CN110162711A - A kind of resource intelligent recommended method and system based on internet startup disk method - Google Patents
A kind of resource intelligent recommended method and system based on internet startup disk method Download PDFInfo
- Publication number
- CN110162711A CN110162711A CN201910451324.XA CN201910451324A CN110162711A CN 110162711 A CN110162711 A CN 110162711A CN 201910451324 A CN201910451324 A CN 201910451324A CN 110162711 A CN110162711 A CN 110162711A
- Authority
- CN
- China
- Prior art keywords
- user
- resource
- network
- node
- tag
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/90—Details of database functions independent of the retrieved data types
- G06F16/95—Retrieval from the web
- G06F16/953—Querying, e.g. by the use of web search engines
- G06F16/9535—Search customisation based on user profiles and personalisation
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/90—Details of database functions independent of the retrieved data types
- G06F16/95—Retrieval from the web
- G06F16/953—Querying, e.g. by the use of web search engines
- G06F16/9536—Search customisation based on social or collaborative filtering
Landscapes
- Engineering & Computer Science (AREA)
- Databases & Information Systems (AREA)
- Theoretical Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
Abstract
The invention belongs to network information processing technical fields, disclose a kind of resource intelligent recommended method and system based on internet startup disk method, user network of the building based on user tag, the user network based on user behavior and the resource network based on resource tag;Then user tag network is acted on using internet startup disk method and resource tag network obtains the low-dimensional feature vector of user and resource, and calculate the similar matrix between user between resource using feature vector;Using the feature of user node in the second order proximity characterization user behavior network of LINE model in internet startup disk method, the high-order neighbours by increasing node expand the neighbours of these nodes.The resource recommendation list that the present invention is provided the user with using content-based recommendation algorithm;Based on collaborative filtering recommending list, supplemented by content recommendation lists, the recommendation results of two kinds of algorithms is recommended into user jointly, the higher similar resource of probability in collaborative filtering recommending list is also added in final list.
Description
Technical field
The invention belongs to network information processing technical fields more particularly to a kind of resource intelligent based on internet startup disk method to push away
Recommend method and system.
Background technique
Currently, the immediate prior art:
In recent years, with the rapid development of Internet technology, the mode that people obtain information is more and more, and information is in explosion
Formula increases, this just makes user bring some puzzlements to the selection of various data, and bring is exactly that user can not be opposite therewith
Selection meets the data of oneself requirement in the shorter time.During people progress into the information overload stage, either
Information consumer or information producer encounter very big challenge: as information consumer, how to look for from bulk information
It is a very difficult thing to oneself information needed;And as information producer, how to allow the information of oneself to be shown one's talent,
Concern and a very difficult thing by users.
The birth effective solution of the recommended technology information overload epoch give the puzzlement of user's bring.And recommender system is exactly
Above-mentioned contradictory important tool is solved, recommender system is to be commercialized one of technology the most successful so far.Recommender system
Basic principle is to guess user preferences based on algorithm and big data, while excavating the potential demand of user, and user is helped to determine
Plan.The application scenarios of recommender system are very extensive, and many companies both domestic and external are using recommender system as its commercially produced product
Essential a part.Internet startup disk method is data analysis technique fast-developing in recent years, but it recommends to lead in individual character
Stage of the domain also in ground zero.
Traditional collaborative filtering is that similarity between user is calculated according to the historical behavior data of user, passes through similar use
Family carries out resource recommendation.This proposed algorithm not can effectively solve the cold start-up of the user in resource recommendation problem and resource is cold
Starting problem, while the problem of there is also Sparses.
In conclusion problem of the existing technology is:
(1) prior art proposed algorithm not can effectively solve the cold start-up of the user in resource recommendation problem and resource is cold opens
Dynamic problem, the Deta sparseness that user data rareness easily causes.
(2) with gradually huge, the growth of data volume of system, existing proposed algorithm not can effectively solve algorithm and calculate again
Miscellaneous degree problem.
(3) diversity of recommendation results and novelty are also problem in need of consideration in recommender system.
Solve the difficulty of above-mentioned technical problem:
As internet is more universal and mature, data largely increase in network, and it is latent that recommended technology mainly excavates user
In demand, user is helped to carry out decision.Traditional collaborative filtering and content-based recommendation algorithm all can not be solved effectively
Certainly cold start-up and data sparsity problem, while in the biggish situation of data volume, since algorithm complexity is higher, calculate cost
Also it increases accordingly.Paper " Learning social network embeddings for predicting information
Diffusion " in propose CSDK algorithm the diffusion of information problem in network be converted into the diffusion of information on lower dimensional space and ask
Topic, and nodal information is indicated using low-dimensional vector, but the complexity of algorithm can achieve a cube magnitude, so being not suitable for analysis
Large-scale data;It will be by net list in paper " Probabilistic Latent Document Network Embedding "
The adjacency matrix shown enters ginseng as svd algorithm, and node is indicated with low-dimensional, but performance is not in data mining for this expression
It is ideal.
Solve the meaning of above-mentioned technical problem:
Recommender system can allow user is cracking to find oneself in vast resources with effective solution data overload problem
It is required.The present invention efficiently solves cold start-up and Sparse Problem and the algorithm meter under big data quantity in recommender system
Calculate complexity issue.Faster and better the recommendation list with diversity and novelty can be provided for user.
