CN104794367B - Medical treatment resource scoring based on hidden semantic model is with recommending method - Google Patents
Medical treatment resource scoring based on hidden semantic model is with recommending method Download PDFInfo
- Publication number
- CN104794367B CN104794367B CN201510240697.4A CN201510240697A CN104794367B CN 104794367 B CN104794367 B CN 104794367B CN 201510240697 A CN201510240697 A CN 201510240697A CN 104794367 B CN104794367 B CN 104794367B
- Authority
- CN
- China
- Prior art keywords
- matrix
- user
- resource items
- resource
- hidden
- 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.)
- Expired - Fee Related
Links
Landscapes
- Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
Abstract
The invention provides a kind of medical treatment resource scoring based on hidden semantic model with recommending method, step：The first step, obtain and recommend resource item data, and carry out data filtering and cleaning；Second step, use the parameters in the learning automaton training SVD++ models of continuous action；3rd step, for user, to all resource items in set, score in predicting value of the user to each resource items is calculated based on hidden semantic model respectively；4th step, TopN recommendation lists are obtained using the sort algorithm after improvement.The present invention will predict appraisal result and to the degree of correlation of most related hidden class as the foundation of generation recommendation list, it reference is also made to the calibration coefficients of resource items in itself simultaneously, the resource items that can ensure to recommend are more likely to cause the interest of patient, while it is also ensured that resource items have sufficiently high scoring to go to obtain liking for patient.
Description
Technical field
The present invention relates to the enigmatic language justice model score prediction of medical treatment resource in medical big data field and TopN to recommend application,
Specifically, refer to be based on enigmatic language justice in medical big data with what gradient descent algorithm was combined based on learning automaton
The medical treatment resource scoring of model is with recommending method.
Background technology
Scoring is widely used already with recommendation function in current internet product, commercial product recommending such as Taobao,
Facebook friend recommendation, news recommendation of Baidu etc., the either associated topic of search engine web site, picture are recommended, electricity
The hot-sale products of sub- business web site is recommended, personalized product is recommended, or the friend recommendation of various social network sites, popular application push away
Recommend etc., recommend it is really omnipresent, it is all-pervasive, and the proportion due to recommendation shared by caused value is always all
It is not in any more.For above-mentioned recommendation problem, proposed algorithm is core technology.Proposed algorithm is the historical behavior according to user
With the feature of resource items, certain recommended models is established, with certain mathematics or statistical method, thus it is speculated that it is most possible to go out user
Like or resource items interested.For many years, there are a considerable amount of proposed algorithms, such as：Pushing away based on collaborative filtering
Recommend algorithm, the proposed algorithm based on graph model, the proposed algorithm based on correlation rule, the proposed algorithm based on hidden semantic model,
Score in predicting algorithm based on neighborhood etc..Because various algorithms have the advantage and disadvantage of its own property decision, thus suitable for not
Same environment and scene.However, in general, proposed algorithm is still a problem for needing to be further spread out research.
Through retrieval, currently without the explanation or report found with similar techniques of the present invention, domestic and international class is also not yet collected into
As data.
The content of the invention
The present invention is directed to the deficiencies in the prior art, and problem to be solved is that hidden semantic model is commented in medical resource
Divide and the application effect in TopN recommendations, it is proposed that a kind of medical treatment resource scoring based on hidden semantic model is with recommending method.This
Invention is improved in the sequence link of classical enigmatic language justice model training algorithm first, is added on the basis of former enigmatic language justice algorithm
The hidden Coefficients of class correlation factor is entered.Hidden semantic model after improvement proposed by the present invention is recommended by score in predicting and TopN can be with
The recall rate under every kind of recommendation list length is improved, and the article that can to shoot straight moves forward in lists, is grown in list
Spend it is limited in the case of, improve recommendation effect.
To achieve the above object, the present invention is achieved by the following technical solutions.
