CN110020896A - A kind of business revenue prediction, model training method, device, equipment and storage medium - Google Patents
A kind of business revenue prediction, model training method, device, equipment and storage medium Download PDFInfo
- Publication number
- CN110020896A CN110020896A CN201910303761.7A CN201910303761A CN110020896A CN 110020896 A CN110020896 A CN 110020896A CN 201910303761 A CN201910303761 A CN 201910303761A CN 110020896 A CN110020896 A CN 110020896A
- Authority
- CN
- China
- Prior art keywords
- business revenue
- data
- recognition
- neural network
- recurrent neural
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
- G06Q30/00—Commerce
- G06Q30/02—Marketing; Price estimation or determination; Fundraising
- G06Q30/0201—Market modelling; Market analysis; Collecting market data
- G06Q30/0202—Market predictions or forecasting for commercial activities
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- Business, Economics & Management (AREA)
- General Physics & Mathematics (AREA)
- Strategic Management (AREA)
- Finance (AREA)
- Development Economics (AREA)
- Accounting & Taxation (AREA)
- Data Mining & Analysis (AREA)
- Computational Linguistics (AREA)
- Biomedical Technology (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- General Health & Medical Sciences (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Evolutionary Computation (AREA)
- Biophysics (AREA)
- Molecular Biology (AREA)
- Artificial Intelligence (AREA)
- Entrepreneurship & Innovation (AREA)
- Life Sciences & Earth Sciences (AREA)
- Health & Medical Sciences (AREA)
- Game Theory and Decision Science (AREA)
- Economics (AREA)
- Marketing (AREA)
- General Business, Economics & Management (AREA)
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
Abstract
The embodiment of the invention provides a kind of business revenue prediction, model training method, device, equipment and storage mediums, wherein, the business revenue prediction technique includes: the virtual present information for obtaining the main broadcaster under target category and receiving in unit time, and the business revenue data time series of target category are determined based on the virtual present information;The business revenue data time series are input in the Recognition with Recurrent Neural Network model of pre-training, the business revenue prediction data time series of the target category is obtained, prediction accuracy can be improved, the realization difficulty of prediction can be reduced, when the business revenue data to each category are predicted, migration is stronger.
Description
Technical field
The present embodiments relate to electric powder prediction more particularly to a kind of prediction of business revenue, model training method, device,
Equipment and storage medium.
Background technique
Under normal circumstances, category refers to the classification or type of object, for example, in game technical field, category is
Title of game, such as heroic alliance, king's honor etc. are all categories.Being broadcast live on platform has various categories, and user can give
Main broadcaster under each category gives virtual present, to make each category, there are business revenues.Wherein, the business revenue of each category depends on
Various factors, for example, it may be the factors such as category own content, main broadcaster.In order to realize the development of live streaming platform, need pair
Each category business revenue is predicted, to formulate the benign development that corresponding strategies promote live streaming platform.
Need to consider various factors to the prediction of each category business revenue, for example, time, season, high-quality main broadcaster's number
The factors such as amount, category content in the related technology, tend not to examine when predicting business revenue since the factor of consideration is more
Consider whole factors, prediction result is caused to be inaccurate;In addition, in the related technology, although the prediction to each category business revenue
More factor is considered, but prediction technique realizes that difficulty is larger, when predicting so as to cause the business revenue to each category, moves
Shifting property is poor.
Summary of the invention
The embodiment of the invention provides a kind of business revenue prediction technique, device, equipment and storage mediums, and it is quasi- that prediction can be improved
Exactness can reduce the realization difficulty of prediction, and when the business revenue data to each category are predicted, migration is stronger.
In a first aspect, the embodiment of the invention provides a kind of business revenue prediction techniques, comprising:
The virtual present information that the main broadcaster under target category receives in unit time is obtained, the virtual present is based on
Information determines the business revenue data time series of target category;
The business revenue data time series are input in the Recognition with Recurrent Neural Network model of pre-training, the target product are obtained
The business revenue prediction data time series of class.
Second aspect, the embodiment of the invention also provides a kind of Recognition with Recurrent Neural Network model training methods, comprising:
The virtual present information that main broadcaster receives within each history unit time under target category is obtained, the virtual gift is based on
Object information determines the business revenue data sample time sequence of target category in training set;
The business revenue data sample time sequence of target category is subjected to equal part according to setting sequence length;Wherein, division
Data bulk in every part of business revenue data sample time sequence is N;
For every part of business revenue data sample time sequence of division, preceding N-1 data are input to Recognition with Recurrent Neural Network mould
In type, the predicted value of n-th data is obtained;
The true value of n-th data in every part of business revenue data sample time sequence of division is input to the predicted value
In loss function, penalty values are obtained, and adjust the network parameter of the Recognition with Recurrent Neural Network model based on the penalty values.
The third aspect, the embodiment of the invention also provides a kind of Recognition with Recurrent Neural Network model training methods, comprising:
The virtual present information that main broadcaster receives within each history unit time under target category is obtained, the virtual gift is based on
Object information determines the business revenue data sample time sequence of target category in training set;
The business revenue data sample time sequence of the target category is subjected to equal part according to setting sequence length;What is divided is every
The data bulk of part business revenue data sample time sequence is M;
For every part of business revenue data sample time sequence of division, preceding M/2 data are input to Recognition with Recurrent Neural Network mould
In type, the predicted value of rear M/2 data is obtained;
By the true value of M/2 data rear in every part of business revenue data sample time sequence of division and rear M/2 data
Predicted value is input in loss function, obtains penalty values, and adjust the Recognition with Recurrent Neural Network model based on the penalty values
Network parameter.
Fourth aspect, the embodiment of the invention provides a kind of business revenue prediction techniques, comprising:
Determine the business revenue data time series of target category;
The business revenue data time series are input in the Recognition with Recurrent Neural Network model of pre-training, the target product are obtained
The business revenue prediction data time series of class.
5th aspect, the embodiment of the invention provides a kind of Recognition with Recurrent Neural Network model training methods, comprising:
The business revenue data sample time sequence of target category in training set is subjected to equal part according to setting sequence length;Its
In, the data bulk in every part of business revenue data sample time sequence of division is N;
For every part of business revenue data sample time sequence of division, preceding N-1 data are input to Recognition with Recurrent Neural Network mould
In type, the predicted value of n-th data is obtained;
The true value of n-th data in every part of business revenue data sample time sequence of division is input to the predicted value
In loss function, penalty values are obtained, and adjust the network parameter of the Recognition with Recurrent Neural Network model based on the penalty values.
6th aspect, the embodiment of the invention provides a kind of Recognition with Recurrent Neural Network model training methods, comprising:
The business revenue data sample time sequence of target category in training set is subjected to equal part according to setting sequence length;It divides
Every part of business revenue data sample time sequence in data bulk be M;
For every part of business revenue data sample time sequence of division, preceding M/2 data are input to Recognition with Recurrent Neural Network mould
In type, the predicted value of rear M/2 data is obtained;
By the true value of M/2 data rear in every part of business revenue data sample time sequence of division and rear M/2 data
Predicted value is input in loss function, obtains penalty values, and adjust the Recognition with Recurrent Neural Network model based on the penalty values
Network parameter.
7th aspect, the embodiment of the invention provides a kind of equipment, comprising:
One or more processors;
Storage device, for storing one or more programs,
When one or more of programs are executed by one or more of processors, so that one or more of processing
Device realizes a kind of business revenue prediction technique provided in an embodiment of the present invention or a kind of circulation nerve net provided in an embodiment of the present invention
Network model training method.
Eighth aspect, the embodiment of the invention provides a kind of computer readable storage mediums, are stored thereon with computer journey
Sequence, which is characterized in that the program realizes a kind of business revenue prediction technique provided in an embodiment of the present invention when being executed by processor, or
A kind of Recognition with Recurrent Neural Network model training method provided in an embodiment of the present invention.
Technical solution provided in an embodiment of the present invention, by the way that target category business revenue data time series are input to pre-training
Recognition with Recurrent Neural Network model in, the business revenue prediction data time series of target category is obtained, by using Recognition with Recurrent Neural Network
Model predicts business revenue data time series, due to Recognition with Recurrent Neural Network model consider in time series front and back data it
Between influence, the influence between the data of front and back may include influence of whole factors to data, so as to improve prediction
Accuracy, and can be omitted the inspection of unit root test and the stationarity to sequence before predicting, it can save more
Time can reduce the difficulty that prediction is realized, predict when using Recognition with Recurrent Neural Network model the business revenue data of each category
When, migration is stronger.
