CN109151073A - Mobile phone application software recommended method and system - Google Patents
Mobile phone application software recommended method and system Download PDFInfo
- Publication number
- CN109151073A CN109151073A CN201811267606.6A CN201811267606A CN109151073A CN 109151073 A CN109151073 A CN 109151073A CN 201811267606 A CN201811267606 A CN 201811267606A CN 109151073 A CN109151073 A CN 109151073A
- Authority
- CN
- China
- Prior art keywords
- vector
- order
- embedded
- low order
- feature
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
Classifications
-
- 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
- H04L67/50—Network services
- H04L67/535—Tracking the activity of the user
-
- 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
- H04L67/50—Network services
- H04L67/55—Push-based network services
Landscapes
- Engineering & Computer Science (AREA)
- Computer Networks & Wireless Communication (AREA)
- Signal Processing (AREA)
- Computer Hardware Design (AREA)
- General Engineering & Computer Science (AREA)
- Telephonic Communication Services (AREA)
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
Abstract
The present invention discloses a kind of mobile phone application software recommended method and system, belongs to technical field of information communication, and the downloading data of user need to be statisticallyd analyze based on application shop by solving application software recommended technology in the prior art, the problem of using limitation.This method comprises: the mobile phone application data of acquisition user and the essential information data of user, construct sparse features vector;Sparse features vector is converted to obtain the low order insertion non-embedded vector of vector sum high-order;Using the processing low order insertion vector capture feature low order linear expression of low order incorporation model, and using the non-embedded vector capture feature high-order nonlinear expression of the non-embedded model treatment high-order of high-order;Splicing fusion is carried out to the expression of feature low order linear and the expression of feature high order linear, and handles the fused result of splicing using fully-connected network and obtains mobile phone application recommendation results.The system includes the method that above-mentioned technical proposal is mentioned.
Description
Technical field
The present invention relates to technical field of information communication more particularly to a kind of mobile phone application software recommended methods and system.
Background technique
With the development of internet technology and the quick of intelligent terminal is popularized, and the quantity of application software increases in explosion type
It is long.In the application software of magnanimity, when user does not know required application software title, it is difficult in numerous application software fast
Speed finds the application software needed for oneself, is based on this, application software recommended technology comes into being.
Existing application software recommended technology mainly relies on application shop to be just able to achieve, i.e., is applying quotient by acquisition user
Downloading data in shop is user's simulated portrait after monitoring analysis user is to the usage behavior for the application software downloaded, so
Recommend the application software being consistent with its interest preference to user based on simulated portrait afterwards, recommends to apply automatically to realize to user
The effect of software.
It is just able to achieve since existing application software recommended technology must be based on application shop, if not installed in user mobile phone
Application shop then can not be achieved the automatic recommendation of application software, it is seen then that existing application software recommended technology has biggish
Limitation reduces the usage experience of user.
Summary of the invention
The purpose of the present invention is to provide a kind of mobile phone application software recommended method and systems, solve in the prior art
Application software recommended technology need to statistically analyze the downloading data of user based on application shop, the problem of using limitation.
To achieve the goals above, an aspect of of the present present invention provides a kind of mobile phone application software recommended method, comprising:
The mobile phone application data of user and the essential information data of user are acquired, sparse features vector is constructed;
Sparse features vector is converted to obtain the low order insertion non-embedded vector of vector sum high-order;
Using the processing low order insertion vector capture feature low order linear expression of low order incorporation model, and use high-order non-embedding
Enter the non-embedded vector capture feature high-order nonlinear expression of model treatment high-order;
Splicing fusion is carried out to the expression of feature low order linear and the expression of feature high order linear, and is handled using fully-connected network
Splice fused result and obtains mobile phone application recommendation results.
Preferably, the mobile phone application data of the acquisition user and the essential information data of user, construct sparse features
The method of vector includes:
Obtain the application installation data in user's current phone and monitor it is therein using situation, generate using
Preference vector;
The discrete features of essential information data acquisition user based on user, after to the discrete features one-hot coding
It generates user interest and simulates vector;
It is fitted to obtain the sparse features vector using preference vector and user interest simulation vector by described.
