuning：Submodel training is carried out for single features dimension, is intersected according to evaluation index and tested
Tuning model is demonstrate,proved, submodel output O is obtainedci, a total of n submodel and output；
(3) diversity and standardization：It is different for submodel output format, submodel is exported before fusion
It is standardized；
(4) stacked combination submodel：Submodel exports OciIt is combined, is carried out by Stacking modes with primitive character
Model Fusion；Stacking algorithms are divided into 2 layers, first layer with different algorithm T Weak Classifiers of formation, while produce one and
Original data set size identical new data set, the grader of the second layer is constituted using the new data set and a new algorithm；First
Layer includes two steps；
The first step, collects each basic learning device and outputs information in a new data set；It is each to original training set
For data item, the new data set represents each basic learning device to the prediction of the affiliated class of each data item and the data of training set
True classification results；It is not in data item of problems to ensure that the training data of generation basic learning device is concentrated；Then,
Using new data set as new learner training dataset；
Second step, step-sizing training set and learner；Stacking algorithms are by initial data set and by primary data
The basic learning device of collection generation is referred to as one-level training dataset and first-level class device, corresponding, the output of first-level class device
Two grades of training datasets are referred to as with the data set of result composition and the classification learning algorithm of second stage truly classified
And secondary classifier.
Further to realize the object of the invention, it is preferable that the tuning model includes：
1) sample tuning：It is big for two class number of samples differences, lay down a regulation and filter some low contribution samples, can using height
By label data；
2) feature tuning：Feature Selection is carried out using tree-model, selection standard includes feature Distribution value and correlation, feature
Information gain size, feature call the influence that frequency, feature are knocked out；
3) model and arameter optimization：By sample training model, and test obtains each modelling effect, chooses the good mould of performance
Type carries out arameter optimization.
The method that the arameter optimization uses fixed variable adjusting parameter.
It is of the invention compared with existing algorithm, with following remarkable advantage：
(1) each submodel otherness is can guarantee that in multiple features dimension heap blending algorithm, while also can guarantee that a submodel phase
It is mutually independent, each submodel can parallel training, be then combined again.Spend the time few so in each round iteration renewal process,
Model is easily extended to large-scale dataset.
(2) multiple features dimension heap blending algorithm belongs to the integrated learning approach of innovation, can be applied to different pieces of information collection, difference
Business scenario simultaneously can obtain good result.
(3) multiple features dimension heap blending algorithm separately models feature according to business dimension, and these submodels only possess list
One characteristic dimension is portrayed, and performance is general, but each dimensional model is combined, then can portray whole dimensional characteristics comprehensively, in advance
Survey performance more excellent.
Embodiment
To more fully understand the present invention, with reference to embodiment, the invention will be further described, but application claims are protected
The scope of shield is not limited thereto.
A kind of data digging method of multiple features dimension heap fusion, including step：
(1) multiple characteristic dimensions are divided：Feature architecture is classified according to business dimension, between each characteristic dimension mutually
Supplement, design feature pond featurepool={ fc1, fc2..., fcn, fciRepresent single features dimension.
(2) submodel training and tuning.Submodel training is carried out for single features dimension, is intersected according to evaluation index and tested
Tuning model is demonstrate,proved, submodel output O is obtainedci, a total of n submodel and output.Model carries out tuning in terms of three：Sample,
Feature, model and parameter.
Sample tuning.It is big for classification imbalance problems, i.e. two class number of samples differences.Formulate rule
Some low contribution samples are then filtered, highly reliable label data is used；
Feature tunings.Preferred feature, selection standard include feature Distribution value and correlation, characteristic information gain size,
Feature calls frequency, influence of feature knockout etc.；
Model and arameter optimization.By sample training model, and test obtains each modelling effect, chooses the mould of better performances
Type carries out the method that can use fixed variable adjusting parameter under arameter optimization, multiparameter case.
