CN110390561A - User-financial product of stochastic gradient descent is accelerated to select tendency ultra rapid predictions method and apparatus based on momentum - Google Patents
User-financial product of stochastic gradient descent is accelerated to select tendency ultra rapid predictions method and apparatus based on momentum Download PDFInfo
- Publication number
- CN110390561A CN110390561A CN201910598139.3A CN201910598139A CN110390561A CN 110390561 A CN110390561 A CN 110390561A CN 201910598139 A CN201910598139 A CN 201910598139A CN 110390561 A CN110390561 A CN 110390561A
- Authority
- CN
- China
- Prior art keywords
- user
- financial product
- gradient
- convergence
- momentum
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F17/00—Digital computing or data processing equipment or methods, specially adapted for specific functions
- G06F17/10—Complex mathematical operations
- G06F17/15—Correlation function computation including computation of convolution operations
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F17/00—Digital computing or data processing equipment or methods, specially adapted for specific functions
- G06F17/10—Complex mathematical operations
- G06F17/16—Matrix or vector computation, e.g. matrix-matrix or matrix-vector multiplication, matrix factorization
-
- 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
- G06Q30/0201—Market modelling; Market analysis; Collecting market data
- G06Q30/0202—Market predictions or forecasting for commercial activities
-
- 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
- G06Q40/00—Finance; Insurance; Tax strategies; Processing of corporate or income taxes
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Business, Economics & Management (AREA)
- Theoretical Computer Science (AREA)
- Mathematical Physics (AREA)
- Data Mining & Analysis (AREA)
- Computational Mathematics (AREA)
- Strategic Management (AREA)
- Accounting & Taxation (AREA)
- Pure & Applied Mathematics (AREA)
- Mathematical Optimization (AREA)
- Finance (AREA)
- Mathematical Analysis (AREA)
- Development Economics (AREA)
- Entrepreneurship & Innovation (AREA)
- Computing Systems (AREA)
- General Business, Economics & Management (AREA)
- Marketing (AREA)
- Economics (AREA)
- Algebra (AREA)
- Databases & Information Systems (AREA)
- Software Systems (AREA)
- General Engineering & Computer Science (AREA)
- Game Theory and Decision Science (AREA)
- Technology Law (AREA)
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
- Financial Or Insurance-Related Operations Such As Payment And Settlement (AREA)
Abstract
User-financial product of stochastic gradient descent is accelerated to select tendency ultra rapid predictions method and apparatus based on momentum the invention discloses a kind of, which includes: acquisition user-financial product score data and be configured to | M | row, | N | the rating matrix of column；It constructs and initializes hidden eigenmatrix, and construct solution objective function；Determine currently to update the most fast decision parameters gradient of convergence rate；Determine whether prediction model meets condition to the limitation parameter of gradient；Judge whether prediction model reaches convergence；After prediction model reaches convergence, obtained hidden eigenmatrix is trained in output；Obtained hidden eigenmatrix is used to predict the missing values in rating matrix.The present invention provides user-financial product that momentum accelerates stochastic gradient descent and selects tendency ultra rapid predictions as a result, providing personalized service for user, quickly provides safe and reliable, rigorous financial service for potential user in family finance e-commerce platform.
Description
Technical field
The present invention relates to computer data processing technology fields, particularly a kind of to be accelerated under stochastic gradient based on momentum
The user of drop-financial product selects tendency ultra rapid predictions method and apparatus.
Background technique
With the active development of computer technology to make rapid progress with social economy, financial electric business platform has goed deep into me
Life.More and more users select different financial products to invest by financial electric business platform, usual user
It can all be reported by investment after buying product in financial platform and personal experience evaluates it accordingly.Therefore, it is formed
It is leading electric business consumption mode with user.But the quantity with the increasingly mature and financial product of financial platform sharply increases
Add, the product for causing user oneself cannot be selected most to admire in numerous financial products.In order to provide a user accurately
Relationship between user and financial product can be indicated by personalized service with a user-financial product rating matrix,
The height of middle user's scoring indicates that user to the height of the degree of recognition of the product, is closed commonly used in measuring user-financial product
The matrix of system is a high-dimensional and extremely sparse matrix.
According to the history scoring of user's financial platform statistics, we will be seen that and analyze user to the preference of financial product
Rule establishes effective user-financial product preference prediction model on this basis.And scored by user financial product
Simulated environment simulates true environment, so that the marketing strategy for financial product provides important scientific basis.
It is had been carried out about user-financial product preference prediction technique, but these methods are in user-financial product building
The model training time, precision of prediction the problems such as on have obvious deficiency, such as: during model training, the training of model
Time spends very high and model precision of prediction lower, timely cannot provide safe and reliable high repayment finance for user and produce
Product.But due to the particularity of financial industry, it is most important that high repayment product is provided for user in time.Therefore, existing method exists
There is very big drawback in true financial relevant electric business platform application.
Summary of the invention
The present invention is directed to the above-mentioned problems in the prior art, provides and a kind of accelerates stochastic gradient descent based on momentum
User-financial product selects tendency ultra rapid predictions method and apparatus, can be scored and be realized according to history of the user to financial product:
(1) to user-financial product rating matrix high-precision forecast；(2) to user-financial product rating matrix ultra rapid predictions.
In order to solve the above technical problems, according to the first aspect of the invention, providing a kind of based on momentum acceleration stochastic gradient
The user of decline-financial product selects tendency ultra rapid predictions device, which includes:
Data preprocessing module: data are collected by financial platform server, processing data into model training can be straight
The data format used is connect, and the data handled well are put into memory module；
Data memory module stores and accelerates user-financial product of stochastic gradient descent to select tendency high speed based on momentum
The numbers such as the hidden eigenmatrix that temporary variable value, the corresponding value of initialization unit and the final training generated during prediction obtains
According to；
Data initialization module, from memory module obtain user-financial product data, initialization model training in needed for
Hidden eigenmatrix；
High-speed convergence direction selection module, for receiving the hidden eigenmatrix of initialization, and during model training
Determine high-speed convergence direction；
Prediction data generation module is used for ultra rapid predictions characteristic, and obtained feature prediction data is stored to number
According in unit；
Data outputting module, for exporting the finally obtained hidden feature of training and being commented with obtained hidden feature prediction is unknown
Point.
Further, the high-speed convergence direction selection module includes receiving corresponding initialization data unit and high-speed convergence side
To selecting unit, the corresponding initialization data unit of the reception is for receiving initialization data required when model training；It is described
High-speed convergence direction selection unit controls gradient magnitude with the gradient self adaptive control factor and damped expoential, passes through current gradient
Linear combination with accumulated gradient obtains current decision parameter and restrains most fast direction gradient, gradient control parameter β and gradient decaying
Exponent gamma remains the virtual value in (0,1) section.
Further, the data outputting module includes exporting hidden feature unit and the unknown scoring unit of prediction, the output
Hidden feature unit is for exporting the finally obtained hidden feature of training；Predict unknown scoring unit for predicting unknown scoring according to defeated
The hidden feature that hidden feature unit obtains out.
It is according to the present invention to accelerate user-financial product of stochastic gradient descent to select tendency high speed pre- based on momentum
The preferred embodiment of device is surveyed, the data preprocessing module includes:
Data receipt unit obtains user-financial product related data from financial platform, and user's set on platform is denoted as
M, commodity set are denoted as N, establish one | M | row, | N | the matrix of column is as user-financial product rating matrix R.