Detailed description of the invention
Fig. 1 is a kind of business revenue prediction technique flow chart provided in an embodiment of the present invention;
Fig. 2 is a kind of Recognition with Recurrent Neural Network model training method flow chart provided in an embodiment of the present invention;
Fig. 3 a is a kind of Recognition with Recurrent Neural Network model training method flow chart provided in an embodiment of the present invention;
Fig. 3 b is the verifying between a kind of business revenue prediction data provided in an embodiment of the present invention and the true value of business revenue data
Comparison schematic diagram;
Fig. 4 is a kind of business revenue prediction technique flow chart provided in an embodiment of the present invention;
Fig. 5 is a kind of Recognition with Recurrent Neural Network model training method flow chart provided in an embodiment of the present invention;
Fig. 6 is a kind of Recognition with Recurrent Neural Network model training method flow chart provided in an embodiment of the present invention;
Fig. 7 is a kind of business revenue prediction meanss structural block diagram provided in an embodiment of the present invention;
Fig. 8 is a kind of Recognition with Recurrent Neural Network model training apparatus structural block diagram provided in an embodiment of the present invention;
Fig. 9 is a kind of Recognition with Recurrent Neural Network model training apparatus structural block diagram provided in an embodiment of the present invention;
Figure 10 is a kind of business revenue prediction meanss structural block diagram provided in an embodiment of the present invention;
Figure 11 is a kind of Recognition with Recurrent Neural Network model training apparatus structural block diagram provided in an embodiment of the present invention;
Figure 12 is a kind of Recognition with Recurrent Neural Network model training apparatus structural block diagram provided in an embodiment of the present invention;
Figure 13 is a kind of device structure schematic diagram provided in an embodiment of the present invention.
Specific embodiment
The present invention is described in further detail with reference to the accompanying drawings and examples.It is understood that this place is retouched
The specific embodiment stated is used only for explaining the present invention rather than limiting the invention.It also should be noted that in order to just
Only the parts related to the present invention are shown in description, attached drawing rather than entire infrastructure.
Fig. 1 is a kind of business revenue prediction technique flow chart provided in an embodiment of the present invention, and the method can be predicted by business revenue
Device executes, and described device can be executed by software and/or hardware, described device can be only fitted to computer, server
In equal electronic equipments.The method can be applied in the scene predicted the business revenue of each category in live streaming platform.
Specifically, the electronic equipment of configuration business revenue prediction meanss can be with main broadcaster end or sight in an application scenarios
Many ends are communicated.Wherein, main broadcaster end and viewer end are mounted on live streaming application, and main broadcaster can be by being mounted on the straight of main broadcaster end
It broadcasts using being broadcast live, spectators can enter the direct broadcasting room of main broadcaster, and and main broadcaster by being mounted on the live streaming application of viewer end
It is interacted.Wherein, spectators can give virtual present to main broadcaster.Electronic equipment configured with business revenue prediction meanss is available
The virtual present information that each main broadcaster receives in unit time under target category, and target is determined according to virtual present information
Category business revenue data time series, and business revenue data time series are input in the Recognition with Recurrent Neural Network model of pre-training, it obtains
To the business revenue prediction data time series of target category.
As shown in Figure 1, technical solution provided in an embodiment of the present invention includes:
S110: obtaining the virtual present information that the main broadcaster under target category receives in unit time, is based on the void
Quasi- present information determines the business revenue data time series of target category.
In embodiments of the present invention, category can be the type or type of object.In live streaming platform, category be can be
The group formed with one species or the object of type.Wherein, there can be at least one main broadcaster under each category, generally
In the case of, there are multiple main broadcasters under each category.
In embodiments of the present invention, the unit time can be daily, per hour, or per minute etc..Time series refers to
By each numerical value of certain phenomenon some statistical indicator on different time, in chronological sequence sequence arrange and the sequence that is formed
Column.Business revenue data time series refer to the sequence that business revenue data are arranged according to chronological order and are formed.
In embodiments of the present invention, virtual present information may include the quantity and information of virtual present, wherein
Every kind of virtual present has certain price, and the virtual present type that main broadcaster receives is different, the business revenue data of main broadcaster also just not phase
Together;Or the quantity of virtual present that main broadcaster receives is not identical, the business revenue data of main broadcaster are not also identical.It can be according to target category
Under the virtual present information that receives in each main broadcaster's unit time, determine the business revenue data in each main broadcaster's unit time respectively,
The value of business revenue data in each main broadcaster's unit time is added, it can obtain target category belonging to each main broadcaster in unit
Business revenue data in time.Business revenue data of the target category in unit time are arranged to form battalion according to chronological order
Receive data time series.
It optionally, can also be by the business revenue data in the unit time of target category within a preset time according to time elder generation
Sequence arrangement afterwards forms business revenue data time series.For example, preset time period can be 7 days, the unit time be can be daily, then
The business revenue data that business revenue data time series can be target category 7 days arrange the sequence to be formed according to chronological order.
In an embodiment of the embodiment of the present invention, optionally, the data in business revenue data time series greater than 0 are super
Preset quantity is crossed, preset quantity, which can according to need, to be determined.If business revenue data time series are 7 days business revenues of target category
Data arrange the sequence to be formed according to chronological order, then preset quantity can be 4 days, i.e., have in business revenue data time series
4 days data are greater than 0.Business revenue data time series are selected by the reasonable quantity of the data greater than 0 as a result, can be improved pre-
The accuracy of survey.
In an embodiment of the embodiment of the present invention, optionally, the business revenue data time series are input to pre-
It can also include: that interpolation processing is carried out to the business revenue data time series before in trained Recognition with Recurrent Neural Network model;It is right
Outlier in the business revenue data time series is smoothed.
Specifically, carrying out interpolation processing to business revenue data time series may is that exist when in business revenue data time series
When null value, the corresponding preset time period of business revenue data time series is determined, and determine that the null value corresponds to the time in preset time period
In position;It is chosen and position pair of the null value in preset time period from the history business revenue data of top n preset time period
The business revenue data answered, the business revenue data based on selection calculate average value, using the average value as the data at the null value.For example,
Target category forms business revenue data time series when the business revenue data of the last week (7 days), wherein the business revenue data of Monday are sky
Value then from choosing the business revenue data of Monday in first 4 weeks in history business revenue data, and calculates average value, using the average value as working as
The business revenue data of Monday in the last week.It is available relatively accurate as a result, by carrying out interpolation processing to business revenue data time series
Business revenue data time series, so as to accurately obtain business revenue prediction data time series.
If judging business revenue data specifically, being smoothed and may is that the outlier in business revenue data time series
The difference between target data and other data in time series is greater than preset value, it is determined that the target data is outlier.
Then, it is determined that the corresponding preset time period of business revenue data time series, and determine that the outlier corresponds to the time in preset time period
In position;It is chosen and position of the outlier in preset time period from the history business revenue data of top n preset time period
Corresponding business revenue data, the business revenue data based on selection calculate average value, using the average value as the data at the outlier.By
This, is by smoothly handling the outlier progress in business revenue data time series, available accurate business revenue data
Time series, so as to accurately obtain business revenue prediction data time series.
It should be noted that carrying out interpolation processing to business revenue data time series and in business revenue data time series
The concrete mode that outlier is smoothed is not limited to above-mentioned mode, can also be using other modes.
S120: the business revenue data time series are input in the Recognition with Recurrent Neural Network model of pre-training, are obtained described
The business revenue prediction data time series of target category.
In embodiments of the present invention, business revenue prediction data time series refers to business revenue prediction data according to chronological order
The sequence of arrangement and formation.Wherein, the sequence length of business revenue prediction data time series can be with the sequence of business revenue time series
Length is identical, or can not also be identical;Optionally, the sequence length of business revenue data time series can be greater than 1, correspondingly,
The sequence length of business revenue prediction data time series can be 1;For example, can be defeated by the business revenue data time series for working as the last week
Enter into the Recognition with Recurrent Neural Network model of pre-training, obtains following one week business revenue prediction data time series.Or it can also be with
To be input in the Recognition with Recurrent Neural Network model of pre-training when the business revenue data time series of the last week, obtain following 10 days or
Following one day business revenue prediction data time series.
In embodiments of the present invention, Recognition with Recurrent Neural Network model includes input layer, hidden layer and output layer, and hidden layer includes
At least one shot and long term remembers lstm layers and at least one dropout layers;Wherein, the quantity of lstm layers and dropout layers can be with
It is determined in the training process of Recognition with Recurrent Neural Network model, the specific training program of Recognition with Recurrent Neural Network model can refer to
The introduction of following embodiments.Lstm layer in the embodiment of the present invention can also recycle GRU layers of (Gated Recurrent using door
Unit it) is replaced.