Summary of the invention
In view of the problems of the existing technology, the present invention provides a kind of resource intelligent recommendation sides based on internet startup disk method
Method and system.
The invention is realized in this way a kind of resource intelligent recommended method based on internet startup disk method includes:
Step 1 utilizes user tag data, user's historical behavior data in systems and the resource tag in system
Data construct the user network based on user tag, the user network based on user behavior and the resource network based on resource tag
Network;
Step 2, acts on user tag network using internet startup disk method and resource tag network obtains user and resource
Low-dimensional feature vector, and the similar matrix between user between resource is calculated using feature vector;
Step 3 utilizes user node in the second order proximity characterization user behavior network of LINE model in internet startup disk method
Feature, and by increase user node high-order neighbours expand user node neighbours;It whether is new user according to user, into
The recommendation of row different modes.
Further, in step 1, the method for obtaining user tag data includes: for user u, user tag collection TuIt indicates
The tag set of one user u itself<u,t>, data set T1={ Tu: u ∈ U } be all users label record set;It obtains
User behavior data, the behavior b triple that a user u generates resource h<u,h,b>it indicates, triple is known as Bu;Number
It is the behavioral data set that all users generate resource according to collection B;Resource tag data are obtained, for resource I, resource tag collection
TIIndicate tag set<h of resource I itself, i>, data set T2={ Ti: i ∈ I } be all resources label record collection
It closes;
Acquisition user is to the behavior that historical behavior data method caused by resource includes: that a user u generates resource h
B triple<u,h,b>it indicates, triple is known as Bu;Data set B is the behavioral data collection that all users generate resource
It closes;
Obtain resource tag data method include:
For resource I, resource tag collection TIIndicate tag set<h of resource I itself, i>, data set T2={ Ti: i
∈ I } be all resources label record set;
User is calculated to the weighted value of each behavior of resource using the comparison scale method of weighting;The original of analytic hierarchy process (AHP) is pressed first
Some classified variable is then subjected to assignment according to certain rule, then according to the significance level between each class, finally according to
Scale Method from 1 to 9, compares scale one by one, constructs following judgment matrix;It can be marked when the i-th class is compared with jth class
Spend dij, then when jth class and the i-th class compare, dji=1/dij;As i=j, dij=1;The geometry of each row element of calculating matrix is equal
Number, calculates following formula;
Normalized is done, all kinds of weights is obtained, calculation formula is as follows;
The concept of user's co-occurrence is defined, i.e., when two users have a same label or all have row to same resource
For when, be the two user's co-occurrences.
Further, in step 1, be built into user-user network and resource-resource network method the following steps are included:
Step 1, exploitation right rescaling method calculates the weighted value of each behavior of user;
Step 2, the user-user network based on user tag is constructed, by user-user network representation at G='s (V, E)
Weighted undirected graph form, each node on behalf a user, V indicate network in nodal information;E indicates nodes
And the nonoriented edge of node, that is, between user and user, represent co-occurrence information between user.Node ViWith node VjBetween side weight
WI, jThe co-occurrence number for being user i and user j in T;
Step 3, the user-user network based on user behavior is constructed, by user-user network representation at G='s (V, E)
Weighted undirected graph form, each node on behalf a user, V indicate network in nodal information.E indicates nodes
And the nonoriented edge of node, that is, between user and user, represent co-occurrence information between user;Node ViWith node VjBetween side weight
WI, jFor user i and user j to the behavior weight sum of same resource in B;
Step 4, the resource network based on resource tag is constructed, it is similar with the user-user network based on user tag.
Further, in step 2, include: using the method that three classes network calculates separately the feature vector value of user and resource
The first step calculates the single order proximity of three classes nodes.It is fixed to a nonoriented edge (i, j) in network first
Two node v of justiceiAnd vjShared probability such as following formula:
U in formulaiAnd ujIt is the vectorization representation of nodes i and j, describes to save from the angle of Embedding
Intimate degree between point;
W in formulaijIt is the weight on side between node i and j, W refers to the sum of all side right values in network, W=∑(i, j) ∈ Ewij;
By objective function optimization such as following formula
Distributional difference between two probability is by dOFunction is measured, and selects KL divergence, target function type is optimized for following formula
O1=-∑(i, j) ∈ Ewijlog p1(vi, vj);
Finally searching minimizesObtain the single order proximity of each point in network;
Second step calculates the second order proximity of two class user-user nodes;Introduce two vectorsWithWork as ui
When being treated as node,It is uiExpression;Work as uiWhen context processing,It is uiExpression;For side (i, j), by node vi
Generate context vjDefinition of probability.Are as follows:
In formula | V | it is the quantity of vertex or context;Make context after dimensionality reduction Probability p (| ui) connect as far as possible
Nearly actual probabilitiesIt indicates are as follows:
α in formulaiIndicate the significance level of nodes i;Actual probabilities are defined asWherein
wijIt is the weight on side (i, j), N (i) is node uiNeighbouring node collection;It, will using KL- divergence as distance functionIt is optimized for O2=-∑(i, j) ∈ Ewijlog p2(vj|vi);
By the learning training of model, minimizedPass through the d vector tieed upIndicate each top
Point ui;
Third step, each user node uiInsertion vector indicateFor the vector under single order proximityIt is neighbouring with second order
Spend lower vectorLinear combination,
Further, in step 2, the feature vector of user and resource is calculated by cosine similarity, show that user is similar
Matrix and resource similar matrix, specifically include:
Respectively obtain each user node in the user network based on user tag and user behaviorAfterwards, using cosine phase
The similarity between user is calculated like degree, calculation formula is as follows;
Further, according to the whether new user of user, the different ways of recommendation are provided for it.By above-mentioned calculating, it is based on
Vector in the user network (hereinafter referred to as first kind user network) of user tag under the single order of user node and second order proximity
It indicates, the single order and second order proximity of user node in the user network (hereinafter referred to as the second class user network) based on user behavior
Under vector indicate that the vector in the resource network based on resource tag under resource node single order proximity indicates.The method
Include:
If user is the new user in recommender system, the single order and second order of user node in first kind user network are used
Similar matrix between the user that vector under proximity calculates, at the same using under resource node single order proximity in resource network to
Then similar matrix between the resource that amount calculates is recommended using mixed recommendation to user.