(3) diversity and standardization.Model Fusion effect depends on the performance of single model and the weight of each model output
Right, under the premise of diversity is met, the model of effect difference can also be merged.Influenceed by different model output formats,
Single model output is standardized (criterion score such as z-score, ranking-score metering method) before fusion, son
Otherness is embodied in following aspect between model：
The species of basic classification device in itself, i.e. its composition algorithm are different；
Data carry out different disposal, including boosting (error rate weighted sample method), bagging (mean sample sides
Method), cross-validation (cross validation), hold-out test (model verification method) etc.；
Data characteristics processing and selection；
Output result processing；
Introduce randomness.
(4) stacked combination submodel.Submodel exports OciIt is combined, passes through with primitive character featurepool
Stacking modes carry out Model Fusion.Stacking algorithms are divided into 2 layers, and first layer forms T weak typing with different algorithms
Device, while producing one and original data set size identical new data set, is constituted using this new data set and a new algorithm
The grader of the second layer.
The first step, collects each basic learning device and outputs information in a new data set.It is each to original training set
For data item, the new data set represents each basic learning device to the prediction of the affiliated class of each data item and the data of training set
True classification results.It should be noted that must assure that the training data of generation basic learning device is concentrated is not in exist to ask
The data item of topic.Then, using new data set as new learner training dataset.
Second step, step-sizing training set and learner.Stacking algorithms are by initial data set and by primary data
The basic learning device of collection generation is referred to as one-level training dataset and first-level class device, corresponding, the output of first-level class device
Two grades of training datasets are referred to as with the data set of result composition and the classification learning algorithm of second stage truly classified
And secondary classifier.
Embodiment：2015 Ali movement is recommended and Guangdong Communication citizens' activities prediction
Algorithm using classical accuracy (precision), recall rate (recall) and F1 values as evaluating standard, finally
Scoring is ranked up according to F1 values.Specific formula for calculation is as follows：
In mobile recommendation analysis, PredictionSet is prediction purchase data acquisition system, and ReferenceSet is true purchase
Buy data acquisition system.In trip prediction, PredictionSet is that data acquisition system is taken in prediction, and ReferenceSet is truly to take
Multiply data acquisition system.
Online Judge number of times is once a day, to ensure model robustness, to devise offline evaluation metricses.One it is good from
Line evaluation and test can should as far as possible simulate Online Judge environment, accomplish that evaluate and test achievement offline ensures consistent with Online Judge achievement, subtracts
Few dependence to being evaluated and tested on line.
The following is contrast of the offline evaluation and test achievement with evaluating and testing achievement on line in trip prediction and mobile recommendation two datasets：
The trip prediction of table 1 is offline with evaluating and testing achievement on line
The movement of table 2 is recommended offline with evaluating and testing achievement on line
Trip prediction and the on-line off-line data set of mobile recommendation prediction are compared, it is found that offline achievement and online achievement are basic
Correlation, if achievement has been lifted in that is, offline evaluation and test, then the achievement of correspondence Online Judge can also be lifted, simply
Lifting amplitude, which has, slightly to be changed.Also illustrate that model shows stable on different pieces of information collection simultaneously, be not in over-fitting feelings
Condition, generalization ability is good.
Multiple features dimension heap blending algorithm (DFSE) is as submodel Integrated Algorithm, in trip prediction and mobile recommendation prediction
In F1 values (being shown in Table 3) and change curve it is as follows.
The multiple features dimension heap blending algorithm of table 3 is contrasted
Base Model (basic model) are based on time preference's feature architecture and sliding window sample is built, multiple features dimension
Heap blending algorithm (DFSE) collects each characteristic dimension model feature, maximizes favourable factors and minimizes unfavourable ones, and the difference between model is fusion lifting F1 achievements
Key.