Preferably, the method for converting to obtain the low order insertion non-embedded vector of vector sum high-order to sparse features vector includes:
Extract each element feature feature in sparse features vector1To featurei, the i expression sparse features vector
In number of elements;
Successively each element feature is converted using low order embedding formula to obtain low order insertion element, and non-embedding using high-order
Enter formula each element feature is converted to obtain high-order insertion element, wherein the low order embedding formula is xi=Wifeaturei,
The non-embedded formula of high-order is zi=Hifeaturei, WiFor sparse features featureiLow order embeded matrix, xiFor low order
It is embedded in element, the HiFor sparse features featureiHigh-order embeded matrix, ziFor the non-embedded element of high-order;
Splicing low order insertion element obtains low order insertion vector, and to obtain high-order non-embedded for the splicing non-embedded element of high-order
Vector.
Preferably, the training method of the low order incorporation model includes;
Step S11 obtains multiple low order insertion vector data collection and the insertion vector verifying of multiple low orders from database
Collection;
Step S12, point-by-point multiplication algorithm successively obtains the corresponding feature of multiple low order insertion vector data collection two-by-two for use
Low order linear expression;
Step S13, based on multiple feature low order linears expression initialization low order incorporation model;
Step S14 successively transfers multiple low order insertion vector verifying collection input initialization low order incorporation model correspondences and obtains
Multiple low orders are embedded in vector verification result;
More than upper low order insertion vector verification result whether step S15 judge current low order insertion vector verification result
Accurately, if then return step S11, if it is not, to save as the low order embedding for the initialization low order incorporation model in then will be last round of
Enter model.
Optionally, the calculation formula of the multiplication algorithm point-by-point two-by-two isIts
In,For the symbol that is multiplied point by point, wijWeight is interacted for vector,Indicate the expression of feature low order linear.
Preferably, the training method of the non-embedded model of the high-order includes;
Step S21 obtains the non-embedded vector data collection of multiple high-orders and the non-embedded vector of multiple high-orders from database
Verifying collection;
Step S22 successively obtains the corresponding feature of the non-embedded vector data collection of multiple high-orders using multi-layer perception (MLP) algorithm
High-order nonlinear expression;
Step S23, based on multiple feature high-order nonlinears expression initialization non-embedded model of high-order;
It is corresponding successively to transfer the non-embedded model of the non-embedded vector verifying collection input initialization high-order of multiple high-orders by step S24
Obtain the non-embedded vector verification result of multiple high-orders;
Step S25 judges the non-embedded vector verification result of current high-order vector verifying knot whether more non-embedded than a upper high-order
Fruit is more acurrate, if then return step S1, if it is not, the non-embedded model of initialization high-order in then will be last round of saves as the height
The non-embedded model of rank.
Optionally, the formula of the multi-layer perception (MLP) algorithm is
It is described
Wherein, the z indicates the non-embedded vector of high-order being spliced to form, layeriIndicate i-th layer of multi-layer perception (MLP) of table
It reaching, f is activation primitive,Indicate i-th layer of weight matrix, biIndicate i-th layer of corresponding bias vector of perceptron.
Preferably, described that splicing fusion is carried out to the expression of feature low order linear and the expression of feature high order linear, using connecting entirely
Connect the method that the vector result of network processes splicing fusion obtains mobile phone application recommendation results are as follows:
UsingThe expression of formula feature low order linear and the non-high order linear expression of feature carry out splicing fusion;
Splicing fusion results are calculated by fully-connected network formula, user is obtained and is inclined to the corresponding mobile phone application of downloading
Probability value, the fully-connected network formula isWhereinFor the weight square of output layer
Battle array, the boutFor the bias vector of output layer;
A maximum value is chosen from probability value obtains mobile phone application recommendation results to user's push.
Compared with prior art, mobile phone application software recommended method provided by the invention has the advantages that
In mobile phone application software recommended method provided by the invention, the mobile phone application data of acquisition user and user first
Essential information data construct sparse features vector, and wherein mobile phone application data include the application software title installed in user mobile phone
With in a period of time to the statistics of installed application software service condition, the essential information data of user include the age of user,
The discrete features such as educational background, hobby, habit, by by application software service condition and discrete features vectorization set up sparse features to
Then obtained sparse features vector is converted to the identifiable dense characteristic vector of model namely low order is embedded in vector sum by amount
The non-embedded vector of high-order, and the feature low order linear between low order insertion vector input low order incorporation model capture feature is expressed,
And express the feature high-order nonlinear between the non-embedded model capture feature of the non-embedded vector input high-order of high-order, finally incite somebody to action
The feature low order linear expression arrived and feature high order linear express anastomosing and splicing, and it is fused to handle splicing using fully-connected network
As a result the mobile phone applied probability of tendency downloading is obtained, and extracts the wherein highest mobile phone of probability and applies to user and recommend to download.
