CN107770365A - A kind of prediction of multiple features APP streams and recommendation method based on Android - Google Patents
A kind of prediction of multiple features APP streams and recommendation method based on Android Download PDFInfo
- Publication number
- CN107770365A CN107770365A CN201710762814.2A CN201710762814A CN107770365A CN 107770365 A CN107770365 A CN 107770365A CN 201710762814 A CN201710762814 A CN 201710762814A CN 107770365 A CN107770365 A CN 107770365A
- Authority
- CN
- China
- Prior art keywords
- app
- prediction
- item
- abs
- current
- 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
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04M—TELEPHONIC COMMUNICATION
- H04M1/00—Substation equipment, e.g. for use by subscribers
- H04M1/72—Mobile telephones; Cordless telephones, i.e. devices for establishing wireless links to base stations without route selection
- H04M1/724—User interfaces specially adapted for cordless or mobile telephones
- H04M1/72403—User interfaces specially adapted for cordless or mobile telephones with means for local support of applications that increase the functionality
- H04M1/72406—User interfaces specially adapted for cordless or mobile telephones with means for local support of applications that increase the functionality by software upgrading or downloading
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04L—TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
- H04L67/00—Network arrangements or protocols for supporting network services or applications
- H04L67/34—Network arrangements or protocols for supporting network services or applications involving the movement of software or configuration parameters
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04M—TELEPHONIC COMMUNICATION
- H04M1/00—Substation equipment, e.g. for use by subscribers
- H04M1/72—Mobile telephones; Cordless telephones, i.e. devices for establishing wireless links to base stations without route selection
- H04M1/724—User interfaces specially adapted for cordless or mobile telephones
- H04M1/72448—User interfaces specially adapted for cordless or mobile telephones with means for adapting the functionality of the device according to specific conditions
- H04M1/72451—User interfaces specially adapted for cordless or mobile telephones with means for adapting the functionality of the device according to specific conditions according to schedules, e.g. using calendar applications
Abstract
The invention discloses a kind of multiple features APP streams prediction based on Android and recommend method, be stored with a plurality of instruction, suitable for the loading of multiple application APPs connected each other realized on a handheld device and be performed in unison with, including：Step 1: structure APP abstract models；Step 2: APP relevant minings, excavate current application AppiCooperation relation between being flowed APP and with previous apply Appi‑1Between existing competitive relation, be divided into cooperation relation calculate and competitive relation calculate；Step 3: realizing that predicted characteristics select, by used APP in the short time and its using being sequentially abstracted, the prediction based on application stream current state, including the selection of APP correlative characters, contextual feature selection and application stream current state feature selecting are realized；Step 4: collaboration Nowcasting and Forecasting realizes the prediction of APP streams and recommended.Compared with prior art, the present invention can effectively lift the accuracy recommended for APP streams, more meet the actual use needs of user, also have positive role to improving the APP cold start-up problems in predicting.
Description
Technical Field
The invention relates to the technical field of Android algorithm application and mobile computing, in particular to a multi-feature APP flow prediction and recommendation method.
Background
The openness of the Android system enables different application developers to conveniently develop mobile application APP. However, the APP is rapidly developed to present an obvious fragmentation characteristic, a large number of APPs in an application market are only concentrated on specific function requirements, user requirements are complex and various, the user requirements exceed a service boundary which can be provided by one APP, the user requirements need to be met by combining, cooperating, competing and other modes in a coordinated mode, and APP stream recommendation enables automatic multiple APP collaborative provision services to be possible.
At present, many researchers have conducted intensive research in the field of user behavior or APP prediction and recommendation, and through a large amount of research and analysis of historical data such as user logs, application use habits of users are mined and modeled, and APP prediction and recommendation based on context are provided. The method comprises the steps of performing collaborative filtering algorithm or constructing respective matrixes of a user and an APP for recommendation, or representing historical use records into a form of < time, place, behavior and APP > quadruple, calculating the historical record closest to the current moment based on the quadruple, recommending the APP used by the historical record, judging the user behavior by adopting a Bayesian network, and performing discretization processing on the time, the place and the behavior for calculating similarity (Journal of Systems Architecture,2014,60 (8): 702-710); or expressing the historical usage records as an "event-condition-action (ECA)" rule, wherein ECAs respectively express a series of user operations, context environment Information and specific APP usage, and selecting an ECA record with the highest similarity to the current time condition for APP recommendation (Information Sciences,2014, 277; or the sensor data is used as an explicit feature, the APP use sequence is used as an implicit feature, in order to avoid huge prediction calculation scale and time consumption caused by excessive features, a personalized feature selection algorithm is designed, and the highest prediction precision (IEEE, 2013. In addition, in the aspect of prediction of APP, besides the Foresting mode, due to the fact that APP is dynamic, a collaborative Nowcasting model appears, and aiming at the characteristic of mobile data sparseness, the Nowcasting concept is introduced into the field of intention monitoring of the mobile internet (International World Wide Web Consumers Steel Committee, 2016. However, different recommendation methods have different selected characteristics and have different advantages and disadvantages on the influence on the recommendation result, so that the APP flow prediction and recommendation are cooperatively performed by combining multiple characteristics, and the user experience can be more comprehensively and truly improved.
The existing APP prediction and recommendation method cannot meet the requirement of accurately predicting the use requirement of a user, and the existing APP prediction and recommendation method has the following fundamental problems: context environment information and historical data are used as relevant features, only the accuracy of prediction and recommendation results is improved, interaction of user input layers among APPs in the APP stream is further achieved after recommendation is not researched, and the dynamic property of the APP is not considered. There is therefore a need for improvements to existing APP stream prediction and recommendation methods to more closely approximate the real usage needs of the user.
Disclosure of Invention
In order to overcome the defects in the prior art, the invention provides a multi-feature APP flow prediction and recommendation method based on Android.