If user has behavioral data in recommender system, the single order and two of user node in the second class user network is used
Similar matrix between the user that vector under rank proximity calculates, while using under resource node single order proximity in resource network
Then similar matrix between the resource that vector calculates is recommended using mixed recommendation to user.
Further, the mentioned mixed recommendation of the present invention is based on collaborative filtering, supplemented by content-based recommendation.It is specific to recommend
Steps are as follows:
It step 1, is that user recommends resource, and calculates user u and be selected as its recommendation by the collaborative filtering based on user
The Probability p of resource i in list.Calculation formula is p (u, i)=∑v∈s(u)∩N(h)Sim (u, v) × rvi, wherein sim (u, v) is indicated
Similarity between user u and user v, S (u) are the similar users collection of user u, and N (i) is the user's collection that selected resource i,
rviIndicate that the similar users v of user u generates the weight of behavior to resource h.Recommendation list is obtained from high to low according to Probability p.
Step 2, it by obtained resource similar matrix, is provided the user with using content-based recommendation algorithm
Resource recommendation list.Finally based on collaborative filtering recommending list, based on supplemented by content recommendation lists, by pushing away for two kinds of algorithms
It recommends result and recommends user jointly, while the higher resource of Probability p in collaborative filtering recommending list is also added in final list
Similar resource
Another object of the present invention is to provide described in a kind of implementation based on the resource intelligent recommended method of internet startup disk method
The resource intelligent recommender system based on internet startup disk method.
Another object of the present invention is to provide described in a kind of implementation based on the resource intelligent recommended method of internet startup disk method
The information processing terminal.
Another object of the present invention is to provide a kind of computer readable storage mediums, including instruction, when it is in computer
When upper operation, so that computer executes the resource intelligent recommended method based on internet startup disk method.
In conclusion advantages of the present invention and good effect are as follows:
Internet startup disk method of the invention is the node indicated in network with low-dimensional, dense, real value feature vector, advantage
It is easy for calculating and store, do not need manual extraction feature, and it is empty that heterogeneous information can be projected to the same low-dimensional
Between in go carry out downstream calculating.First in system user tag data, user's historical behavior data in systems and
Resource tag data construct three kinds of networks, the user network respectively based on user tag, the user network based on user behavior
With the resource network based on resource tag;Then user tag network is acted on using internet startup disk method and resource tag network obtains
The similar matrix between user between resource is calculated to the low-dimensional feature vector of user and resource, and using feature vector, to solve
Certainly user and resource are cold-started problem;Finally user behavior is characterized using the second order proximity of LINE model in internet startup disk method
The feature of user node in network expands the neighbour of these nodes by increasing the high-order neighbours (neighbours of such as neighbours) of node
It occupies, to solve the problems, such as sparsity caused by user data is rare.By the above-mentioned mixing proposed algorithm based on internet startup disk method,
Property can be promoted.
The present invention passes through obtained resource similar matrix, is provided the user with using content-based recommendation algorithm
Resource recommendation list.Finally based on collaborative filtering recommending list, based on supplemented by content recommendation lists, by pushing away for two kinds of algorithms
It recommends result and recommends user jointly, while the higher resource of Probability p in collaborative filtering recommending list is also added in final list
Similar resource.
Detailed description of the invention
Fig. 1 is the resource intelligent recommended method flow chart provided in an embodiment of the present invention based on internet startup disk method.
Fig. 2 is the user-user network provided in an embodiment of the present invention based on user tag.
Fig. 3 is the user-user network provided in an embodiment of the present invention based on user behavior.
Fig. 4 is the resource network figure provided in an embodiment of the present invention based on resource tag.
Fig. 5 is F1 value change curve (S (u)=5) provided in an embodiment of the present invention
Fig. 6 is F1 value change curve (S (u)=10) provided in an embodiment of the present invention
Specific embodiment
In order to make the objectives, technical solutions, and advantages of the present invention clearer, with reference to embodiments, to the present invention
It is further elaborated.It should be appreciated that the specific embodiments described herein are merely illustrative of the present invention, it is not used to
Limit the present invention.
Prior art proposed algorithm not can effectively solve the cold start-up of the user in resource recommendation problem and resource cold start-up
Problem, the Deta sparseness that user data rareness easily causes.
To solve the above problems, making description in detail to the present invention below with reference to concrete scheme.
As shown in Figure 1, the resource intelligent recommended method provided in an embodiment of the present invention based on internet startup disk method includes:
Step 1 obtains user tag data.
