CN109902201A - A kind of recommended method based on CNN and BP neural network - Google Patents
A kind of recommended method based on CNN and BP neural network Download PDFInfo
- Publication number
- CN109902201A CN109902201A CN201910175261.XA CN201910175261A CN109902201A CN 109902201 A CN109902201 A CN 109902201A CN 201910175261 A CN201910175261 A CN 201910175261A CN 109902201 A CN109902201 A CN 109902201A
- Authority
- CN
- China
- Prior art keywords
- user
- cnn
- model
- matrix
- neural 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
Landscapes
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
Abstract
A kind of recommended method based on CNN and BP neural network, to the effect that estimates the hobby of target user, the item sequence that user may like recommends target user by a series of behaviors of target user.In recommended models, the content of core is the model based on score in predicting.The new model CNN-BP proposed through the invention predicts next consumer behavior of user, was analyzed before this using convolutional neural networks CNN behavior sequence, and obtained every kind of elementary probability for not watching project；Then final probability is calculated using reverse Propagation Neural Network BP；Finally, N before being selected according to final probability set, and the project recommendation watched using this preceding N project as target user's most probable next time is to user.This method be by taking the recommended models of film as an example, the experimental results showed that, CNN-BP model can more accurately predict the watching behavior next time of target user, and provide superior performance for recommender system.
Description
Technical field
The present invention relates to recommended models, specifically provide a kind of recommended method based on CNN and BP neural network.
Background technique
Have much in previous various traditional proposed algorithms for the various problems that recommender system faces
Research achievement is simultaneously widely applied by each field, such as content-based recommendation algorithm, Knowledge based engineering proposed algorithm and mixing
Proposed algorithm etc..In recent years, deep learning suffers from good effect in natural language processing and image procossing, mainly has
Convolutional neural networks CNN, recurrent neural network and BP neural network etc., and then recommending Application of Neural Network there are many scholar
In system.
At abroad, Falk J [1] et al. proposes a kind of recommender system based on feedforward neural network, finally counts and assess
The performance of recommender system.Netflix [2] et al. proposes a kind of video recommendation system based on deep neural network DNN, the model
Input be the historical behavior of user's viewing and the information of user, then learnt using the DNN network model of multilayer, it is last straight
Tap into row Top-N recommendation.And Murat Aykanat [3] et al. is then to utilize convolutional neural networks CNN and support vector machines SVW
To realize accurate classification.At home, practice thread precious [4] and propose and a kind of pushed away based on what convolutional neural networks CNN and sequence learnt
System is recommended, on electric business platform, the commodity that user needs to buy next time, related experiment are predicted using convolutional neural networks CNN
It shows this method and contribution is made that the quality of recommendation to a certain extent.Wu's alkene [5] be then propose it is a kind of using convolution
The principle of neural network CNN reaches matched recommendation, and main purpose reduced as far as possible under the premise of precision does not decline
The time of algorithm and the accuracy for improving algorithm, and this method and traditional algorithm are analyzed compared with.Sun Qian [6] research
By Application of Neural Network in proposed algorithm, primarily directed to how alleviating data sparsity problem, therefore in score in predicting
When introduce neural network, and the nearest-neighbors collection of user is found according to the shared part of user's scoring, rating matrix it is dilute
Thin property is extenuated.
It is found in Revision status, with growing information resources, traditional recommender system is can not be complete
Full reply, and after the continuous research to deep learning, there is breakthrough in fields such as image, voice and text classifications
The achievement of property.Since neural network has the characteristics that robustness, learning ability, non-linear and concurrency, and a large amount of scholars
Research practice all has been described above neural network and can be applied in recommender system, and neural network has better effect.But
Recommender system based on deep learning still primarily directed to the scoring behavior of user, and mainly by single neural network come
Training study.So this method joined more fully user behavior, single model is not then used, but is used
The fusion of two models finally carries out the score in predicting of film with trained new model.
The calculation formula of Classical forecast scoring has certain limitation, and it is not very accurate for causing the precision recommended, so
The precision of recommendation is promoted method proposes a new model.Have in mainly studying: firstly, in above-mentioned three kinds of history rows
For on the basis of, method introduces the behavior instantly of a fashion elements, these four behaviors are then based on CNN model to instruct
Practice, using four kinds of probability of output as the input of BP model, finally exports a final probability.This method using neural network come
Prediction scoring is carried out to film, primarily to promoting the precision recommended.