A kind of medical treatment resource scoring based on hidden semantic model comprises the following steps with recommending method：
The first step, in medical big data, obtain and recommend resource item data, and carry out data filtering and cleaning, form money
Source item set；
Second step, using the parameters in the learning automaton training SVD++ models of continuous action；
3rd step, based on hidden semantic model, user-resource items rating matrix is had according to the SVD++ models Central Plains after training
R, score in predicting value of the user to all resource items in resource items set is calculated respectively, it is pre- to obtain user-resource item rating
Measured value matrix R '；
4th step, according to the user obtained in the 3rd step-resource items score in predicting value matrix R ', obtain TopN after sequence and push away
Recommend list；
Wherein,
The SVD++ models are specially：
Including following parameter：
Matrix P, for having weighed preference of the user for each hidden class, wherein, PuFor the parameter in matrix P, table
Show and represent preference coefficients of the user u to hidden class；
Matrix Q, resource items and the degree of correlation of each hidden class are represented, wherein,For the parameter in matrix Q, resource is represented
Coefficient correlation between item i and hidden class；
Matrix Y, the factor of influence for resource items in user's history record to hidden class, wherein, yJ, fFor the parameter in matrix Y,
Represent that j-th of resource items is to the factor of influence of f-th of hidden class in user's history record；
Vectorial bu, biased as user, weigh user u scoring benchmark；
Vectorial bi, biased as resource items, weigh resource items i scoring benchmark；
M, as integral biased, weigh the scoring benchmark of whole score-system；
N (u), represent user u historical viewings or the resource items list of purchase.
Preferably, the second step specifically comprises the following steps：
Step 1, it is matrix P, matrix Q, matrix Y and vectorial bu, vectorial biIn independent continuous of each parameter configuration one
Action learning automatic machine, respective (μ, the σ) parameter pair of each parameter is safeguarded, form training set；Wherein, μ is to become in each parameter
The average of secondary element, σ are the variance of variable element in each parameter；
Step 2, by iterating to calculate Filled function parameter, in each iteration, any row record in training set is read；
Step 3, using corresponding square in the renewal changing ideas training set any row record of continuous action learning automaton
Battle array Pu, matrix Qi, matrix Yi, vectorial buWith vectorial biParameter, model training is completed after iteration for several times.
Preferably, the renewal thought of the continuous action learning automaton is specially：Continuous action learning automaton
Candidate actions set is number axis the preceding paragraph successive value, and its probability density function is Gaussian Profile, and its mean μ and variances sigma are all
It can change with different feed back of environment in each iteration.
Preferably, the 3rd step is specially：
All resource items in enigmatic language justice model hypothesis resource items set are divided into F hidden classes, and all resource items can be distinguished
It is grouped into this F hidden classes；
Original user-resource items rating matrix R is resolved into matrix P and matrix Q product；
After carrying out singular value decomposition to matrix P and matrix Q, matrix P ' and matrix Q ' are obtained, makes matrix P ' and matrix Q ' phases
Multiply, obtain restructuring matrix user-resource items score in predicting value matrix R ', scoring of the user to all resource items is predicted with this.
Preferably, the 4th step is specially：
In the middle taking-ups of user-resource items score in predicting value matrix R ' and the hidden class f of user's degree of correlation highest, as arranging again
The foundation of sequence；
Using hidden class f as calibration coefficients, all resource items are resequenced according to hidden class f coefficient correlation, as last
Recommendation list improve, formed TopN recommendation lists.
Preferably, the 4th step includes as follows：
Step I ', willArranged according to descending, resource number corresponding to TopN scoring before taking-up, wherein,For score in predicting value；
Step II ', for active user u, traverse user u matrix P corresponds to row PuIn each element, find out maximum
Value Pu max, and obtain Pu maxCorresponding hidden class-mark kmax；
Step III ', for each resource items i in TopN, take out resource items i and correspond to kthmaxThe phase relation of individual hidden class
NumberTake out calibration coefficients b corresponding to resource items i simultaneouslyi；
Step IV ', calculate each resource items i and user u degree of correlation coefficientAnd by TopN
In resource according to CI, uDescending rearrange, as consequently recommended order.
Compared with prior art, the invention has the advantages that：
1st, the present invention declines training algorithm resource item rating number is less, scoring based on the original gradient of hidden semantic model
In the sparse data set of matrix, the deficiency that test error rises, prediction accuracy declines be present, it is proposed that CALA-TM models
Training algorithm, the learning automaton for introducing continuous action replace original gradient to decline training algorithm progress model training, and should
Use in score in predicting, in practice it has proved that CALA-TM models remain to converge to relatively low in the case where user's rating matrix is sparse
Test error, compensate for original gradient and decline the deficiency that training algorithm prediction accuracy declines.