Wherein, lstm layers may include Multilever neuron, neuron, for based on input current business revenue data and
The memory result of the neuron of level-one determines the memory of Current neural member as a result, and being sent to the memory result of Current neural member
Next stage neuron.Dropout layers, the memory result that can be used for the neuron in lstm layers of STOCHASTIC CONTROL is 0.Wherein, lead to
The problem of gradient that crossing can solve using lstm layers and occur in Recognition with Recurrent Neural Network model iterative process disappears, by using
Dropout layers can be to avoid circulation nerve
The problem of network model over-fitting.
In the related technology, it is established back using series of factors such as time, season, high-quality main broadcaster quantity, category contents
Return the variation tendency and business revenue volume of model prediction future business revenue data, but since the factor of consideration is more, regression model is being built
Immediately by the way of Factor Selection, prediction result can not be caused to be inaccurate in view of whole factors.The present invention is implemented
The technical solution that example provides, by using Recognition with Recurrent Neural Network model to business revenue number it was predicted that due to Recognition with Recurrent Neural Network model
Consider the influence between the data of front and back, for example, the business revenue data on the same day and non-singleton, but with current preceding business revenue number
Accordingly and the business revenue data in future on the same day are there are relationship, and the influence between the data of front and back may include whole factors to data
It influences, so as to improve the accuracy of prediction.
Be broadcast live each category of platform business revenue on the one hand depend on category content and high-quality main broadcaster, on the other hand dependent on good
The activity of good, right times categories itself, the shorter consumption fatigue that will lead to spectators' certain time of movable time interval;
The another aspect activity frequency is too long, is also unable to satisfy the demand of live streaming platform self-growth, therefore, it is necessary to a kind of methods to shift to an earlier date
It predicts the trend of following a period of time category business revenue, and then provides effective reference for activity time or each category business revenue strategy.
In the related technology, moving average model (Autoregressive Integrated is integrated using autoregression
Moving AverageModel, ARIMA) time series to be predicted, the process of prediction includes: the trend of time series, season
Section, Stochastic analysis, ADF unit root test, PQ determine rank, and modelling effect is examined;Wherein, ARIMA model is being modeled and was being predicted
Cheng Zhong needs to test to ADF unit root test and to the stationarity of sequence, when modeling and prediction need to spend more
Between, to make to model or predict that there is biggish realization difficulty, when the business revenue data to each category are predicted, between category
Migration it is poor.Technical solution provided in an embodiment of the present invention, using Recognition with Recurrent Neural Network model to business revenue number it was predicted that
The inspection of ADF unit root test and the stationarity to sequence is omitted during the modeling and prediction of Recognition with Recurrent Neural Network model
It tests, the more time can be saved, the realization difficulty of modeling and prediction can be reduced, when the business revenue data to each category carry out in advance
When survey, migration is stronger between category.
On the basis of the above embodiments, technical solution provided in an embodiment of the present invention further include: to the pre-training
Recognition with Recurrent Neural Network model is initialized;It is received within each history unit time based on the main broadcaster under non-targeted category virtual
Present information determines the business revenue data sample time sequence of non-targeted category;Using the business revenue data sample of the non-targeted category
This time series carries out re -training to the Recognition with Recurrent Neural Network model, updates the Recognition with Recurrent Neural Network model of pre-training;It will
The business revenue data time series of non-targeted category are input in the Recognition with Recurrent Neural Network model of update, obtain the non-targeted category
Business revenue prediction data time series.
Wherein, after business revenue data of the Recognition with Recurrent Neural Network model of pre-training to target category are predicted, when
When needing to predict the business revenue data of non-targeted category, need to initialize Recognition with Recurrent Neural Network model, use is non-
The business revenue data sample time sequence of target category carries out re -training to Recognition with Recurrent Neural Network model, to non-targeted category
Business revenue data are predicted, when replacing category, the process that needs to model Recognition with Recurrent Neural Network model again and again pre-
The process of survey does not need the inspection of the inspection of unit root and the stationarity of sequence, can save the more time, can drop
Low realization difficulty, migration is stronger between category.
On the basis of the above embodiments, technical solution provided in an embodiment of the present invention further include: pre- according to the business revenue
Inflection point data and the inflection point the data corresponding time in measured data time series generates action message.Wherein, inflection point number
According to the minimum point that can be business revenue data.The action message generation time can be time corresponding in inflection point data, action message
Content can the numerical value based on inflection point data and generate, for example, action message content can be with information of discount etc., when inflection point data
Numerical value get over hour, the discount of discounting is bigger.As a result, by according in business revenue prediction data time series inflection point data and
The inflection point data corresponding unit time generates action message, can contribute to the business revenue benign development of live streaming platform.
Technical solution provided in an embodiment of the present invention, by the way that target category business revenue data time series are input to pre-training
Recognition with Recurrent Neural Network model in, the business revenue prediction data time series of target category is obtained, by using Recognition with Recurrent Neural Network
Model predicts business revenue data time series, due to Recognition with Recurrent Neural Network model consider in time series front and back data it
Between influence, the influence between the data of front and back may include influence of whole factors to data, so as to improve prediction
Accuracy, and can be omitted the inspection of unit root test and the stationarity to sequence before predicting, it can save more
Time can reduce the difficulty that prediction is realized, predict when using Recognition with Recurrent Neural Network model the business revenue data of each category
When, migration is stronger.
Fig. 2 is a kind of Recognition with Recurrent Neural Network model training method flow chart provided in an embodiment of the present invention, and the method can
To be executed by Recognition with Recurrent Neural Network model training apparatus, described device can be only fitted to the electronic equipments such as server, computer
In, method provided in an embodiment of the present invention can be trained Recognition with Recurrent Neural Network model provided by the above embodiment, pass through
The Recognition with Recurrent Neural Network model of providing method of embodiment of the present invention training can be applied to the battalion to each category in live streaming platform
Receive the case where data are predicted.
As shown in Fig. 2, technical solution provided in an embodiment of the present invention includes:
S210: obtaining the virtual present information that main broadcaster receives within each history unit time under target category, based on described
Virtual present information determines the business revenue data sample time sequence of target category.
In embodiments of the present invention, the history unit time can be it is past daily, per hour or per minute etc..Time
Sequence refers to each numerical value by certain phenomenon some statistical indicator on different time, in chronological sequence sequence arrangement and shape
At sequence.Business revenue data sample time sequence refers to the sequence that history business revenue data are arranged according to chronological order and are formed
Column.
Wherein, virtual present information may include the quantity and information of virtual present, wherein every kind of virtual present
With certain price, the virtual present type that main broadcaster receives is different, and the business revenue data of main broadcaster are not also just identical;Or main broadcaster receives
The quantity of the virtual present arrived is not identical, and the business revenue data of main broadcaster are not also identical.It can be gone through according to main broadcaster each under target category
The virtual present information received in the history unit time determines the history business revenue data in each main broadcaster's history unit time respectively,
The value of history business revenue data in each main broadcaster's history unit time is added, it can obtain target product belonging to each main broadcaster
History business revenue data of the class within the history unit time.History business revenue data of the target category within each history unit time are pressed
It arranges to form business revenue data sample time sequence according to chronological order.
Optionally, the history business revenue data in each history unit time of target category within a preset time can also be pressed
It arranges to form business revenue data sample time sequence according to chronological order.For example, preset time period can be 7 days, history unit
Time can be over daily, then business revenue data sample time sequence can be target category 7 days history business revenue data according to
Chronological order arranges the sequence to be formed.
In an embodiment of the embodiment of the present invention, optionally, business revenue data sample time sequence is greater than 0 data
More than preset quantity, preset quantity, which can according to need, to be determined.If business revenue data sample time sequence is target category 7 days
History business revenue data the sequence to be formed is arranged according to chronological order, then preset quantity can be 4 days, i.e. business revenue data sample
There are 4 days data to be greater than 0 in this time series.When selecting business revenue data sample by the reasonable quantity of the data greater than 0 as a result,
Between sequence, the precision of model can be improved.
It can also include: to the business revenue data sample time sequence in an embodiment of the embodiment of the present invention
Carry out interpolation processing;Outlier in the business revenue data sample time sequence is smoothed.
Specifically, carrying out interpolation processing to business revenue data sample time sequence may is that when business revenue data sample time sequence
There are when null value, determining the corresponding preset time period of business revenue data sample time sequence in column, and determine that the null value corresponds to the time
Position in preset time period;It chooses with the null value from the history business revenue data of top n preset time period in preset time
The corresponding history business revenue data in position in section, history business revenue data based on selection calculate average value, using the average value as
History business revenue data at the null value.For example, target category nearest one month history business revenue data form business revenue data sample
Time series, wherein this month first day history business revenue data are null value, then preceding 4 middle of the month are chosen from history business revenue data
First day history business revenue data, and calculate average value, using the average value as first day nearest one month history business revenue number
According to.As a result, by carrying out interpolation processing to business revenue data sample time sequence, when available accurate business revenue data sample
Between sequence, thus the precision of model.