As it can be seen that the present invention obtains the mobile phone application data and essential information data of user by big data technology, and use
Depth learning technology is expressed to capture the expression of the feature low order linear between feature and feature high-order nonlinear, significantly improves model
Expansion performance and ability to express, the inherent complex relationship of mass data can be accurately captured, to precisely recommend to user
It is inclined to the mobile phone application of downloading, therefore is compared to the prior art and can only just be able to achieve mobile phone by mounted application shop and answer
For being recommended automatically with software, the present invention breaches use limitation, improves user experience.
Another aspect of the present invention provides a kind of mobile phone application software recommender system, applied to described in above-mentioned technical proposal
In mobile phone application software recommended method, the system comprises:
Data acquisition unit constructs dilute for acquiring the mobile phone application data of user and the essential information data of user
Dredge feature vector;
Vector scaling unit obtains the low order insertion non-embedded vector of vector sum high-order for converting to sparse features vector;
Expression unit, for capturing the expression of feature low order linear using low order incorporation model processing low order insertion vector, with
And using the non-embedded vector capture feature high-order nonlinear expression of the non-embedded model treatment high-order of high-order;
Recommendation unit, for capturing the expression of feature low order linear using low order incorporation model processing low order insertion vector, with
And using the non-embedded vector capture feature high-order nonlinear expression of the non-embedded model treatment high-order of high-order.
Preferably, the data acquisition unit includes the first acquisition module and the second acquisition module;
First acquisition module is used to obtain the application installation data in user's current phone and monitors application therein
Service condition is generated using preference vector;
Second acquisition module is used for the discrete features of the essential information data acquisition user based on user, by institute
User interest simulation vector is generated after stating discrete features one-hot coding;
The fitting module is used to be fitted to obtain using preference vector and user interest simulation vector by described
The sparse features vector.
Compared with prior art, the beneficial effect of mobile phone application software recommender system provided by the invention and above-mentioned technical side
The beneficial effect for the mobile phone application software recommended method that case provides is identical, and this will not be repeated here.
Detailed description of the invention
The drawings described herein are used to provide a further understanding of the present invention, constitutes a part of the invention, this hair
Bright illustrative embodiments and their description are used to explain the present invention, and are not constituted improper limitations of the present invention.In the accompanying drawings:
Fig. 1 is the flow diagram of mobile phone application software recommended method in the embodiment of the present invention one;
Fig. 2 is the model schematic of capture feature low order linear expression in the embodiment of the present invention one;
Fig. 3 is the model schematic of capture feature high-order nonlinear expression in the embodiment of the present invention one;
Fig. 4 is the structural block diagram of mobile phone application software recommender system in the embodiment of the present invention two.
Appended drawing reference:
1- data acquisition unit, 2- vector scaling unit;
3- expression unit, 4- recommendation unit;
The first acquisition module of 11-, the second acquisition module of 12-;
13- fitting module.
Specific embodiment
In order to make the foregoing objectives, features and advantages of the present invention clearer and more comprehensible, implement below in conjunction with the present invention
Attached drawing in example, technical scheme in the embodiment of the invention is clearly and completely described.Obviously, described embodiment
Only a part of the embodiment of the present invention, instead of all the embodiments.Based on the embodiments of the present invention, the common skill in this field
Art personnel all other embodiment obtained without creative labor belongs to the model that the present invention protects
It encloses.
Embodiment one
Fig. 1-Fig. 3 is please referred to, the present embodiment provides a kind of mobile phone application software recommended methods, comprising:
The mobile phone application data of user and the essential information data of user are acquired, sparse features vector is constructed;
Sparse features vector is converted to obtain the low order insertion non-embedded vector of vector sum high-order;
Using the processing low order insertion vector capture feature low order linear expression of low order incorporation model, and use high-order non-embedding
Enter the non-embedded vector capture feature high-order nonlinear expression of model treatment high-order;
Splicing fusion is carried out to the expression of feature low order linear and the expression of feature high order linear, and is handled using fully-connected network
Splice fused result and obtains mobile phone application recommendation results.