Claims (3)
1. a kind of data digging method of multiple features dimension heap fusion, it is characterised in that including step：
(1) multiple characteristic dimensions are divided：Feature architecture is classified according to business dimension, it is mutually complementary between each characteristic dimension
Fill, design feature pond featurepool={ fc1, fc2..., fcn, fciRepresent single features dimension；
(2) submodel training and tuning：Submodel training is carried out for single features dimension, is adjusted according to evaluation index cross validation
Excellent model, obtains submodel output Oci, a total of n submodel and output；
(3) diversity and standardization：It is different for submodel output format, submodel is exported and carried out before fusion
Standardization；
(4) stacked combination submodel：Submodel exports OciIt is combined with primitive character, model is carried out by Stacking modes
Fusion；Stacking algorithms are divided into 2 layers, and first layer is with different algorithm T Weak Classifiers of formation, while producing one and former number
According to collection size identical new data set, the grader of the second layer is constituted using the new data set and a new algorithm；First layer bag
Include two steps；
The first step, collects each basic learning device and outputs information in a new data set；To each data of original training set
For, the new data set represents each basic learning device to the true of the data of the prediction of the affiliated class of each data item and training set
Real classification results；It is not in data item of problems to ensure that the training data of generation basic learning device is concentrated；Then, will be new
Data set as new learner training dataset；
Second step, step-sizing training set and learner；Stacking algorithms are by initial data set with being given birth to by initial data set
Into basic learning device be referred to as one-level training dataset and first-level class device, it is corresponding, the output of first-level class device with it is true
The data set of result composition and the classification learning algorithm of second stage of real classification are referred to as two grades of training datasets and two
Level grader.
2. the data digging method of multiple features dimension heap fusion according to claim 1, it is characterised in that the tuning mould
Type includes：
1) sample tuning：It is big for two class number of samples differences, lay down a regulation and filter some low contribution samples, use highly reliable mark
Sign data；
2) feature tuning：Feature Selection is carried out using tree-model, selection standard includes feature Distribution value and correlation, characteristic information
Gain size, feature call the influence that frequency, feature are knocked out；
3) model and arameter optimization：By sample training model, and test obtains each modelling effect, chooses the good model of performance and enters
Row arameter optimization.
3. the data digging method of multiple features dimension heap fusion according to claim 2, it is characterised in that the parameter is adjusted
The method of excellent use fixed variable adjusting parameter.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710458628.XA CN107301221A (en) | 2017-06-16 | 2017-06-16 | A kind of data digging method of multiple features dimension heap fusion |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710458628.XA CN107301221A (en) | 2017-06-16 | 2017-06-16 | A kind of data digging method of multiple features dimension heap fusion |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN107301221A true CN107301221A (en) | 2017-10-27 |
Family
ID=60135424
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201710458628.XA Pending CN107301221A (en) | 2017-06-16 | 2017-06-16 | A kind of data digging method of multiple features dimension heap fusion |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN107301221A (en) |
Cited By (18)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107833216A (en) * | 2017-11-10 | 2018-03-23 | 西安电子科技大学 | Polarimetric SAR Image change detecting method based on depth song ripple difference DSN |
| CN108281184A (en) * | 2018-02-01 | 2018-07-13 | 中国科学院上海微系统与信息技术研究所 | Myocardial ischemia based on machine learning examines method, storage medium and electronic equipment in advance |