The invention relates to an Android-based APP flow prediction and recommendation method, which is suitable for loading and cooperatively executing a plurality of applications APP connected with each other on a handheld device, and comprises the following steps:
step one, constructing an APP abstract model, representing one APP as a user input set ItemSet in the abstract model, and driving various functions of the APP by various different user input sets respectively; each user input Item contained in the user input set ItemSet is a triple formed by an Item-Storage part, an Item-Semantic part and an Item-Syntax part, wherein the triple comprises a < Storage, semantic and Syntax > triple; wherein Item-Storage represents a physical location of the user input Item in the local Storage; item-Semantic representation describes the meaning of the user input Item, and is used for adaptive parameter matching between different APPs; item-Syntax represents the data format of different types of content customized internally by APP; each Item supports two operations, a fetch operation and an inject operation; the Item instance represents a data instance directly obtained from the local storage inside the APP; (ii) a
Step two, APP correlation mining, namely mining the current application App i Collaboration with APP streams, and with the previous application App i-1 The competition relationship is divided into cooperative relationship calculation and competition relationship calculation:
and (3) calculating a cooperation relation: by computing arbitrary mobile application apps t Semantic similarity score with all APPs in current application stream t And obtaining the APP cooperative relationship as shown in the following formula:
wherein, abs t Representative App t The abstraction of (2); reused (item) represents a reusable relationship between a concept used when specifying semantic annotation of the user input item and semantics of all items (i.e., parameters) of all APPs in the current application flow current; the calculation formula of reusable operation Reused is as follows:
wherein S represents the set of all items in APPs with which items are compared in quantifying reusability of Item instances; s i Representing one item of all items; reused _ single (S i Item) represents S i The semantic similarity between the two parameters and the concept used by the item during semantic annotation is calculated according to the following formula:
wherein, reused _ single (p) 1 ,p 2 ) It can be represented by 5 relations, and accordingly can take 5 values. When taking 1, represents the parameter p 1 And parameter p 2 Is completely the same (same as); the parameter p is expressed when 0.8 is taken 1 And p 2 The description of (1) is semantically equivalent (equivalent to); the parameter p is expressed when 0.6 is taken 1 Is semantically attributed to parameter p 2 A semantic description of, e.g., a destination belonging to a place; when 0.5 is taken, the parameter p is expressed 2 Is attributed to the parameter p 1 The semantic description of (1); otherwise, the value is 0;
sequentially forming one-dimensional vectors by the zoops of all apps installed in the equipment to serve as a prediction characteristic;
and (3) calculating the competitive relationship: by calculating and comparing App i-1 And App t Parameter semantic similarity Scomp t The purpose of obtaining APP competition relationship and semantic similarity is to provide the data similar to App i-1 Similar services; as shown in the following formula.
Wherein, abs i-1 Representing App i-1 The abstraction of (2); abs t Representing App t The abstraction of (2); reused (App) i-1 Item) representation based on App i-1 Quantized Item reusability; app i-1 Indicating the last APP, app in the application stream just finished executing t Represents any mobile application where i-1 and t cannot take the same value; wherein Abs i-1 ∪Abs t Represents Abs t And Abs i-1 ∩Abs t Difference result and Abs i-1 Is as followsThe formula is shown in the specification.
Abs i-1 ∪Abs t ＝Abs t /(Abs i-1 ∩Abs t )∪Abs i-1
Wherein, abs i-1 ∩Abs t Represents Abs t Can reuse the Item contained therein from the App i-1 A user input value of (a);
Abs i-1 ∩Abs t ＝{item∈Abs t |Resued(App i-1 ，item)＞0}
sequentially forming one-dimensional vectors by the Scomps of all APPs as a prediction characteristic;
thirdly, realizing prediction feature selection, abstracting the APP used in a short time and the use sequence thereof, and realizing prediction based on the current state of the application flow, wherein the prediction comprises APP correlation feature selection, context feature selection and application flow current state feature selection; and selecting the APP relevance features, wherein one-dimensional vectors respectively formed by the Scoop and the Scomp of all the APPs in the step two are used as the two APP relevance features. The calculation formulas of the Scoop correlation characteristic Scoop vector and the Scomp correlation characteristic ScompVector are respectively expressed as follows:
among them, app i Representing any one of the APP, APP, in the mobile device 0 ,app 1 ,….app i-1 Representing the current APP stream;
selecting context characteristics, wherein the selected content comprises a current date, a current time, a current position and an APP used last time;
the method comprises the following steps of selecting current state characteristics of an application flow, wherein the specific calculation process is as follows: will { app } 0 ,app 1 ,…,app x ,…,app i-1 Expressed as<p 1t ,p 2t ,…,p xt ,…,p nt >，p xt Indicates at the end of app x Using APPs directly or after several APPs t The probability of (d); when the current state of the application flow is calculated, flow vector characteristics of a training set and a test set are calculated respectively; for the training set, according to { app } 0 ,app 1 ,…app i-1 Reach APP in conjunction with each APP in an AUG computing device i The probability of (d); for test set data, when predicting app i First, for each App installed in the device t Calculating the current app i ＝App t A flow vector of time, then assume P t Is app i ＝App t Is calculated by using an iterative update algorithm i Weighted flow vectors up to P t Converging;
and step four, APP flow prediction and recommendation are achieved, based on the three characteristics of the step three, nowcasting and Forecasting calculation are respectively carried out in the prediction process, and after the results of the Nowcasting and the Forecasting are combined, the Android client side recommends the results to the user in a system popup mode.
The fourth step specifically comprises the following steps:
performing APP recommendation for multiple times based on the application flow state characteristics: at the beginning of the new flow, with the first application app used by the user 0 Performing APP recommendation for the current state of the application stream by combining the current context characteristics and the APP correlation characteristics, wherein the recommendation result is the APP i (ii) a Secondly, if the user selects from the recommendation list, the selected APP is used as the APP i Updating the current state of the APP flow, and recommending the next time after the user quits from the APP; if the user ignores the recommendation popup, the application stream recommendation is finished; and, the single APP recommendation algorithm includes the following processes: carrying out APP prediction by cooperating with a Forecasting mode and a Nowcasting mode;
the method comprises the following steps of collecting user data for several weeks in a Forecasting mode, modeling by mining historical data, and predicting the demand of a user on an APP at the current moment;
the calculation process of the Nowcasting mode comprises the following steps:
construct Panel, which will predict app in the problem i Regarding each APP installed in the mobile phone as an APP for the user's intention i Selecting 5 APP usage records nearest to the prediction time, and representing the prediction characteristics and intention of the 5 records and the prediction characteristics of the current time into a time sequence to form a Panel matrix.