[1] Matthias J ü ttner, Falk J, Rucker W M.A neural network based
recommendation system for solvers and preconditioners for systems of linear
Equations [C] .Electromagnetic Field Computation, IEEE, 2017
[2] Hidasi B, Karatzoglou A, Baltrunas L, et al.Session-based
Recommendations with Recurrent Neural Networks [J] .Computer Science, 2015
[3] Aykanat M, Ozkan K, Kurt B, et al.Classification of lung sounds using
Convolutional neural networks [J] .Eurasip Journal on Image&Video Processing,
2017, (1): 65
[4] practice proposed algorithm research [D] Dalian University of Technology of the thread treasured based on sequence study and convolutional neural networks,
2016
[5] proposed algorithm [D] South China Science & Engineering University of Wu's alkene based on convolutional neural networks, 2016
[6] research of proposed algorithm of the Sun Qian based on LM-BP neural network and application [D] Beijing Jiaotong University, 2016
Summary of the invention
The present invention proposes a kind of new score in predicting model, it is therefore an objective to promote the precision of recommendation.Therefore it mentions in the method
A new model CNN-BP predicts next watching behavior of user out, then recommends prediction project again for user.This method
In analyzed before this using four kind behavior sequences of the convolutional neural networks CNN to user, i.e., user subjective scoring behavior, use
The objective behavior in family, time-based user preference behavior and the behavior of fashion elements instantly so that it can be concluded that every kind not
Four kinds of elementary probabilities of viewing project.Then corresponding final probability is calculated using reverse Propagation Neural Network BP.Finally, according to
Choose preceding N projects of the final probability of highest among these according to the final probability set calculated, and using this preceding N project as
They are recommended user by the project of target user's most probable next time viewing one by one.This method is to be with the recommended models of film
Example, the experimental results showed that, CNN-BP model can more accurately predict the watching behavior next time of target user, and to recommend system
System provides superior performance.
Recommended method provided by the invention based on CNN and BP neural network mainly includes following committed step:
1st, user's different type behavior:
1.1st, initial data set is handled, obtains user-user attribute matrix, user-project general comment sub-matrix
With fashion elements behavioural matrix；
1.2nd, user preference matrix is constructed, to indicate user to the preference of film native；
2nd, the hidden layer of CNN model:
2.1st, feature is extracted using convolutional layer；
2.2nd, using pond layer come compressive features；
2.3rd, full articulamentum；
2.4th, overfitting problem in CNN network；
3rd, the score in predicting based on CNN network:
3.1st, the algorithmic procedure of the score in predicting based on CNN network；
4th, final probability is obtained using BP:
4.1st, the calculation of final probability.
The advantages and positive effects of the present invention are:
A kind of recommender system based on CNN and BP neural network of this method major design, this method are based on two lists
Only model carries out fusion recommendation.The calculation formula of Classical forecast scoring has certain limitation, and the precision recommended is caused not to be
It is highly desirable, so this method proposes a kind of new model to promote the precision of recommendation.Firstly, method introduces four different types
User behavior, be then based on these four Behavioral availabilities CNN model to train, using four kinds of probabilities of output as BP model
Input, finally export a final probability.This method carries out prediction scoring to film using neural network, primarily to
Promote the precision recommended.
Detailed description of the invention
Fig. 1 is the process of CNN-BP model；
Fig. 2 is the local sensing in CNN model convolutional layer；
Fig. 3 is the parameter sharing in CNN model convolutional layer；
Fig. 4 is the process of convolution in CNN model；
Fig. 5 is the process in maximum pond in CNN model basin layer；
Fig. 6 is the model of BP；
Fig. 7 is machine learning and the comparison of context of methods；
Fig. 8 is the comparison of traditional algorithm and this paper algorithm.
Specific embodiment
Based on CNN model and BP model, this method passed through data prediction before this and was used the method that the present invention designs
Then four kinds of behaviors at family carry out K-Means cluster using the first three behavior for not including fashion elements behavior, then for each
A class cluster carries out the training and study of neural network.Selection in relation to neural network, this method are the fusions using CNN and BP
Model considers four kinds of behaviors of above-mentioned user in model, obtained original probability by the training of CNN model before this, recycles
BP model obtains final probability, and preceding N high films of probability are recommended user.Experimental result shows that the precision of recommendation obtains
Promotion is arrived, the effect of recommendation is also improved.
The method of the present invention mainly includes following committed step:
1st, fundamental matrix:
1.1st, by handling initial data set, user-user attribute matrix, user-project overall score are obtained
Its corresponding matrix is denoted as V, F ' and N respectively by matrix and fashion elements behavioural matrix.
The form of V is as follows:
Wherein: k represents the number of user property, attribute, that is, user ID of user, gender, age bracket, occupation etc.；Matrix
In elementRepresent user uiIn the objective characteristics attribute g of its ownkOn value.
The form of F ' is as follows:
Wherein: element in matrixRepresent user uiTo watching and contained characteristic attribute gkAll films
Overall score, calculation formula are as follows:
Wherein:Represent user uiTo each film mjScoring；It is made of number 0 and 1,0 represents
Film mjNot comprising item characteristic attribute gk；1 represents film mjInclude item characteristic attribute gk。
The form of N is as follows:
Wherein:It is made of number 0 and 1,0 represents film mjNot comprising fashion elements gk；1 represents film mj
Include item characteristic attribute gk。
1.2nd, user preference matrix is constructed, to indicate that user to the preference of film native, is denoted as P ', the form of P ' is such as
Under:
Wherein: element in matrixRepresent the user u under time factoriTo item characteristic attribute gkPreference journey
Degree；Steps are as follows for calculating:
First a forgetting function is defined according to this great forgetting curve of Chinese mugwort guest:
Then the preference of user is constructed using TF-IDF algorithm and forgetting function, formula is as follows:
Wherein:And be made of number 0 and 1,0 represents user uiTo film mjIt does not score, i.e., user does not have
Watched the film；1 opposite.