2nd, the present invention declines training algorithm each in the case of advantage, by two kinds considering CALA and original gradient
Algorithm is combined, and furthermore present LAGD-TM model training algorithms, and optimal initial value point is found first by learning automaton,
Then decline training algorithm using gradient and carry out iteratively faster continuation Optimal Parameters, in practice it has proved that LAGD-TM algorithms can enter one
Step reduces test error, while reduces the model training time compared to CALA-TM algorithms.
3rd, the present invention realizes application of the hidden semantic model in TopN recommendations, the best LAGD-TM of combined training effect
Model training algorithm, and improvement is proposed to the sequence link of former proposed algorithm (gradient decline training algorithm), by adding
Hidden class correlation factors reorder to former recommendation list, generally improve the recall rate under every kind of recommendation list length,
And the resource items for shoot straight move forward in lists, the ratio that Top10 and Top20 recommends hit is added, is grown in list
Spend it is limited in the case of, improve recommendation effect.
4th, the present invention using predict appraisal result and to the degree of correlation of most related hidden class all as generate recommendation list according to
According to, while reference is also made to the calibration coefficients of resource items in itself, it is ensured that the resource items of recommendation are more likely to cause the emerging of patient
Interest, while it is also ensured that resource items have sufficiently high scoring to go to obtain liking for patient.
Embodiment
Embodiments of the invention are elaborated below：The present embodiment is carried out lower premised on technical solution of the present invention
Implement, give detailed embodiment and specific operating process.It should be pointed out that to one of ordinary skill in the art
For, without departing from the inventive concept of the premise, various modifications and improvements can be made, these belong to the guarantor of the present invention
Protect scope.
Present embodiments provide a kind of medical treatment resource scoring based on hidden semantic model and recommendation method, including following step
Suddenly：
The first step, in medical big data, obtain and recommend resource item data, and carry out data filtering；
Second step, the parameters in SVD++ models are trained using innovatory algorithm；
3rd step, for user, to all resource items in set, the user is calculated respectively using new algorithm to each money
The score in predicting value of source item；
4th step, TopN recommendation lists are obtained using the sort algorithm after improvement.
Specially：
All resource items of enigmatic language justice model hypothesis (article) can be divided into F hidden classes (potential classification), all resources
Item can be grouped into this F hidden classes respectively.Original user-resource items rating matrix R is resolved into matrix P and square by new algorithm
Battle array Q product, wherein matrix P have weighed preference of the user for each hidden class, and matrix Q represents resource items and each hidden class
Degree of correlation.Matrix P after singular value decomposition is multiplied with matrix Q, obtains restructuring matrix R ', user is predicted with this
Scoring to all resource items.Due to containing the resource items in all set in rating matrix R, and each user may be only right
The resource items of wherein only a few have scoring, so matrix R is a sparse matrix, matrix R is often one large-scale under actual environment
Sparse matrix, now problem translated into the singular value decomposition of Large sparse matrix.
In a practical situation, scoring datum line of each different user to same asset item is different, and each resource items obtain
The datum line of scoring is also different, and (such as resource items of the most people to high score even if some user does not like, will not also be given
Too low fraction), and different types of resource item rating datum line is also different, therefore parameter m is introduced as integral biased, measurement
The scoring benchmark of whole score-system；Introduce vectorial buBiased as user, weigh user u scoring benchmark；Introduce vectorial biMake
Biased for resource items, weigh resource items i scoring benchmark.
The present embodiment Improvement training algorithm based on SVD++ models, the gradient based on classics decline training algorithm
The characteristics of and deficiency, on this basis, it is proposed that the training algorithm based on learning automaton.Learning automaton has anti-noise acoustic energy
The advantages of power is strong, global convergence performance is high, wherein, compared to the learning automaton of discrete movement, the study of continuous action is automatic
Machine (CALA) is more suitable in continuously spatially majorized function.The medical treatment resource based on hidden semantic model that the present embodiment proposes is commented
Divide with recommending method, be a kind of enigmatic language justice model training innovatory algorithm based on continuous action learning automaton.Therefore, second step
Parameters (CALA-TM model trainings algorithm) in training SVD++ models specifically comprise the following steps：
Step 1, it is matrix P, matrix Q, matrix Y and vectorial bu, vectorial biIn each parameter configuration one it is independent
CALA, respective (μ, σ) parameter pair is safeguarded, form training set；Wherein, μ is the average of variable element in each parameter, and σ is every
The variance of variable element in individual parameter；
Step 2, parameter is continued to optimize by iterative calculation, in each iteration, reads a line record in training set；It is right
In some user u and resource items i, obtain user-resource items score in predicting value matrix R ', and with original user-resource item rating
Matrix R relatively obtains error eU, i；
Step 3, matrix P is recorded using any row in training set corresponding to CALA renewal changing ideasu, matrix Qi, square
Battle array Yi, vectorial buWith vectorial biIn parameter, by certain number iteration complete model training；Wherein, the renewal of the CALA
The candidate actions set that thought is specially CALA is number axis the preceding paragraph successive value, and its probability density function is Gaussian Profile,
Its mean μ and variances sigma can all change with different feed back of environment in each iteration.