A composition Tensor which is used for selecting 5 continuous data with almost similar week and time from historical data of the last two weeks of a user to form a plurality of panels of different weeks, and forming a three-dimensional matrix V together with the current panels, wherein the V is expressed as a Tensor X by using an X = Tensor (V) function of a Tensor-box tool;
the method comprises the steps of CP Tensor decomposition, namely obtaining cooperative latent factor matrixes of different cycles by using a CP Tensor decomposition method, decomposing a Tensor X into R one-dimensional vectors U, V and W triples by using a CP Tensor decomposition function P = parafacc _ als (X, R) provided by a sensor-box, wherein R is set as 4;
kalman filtering, namely calculating necessary parameters A and B of simulation according to U, V and W, and then using the Kalman filtering to obtain potential factors of the current moment, wherein the Matlab of the Kalman filtering is realized;
performing least square regression, namely performing linear regression on the potential cooperative factors and the actual APP use condition in a short term by using the least square method, and realizing Matlab;
real-time prediction is carried out, after a linear relation between the potential factors and the APP intents is obtained, according to the potential factor prediction intents at the current moment, the estimation values of all the APP intents at the current moment are sequenced, and a prediction result of the current APP is obtained;
the calculation process of the Forecasting mode comprises the following steps: calculating feature similarity, namely: respectively calculating the distance between each record in the training set and the current data; adopting a standardized Euclidean distance, wherein the Euclidean distance between vectors is calculated by the current state of the APP flow in a vector form and the correlation characteristics; the context information with discrete values is endowed with a distance value of 1 or 0 according to whether the value-taking result is consistent; calculating the difference value of two moments and normalizing the difference value of the two moments for the time characteristic distance; calculating the normalized Euclidean distance of the results as the distance between the two features; selecting the training data with the nearest distance of the first 20 percent to form a RecordSet; taking APP as a Class Label, and accumulating the characteristic distance reciprocal of records with the same Class Label; sequencing from big to small to obtain a Forecasting result of the Forecasting;
mixing the results of Nowcasting and Forecasting to be used as the APP list finally recommended to the user.
The Matlab implementation process of Kalman filtering comprises the following steps: respectively an initialized posterior factor, an initialized posterior measurement error covariance matrix and an initialized and corrected personalized factor matrix; a correction factor; calculating a prior factor of the next recording moment, calculating a prior measurement error covariance matrix of the next recording moment, acquiring Kalman gain, correcting a posterior factor FTG obtained by the prior factor ftc through panel, assigning the estimated personalized factor to FTG, and calculating the posterior measurement error covariance matrix of the next recording moment.
The Matlab implementation process of the least square regression comprises the following steps: solving the average value of each row vector and the average value of the intention vector of the matrix f, initializing the row number and the column number of the matrix f, and initializing the temporary variable t 1 For calculating the denominator of the regression parameter bt, initializing the temporary variable t 2 The numerator used to calculate the regression parameter bt, and output the regression parameter for calculating the regression parameter af.
Compared with the prior art, the method for predicting and recommending the APP flow based on the Android is provided by integrating different characteristics and analyzing the high dynamic property of the APP, and the prediction effect of the APP flow is improved. The expected benefits include:
1. the method has the advantages that the correlation of user input data among different APPs is considered, the APP abstract information is the inherent attribute of the APPs, the prediction accuracy is improved in an APP stream mode, and the method has a positive effect on improving the APP cold start problem in prediction;
2. by adopting the APP flow prediction and recommendation method combining Forecasting and Nowcasting, the accuracy of APP flow recommendation can be effectively improved, and the APP flow recommendation method can better meet the actual use requirements of users.
Drawings
Fig. 1 is a schematic overall flow diagram of an APP stream prediction and recommendation method based on Android of the present invention;
FIG. 2 is a schematic diagram of an APP flow recommendation process;
fig. 3 is a schematic diagram of the calculation process of Nowcasting.
Detailed Description
Embodiments of the present invention will be described in further detail below with reference to the accompanying drawings.
The APP stream recommendation problem is described as based on the current context, current state of the application stream { APP } 0 ,app 1 ,…,app i-1 Correlation between APP and user, predicting APP most needed by user at the current moment i And recommending the prediction result to the terminal user in a good interactive mode.
As shown in fig. 1, the whole process of the APP stream prediction and recommendation method based on Android of the present invention includes the following four steps:
the first step is as follows: constructing an APP abstract model, namely: an APP concept is represented in the abstract model as a set ItemSet, which contains multiple functions, each driven by a different set of user inputs. Each Item then represents a user input within an APP (within an APP, multiple user inputs are often required to drive the execution of a certain function).
Each user input Item is composed of three parts of Item-Storage, item-Semantic and Item-Syntax, and the < Storage, semantic and Syntax > triple. Wherein Item-Storage represents a physical location of the user input Item in the local Storage; item-Semantic represents the meaning of describing the user input data, is used for adaptive parameter matching among different APPs, and considers the equivalence relation and the inclusion relation of Semantic concepts for the concept relation of the user input data among the different APPs; two concepts with equivalence have similar meanings in some scenarios, and examples of them can be shared. The inclusion relationship refers to the parent-child relationship among concepts, the child concepts belong to the parent concept, and the parent concept can reuse the instances of the child concepts; item-Syntax represents the data format of different types of content customized internally by APP.
The second step: APP relevance mining, like traditional service composition, APPs i There may be a cooperative relationship with the APP stream and also with the APP i-1 There is a competing relationship. Both relationships are represented in vector form to participate in the prediction process. The step is divided into cooperative relation calculation and competitive relation calculation.
(11) And (3) calculating a cooperation relation:
by computing arbitrary mobile application apps t Semantic similarity score with all APPs in current application stream t Obtaining the APP cooperation relationship as shown in the following formula:
wherein, abs t Representative App t The abstraction of (2); reused (Item) represents the reusable relationship between the concept used when specifying semantic annotation of the Item by the user and the semantics of all items (i.e. parameters) of all APPs in the current application flow currentflow, used to quantify the reusability of the Item instance. In fact, scoop represents the average probability of reuse of an Item instance.