In mobile phone application software recommended method provided in this embodiment, mobile phone application data and the user of user are acquired first
Essential information data construct sparse features vector, wherein mobile phone application data include the application software name installed in user mobile phone
Claim and in a period of time to the statistics of installed application software service condition, the essential information data of user include the year of user
The discrete features such as age, educational background, hobby, habit, by the way that application software service condition and discrete features vectorization are set up sparse spy
Levy vector, then by obtained sparse features vector be converted to the identifiable dense characteristic vector of model namely low order be embedded in
Amount and the non-embedded vector of high-order, and low order insertion vector input low order incorporation model is captured into the feature low order linear table between feature
It reaches, and the feature high-order nonlinear between the non-embedded model capture feature of the non-embedded vector input high-order of high-order is expressed, finally
Anastomosing and splicing is expressed into the expression of obtained feature low order linear and feature high order linear, handle splicing fusion using fully-connected network
Result afterwards obtains the mobile phone applied probability of tendency downloading, and extracts the wherein highest mobile phone of probability and apply under user's recommendation
It carries.
As it can be seen that the present embodiment obtains the mobile phone application data and essential information data of user by big data technology, and adopt
The expression of feature low order linear and the expression of feature high-order nonlinear between feature are captured with depth learning technology, significantly improve mould
The expansion performance and ability to express of type, can accurately capture the inherent complex relationship of mass data, to precisely push away to user
The mobile phone application of tendency downloading is recommended, therefore be compared to the prior art can only just be able to achieve mobile phone by mounted application shop
For application software is recommended automatically, the present embodiment breaches use limitation, improves user experience.
It is understood that the acquisition modes of mobile phone application data and essential information data are varied, for example, can be from
It transfers in the database of server, or is directly acquired from the mobile phone terminal of user.
The mobile phone application data of user and the essential information data of user are acquired in above-described embodiment, construct sparse features
The method of vector includes:
Obtain the application installation data in user's current phone and monitor it is therein using situation, generate using
Preference vector;The discrete features of essential information data acquisition user based on user, by raw after discrete features one-hot coding
Vector is simulated at user interest;It will be fitted to obtain sparse features vector using preference vector and user interest simulation vector.
When it is implemented, based in current phone using installation data initialize a known length number of applications to
This was applied corresponding vector position if user uses some to apply whithin a period of time by amount, Initial Value definition 0
Place is changed to 1 by 0, and the vector after changing at this time is using preference vector, for example, being mounted with 5 applications in mobile phone, then
Initial vector be (0,0,0,0,0), if user only used first application whithin a period of time, at this time using preference to
Amount is (1,0,0,0,0).The essential information data of user include the discrete features such as age, educational background, hobby, the habit of user, are led to
It crosses and these discrete features one-hot codings is produced to user interest simulation vector, in addition, discrete features one-hot coding is ability
Field technique personnel's common technology, fitting obtain sparse features vector only and are to using preference vector and user interest to simulate vector
Simple combination is also those skilled in the art's common technology, and the present embodiment no longer traces herein.
Specifically, sparse features vector is converted in above-described embodiment to obtain the low order insertion non-embedded vector of vector sum high-order
Method include:
Extract each element feature feature in sparse features vector1To featurei, i indicate sparse features vector in
Number of elements;Successively each element feature is converted using low order embedding formula to obtain low order insertion element, and non-using high-order
Embedding formula converts each element feature to obtain high-order insertion element, wherein low order embedding formula is xi=Wifeaturei, high
The non-embedded formula of rank is zi=Hifeaturei, WiFor sparse features featureiLow order embeded matrix, also as low order is sparse
The weight of feature, xiElement, H are embedded in for low orderiFor sparse features featureiThe non-embedded matrix of high-order, also as high-order is dilute
Dredge the weight of feature, ziFor the non-embedded element of high-order;Splicing low order insertion element obtains low order insertion vector, and splicing high-order
Non-embedded element obtains the non-embedded vector of high-order, above-mentioned splicing low order insertion element refer to the low order insertion element that will acquire according to
Sequence merges composition low order and is embedded in vector, and the above-mentioned non-embedded element of splicing high-order refers to the non-embedded element of the high-order that will acquire sequentially
Merge the composition non-embedded vector of high-order.In addition, WiAnd HiBoth it can manually be arranged, and can also be analyzed and be adopted by server database
The data collected are derived automatically from.
