CN109447244A - A kind of advertisement recommended method of combination gating cycle unit neural network - Google Patents
A kind of advertisement recommended method of combination gating cycle unit neural network Download PDFInfo
- Publication number
- CN109447244A CN109447244A CN201811183803.XA CN201811183803A CN109447244A CN 109447244 A CN109447244 A CN 109447244A CN 201811183803 A CN201811183803 A CN 201811183803A CN 109447244 A CN109447244 A CN 109447244A
- Authority
- CN
- China
- Prior art keywords
- neural network
- text
- model
- advertisement
- network
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/044—Recurrent networks, e.g. Hopfield networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
- G06Q30/00—Commerce
- G06Q30/02—Marketing; Price estimation or determination; Fundraising
- G06Q30/0241—Advertisements
- G06Q30/0251—Targeted advertisements
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Computational Linguistics (AREA)
- Health & Medical Sciences (AREA)
- Biophysics (AREA)
- Artificial Intelligence (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- Life Sciences & Earth Sciences (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Biomedical Technology (AREA)
- Business, Economics & Management (AREA)
- Accounting & Taxation (AREA)
- Development Economics (AREA)
- Finance (AREA)
- Strategic Management (AREA)
- Entrepreneurship & Innovation (AREA)
- Game Theory and Decision Science (AREA)
- Economics (AREA)
- Marketing (AREA)
- General Business, Economics & Management (AREA)
- Character Discrimination (AREA)
Abstract
The present invention provides a kind of advertisement recommended method of combination gating cycle unit neural network, the text feature for the advertisement that this method extracts gating cycle unit neural network is inputted as the part of network, information in advertising words and advertisement descriptive text can be extracted, special disposal text feature.Gating cycle neural network can then be obtained by pre-training, and using trained network as one layer of advertisement recommended models, i.e., Text character extraction layer is trained end to end to realize one.New method can preferably extract the feature in advertisement text compared to the model of no Text character extraction layer, it was found that the semantic structure information in text between word and word, sentence and sentence, the problem of can disappearing simultaneously to avoid traditional Recognition with Recurrent Neural Network bring training gradient, and GRU network may be implemented once to input directly as a network layer of entire training pattern, output result can be obtained, completion is trained end to end, reduces intermediate steps.
Description
Technical field
The invention belongs to deep learning fields, gating cycle neural network are combined with range depth model end-to-end
Advertisement recommended method.
Background technique
In recent years, machine learning techniques continue to develop, and machine learning method has early had been supplied in the every field in life.
These living scenes closely bound up with us of webpage sorting, dining room recommendation etc. are all because the addition of machine learning becomes more
Efficiently and accurately.As our requirements for machine learning model are higher and higher, again by manually extraction, assemblage characteristic in the past
It goes cost of labor required for the mode of training machine learning model and calculating cost all higher and higher, and is generally unattainable
The effect that we want, and the artificial work for carrying out Feature Engineering can replace by tree-model as GBDT to a certain extent
Make, but calculation amount is still very big.
In this context, deep neural network high speed development gets up, and by the connection of neural network between layers, realizes
The more feature combination of high-order, saves the cost of a large amount of manual construction features.Therefore at present in industry under many scenes all
It is to be learnt using deep neural network.
But simple deep neural network, there are also problem, feature, which enters deep neural network, can compare high-order
Feature combination, the interpretation of this combination is not strong, and combined mode is also uncontrollable.Especially known to us
In the case where having extraordinary effect after certain feature combinations, we not can guarantee this feature energy yet in deep neural network
As we are combined with being willing to.Therefore, Google propose it is a kind of can be by the Memorability of linear model and depth nerve net
The model that the generalization ability of network combines --- extensiveness and intensiveness model.
On the other hand, under the scene that advertisement is recommended, be often related to the text information of some advertisements, traditional processing this
The method of a little text informations is usually to segment to text, then calculate its TF-IDF value, extracts text feature with this, or
The term vector of text information is obtained by the method for word2vec.The meeting when handling the semantic structure information of text of such method
Some shortcomings, while needing individually to handle text information in advance every time, it can not be trained, be increased end to end
Intermediate steps.
Gating cycle unit neural network (GRU) is the variant of length memory Recognition with Recurrent Neural Network, in natural language processing
There is good effect in field.More particularly to semantic structure etc. has very big advantage when having the problem of time structure.