Recommender system sets the interest tags that every user has oneself selected in registration, for user u, Yong Hubiao
Label collection TuIt can indicate the tag set of user u itself<u,t>, data set T1={ Tu: u ∈ U } be all users label
Set of records ends.
Step 2 obtains user to historical behavior data caused by resource.
User can generate various actions, such as browsing, collection, purchase etc. when using recommender system, and one u pairs of user
The behavior b that resource h is generated can use triple<u,h,b>it indicates, which is known as Bu.Data set B is all users
The behavioral data set that resource is generated.
Step 3 obtains resource tag data.
When resource is online in recommender system, the information of backstage manager's typing resource in background system is had,
In include the resource feature tag.For resource I, resource tag collection TIIt can indicate the label of resource I itself
Set<h, i>, data set T2={ Ti: i ∈ I } be all resources label record set.
User is calculated to the weighted value of each behavior of resource using the comparison scale method of weighting.The original of analytic hierarchy process (AHP) is pressed first
Some classified variable is then subjected to assignment according to certain rule, then according to the significance level between each class, finally according to
Scale Method from 1 to 9, compares scale one by one, constructs following judgment matrix.As shown in Table 1 and Table 2.In the i-th class and jth
Scale d can be obtained when class comparesij, then when jth class and the i-th class compare, dji=1/dij;As i=j, dij=1.Calculating matrix are every
The geometric mean of one row element, shown in calculation formula such as formula (1);
Normalized is done, all kinds of weights is obtained, shown in calculation formula such as formula (2);
The concept of user's co-occurrence is defined, i.e., when two users have a same label or all have row to same resource
For when, then being known as the two user's co-occurrences.
Step 4 constructs the user-user network based on user tag, such as Fig. 2.By user-user network representation at G=
The weighted undirected graph form of (V, E), each node on behalf a user, V indicate network in nodal information.E indicates net
The nonoriented edge of network interior joint and node, that is, between user and user, represents co-occurrence information between user.Node ViWith node VjBetween
The weight W on sideI, jThe co-occurrence number for being user i and user j in T.
Step 5 constructs the user-user network based on user behavior, such as Fig. 3.By user-user network representation at G=
The weighted undirected graph form of (V, E), each node on behalf a user, V indicate network in nodal information.E indicates net
The nonoriented edge of network interior joint and node, that is, between user and user, represents co-occurrence information between user.Node ViWith node VjBetween
The weight W on sideI, jFor user i and user j to the behavior weight sum of same resource in B.
Step 6 constructs the resource network based on resource tag, such as Fig. 4.With the user-user network based on user tag
It is similar.The single order proximity of three classes network is calculated using internet startup disk method, calculating process is as follows;
Two node v are defined to a nonoriented edge (i, j) in network firstiAnd vjShared probability such as formula (3):
U in formula (3)iAnd ujIt is the vectorization representation of nodes i and j, is equivalent to from Embedding's
Angle describes the intimate degree between node.
W in formula (4)ijIt is the weight on side between node i and j, W refers to the sum of all side right values in network, i.e. W=
∑(i, j) ∈ Ewij.For the reliability for guaranteeing single order proximity, p1WithBetween distributional difference it is the smaller the better, then by objective function
Optimization is such as formula (5)
Distributional difference between two probability is by dOFunction is measured, and the present invention selects KL divergence, formula (5) can be optimized
For formula (6).
O1=-∑(i, j) ∈ Ewijlogp1(vi, vj) (6)
(6) formula minimum can be made by finally findingThe single order that can be obtained by each point in network is neighbouring
Degree.Each section in the user network based on user tag and user behavior is calculated separately after obtaining the single order proximity of three classes network
The second order proximity of point.
Second order is neighbouring to assume a user node in user-user network and many connections of other users nodes sharing
Node is similar to each other.Second order proximity is to think have two nodes of identical neighbor node similar to each other with other nodes.?
In this case, each node is also considered as one specific " context ", and assume that there is similar distribution on " context "
Node be similar.Therefore there are two types of roles: specific " context " of node itself and other nodes for each node tool.Draw
Enter two vectorsWithWork as uiWhen being treated as node,It is uiExpression;Work as uiWhen " context " being taken as to handle,It is uiExpression.For side (i, j), by node viGenerate " context " vjDefinition of probability are as follows:
In formula (7) | V | it is the quantity on vertex or " context ".In order to guarantee second order similarity information between node, then should
Make context after dimensionality reduction Probability p (| ui) as close as actual probabilitiesIt indicates are as follows:
Since the importance of network node may be different, so the α in formula (8)iIndicate the weight of nodes i
Want degree.Actual probabilities are defined asWherein wijIt is the weight on side (i, j), N (i) is node uiNeighbouring
Node collection.Here formula (8) is optimized for as distance function by formula (9) using KL- divergence.
O2=-∑(i, j) ∈ Ewijlog p2(vj|vi) (g)
It is available to minimize (9) formula by the learning training of modelCan by d dimension to
AmountIndicate each vertex ui.Finally, each user node uiInsertion vector indicateFor the vector under single order proximityWith vector under second order proximityLinear combination, it may be assumed that
Respectively obtain each user node in the user network based on user tag and user behaviorAfterwards, using cosine phase
The similarity between user, calculation formula such as formula (11) are calculated like degree.Resource network only considers the vector under single order proximity,
Between similarity also use cosine similarity, with calculate user between the method for similarity it is identical.