| CN108776805A (en) * | 2018-05-03 | 2018-11-09 | 北斗导航位置服务（北京）有限公司 | It is a kind of establish image classification model, characteristics of image classification method and device |
| CN109117864A (en) * | 2018-07-13 | 2019-01-01 | 华南理工大学 | Coronary heart disease risk prediction technique, model and system based on heterogeneous characteristic fusion |
| CN109145175A (en) * | 2018-09-03 | 2019-01-04 | 四川佳联众合企业管理咨询有限公司 | Space-time data prediction technique based on stacking Ensemble Learning Algorithms |
| CN109241227A (en) * | 2018-09-03 | 2019-01-18 | 四川佳联众合企业管理咨询有限公司 | Space-time data based on stacking Ensemble Learning Algorithms predicts modeling method |
| CN109460420A (en) * | 2018-09-28 | 2019-03-12 | 北京赛博贝斯数据科技有限责任公司 | The Knowledge Discovery Method and system of structural data |
| CN109657883A (en) * | 2019-01-28 | 2019-04-19 | 重庆邮电大学 | A kind of bank branches recommended method based on multi-source data driving |
| CN109685104A (en) * | 2018-11-15 | 2019-04-26 | 同盾控股有限公司 | Determination method and device for recognition model |
| WO2019105258A1 (en) * | 2017-11-30 | 2019-06-06 | 中国银联股份有限公司 | Data classification method and system based on machine learning |
| CN109961296A (en) * | 2017-12-25 | 2019-07-02 | 腾讯科技（深圳）有限公司 | Merchant type recognition methods and device |
| CN110264462A (en) * | 2019-06-25 | 2019-09-20 | 电子科技大学 | A kind of breast ultrasound tumour recognition methods based on deep learning |
| CN110365535A (en) * | 2019-07-15 | 2019-10-22 | 阿里巴巴集团控股有限公司 | A kind of model loading method and device based on multi-party computations |
| CN111199343A (en) * | 2019-12-24 | 2020-05-26 | 上海大学 | Multi-model fusion tobacco market supervision abnormal data mining method |
| CN111222575A (en) * | 2020-01-07 | 2020-06-02 | 北京联合大学 | KLXS multi-model fusion method and system based on HRRP target recognition |
| CN111507507A (en) * | 2020-03-24 | 2020-08-07 | 重庆森鑫炬科技有限公司 | Big data-based monthly water consumption prediction method |
| CN113780339A (en) * | 2021-08-03 | 2021-12-10 | 阿里巴巴（中国）有限公司 | Model training, predicting and content understanding method and electronic equipment |
| CN113961774A (en) * | 2021-12-22 | 2022-01-21 | 四川新网银行股份有限公司 | Recommendation method for multi-feature combination strategy |
-
2017
- 2017-06-16 CN CN201710458628.XA patent/CN107301221A/en active Pending
Cited By (25)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107833216B (en) * | 2017-11-10 | 2019-10-25 | 西安电子科技大学 | Polarimetric SAR image change detecting method based on depth song wave difference DSN |
| CN107833216A (en) * | 2017-11-10 | 2018-03-23 | 西安电子科技大学 | Polarimetric SAR Image change detecting method based on depth song ripple difference DSN |
| WO2019105258A1 (en) * | 2017-11-30 | 2019-06-06 | 中国银联股份有限公司 | Data classification method and system based on machine learning |
| CN109961296A (en) * | 2017-12-25 | 2019-07-02 | 腾讯科技（深圳）有限公司 | Merchant type recognition methods and device |
| CN108281184A (en) * | 2018-02-01 | 2018-07-13 | 中国科学院上海微系统与信息技术研究所 | Myocardial ischemia based on machine learning examines method, storage medium and electronic equipment in advance |
| CN108776805A (en) * | 2018-05-03 | 2018-11-09 | 北斗导航位置服务（北京）有限公司 | It is a kind of establish image classification model, characteristics of image classification method and device |
| CN109117864A (en) * | 2018-07-13 | 2019-01-01 | 华南理工大学 | Coronary heart disease risk prediction technique, model and system based on heterogeneous characteristic fusion |
| CN109241227A (en) * | 2018-09-03 | 2019-01-18 | 四川佳联众合企业管理咨询有限公司 | Space-time data based on stacking Ensemble Learning Algorithms predicts modeling method |
| CN109145175A (en) * | 2018-09-03 | 2019-01-04 | 四川佳联众合企业管理咨询有限公司 | Space-time data prediction technique based on stacking Ensemble Learning Algorithms |
| CN109145175B (en) * | 2018-09-03 | 2023-05-30 | 成都卡普数据服务有限责任公司 | Spatiotemporal data prediction method based on stacking integrated learning algorithm |
| CN109460420A (en) * | 2018-09-28 | 2019-03-12 | 北京赛博贝斯数据科技有限责任公司 | The Knowledge Discovery Method and system of structural data |
| CN109685104A (en) * | 2018-11-15 | 2019-04-26 | 同盾控股有限公司 | Determination method and device for recognition model |
| CN109657883A (en) * | 2019-01-28 | 2019-04-19 | 重庆邮电大学 | A kind of bank branches recommended method based on multi-source data driving |
| CN110264462A (en) * | 2019-06-25 | 2019-09-20 | 电子科技大学 | A kind of breast ultrasound tumour recognition methods based on deep learning |