It is according to the present invention to accelerate user-financial product of stochastic gradient descent to select tendency high speed pre- based on momentum
The preferred embodiment of device is surveyed, the data memory module includes:
Storage unit, for storing received score data in the form of triple.Triple representation is U
=(m, n, r), wherein m, n respectively indicate user and financial product in user-financial product relational matrix, and r indicates scoring,
That is scoring of the user m to financial product n.The characteristic of user and financial product is stored as two hidden feature squares in the matrix form
Battle array B and J, and the two hidden eigenmatrix characteristic dimension f having the same.
It is according to the present invention to accelerate user-financial product of stochastic gradient descent to select tendency high speed pre- based on momentum
The preferred embodiment of device is surveyed, initialization unit user initialization accelerates user-financial product of stochastic gradient descent based on momentum
Relevant parameter involved in tendency ultra rapid predictions process is selected, is specifically included:
Initialize two hidden eigenmatrix B and J；Initialization feature dimension f；It initializes accumulation gradient and balances controlling elements β；
Initialize gradient damped expoential γ；Initialization convergence terminates threshold tau；The maximum training iteration wheel number L of initialization；Initialization was trained
Iteration wheel number controls variable l in journey；Initialize regularization factors λ2；Wherein characteristic dimension f determines each hidden eigenmatrix
Feature space dimension, is initialized as positive integer；Two hidden eigenmatrix B and J, i.e. B are | M | hidden eigenmatrix, the J of row f column be |
N | the hidden eigenmatrix of row f column initializes hidden eigenmatrix with random lesser positive number respectively；Maximum training iteration
Wheel number L is the variable for controlling the iterative process upper limit, is initialized as biggish positive integer；Iteration wheel number control variable l is initialized as
0；Whether it is to judge iterative process convergent parameter that convergence terminates threshold tau, is initialized with minimum positive number；Regularization factors λ2
That control is extracted in iterative process, the constant of the regularization effect of corresponding hidden eigenmatrix B and J coherent element, be initialized as compared with
Small positive number.It is according to the present invention to accelerate user-financial product of stochastic gradient descent to select tendency high speed based on momentum
The preferred embodiment of prediction meanss, for combining initialized relevant parameter and construction high-speed convergence direction selection；It specifically includes:
According to the known score data set Λ in user-financial product relational matrix R, target loss function is constructedIt is expressed as follows:
Wherein R(Λ)Indicate that user is to score data set known to financial product in user-financial product rating matrix R；rm,n
Entity relationship of the meaning of expression between user m and financial product n is scoring of the user m to financial product n；It indicates
Scoring of the user m to commodity n in known score data set Λ；bm′Indicate that m-th of user is corresponding in the hidden eigenmatrix B of user
Hidden feature；jn′Indicate the corresponding hidden feature of n-th of commodity in the hidden eigenmatrix J of financial product.
In order to enhance the Generalization Capability of model, usually in target loss functionMiddle L2Regularization term uses L2Just
Then change, optimization process is constrained, prevents occurring the problem of over-fitting in optimization process.Therefore, it is added into objective function
L2After regularization term, objective function can be obtained are as follows:
Wherein R(Λ)Indicate that user is to score data set known to financial product in user-financial product rating matrix R；rm,n
Entity relationship of the meaning of expression between user m and financial product n is the scoring of user m and financial product n；It indicates
Scoring of the user m to financial product n in known score data set Λ；bm′Indicate m-th of user couple in the hidden eigenmatrix B of user
The hidden feature answered；jn′Indicate the corresponding hidden feature of n-th of commodity in the hidden eigenmatrix J of financial product.λ2Indicate hidden eigenmatrix
Regularization parameter, measure L2Restriction effect of the regularization term to model.
Control B and J, which meets, closes accumulated error in collectionMinimum is optimized using the stochastic gradient descent of high-speed convergence
Algorithm is in above-mentioned accumulated errorOn matrix B and J are trained, obtain matrix B and the globally optimal solution of J.In training
In the process by the selection to user and the hidden feature high speed more new direction of financial product, according to the gradient equilibrium factor and exponential damping
Gradient and the current linear combination for updating gradient, calculate current update b before the factor is realizedm′And jn′Convergence rate is most fast
When corresponding gradient value, formula expression is as follows:
Wherein β is gradient equilibrium controlling elements, and γ is gradient damped expoential.Parameter κ(t)With parameter ν(t)Respectively single element
bm,kAnd jn,kFinal gradient magnitude when updating for the t times, they consist of two parts respectively: κ(t-1)And ν(t-1)Correspondence is determined
The final gradient updated before plan parameter and current gradient accumulated value.Parameter κ(t-1)With parameter ν(t-1)Respectively single element bm,kWith
jn,kThe t-1 times update when final gradient magnitude, as the t-2 time update when final gradient and the t-1 times corresponding gradient
It is cumulative.Wherein bm,kAnd jn,kRespectively bm′And jn′K-th of element of vector.WithRespectively
Current gradient value when for the t times update.Its expression formula are as follows:
It is according to the present invention to accelerate user-financial product of stochastic gradient descent to select tendency high speed pre- based on momentum
The preferred embodiment for surveying device is generated for ultra rapid predictions data；It specifically includes:
Current update b can be calculated according to high-speed convergence direction selection unitm′And jn′It is corresponding when convergence rate is most fast
Gradient value, therefore, it is available based on momentum accelerate stochastic gradient descent user-financial product select tendency ultra rapid predictions
The update rule of model:
About bm,kMore new formula it is as follows:
About jn,kMore new formula it is as follows:
WhereinIndicate scoring of the user m to financial product n in known score data set Λ；β is gradient equilibrium control
The factor；γ is gradient damped expoential；Parameter k(t)With parameter ν(t)Respectively single element bm,kAnd jn,kIt is final when updating for the t times
Gradient magnitude, they consist of two parts respectively: k(t-1)And ν(t-1)The final gradient and work as that corresponding decision parameters update before
Preceding gradient accumulated value.Parameter κ(t-1)With parameter ν(t-1)Respectively single element bm,kAnd jn,kFinal gradient when updating for the t-1 times
Size, final gradient and the t-1 time corresponding gradient when updating for as the t-2 times add up.Wherein bm,kAnd jn,kRespectively bm′
And jn′K-th of element of vector.WithRespectively the t times gradient value current when updating.
Above-mentioned training process is repeated on Λ, untilTo convergence on Λ, restraining decision condition is training iteration
What wheel number control variable r was calculated after reaching maximum training iteration wheel number L or epicycle iterationValue and upper one
WheelThe absolute value of the difference of value terminates threshold tau already less than convergence.
It is according to the present invention to accelerate user-financial product of stochastic gradient descent to select tendency high speed pre- based on momentum
Survey device, it is characterised in that: output unit is for exporting hidden feature unit and predicting unknown scoring unit for predicting unknown comment
Point；Specifically:
Output unit exports target loss functionReach two hidden eigenmatrixes when minimum, predicts unknown scoring
Unknown scoring in hidden eigenmatrix prediction user-financial product matrix that unit output unit obtains.
According to the second aspect of the invention, a kind of user-financial product being accelerated stochastic gradient descent based on momentum is provided
Select tendency ultra rapid predictions method, it is characterised in that: this method comprises the following steps:
S1: by financial platform collection of server user-financial product score data, the user in the financial platform collects
Conjunction is denoted as M, and financial product set is denoted as N, establishes one | M | row, | N | the matrix of column is as user-financial product rating matrix
R.It is sent to and accelerates user-financial product of stochastic gradient descent to select tendency ultra rapid predictions device based on momentum.
S2: utilizing user-financial product score data, initializes corresponding hidden eigenmatrix, passes through known score data
Objective function is constructed with corresponding predicted value.