2nd, the hidden layer of CNN model:
2.1st, feature is extracted using convolutional layer:
It is local sensing first, as shown in Fig. 2, its theoretical basis is derived from and grinds to animal nerve unit connection type
Study carefully.All connected entirely between the neuron on each layer in fact, but with later slowly the study found that part connection
Can equally obtain it is ideal as a result, and the parameter in network can also be reduced, also just produce local sensing.
In general, people is also then to converge to the overall situation since locally to the cognition of extraneous things.Just as neuron one
Sample, the neuron correlation being closer can be relatively strong, and the correlation between farther away neuron is obvious
It is weaker, so, It is not necessary to be connected entirely between them, it is only necessary to local sensing is carried out, finally by upper one layer
Neuron converge those local sensings, thus information to the end.
Followed by parameter sharing, because above-mentioned local sensing is only the reduction of a part of parameter, for remaining entirety
Parameter, quantity are still very big, so can reach largely reducing for number of parameters using parameter sharing.Parameter sharing
Figure is as shown in figure 3, the connection arrow of each same type in figure represents each connection weight, the original that can be done so
Because there is two aspects, on the one hand, duplicate neural unit can carry out the identification of feature in the case where not considering position；Another party
Face, parameter sharing can still make our available effective spies in the case where reducing the variable quantity for needing to learn
Levy the result extracted.
It is finally convolution, the volume of the calculating operation actually carried out on each layer, i.e. input layer and weight parameter
Long-pending and convolutional neural networks CNN name source.Illustrate the process of convolution with an example, understand for convenience, herein
The image of 5 × 5 sizes and the convolution kernel of 3 × 3 sizes are had chosen, and convolution kernel has 9 parameters, also
It is one to share 9 corresponding neurons on convolution kernel.The detailed process of convolution is as shown in Figure 4.
Remaining calculating is carried out according to above-mentioned rule, after convolution, forms the spy of 3 × 3 sizes
Sign figure.Each convolution represents a kind of mode, the i.e. extraction of feature in fact, the whole process of convolution, like a sieve,
Our parts of requirement will be met all to screen.
2.2nd, using pond layer come compressive features:
After convolutional calculation, feature is obtained, is next exactly to go to classify for foundation with these features, when
So, it theoretically can use what classifier went to realize, such as classifier softmax, but have the appearance of overfitting problem.Example
Such as: the image for being 50*50 now with a pixel has passed through study at present and has obtained 100 features, and being is being defeated with 4*4
On the basis of entering, then, any one obtained feature can generate a dimension after carrying out convolution with image as (50-4+1)
The convolution feature of × (50-4+1)=2209, and because there is 100 features, finally will form a dimension be 2209 ×
The vector of 100=220900 convolution feature.If the input feature vector quantity of a classifier is bigger, it is very not square that study, which is got up,
Just, and also easily there is overfitting problem.
Pond is in CNN network or has critically important status, pondization namely down-sampling, its most important purpose
It is to reduce characteristic pattern, that is, the reduction by the dimension of feature as far as possible can also be fine so as to reduce the complexity of calculating
Avoid over-fitting this problem.In the layer of pond there are two types of the operations of most Chang Shiyong, one is maximum ponds, i.e., in selection pond
Maximum value；Another kind is mean value pond, i.e., the mean value of all the points in selection pond.
Maximum pond process is as shown in figure 5, be to be based on the assumption that the scale of pond layer is 2*2, stride is 2, i.e., to four
The maximum pond of number, if the input in the layer of pond is not 2 multiple, it is only necessary to Chi Huaqian using edge zero padding this
The unit of input is mended 2 multiple at needs by method.
2.3rd, full articulamentum:
The front layer of CNN model is mainly to be made of many a convolution sum ponds, and the last layer is full articulamentum, Quan Lian
It connects layer and is equivalent to a classifier.From the foregoing description it is known that convolutional layer is the feature for extracting input data, pond
Changing layer is that the feature that will be extracted is compressed so that available more relatively important feature, and full articulamentum is then whole
The feature being previously obtained is closed, is normalized, a probability to every kind of situation of classifying, last classifier then can be exported
It can be that foundation is classified with this probability.For the classifier of full articulamentum, softmax recurrence, logistic regression and supporting vector
Machine can be used.
2.4th, overfitting problem in CNN network；
In the training carried out to CNN network, if there is parameter when excessive or repetition training data when
It waits, with regard to it is easy to appear overfitting problems and time-consuming problem.It is usually all using to relevant in model for both of these problems
Partial parameters are abandoned, the model to reduce the quantity of Model Parameter, that is, before meaning present model relatively
A degree of simplification has been obtained, that is, has reduced the probability for overfitting problem occur.The Dropout used herein is also to improve
One of the method for CNN model generalization ability, can not only be such that overfitting problem is effectively alleviated, but also in certain journey
Also the effect of regularization may be implemented on degree.