By above-mentioned calculating, obtains user in the user network (hereinafter referred to as first kind user network) based on user tag and save
The single order of point and the vector expression under second order proximity, the user network (hereinafter referred to as the second class user network) based on user behavior
The single order of middle user node is with the vector expression under second order proximity, resource node single order in the resource network based on resource tag
Vector under proximity indicates.
If user is the new user in recommender system, the single order and second order of user node in first kind user network are used
Similar matrix between the user that vector under proximity calculates, at the same using under resource node single order proximity in resource network to
Then similar matrix between the resource that amount calculates is recommended using mixed recommendation to user.
If user has behavioral data in recommender system, the single order and two of user node in the second class user network is used
Similar matrix between the user that vector under rank proximity calculates, while using under resource node single order proximity in resource network
Then similar matrix between the resource that vector calculates is recommended using mixed recommendation to user.
Mixed recommendation is based on collaborative filtering, supplemented by content-based recommendation.Mainly in the base of collaborative filtering recommending
Content-based recommendation is added on plinth, two kinds of algorithms complement each other, not only remained the characteristic of collaborative filtering but also realized for
Across the classification recommendation resource of user, improves the precision of recommendation.Mixed recommendation flow chart such as Fig. 1, process are as follows.
It is that user recommends resource, and calculates user u and be selected as in the list of its recommendation by the collaborative filtering based on user
The Probability p of resource i.Calculation formula is p (u, i)=∑v∈S(u)∩N(h)Sim (u, v) × rvi, wherein sim (u, v) indicate user u with
Similarity between user v, S (u) are the similar users collection of user u, and N (i) is the user's collection that selected resource i, rviIt indicates
The similar users v of user u generates the weight of behavior to resource h.Recommendation list is obtained from high to low according to Probability p.
By obtained resource similar matrix, the resource provided the user with using content-based recommendation algorithm
Recommendation list.Finally based on collaborative filtering recommending list, supplemented by content recommendation lists, by the recommendation results of two kinds of algorithms
User is recommended jointly, while the similar money of the higher resource of Probability p in collaborative filtering recommending list being also added in final list
Source.
Table 1
Table 2
Prove that (specific embodiment/experiment/emulation credit is analysed/is able to demonstrate that the front experiment number of the invention for part
According to, evidence material, probation report, business data, research and development evidence, business associate evidence etc.)
To verify effectiveness of the invention, the user tag in website of being booked rooms using certain people place, user behavior data and
Source of houses label data is verified.Its user tag, user behavior and source of houses label data are both from actual items.Synthesis is commented
Valence index (F1 value) its calculation formula isWherein p is accuracy rateR is recall rateAs similar users collection S (u)=5, that is, 5 users similar with recommended user is taken to recommend respectively
The 6-10 source of houses provides the source of houses with traditional Collaborative Filtering Recommendation Algorithm using the recommended method in the present invention respectively for user and pushes away
List is recommended, the comprehensive evaluation index (F1 value) of above two algorithm, the result that two kinds of algorithms obtain at this time such as 3 institute of table are finally calculated
Show, it is as shown in Figure 5 to be converted to diagrammatic form.
3 experimental result of table (S (u)=5)
As similar users collection S (u)=10, that is, 10 users similar with recommended user is taken to recommend 6-10 room respectively
Source provides source of houses recommendation list with traditional Collaborative Filtering Recommendation Algorithm using the recommended method in the present invention respectively for user,
The F1 value of above two algorithm is finally calculated, two kinds of algorithms obtain that the results are shown in Table 4 at this time, are converted to diagrammatic form as schemed
Shown in 6.
4 experimental result of table (S (u)=10)
From in experimental result as can be seen that recommended method refers in overall merit than traditional collaborative filtering in the present invention
It puts on more excellent.By above-mentioned experiment, recommended method can relative to the F1 value of traditional Collaborative Filtering Recommendation Algorithm in the present invention
Promote 20%.
In the above-described embodiments, can come wholly or partly by software, hardware, firmware or any combination thereof real
It is existing.When using entirely or partly realizing in the form of a computer program product, the computer program product include one or
Multiple computer instructions.When loading on computers or executing the computer program instructions, entirely or partly generate according to
Process described in the embodiment of the present invention or function.The computer can be general purpose computer, special purpose computer, computer network
Network or other programmable devices.The computer instruction may be stored in a computer readable storage medium, or from one
Computer readable storage medium is transmitted to another computer readable storage medium, for example, the computer instruction can be from one
A web-site, computer, server or data center pass through wired (such as coaxial cable, optical fiber, Digital Subscriber Line (DSL)
Or wireless (such as infrared, wireless, microwave etc.) mode is carried out to another web-site, computer, server or data center
Transmission).The computer-readable storage medium can be any usable medium or include one that computer can access
The data storage devices such as a or multiple usable mediums integrated server, data center.The usable medium can be magnetic Jie
Matter, (for example, floppy disk, hard disk, tape), optical medium (for example, DVD) or semiconductor medium (such as solid state hard disk Solid
State Disk (SSD)) etc..
The foregoing is merely illustrative of the preferred embodiments of the present invention, is not intended to limit the invention, all in essence of the invention
Made any modifications, equivalent replacements, and improvements etc., should all be included in the protection scope of the present invention within mind and principle.