S3: solving decision parameters according to the objective function of foundation, finds out the decision that will currently update in epicycle
The corresponding gradient of parameter；
S4: determine currently to update the most fast decision parameters gradient of convergence rate；
S5: determine that user-financial product of stochastic gradient descent is accelerated to select tendency ultra rapid predictions model pair based on momentum
Whether the limitation parameter of gradient meets condition: if satisfied, thening follow the steps S6, otherwise resetting illegal parameter and executing step again
Rapid S5；
S6: ultra rapid predictions model is inclined to according to accelerating user-financial product of stochastic gradient descent to select based on momentum
The condition of convergence comes whether judgment models reach convergence: if reaching convergence, thening follow the steps S7, if not reaching convergence, returns
Receipt row step S3；
S7: after prediction model training has reached convergence, output model trains the obtained hidden eigenmatrix of correspondence；
S8: the hidden eigenmatrix obtained by matrix disassembling method to missing values in user-financial product rating matrix into
Row prediction.
A kind of user-financial product selection tendency height accelerating stochastic gradient descent based on momentum according to the present invention
The preferred embodiment of fast prediction technique, step S2 include:
S21 initializes relevant parameter:
Initialize two hidden eigenmatrix B and J；Initialization feature dimension f；It initializes accumulation gradient and balances controlling elements β；
Initialize gradient damped expoential γ；The maximum training iteration wheel number L of initialization；It initializes iteration wheel number in training process and controls variable
L, convergence terminates threshold tau；Initialize regularization factors λ2；
It is as follows to initialize content:
Characteristic dimension f determines the feature space dimension of user and the hidden eigenmatrix of financial product, is initialized as positive integer.
The size of two hidden eigenmatrix B and J is respectively as follows: B and is | M | row f column hidden eigenmatrix, J be | N | what row f was arranged
Hidden eigenmatrix, eigenmatrix hidden for two are initialized with random lesser positive number respectively.
Initialization accumulation gradient balance controlling elements β is initialized as lesser positive number.
Initialization gradient damped expoential γ is initialized as lesser positive number.
Maximum training iteration wheel number L is the variable for controlling the iterative process upper limit, is initialized as biggish positive integer.
Iteration wheel number control variable l is initialized as 0.
Whether it is to judge iterative process convergent parameter that convergence terminates threshold tau, is initialized with minimum positive number.
Regularization factors λ2It is that control is extracted in iterative process, the regularization effect of corresponding hidden eigenmatrix B and J coherent element
The constant answered is initialized as lesser positive number.
S22 constructs target loss function according to the known score data set Λ in user-financial product relational matrix RIt is expressed as follows:
Wherein using Euclidean distance as optimization aim；Use L2Regularization constrains optimization process, prevents from optimizing
Occurs the problem of over-fitting in journey.
A kind of user-financial product selection tendency height accelerating stochastic gradient descent based on momentum according to the present invention
The preferred embodiment of fast prediction technique, step S3 include:
S31 finds out the corresponding gradient of decision parameters for currently needing to update according to objective function, and formula is expressed as follows:
A kind of user-financial product selection tendency height accelerating stochastic gradient descent based on momentum according to the present invention
The preferred embodiment of fast prediction technique, step S4 include:
S41 obtains the corresponding gradient of current update decision parameters according to step S3With
S42 passes through accumulated gradient and current gradient linearity according to balance factor and damped expoential to the restriction rule of gradient
Combination codetermines out the most fast decision parameters gradient of current update convergence rate, and formula is as follows:
Wherein β is gradient equilibrium controlling elements, and γ is gradient damped expoential.Parameter κ(t)With parameter v(t)Respectively single element
bm,kAnd jn,kFinal gradient magnitude when updating for the t times, they consist of two parts respectively: κ(t-1)And ν(t-1)Correspondence is determined
The final gradient updated before plan parameter and current gradient accumulated value.Parameter κ(t-1)With parameter ν(t-1)Respectively single element bm,kWith
jn,kThe t-1 times update when final gradient magnitude, as the t-2 time update when final gradient and the t-1 times corresponding gradient
It is cumulative.Wherein bm,kAnd jn,kRespectively bm′And jn′K-th of element of vector.WithRespectively
Current gradient value when for the t times update.
A kind of user-financial product selection tendency height accelerating stochastic gradient descent based on momentum according to the present invention
The preferred embodiment of fast prediction technique, step S6 include:
S61 accelerates user-financial product of stochastic gradient descent to select tendency ultra rapid predictions model more based on momentum
It is new:
Wherein about bm,kMore new formula it is as follows:
Wherein about jn,kMore new formula it is as follows:
WhereinIndicate scoring of the user m to financial product n in known score data set Λ；β is gradient equilibrium control
The factor；γ is gradient damped expoential；Parameter κ(t)With parameter ν(t)Respectively single element bm,kAnd jn,kIt is final when updating for the t times
Gradient magnitude, they consist of two parts respectively: κ(t-1)And ν(t-1)The final gradient and work as that corresponding decision parameters update before
Preceding gradient accumulated value.Parameter κ(t-1)With parameter ν(t-1)Respectively single element bm,kAnd jn,kFinal gradient when updating for the t-1 times
Size, final gradient and the t-1 time corresponding gradient when updating for as the t-2 times add up.Wherein bm,kAnd jn,kRespectively bm′
And jn′K-th of element of vector.
S62 is inclined to ultra rapid predictions model according to accelerating user-financial product of stochastic gradient descent to select based on momentum
Whether the condition of convergence comes whether judgment models reach convergence, as judge when parameter current updates already less than threshold tau.Usually set
Precision of prediction and the last precision of prediction for being set to this differ by more than 1 × 10-5When, meet parameter more new demand, otherwise
Reach the model condition of convergence.
A kind of user-financial product based on momentum acceleration stochastic gradient descent of the present invention selects tendency high speed pre-
The beneficial effect for surveying method and apparatus is: the present invention provides a kind of user-finance for accelerating stochastic gradient descent based on momentum
Product selects tendency ultra rapid predictions method and apparatus, analyzes the data generated between financial product and user, by dividing
Analysis result predicts user's relationship unknown with financial product.The strategy of prediction is to accelerate analysis thought pair using hidden feature
The inherent statistical law of known users-commodity score data is analyzed, to provide the use that momentum accelerates stochastic gradient descent
Family-financial product selects tendency ultra rapid predictions as a result, providing personalized service for user, is in family finance e-commerce platform
Potential user quickly provides safe and reliable, rigorous financial service, realizes better Wealthy distribution and family assets configuration.
Detailed description of the invention
Fig. 1 is the knot for accelerating user-financial product of stochastic gradient descent to select tendency ultra rapid predictions device based on momentum
Structure schematic diagram.
Fig. 2 is the stream for accelerating user-financial product of stochastic gradient descent to select tendency ultra rapid predictions method based on momentum
Journey schematic diagram.
Prediction model before Fig. 3 application present invention and after the present invention executes time comparison diagram.
Prediction model precision of prediction comparison diagram before Fig. 4 application present invention and after the present invention, wherein RMSE is prediction error
Measurement index, RMSE is smaller, and precision of prediction is higher.
Specific embodiment
It is clear to be more clear the objectives, technical solutions, and advantages of the present invention, referring to the drawings and by embodiment,
Invention is further described in detail.It is exemplary below by way of the embodiment being described with reference to the drawings, is only used for explaining this
Invention, and be not considered as limiting the invention.
Embodiment 1:
It shows to select the present invention is based on user-financial product of momentum acceleration stochastic gradient descent referring to Fig. 1, Fig. 1 and incline
To ultra rapid predictions device, which includes:
Data preprocessing module 510, for receiving global user-family finance product score data, at these data
The data format that can be used directly in model training is managed into, and the data handled well are put into data memory module 520.
Data memory module 520, the interim change for storing pretreated input data, generating in model predictive process
The hidden eigenmatrix that amount, the corresponding value of initialization unit and final training obtain.