Summary of the invention
The present invention provides a kind of method that the advertisement for combining gating cycle neural network with range depth model is recommended,
This method can preferably extract text feature in advertisement, and realize training end to end and prediction.
In order to reach above-mentioned technical effect, technical scheme is as follows:
A kind of advertisement recommended method of combination gating cycle unit neural network, comprising:
S1: pre-training gating cycle unit neural network obtains trained model as the advertisement recommended models
Text character extraction layer；
Door control unit neural network is a mutation of length memory Recognition with Recurrent Neural Network (LSTM).Therefore GRU is in a side
Face inherits advantage of the LSTM on the problem of processing feature length relies on, while can be avoided gradient and disappearing and gradient explosion,
On the other hand, GRU simplifies the structure of LSTM, more simpler than LSTM；Specifically, needing to GRU layers of progress pre-training, training
Good network will be as entire model network layer handles text feature.Pre-training needs to carry out GRU network preceding to biography
It broadcasts, calculates the weight and biasing that each layer is connect with neuron.Input reaches output layer by hidden layer, can obtain and really export
Error.Connecting quantity is updated by error Back-Propagation, final trained GRU network is negative as a layer network of final mask
Duty processing text feature；
S2: the input layer of other features of input in addition to advertisement text feature to advertisement recommended models；
Specifically, common continuous feature, discrete features are directly inputted to extensiveness and intensiveness model and are trained；
S3: input advertisement text content to Text character extraction layer；
S4: text feature, and the input as advertisement recommended models are extracted with Text character extraction layer；
Specifically, the text information of advertisement is input to GRU layer trained in advance, extracts and be transferred to after text feature extensively
Degree is trained with depth model；
S5: extensiveness and intensiveness model is trained, and wherein deep neural network part is propagated forward, and is calculated
Error, linear segment is directly trained, and error is calculated；
Specifically, deep neural network part to input data carry out a propagated forward, input data from input layer to
Hidden layer finally reaches output layer, by the calculating between each layer, is exported, at the same calculate reality output and truthful data it
Between error, the update for subsequent network parameter.Linear model part is calculated by using gradient descent method to loss function
Acquire optimal parameter；
S6: training error obtained in step S5, which can be fed back simultaneously in linear model and deep neural network model, to carry out
Parameter updates, and wherein deep neural network model passes through error Back-Propagation undated parameter；
Specifically, parameter and error that deep neural network part is returned according to linear segment carry out carry out Back-propagation
It broadcasts, updates every layer of parameter；
S7: executing S5 to S6 step repeatedly, until model is restrained.
Specifically, the output of linear segment and the output of deep neural network part are integrated by a weighting function
One output, provides the prediction result of model；
Further, the error Back-Propagation in the step S1 and S6 needs to seek parameters error respectively
It leads.
Further, the gradient descent method of linear model needs the parameter derivation to loss function in the step S5,
Direction of the direction of negative derivative as fitting is chosen, it can more rapid convergence.
Compared with prior art, the beneficial effect of technical solution of the present invention is:
The present invention realizes the training method end to end from text information to prediction output, and the prior art needs to mention mostly
It is preceding that text information is handled, text feature is extracted, this undoubtedly increases the intermediate steps of model training link, does not meet end
To the thought of end training.The present invention is pre-processed the GRU network of a pre-training as a text information before input layer
Layer, i.e. Text character extraction layer, then enter extensiveness and intensiveness model training after integrating other input datas.Wherein GRU network is
The mutation of LSTM network can not only extract the context relation in text, solve the problems, such as that shot and long term relies on, prevent gradient from disappearing
The case where gradient of becoming estranged is exploded, is more a simplified the structure of LSTM, original forgetting door, input gate and out gate is reduced to more
New door and resetting door, more succinctly.In addition to this, linear model and depth have then been merged as the extensiveness and intensiveness model of master cast
The advantages of spending both neural networks not only has extremely strong memory capability, while also having extremely strong generalization ability, so that entire mould
Type is excellent in.
Detailed description of the invention
Fig. 1 is flow chart of the invention；
Fig. 2 is GRU unit block structural diagram of the present invention；
The position Fig. 3 extensiveness and intensiveness model structure of the present invention.
Specific embodiment
The attached figures are only used for illustrative purposes and cannot be understood as limitating the patent；
In order to better illustrate this embodiment, the certain components of attached drawing have omission, zoom in or out, and do not represent actual product
Size；
To those skilled in the art, it is to be understood that certain known features and its explanation, which may be omitted, in attached drawing
's.