If judging business revenue data specifically, being smoothed and may is that the outlier in business revenue data time series
The difference between target data and other data in sample time-series is greater than preset value, it is determined that the target data is to peel off
Value.Then, it is determined that the corresponding preset time period of business revenue data sample time sequence, and determine the outlier corresponding time pre-
If the position in the period;It chooses with the outlier from the history business revenue data of top n preset time period in preset time period
In the corresponding history business revenue data in position, history business revenue data based on selection calculate average value, regard the average value as this
Data at outlier.It, can be with as a result, by smoothly handling the outlier progress in business revenue data sample time sequence
Accurate business revenue data sample time sequence is obtained, the precision of model can be improved.
S220: the business revenue data sample time sequence of target category is subjected to equal part according to setting sequence length;Wherein, it draws
The data bulk for the every part of business revenue data sample time sequence divided is N.
In embodiments of the present invention, setting sequence length, which can according to need, is set, wherein when business revenue data sample
Between the part that cannot be divided of end of sequence can be removed.
S230: for every part of business revenue data sample time sequence of division, preceding N-1 data are input to circulation nerve net
In network model, the predicted value of n-th data is obtained.
In embodiments of the present invention, training Recognition with Recurrent Neural Network model before, can to Recognition with Recurrent Neural Network model into
Row parameter setting, wherein parameter includes the sequence length of business revenue data sample time sequence: for example, sequence_length=
14;When model training, the sample size of sample time-series used in a subgradient is calculated: such as, Batch_size=13;Mould
Type iterative steps: such as, epochs=1;Steps_per_epoch=1/13.
Wherein, Recognition with Recurrent Neural Network model includes input layer, hidden layer and output layer, and hidden layer includes at least one length
Phase remembers lstm layers and at least one dropout layers.Wherein, hidden layer may include five layers, lstm layers of first layer, may include
100 neurons, dropout layers of the second layer, dropout ratio 10%, third layer is lstm layers, may include 100 nerves
Member;4th layer identical as third layer, and layer 5 is identical as the second layer.Wherein, the number of plies of the hidden layer of Recognition with Recurrent Neural Network model
It can according to need carry out initial setting up.
S240: by the true value of n-th data in every part of business revenue data sample time sequence of division and the predicted value
It is input in loss function, obtains penalty values, and adjust the network ginseng of the Recognition with Recurrent Neural Network model based on the penalty values
Number.
Wherein, loss function can be the function established based on mean square error, and the size of penalty values can be used for judging N
Gap between the true value of a data and the predicted value of n-th data can characterize the gap of the two when penalty values are bigger
It is bigger;When penalty values are smaller, the gap for characterizing the two is smaller.The net of Recognition with Recurrent Neural Network model can be adjusted according to penalty values
Network parameter, the method specifically adjusted can use adam algorithm.
In embodiments of the present invention, after the completion of Recognition with Recurrent Neural Network model training, the sample of training set can be used
Time series verifies Recognition with Recurrent Neural Network model, to determine the training result of Recognition with Recurrent Neural Network model, can will draw
The 80% business revenue data sample time sequence divided, can be by the business revenue of the 20% of division for training Recognition with Recurrent Neural Network model
Data sample time sequence forms inspection set, for detecting Recognition with Recurrent Neural Network model.
Method provided in an embodiment of the present invention carries out Recognition with Recurrent Neural Network model by business revenue data sample time sequence
Training, to model to Recognition with Recurrent Neural Network model, forms the Recognition with Recurrent Neural Network model of pre-training, can accurately predict not
The business revenue data for carrying out 1 unit time, during model modeling, without carrying out the inspection of unit root and the stationarity of sequence
It examines, the more time can be saved, reduce modeling difficulty and model maintenance difficulty.
Fig. 3 a is a kind of Recognition with Recurrent Neural Network model training method flow chart provided in an embodiment of the present invention, and the method can
To be executed by Recognition with Recurrent Neural Network model training apparatus, described device can be only fitted to the electronic equipments such as server, computer
In, method provided in an embodiment of the present invention can be trained Recognition with Recurrent Neural Network model provided by the above embodiment, pass through
The Recognition with Recurrent Neural Network model of providing method of embodiment of the present invention training can be applied to the battalion to each category in live streaming platform
Receive the case where data are predicted.
As shown in Figure 3a, technical solution provided in an embodiment of the present invention includes:
S310: obtaining the virtual present information that main broadcaster receives within each history unit time under target category, based on described
Virtual present information determines the business revenue data sample time sequence of target category.
Specific introduce of this step can be with reference to the introduction of S210 step in above-described embodiment.
S320: the business revenue data sample time sequence of the target category is subjected to equal part according to setting sequence length;It draws
The data bulk for the every part of business revenue data sample time sequence divided is M.
Wherein, setting sequence length, which can according to need, is set, wherein the end of business revenue data sample time sequence
The part that cannot be divided can be removed.
S330: for every part of business revenue data sample time sequence of division, preceding M/2 data are input to circulation nerve net
In network model, the predicted value of rear M/2 data is obtained.
It, can be to Recognition with Recurrent Neural Network model before training Recognition with Recurrent Neural Network model in embodiments of the present invention
Carry out parameter setting, wherein parameter includes the sequence length of business revenue data sample time sequence: for example, sequence_length
=14;When model training, the sample size of sample time-series used in a subgradient is calculated: such as, Batch_size=13;
Model iterative steps: such as, epochs=1;Steps_per_epoch=1/13.
Wherein, Recognition with Recurrent Neural Network model includes input layer, hidden layer and output layer, and hidden layer includes at least one length
Phase remembers lstm layers and at least one dropout layers.Wherein, hidden layer may include five layers, lstm layers of first layer, may include
100 neurons, dropout layers of the second layer, dropout ratio 10%, third layer is lstm layers, may include 100 nerves
Member;4th layer identical as third layer, and layer 5 is identical as the second layer.Wherein, the number of plies of the hidden layer of Recognition with Recurrent Neural Network model
It can according to need carry out initial setting up.
S340: by the true value and rear M/2 number of M/2 data rear in every part of business revenue data sample time sequence of division
According to predicted value be input in loss function, obtain penalty values, and adjust the Recognition with Recurrent Neural Network mould based on the penalty values
The network parameter of type.
Wherein, loss function can be the function established based on mean square error, after the size of penalty values can be used for judgement
Both gap between the true value of M/2 data and the predicted value of rear M/2 data, when penalty values are bigger, can characterize
Gap it is bigger;When penalty values are smaller, the gap for characterizing the two is smaller.Recognition with Recurrent Neural Network mould can be adjusted according to penalty values
The network parameter of type, the method specifically adjusted can use adam algorithm.
It in embodiments of the present invention, can be using business revenue in training set after the completion of Recognition with Recurrent Neural Network model training
The sample time-series of data verify Recognition with Recurrent Neural Network model, to determine the training knot of Recognition with Recurrent Neural Network model
Fruit can be used for the business revenue data sample time sequence of the 80% of division to train Recognition with Recurrent Neural Network model, can will divide
20% business revenue data sample time sequence formed inspection set, for verifying Recognition with Recurrent Neural Network model.As shown in Figure 3b, a
For the curve that business revenue prediction data is formed, b is the curve that business revenue data true value is formed, and passes through two songs in comparison chart 3b
Line, the obtained model of training can the business revenue data very well to target category predict, can predict in business revenue data most
Low spot so as to generate action message in time helps that the benign development of platform business revenue is broadcast live.
Method provided in an embodiment of the present invention carries out Recognition with Recurrent Neural Network model by business revenue data sample time sequence
Training, to model to Recognition with Recurrent Neural Network model, forms the Recognition with Recurrent Neural Network model of pre-training, can accurately predict not
Carry out the business revenue data in multiple unit time, during model modeling, without carrying out the inspection of unit root and putting down for sequence
Stability is examined, and the more time can be saved, and is reduced modeling difficulty, is reduced model maintenance difficulty.
Fig. 4 is a kind of business revenue prediction technique flow chart provided in an embodiment of the present invention, and the method can be predicted by business revenue
Device executes, and described device can be executed by software and/or hardware, described device can be only fitted to computer, server
In equal electronic equipments.The method can be applied in the scene predicted the business revenue of each category in live streaming platform,
It can be applied in the business revenue being able to carry out in other platforms of consumer sale prediction scene, for example, to each category of businessman
The scene predicted of business revenue in.