Further, in the step 1, concretely comprise the following steps：
Step 1.1, each group of matrix P element of random initializtion：μp(u, k), σp(u, k), matrix Q elements：μq(i, k), σq
(i, k) and matrix Y element：μy(i, k), σy(i, k), wherein, matrix P and matrix Q μ values initial value are between (0, max_
Rating/sqrt (F)) between random number, max_rating is the peak of prediction scoring, and F is the number of hidden class, matrix Y
μ be initialized as 0；σ initial value is uniformly set to 1/sqrt (F)；I is resource item number, and k is hidden class number, and u is number of users；
Step 1.2, b is initializeduVector element (μbu(u), σbu) and b (u)iVector element (μbi(i), σbi(i)), wherein,
μ is uniformly initialized as max_rating/4, and σ is uniformly arranged to max_rating；
Further, in the step 2, concretely comprise the following steps：
Step 2.1, for some user u and resource items i, according to the probability distribution of each parameter, be randomly derived p (u, k),
Q (i, k), buAnd b (u)i(i) value, for all neighborhood article j ∈ N (u), y (j, k) value, wherein k ∈ are randomly derived
[1, F], y (j, k) are matrix Y parameter, and p (u, k) is matrix P parameters, and q (i, k) is matrix Q parameters, bu(u) biased for user, bi
(i) biased for resource items；
Step 2.2, prediction score value is calculated using the sample point x and mean μ obtained at random respectively
Wherein：For sample point x prediction score value,Score value, μ are predicted for mean μy(j, k) is average
μ calculates middle entry；
Step 2.3, using true score value and prediction score value computing environment feedback β：
Wherein, βxThe environmental feedback calculated for prediction score value, βμThe environmental feedback calculated for true score value；
Step 2.4, according to the mean μ and variances sigma of each parameter of CAlA Policy Updates：
Wherein φ (σ)=σl；forσ≤σl
φ (σ)=σ；For σ ＞ σl
Parameter C is a constant positive number, parameter σlMinimum lower limit as σ is used for preventing from converging to non-optimal point, λ ginsengs
Number is learning rate, and its size needs to determine by testing,For σ value functions.
Further, in the step 3, concretely comprise the following steps：
Step 3.1, iteration is continued cycling through, until βμValue converges to sufficiently small.For single CALA, with iterations
Increase, μ has a very big convergence in probability (point obtains optimal environmental feedback) at optimum point, and σ can converge to it is minimum
Lower limit σl.Above-mentioned model, in order to carry out score in predicting, it is necessary to obtain scoring in advance using formula (3-8) after the completion of model training
Measured value.For above-mentioned CALA-TM algorithms, because each parameter is trained using independent CALA, therefore its is iterative
Form be all identical, and in classical GD-TM algorithms (gradient decline training algorithm), because y (i, k) parameter is added according to i
Sum, it is also required to carry out in its iteration recurrence formula once to add and one-level complexity is increased in total algorithm, still
The problem is not present in CALA-TM algorithms.It is automatic compared to the optimal way that gradient declines training algorithm straight line descending manner, study
The iterative process of machine is screw type, due to the randomness of the uncertain and selection action of environmental feedback, error in an iterative process
Can't straight line decline, there may come a time when to go up, but in condition compared with harsh environment, this iterative manner can be avoided effectively
Off-target value and be absorbed in local optimum, this advantage of learning automaton just compensate for the deficiency of gradient descent algorithm.It is real
Proof is trampled in the case where scoring number is less, the constringency performance of CALA-TM training algorithms is better than classical GD-TM training and calculated
Method.