The calculation formula for the reusable operation Reused is as follows:
wherein S represents the set of all items in APPs with which items are compared in quantifying reusability of Item instances; s. the i Representing one item of all items; used _ single (S) i Item) represents S i The semantic similarity between the two parameters and the concept used by the item during semantic annotation is calculated according to the following formula:
wherein, reused _ single (p) 1 ,p 2 ) It can be represented by 5 relations, and accordingly can take 5 values. When taking 1, represents the parameter p 1 And parameter p 2 Is completely the same (same as); the parameter p is expressed when 0.8 is taken 1 And p 2 The description of (1) is semantically equivalent (equivalent to); the parameter p is expressed when 0.6 is taken 1 Is semantically attributed to a parameter p 2 A semantic description of, e.g., a destination belonging to a place; the parameter p is expressed when 0.5 is taken 2 Is attributed to the parameter p 1 The semantic description of (1); otherwise, the value is 0.
And sequentially forming one-dimensional vectors by the zoops of all apps in the equipment as a prediction characteristic.
(12) Contention computation
By computing comparison App i-1 And App t Parameter semantic similarity Scomp t The purpose of obtaining APP competition relation and semantic similarity is to provide the application i-1 Similar services; as shown in the following formula.
Wherein, abs i-1 Representing App i-1 The abstraction of (2); abs t Representing App t The abstraction of (2); reused (App) i-1 Item) representation based on App i-1 Quantized Item reusability; app i-1 Indicating the last APP, app in the application stream just finished executing t Represents any mobile application where i-1 and t cannot take the same value; wherein Abs i-1 ∪Abs t Represents Abs t And Abs i-1 ∩Abs t Difference set results are then compared with Abs i-1 The union of (1) is shown in the following formula.
Abs i-1 ∪Abs t ＝Abs t /(Abs i-1 ∩Abs t )∪Abs i-1
Wherein, abs i-1 ∩Abs t Represents Abs t Can reuse the Item from the App i-1 A user input value of (a);
Abs i-1 ∩Abs t ＝{item∈Abs t |Resued(App i-1 ，item)＞0}
and sequentially forming one-dimensional vectors by the Scomps of all Apps as a prediction feature.
The third step: selecting prediction characteristics of three categories, including APP correlation characteristic selection, context characteristic selection and application flow current state characteristic selection;
(31) And selecting the APP relevance features, wherein the one-dimensional vectors of the Scoop and the Scomp of all the APPs in the second step are used as two kinds of APP relevance features. The calculation formulas of the Scoop correlation characteristic Scoop vector and the Scomp correlation characteristic ScompVector are respectively expressed as follows:
wherein, app i Representing any one of the APPs, APPs in the mobile device 0 ，app 1 ，....app i-1 Representing the current APP stream;
and the context feature contains important environmental information. Many hardware sensors are available in mobile devices to convey the user's real-world environment information. The method is also a problem that much research work is dedicated to solve by exploring the importance degree of various types of context data in prediction and acquiring higher prediction accuracy by using less context data as far as possible. Studies have shown that spatio-temporal distribution information plays a crucial role in APP prediction. Thus, the context features employed in the prediction by the present invention are: current date, current time, current location and the last APP used.
(33) The current state characteristic of the application flow comprises the following specific calculation processes: will { app 0 ，app 1 ，…，app x ，…，app i-1 Expressed as<p 1t ，p 2t ，…，p xt ，…，p nt >, wherein p xt Indicating that app is ending x Using APPs directly or after several APPs t The probability of (c). The invention refers to the calculation method of implicit characteristics in the existing research, extracts data in a history log and constructs an APP Usage transformation graph (AUG) to describe the transformation relation of the APP on use. In the AUG, the nodes are APPs, the directed edges connecting the nodes represent the switching directions of the APPs, and the weight of the edges is a function of the starting time interval of the two APPs and represents the transition probability. And when the current state of the application flow is calculated, calculating the flow vector characteristics of the training set and the test set respectively.
App for training set i As known, only need to be according to { app 0 ,app 1 ,…app i-1 Reach APPs in conjunction with each APP in an AUG computing device i The probability of (2) is just required.
For test set data, when app is predicted i First, for each App installed in the device t Calculate current app i ＝App t A flow vector of time, then assume P t Representing apps i ＝APP t Is calculated by using an iterative update algorithm i Weighted flow vectors up to P t And (6) converging.
Because each App installed on the mobile phone is taken as a candidate App when the prediction is started i Thus P is t Is that all apps have equal probability. Considering that APPs with similar semantic parameters are more likely to appear in the same application stream, the invention normalizes the score representing the reusability of the current stream parameters as the initial P t The iterative update calculation is carried out instead of the average starting probability, so that the P can be accelerated t And converging and reducing the iteration times. Updated P t The final result is not affected because the AUG is not changed.
The fourth step: the method comprises the following steps of APP flow recommendation, wherein the traditional APP flow recommendation is performed for one time after user environment characteristics are evaluated, and when the complex requirements of a user are responded, the method cannot meet the requirement of multi-APP cooperative flow execution. The present invention performs multiple APP recommendations based on application stream state characteristics, as shown in fig. 2, which is an APP stream recommendation process for multiple APP recommendations.
First, a new process is initiated with the first application app used by the user 0 Performing APP recommendation for the current state of the application stream by combining the current context characteristics and the APP correlation characteristics, wherein the recommendation result is the APP i (ii) a Secondly, if the user selects from the recommendation list, the selected APP is used as the APP i Updating the current state of the APP flow, and recommending the next time after the user quits from the APP; and if the user ignores the recommendation popup, ending the application stream recommendation.
The single APP recommendation algorithm is as follows:
the existing APP recommendation modes can be regarded as a Forecasting mode, user data of several weeks generally need to be collected, modeling is carried out by mining historical data, and the demands of users on APPs at the current moment are recommended. However, this approach does not fully accommodate the dynamics of the handset application. Compared with Forecasting, the Nowcasting mode widely used in meteorology and economics has better utilization on real-time data, so that the invention combines the Nowcasting and the Forecasting to realize the recommendation of APP.
(41)Nowcasting
The invention applies a synergistic Nowcasting model to the APP prediction problem. The calculation process of Nowcasting is shown in fig. 3.