Data initialization module 530, the hidden eigenmatrix for initialization model training.
High-speed convergence direction selection module 540, for receiving the hidden eigenmatrix of initialization, and in model training process
Middle determining high-speed convergence direction.
The high-speed convergence direction selection module 540 includes the unit 541 and high-speed convergence for receiving corresponding initialization data
Direction selection unit 542.
The unit 541 of corresponding initialization data is received for receiving initialization data required when model training.
High-speed convergence direction selection unit 542 controls gradient magnitude with the gradient self adaptive control factor and damped expoential.
Current decision parameter, which is obtained, by the linear combination of current gradient and accumulated gradient restrains most fast direction gradient, gradient control parameter
β and gradient damped expoential γ remains the virtual value in (0,1) section.
Prediction data generation module 550 is used for ultra rapid predictions characteristic, and obtained feature prediction data storage is arrived
In data cell.
The prediction data generation module 550 include ultra rapid predictions data generating unit 551, for by module 540 not
Linear combination is carried out with the corresponding gradient of update and current gradient before the cumulative decision parameters of degree and obtains final gradient value, and pressing should
Gradient value is updated operation to decision parameters.
Output module 560 is used for the finally obtained hidden eigenmatrix of output model training by matrix decomposition technology
Inner product of vectors acquires the predicted value of unknown-value in user-financial product rating matrix in corresponding hidden eigenmatrix.
In a particular embodiment, the data preprocessing module 510 includes:
Data receipt unit 511, for obtaining user-financial product related data from financial platform, the user on platform
Set is denoted as M, and commodity set is denoted as N, establishes one | M | row, | N | the matrix of column is as user-financial product rating matrix R.
In a particular embodiment, the data memory module 520 includes:
Storage unit 521, for storing received score data in the form of triple.Triple representation
For U=(u, i, r), wherein u, i respectively indicate user and financial product in user-financial product relational matrix, and r expression is commented
Point, i.e. scoring of the user u to financial product i.The characteristic of user and financial product is stored as two hidden spies in the matrix form
Levy matrix B and J, and the two hidden eigenmatrix characteristic dimension f having the same.
In a particular embodiment, initialization unit 531 is for initializing hidden eigenmatrix and accelerating boarding steps based on momentum
The user for spending decline-financial product selects relevant parameter involved in tendency ultra rapid predictions process；It specifically includes:
Initialize two hidden eigenmatrix B and J；Initialization feature dimension f；It initializes accumulation gradient and balances controlling elements β；
Initialize gradient damped expoential γ；Initialization convergence terminates threshold tau；The maximum training iteration wheel number L of initialization；Initialization was trained
Iteration wheel number controls variable l in journey；Initialize regularization factors λ2；Wherein characteristic dimension f determines each hidden eigenmatrix
Feature space dimension, is initialized as positive integer；Two hidden eigenmatrix B and J, i.e. B are | M | hidden eigenmatrix, the J of row f column be |
N | the hidden eigenmatrix of row f column initializes hidden eigenmatrix with random lesser positive number respectively；Maximum training iteration
Wheel number L is the variable for controlling the iterative process upper limit, is initialized as biggish positive integer；Iteration wheel number control variable l is initialized as
0；Whether it is to judge iterative process convergent parameter that convergence terminates threshold tau, is initialized with minimum positive number；Regularization factors λ2
That control is extracted in iterative process, the constant of the regularization effect of corresponding hidden eigenmatrix B and J coherent element, be initialized as compared with
Small positive number.
In a particular embodiment, high-speed convergence direction selection unit 542 is used to combine initialized relevant parameter and structure
Make high-speed convergence direction selection；It specifically includes:
According to the known score data set Λ in user-financial product relational matrix R, target loss function is constructedIt is expressed as follows:
Wherein R(Λ)Indicate that user is to score data set known to financial product in user-financial product rating matrix R；rm,n
Entity relationship of the meaning of expression between user m and financial product n, as scoring of the user m to financial product n；It indicates
Scoring of the user m to financial product n in known score data set Λ；bm′Indicate m-th of user couple in the hidden eigenmatrix B of user
The hidden feature answered；jn′Indicate the corresponding hidden feature of n-th of commodity in the hidden eigenmatrix J of financial product.
In order to enhance the Generalization Capability of model, usually in target loss functionMiddle addition L2Regularization term uses L2
Regularization constrains optimization process, prevents occurring the problem of over-fitting in optimization process.Therefore, by objective function
Middle L2After regularization term, objective function can be obtained are as follows:
Wherein R(Λ)Indicate that user is to score data set known to financial product in user-financial product rating matrix R；λ2
It indicates the regularization parameter of hidden eigenmatrix, measures L2Restriction effect of the regularization term to model.
Control B and J, which meets, closes accumulated error in collectionMinimum is optimized using the stochastic gradient descent of high-speed convergence
Algorithm is in above-mentioned accumulated errorOn matrix B and J are trained, obtain matrix B and the globally optimal solution of J.In training
In the process by the selection to user and the hidden feature high speed more new direction of financial product, according to the gradient equilibrium factor and exponential damping
Gradient and the current linear combination for updating gradient, calculate current update b before the factor is realizedm′And jn′Convergence rate is most fast
When corresponding gradient value, formula expression is as follows:
Wherein β is gradient equilibrium controlling elements, and γ is gradient damped expoential.Parameter κ(t)With parameter ν(t)Respectively single element
bm,kAnd jn,kFinal gradient magnitude when updating for the t times, they consist of two parts respectively: κ(t-1)And ν(t-1)Correspondence is determined
The final gradient updated before plan parameter and current gradient accumulated value.Parameter κ(t-1)With parameter v(t-1)Respectively single element bm,kWith
jn,kThe t-1 times update when final gradient magnitude, as the t-2 time update when final gradient and the t-1 times corresponding gradient
It is cumulative.Wherein bm,kAnd jn,kRespectively bm′And jn′K-th of element of vector.WithRespectively
Current gradient value when for the t times update.Its expression formula are as follows:
In a particular embodiment, ultra rapid predictions data generating unit 551 is generated for ultra rapid predictions data；It specifically includes:
Current update b can be calculated according to high-speed convergence direction selection unit 551m′And jn′When convergence rate is most fast
Corresponding gradient value, it is therefore, available to accelerate user-financial product of stochastic gradient descent to select tendency high speed based on momentum
The update rule of prediction model:
About bm,kMore new formula it is as follows:
About jn,kMore new formula it is as follows:
WhereinIndicate scoring of the user m to financial product n in known score data set Λ；β is gradient equilibrium control
The factor；γ is gradient damped expoential；Parameter κ(t)With parameter ν(t)Respectively single element bm,kAnd jn,kIt is final when updating for the t times
Gradient magnitude, they consist of two parts respectively: κ(t-1)And υ(t-1)The final gradient and work as that corresponding decision parameters update before
Preceding gradient accumulated value.Parameter κ(t-1)With parameter ν(t-1)Respectively single element bm,kAnd jn,kFinal gradient when updating for the t-1 times
Size, final gradient and the t-1 time corresponding gradient when updating for as the t-2 times add up.Wherein bm,kAnd jn,kRespectively bm′
And jn′K-th of element of vector.WithRespectively the t times gradient value current when updating.
Above-mentioned training process is repeated on Λ, untilTo convergence on Λ, restraining decision condition is training iteration
What wheel number control variable r was calculated after reaching maximum training iteration wheel number L or epicycle iterationValue and upper one
WheelThe absolute value of the difference of value terminates threshold tau already less than convergence.