Optionally, in above-described embodiment, the training method of low order incorporation model includes;
Step S11 obtains multiple low order insertion vector data collection and the insertion vector verifying of multiple low orders from database
Collection;Step S12, point-by-point multiplication algorithm successively obtains the corresponding feature low order line of multiple low order insertion vector data collection two-by-two for use
Property expression;Step S13, based on multiple feature low order linears expression initialization low order incorporation model;Step S14 is successively transferred more
A low order insertion vector verifying collection input initialization low order incorporation model is corresponding to obtain multiple low order insertion vector verification results;Step
Rapid S15 judges whether current low order insertion vector verification result is more more acurrate than upper low order insertion vector verification result, if then
Return step S11, if it is not, the initialization low order incorporation model in then will be last round of saves as low order incorporation model.Wherein, two-by-two
The calculation formula of multiplication algorithm is point by pointWherein,For the symbol that is multiplied point by point, wijFor
Vector interacts weight,Indicate the expression of feature low order linear.Wherein, wijBoth it can manually be arranged, server count can also be passed through
Collected data are analyzed according to library to be derived automatically from.
When it is implemented, the training speed of low order incorporation model and accuracy in order to balance, introduces early stop in step S15
Number judgment mechanism is taken turns, judges whether current low order insertion vector verification result is more quasi- than upper low order insertion vector verification result
Really, if so, return step S11 continues to train after saving currently trained low order incorporation model, if it is not, then judging whether to reach
Morning stops taking turns number, when do not reach it is early stop taking turns number when return step S11 continue to train, when having reached morning and stopping taking turns number, directly terminate
Training, the low order incorporation model that last time is saved export.As it can be seen that the present embodiment is embedding using back-propagation algorithm training low order
The training speed of model can either be guaranteed by entering model, and to the accuracy for enough guaranteeing model training.
Similarly, the training method of the non-embedded model of high-order includes: in above-described embodiment
Step S21 obtains the non-embedded vector data collection of multiple high-orders and the non-embedded vector of multiple high-orders from database
Verifying collection;It is high successively to obtain the corresponding feature of the non-embedded vector data collection of multiple high-orders using multi-layer perception (MLP) algorithm by step S22
Rank non-linear expression;Step S23, based on multiple feature high-order nonlinears expression initialization non-embedded model of high-order;Step S24,
Successively transferring the non-embedded model of multiple high-orders non-embedded vector verifying collection input initialization high-order, corresponding to obtain multiple high-orders non-embedding
Incoming vector verification result;Step S25 judges the non-embedded vector verification result of current high-order vector whether more non-embedded than a upper high-order
Verification result is more acurrate, if then return step S1, if it is not, the non-embedded model of initialization high-order in then will be last round of saves as
The non-embedded model of high-order.The formula of above-mentioned multi-layer perception (MLP) algorithm is
It is described
Wherein, z indicates the non-embedded vector of high-order being spliced to form, layeriIndicate i-th layer of multi-layer perception (MLP) of expression, f
For activation primitive,Indicate i-th layer of weight matrix, biIndicate i-th layer of corresponding bias vector of perceptron.
When it is implemented, the training speed of the non-embedded model of high-order and accuracy in order to balance, introduces early in step s 25
Stop taking turns number judgment mechanism, judge that whether more non-embedded than a upper high-order the non-embedded vector of current high-order vector be more acurrate, if saving
Currently return step S21 continues to train after the trained non-embedded model of high-order, then judges whether that having reached morning stops taking turns number if not,
When do not reach it is early stop taking turns number when return step S21 continue to train, when having reached morning and stopping taking turns number, directly terminate to train, and will most
The non-embedded model output of the high-order once saved afterwards.As it can be seen that the present embodiment utilizes the back-propagation algorithm training non-embedded mould of high-order
Type can either guarantee the training speed of model, and to the accuracy for enough guaranteeing model training.
Specifically, splicing fusion is carried out to the expression of feature low order linear and the expression of feature high order linear in above-described embodiment,
The method for obtaining mobile phone application recommendation results using the vector result that fully-connected network handles splicing fusion are as follows:
UsingThe expression of formula feature low order linear and the expression of feature high-order nonlinear carry out splicing fusion;
Splicing fusion results are calculated by fully-connected network formula, user is obtained and is inclined to the corresponding mobile phone application of downloading
Probability value, fully-connected network formula isWhereinFor the weight matrix of output layer,
boutFor the bias vector of output layer,Indicate that user is inclined to the probability value for downloading corresponding mobile phone application, the probability value be it is multiple, so
A maximum value is chosen from probability value afterwards and obtains mobile phone application recommendation results to user's push.Wherein,And boutIt both can be with people
Work setting, collected data can also be analyzed by server database and are derived automatically from.