The following further describes the technical solution of the present invention with reference to the accompanying drawing.
As shown in Figure 1, a kind of advertisement recommended method of combination gating cycle unit neural network, the specific steps are as follows:
A. pre-training GRU network layer, i.e. Text character extraction layer；
The structure of GRU network is similar with RNN, but in GRU network, constitutes biography by the block that multiple GRU units are constituted
Hiding layer network in system network.The structure of GRU unit as shown in Fig. 2, GRU network unit by three door letters in LSTM
It has been melted into two: having updated door (by forgetting that door and input gate synthesize) and resetting door, i.e. z in figuretAnd rt.Door is updated for controlling
The status information of previous moment is brought into the degree in current state, updates the bigger state letter for illustrating previous moment of value of door
Breath is brought into more.Resetting door is used to control the degree for the status information for ignoring previous moment, and the smaller explanation of value for resetting door is ignored
It obtains more.Presence due to door enables GRU to store and access prolonged information plus cell state etc. is incorporated,
Prevent gradient from disappearing and gradient explosion issues.
Shown in Fig. 2, xtIndicate the input of t moment network, ht-1Indicate the hidden state at t-1 moment, htThen indicate t moment
Hidden state, rtRepresent the resetting door of t moment, ztRepresent the update door of t moment.Calculation is such as in the propagated forward of GRU
Under:
Reset door:
rt=σ (Wr·[ht-1, xt])
Update door:
zt=σ (Wz·[ht-1, xt])
Candidate hidden state:
Hidden state:
Output:
yt=σ (Wo·ht)
Wherein [] indicates that two vectors are connected, and * representing matrix element multiplication, σ indicates sigmoid function.
According to the calculation method of GRU propagated forward, that need training is parameter WrWzWhWo, wherein first three weight all belongs to
In splicing weight, it is as follows to need to do other processing in study:
Wr=Wrx+Wrh
For the output layer of GRU network, input are as follows:
The output of output layer are as follows:
According to output, the loss function of t moment, i.e. gap between model predication value and true value can be calculated:
Therefore the error that sample can be calculated updates network parameter for subsequent error Back-Propagation:
In the error Back-Propagation the step of, the error acquired to previous step is needed to seek local derviation to parameters respectively, in addition to
Including above-mentionedIt further include the neuron weight W of output layero, error Back-Propagation knot
Fruit is as follows:
Wherein:
The derivative of σ ' expression sigmoid function.
Network parameter is constantly updated by above-mentioned propagated forward and error Back-Propagation, a given text is obtained after the completion of training
This input exports the network of text feature, using this network as the Text character extraction layer of entire model, completes GRU pre-training.
B. the text data in input data is to Text character extraction layer；
Input data is divided into two parts, and a part is text data, and another part is general data.Its text data is just
It individually enters the good Text character extraction layer of pre-training in the step A and extracts text feature, since this layer of GRU network is pre-
Training is completed, therefore can obtain text feature quickly.
C. text feature and other features are input to extensiveness and intensiveness model；
The text feature extracted in the step B by Text character extraction layer, it is such as continuous special with other general features
Sign, discrete special data characteristics etc. enter extensiveness and intensiveness model training as input together.
D. extensiveness and intensiveness model training, undated parameter；
Depth and broadness model is combined both linear model and deep neural network, while linear mould is utilized
The good memory capability of type and the powerful generalization ability of deep neural network, are mutually transmitted by parameter, error, realize two
The joint training of model allows two models to be respectively dedicated to the work being most good at, to obtain good effect.
Linear model part uses logistic regression, is carried out using the method that gradient declines to objective function in logistic regression excellent
Change iteration, find so that objective function i.e. loss function minimum when parameter value, the final argument as model.Method is as follows:
By maximizing, likelihood function can be obtained:
To the expression formula that likelihood function logarithmetics negate, i.e. loss function expression formula are as follows:
The form for being written as matrix is as follows:
J (θ)=- Y ⊙ loghθ(X)-(E-Y)⊙log(E-hθ(X))
Parameter θ derivation is obtained using gradient descent method:
It can be obtained after abbreviation:
Therefore, the iterative manner of parameter θ is as follows:
θ=θ-α XT(hθ(X)-Y)
Wherein, α is step-length.