It is according to the present invention to accelerate user-financial product of stochastic gradient descent to select tendency high speed pre- based on momentum
Survey device, it is characterised in that: output unit is for exporting hidden feature unit and predicting unknown scoring unit for predicting unknown comment
Point；Specifically:
Output unit exports target loss functionReach two hidden eigenmatrixes when minimum, predicts unknown scoring
Unknown scoring in hidden eigenmatrix prediction user-financial product matrix that unit output unit obtains.
The present invention is acting exclusively on the ultra rapid predictions that user-financial product selects tendency.
Embodiment 2
Referring to fig. 2, user-financial product of stochastic gradient descent is accelerated to select based on momentum Fig. 2 shows of the invention
It is inclined to ultra rapid predictions method, this method comprises the following steps:
S1: collection of server user-financial product score data, user's set on platform are denoted as M, and commodity set is denoted as
N establishes one | M | row, | N | the matrix of column is as user-financial product rating matrix R.It is sent to and is accelerated at random based on momentum
User-financial product of gradient decline selects tendency ultra rapid predictions device.
S2: utilizing user-financial product score data, initializes corresponding hidden eigenmatrix, passes through known score data
Objective function is constructed with corresponding predicted value.
S3: solving objective function according to the objective function of foundation, finds out the decision that will currently update in epicycle
The corresponding gradient of parameter；
S4: determine currently to update the most fast decision parameters gradient of convergence rate；
S5: determine that user-financial product of stochastic gradient descent is accelerated to select tendency ultra rapid predictions model pair based on momentum
Whether the limitation parameter of gradient meets condition: if satisfied, thening follow the steps S6, otherwise resetting illegal parameter and executing step again
Rapid S5.
S6: ultra rapid predictions model is inclined to according to accelerating user-financial product of stochastic gradient descent to select based on momentum
The condition of convergence comes whether judgment models reach convergence.S6 in Fig. 2: user-finance of stochastic gradient descent is accelerated to produce based on momentum
Product select tendency ultra rapid predictions model, determine whether current control parameter situation meets the update condition of parameter: if meeting parameter
Update condition, then return step S3；If it is not, indicating to reach the model condition of convergence, S7 is thened follow the steps.
S7: after prediction model training has reached convergence, output model trains the obtained hidden eigenmatrix of correspondence；
S8: the hidden eigenmatrix obtained by matrix disassembling method to missing values in user-financial product rating matrix into
Row prediction.
In a particular embodiment, the hidden eigenmatrix of step S2 initialising subscriber and financial product passes through known scoring number
According to objective function is constructed with corresponding predicted value, specifically include:
S21 initializes relevant parameter:
Initialize two hidden eigenmatrix B and J；Initialization feature dimension f；It initializes accumulation gradient and balances controlling elements β；
Initialize gradient damped expoential γ；The maximum training iteration wheel number L of initialization；It initializes iteration wheel number in training process and controls variable
L, convergence terminates threshold tau；Initialize regularization factors λ2；
It is as follows to initialize content:
Characteristic dimension f determines the feature space dimension of user and the hidden eigenmatrix of financial product, is initialized as positive integer；
Such as characteristic dimension f is initialized as 20.
The size of two hidden eigenmatrix B and J is respectively as follows: B and is | M | row f column hidden eigenmatrix, J be | N | what row f was arranged
Hidden eigenmatrix, eigenmatrix hidden for two are initialized with random lesser positive number respectively；Such as with set (0,
0.005) random normal number in initializes two hidden eigenmatrix B and J.
Initialization accumulation gradient balance controlling elements β is initialized as lesser positive number；Such as accumulation gradient is balanced and is controlled
Factor-beta is initialized as the positive number in set (0,1).
Initialization gradient damped expoential γ is initialized as lesser positive number；Such as gradient damped expoential γ is initialized as collecting
Close the positive number in (0,1).
Maximum training iteration wheel number L is the variable for controlling the iterative process upper limit, is initialized as biggish positive integer；Such as it will
Maximum training iteration wheel number L is initialized as 1000.
Iteration wheel number control variable l is initialized as 0.
Whether it is to judge iterative process convergent parameter that convergence terminates threshold tau, is initialized with minimum positive number；Such as it will
Convergence terminates threshold tau and is initialized as 1 × 10-5。
Regularization factors λ2It is that control is extracted in iterative process, the regularization effect of corresponding hidden eigenmatrix B and J coherent element
The constant answered is initialized as lesser positive number；Such as by regularization factors λ2It is initialized as in set [0.005,0.05] just
Number.
S22 constructs target loss function according to the known score data set Λ in user-financial product relational matrix RIt is expressed as follows:
Wherein using Euclidean distance as optimization aim；Use L2Regularization constrains optimization process, prevents from optimizing
Occurs the problem of over-fitting in journey.
In a particular embodiment, step S3 includes:
S31 finds out the corresponding gradient of decision parameters for currently needing to update according to objective function, and formula is expressed as follows:
In a particular embodiment, step S4 includes:
S41 obtains the corresponding gradient of current update decision parameters according to step S3With
S42 passes through accumulated gradient and current gradient linearity according to balance factor and damped expoential to the restriction rule of gradient
Combination codetermines out the most fast decision parameters gradient of current update convergence rate, and formula is as follows:
Wherein β is gradient equilibrium controlling elements, and γ is gradient damped expoential.Parameter k(t)With parameter ν(t)Respectively single element
bm,kAnd jn,kFinal gradient magnitude when updating for the t times, they consist of two parts respectively: k(t-1)And ν(t-1)Correspondence is determined
The final gradient updated before plan parameter and current gradient accumulated value.Parameter κ(t-1)With parameter ν(t-1)Respectively single element bm,kWith
jn,kThe t-1 times update when final gradient magnitude, as the t-2 time update when final gradient and the t-1 times corresponding gradient
It is cumulative.Wherein bm,kAnd jn,kRespectively bm′And jn′K-th of element of vector.WithRespectively
Current gradient value when for the t times update.
In a particular embodiment, step S6 user-financial product selection tendency ultra rapid predictions include:
S61 accelerates user-financial product of stochastic gradient descent to select tendency ultra rapid predictions model more based on momentum
It is new:
Wherein about bm,kMore new formula it is as follows:
Wherein about jn,kMore new formula it is as follows:
WhereinIndicate scoring of the user m to financial product n in known score data set Λ；β is gradient equilibrium control
The factor；γ is gradient damped expoential；Parameter κ(t)With parameter ν(t)Respectively single element bm,kAnd jn,kIt is final when updating for the t times
Gradient magnitude, they consist of two parts respectively: κ (t-1)And ν(t-1)The final gradient and work as that corresponding decision parameters update before
Preceding gradient accumulated value.Parameter κ(t-1)With parameter ν(t-1)Respectively single element bm,kAnd jn,kFinal gradient when updating for the t-1 times
Size, final gradient and the t-1 time corresponding gradient when updating for as the t-2 times add up.Wherein bm,kAnd jn,kRespectively bm′
And jn′K-th of element of vector.
S62 is inclined to ultra rapid predictions model according to accelerating user-financial product of stochastic gradient descent to select based on momentum
Whether the condition of convergence comes whether judgment models reach convergence, as judge when parameter current updates already less than threshold tau.Usually set
Precision of prediction and the last precision of prediction for being scheduled on this differ by more than 1 × 10-5When, meet parameter more new demand, otherwise
Reach the model condition of convergence.