Claims (10)
1. a kind of resource intelligent recommended method based on internet startup disk method, which is characterized in that the money based on internet startup disk method
Source intelligent recommendation method includes:
Step 1 utilizes user tag data, the historical behavior data and resource tag data of user in systems in system
Construct the user network based on user tag, the user network based on user behavior and the resource network based on resource tag;
Step 2 acts on user tag network using internet startup disk method and resource tag network obtains the low-dimensional of user and resource
Feature vector, and the similar matrix between user between resource is calculated using feature vector;
Step 3 utilizes the spy of user node in the second order proximity characterization user behavior network of LINE model in internet startup disk method
Sign, and the high-order neighbours by increasing user node expand the neighbours of user node;Whether it is new user according to user, carries out not
With the recommendation of mode.
2. as described in claim 1 based on the resource intelligent recommended method of internet startup disk method, which is characterized in that
In step 1, the method for obtaining user tag data includes: for user u, user tag collection TuIndicate a user u certainly
The tag set of body<u,t>, data set T1={ Tu: u ∈ U } be all users label record set;Obtain user behavior number
According to the behavior b triple that a user u generates resource h<u,h,b>it indicates, triple is known as Bu;Data set B is institute
The behavioral data set for thering is user to generate resource;Resource tag data are obtained, for resource I, resource tag collection TlIndicate one
Tag set<h of a resource I itself, i>, data set T2={ Ti: i ∈ I } be all resources label record set;
Obtaining user includes: that a user u uses the behavior b that resource h is generated to historical behavior data method caused by resource
Triple<u,h,b>it indicates, triple is known as Bu;Data set B is the behavioral data set that all users generate resource;
Obtain resource tag data method include:
For resource I, resource tag collection TIIndicate tag set<h of resource I itself, i>, data set T2={ Ti: i ∈ I }
It is the label record set of all resources;
User is calculated to the weighted value of each behavior of resource using the comparison scale method of weighting;It first will by the principle of analytic hierarchy process (AHP)
Some classified variable carries out assignment according to certain rule, then according to the significance level between each class, finally according to from 1 to
9 Scale Method, compares scale one by one, constructs following judgment matrix;Scale d can be obtained when the i-th class is compared with jth classij,
When then jth class and the i-th class compare, dji=1/dij;As i=j, dij=1;The geometric mean of each row element of calculating matrix, meter
Calculate following formula;
Normalized is done, all kinds of weights is obtained, calculation formula is as follows;
The concept of user's co-occurrence is defined, i.e., when two users have a same label or all have behavior to same resource
When, it is the two user's co-occurrences.
3. as described in claim 1 based on the resource intelligent recommended method of internet startup disk method, which is characterized in that in step 1,
Be built into user-user network and resource-resource network method the following steps are included:
Step 1, exploitation right rescaling method calculates the weighted value of each behavior of user;
Step 2, the user-user network based on user tag is constructed, by user-user network representation at the weighting of G=(V, E)
Undirected diagram form, each node on behalf a user, V indicate network in nodal information;E indicates nodes and section
Nonoriented edge of the point i.e. between user and user, represents co-occurrence information between user.Node ViWith node VjBetween side weight WI, j
The co-occurrence number for being user i and user j in T;
Step 3, the user-user network based on user behavior is constructed, by user-user network representation at the weighting of G=(V, E)
Undirected diagram form, each node on behalf a user, V indicate network in nodal information.E indicates nodes and section
Nonoriented edge of the point i.e. between user and user, represents co-occurrence information between user;Node ViWith node VjBetween side weight WI, j
For user i and user j to the behavior weight sum of same resource in B;
Step 4, the resource network based on resource tag is constructed, it is similar with the user-user network based on user tag.
4. as described in claim 1 based on the resource intelligent recommended method of internet startup disk method, which is characterized in that in step 2,
Include: using the method that three classes network calculates separately the feature vector value of user and resource
The first step calculates the single order proximity of three classes nodes.Two are defined to a nonoriented edge (i, j) in network first
A node viAnd vjShared probability such as following formula:
U in formulaiAnd ujThe vectorization representation of nodes i and j, described from the angle of Embedding node it
Between intimate degree;
W in formulaijIt is the weight on side between node i and j, W refers to the sum of all side right values in network, W=∑(i, j) ∈ Ewij;By mesh
Scalar functions optimization such as following formula
Distributional difference between two probability is measured by d () function, selects KL divergence, target function type is optimized for following formula
O1=-∑(i, j) ∈ Ewijlogp1(vi, vj);
Finally searching minimizesObtain the single order proximity of each point in network;
Second step calculates the second order proximity of two class user-user nodes;Introduce two vectorsWithWork as uiLocated
When reason is node,It is uiExpression;Work as uiWhen context processing,It is uiExpression;For side (i, j), by node viIt generates
Context vjDefinition of probability.Are as follows:
In formula | V | it is the quantity of vertex or context;Make context after dimensionality reduction Probability p (| ui) as close as reality
ProbabilityIt indicates are as follows:
α in formulaiIndicate the significance level of nodes i;Actual probabilities are defined asWherein wijIt is
The weight on side (i, j), N (i) are node uiNeighbouring node collection;It, will using KL- divergence as distance functionIt is optimized for O2=-∑(i, j) ∈ Ewijlogp2(vj|vi);
By the learning training of model, minimizedPass through the d vector tieed upIndicate each vertex
ui;
Third step, each user node uiInsertion vector indicateFor the vector under single order proximityUnder second order proximity
VectorLinear combination,
5. as described in claim 1 based on the resource intelligent recommended method of internet startup disk method, which is characterized in that in step 2,
The feature vector of user and resource is calculated by cosine similarity, obtains user's similar matrix and resource similar matrix, specifically
Include:
Respectively obtain each user node in the user network based on user tag and user behaviorAfterwards, using cosine similarity
The similarity between user is calculated, calculation formula is as follows;
6. as claimed in claim 5 based on the resource intelligent recommended method of internet startup disk method, which is characterized in that be according to user
No new user, provides the different ways of recommendation;By using cosine similarity to calculate the similarity calculation between user, it is based on
Vector in the first kind user network of user tag under the single order of user node and second order proximity indicates, is based on user behavior
The second class user network in user node single order and second order proximity under vector indicate, the resource network based on resource tag
Vector in network under resource node single order proximity indicates;It specifically includes:
It is neighbouring using the single order and second order of user node in first kind user network if user is the new user in recommender system
Similar matrix between the user that the lower vector of degree calculates, at the same in use resource network under resource node single order proximity to meter
Then similar matrix between the resource of calculation is recommended using mixed recommendation to user;
It is adjacent using the single order and second order of user node in the second class user network if user has behavioral data in recommender system
Similar matrix between the user that vector under recency calculates, while using the vector under resource node single order proximity in resource network
Then similar matrix between the resource of calculating is recommended using mixed recommendation to user.