1) Construct Panel, which will predict app in the problem i The user's intention is considered, and each APP installed in the mobile phone is an APP i Each APP in the device is thus a likely intention of the user at the time of the predicted occurrence, and no more than one APP intention will occur at the same time. Selecting 5 APP use records nearest to the prediction time, and representing the content (prediction characteristics and intention) of the 5 records and the prediction characteristics of the current time into a time sequence to form a Panel matrix. Rules such asThe following:
the APP correlation characteristics and the current state of the APP flow are in a vector form, and each vector element of the APP correlation characteristics and the current state of the APP flow form a line sequence in time sequence; each discretized value of the place label in the context feature forms a line sequence in time order; the possible values of LastAPP in the context characteristics are all APPs in the mobile phone, and each APP forms a line sequence according to the time sequence; each APP is also an intended valid value, so each APP again constitutes a time series as a possible intention.
TABLE 2 Panel example
| TimeSTEP | 10:30 | 11:00 | 11:30 | 12:00 | Now12:20 |
| OFFICE | 1 | 0 | 0 | 0 | 0 |
| HOME | 0 | 0 | 0 | 1 | 1 |
| FLOW-FEA(APP x ) | 0 | 0.3 | 0.5 | 0.2 | 0.3 |
| FLOW-FEA(APP y ) | 0 | 0 | 0 | 0.6 | 0 |
| LastAPP(APP 1 ) | 0 | 0 | 0 | 1 | 1 |
| LastAPP(APP 2 ) | 1 | 0 | 0 | 0 | 0 |
| APP1 | 0 | 0 | 0 | 1 | ？？ |
| APP2 | 1 | 0 | 0 | 0 | ？？ |
As shown in table 2, is an example of Panel. Wherein "OFFICE" and "HOME" represent contextual location tags; "LastAPP" represents the APP that was used last time in the context, and its discretization value is all APPs installed in the device; FLOW-FEA represents the current state of the APP stream, and each value forms a time sequence, and the APP correlation characteristics in a vector form are similar; "APP 1 ”、“APP 2 "indicates intention, and the intention value at the current time is calculated by Nowcasting.
2) The composition tensor and a large number of researches show that the use of APP presents periodicity with a period of one week, so that in the same existence of Panel matrix sparseness problem, firstly, 5 pieces of continuous data with approximate week and time are respectively selected from historical data of the last two weeks of a user to form a plurality of Panel with different weeks, and the Panel and the current Panel are combined into a three-dimensional matrixWhere N represents a line in Panel, i.e., a predicted feature and intent; t represents a column in Panel, i.e., a different time instant; m represents different weeks. V is expressed as Tensor X using the X = Tensor (V) function of the sensor-box tool.
3) CP tensor decomposition, and obtaining cooperative latent factor matrixes of different cycles by using CP tensor decomposition methodWherein R represents the number of factors in PanelT denotes different times in Panel, and the Tensor X is decomposed into R one-dimensional vector U, V, W triplets, with R set to 4, using CP Tensor decomposition function P = parafacc _ als (X, R) provided by the sensor-box.
4) And (3) calculating necessary parameters A and B of simulation according to U, V and W due to insufficient collaborative latent factors obtained in CP tensor decomposition in the step (3), and then obtaining the latent factors at the current moment by using Kalman filtering, wherein the Matlab of the Kalman filtering is realized as follows.
And lines 1-3 are an initialized posterior factor, an initialized posterior measurement error covariance matrix and an initialized and corrected personalized factor matrix respectively.
4-11 act as a correction factor. 5, calculating a prior factor of the next recording time, 6, obtaining Kalman gain, 8, obtaining a posterior factor FTG by correcting the prior factor ftc through panel, 9, assigning the estimated personalized factor to FTG, and 10, calculating the posterior measure error covariance matrix of the next recording time.
5) And (3) performing least square regression, namely performing linear regression on the potential cooperative factors and the actual APP use condition in a short term by using the least square method, wherein Matlab is realized as follows.
1, calculating the average value of each row vector of a row matrix f, 2, calculating the average value of a row intention vector, 3, rows and columns of an initialization matrix f, 4, 5, 6-10, 11, and 12 rows of output regression parameters, wherein the initialization temporary variable t1 is used for calculating the denominator of the regression parameter bt, the initialization temporary variable t2 is used for calculating the numerator of the regression parameter bt, and the initialization temporary variable t2 is used for calculating the regression parameter bt.
6) And (3) predicting in real time, and after a linear relation between the potential factors and the APP intents is obtained, sequencing the estimated value of each APP intention at the current moment according to the potential factor prediction intention at the current moment to obtain a prediction result of the current APP.
(42) Forecasting, in the Forecasting part, the invention uses APP itself as Class Label, selects simpler k-nearest neighbor (KNN) classification algorithm to construct a simple bipartite graph of user historical use data, wherein k =20%, namely selects the data record with the top 20% of the graph (data set) closest to the current prediction characteristic. The calculation process of Forecasting described above is as follows.
Calculating the distance between each record in the training set and the current data; the method comprises the steps of adopting a standardized Euclidean distance, wherein the Euclidean distance between vectors is calculated by the characteristics (the current state and the correlation characteristics of the APP flow) in the form of vectors; context information (week, APP used at the last time and place) with discrete values is endowed with a distance value of 1 or 0 according to whether the value-taking result is consistent; for the time characteristic distance, the difference between the two time instants is calculated and normalized. Calculating the normalized Euclidean distance of the results as the distance between the two features; selecting the training data with the nearest distance of the first 20 percent to form a RecordSet; taking APP as a Class Label, and accumulating the reciprocal of the characteristic distance of records with the same Class Label; and sequencing from big to small to obtain the prediction result of Forecasting.
Forecasting represents an APP prediction and recommendation mode based on context environment information and historical data; nowcasting indicates that, on the basis of context information and historical data, a correlation between the current state of an APP stream and an APP is taken as a prediction factor, and an APP (i.e., an intention) to be used by a user in the near future is predicted.
Mixing the results of Nowcasting and Forecasting as the list of APPs ultimately recommended to the user. By experiment, nowcasting shows better prediction accuracy. Therefore, when the Nowcasting result and the Forecasting result are synthesized, the APP sequencing is carried out by taking the Nowcasting result as a main result and taking the Forecasting result as an auxiliary result.
Since abstract information is an inherent property of APP, APP correlation is not affected by the APP cold start problem, which also ensures that prediction of APP flow will not fail completely even if sufficient amount of historical data is not collected.