It is using the process of Dropout: firstly, the case where input layer and output layer epineural member quantity remain unchanged
Under, the interim neuron for having hidden half；Secondly, before being carried out using the model for concealing half neuron to the data of input
To propagating and backpropagation, and undated parameter (W, b)；Then, the neuron being hidden is shown, the neuron being hidden at this time does not have
There is any change, and update has been obtained in the neuron stayed before；Finally, needing the continuous iteration above process.
3rd, the score in predicting based on CNN network:
3.1st, the algorithmic procedure of the score in predicting based on CNN network:
In recommender system based on film, the data volume of user behavior be in fact it is very huge, behavioural characteristic is also relatively multiple
Miscellaneous, when carrying out traditional similarity calculation with prediction scoring, related Traditional calculating methods cannot be according to the similarity of user
Carry out calculating rapidly and efficiently.And CNN network has efficient parallel processing capability, can recognize related data and prediction energy
The nonlinear model of the height of power, so CNN network application is herein defined as feasible.
In the recommender system of entire film, the score in predicting of film is a part of core, passes through the row of target user
To predict it for not watching the scoring flashed back past events.It is the problem of for CNN network, in fact most important to seek to pass through study
Obtain a weight matrix W and its biasing b of most suitable this paper, so need to set some parameters before training, wherein
The quantity and size, the step-length of convolution, the quantity of pond layer, Chi Hua of convolution kernel in quantity, each convolutional layer including convolutional layer
Method, the method etc. that calculates of the scale in pond, the selection of activation primitive, loss function.And the core concept of CNN model is benefit
It is trained with continuous iteration, then reduces error function by the use of gradient descent method, and it is anti-to update result
To each layer is transmitted to, the bias b of weighted value W and it are finally constantly updated using difference.
In summary the algorithmic procedure of introduction, the score in predicting herein based on CNN network is as follows:
1) clustering cluster and behavioral data, the data of convection current row element after obtaining cluster are pre-processed；
2) data in clustering cluster are divided into two parts, a part is used as training set, and another part is used as surveying
Examination collection.And this training sample is then random n input datas and output number corresponding thereto to be selected from training set
According to being denoted as respectively:
X (n)=(x1(n),x2(n),......,xk(n)) (8)
D (n)=(d1(n),d2(n),......,dt(n)) (9)
3) CNN network is initialized, defines various network parameters mentioned above first, then utilize random side
Method is to the weighted value and bias assignment in CNN network；
4) the selected part data from training set, as input data, the input layer being input in CNN is arrived and is hidden
The formula of layer are as follows:
hi=Relu (Wxi+b) (10)
Wherein: Relu is activation primitive, and W weighted value, b is bias；
5) subsequently into pond layer, pond method is used to it, herein using maximum pond method.Maximum Chi Huafang
The formula of method is as follows:
6) after CNN network exports result, pass through the difference of output result and practical CNN network output result in training set
Value constructs error function, then seeks local derviation to error function respectively, about the local derviation of weighted value W and bias b, passes through calculating
As a result come reversely to adjust the W and b in each layer in front.The formula of error function is as follows:
7) weighted value W and bias b are updated by continuous iteration；
8) if the value of error function meets the requirement of setting or the number of iteration has reached the upper limit of setting, training is represented
Terminate, otherwise returns and continue iteration.
4th, final probability is obtained using BP:
4.1st, the calculation of final probability:
The behavior of user's objective attribute, the behavior of item attribute overall score, user preference are being based on from a upper trifle is available
In the case where behavior and fashion elements behavior these four behaviors, target user is to the original probability for not watching film project, herein
It is the original probability that above-mentioned four kinds of behaviors are calculated by CNN, it may be assumed that
PV=CNN (behavior of user's objective attribute) (13)
PF′=CNN (behavior of item attribute overall score) (14)
PP′=CNN (user preference behavior) (15)
PN=CNN (fashion elements behavior) (16)
Using be calculated four kinds of original probabilities as the input of BP model, suitable weighted value is trained by BP model
Then W and bias b can calculate user in different behaviors by updated weighted value, bias and original probability
Under the influence of to the final probability of each possible film project.
For BP model as shown in fig. 6, what is finally exported is final probability, the calculating of mode is as follows:
Above-mentioned calculating is the calculating of a part of neuron, be left neuron calculate can according to following formula into
Row is further to be calculated, and calculation formula is as follows:
Z(l)=W(l)A(l-1)+b(l) (23)
Wherein, function f is activation primitive, is to have selected Sigmoid function for activation primitive herein.
The error function formula of user in BP model are as follows:
Wherein, Pr iFor the real probability of customer consumption project i, value represents user for 0 or 1,0 and does not consume this
Mesh, 1 represents the customer consumption project；PiFor the final prediction probability after neural metwork training.