Fig. 3 is using the user-financial product selection tendency ultra rapid predictions for accelerating stochastic gradient descent based on momentum and not have
Have and shows using the execution time comparison for accelerating user-financial product of stochastic gradient descent to select tendency ultra rapid predictions based on momentum
It is intended to.From the figure 3, it may be seen that, for user-financial product relation data, the runing time of model is remote after applying the technology of the present invention
The case where less than the technology is not applied.Specifically, available from Fig. 3, after having used the technology of the present invention, without using it is of the invention when
Between about use 6 times of the technology of the present invention.I.e. with after the technology of the present invention, the runing time of model than improving more than 6 times originally,
It allows users to quickly select the product oneself admired in numerous financial products.
Fig. 4 is using the user-financial product selection tendency ultra rapid predictions for accelerating stochastic gradient descent based on momentum and not have
In the case where having using the user-financial product selection tendency ultra rapid predictions for accelerating stochastic gradient descent based on momentum, RMSE pairs
Than schematic diagram, RMSE is the measurement index for predicting error, and RMSE value is bigger, and precision of prediction is lower；RMSE value is smaller, precision of prediction
It is higher.
As shown in figure 4, lower more than 3 times than unused the technology of the present invention using RMSE after the technology of the present invention, RMSE value is got over
Small, precision of prediction is higher.That is, substantially increasing model using the technology of the present invention selects the essence in tendency prediction in financial product
Degree.In practical applications, can preferably for user provide it is safe and reliable, have and actually guarantee and meet users ' individualized requirement
Financial product.
As seen from the above technical solution, the present invention provides a kind of user-gold for accelerating stochastic gradient descent based on momentum
Melt product and select tendency ultra rapid predictions method, to accelerate stochastic gradient descent prediction model to financial product and user based on momentum
Between the data that generate analyzed, user's relationship unknown with financial product is predicted by analyzing result.Prediction
Strategy is that analysis thought is accelerated to analyze the inherent statistical law of known users-commodity score data using hidden feature, from
And user-financial product that momentum accelerates stochastic gradient descent is provided and selects tendency ultra rapid predictions as a result, providing individual character for user
Change service, quickly provides safe and reliable, rigorous financial service for potential user in family finance e-commerce platform, realize
Better Wealthy distribution and family assets configuration.
Although an embodiment of the present invention has been shown and described, it will be understood by those skilled in the art that: not
A variety of change, modification, replacement and modification can be carried out to these embodiments in the case where being detached from the principle of the present invention and objective, this
The range of invention is defined by the claims and their equivalents.
Claims (10)
1. a kind of accelerate user-financial product of stochastic gradient descent to select tendency ultra rapid predictions method, feature based on momentum
It is, includes the following steps:
S1 acquires user-financial product score data, and user-financial product score data is configured to | M | row, | N | column
User-financial product rating matrix R, wherein M is user's set, and N is financial product set；
S2 constructs and initializes hidden eigenmatrix, constructs objective function by known score data and predicted value；
S3 solves objective function according to the objective function of foundation, finds out the decision parameters that will currently update in epicycle
Corresponding gradient；
S4 determines currently to update the most fast decision parameters gradient of convergence rate；
S5 determines that user-financial product of stochastic gradient descent is accelerated to select tendency ultra rapid predictions model to gradient based on momentum
Limitation parameter whether meet condition: if satisfied, S6 is thened follow the steps, if not satisfied, then resetting illegal parameter and holding again
Row step S5；
S6, according to the user-financial product selection tendency ultra rapid predictions model convergence for accelerating stochastic gradient descent based on momentum
Condition comes whether judgment models reach convergence: if reaching convergence, thening follow the steps S7, if not reaching convergence, return is held
Row step S3；
S7, after prediction model training has reached convergence, output model trains obtained hidden eigenmatrix；
Obtained hidden eigenmatrix is used to predict lacking in user-financial product rating matrix by matrix disassembling method by S8
Mistake value.
2. according to claim 1 accelerate user-financial product of stochastic gradient descent to select tendency high speed based on momentum
Prediction technique, which is characterized in that the step S2 includes:
S21 initializes relevant parameter:
Characteristic dimension f is initialized as positive integer；
Two hidden eigenmatrix B and J of initialization: matrix B is | M | the hidden eigenmatrix of user of row f column, matrix J be | and N | row f column
The hidden eigenmatrix of financial product, two hidden eigenmatrixes are initialized with random lesser positive number respectively；
Accumulation gradient balance controlling elements β is initialized as lesser positive number；
Gradient damped expoential γ is initialized as lesser positive number；
Maximum training iteration wheel number L is initialized as biggish positive integer；
Iteration wheel number control variable l is initialized as 0；
Convergence is terminated the minimum positive number of threshold tau to initialize；
By regularization factors λ2It is initialized as lesser positive number,
S22 constructs objective function according to the known score data set Λ in user-financial product relational matrix R
To objective functionUsing Euclidean distance as optimization aim, and use L2Regularization, obtained objective function are expressed as follows:
Wherein, R(Λ)Indicate that user is to score data set known to financial product in user-financial product rating matrix R；bm′It indicates
The corresponding hidden feature of m-th of user in the hidden eigenmatrix B of user；jn′Indicate n-th of commodity pair in the hidden eigenmatrix J of financial product
The hidden feature answered；rm,nIndicate scoring of the user m to financial product n；Indicate that user m is to quotient in known score data set Λ
The scoring of product n；λ2Indicate the regularization factors of hidden eigenmatrix.
3. according to claim 2 accelerate user-financial product of stochastic gradient descent to select tendency high speed based on momentum
Prediction technique, which is characterized in that the step S3 includes:
According to objective function, the corresponding gradient of decision parameters for currently needing to update is found out, formula is expressed as follows:
Wherein, bm,kAnd jn,kRespectively bm′And jn′K-th of element of vector；WithRespectively
B when the t times updatem,kAnd jn,kCurrent gradient value.
4. according to claim 3 accelerate user-financial product of stochastic gradient descent to select tendency high speed based on momentum
Prediction technique, which is characterized in that the step S4 includes:
S41 obtains the corresponding gradient of current update decision parameters according to step S3With
S42, according to accumulation gradient balance factor β and gradient damped expoential γ to the restriction rule of gradient, by accumulated gradient and
Current gradient linearity combination codetermines out the most fast decision parameters gradient of current update convergence rate, and formula is as follows:
Wherein, κ(t)And ν(t)Respectively single element bm,kAnd jn,kFinal gradient magnitude when updating for the t times；k(t-1)And ν(t-1)Point
It Wei not single element bm,kAnd jn,kFinal gradient magnitude when updating for the t-1 times.
5. according to claim 4 accelerate user-financial product of stochastic gradient descent to select tendency high speed based on momentum
Prediction technique, which is characterized in that the step S6 includes:
S61 accelerates user-financial product of stochastic gradient descent to select the update b of tendency ultra rapid predictions model based on momentumm,k
And jn,k:
Wherein about bm,kMore new formula it is as follows:
Wherein about jn,kMore new formula it is as follows:
Wherein,Indicate scoring of the user m to financial product n in known score data set Λ；β is accumulation gradient balance control
The factor processed；γ is gradient damped expoential；bm,kAnd jn,kRespectively bm′And jn′K-th of element of vector；WithRespectively the t times gradient value current when updating；k(t)And ν(t)Respectively single element bm,kAnd jn,kAt the t times
Final gradient magnitude when update；κ(t-1)And ν(t-1)Respectively single element bm,kAnd jn,kGradient when updating for the t-1 times is big
It is small,
S62, according to the user-financial product selection tendency ultra rapid predictions model convergence for accelerating stochastic gradient descent based on momentum
Whether condition comes whether judgment models reach convergence: judging when parameter current updates already less than threshold tau, if it is less, reaching
To the model condition of convergence.