Embodiment two
Fig. 1 and Fig. 4 are please referred to, the present embodiment provides a kind of mobile phone application software recommender systems, comprising:
Data acquisition unit 1 constructs dilute for acquiring the mobile phone application data of user and the essential information data of user
Dredge feature vector;
Vector scaling unit 2 obtains the low order insertion non-embedded vector of vector sum high-order for converting to sparse features vector;
Expression unit 3, for capturing the expression of feature low order linear using low order incorporation model processing low order insertion vector, with
And using the non-embedded vector capture feature high-order nonlinear expression of the non-embedded model treatment high-order of high-order;
Recommendation unit 4, for capturing the expression of feature low order linear using low order incorporation model processing low order insertion vector, with
And using the non-embedded vector capture feature high-order nonlinear expression of the non-embedded model treatment high-order of high-order.
Preferably, data acquisition unit 1 includes the first acquisition module 11 and the second acquisition module 12;
First acquisition module 11, which is used to obtain the application installation data in user's current phone and monitors application therein, to be made
With situation, generate using preference vector;
Second acquisition module 12 is used for the discrete features of the essential information data acquisition user based on user, by discrete
User interest is generated after feature one-hot coding simulates vector;
Fitting module 13 for will be fitted to obtain using preference vector and user interest simulation vector sparse features to
Amount.
Compared with prior art, the beneficial effect of mobile phone application software recommender system provided in an embodiment of the present invention with it is above-mentioned
The beneficial effect for the mobile phone application software recommended method that embodiment one provides is identical, and this will not be repeated here.
It will appreciated by the skilled person that realizing that all or part of the steps in foregoing invention method is can to lead to
Program is crossed to instruct relevant hardware and complete, above procedure can store in computer-readable storage medium, the program
When being executed, each step including above-described embodiment method, and the storage medium may is that ROM/RAM, magnetic disk, CD,
Storage card etc..
The above description is merely a specific embodiment, but scope of protection of the present invention is not limited thereto, any
Those familiar with the art in the technical scope disclosed by the present invention, can easily think of the change or the replacement, and should all contain
Lid is within protection scope of the present invention.Therefore, protection scope of the present invention should be based on the protection scope of the described claims.
Claims (10)
1. a kind of mobile phone application software recommended method characterized by comprising
The mobile phone application data of user and the essential information data of user are acquired, sparse features vector is constructed;
Sparse features vector is converted to obtain the low order insertion non-embedded vector of vector sum high-order;
Using the processing low order insertion vector capture feature low order linear expression of low order incorporation model, and use the non-embedded mould of high-order
Type handles the non-embedded vector capture feature high-order nonlinear expression of high-order;
Splicing fusion is carried out to the expression of feature low order linear and the expression of feature high order linear, and uses fully-connected network processing splicing
Fused result obtains mobile phone application recommendation results.
2. the method according to claim 1, wherein mobile phone application data and the user of the acquisition user
Essential information data, the method for constructing sparse features vector include:
It obtains the application installation data in user's current phone and monitors therein using situation, generate using preference
Vector;
The discrete features of essential information data acquisition user based on user, by being generated after the discrete features one-hot coding
User interest simulates vector;
It is fitted to obtain the sparse features vector using preference vector and user interest simulation vector by described.
3. the method according to claim 1, wherein converting to obtain low order insertion vector sum to sparse features vector
The method of the non-embedded vector of high-order includes:
Extract each element feature feature in sparse features vector1To featurei, the i indicates in sparse features vector
Number of elements;
Successively each element feature is converted using low order embedding formula to obtain low order insertion element, and uses the non-embedded public affairs of high-order
Formula converts each element feature to obtain high-order insertion element, wherein the low order embedding formula is xi=Wifeaturei, described
The non-embedded formula of high-order is zi=Hifeaturei, WiFor sparse features featureiLow order embeded matrix, xiFor low order insertion
Element, the HiFor sparse features featureiThe non-embedded matrix of high-order, ziFor the non-embedded element of high-order;
Splicing low order insertion element obtain low order insertion vector, and splicing the non-embedded element of high-order obtain high-order it is non-embedded to
Amount.
4. the method according to claim 1, wherein the training method of the low order incorporation model includes:
Step S11 obtains multiple low order insertion vector data collection and multiple low orders insertion vector verifying collection from database;
Step S12, point-by-point multiplication algorithm successively obtains the corresponding feature low order of multiple low order insertion vector data collection two-by-two for use
Linear expression;
Step S13, based on multiple feature low order linears expression initialization low order incorporation model;
Step S14, successively transfer multiple low orders insertion vectors verifying collection input initialization low order incorporation models it is corresponding obtain it is multiple
Low order is embedded in vector verification result;
Step S15 judges whether current low order insertion vector verification result is more more acurrate than upper low order insertion vector verification result,
If then return step S11, if it is not, the initialization low order incorporation model in then will be last round of saves as the low order insertion mould
Type.