BP optimizes weight parameter and the offset from hidden layer to output layer using gradient descent method.Then pass through above-mentioned error
The weight and offset parameter from input layer to hidden layer can also be optimized.By the above-mentioned BP model of training, it can predict that target is used
Family is under the influence of different behaviors to the final probability of each possible consumption project.In the data set of test, it is recognized herein that tool
Preceding N of maximum final probability P are that user may consume most projects.In other words, that is, these first N is user
The project that next time most possibly consumes, recommends this N first to target user.
Example 1:
The validity of CNN-BP model proposed in this paper is verified by the comparison and analysis of experiment, comparative experiments is mainly wrapped
Include the comparison of machine learning algorithm Yu this paper algorithm, the comparison of traditional algorithm and this paper algorithm.This experimental selection is 100K
MovieLens data set is as collected by the GroupLens research team of Minnesota university, and wherein file u.data includes
943 users are to 100,000 scorings of 1682 films and timestamp.Each user at least 20 scorings, scoring
Value range is integer 1-5, and value is bigger, represents user and more likes the film.File u.user contains the ID of user, year
Age, gender and occupation.7 classes and 21 classes have been respectively divided in age and occupation.File u.item includes ID, title, the hair of film
Cloth date, the date of showing, film types.Film types is divided into 19 classes, and the ID and type of film are mainly used in experiment.For
When data prediction, need to select some features, and these features not only need the reasonability and validity in view of itself, and
And it is also required to effective expression in view of them.Herein, four parts, user's objective attribute number have mainly been divided the data into
According to, item attribute general comment divided data, the data of user preference data and fashion elements, can be obtained by data prediction.Phase
It is compared to traditional recommendation, the aspect considered herein is more extensive.
Experiment knot is mainly analyzed using accuracy rate (Accuracy) and error amount (Error) the two evaluation indexes herein
Fruit.
Accuracy rate (Accuracy) is the accurate degree for evaluating proposed algorithm in recommender system, is theoretically
The higher the better for the value of accuracy rate, also represents the effect of proposed algorithm more preferably and the precision of proposed algorithm is higher, so, this literary grace
Use accuracy rate as one of evaluation index.Certainly, accuracy rate will not be directly displayed to target user, be shown to of user
The list recommended, so even if recommendation list successfully can be shown to user, do not represent yet accuracy rate be bound to it is very high,
It could also be possible that barely satisfactory.The calculation formula of accuracy rate are as follows:
Error rate (Error) is primarily referred to as the degree of a deviation between predicted value and actual value, commonly used to calculate
There are two ways to error amount, one is mean square error (MSE), another kind is root-mean-square error (RMSE), their calculating is public
Formula is defined as:
Wherein: XiRepresent predicted value, YiRepresent actual value；N represents the sum of project.
Since the result of calculating is used only to classify, and there is no any mathematical meanings, so combining the reality of this paper
Border situation is finally used as evaluation index only with mean square error (MSE).Certainly, when the value of MSE is smaller, prediction is represented
The degree of deviation is smaller between value and actual value, that is, the precision recommended is higher.It regard 80% in data as training set simultaneously, remains
Under be used as test set.
The comparison and analysis of experimental result are influence of the different parameters to accuracy rate in CNN first.
In CNN model, parameter is relatively more, and different parameters can generate different influences to the result of output, so herein
It is to use combined method, using accuracy rate come the effect of measure algorithm.This experiment is primarily directed to batch size, volume
Influence of the number of plies and learning rate of lamination to accuracy rate, the results are shown in Table 1:
The parameter combination of 1 part CNN of table
As can be seen from the table, the influence of the number of plies and learning rate of convolutional layer to accuracy rate, batch size pair are compared to
The influence of accuracy rate is the largest.In fact, in the identical situation of the number of plies and learning rate of convolutional layer, when batch size is chosen
When 100, accuracy rate can all decline 10% or more.Such situation is generated to be primarily due to choose excessively when batch size
If small, the data volume of input will become larger very much, thus be easier to generate overfitting problem, to influence accuracy rate
Output, so batch size selection or it is critically important.This experiment is carried out by the combination of partial parameters, is passed through
Multiple training and verifying, is finally selected as 6 for the number of plies of convolutional layer, and modelling effect at this time is best.Partial parameters in CNN model
Setting it is as shown in table 2:
The parameter setting of 2 CNN model of table
The followed by comparison of different machines learning algorithm and this paper algorithm.Influence in addition to the parameter inside CNN to model,
It also needs to compare with other algorithm herein, has actual effect to verify new model proposed in this paper.This trifle was before this
Have chosen decision-tree model (the Decision Tree in the algorithm in the machine learning of part, such as hackberry bayesian algorithm
Model, DTM) and model-naive Bayesian (Naive Bayesian Model, NBM)；Then the two models and this paper are utilized
Model compare test；Finally select evaluation index of the accuracy rate (Accuracy) as current comparative experiments.