As shown in figure 4, technical solution provided in an embodiment of the present invention includes:
S410: the business revenue data time series of target category are determined.
S420: the business revenue data time series are input in the Recognition with Recurrent Neural Network model of pre-training, are obtained described
The business revenue prediction data time series of target category.
In embodiments of the present invention, the step introduction in S410 and S420 may refer to S110 and S120 in above-described embodiment
In introduction.Wherein, principle is identical with the mode of data processing.
Technical solution provided in an embodiment of the present invention, by the way that target category business revenue data time series are input to pre-training
Recognition with Recurrent Neural Network model in, the business revenue prediction data time series of target category is obtained, by using Recognition with Recurrent Neural Network
Model predicts business revenue data time series, due to Recognition with Recurrent Neural Network model consider in time series front and back data it
Between influence, the influence between the data of front and back may include influence of whole factors to data, so as to improve prediction
Accuracy, and can be omitted the inspection of unit root test and the stationarity to sequence before predicting, it can save more
Time can reduce the difficulty that prediction is realized, predict when using Recognition with Recurrent Neural Network model the business revenue data of each category
When, migration is stronger.
Fig. 5 is a kind of Recognition with Recurrent Neural Network model training method provided in an embodiment of the present invention, and the method can be by following
Ring neural network model training device executes, and described device can be only fitted in the electronic equipments such as server, computer, this hair
The Recognition with Recurrent Neural Network model that the method that bright embodiment provides can provide method shown in Fig. 4 is trained, through the invention
The Recognition with Recurrent Neural Network model for the method training that embodiment provides can be applied to the business revenue number to each category in live streaming platform
The case where according to being predicted, or also can be applied to the business revenue prediction case in other platforms of consumer sale.
S510: the business revenue data sample time sequence of target category in training set is carried out according to setting sequence length etc.
Point;Wherein, the data bulk of every part of business revenue data sample time sequence of division is N;
S520: for every part of business revenue data sample time sequence of division, preceding N-1 data are input to circulation nerve net
In network model, the predicted value of n-th data is obtained;
S530: by the true value of n-th data in every part of business revenue data sample time sequence of division and the predicted value
It is input in loss function, obtains penalty values, and adjust the network ginseng of the Recognition with Recurrent Neural Network model based on the penalty values
Number.
Being discussed in detail for S510-530 provided in an embodiment of the present invention can be with reference to being discussed in detail in S210-S230.
Technical solution provided in an embodiment of the present invention, by business revenue data sample time sequence to Recognition with Recurrent Neural Network model
It is trained, to model to Recognition with Recurrent Neural Network model, forms the Recognition with Recurrent Neural Network model of pre-training, it can be accurately pre-
The business revenue data for surveying following 1 unit time, during model modeling, without carrying out the inspection of unit root and putting down for sequence
Stability is examined, and the more time can be saved, and reduces modeling difficulty and model maintenance difficulty.
Fig. 6 is a kind of Recognition with Recurrent Neural Network model training method flow chart provided in an embodiment of the present invention, and the method can
To be executed by Recognition with Recurrent Neural Network model training apparatus, described device can be only fitted to the electronic equipments such as server, computer
In, the Recognition with Recurrent Neural Network model that method provided in an embodiment of the present invention can provide method shown in Fig. 4 is trained, and is passed through
The neural network model of providing method of embodiment of the present invention training can be applied to the business revenue number to each category in live streaming platform
The case where according to being predicted, or also can be applied to the business revenue prediction case in other platforms of consumer sale.
S610: the business revenue data sample time sequence of target category in training set is carried out according to setting sequence length etc.
Point;The data bulk of the every part of business revenue data sample time sequence divided is M.
S620: for every part of business revenue data sample time sequence of division, preceding M/2 data are input to circulation nerve net
In network model, the predicted value of rear M/2 data is obtained.
S630 is by the true value of M/2 data after in every part of business revenue data sample time sequence of division and rear M/2 number
According to predicted value be input in loss function, obtain penalty values, and adjust the Recognition with Recurrent Neural Network mould based on the penalty values
The network parameter of type.
Being discussed in detail for S610-630 provided in an embodiment of the present invention can be with reference to being discussed in detail in S310-S330.
Technical solution provided in an embodiment of the present invention, by business revenue data sample time sequence to Recognition with Recurrent Neural Network model
It is trained, to model to Recognition with Recurrent Neural Network model, forms the Recognition with Recurrent Neural Network model of pre-training, it can be accurately pre-
The business revenue data of surveying following multiple unit time, during model modeling, without carrying out the inspection and sequence of unit root
Stationary test can save the more time, reduce modeling difficulty and model maintenance difficulty.
Fig. 7 is a kind of business revenue prediction meanss structural block diagram provided in an embodiment of the present invention, and institute is as shown in fig. 7, described device
It include: the first determining module 710 and the first prediction module 720.
First determining module 710, the virtual present received in unit time for obtaining the main broadcaster under target category
Information determines the business revenue data time series of target category based on the virtual present information;
First prediction module 720, for the business revenue data time series to be input to the Recognition with Recurrent Neural Network of pre-training
In model, the business revenue prediction data time series of the target category is obtained.
Optionally, the sequence length of the business revenue data time series is greater than 1, correspondingly, business revenue prediction data time sequence
The sequence length of column is 1;
The sequence length phase of the sequence length of the business revenue data time series and the business revenue prediction data time series
Together.
Optionally, the Recognition with Recurrent Neural Network model includes input layer, hidden layer and output layer, and the hidden layer includes extremely
A few shot and long term remembers lstm layers and at least one dropout layers.
Optionally, described device further includes preprocessing module, is used for:
Interpolation processing is carried out to the business revenue data time series;
Outlier in the business revenue data time series is smoothed.
Optionally, the data in the business revenue data time series greater than 0 are more than preset quantity.
Optionally, described device further include: action message generation module, for according to the business revenue prediction data time sequence
Inflection point data and the inflection point the data corresponding time in column generates action message.
Optionally, described device further include: retraining/prediction module is used for:
The Recognition with Recurrent Neural Network model of the pre-training is initialized;
Based on the virtual present information that the main broadcaster under non-targeted category receives within each history unit time, determine non-targeted
The business revenue data sample time sequence of category;
Weight is carried out to the Recognition with Recurrent Neural Network model using the business revenue data sample time sequence of the non-targeted category
New training, updates the Recognition with Recurrent Neural Network of the pre-training;
The business revenue data time series of non-targeted category are input in the Recognition with Recurrent Neural Network of update, the non-mesh is obtained
Mark the business revenue prediction data time series of category.
Method provided by any embodiment of the invention can be performed in above-mentioned apparatus, has the corresponding functional module of execution method
And beneficial effect.
Fig. 8 is a kind of Recognition with Recurrent Neural Network model training apparatus structural block diagram, as shown in figure 8, the embodiment of the present invention provides
Device include: the sub-modules 820 such as first sample determining module 810, first, the first input module 830 and the first parameter adjustment
Module 840.
First sample determining module 810, for obtaining the void that main broadcaster receives within each history unit time under target category
Quasi- present information, the business revenue data sample time sequence of target category in training set is determined based on the virtual present information;
First equal sub-modules 820, for by the business revenue data sample time sequence of target category according to setting sequence length
Carry out equal part;Wherein, the data bulk in every part of business revenue data sample time sequence of division is N;
First input module 830, for every part of business revenue data sample time sequence for division, by preceding N-1 data
It is input in Recognition with Recurrent Neural Network model, obtains the predicted value of n-th data;
First parameter adjustment module 840, n-th data in every part of business revenue data sample time sequence for that will divide
True value and the predicted value are input in loss function, obtain penalty values, and adjust the circulation mind based on the penalty values
Network parameter through network model.
Method provided by any embodiment of the invention can be performed in above-mentioned apparatus, has the corresponding functional module of execution method
And beneficial effect.
Fig. 9 is a kind of Recognition with Recurrent Neural Network model training apparatus structural block diagram, as shown in figure 9, described device includes: second
The sub-modules 920 such as sample determining module 910, second, the second input module 930 and the second parameter adjustment module 940.
Second sample determining module 910, for obtaining the void that main broadcaster receives within each history unit time under target category
Quasi- present information, the business revenue data sample time sequence of target category in training set is determined based on the virtual present information;
Second equal sub-modules 920, for by the business revenue data sample time sequence of the target category according to setting sequence
Length carries out equal part;The data bulk of the every part of business revenue data sample time sequence divided is M;
Second input module 930, for every part of business revenue data sample time sequence for division, by preceding M/2 data
It is input in Recognition with Recurrent Neural Network model, obtains the predicted value of rear M/2 data;
Second parameter adjustment module 940, in every part of business revenue data sample time sequence for that will divide after M/2 data
True value and the predicted values of rear M/2 data be input in loss function, obtain penalty values, and based on penalty values adjustment
The network parameter of the Recognition with Recurrent Neural Network model.