The prediction mainly occurs when a user wants to obtain service through equipment, and the method specifically designs three moments, namely when the equipment is started, the equipment is unlocked or a screen is lightened, and when an APP is exited to a system main interface. The Android system can send out corresponding Broadcast events when starting up the equipment, unlocking the equipment and lighting a screen, and only needs to register corresponding Broadcast receivers to receive and process broadcasts.
The fifth step: based on APP flow prediction and recommendation result evaluation of Android, the method adopts Average Recall ratio Average Recall and normalized breakage cumulative gain nDCG to evaluate the prediction result.
The average recall rate is calculated as shown below,
wherein C represents a set of test cases, I (Rec) i ·App i ) The function is represented in a recommendation list Rec i Hit true result App i Get 1 if not 0. Wherein C is i Is a certain test case in C.
The normalized break-up cumulative gain nDCG is calculated as shown below,
wherein C represents a test case set; p represents the sequencing position of the hit result in the recommendation list, and nDCG is 0 when no hit occurs; c i Is a certain test case in C.
The Android-based APP flow prediction and recommendation method disclosed by the invention shows that the Nowcasting mode is better than the Forecasting mode, and through manual comparison and analysis, when the APP in the equipment is changed, such as new installation or unloading, the Forecasting result is obviously lower than the Nowcasting result. This indicates that Nowcasting can better adapt to the dynamic of the mobile end, and is more suitable for APP prediction of the mobile end. Due to the fact that the mobile phone screen is limited, the method recommends 6 APPs for the user, and the recall rate can reach 0.78.
The correlation between the APPs has an obvious effect on improving the analysis and prediction effect of a user input data layer, and the APP correlation is an internal attribute irrelevant to the use record and can optimize the cold start problem of the APPs in prediction.
Therefore, the problem that a user frequently opens different APPs due to fragmentation of the mobile application is solved by reusing user input data with high correlation among the APPs through prediction recommendation.
Claims (4)
1. A multi-feature APP flow prediction and recommendation method based on Android stores a plurality of instructions and is suitable for loading and cooperatively executing a plurality of Applications (APP) which are connected with each other and are realized on a handheld device, and the method is characterized by comprising the following steps:
step one, constructing an APP abstract model, representing one APP as a user input set ItemSet in the abstract model, and driving various functions of the APP by various different user input sets respectively; each user input Item contained in the user input set ItemSet is a triple formed by an Item-Storage part, an Item-Semantic part and an Item-Syntax part, wherein the triple comprises a < Storage, semantic and Syntax > triple; wherein Item-Storage represents a physical location of the user input Item in the local Storage; item-Semantic representation describes the meaning of the user input Item, and is used for adaptive parameter matching among different APPs; item-Syntax represents the data format of different types of content customized internally by the APP; each Item supports two operations, a fetch operation and an inject operation; the Item instance represents a data instance directly obtained from the local storage inside the APP; (ii) a
Step two, APP correlation mining, namely mining the current application App i Collaboration with APP streams, and with the previous application App i-1 The competition relationship is divided into cooperative relationship calculation and competition relationship calculation:
and (3) calculating a cooperation relation: by computing arbitrary mobile application apps t Semantic similarity score with all APPs in current application stream t And obtaining the APP cooperative relationship as shown in the following formula:
wherein, abs t Representative App t The abstraction of (2); reused (item) represents a reusable relationship between a concept used when specifying semantic annotation of the user input item and semantics of all items (i.e., parameters) of all APPs in the current application flow current; the calculation formula of reusable operation Reused is as follows:
wherein S represents the set of all items in APPs with which items are compared in quantifying reusability of Item instances; s i Representing one item of all items; reused _ single (S) i Item) representationS i The semantic similarity between the concepts used by the two parameters and the item during semantic annotation is calculated according to the following formula:
wherein, reused _ single (p) 1 ,p 2 ) It can be represented by 5 relations, and accordingly can take 5 values. Representing the parameter p when taking 1 1 And parameter p 2 Is completely the same (same as); when 0.8 is taken, the parameter p is expressed 1 And p 2 Is semantically equivalent (equivalent to); the parameter p is expressed when 0.6 is taken 1 Is semantically attributed to a parameter p 2 A semantic description of, e.g., a destination belonging to a place; the parameter p is expressed when 0.5 is taken 2 Is attributed to the parameter p 1 The semantic description of (c); the value of other conditions is 0;
sequentially forming one-dimensional vectors by the zoops of all apps installed in the equipment to serve as a prediction characteristic;
calculating the competitive relationship: by calculating and comparing App i-1 And App t Parameter semantic similarity Scomp t The purpose of obtaining APP competition relationship and semantic similarity is to provide the data similar to App i-1 Similar services; as shown in the following formula.
Wherein, abs i-1 Representing App i-1 The abstraction of (2); abs t Representing App t The abstraction of (2); reused (App) i-1 Item) representation based on App i-1 Quantized Item reusability; app i -1 denotes the last APP, APP in the application stream just finished executing t Represents any mobile application where i-1 and t cannot take the same value; wherein Abs i-1 ∪Abs t Represents Abs t And Abs i-1 ∩Abs t Difference set results are then compared with Abs i-1 The union of (1) is shown in the following formula.