Comparing result is as shown in fig. 7, the accuracy rate of new model CNN-BP proposed in this paper is most as can be observed from Figure
High.The especially accuracy rate of the DTM half that is almost new model CNN-BP accuracy rate, and when the value of accuracy rate is bigger, also
Be recommend precision it is higher, the impact of performance of algorithm is better.The accuracy rate of new model CNN-BP is accurate also greater than CNN model
Rate, it can be seen that the effect of this paper Fusion Model is better than the effect of single model.So being compared to other machine learning
Benchmark algorithm, the new model constructed herein can more accurately go next behavior of prediction user, to promote expiring for user
Meaning degree.
It is finally the comparison of traditional algorithm Yu this paper algorithm, this part is by methods herein, based on the collaboration of user
Filter algorithm (UserCF) and the comparative experiments that is carried out based on traditional these three methods of CNN algorithm, and from index mean square error
Poor (MSE) is come what is be compared.Comparing result is as shown in Figure 8.
From the description to evaluation index mean square error (MSE), it can learn that the value of MSE is the smaller the better, when being worth smaller,
The degree of deviation between the predicted value and actual value calculated is represented with regard to smaller, that is, the precision recommended is higher.Such as institute on figure
Show, it can be seen that the value of the MSE of new model CNN-BP proposed in this paper is respectively less than first two algorithm, i.e. the MSE value of this paper algorithm
Minimum also means that for the precision of the precision relative datum algorithm of this paper algorithm be relatively high.
In conclusion can be seen that from the comparison and analysis of above-mentioned experiment relative to common machine algorithm, this paper algorithm
Accuracy rate have been significantly improved；And the CNN relative to common algorithm and single model, the new model merged herein
The mean square error of CNN-BP also decreases；Demonstrating the different types of behavior of user has front to the precision and result of recommendation
Influence, also demonstrate high efficiency of the Fusion Model than single model of this paper.
Claims (8)
1. a kind of recommended method based on CNN and BP neural network, it is characterised in that this method mainly includes following committed step:
1st, user's different type behavior:
1.1st, initial data set is handled, obtains user-user attribute matrix, user-project general comment sub-matrix and stream
Row element behavioural matrix；
1.2nd, user preference matrix is constructed, to indicate user to the preference of film native；
2nd, the hidden layer of CNN model:
2.1st, feature is extracted using convolutional layer；
2.2nd, using pond layer come compressive features；
2.3rd, full articulamentum；
2.4th, overfitting problem in CNN network；
3rd, the score in predicting based on CNN network:
3.1st, the algorithmic procedure of the score in predicting based on CNN network；
4th, final probability is obtained using BP:
4.1st, the calculation of final probability.
2. the recommended method according to claim 1 based on CNN and BP neural network, it is characterised in that:
User-user attribute matrix described in step the 1.1, user-project general comment sub-matrix and fashion elements behavioural matrix,
Its corresponding matrix is denoted as V, F ' and N respectively,
The form of V is as follows:
Wherein: k represents the number of user property, attribute, that is, user ID of user, gender, age bracket, occupation etc.；In matrix
ElementRepresent user uiIn the objective characteristics attribute g of its ownkOn value；
The form of F ' is as follows:
Wherein: element in matrixRepresent user uiTo watching and contained characteristic attribute gkAll films general comment
Point；
The form of N is as follows:
Wherein:It is made of number 0 and 1,0 represents film mjNot comprising fashion elements gk；1 represents film mjInclude
Item characteristic attribute gk。
3. the recommended method according to claim 1 based on CNN and BP neural network, it is characterised in that:
User preference matrix is denoted as P ' in step the 1.2, and the form of P ' is as follows:
Wherein: element in matrixRepresent the user u under time factoriTo item characteristic attribute gkPreference.
4. the recommended method according to claim 1 based on CNN and BP neural network, it is characterised in that:
In step the 2.1 the following steps are included:
It is local sensing first；
Followed by parameter sharing reaches largely reducing for number of parameters using parameter sharing, obtains effective feature extraction
As a result；
It is finally convolution, the extraction for feature.
5. the recommended method according to claim 1 based on CNN and BP neural network, it is characterised in that:
Step the 2.3 is the feature integrating Chi Huahou and obtaining, and is normalized, and then exports one to every kind of classification feelings
The probability of condition is finally that foundation is classified with this probability.
6. the recommended method according to claim 1 based on CNN and BP neural network, it is characterised in that:
Overfitting problem includes overfitting problem and time-consuming problem in CNN network in step the 2.4, is using to related in model
Partial parameters abandoned, to reduce the quantity of Model Parameter, reduce the probability for overfitting problem occur.