7. as claimed in claim 6 based on the resource intelligent recommended method of internet startup disk method, which is characterized in that: mixed recommendation
Based on collaborative filtering, supplemented by content-based recommendation;The following steps are included:
Step I is that user recommends resource, and calculates the column that user u is selected as user u recommendation by the collaborative filtering based on user
The Probability p of resource i in table;Calculation formula is p (u, i)=∑v∈S(u)∩N(h)Sim (u, v) × rvi, wherein sim (u, v) indicates to use
Similarity between family u and user v, S (u) are the similar users collection of user u, and N (i) is the user's collection that selected resource i, rvi
Indicate that the similar users v of user u generates the weight of behavior to resource h;Recommendation list is obtained from high to low according to Probability p.
Step II is provided the user with by obtained resource similar matrix using content-based recommendation algorithm
Resource recommendation list, the core concept of content-based recommendation algorithm are exactly the feature by analyzing article itself, then basis
User preferences, which are recorded, recommends similar article in hobby record to it;Finally based on collaborative filtering recommending list, it is based on content
Supplemented by recommendation list, user, while final column will be recommended jointly based on content and the recommendation results based on collaborative filtering
Also the similar resource of the higher resource of Probability p in collaborative filtering recommending list is added in table.
8. it is a kind of implement the resource intelligent recommended method based on internet startup disk method described in claim 1 based on internet startup disk method
Resource intelligent recommender system.
9. a kind of information for implementing the resource intelligent recommended method based on internet startup disk method described in claim 1~7 any one
Processing terminal.
10. a kind of computer readable storage medium, including instruction, when run on a computer, so that computer executes such as
Based on the resource intelligent recommended method of internet startup disk method described in claim 1-7 any one.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910451324.XA CN110162711B (en) | 2019-05-28 | 2019-05-28 | Resource intelligent recommendation method and system based on network embedding method |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910451324.XA CN110162711B (en) | 2019-05-28 | 2019-05-28 | Resource intelligent recommendation method and system based on network embedding method |
Publications (2)
Publication Number | Publication Date |
---|---|
CN110162711A true CN110162711A (en) | 2019-08-23 |
CN110162711B CN110162711B (en) | 2023-07-28 |
Family
ID=67629481
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201910451324.XA Active CN110162711B (en) | 2019-05-28 | 2019-05-28 | Resource intelligent recommendation method and system based on network embedding method |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN110162711B (en) |
Cited By (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN111125556A (en) * | 2019-12-25 | 2020-05-08 | 青梧桐有限责任公司 | Recommendation sorting method and system, electronic device and readable storage medium |
CN111143714A (en) * | 2019-12-25 | 2020-05-12 | 青梧桐有限责任公司 | Recommendation sorting method and system, electronic device and readable storage medium |
CN111325578A (en) * | 2020-02-20 | 2020-06-23 | 深圳市腾讯计算机系统有限公司 | Prediction model sample determination method, prediction model sample determination device, prediction model sample determination medium, and prediction model sample determination device |
CN111931041A (en) * | 2020-07-03 | 2020-11-13 | 武汉卓尔数字传媒科技有限公司 | Label recommendation method and device, electronic equipment and storage medium |
CN113641791A (en) * | 2021-08-12 | 2021-11-12 | 卓尔智联(武汉)研究院有限公司 | Expert recommendation method, electronic device and storage medium |
CN114331142A (en) * | 2021-12-30 | 2022-04-12 | 上海淇玥信息技术有限公司 | Internet service determination method and device based on enterprise relational network |
Citations (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20100161619A1 (en) * | 2008-12-18 | 2010-06-24 | Lamere Paul B | Method and Apparatus for Generating Recommendations From Descriptive Information |
US20120233191A1 (en) * | 2010-11-22 | 2012-09-13 | Salesforce.Com, Inc. | Method and system for making content-based recommendations |
CN103959329A (en) * | 2011-11-23 | 2014-07-30 | 皇家飞利浦有限公司 | Image domain de-noising |
CN107491540A (en) * | 2017-08-24 | 2017-12-19 | 济南浚达信息技术有限公司 | A kind of combination depth Bayesian model and the film of collaboration Heterogeneous Information insertion recommend method |
CN107633100A (en) * | 2017-10-23 | 2018-01-26 | 苏州大学 | A kind of point of interest based on incorporation model recommends method and device |