Preferably, the 3rd step associative learning automatic machine and the respective advantage of gradient descent algorithm, give LAGD-TM
Model training algorithm.Optimum value can be accurately converged in view of CALA-TM algorithms, and classical GD-TM algorithms can be fast
The low training error of prompt drop, so first find best base on schedule using the learning automaton of continuous action in LAGD-TM algorithms,
Gradient descent algorithm is recycled to carry out iteratively faster convergence.
LAGD-TM algorithms comprise the following steps that (implication of wherein each parameter is same as above)：
Step 1, P, Q, Y matrix and b are initialized according to CALA-TM algorithmic rulesu、biEach variable element is equal in vector
Value μ and variances sigma；
Step 2, setting iterations are N, are iterated according to CALA-TM algorithmic rules and seek ginseng；
Step 3, matrix P, matrix Q, matrix Y and vectorial b are setu, vectorial biIn the final of each variable element converge to
Initial value of the average as next step algorithm (step 4 and step 5)；
Step 4, reset learning rate α；
Step 5, setting iterations are M, are iterated according to classical GD-TM algorithmic rules and seek ginseng；Wherein N's and M takes
Value needs to determine optimum value by testing repeatedly.
For score in predicting application, after the completion of above-mentioned model training, can equally be entered using preference pattern (3-8) formula
Row score in predicting.
By above-mentioned arthmetic statement, in LAGD-TM algorithms, CALA-TM algorithms are employed first and find best base
On schedule, when M values are larger, for the learning automaton of each parameter, substantially close to convergence, can avoid connecing
Local optimum is absorbed in the training got off herein, and algorithm is further iterated using classical GD-TM algorithms, can to instruct
Practice error on this basis further to decline, so as to improve final test error.Further, since the speed of classical GD-TM algorithms
Degree is also faster than CALA-TM, therefore LAGD-TM is also better than CALA-TM algorithms on the training time.
Preferably, in above-mentioned 4th step based on TopN typical case's proposed algorithm SVD++ of hidden semantic model by four processes group
Into being model determination, model training, score in predicting and sequence respectively.The present embodiment recommends method to make with the TopN based on SVD++
For the starting point of research, the performance that hidden semantic model TopN recommends each link is analyzed.TopN based on hidden semantic model recommends to calculate
The committed step of method is as follows：
Step I, use the parameters in LAGD-TM Algorithm for Training SVD++ models；
Step II, for user u, to all items i ∈ I in set, the user is calculated respectively to every using formula (3-9)
The score in predicting value of individual article
Step III, TopN recommendation lists are obtained using the sort algorithm after improvement.
By above step description it can be found that the model training link and last sequence link that TopN recommends all can be right
Last recommendation results have an impact.For above-mentioned algorithm steps, it is contemplated that LAGD-TM model training algorithms have preferably property
Can, therefore hidden semantic model is trained using LAGD-TM algorithms in step 1, illustrate to change sequence in step 3 below
Enter.
Preferably, in above-mentioned 3rd step for the link that sorts, as it was previously stated, TopN proposed algorithms are by the pre- of models fitting
Test and appraisal divide RuiAs unique sequence index, its limitation has so been done, and in general, a user always has its preference
Type of items, also just have coefficient correlation pufThe hidden class f of highest one, and with the hidden Coefficients of class correlation QifRelatively high resource
Item i will be more likely to be paid close attention to by user, meanwhile, sufficiently high prediction scoring RuiAnd ensure that resource items obtain the weight that user likes
Want factor, it is therefore desirable to which coefficient correlation and prediction two aspect factors of scoring are combined into consideration.Analyzed based on more than, this implementation
Example gives a kind of improved sort algorithm RCSM, the algorithm taken out with the user hidden class f of u degree of correlation highests, as again
The foundation of sequence, all items are according to the coefficient correlation Q with the hidden classifRearrangement, as last recommendation list.
TopN after improvement recommends comprising the following steps that for sort algorithm RCSM：
Step I ', willArranged according to descending, Item Number corresponding to TopN scoring before taking-up；
Step II ', for active user u, travel through its matrix P and correspond to row PuIn each element, find out maximum
Pu max, and obtain hidden class-mark k corresponding to itmax；
Step III ', for each resource items i in TopN, take out its corresponding kthmaxThe coefficient correlation of individual hidden classTake out calibration coefficients b corresponding to resource items i simultaneouslyi；
Step IV ', calculate each resource items i and user u degree of correlation coefficientAnd by TopN
In article according to CI, uDescending rearrange, as consequently recommended order.