6. a kind of accelerate user-financial product of stochastic gradient descent to select tendency ultra rapid predictions device, feature based on momentum
It is, comprising:
Data preprocessing module is configured to acquisition user-financial product data, and user-financial product data is processed into mould
The data format that can be used directly in type training；
Data memory module is configured for storing the data in the prediction meanss；
Data initialization module: it is configured to obtain user-financial product data from the data memory module, and initializes mould
Hidden eigenmatrix needed for type training；
High-speed convergence direction selection module is configured for receiving the hidden eigenmatrix of initialization, and in model training process
Middle determining high-speed convergence direction；
Prediction data generation module is configured for ultra rapid predictions characteristic, and obtained feature prediction data storage is arrived
In the data memory module；
Data outputting module is configured as output to train finally obtained hidden feature and predicts unknown scoring with obtained hidden feature.
7. according to claim 6 accelerate user-financial product of stochastic gradient descent to select tendency high speed based on momentum
Prediction meanss, which is characterized in that
The data preprocessing module is configured to establish one | M | row, | N | the matrix of column is as user-financial product scoring square
Battle array R, wherein M is user's set, and N is financial product set.
8. according to claim 7 accelerate user-financial product of stochastic gradient descent to select tendency high speed based on momentum
Prediction meanss, which is characterized in that
The data initialization module includes initialization unit, the initialization unit be configured to initialization based on momentum accelerate with
User-financial product of machine gradient decline selects relevant parameter involved in tendency ultra rapid predictions process,
The initialization of related relevant parameter includes:
Characteristic dimension f is initialized as positive integer；
Two hidden eigenmatrix B and J of initialization: matrix B is | M | the hidden eigenmatrix of user of row f column, matrix J be | and N | row f column
The hidden eigenmatrix of financial product, two hidden eigenmatrixes are initialized with random lesser positive number respectively；
Accumulation gradient balance controlling elements β is initialized as lesser positive number；
Gradient damped expoential γ is initialized as lesser positive number；
Maximum training iteration wheel number L is initialized as biggish positive integer；
Iteration wheel number control variable l is initialized as 0；
Convergence is terminated the minimum positive number of threshold tau to initialize；
By regularization factors λ2It is initialized as lesser positive number.
9. according to claim 8 accelerate user-financial product of stochastic gradient descent to select tendency high speed based on momentum
Prediction meanss, which is characterized in that
The high-speed convergence direction selection module includes receiving corresponding initialization data unit and high-speed convergence direction selection unit,
The corresponding initialization data unit of the reception is configured to receive initialized relevant parameter；
The high-speed convergence direction selection unit is configured to execute following steps:
According to the known score data set Λ in user-financial product relational matrix R, objective function is constructedAnd it is right
Objective functionUsing Euclidean distance as optimization aim and use L2Regularization, obtained objective function are expressed as follows:
Wherein, R(Λ)Indicate that user is to score data set known to financial product in user-financial product rating matrix R；bm′It indicates
The corresponding hidden feature of m-th of user in the hidden eigenmatrix B of user；jn′Indicate n-th of commodity pair in the hidden eigenmatrix J of financial product
The hidden feature answered；rm,nIndicate scoring of the user m to financial product n；Indicate that user m is to gold in known score data set Λ
Melt the scoring of product n；λ2Indicate the regularization factors of hidden eigenmatrix；
Using the stochastic gradient descent optimization algorithm of high-speed convergence in above-mentioned accumulated errorOn matrix B and J are instructed
Practice, obtains matrix B and the globally optimal solution of J；
Gradient and current linear group for updating gradient before being realized according to accumulation gradient balance factor β and gradient damped expoential γ
It closes, calculates current update bm′And jn′Corresponding gradient value, formula expression are as follows when convergence rate is most fast:
Wherein, κ(t)And v(t)Respectively single element bm,kAnd jn,kFinal gradient magnitude when updating for the t times；κ(t-1)And v(t-1)
Respectively single element bm,kAnd jn,kGradient magnitude when updating for the t-1 times；WithRespectively
B when the t times updatem,kAnd jn,kCurrent gradient value, its calculation formula is:
10. according to claim 9 accelerate user-financial product of stochastic gradient descent to select tendency high speed based on momentum
Prediction meanss, which is characterized in that
The prediction data generation module is configured to execute following steps:
Pass through calculated current update bm′And jn′Corresponding gradient value when convergence rate is most fast obtains accelerating based on momentum random
User-financial product of gradient decline selects the update rule of tendency ultra rapid predictions model:
About bm,kMore new formula it is as follows:
About jn,kMore new formula it is as follows:
Wherein,Indicate scoring of the user m to financial product n in known score data set Λ；β is accumulation gradient balance control
The factor processed；γ is gradient damped expoential；bm,kAnd jn,kRespectively bm′And jn′K-th of element of vector；WithRespectively the t times gradient value current when updating；κ(t)And v(t)Respectively single element bm,kAnd jn,kAt the t times
Final gradient magnitude when update；κ(t-1)And v(t-1)Respectively single element bm,kAnd jn,kGradient when updating for the t-1 times is big
It is small,
Above-mentioned training process is repeated on Λ, untilTo convergence on Λ, restraining decision condition is training iteration wheel number
What control variable r was calculated after reaching maximum training iteration wheel number L or epicycle iterationValue with it is last round ofThe absolute value of the difference of value is less than convergence and terminates threshold tau.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910598139.3A CN110390561B (en) | 2019-07-04 | 2019-07-04 | User-financial product selection tendency high-speed prediction method and device based on momentum acceleration random gradient decline |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910598139.3A CN110390561B (en) | 2019-07-04 | 2019-07-04 | User-financial product selection tendency high-speed prediction method and device based on momentum acceleration random gradient decline |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN110390561A true CN110390561A (en) | 2019-10-29 |
| CN110390561B CN110390561B (en) | 2022-04-29 |
Family
ID=68286137
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910598139.3A Active CN110390561B (en) | 2019-07-04 | 2019-07-04 | User-financial product selection tendency high-speed prediction method and device based on momentum acceleration random gradient decline |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN110390561B (en) |
Cited By (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111311324A (en) * | 2020-02-18 | 2020-06-19 | 电子科技大学 | User-commodity preference prediction system and method based on stable neural collaborative filtering |
| CN111563203A (en) * | 2020-05-08 | 2020-08-21 | 深圳市万佳安人工智能数据技术有限公司 | Intelligent household user-service interest degree personalized prediction device and method based on rapid non-negative implicit characteristic analysis |
| CN111564200A (en) * | 2020-05-08 | 2020-08-21 | 深圳市万佳安人工智能数据技术有限公司 | Old people diet feature extraction device and method based on rapid random gradient descent |
| CN112037850A (en) * | 2020-09-04 | 2020-12-04 | 中国科学院重庆绿色智能技术研究院 | Momentum acceleration-based device and method for predicting interaction between missing proteins |