7. the recommended method according to claim 1 based on CNN and BP neural network, it is characterised in that:
The algorithmic procedure of score in predicting based on CNN network in step the 3.1 is as follows:
1) clustering cluster and behavioral data, the data of convection current row element after obtaining cluster are pre-processed；
2) data in clustering cluster are divided into two parts, a part is used as training set, and another part is used as test set；
3) CNN network is initialized, defines various network parameters mentioned above first, is then given using random method
Weighted value and bias assignment in CNN network；
4) the selected part data from training set, as input data, the input layer being input in CNN arrives hidden layer
Formula:
hi=Relu (Wxi+b) (10)
Wherein: Relu is activation primitive, and W weighted value, b is bias；
5) subsequently into pond layer, pond method is used to it, herein using maximum pond method；
6) after CNN network exports result, by the difference of output result in training set and practical CNN network output result come
Error function is constructed, local derviation then is asked to error function respectively, about the local derviation of weighted value W and bias b, passes through calculated result
Reversely to adjust the W and b in each layer in front；
7) weighted value W and bias b are updated by continuous iteration；
If 8) value of error function meets the requirement of setting or the number of iteration has reached the upper limit of setting, training knot is represented
Otherwise beam returns and continues iteration.
8. the recommended method according to claim 1 based on CNN and BP neural network, it is characterised in that:
Step the 4.1 is the input of four kinds of original probabilities will being calculated as BP model, and it is suitable to be trained by BP model
Weighted value W and bias b, user is then calculated by updated weighted value, bias and original probability and is not being gone together
To the final probability of each possible film project under the influence of；Using be calculated four kinds of original probabilities as BP model
Input, suitable weighted value W and bias b is trained by BP model, then can pass through updated weighted value, biasing
Value and original probability calculate user under the influence of different behaviors to the final probability of each possible film project；
Wherein four kinds of original probabilities are based on the behavior of user's objective attribute, the behavior of item attribute overall score, user preference behavior
In the case where these four behaviors of fashion elements behavior, target user is to the original probability for not watching film project.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910175261.XA CN109902201A (en) | 2019-03-08 | 2019-03-08 | A kind of recommended method based on CNN and BP neural network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910175261.XA CN109902201A (en) | 2019-03-08 | 2019-03-08 | A kind of recommended method based on CNN and BP neural network |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN109902201A true CN109902201A (en) | 2019-06-18 |
Family
ID=66946860
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910175261.XA Pending CN109902201A (en) | 2019-03-08 | 2019-03-08 | A kind of recommended method based on CNN and BP neural network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109902201A (en) |
Cited By (9)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110442618A (en) * | 2019-07-25 | 2019-11-12 | 昆明理工大学 | Merge convolutional neural networks evaluation expert's recommended method of expert info incidence relation |
| CN111125338A (en) * | 2019-11-08 | 2020-05-08 | 中山赛伯坦智能科技有限公司 | Book purchasing recommendation method and device based on convolutional neural network |
| CN111401297A (en) * | 2020-04-03 | 2020-07-10 | 天津理工大学 | Triphibian robot target recognition system and method based on edge calculation and neural network |
| CN111640483A (en) * | 2020-06-09 | 2020-09-08 | 南京邮电大学 | Fitness scheme recommendation method based on AKC model |
| CN111723287A (en) * | 2020-06-03 | 2020-09-29 | 开普云信息科技股份有限公司 | Content and service recommendation method and system based on large-scale machine learning |
| CN111967578A (en) * | 2020-08-04 | 2020-11-20 | 厦门大学 | Construction method of depth recommendation system framework based on uncompensated decision mechanism |
| CN112749345A (en) * | 2021-02-09 | 2021-05-04 | 上海海事大学 | K nearest neighbor matrix decomposition recommendation method based on neural network |
| CN113139088A (en) * | 2021-05-14 | 2021-07-20 | 西安建筑科技大学 | Movie recommendation method, medium, device and system of IDF (inverse discrete function) model collaborative filtering model |
| CN117236900A (en) * | 2023-10-23 | 2023-12-15 | 广东电网有限责任公司 | Individual tax data processing method and system based on flow automation |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108038629A (en) * | 2017-12-30 | 2018-05-15 | 北京工业大学 | A kind of optimization method based on collaborative filtering |
| CN109003132A (en) * | 2017-10-30 | 2018-12-14 | 上海寒武纪信息科技有限公司 | Advertisement recommended method and Related product |
| CN109241203A (en) * | 2018-09-27 | 2019-01-18 | 天津理工大学 | A kind of user preference and distance weighted clustering method of time of fusion factor |
-
2019
- 2019-03-08 CN CN201910175261.XA patent/CN109902201A/en active Pending
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109003132A (en) * | 2017-10-30 | 2018-12-14 | 上海寒武纪信息科技有限公司 | Advertisement recommended method and Related product |