For deep neural network part, network parameter is updated using propagated forward and Back-propagation broadcasting method, forward direction passes
Broadcasting method is as follows:
Total number of plies L, all hidden layers and the corresponding matrix W of output layer, bias vector b, input value vector x；Initialize a1
=x
For subsequent each layer, calculate:
al=σ (zl)=σ (Wlal-1+bl)
Then aLIt as exports, according to output valve, we can calculate the error between true value:
By al=σ (zl)=σ (Wlal-1+bl) above formula and to W is substituted into, b derivation obtains:
Using chain rule, derivation can be carried out to the parameter of the hidden layer of front:
It enablesThen:
Declined simultaneously using gradient, update each layer of parameter:
Operation as above is repeated, until two parts model is restrained.By such method, linear segment and depth mind
The training of parameter can be completed through network portion, after the completion of two parts training, by by the function of a weighting, will most be terminated
Fruit is calculated.
The present invention is directed to propose a kind of advertisement recommended method of combination gating cycle unit neural network.Method mainly exists
A door control unit neural network is added before extensiveness and intensiveness model as the feature extraction layer for extracting text information, is realized to text
The extraction of this information, to complete to train end to end.
Its features and advantages are as follows: the text feature in advertisement can be preferably extracted compared to general advertisement recommended method, and
And the pre-training of GRU network, it realizes from text to the end-to-end realization finally predicted, reduces intermediate steps, while GRU net
Network compared with the methods of other TF-IDF, can preferably extract shot and long term in text rely on, context relation, extract
Text feature is more accurate.Extensiveness and intensiveness model can play the memory capability and deep neural network of linear model simultaneously
Superpower generalization ability, Each performs its own functions, and accuracy is higher.
Therefore GRU network is combined with extensiveness and intensiveness model, the text feature in advertisement can be extracted well, it is real
Existing end-to-end training, and generalization ability can be combined with memory capability, excavate the relationship between feature.
Described in attached drawing positional relationship for only for illustration, should not be understood as the limitation to this patent.
Claims (4)
1. a kind of advertisement recommended method of combination gating cycle unit neural network, which is characterized in that the advertisement recommended models
For extensiveness and intensiveness model (wide and deep), model includes linear segment and deep neural network part, including following step
It is rapid:
S1: pre-training gating cycle unit neural network obtains text of the trained model as the advertisement recommended models
Feature extraction layer；
S2: the input layer of other features of input in addition to advertisement text feature to advertisement recommended models；
S3: input advertisement text content to Text character extraction layer；
S4: text feature, and the input as advertisement recommended models are extracted with Text character extraction layer；
S5: extensiveness and intensiveness model is trained, and wherein deep neural network part is propagated forward, and mistake is calculated
Difference, linear segment is directly trained, and error is calculated；
S6: training error obtained in step S5 can feed back progress parameter update in linear model and dnn model simultaneously, wherein
Dnn model passes through error Back-Propagation undated parameter；
S7: executing S5 to S6 step repeatedly, until model is restrained.
2. the advertisement recommended method of combination gating cycle unit neural network according to claim 1, which is characterized in that step
The pre-training of gating cycle neural network in rapid S1 calculates the error for obtaining network by the propagated forward of neural network, then leads to
It crosses error Back-Propagation and updates the parameter of each layer of network, until obtaining convergent result.
3. the advertisement recommended method of combination gating cycle unit neural network according to claim 1, which is characterized in that step
Text character extraction layer in rapid S4 can be extracted the semantic structure information in text, be better than by GRU network implementations
The result that common bag of words are connected to.