| CN112036963A (en) * | 2020-09-24 | 2020-12-04 | 深圳市万佳安物联科技股份有限公司 | Webpage advertisement putting device and method based on multilayer random hidden feature model |
| CN112258263A (en) * | 2020-09-28 | 2021-01-22 | 中国科学院重庆绿色智能技术研究院 | Generalized momentum-based intelligent product recommendation device and method |
| CN112488183A (en) * | 2020-11-27 | 2021-03-12 | 平安科技（深圳）有限公司 | Model optimization method and device, computer equipment and storage medium |
Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20110238550A1 (en) * | 2010-03-24 | 2011-09-29 | Joshua Reich | Systems and methods for predicting financial behaviors |
| CN102930341A (en) * | 2012-10-15 | 2013-02-13 | 罗辛 | Optimal training method of collaborative filtering recommendation model |
| CN103235893A (en) * | 2013-05-06 | 2013-08-07 | 重庆大学 | User-commodity click rate self-adaptive prediction device and method |
| CN106776479A (en) * | 2016-12-16 | 2017-05-31 | 北京理工大学 | A kind of score in predicting method towards many attribute ratings systems |
| CN109615452A (en) * | 2018-10-29 | 2019-04-12 | 华中科技大学 | A kind of Products Show method based on matrix decomposition |
| CN109872006A (en) * | 2019-03-11 | 2019-06-11 | 西南石油大学 | A kind of scoring distribution forecasting method and device |
-
2019
- 2019-07-04 CN CN201910598139.3A patent/CN110390561B/en active Active
Patent Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20110238550A1 (en) * | 2010-03-24 | 2011-09-29 | Joshua Reich | Systems and methods for predicting financial behaviors |
| CN102930341A (en) * | 2012-10-15 | 2013-02-13 | 罗辛 | Optimal training method of collaborative filtering recommendation model |
| CN103235893A (en) * | 2013-05-06 | 2013-08-07 | 重庆大学 | User-commodity click rate self-adaptive prediction device and method |
| CN106776479A (en) * | 2016-12-16 | 2017-05-31 | 北京理工大学 | A kind of score in predicting method towards many attribute ratings systems |
| CN109615452A (en) * | 2018-10-29 | 2019-04-12 | 华中科技大学 | A kind of Products Show method based on matrix decomposition |
| CN109872006A (en) * | 2019-03-11 | 2019-06-11 | 西南石油大学 | A kind of scoring distribution forecasting method and device |
Non-Patent Citations (5)
| Title |
|---|
| SAROJ KUMAR LENKA 等: "Gradient descent with momentum based neural network pattern classification for the prediction of soil moisture content in precision agriculture", 《2015 IEEE INTERNATIONAL SYMPOSIUM ON NANOELECTRONIC AND INFORMATION SYSTEMS》 * |
| 孟利民 等: "评分预测问题中个性化推荐模型的研究", 《浙江工业大学学报》 * |
| 王德贤 等: "结合L1和L2正则化约束的隐语义预测模型研究", 《计算机工程与应用》 * |
| 黄晓凤: "并行协同过滤推荐模型的研究", 《中国优秀硕士学位论文全文数据库 信息科技辑》 * |
| 黎新志等: "基于用户分类的隐含因子模型研究", 《计算机应用研究》 * |
Cited By (12)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111311324A (en) * | 2020-02-18 | 2020-06-19 | 电子科技大学 | User-commodity preference prediction system and method based on stable neural collaborative filtering |
| CN111311324B (en) * | 2020-02-18 | 2022-05-20 | 电子科技大学 | User-commodity preference prediction system and method based on stable neural collaborative filtering |
| CN111563203A (en) * | 2020-05-08 | 2020-08-21 | 深圳市万佳安人工智能数据技术有限公司 | Intelligent household user-service interest degree personalized prediction device and method based on rapid non-negative implicit characteristic analysis |
| CN111564200A (en) * | 2020-05-08 | 2020-08-21 | 深圳市万佳安人工智能数据技术有限公司 | Old people diet feature extraction device and method based on rapid random gradient descent |
| CN112037850A (en) * | 2020-09-04 | 2020-12-04 | 中国科学院重庆绿色智能技术研究院 | Momentum acceleration-based device and method for predicting interaction between missing proteins |
| CN112037850B (en) * | 2020-09-04 | 2021-08-27 | 中国科学院重庆绿色智能技术研究院 | Momentum acceleration-based device and method for predicting interaction between missing proteins |
| CN112036963A (en) * | 2020-09-24 | 2020-12-04 | 深圳市万佳安物联科技股份有限公司 | Webpage advertisement putting device and method based on multilayer random hidden feature model |
| CN112036963B (en) * | 2020-09-24 | 2023-12-08 | 深圳市万佳安物联科技股份有限公司 | Webpage advertisement putting device and method based on multilayer random hidden feature model |
| CN112258263A (en) * | 2020-09-28 | 2021-01-22 | 中国科学院重庆绿色智能技术研究院 | Generalized momentum-based intelligent product recommendation device and method |
| CN112488183A (en) * | 2020-11-27 | 2021-03-12 | 平安科技（深圳）有限公司 | Model optimization method and device, computer equipment and storage medium |
| WO2022110640A1 (en) * | 2020-11-27 | 2022-06-02 | 平安科技（深圳）有限公司 | Model optimization method and apparatus, computer device and storage medium |
| CN112488183B (en) * | 2020-11-27 | 2024-05-10 | 平安科技（深圳）有限公司 | Model optimization method, device, computer equipment and storage medium |
Also Published As
| Publication number | Publication date |
|---|---|
| CN110390561B (en) | 2022-04-29 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110390561A (en) | User-financial product of stochastic gradient descent is accelerated to select tendency ultra rapid predictions method and apparatus based on momentum | |
| CN109902706B (en) | Recommendation method and device | |
| CN109446430A (en) | Method, apparatus, computer equipment and the readable storage medium storing program for executing of Products Show | |
| CN108182634A (en) | A kind of training method for borrowing or lending money prediction model, debt-credit Forecasting Methodology and device | |
| CN109118013A (en) | A kind of management data prediction technique, readable storage medium storing program for executing and forecasting system neural network based | |
| CN109472700A (en) | Prediction technique, server and the storage medium of stock price | |
| CN108052979A (en) | The method, apparatus and equipment merged to model predication value | |
| CN111311338A (en) | User value prediction method and user value prediction model training method | |
| CN107122327A (en) | The method and training system of a kind of utilization training data training pattern | |
| CN112182362A (en) | Method and device for training model for online click rate prediction and recommendation system | |
| CN106529185B (en) | A kind of combination forecasting method and system of ancient building displacement | |
| CN107169574A (en) | Using nested machine learning model come the method and system of perform prediction | |
| CN108959409A (en) | The matrix decomposition proposed algorithm of theme and emotion information in a kind of combination comment | |
| CN107832426A (en) | A kind of APP recommendation method and system based on using sequence context | |
| CN110413878A (en) | User based on adaptive elastomeric network-commodity preference prediction meanss and method | |
| CN110138766A (en) | Longicorn must be in conjunction with the network inbreak detection method of random forest | |
| WO2005048184A1 (en) | Active learning method and system | |
| Li et al. | A novel approach based on the Gauss-v SVR with a new hybrid evolutionary algorithm and input vector decision method for port throughput forecasting | |
| CN110796261A (en) | Feature extraction method and device based on reinforcement learning and computer equipment | |
| CN109902823A (en) | A kind of model training method and equipment based on generation confrontation network | |
| Ayyadevara | Neural Networks with Keras Cookbook: Over 70 recipes leveraging deep learning techniques across image, text, audio, and game bots | |
| CN111191722B (en) | Method and device for training prediction model through computer | |
| Barr et al. | A computational theory of the firm | |
| CN111414555A (en) | Personalized recommendation method based on collaborative filtering | |
| CN106529726A (en) | Method of performing classification and recommendation based on stock prediction trends |
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 | ||
| TA01 | Transfer of patent application right | ||
| TA01 | Transfer of patent application right |
Effective date of registration: 20220408 Address after: 518000 west block, phase II, Tian'an innovation and Technology Plaza, Tairan 10th Road, Tian'an community, Shatou street, Futian District, Shenzhen, Guangdong 2001 and 2002 Applicant after: Yirong station information technology (Shenzhen) Co.,Ltd. Address before: 610213 No. 1, floor 3, unit 8, building 28, No. 158, Section 2, new hope Avenue, Huayang street, Tianfu new area, Chengdu, Sichuan Applicant before: Sichuan jinzan Technology Co.,Ltd. |
|
| GR01 | Patent grant | ||
| GR01 | Patent grant |