US20190079977A1 (en) * | 2016-05-12 | 2019-03-14 | Alibaba Group Holding Limited | Method for determining user behavior preference, and method and device for presenting recommendation information |
-
2019
- 2019-05-28 CN CN201910451324.XA patent/CN110162711B/en active Active
Patent Citations (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20100161619A1 (en) * | 2008-12-18 | 2010-06-24 | Lamere Paul B | Method and Apparatus for Generating Recommendations From Descriptive Information |
US20120233191A1 (en) * | 2010-11-22 | 2012-09-13 | Salesforce.Com, Inc. | Method and system for making content-based recommendations |
CN103959329A (en) * | 2011-11-23 | 2014-07-30 | 皇家飞利浦有限公司 | Image domain de-noising |
US20190079977A1 (en) * | 2016-05-12 | 2019-03-14 | Alibaba Group Holding Limited | Method for determining user behavior preference, and method and device for presenting recommendation information |
CN107491540A (en) * | 2017-08-24 | 2017-12-19 | 济南浚达信息技术有限公司 | A kind of combination depth Bayesian model and the film of collaboration Heterogeneous Information insertion recommend method |
CN107633100A (en) * | 2017-10-23 | 2018-01-26 | 苏州大学 | A kind of point of interest based on incorporation model recommends method and device |
Non-Patent Citations (1)
Title |
---|
杨宇凌;王澳蓉;吴浩;董琳;何鹏;: "基于用户关系网络表征学习的服务推荐方法", 重庆大学学报, no. 07 * |
Cited By (7)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN111125556A (en) * | 2019-12-25 | 2020-05-08 | 青梧桐有限责任公司 | Recommendation sorting method and system, electronic device and readable storage medium |
CN111143714A (en) * | 2019-12-25 | 2020-05-12 | 青梧桐有限责任公司 | Recommendation sorting method and system, electronic device and readable storage medium |
CN111325578A (en) * | 2020-02-20 | 2020-06-23 | 深圳市腾讯计算机系统有限公司 | Prediction model sample determination method, prediction model sample determination device, prediction model sample determination medium, and prediction model sample determination device |
CN111325578B (en) * | 2020-02-20 | 2023-10-31 | 深圳市腾讯计算机系统有限公司 | Sample determination method and device of prediction model, medium and equipment |
CN111931041A (en) * | 2020-07-03 | 2020-11-13 | 武汉卓尔数字传媒科技有限公司 | Label recommendation method and device, electronic equipment and storage medium |
CN113641791A (en) * | 2021-08-12 | 2021-11-12 | 卓尔智联(武汉)研究院有限公司 | Expert recommendation method, electronic device and storage medium |
CN114331142A (en) * | 2021-12-30 | 2022-04-12 | 上海淇玥信息技术有限公司 | Internet service determination method and device based on enterprise relational network |
Also Published As
Publication number | Publication date |
---|---|
CN110162711B (en) | 2023-07-28 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
Moosavi et al. | Community detection in social networks using user frequent pattern mining | |
Vega-Oliveros et al. | A multi-centrality index for graph-based keyword extraction | |
Bedi et al. | Community detection in social networks | |
Li et al. | Opinion community detection and opinion leader detection based on text information and network topology in cloud environment | |
Mao et al. | Multiobjective e-commerce recommendations based on hypergraph ranking | |
CN110162711A (en) | A kind of resource intelligent recommended method and system based on internet startup disk method | |
CN103577549B (en) | Crowd portrayal system and method based on microblog label | |
Zhang et al. | User community discovery from multi-relational networks | |
Huang et al. | Exploring expert cognition for attributed network embedding | |
Guo et al. | Exploiting pre-trained network embeddings for recommendations in social networks | |
Huang et al. | Joint weighted nonnegative matrix factorization for mining attributed graphs | |
CN112633973A (en) | Commodity recommendation method and related equipment thereof | |
Dougnon et al. | Inferring social network user profiles using a partial social graph | |
Peng et al. | Multi-level preference regression for cold-start recommendations | |
Chen et al. | Dynamic evolutionary clustering approach based on time weight and latent attributes for collaborative filtering recommendation | |
CN110069713A (en) | A kind of personalized recommendation method based on user's context perception | |
Singh et al. | A survey on the generation of recommender systems | |
KR102498294B1 (en) | Method and device for extracting key keywords based on keyword joint appearance network | |
Xu et al. | Latent interest and topic mining on user-item bipartite networks | |
Liu | Personalized recommendation system design for library resources through deep belief networks | |
Shirgahi et al. | Trust estimation of the semantic web using semantic web clustering | |
Yuan et al. | Sign prediction on unlabeled social networks using branch and bound optimized transfer learning | |
Zhang et al. | Matrix factorization meets social network embedding for rating prediction | |
Zhang et al. | Evaluation for machine tool components importance based on improved LeaderRank | |
Ye et al. | CDCN: A New NMF‐Based Community Detection Method with Community Structures and Node Attributes |
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 |