| CN108038629A (en) * | 2017-12-30 | 2018-05-15 | 北京工业大学 | A kind of optimization method based on collaborative filtering |
| CN109241203A (en) * | 2018-09-27 | 2019-01-18 | 天津理工大学 | A kind of user preference and distance weighted clustering method of time of fusion factor |
Non-Patent Citations (2)
| Title |
|---|
| ZHAOWEI QU.ET: ""Converged Recommendation System Based on RNN and BP Neural Networks"", 《2018 IEEE INTERNATIONAL CONFERENCE ON BIG DATA AND SMART COMPUTING (BIGCOMP)》 * |
| 陈增艳: ""基于卷积神经网络的电影推荐模型研究"", 《 CNKI优秀硕士学位论文全文库》 * |
Cited By (14)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110442618A (en) * | 2019-07-25 | 2019-11-12 | 昆明理工大学 | Merge convolutional neural networks evaluation expert's recommended method of expert info incidence relation |
| CN110442618B (en) * | 2019-07-25 | 2023-04-18 | 昆明理工大学 | Convolutional neural network review expert recommendation method fusing expert information association relation |
| CN111125338A (en) * | 2019-11-08 | 2020-05-08 | 中山赛伯坦智能科技有限公司 | Book purchasing recommendation method and device based on convolutional neural network |
| CN111401297A (en) * | 2020-04-03 | 2020-07-10 | 天津理工大学 | Triphibian robot target recognition system and method based on edge calculation and neural network |
| CN111723287A (en) * | 2020-06-03 | 2020-09-29 | 开普云信息科技股份有限公司 | Content and service recommendation method and system based on large-scale machine learning |
| CN111640483B (en) * | 2020-06-09 | 2022-10-28 | 南京邮电大学 | Fitness scheme recommendation method based on AKC model |
| CN111640483A (en) * | 2020-06-09 | 2020-09-08 | 南京邮电大学 | Fitness scheme recommendation method based on AKC model |
| CN111967578B (en) * | 2020-08-04 | 2022-06-21 | 厦门大学 | Construction method of depth recommendation system framework based on uncompensated decision mechanism |
| CN111967578A (en) * | 2020-08-04 | 2020-11-20 | 厦门大学 | Construction method of depth recommendation system framework based on uncompensated decision mechanism |
| CN112749345A (en) * | 2021-02-09 | 2021-05-04 | 上海海事大学 | K nearest neighbor matrix decomposition recommendation method based on neural network |
| CN112749345B (en) * | 2021-02-09 | 2023-11-14 | 上海海事大学 | K neighbor matrix decomposition recommendation method based on neural network |
| CN113139088A (en) * | 2021-05-14 | 2021-07-20 | 西安建筑科技大学 | Movie recommendation method, medium, device and system of IDF (inverse discrete function) model collaborative filtering model |
| CN117236900A (en) * | 2023-10-23 | 2023-12-15 | 广东电网有限责任公司 | Individual tax data processing method and system based on flow automation |
| CN117236900B (en) * | 2023-10-23 | 2024-03-29 | 广东电网有限责任公司 | Individual tax data processing method and system based on flow automation |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109902201A (en) | A kind of recommended method based on CNN and BP neural network | |
| CN109299396B (en) | Convolutional neural network collaborative filtering recommendation method and system fusing attention model | |
| Zhang et al. | MOOCRC: A highly accurate resource recommendation model for use in MOOC environments | |
| CN108763362B (en) | Local model weighted fusion Top-N movie recommendation method based on random anchor point pair selection | |
| Zuo et al. | Tag-aware recommender systems based on deep neural networks | |
| Noviantoro et al. | Investigating airline passenger satisfaction: Data mining method | |
| CN111680721A (en) | Accurate and interpretable classification with hard attention | |
| CN111242729A (en) | Serialization recommendation method based on long-term and short-term interests | |
| CN112163165A (en) | Information recommendation method, device, equipment and computer readable storage medium | |
| Choudhary et al. | SARWAS: Deep ensemble learning techniques for sentiment based recommendation system | |
| CN115917535A (en) | Recommendation model training method, recommendation device and computer readable medium | |
| WO2021139415A1 (en) | Data processing method and apparatus, computer readable storage medium, and electronic device | |
| Alatrash et al. | Augmented language model with deep learning adaptation on sentiment analysis for E-learning recommendation | |
| Meena et al. | Sentiment analysis from images using vgg19 based transfer learning approach | |
| CN112699310A (en) | Cold start cross-domain hybrid recommendation method and system based on deep neural network | |
| Ibrahim et al. | An intelligent hybrid neural collaborative filtering approach for true recommendations | |
| Alfarhood et al. | DeepHCF: a deep learning based hybrid collaborative filtering approach for recommendation systems | |
| CN113779387A (en) | Industry recommendation method and system based on knowledge graph | |
| Arık et al. | Hybrid course recommendation system design for a real-time student automation application | |
| Sajeevan et al. | An enhanced approach for movie review analysis using deep learning techniques | |
| Pughazendi et al. | Graph sample and aggregate attention network optimized with barnacles mating algorithm based sentiment analysis for online product recommendation | |
| Li et al. | Meta hierarchical reinforced learning to rank for recommendation: a comprehensive study in moocs | |
| Patel et al. | CNNRec: Convolutional Neural Network based recommender systems-A survey | |
| Lu et al. | Artificial immune network with feature selection for bank term deposit recommendation | |
| Kumar et al. | A Recommendation System & Their Performance Metrics using several ML Algorithms |
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 |
Application publication date: 20190618 |
|
| WD01 | Invention patent application deemed withdrawn after publication |