5. according to the method described in claim 4, it is characterized in that, the calculation formula of the multiplication algorithm point-by-point two-by-two isWherein,For the symbol that is multiplied point by point, wijWeight is interacted for vector,Indicate feature
Low order linear expression.
6. the method according to claim 1, wherein the training method of the non-embedded model of the high-order includes:
Step S21 obtains the non-embedded vector data collection of multiple high-orders and the non-embedded vector verifying of multiple high-orders from database
Collection;
Step S22 successively obtains the corresponding feature high-order of the non-embedded vector data collection of multiple high-orders using multi-layer perception (MLP) algorithm
Non-linear expression;
Step S23, based on multiple feature high-order nonlinears expression initialization non-embedded model of high-order;
Step S24 successively transfers the non-embedded non-embedded model correspondence of vector verifying collection input initialization high-order of multiple high-orders and obtains
Multiple non-embedded vector verification results of high-order;
Step S25 judges the non-embedded vector verification result of current high-order vector verification result whether more non-embedded than a upper high-order more
Accurately, if then return step S1, if it is not, to save as the high-order non-for the non-embedded model of initialization high-order in then will be last round of
Incorporation model.
7. according to the method described in claim 6, it is characterized in that, the formula of the multi-layer perception (MLP) algorithm isIt is described
Wherein, the z indicates the non-embedded vector of high-order being spliced to form, layeriIndicate i-th layer of multi-layer perception (MLP) of expression, f is
Activation primitive,Indicate i-th layer of weight matrix, biIndicate i-th layer of corresponding bias vector of perceptron.
8. the method according to claim 1, wherein described to the expression of feature low order linear and feature high order linear
Expression carries out splicing fusion, obtains the side of mobile phone application recommendation results using the vector result that fully-connected network handles splicing fusion
Method are as follows:
UsingThe expression of formula feature low order linear and the expression of feature high-order nonlinear carry out splicing fusion;
Splicing fusion results are calculated by fully-connected network formula, user is obtained and is inclined to the general of the corresponding mobile phone application of downloading
Rate value, the fully-connected network formula areWhereinFor the weight matrix of output layer, institute
State boutFor the bias vector of output layer;
A maximum value is chosen from probability value obtains mobile phone application recommendation results to user's push.
9. a kind of mobile phone application software recommender system characterized by comprising
Data acquisition unit constructs sparse spy for acquiring the mobile phone application data of user and the essential information data of user
Levy vector;
Vector scaling unit obtains the low order insertion non-embedded vector of vector sum high-order for converting to sparse features vector;
Expression unit captures the expression of feature low order linear for being embedded in vector using low order incorporation model processing low order, and adopts
With the non-embedded vector capture feature high-order nonlinear expression of the non-embedded model treatment high-order of high-order;
Recommendation unit captures the expression of feature low order linear for being embedded in vector using low order incorporation model processing low order, and adopts
With the non-embedded vector capture feature high-order nonlinear expression of the non-embedded model treatment high-order of high-order.
10. system according to claim 9, which is characterized in that the data acquisition unit include the first acquisition module and
Second acquisition module;
First acquisition module be used for obtain in user's current phone application installation data and monitor it is therein using
Situation is generated using preference vector;
Second acquisition module be used for the essential information data acquisition user based on user discrete features, by it is described from
User interest simulation vector is generated after dissipating feature one-hot coding;
The fitting module is used to be fitted to obtain using preference vector and user interest simulation vector by described described
Sparse features vector.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201811267606.6A CN109151073B (en) | 2018-10-29 | 2018-10-29 | Mobile phone application software recommendation method and system |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201811267606.6A CN109151073B (en) | 2018-10-29 | 2018-10-29 | Mobile phone application software recommendation method and system |
Publications (2)
Publication Number | Publication Date |
---|---|
CN109151073A true CN109151073A (en) | 2019-01-04 |
CN109151073B CN109151073B (en) | 2021-07-16 |
Family
ID=64806574