Abs i-1 ∪Abs t ＝Abs t /(Abs i-1 ∩Abs t )∪Abs i-1
Wherein, abs i-1 ∩Abs t Represents Abs t Can reuse the Item contained therein from the App i-1 A user input value of (a);
sequentially forming one-dimensional vectors by the Scomps of all the APPs as a prediction characteristic;
thirdly, realizing prediction feature selection, abstracting the APP used in a short time and the use sequence thereof, and realizing prediction based on the current state of the application flow, wherein the prediction comprises APP correlation feature selection, context feature selection and application flow current state feature selection; and selecting the APP correlation characteristics, wherein one-dimensional vectors respectively formed by the Scoop and the Scomp of all the applications of the APP in the step two are used as the two APP correlation characteristics. The calculation formulas of the Scoop correlation characteristic Scoop vector and the Scomp correlation characteristic ScompVector are respectively expressed as follows:
among them, app i Representing any one of the APP, APP, in the mobile device 0 ,app 1 ,….app i-1 Representing the current APP stream;
selecting context characteristics, wherein the selected content comprises a current date, a current time, a current position and an APP used last time;
the method comprises the following steps of selecting current state characteristics of an application flow, wherein the specific calculation process is as follows: will { app } 0 ,app 1 ,…,app x ,…,app i-1 Expressed as<p 1t ,p 2t ,…,p xt ,…,p nt >，p xt Indicating that app is ending x Using APPs directly or after several APPs t The probability of (d); when the current state of the application flow is calculated, flow vector characteristics of a training set and a test set are calculated respectively; for the training set, according to { app } 0 ,app 1 ,…app i-1 Reach APPs in conjunction with each APP in an AUG computing device i The probability of (d); for test set data, when predicting app i First, for each App installed in the device t Calculate current app i ＝App t A flow vector of time, then assume P t Is app i ＝App t Calculating the app by using an iterative update algorithm i Weighted flow vectors up to P t Converging;
and fourthly, forecasting and recommending the APP flow, respectively performing Nowcasting and Forecasting calculation in the Forecasting process based on the three characteristics of the third step, combining the results of the Nowcasting and the Forecasting, and recommending the result to the user by the Android client side in a system popup mode.
2. The Android-based multi-feature APP stream prediction and recommendation method of claim 1, wherein the fourth step specifically includes the following steps:
performing APP recommendation for multiple times based on the application flow state characteristics: at the beginning of the new flow, with the first application app used by the user 0 Performing APP recommendation for the current state of the application stream by combining the current context characteristics and the APP correlation characteristics, wherein the recommendation result is the APP i (ii) a Secondly, if the user selects from the recommendation list, the selected APP is used as the APP i Updating the current state of the APP flow, and recommending the next time after the user quits from the APP; if the user ignores the recommendation popup, the application stream recommendation is finished; and, the single APP recommendation algorithm includes the following processes: carrying out APP prediction by cooperating with a Forecasting mode and a Nowcasting mode;
the method comprises the steps of collecting user data of several weeks in a Forecasting mode, and Forecasting the demand of a user on an APP at the current moment by mining historical data for modeling;
the calculation process of the Nowcasting mode comprises the following steps:
construct Panel, which will predict app in the problem i Regarding each APP installed in the mobile phone as an APP, which is the intention of the user i Selecting 5 APP usage records nearest to the prediction time, and representing the prediction characteristics and intention of the 5 records and the prediction characteristics of the current time into a time sequence to form a Panel matrix.
A composition Tensor which is used for selecting 5 continuous data with almost similar week and time from historical data of the last two weeks of a user to form a plurality of panels of different weeks, and forming a three-dimensional matrix V together with the current panels, wherein the V is expressed as a Tensor X by using an X = Tensor (V) function of a Tensor-box tool;
the method comprises the steps of CP Tensor decomposition, namely obtaining cooperative latent factor matrixes of different weeks by using a CP Tensor decomposition method, decomposing a Tensor X into R one-dimensional vectors U, V and W triples by using a CP Tensor decomposition function P = parafacc _ als (X, R) provided by a sensor-box, wherein R is set as 4;
kalman filtering, namely calculating necessary parameters A and B for simulation according to U, V and W, and then obtaining potential factors at the current moment by using Kalman filtering, wherein Matlab of Kalman filtering is realized;
performing least square regression, namely performing linear regression on the potential cooperative factors and actual APP use conditions in a short term by using the least square method, and realizing Matlab;
real-time prediction is carried out, after a linear relation between the potential factors and the APP intents is obtained, according to the potential factor prediction intents at the current moment, the estimation values of all the APP intents at the current moment are sequenced, and a prediction result of the current APP is obtained;
the calculation process of the Forecasting mode comprises the following steps: calculating feature similarity, namely: respectively calculating the distance between each record in the training set and the current data; adopting a standardized Euclidean distance, wherein the Euclidean distance between the vectors is calculated by the current state and the correlation characteristics of the APP flow in the form of the vectors; the context information with discrete values is endowed with a distance value of 1 or 0 according to whether the value-taking result is consistent; calculating the difference value of two moments and normalizing the difference value of the two moments for the time characteristic distance; calculating the normalized Euclidean distance of the results as the distance between the two features; selecting the training data with the nearest distance of the first 20 percent to form a RecordSet; taking APP as a Class Label, and accumulating the reciprocal of the characteristic distance of records with the same Class Label; sequencing from big to small to obtain a forecast result of Forecasting;
mixing the results of Nowcasting and Forecasting as the list of APPs ultimately recommended to the user.
3. The Android-based multi-feature APP stream prediction and recommendation method of claim 1, wherein the Matlab implementation process of Kalman filtering comprises: respectively an initialized posterior factor, an initialized posterior measurement error covariance matrix and an initialized and corrected personalized factor matrix; a correction factor; calculating a prior factor of the next recording moment, calculating a prior measurement error covariance matrix of the next recording moment, acquiring Kalman gain, correcting a posterior factor FTG obtained by the prior factor ftc through panel, assigning the estimated personalized factor to FTG, and calculating the posterior measurement error covariance matrix of the next recording moment.