Arthmetic statement more than, algorithm after improvement is by prediction appraisal result and correlation to most related hidden class
Foundation of the degree all as generation recommendation list, while reference is also made to the calibration coefficients b of resource items in itselfi, it is ensured that recommend
Resource items be more likely to cause the interest of user, while it is also ensured that resource items have sufficiently high scoring to go to obtain user's
Like.
In the present embodiment：
Still have some limitations because gradient declines training algorithm, when sample matrix is sparse, more likely fall into
Enter local optimum, the innovatory algorithm that the present embodiment proposes has that noise resisting ability is strong, global convergence performance based on learning automaton
The advantages of high, wherein, compared to the learning automaton of discrete movement, the learning automaton of continuous action is more suitable for continuous empty
Between upper majorized function；
Innovatory algorithm improves to some extent in prediction accuracy, but due to algorithm self character determine its iteration speed compared with
Slowly, the model training time is longer.But for commending system, that is, offline recommendation is used in, improves the training of model as far as possible
Time is also valuable, so angle of the present embodiment from training for promotion speed, takes into account the requirement of the lifting degree of accuracy, it is right
Training algorithm is further improved, and it is applied into score in predicting；
Sort algorithm after improvement taken out with the hidden class of user's degree of correlation highest, as the foundation of rearrangement, own
Resource items are resequenced according to the coefficient correlation of the hidden class, are improved as last recommendation list.Algorithm after improvement will be pre-
Survey appraisal result and to the degree of correlation of most related hidden class all as generating the foundation of recommendation list, while reference is also made to resource items
The calibration coefficients of itself, it is ensured that the resource items of recommendation are more likely to cause the interest of user, while it is also ensured that resource
Item has sufficiently high scoring to go to obtain liking for user.
The specific embodiment of the present invention is described above.It is to be appreciated that the invention is not limited in above-mentioned
Particular implementation, those skilled in the art can make various deformations or amendments within the scope of the claims, this not shadow
Ring the substantive content of the present invention.
Claims (4)
1. a kind of medical treatment resource scoring based on hidden semantic model is with recommending method, it is characterised in that comprises the following steps：
The first step, in medical big data, obtain and recommend resource item data, and carry out data filtering and cleaning, form resource items
Set；
Second step, using the parameters in the learning automaton training SVD++ models of continuous action；
3rd step, based on hidden semantic model, user-resource items rating matrix R is had according to the SVD++ models Central Plains after training, point
Score in predicting value of the user to all resource items in resource items set is not calculated, obtains user-resource items score in predicting value
Matrix R'；
4th step, according to the user obtained in the 3rd step-resource items score in predicting value matrix R', TopN is obtained after sequence and recommends row
Table；
Wherein,
The SVD++ models preference (u, i) is specially：
Including following parameter：
Matrix P, for having weighed preference of the user for each hidden class, wherein, PuFor the parameter in matrix P, user is represented
Preference coefficients of the u to hidden class；
Matrix Q, resource items and the degree of correlation of each hidden class are represented, wherein,For the parameter in matrix Q, represent resource items i with
Coefficient correlation between hidden class；
Matrix Y, the factor of influence for resource items in user's history record to hidden class, wherein, yj,fFor the parameter in matrix Y, represent
Factor of influence of j-th of resource items to f-th of hidden class in user's history record；
Vectorial bu, biased as user, weigh user u scoring benchmark；
Vectorial bi, biased as resource items, weigh resource items i scoring benchmark；
M, as integral biased, weigh the scoring benchmark of whole score-system；
N (u), represent user u historical viewings or the resource items list of purchase；
4th step is specially：
Taking-up and the hidden class f of user's degree of correlation highest in user-resource items score in predicting value matrix R', as rearrangement
Foundation；
Using hidden class f as calibration coefficients, all resource items are resequenced according to hidden class f coefficient correlation, are pushed away as last
List improvement is recommended, forms TopN recommendation lists；
4th step comprises the following steps：
Step I ', willArranged according to descending, resource number corresponding to TopN scoring before taking-up, wherein,To comment
Divide predicted value；
Step II ', for active user u, traverse user u matrix P corresponds to row PuIn each element, find out maximum
Pumax, and obtain PumaxCorresponding hidden class-mark kmax；
Step III ', for each resource items i in TopN, take out resource items i and correspond to kthmaxThe coefficient correlation of individual hidden classTake out calibration coefficients b corresponding to resource items i simultaneouslyi；
Step IV ', calculate each resource items i and user u degree of correlation coefficientAnd by TopN
Resource is according to Ci,uDescending rearrange, as consequently recommended order.