| CN110264462B (en) * | 2019-06-25 | 2022-06-28 | 电子科技大学 | Deep learning-based breast ultrasonic tumor identification method |
| CN110365535A (en) * | 2019-07-15 | 2019-10-22 | 阿里巴巴集团控股有限公司 | A kind of model loading method and device based on multi-party computations |
| CN110365535B (en) * | 2019-07-15 | 2022-03-04 | 创新先进技术有限公司 | Model online method and device based on secure multi-party computation |
| CN111199343A (en) * | 2019-12-24 | 2020-05-26 | 上海大学 | Multi-model fusion tobacco market supervision abnormal data mining method |
| CN111222575A (en) * | 2020-01-07 | 2020-06-02 | 北京联合大学 | KLXS multi-model fusion method and system based on HRRP target recognition |
| CN111222575B (en) * | 2020-01-07 | 2023-11-17 | 北京联合大学 | KLXS multi-model fusion method and system based on HRRP target recognition |
| CN111507507A (en) * | 2020-03-24 | 2020-08-07 | 重庆森鑫炬科技有限公司 | Big data-based monthly water consumption prediction method |
| CN113780339A (en) * | 2021-08-03 | 2021-12-10 | 阿里巴巴（中国）有限公司 | Model training, predicting and content understanding method and electronic equipment |
| CN113780339B (en) * | 2021-08-03 | 2024-03-29 | 阿里巴巴（中国）有限公司 | Model training, predicting and content understanding method and electronic equipment |
| CN113961774A (en) * | 2021-12-22 | 2022-01-21 | 四川新网银行股份有限公司 | Recommendation method for multi-feature combination strategy |
| CN113961774B (en) * | 2021-12-22 | 2022-04-15 | 四川新网银行股份有限公司 | Recommendation method for multi-feature combination strategy |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN107301221A (en) | A kind of data digging method of multiple features dimension heap fusion | |
| Zhang et al. | A return-cost-based binary firefly algorithm for feature selection | |
| WO2021007801A1 (en) | Aluminum oxide comprehensive production index decision-making method based on multi-scale deep convolutional network | |
| CN106779087A (en) | A kind of general-purpose machinery learning data analysis platform | |
| CN107451101A (en) | It is a kind of to be layered integrated Gaussian process recurrence soft-measuring modeling method | |
| CN107209873A (en) | Hyper parameter for depth convolutional network is selected | |
| CN102314614B (en) | Image semantics classification method based on class-shared multiple kernel learning (MKL) | |
| CN107644235A (en) | Image automatic annotation method based on semi-supervised learning | |
| CN107563439A (en) | A kind of model for identifying cleaning food materials picture and identification food materials class method for distinguishing | |
| CN106845717A (en) | A kind of energy efficiency evaluation method based on multi-model convergence strategy | |
| CN106709754A (en) | Power user grouping method based on text mining | |
| CN107430703A (en) | Sequential picture sampling and storage to fine tuning feature | |
| CN107851198A (en) | Media categories | |
| CN107924486A (en) | Pressure for classification is sparse | |
| CN103020122A (en) | Transfer learning method based on semi-supervised clustering | |
| CN110009030A (en) | Sewage treatment method for diagnosing faults based on stacking meta learning strategy | |
| CN107301604A (en) | Multi-model fusion estimation system | |
| CN105654196A (en) | Adaptive load prediction selection method based on electric power big data | |
| CN105303195A (en) | Bag-of-word image classification method | |
| CN108256630A (en) | A kind of over-fitting solution based on low dimensional manifold regularization neural network | |
| CN106447133A (en) | Short-term electric load prediction method based on deep self-encoding network | |
| CN104091038A (en) | Method for weighting multiple example studying features based on master space classifying criterion | |
| CN104463221A (en) | Imbalance sample weighting method suitable for training of support vector machine | |
| Chou et al. | Imaging time-series with features to enable visual recognition of regional energy consumption by bio-inspired optimization of deep learning | |
| CN106097094A (en) | A kind of man-computer cooperation credit evaluation new model towards medium-sized and small enterprises |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| WD01 | Invention patent application deemed withdrawn after publication | ||
| WD01 | Invention patent application deemed withdrawn after publication |
Application publication date: 20171027 |