4. the advertisement recommendation side of combination gating cycle unit neural network according to claim 1 and probabilistic neural network
Method, which is characterized in that the error of step S6 feed back between linear model and dnn model be it is mutual, the weight of single model is more
Newly it will receive the end wide and the end deep to the joint effect of model training error.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811183803.XA CN109447244A (en) | 2018-10-11 | 2018-10-11 | A kind of advertisement recommended method of combination gating cycle unit neural network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811183803.XA CN109447244A (en) | 2018-10-11 | 2018-10-11 | A kind of advertisement recommended method of combination gating cycle unit neural network |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN109447244A true CN109447244A (en) | 2019-03-08 |
Family
ID=65545961
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201811183803.XA Pending CN109447244A (en) | 2018-10-11 | 2018-10-11 | A kind of advertisement recommended method of combination gating cycle unit neural network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109447244A (en) |
Cited By (9)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109993668A (en) * | 2019-04-09 | 2019-07-09 | 桂林电子科技大学 | A kind of recommending scenery spot method based on gating cycle unit neural network |
| CN110225460A (en) * | 2019-06-05 | 2019-09-10 | 三维通信股份有限公司 | A kind of indoor orientation method and device based on deep neural network |
| CN110222894A (en) * | 2019-06-06 | 2019-09-10 | 阿里巴巴集团控股有限公司 | Advertisement placement method, device and equipment |
| CN110299194A (en) * | 2019-06-06 | 2019-10-01 | 昆明理工大学 | The similar case recommended method with the wide depth model of improvement is indicated based on comprehensive characteristics |
| CN110716964A (en) * | 2019-09-19 | 2020-01-21 | 卓尔智联(武汉)研究院有限公司 | Newborn naming method based on GRU network, electronic device and storage medium |
| CN111222628A (en) * | 2019-11-20 | 2020-06-02 | 深圳前海微众银行股份有限公司 | Method, device and system for optimizing recurrent neural network training and readable storage medium |
| CN113288050A (en) * | 2021-04-23 | 2021-08-24 | 山东师范大学 | Multidimensional enhanced epileptic seizure prediction system based on graph convolution network |
| CN114462584A (en) * | 2022-04-11 | 2022-05-10 | 北京达佳互联信息技术有限公司 | Recommendation model training method, recommendation device, server and medium |
| CN114500197A (en) * | 2022-01-24 | 2022-05-13 | 华南理工大学 | Method, system, device and storage medium for equalization after visible light communication |
Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106682217A (en) * | 2016-12-31 | 2017-05-17 | 成都数联铭品科技有限公司 | Method for enterprise second-grade industry classification based on automatic screening and learning of information |
| CN107577662A (en) * | 2017-08-08 | 2018-01-12 | 上海交通大学 | Towards the semantic understanding system and method for Chinese text |
| CN107807971A (en) * | 2017-10-18 | 2018-03-16 | 北京信息科技大学 | A kind of automated graphics semantic description method |
| CN108416625A (en) * | 2018-02-28 | 2018-08-17 | 阿里巴巴集团控股有限公司 | The recommendation method and apparatus of marketing product |
| CN108596645A (en) * | 2018-03-13 | 2018-09-28 | 阿里巴巴集团控股有限公司 | A kind of method, apparatus and equipment of information recommendation |
| US20180336889A1 (en) * | 2017-05-19 | 2018-11-22 | Baidu Online Network Technology (Beijing) Co., Ltd . | Method and Apparatus of Building Acoustic Feature Extracting Model, and Acoustic Feature Extracting Method and Apparatus |
-
2018
- 2018-10-11 CN CN201811183803.XA patent/CN109447244A/en active Pending
Patent Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106682217A (en) * | 2016-12-31 | 2017-05-17 | 成都数联铭品科技有限公司 | Method for enterprise second-grade industry classification based on automatic screening and learning of information |
| US20180336889A1 (en) * | 2017-05-19 | 2018-11-22 | Baidu Online Network Technology (Beijing) Co., Ltd . | Method and Apparatus of Building Acoustic Feature Extracting Model, and Acoustic Feature Extracting Method and Apparatus |
| CN107577662A (en) * | 2017-08-08 | 2018-01-12 | 上海交通大学 | Towards the semantic understanding system and method for Chinese text |
| CN107807971A (en) * | 2017-10-18 | 2018-03-16 | 北京信息科技大学 | A kind of automated graphics semantic description method |
| CN108416625A (en) * | 2018-02-28 | 2018-08-17 | 阿里巴巴集团控股有限公司 | The recommendation method and apparatus of marketing product |
| CN108596645A (en) * | 2018-03-13 | 2018-09-28 | 阿里巴巴集团控股有限公司 | A kind of method, apparatus and equipment of information recommendation |
Cited By (14)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109993668B (en) * | 2019-04-09 | 2021-07-13 | 桂林电子科技大学 | Scenic spot recommendation method based on gated cyclic unit neural network |
| CN109993668A (en) * | 2019-04-09 | 2019-07-09 | 桂林电子科技大学 | A kind of recommending scenery spot method based on gating cycle unit neural network |