Figure 10 is a kind of business revenue prediction meanss structural block diagram provided in an embodiment of the present invention, and as shown in Figure 10, the present invention is real
The device for applying example offer includes: the second determining module 1010 and the second prediction module 1020.
Second determining module 1010, for determining the business revenue data time series of target category;
Second prediction module 1020, for the business revenue data time series to be input to the Recognition with Recurrent Neural Network of pre-training
In model, the business revenue prediction data time series of the target category is obtained.
Optionally, the sequence length of the business revenue data time series is greater than 1, correspondingly, business revenue prediction data time sequence
The sequence length of column is 1;
The sequence length phase of the sequence length of the business revenue data time series and the business revenue prediction data time series
Together.
Optionally, the Recognition with Recurrent Neural Network model includes input layer, hidden layer and output layer, and the hidden layer includes extremely
A few shot and long term remembers lstm layers and at least one dropout layers.
Optionally, described device further includes preprocessing module, is used for:
Interpolation processing is carried out to the business revenue data time series;
Outlier in the business revenue data time series is smoothed.
Optionally, the data in the business revenue data time series greater than 0 are more than preset quantity.
Optionally, described device further include: action message generation module, for according to the business revenue prediction data time sequence
Inflection point data and the inflection point the data corresponding time in column generates action message.
Optionally, described device further include: retraining/prediction module is used for:
The Recognition with Recurrent Neural Network model of the pre-training is initialized;
Based on the virtual present information that the main broadcaster under non-targeted category receives within each history unit time, determine non-targeted
The business revenue data sample time sequence of category;
Weight is carried out to the Recognition with Recurrent Neural Network model using the business revenue data sample time sequence of the non-targeted category
New training, updates the Recognition with Recurrent Neural Network model of the pre-training;
The business revenue data time series of non-targeted category are input in the Recognition with Recurrent Neural Network model of update, are obtained described
The business revenue prediction data time series of non-targeted category.
Method provided by any embodiment of the invention can be performed in above-mentioned apparatus, has the corresponding functional module of execution method
And beneficial effect.
Figure 11 is a kind of Recognition with Recurrent Neural Network model training apparatus structural block diagram provided in an embodiment of the present invention, such as Figure 11 institute
Show, described device includes: the sub-modules such as third 1110, third input module 1120 and third parameter adjustment module 1130.
Third division module 1110, by the business revenue data sample time sequence of target category in training set according to setting sequence
Length carries out equal part;Wherein, the data bulk in every part of business revenue data sample time sequence of division is N;
Third input module 1120 inputs preceding N-1 data for every part of business revenue data sample time sequence of division
Into Recognition with Recurrent Neural Network model, the predicted value of n-th data is obtained;
Third parameter adjustment module 1130, by every part of business revenue data sample time sequence of division n-th data it is true
Real value and the predicted value are input in loss function, obtain penalty values, and adjust the circulation nerve based on the penalty values
The network parameter of network model.
Method provided by any embodiment of the invention can be performed in above-mentioned apparatus, has the corresponding functional module of execution method
And beneficial effect.
Figure 12 is a kind of Recognition with Recurrent Neural Network model training apparatus structural block diagram provided in an embodiment of the present invention, such as Figure 12 institute
Show, described device includes: quartering module 1210, the 4th input module 1220 and the 4th parameter adjustment module 1230.
Quartering module 1210, for by the business revenue data sample time sequence of target category in training set according to setting
Sequence length carries out equal part;The data bulk in every part of business revenue data sample time sequence divided is M;
4th input module 1220, for every part of business revenue data sample time sequence for division, by preceding M/2 data
It is input in Recognition with Recurrent Neural Network model, obtains the predicted value of rear M/2 data;
4th parameter adjustment module 1230, in every part of business revenue data sample time sequence for that will divide after M/2 number
According to true value and the predicted values of rear M/2 data be input in loss function, obtain penalty values, and based on the penalty values tune
The network parameter of the whole Recognition with Recurrent Neural Network model.
Method provided by any embodiment of the invention can be performed in above-mentioned apparatus, has the corresponding functional module of execution method
And beneficial effect.
Figure 13 is a kind of device structure schematic diagram provided in an embodiment of the present invention, and as shown in figure 13, which includes:
One or more processors 1310, in Figure 13 by taking a processor 1310 as an example;
Memory 1320;
The equipment can also include: input unit 1330 and output device 1340.
Processor 1310, memory 1320, input unit 1330 and output device 1340 in the equipment can pass through
Bus or other modes connect, in Figure 13 for being connected by bus.
Memory 1320 is used as a kind of non-transient computer readable storage medium, can be used for storing software program, computer
Executable program and module, such as the corresponding program instruction of one of embodiment of the present invention business revenue prediction technique/module (example
Such as, attached first determining module 710 shown in Fig. 7 and the first prediction module 720, or the second determining module as shown in Figure 10
1010 and second prediction module 1020), or such as one of embodiment of the present invention Recognition with Recurrent Neural Network model training method pair
Sequence instruction/the module answered is (for example, the sub-modules 820, first such as attached first sample determining module shown in Fig. 8 810, first input
The sub-modules such as module 830 and the first parameter adjustment module 840 or attached second sample determining module 910, second shown in Fig. 9
920, the sub-modules 1110, third such as third shown in the second input module 930 and the second parameter adjustment module 940 or Figure 11
Quartering module the 1210, the 4th shown in input module 1120 and third parameter adjustment module 1130 or Figure 12 inputs mould
Block 1220 and the 4th parameter adjustment module 1230).Processor 1310 by operation be stored in memory 1320 software program,
Instruction and module, thereby executing the various function application and data processing of computer equipment, i.e. the realization above method is implemented
A kind of business revenue prediction technique of example, it may be assumed that
The virtual present information that the main broadcaster under target category receives in unit time is obtained, the virtual present is based on
Information determines the business revenue data time series of target category;
The business revenue data time series are input in the Recognition with Recurrent Neural Network model of pre-training, the target product are obtained
The business revenue prediction data time series of class.Or;
Determine the business revenue data time series of target category;
The business revenue data time series are input in the Recognition with Recurrent Neural Network model of pre-training, the target product are obtained
The business revenue prediction data time series of class.
Or realize a kind of Recognition with Recurrent Neural Network model training method of above method embodiment, it may be assumed that
The virtual present information that main broadcaster receives within each history unit time under target category is obtained, the virtual gift is based on
Object information determines the business revenue data sample time sequence of target category in training set;
The business revenue data sample time sequence of target category is subjected to equal part according to setting sequence length;Wherein, division
Data bulk in every part of business revenue data sample time sequence is N;
For every part of business revenue data sample time sequence of division, preceding N-1 data are input to Recognition with Recurrent Neural Network mould
In type, the predicted value of n-th data is obtained;
The true value of n-th data in every part of business revenue data sample time sequence of division is input to the predicted value
In loss function, penalty values are obtained, and adjust the network parameter of the Recognition with Recurrent Neural Network model based on the penalty values.
Or;
The virtual present information that main broadcaster receives within each history unit time under target category is obtained, the virtual gift is based on
Object information determines the business revenue data sample time sequence of target category in training set;
The business revenue data sample time sequence of the target category is subjected to equal part according to setting sequence length;What is divided is every
The data bulk of part business revenue data sample time sequence is M;
For every part of business revenue data sample time sequence of division, preceding M/2 data are input to Recognition with Recurrent Neural Network mould
In type, the predicted value of rear M/2 data is obtained;
By the true value of M/2 data rear in every part of business revenue data sample time sequence of division and rear M/2 data
Predicted value is input in loss function, obtains penalty values, and adjust the Recognition with Recurrent Neural Network model based on the penalty values
Network parameter.
Or;
The business revenue data sample time sequence of target category in training set is subjected to equal part according to setting sequence length;Its
In, the data bulk in every part of business revenue data sample time sequence of division is N;
For every part of business revenue data sample time sequence of division, preceding N-1 data are input to Recognition with Recurrent Neural Network mould
In type, the predicted value of n-th data is obtained;
The true value of n-th data in every part of business revenue data sample time sequence of division is input to the predicted value
In loss function, penalty values are obtained, and adjust the network parameter of the Recognition with Recurrent Neural Network model based on the penalty values.