4. The Android-based multi-feature APP flow prediction and recommendation method of claim 1, wherein the Matlab implementation process of least squares regression comprises: solving the average value of each row vector and the average value of the intention vector of the matrix f, initializing the row number and the column number of the matrix f, and initializing the temporary variable t 1 For calculating the denominator of the regression parameter bt, initializing the temporary variable t 2 The numerator used to calculate the regression parameter bt, and output the regression parameter for calculating the regression parameter af.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710762814.2A CN107770365A (en) | 2017-08-30 | 2017-08-30 | A kind of prediction of multiple features APP streams and recommendation method based on Android |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710762814.2A CN107770365A (en) | 2017-08-30 | 2017-08-30 | A kind of prediction of multiple features APP streams and recommendation method based on Android |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN107770365A true CN107770365A (en) | 2018-03-06 |
Family
ID=61265321
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201710762814.2A Pending CN107770365A (en) | 2017-08-30 | 2017-08-30 | A kind of prediction of multiple features APP streams and recommendation method based on Android |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN107770365A (en) |
Cited By (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108536572A (en) * | 2018-04-16 | 2018-09-14 | 浙江大学 | Smart mobile phone App based on AppUsage2Vec models uses prediction technique |
| CN109146602A (en) * | 2018-06-29 | 2019-01-04 | 康美药业股份有限公司 | A kind of medicine selling machine drugs supply method and automatic medicine selling machine based on user behavior |
| CN110119465A (en) * | 2019-05-17 | 2019-08-13 | 哈尔滨工业大学 | Merge the mobile phone application user preferences search method of LFM latent factor and SVD |
| CN111309341A (en) * | 2020-02-17 | 2020-06-19 | 中南大学 | Android application installation flow optimization method based on time-consuming prediction |
| CN111372221A (en) * | 2019-12-26 | 2020-07-03 | 北京蓦然认知科技有限公司 | Generation and sharing method and device of general APP ecological graph |
| CN111753145A (en) * | 2020-06-10 | 2020-10-09 | 西北工业大学 | Mobile application use prediction method based on time sequence mode |
| CN112328935A (en) * | 2020-10-29 | 2021-02-05 | 胡培培 | Application program pushing system and method based on big data |
| US11869015B1 (en) | 2022-12-09 | 2024-01-09 | Northern Trust Corporation | Computing technologies for benchmarking |
Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN104820843A (en) * | 2015-05-29 | 2015-08-05 | 常熟苏大低碳应用技术研究院有限公司 | Method for marking picture semantics based on Gauss mixture model |
| CN105740430A (en) * | 2016-01-29 | 2016-07-06 | 大连理工大学 | Personalized recommendation method with socialization information fused |
-
2017
- 2017-08-30 CN CN201710762814.2A patent/CN107770365A/en active Pending
Patent Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN104820843A (en) * | 2015-05-29 | 2015-08-05 | 常熟苏大低碳应用技术研究院有限公司 | Method for marking picture semantics based on Gauss mixture model |
| CN105740430A (en) * | 2016-01-29 | 2016-07-06 | 大连理工大学 | Personalized recommendation method with socialization information fused |
Non-Patent Citations (1)
| Title |
|---|
| 林美辰: "基于Android的应用自动协同框架研究", 《天津大学硕士学位论文》 * |
Cited By (11)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108536572A (en) * | 2018-04-16 | 2018-09-14 | 浙江大学 | Smart mobile phone App based on AppUsage2Vec models uses prediction technique |
| CN108536572B (en) * | 2018-04-16 | 2020-05-12 | 浙江大学 | Smart phone App use prediction method based on ApUage 2Vec model |
| CN109146602A (en) * | 2018-06-29 | 2019-01-04 | 康美药业股份有限公司 | A kind of medicine selling machine drugs supply method and automatic medicine selling machine based on user behavior |
| CN110119465A (en) * | 2019-05-17 | 2019-08-13 | 哈尔滨工业大学 | Merge the mobile phone application user preferences search method of LFM latent factor and SVD |
| CN110119465B (en) * | 2019-05-17 | 2023-06-13 | 哈尔滨工业大学 | Mobile phone application user preference retrieval method integrating LFM potential factors and SVD |
| CN111372221A (en) * | 2019-12-26 | 2020-07-03 | 北京蓦然认知科技有限公司 | Generation and sharing method and device of general APP ecological graph |
| CN111309341A (en) * | 2020-02-17 | 2020-06-19 | 中南大学 | Android application installation flow optimization method based on time-consuming prediction |
| CN111309341B (en) * | 2020-02-17 | 2023-04-07 | 中南大学 | Android application installation flow optimization method based on time-consuming prediction |
| CN111753145A (en) * | 2020-06-10 | 2020-10-09 | 西北工业大学 | Mobile application use prediction method based on time sequence mode |
| CN112328935A (en) * | 2020-10-29 | 2021-02-05 | 胡培培 | Application program pushing system and method based on big data |
| US11869015B1 (en) | 2022-12-09 | 2024-01-09 | Northern Trust Corporation | Computing technologies for benchmarking |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN107770365A (en) | A kind of prediction of multiple features APP streams and recommendation method based on Android | |
| Nikitin et al. | Automated evolutionary approach for the design of composite machine learning pipelines | |
| CN110263979B (en) | Method and device for predicting sample label based on reinforcement learning model | |
| Nguyen et al. | Using meta-mining to support data mining workflow planning and optimization | |
| CN111723292B (en) | Recommendation method, system, electronic equipment and storage medium based on graph neural network | |
| CN110799997B (en) | Industrial data service, data modeling, and data application platform | |
| US20160217385A1 (en) | Method and apparatus for analyzing missing not at random data and recommendation system using the same | |
| US11886779B2 (en) | Accelerated simulation setup process using prior knowledge extraction for problem matching | |
| EP4024203A1 (en) | System performance optimization | |
| Cui et al. | Graph based multispectral high resolution image segmentation | |
| CN101140513A (en) | Software requirement acquiring system | |
| Pham et al. | Unsupervised training of Bayesian networks for data clustering | |
| KR20160066395A (en) | Method for analyzing data based on matrix factorization model and apparatus therefor | |
| Sakhrawi et al. | Support vector regression for enhancement effort prediction of Scrum projects from COSMIC functional size | |
| Zhang et al. | A tree-structured multi-task model recommender | |
| CN113763031A (en) | Commodity recommendation method and device, electronic equipment and storage medium | |
| US20230409929A1 (en) | Methods and apparatuses for training prediction model | |
| CN112559877A (en) | CTR (China railway) estimation method and system based on cross-platform heterogeneous data and behavior context | |
| CN111324594A (en) | Data fusion method, device, equipment and storage medium for grain processing industry | |
| US20160042277A1 (en) | Social action and social tie prediction | |
| Korhonen et al. | A dimensional decomposition approach to identifying efficient units in large-scale DEA models | |
| CN109711555A (en) | A kind of method and system of predetermined depth learning model single-wheel iteration time | |
| CN113591979A (en) | Industry category identification method, equipment, medium and computer program product | |
| CN112070162A (en) | Multi-class processing task training sample construction method, device and medium | |
| Parvin et al. | Nonnegative matrix factorization regularized with trust relationships for solving cold-start problem in recommender Systems |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| WD01 | Invention patent application deemed withdrawn after publication |
Application publication date: 20180306 |
|
| WD01 | Invention patent application deemed withdrawn after publication |