2. the medical treatment resource scoring according to claim 1 based on hidden semantic model is with recommending method, it is characterised in that institute
Second step is stated to specifically comprise the following steps：
Step 1, it is matrix P, matrix Q, matrix Y and vectorial bu, vectorial biIn the one independent continuous action of each parameter configuration
Learning automaton, respective (μ, the σ) parameter pair of each parameter is safeguarded, form training set；Wherein, μ is variable member in each parameter
The average of element, σ are the variance of variable element in each parameter；
Step 2, by iterating to calculate Filled function parameter, in each iteration, any row record in training set is read；
Step 3, using corresponding matrix P in the renewal changing ideas training set any row record of continuous action learning automatonu、
Matrix Qi, matrix Yi, vectorial buWith vectorial biParameter, model training is completed after iteration for several times.
3. the medical treatment resource scoring according to claim 2 based on hidden semantic model is with recommending method, it is characterised in that institute
The renewal thought of continuous action learning automaton is stated specifically, the candidate actions set of continuous action learning automaton is on number axis
One section of successive value, and its probability density function is Gaussian Profile, its mean μ and variances sigma all can be with environment in each iteration
Different feedbacks and change.
4. the medical treatment resource scoring according to claim 1 based on hidden semantic model is with recommending method, it is characterised in that institute
Stating the 3rd step is specially：
All resource items in enigmatic language justice model hypothesis resource items set are divided into F hidden classes, and all resource items can be grouped into respectively
In this F hidden classes；
Original user-resource items rating matrix R is resolved into matrix P and matrix Q product；
After carrying out singular value decomposition to matrix P and matrix Q, matrix P' and matrix Q' are obtained, matrix P' is multiplied with matrix Q', obtains
To restructuring matrix user-resource items score in predicting value matrix R', scoring of the user to all resource items is predicted with this.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201510240697.4A CN104794367B (en) | 2015-05-12 | 2015-05-12 | Medical treatment resource scoring based on hidden semantic model is with recommending method |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201510240697.4A CN104794367B (en) | 2015-05-12 | 2015-05-12 | Medical treatment resource scoring based on hidden semantic model is with recommending method |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN104794367A CN104794367A (en) | 2015-07-22 |
| CN104794367B true CN104794367B (en) | 2018-01-12 |
Family
ID=53559158
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201510240697.4A Expired - Fee Related CN104794367B (en) | 2015-05-12 | 2015-05-12 | Medical treatment resource scoring based on hidden semantic model is with recommending method |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN104794367B (en) |
Families Citing this family (9)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107169830B (en) * | 2017-05-15 | 2020-11-03 | 南京大学 | Personalized recommendation method based on clustering PU matrix decomposition |
| CN108021708B (en) * | 2017-12-27 | 2020-07-17 | 广州启生信息技术有限公司 | Content recommendation method and device and computer readable storage medium |
| CN108960966B (en) * | 2018-06-07 | 2022-05-03 | 重庆邮电大学 | Matrix decomposition recommendation method based on user preference |
| CN110008404B (en) * | 2019-03-22 | 2022-08-23 | 成都理工大学 | Latent semantic model optimization method based on NAG momentum optimization |
| CN111950016B (en) * | 2019-05-14 | 2023-11-21 | 北京腾云天下科技有限公司 | Method and device for generating data open output model and computing equipment |
| CN110164523A (en) * | 2019-05-21 | 2019-08-23 | 山东大学 | A kind of intelligent health analysis method and system with intelligence function |
| CN110210933B (en) * | 2019-05-21 | 2022-02-11 | 清华大学深圳研究生院 | Latent semantic recommendation method based on generation of confrontation network |
| CN112699302B (en) * | 2021-01-05 | 2023-07-04 | 汕头大学 | Real-time recommendation method based on geodesic distance |
| CN113539457A (en) * | 2021-07-16 | 2021-10-22 | 挂号网（杭州）科技有限公司 | Medical resource recommendation method and device, electronic equipment and storage medium |
Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103473291A (en) * | 2013-09-02 | 2013-12-25 | 中国科学院软件研究所 | Personalized service recommendation system and method based on latent semantic probability models |
| CN104112366A (en) * | 2014-07-25 | 2014-10-22 | 中国科学院自动化研究所 | Method for traffic signal optimization based on latent semantic model |
-
2015
- 2015-05-12 CN CN201510240697.4A patent/CN104794367B/en not_active Expired - Fee Related
Patent Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103473291A (en) * | 2013-09-02 | 2013-12-25 | 中国科学院软件研究所 | Personalized service recommendation system and method based on latent semantic probability models |
| CN104112366A (en) * | 2014-07-25 | 2014-10-22 | 中国科学院自动化研究所 | Method for traffic signal optimization based on latent semantic model |
Non-Patent Citations (3)
| Title |
|---|
| A Learning Automata-Based Singular Value Decomposition and Its Application in Recommendation System;Yuchun Jing et al.;《Intelligent Computing Methodologies》;20140806;第8589卷;第26-32页 * |
| 基于SVD++与行为分析的社交推荐;刘剑波 等;《计算机应用》;20130630;第33卷(第S1期);第82-86页 * |
| 基于协同过滤和行为分析的微博推荐系统;刘剑波;《中国优秀硕士学位论文全文数据库 信息科技辑》;20140715(第07期);论文第18、22-39页 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN104794367A (en) | 2015-07-22 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN104794367B (en) | Medical treatment resource scoring based on hidden semantic model is with recommending method | |
| Yin et al. | A classification algorithm based on ensemble feature selections for imbalanced-class dataset | |
| CN107256241B (en) | Movie recommendation method for improving multi-target genetic algorithm based on grid and difference replacement | |
| CN107657034A (en) | A kind of event social networks proposed algorithm of social information enhancing | |
| Navgaran et al. | Evolutionary based matrix factorization method for collaborative filtering systems | |
| CN108038629A (en) | A kind of optimization method based on collaborative filtering | |
| Zou et al. | Reinforcement learning to diversify top-n recommendation | |
| CN114202061A (en) | Article recommendation method, electronic device and medium based on generation of confrontation network model and deep reinforcement learning | |
| Hu et al. | Hybrid recommendation algorithm based on latent factor model and PersonalRank | |
| CN109902823A (en) | A kind of model training method and equipment based on generation confrontation network | |
| Czajkowski et al. | Steering the interpretability of decision trees using lasso regression-an evolutionary perspective | |
| Chen et al. | Poverty/investment slow distribution effect analysis based on Hopfield neural network | |
| Akkoyun et al. | Estimations of fission barrier heights for, and nuclei by neural networks. | |
| CN110825978B (en) | Multitask collaborative filtering method based on neighbor user feature sharing | |
| CN106777123A (en) | A kind of group based on two-way tensor resolution model recommends method | |
| Parlar et al. | IWD based feature selection algorithm for sentiment analysis | |
| Prusa et al. | Utilizing ensemble, data sampling and feature selection techniques for improving classification performance on tweet sentiment data | |
| CN109885758A (en) | A kind of recommended method of the novel random walk based on bigraph (bipartite graph) | |
| Bing et al. | Prediction of user's purchase intention based on machine learning | |
| M John et al. | Predicting House Prices using Machine Learning and LightGBM. | |
| CN113850317A (en) | Multi-type neighbor aggregation graph convolution recommendation method and system | |
| Dehkordi et al. | Sentiment Classification of reviews with RNNMS and GRU Architecture Approach Based on online customers rating | |
| Yi et al. | A novel recommender algorithm using information entropy and secondary-clustering | |
| Ji et al. | An efficient imperialist competitive algorithm for solving the QFD decision problem | |
| Yuan et al. | On predicting event propagation on weibo |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| C06 | Publication | ||
| PB01 | Publication | ||
| EXSB | Decision made by sipo to initiate substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant | ||
| CF01 | Termination of patent right due to non-payment of annual fee |
Granted publication date: 20180112 Termination date: 20200512 |
|
| CF01 | Termination of patent right due to non-payment of annual fee |