Or;
The business revenue data sample time sequence of target category in training set is subjected to equal part according to setting sequence length;It divides
Every part of business revenue data sample time sequence in data bulk be M;
For every part of business revenue data sample time sequence of division, preceding M/2 data are input to Recognition with Recurrent Neural Network mould
In type, the predicted value of rear M/2 data is obtained;
By the true value of M/2 data rear in every part of business revenue data sample time sequence of division and rear M/2 data
Predicted value is input in loss function, obtains penalty values, and adjust the Recognition with Recurrent Neural Network model based on the penalty values
Network parameter.
Memory 1320 may include storing program area and storage data area, wherein storing program area can store operation system
Application program required for system, at least one function;Storage data area can be stored to be created according to using for computer equipment
Data etc..In addition, memory 1320 may include high-speed random access memory, it can also include non-transitory memory, example
Such as at least one disk memory, flush memory device or other non-transitory solid-state memories.In some embodiments, it deposits
Optional reservoir 1320 includes the memory remotely located relative to processor 1310, these remote memories can be connected by network
It is connected to terminal device.The example of above-mentioned network include but is not limited to internet, intranet, local area network, mobile radio communication and
A combination thereof.
Input unit 1330 can be used for receiving the number or character information of input, and generate the user with computer equipment
Setting and the related key signals input of function control.Output device 1340 may include that display screen etc. shows equipment.
The embodiment of the invention provides a kind of computer readable storage mediums, are stored thereon with computer program, the program
A kind of such as business revenue prediction technique provided in an embodiment of the present invention is realized when being executed by processor:
The virtual present information that the main broadcaster under target category receives in unit time is obtained, the virtual present is based on
Information determines the business revenue data time series of target category;
The business revenue data time series are input in the Recognition with Recurrent Neural Network model of pre-training, the target product are obtained
The business revenue prediction data time series of class.Or;
Determine the business revenue data time series of target category;
The business revenue data time series are input in the Recognition with Recurrent Neural Network model of pre-training, the target product are obtained
The business revenue prediction data time series of class.
Or realize a kind of Recognition with Recurrent Neural Network model training method of above method embodiment, it may be assumed that
The virtual present information that main broadcaster receives within each history unit time under target category is obtained, the virtual gift is based on
Object information determines the business revenue data sample time sequence of target category in training set;
The business revenue data sample time sequence of target category is subjected to equal part according to setting sequence length;Wherein, division
Data bulk in every part of business revenue data sample time sequence is N;
For every part of business revenue data sample time sequence of division, preceding N-1 data are input to Recognition with Recurrent Neural Network mould
In type, the predicted value of n-th data is obtained;
The true value of n-th data in every part of business revenue data sample time sequence of division is input to the predicted value
In loss function, penalty values are obtained, and adjust the network parameter of the Recognition with Recurrent Neural Network model based on the penalty values.
Or;
The virtual present information that main broadcaster receives within each history unit time under target category is obtained, the virtual gift is based on
Object information determines the business revenue data sample time sequence of target category in training set;
The business revenue data sample time sequence of the target category is subjected to equal part according to setting sequence length;What is divided is every
The data bulk of part business revenue data sample time sequence is M;
For every part of business revenue data sample time sequence of division, preceding M/2 data are input to Recognition with Recurrent Neural Network mould
In type, the predicted value of rear M/2 data is obtained;
By the true value of M/2 data rear in every part of business revenue data sample time sequence of division and rear M/2 data
Predicted value is input in loss function, obtains penalty values, and adjust the Recognition with Recurrent Neural Network model based on the penalty values
Network parameter.
Or;
The business revenue data sample time sequence of target category in training set is subjected to equal part according to setting sequence length;Its
In, the data bulk in every part of business revenue data sample time sequence of division is N;
For every part of business revenue data sample time sequence of division, preceding N-1 data are input to Recognition with Recurrent Neural Network mould
In type, the predicted value of n-th data is obtained;
The true value of n-th data in every part of business revenue data sample time sequence of division is input to the predicted value
In loss function, penalty values are obtained, and adjust the network parameter of the Recognition with Recurrent Neural Network model based on the penalty values.
Or;
The business revenue data sample time sequence of target category in training set is subjected to equal part according to setting sequence length;It divides
Every part of business revenue data sample time sequence in data bulk be M;
For every part of business revenue data sample time sequence of division, preceding M/2 data are input to Recognition with Recurrent Neural Network mould
In type, the predicted value of rear M/2 data is obtained;
By the true value of M/2 data rear in every part of business revenue data sample time sequence of division and rear M/2 data
Predicted value is input in loss function, obtains penalty values, and adjust the Recognition with Recurrent Neural Network model based on the penalty values
Network parameter.
It can be using any combination of one or more computer-readable media.Computer-readable medium can be calculating
Machine readable signal medium or computer readable storage medium.Computer readable storage medium for example can be --- but it is unlimited
In system, device or the device of --- electricity, magnetic, optical, electromagnetic, infrared ray or semiconductor, or any above combination.It calculates
The more specific example (non exhaustive list) of machine readable storage medium storing program for executing includes: electrical connection with one or more conducting wires, just
Taking formula computer disk, hard disk, random access memory (RAM), read-only memory (ROM), erasable type may be programmed read-only storage
Device (EPROM or flash memory), optical fiber, portable compact disc read-only memory (CD-ROM), light storage device, magnetic memory device,
Or above-mentioned any appropriate combination.In this document, computer readable storage medium can be it is any include or storage journey
The tangible medium of sequence, the program can be commanded execution system, device or device use or in connection.
Computer-readable signal media may include in a base band or as carrier wave a part propagate data-signal,
Wherein carry computer-readable program code.The data-signal of this propagation can take various forms, including --- but
It is not limited to --- electromagnetic signal, optical signal or above-mentioned any appropriate combination.Computer-readable signal media can also be
Any computer-readable medium other than computer readable storage medium, which can send, propagate or
Transmission is for by the use of instruction execution system, device or device or program in connection.
The program code for including on computer-readable medium can transmit with any suitable medium, including --- but it is unlimited
In --- wireless, electric wire, optical cable, RF etc. or above-mentioned any appropriate combination.
The computer for executing operation of the present invention can be write with one or more programming languages or combinations thereof
Program code, described program design language include object oriented program language-such as Java, Smalltalk, C++,
Further include conventional procedural programming language-such as " C " language or similar programming language.Program code can be with
It fully executes, partly execute on the user computer on the user computer, being executed as an independent software package, portion
Divide and partially executes or executed on a remote computer or server completely on the remote computer on the user computer.?
Be related in the situation of remote computer, remote computer can pass through the network of any kind --- including local area network (LAN) or
Wide area network (WAN)-be connected to subscriber computer, or, it may be connected to outer computer (such as mentioned using Internet service
It is connected for quotient by internet).
Note that the above is only a better embodiment of the present invention and the applied technical principle.It will be appreciated by those skilled in the art that
The invention is not limited to the specific embodiments described herein, be able to carry out for a person skilled in the art it is various it is apparent variation,
It readjusts and substitutes without departing from protection scope of the present invention.Therefore, although being carried out by above embodiments to the present invention
It is described in further detail, but the present invention is not limited to the above embodiments only, without departing from the inventive concept, also
It may include more other equivalent embodiments, and the scope of the invention is determined by the scope of the appended claims.
Claims (14)
1. a kind of business revenue prediction technique characterized by comprising
The virtual present information that the main broadcaster under target category receives in unit time is obtained, the virtual present information is based on
Determine the business revenue data time series of target category;
The business revenue data time series are input in the Recognition with Recurrent Neural Network model of pre-training, the target category is obtained
Business revenue prediction data time series.
2. the method according to claim 1, wherein
The sequence length of the business revenue data time series is greater than 1, correspondingly, the sequence length of business revenue prediction data time series
It is 1;
The sequence length of the business revenue data time series is identical as the sequence length of the business revenue prediction data time series.
3. the method according to claim 1, wherein
The Recognition with Recurrent Neural Network model includes input layer, hidden layer and output layer, and the hidden layer includes at least one length
Phase remembers lstm layers and at least one dropout layers.
4. the method according to claim 1, wherein further include:
Interpolation processing is carried out to the business revenue data time series;
Outlier in the business revenue data time series is smoothed.
5. the method according to claim 1, wherein the data in the business revenue data time series greater than 0 are super
Cross preset quantity.
6. the method according to claim 1, wherein further include:
According to the inflection point data and the corresponding time generation work of the inflection point data in the business revenue prediction data time series
Dynamic information.