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201811267606.6A Active CN109151073B (en) | 2018-10-29 | 2018-10-29 | Mobile phone application software recommendation method and system |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN109151073B (en) |
Cited By (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN111813447A (en) * | 2019-04-12 | 2020-10-23 | 杭州中天微系统有限公司 | Processing method and processing device for data splicing instruction |
Citations (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US7885820B1 (en) * | 2000-07-19 | 2011-02-08 | Convergys Cmg Utah, Inc. | Expert system supported interactive product selection and recommendation |
CN102890711A (en) * | 2012-09-13 | 2013-01-23 | 中国人民解放军国防科学技术大学 | Retrieval ordering method and system |
CN105354330A (en) * | 2015-11-27 | 2016-02-24 | 南京邮电大学 | Sparse data preprocessing based collaborative filtering recommendation method |
US20170236069A1 (en) * | 2016-02-11 | 2017-08-17 | Nec Laboratories America, Inc. | Scalable supervised high-order parametric embedding for big data visualization |
CN107273909A (en) * | 2016-04-08 | 2017-10-20 | 上海市玻森数据科技有限公司 | The sorting algorithm of high dimensional data |
-
2018
- 2018-10-29 CN CN201811267606.6A patent/CN109151073B/en active Active
Patent Citations (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US7885820B1 (en) * | 2000-07-19 | 2011-02-08 | Convergys Cmg Utah, Inc. | Expert system supported interactive product selection and recommendation |
CN102890711A (en) * | 2012-09-13 | 2013-01-23 | 中国人民解放军国防科学技术大学 | Retrieval ordering method and system |
CN105354330A (en) * | 2015-11-27 | 2016-02-24 | 南京邮电大学 | Sparse data preprocessing based collaborative filtering recommendation method |
US20170236069A1 (en) * | 2016-02-11 | 2017-08-17 | Nec Laboratories America, Inc. | Scalable supervised high-order parametric embedding for big data visualization |
CN107273909A (en) * | 2016-04-08 | 2017-10-20 | 上海市玻森数据科技有限公司 | The sorting algorithm of high dimensional data |
Non-Patent Citations (1)
Title |
---|
JIANXUN LIAN ET AL.: "xDeepFM: Combining Explicit and Implicit Feature Interactions for Recommender Systems", 《KDD "18: PROCEEDINGS OF THE 24TH ACM SIGKDD INTERNATIONAL CONFERENCE ON KNOWLEDGE DISCOVERY & DATA MINING》 * |
Cited By (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN111813447A (en) * | 2019-04-12 | 2020-10-23 | 杭州中天微系统有限公司 | Processing method and processing device for data splicing instruction |
CN111813447B (en) * | 2019-04-12 | 2022-11-08 | 杭州中天微系统有限公司 | Processing method and processing device for data splicing instruction |
Also Published As
Publication number | Publication date |
---|---|
CN109151073B (en) | 2021-07-16 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN110147711A (en) | Video scene recognition methods, device, storage medium and electronic device | |
CN113923441B (en) | Video quality assessment method and device and electronic equipment | |
CN109859120A (en) | Image defogging method based on multiple dimensioned residual error network | |
CN109741477A (en) | Work attendance management system, method and electronic equipment | |
CN110363081A (en) | Face identification method, device, equipment and computer readable storage medium | |
CN110895568B (en) | Method and system for processing court trial records | |
CN109246171A (en) | Application program recommended method, recommendation apparatus and terminal device | |
CN107679457A (en) | User identity method of calibration and device | |
CN111009160A (en) | Teaching system based on VR virtual classroom | |
KR20180014259A (en) | System for providing education brokerage service, server and method for providing education brokerage service | |
CN111738010B (en) | Method and device for generating semantic matching model | |
CN108733801A (en) | A kind of moving-vision search method towards digital humanity | |
CN112100221A (en) | Information recommendation method and device, recommendation server and storage medium | |
CN108875539A (en) | Expression matching process, device and system and storage medium | |
CN112906806A (en) | Data optimization method and device based on neural network | |
CN109151073A (en) | Mobile phone application software recommended method and system | |
CN114360018B (en) | Rendering method and device of three-dimensional facial expression, storage medium and electronic device | |
FR2799911A1 (en) | STRESS CONTROL SYSTEM | |
CN109885727A (en) | Data method for pushing, device, electronic equipment and system | |
CN110427998A (en) | Model training, object detection method and device, electronic equipment, storage medium | |
CN111461228B (en) | Image recommendation method and device and storage medium | |
CN110162452A (en) | A kind of analog detection method, device and storage medium for testing and control service | |
CN114339252A (en) | Data compression method and device | |
CN110210205A (en) | Determination method, apparatus, computer equipment and the computer storage medium of logging state | |
CN110381339A (en) | Picture transmission method and device |
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 | ||
GR01 | Patent grant | ||
GR01 | Patent grant |