| CN110225460B (en) * | 2019-06-05 | 2021-03-23 | 三维通信股份有限公司 | Indoor positioning method and device based on deep neural network |
| CN110225460A (en) * | 2019-06-05 | 2019-09-10 | 三维通信股份有限公司 | A kind of indoor orientation method and device based on deep neural network |
| CN110299194A (en) * | 2019-06-06 | 2019-10-01 | 昆明理工大学 | The similar case recommended method with the wide depth model of improvement is indicated based on comprehensive characteristics |
| CN110222894A (en) * | 2019-06-06 | 2019-09-10 | 阿里巴巴集团控股有限公司 | Advertisement placement method, device and equipment |
| CN110716964A (en) * | 2019-09-19 | 2020-01-21 | 卓尔智联(武汉)研究院有限公司 | Newborn naming method based on GRU network, electronic device and storage medium |
| CN111222628A (en) * | 2019-11-20 | 2020-06-02 | 深圳前海微众银行股份有限公司 | Method, device and system for optimizing recurrent neural network training and readable storage medium |
| CN111222628B (en) * | 2019-11-20 | 2023-09-26 | 深圳前海微众银行股份有限公司 | Method, device, system and readable storage medium for optimizing training of recurrent neural network |
| CN113288050A (en) * | 2021-04-23 | 2021-08-24 | 山东师范大学 | Multidimensional enhanced epileptic seizure prediction system based on graph convolution network |
| CN114500197A (en) * | 2022-01-24 | 2022-05-13 | 华南理工大学 | Method, system, device and storage medium for equalization after visible light communication |
| CN114500197B (en) * | 2022-01-24 | 2023-05-23 | 华南理工大学 | Method, system, device and storage medium for equalizing after visible light communication |
| CN114462584A (en) * | 2022-04-11 | 2022-05-10 | 北京达佳互联信息技术有限公司 | Recommendation model training method, recommendation device, server and medium |
| CN114462584B (en) * | 2022-04-11 | 2022-07-22 | 北京达佳互联信息技术有限公司 | Recommendation model training method, recommendation device, server and medium |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109447244A (en) | A kind of advertisement recommended method of combination gating cycle unit neural network | |
| CN111488734B (en) | Emotional feature representation learning system and method based on global interaction and syntactic dependency | |
| CN110163299B (en) | Visual question-answering method based on bottom-up attention mechanism and memory network | |
| CN107239444B (en) | A kind of term vector training method and system merging part of speech and location information | |
| CN106022237B (en) | A kind of pedestrian detection method of convolutional neural networks end to end | |
| CN109299396A (en) | Merge the convolutional neural networks collaborative filtering recommending method and system of attention model | |
| CN110609899B (en) | Specific target emotion classification method based on improved BERT model | |
| CN110728541B (en) | Information streaming media advertising creative recommendation method and device | |
| CN108664632A (en) | A kind of text emotion sorting algorithm based on convolutional neural networks and attention mechanism | |
| CN109740154A (en) | A kind of online comment fine granularity sentiment analysis method based on multi-task learning | |
| CN108038205B (en) | Viewpoint analysis prototype system for Chinese microblogs | |
| CN109597891A (en) | Text emotion analysis method based on two-way length Memory Neural Networks in short-term | |
| CN110083700A (en) | A kind of enterprise's public sentiment sensibility classification method and system based on convolutional neural networks | |
| Chennupati et al. | Auxnet: Auxiliary tasks enhanced semantic segmentation for automated driving | |
| CN109582956A (en) | text representation method and device applied to sentence embedding | |
| CN107066445A (en) | The deep learning method of one attribute emotion word vector | |
| CN111460157B (en) | Cyclic convolution multitask learning method for multi-field text classification | |
| CN110866542A (en) | Depth representation learning method based on feature controllable fusion | |
| CN108427665A (en) | A kind of text automatic generation method based on LSTM type RNN models | |
| CN109284361A (en) | A kind of entity abstracting method and system based on deep learning | |
| CN108647191A (en) | It is a kind of based on have supervision emotion text and term vector sentiment dictionary construction method | |
| CN110263165A (en) | A kind of user comment sentiment analysis method based on semi-supervised learning | |
| CN109101629A (en) | A kind of network representation method based on depth network structure and nodal community | |
| CN109584006A (en) | A kind of cross-platform goods matching method based on depth Matching Model | |
| CN109117943A (en) | Utilize the method for more attribute informations enhancing network characterisation study |
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 | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20190308 |
|
| RJ01 | Rejection of invention patent application after publication |