7. the method according to claim 1, wherein further include:
The Recognition with Recurrent Neural Network model of the pre-training is initialized;
Based on the virtual present information that the main broadcaster under non-targeted category receives within each history unit time, non-targeted category is determined
Business revenue data sample time sequence;
The Recognition with Recurrent Neural Network model is instructed again using the business revenue data sample time sequence of the non-targeted category
Practice, updates the Recognition with Recurrent Neural Network model of the pre-training;
The business revenue data time series of non-targeted category are input in the Recognition with Recurrent Neural Network model of update, the non-mesh is obtained
Mark the business revenue prediction data time series of category.
8. a kind of Recognition with Recurrent Neural Network model training method characterized by comprising
The virtual present information that main broadcaster receives within each history unit time under target category is obtained, is believed based on the virtual present
Cease the business revenue data sample time sequence for determining target category in training set;
The business revenue data sample time sequence of target category is subjected to equal part according to setting sequence length;Wherein, every part of division
Data bulk in business revenue data sample time sequence is N;
For every part of business revenue data sample time sequence of division, preceding N-1 data are input in Recognition with Recurrent Neural Network model,
Obtain the predicted value of n-th data;
The true value of n-th data in every part of business revenue data sample time sequence of division and the predicted value are input to loss
In function, penalty values are obtained, and adjust the network parameter of the Recognition with Recurrent Neural Network model based on the penalty values.
9. a kind of Recognition with Recurrent Neural Network model training method characterized by comprising
The virtual present information that main broadcaster receives within each history unit time under target category is obtained, is believed based on the virtual present
Cease the business revenue data sample time sequence for determining target category in training set;
The business revenue data sample time sequence of the target category is subjected to equal part according to setting sequence length;The every part of battalion divided
The data bulk for receiving data sample time sequence is M;
For every part of business revenue data sample time sequence of division, preceding M/2 data are input in Recognition with Recurrent Neural Network model,
Obtain the predicted value of rear M/2 data;
By the prediction of the true value of M/2 data rear in every part of business revenue data sample time sequence of division and rear M/2 data
Value is input in loss function, obtains penalty values, and the network of the Recognition with Recurrent Neural Network model is adjusted based on the penalty values
Parameter.
10. a kind of business revenue prediction technique characterized by comprising
Determine the business revenue data time series of target category;
The business revenue data time series are input in the Recognition with Recurrent Neural Network model of pre-training, the target category is obtained
Business revenue prediction data time series.
11. a kind of Recognition with Recurrent Neural Network model training method characterized by comprising
The business revenue data sample time sequence of target category in training set is subjected to equal part according to setting sequence length;Wherein, it draws
The data bulk in every part of business revenue data sample time sequence divided is N;
For every part of business revenue data sample time sequence of division, preceding N-1 data are input in Recognition with Recurrent Neural Network model,
Obtain the predicted value of n-th data;
The true value of n-th data in every part of business revenue data sample time sequence of division and the predicted value are input to loss
In function, penalty values are obtained, and adjust the network parameter of the Recognition with Recurrent Neural Network model based on the penalty values.
12. a kind of Recognition with Recurrent Neural Network model training method characterized by comprising
The business revenue data sample time sequence of target category in training set is subjected to equal part according to setting sequence length;What is divided is every
Data bulk in part business revenue data sample time sequence is M;
For every part of business revenue data sample time sequence of division, preceding M/2 data are input in Recognition with Recurrent Neural Network model,
Obtain the predicted value of rear M/2 data;
By the prediction of the true value of M/2 data rear in every part of business revenue data sample time sequence of division and rear M/2 data
Value is input in loss function, obtains penalty values, and the network of the Recognition with Recurrent Neural Network model is adjusted based on the penalty values
Parameter.
13. a kind of equipment characterized by comprising
One or more processors;
Storage device, for storing one or more programs,
When one or more of programs are executed by one or more of processors, so that one or more of processors are real
Now as any one of claim 1-7 perhaps business revenue prediction technique described in any one of claim 10 perhaps claim 8 or 9 or
A kind of Recognition with Recurrent Neural Network model training method described in claim 11 or 12.
14. a kind of computer readable storage medium, is stored thereon with computer program, which is characterized in that the program is by processor
Any one of such as claim 1-7 perhaps business revenue prediction technique or claim 8 described in any one of claim 10 are realized when execution
Or a kind of Recognition with Recurrent Neural Network model training method described in 9 or claim 11 or 12.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910303761.7A CN110020896A (en) | 2019-04-16 | 2019-04-16 | A kind of business revenue prediction, model training method, device, equipment and storage medium |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910303761.7A CN110020896A (en) | 2019-04-16 | 2019-04-16 | A kind of business revenue prediction, model training method, device, equipment and storage medium |
Publications (1)
Publication Number | Publication Date |
---|---|
CN110020896A true CN110020896A (en) | 2019-07-16 |
Family
ID=67191576
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201910303761.7A Pending CN110020896A (en) | 2019-04-16 | 2019-04-16 | A kind of business revenue prediction, model training method, device, equipment and storage medium |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN110020896A (en) |
Cited By (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN111210056A (en) * | 2019-12-25 | 2020-05-29 | 深圳供电局有限公司 | Electricity price scheme determination method and device, computer equipment and storage medium |
CN113537631A (en) * | 2021-08-04 | 2021-10-22 | 北方健康医疗大数据科技有限公司 | Method and device for predicting medicine demand, electronic equipment and storage medium |
CN114363193A (en) * | 2022-01-04 | 2022-04-15 | 北京达佳互联信息技术有限公司 | Method and device for training resource prediction model and method and device for resource prediction |
-
2019
- 2019-04-16 CN CN201910303761.7A patent/CN110020896A/en active Pending
Cited By (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN111210056A (en) * | 2019-12-25 | 2020-05-29 | 深圳供电局有限公司 | Electricity price scheme determination method and device, computer equipment and storage medium |
CN113537631A (en) * | 2021-08-04 | 2021-10-22 | 北方健康医疗大数据科技有限公司 | Method and device for predicting medicine demand, electronic equipment and storage medium |
CN113537631B (en) * | 2021-08-04 | 2023-11-10 | 北方健康医疗大数据科技有限公司 | Medicine demand prediction method, device, electronic equipment and storage medium |
CN114363193A (en) * | 2022-01-04 | 2022-04-15 | 北京达佳互联信息技术有限公司 | Method and device for training resource prediction model and method and device for resource prediction |
CN114363193B (en) * | 2022-01-04 | 2024-01-09 | 北京达佳互联信息技术有限公司 | Training method and device of resource prediction model, and resource prediction method and device |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
US20210049627A1 (en) | System and method for evaluating and optimizing media content | |
US11361342B2 (en) | Methods and apparatus to incorporate saturation effects into marketing mix models | |
CN109961142B (en) | Neural network optimization method and device based on meta learning | |
CN110263244A (en) | Content recommendation method, device, storage medium and computer equipment | |
CN110020896A (en) | A kind of business revenue prediction, model training method, device, equipment and storage medium | |
US11462011B2 (en) | System and method for augmenting casted content with augmented reality content | |
CN108334575A (en) | A kind of recommendation results sequence modification method and device, electronic equipment | |
KR101926404B1 (en) | Promotional image improvement apparatus and method in online shopping mall using artificial intelligence | |
KR101926403B1 (en) | Promotion Performance Prediction and Recommendation Apparatus in Online Shopping Mall Using Artificial Intelligence | |
CN104992347A (en) | Video matching advertisement method and device | |
CN109544197A (en) | A kind of customer churn prediction technique and device | |
US20210241293A1 (en) | Apparatuses, computer-implemented methods, and computer program products for improved model-based determinations | |
CN109697636A (en) | Merchant recommendation method, merchant recommendation device, electronic equipment and medium | |
CN105354056B (en) | APP recommendation information display method and device | |
KR102163412B1 (en) | Method for providing business establishing information report | |
CN113570395A (en) | Information processing method and device, computer readable medium and electronic equipment | |
CN111798280A (en) | Multimedia information recommendation method, device and equipment and storage medium | |
CN113918826B (en) | Processing method of release information, and training method and device of resource prediction model | |
CN109087124A (en) | A kind of application program Value Prediction Methods and device | |
CN106533831A (en) | Method and system for determining bandwidth usage of user | |
KR102149728B1 (en) | Method and system for extracting user-centered design guides of products through artificial intelligence | |
US20160217480A1 (en) | Agent-based demand prediction system for generating weather-dependent product demand predictions | |
CN110533485A (en) | A kind of method, apparatus of object select, storage medium and electronic equipment | |
CN113536103B (en) | Information recommendation method and device, electronic equipment and storage medium | |
Le Cadre et al. | Consumers’ preference modeling to price bundle offers in the telecommunications industry: A game with competition among operators |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
RJ01 | Rejection of invention patent application after publication |
Application publication date: 20190716 |
|
RJ01 | Rejection of invention patent application after publication |