CN105912646A - Keyword retrieval method based on diversity and proportion characteristics - Google Patents

Abstract

The present invention relates to a keyword retrieval method based on diversity and proportion characteristics. For a keyword and the natural number 1 which are input by a user, according to a link relation between the keyword and object information, an algorithm is utilized to return one piece of most comprehensive object information based on the keyword to the user. The keyword retrieval method comprises the steps of (1) designing static off-line ordering evaluation scores according to a link analysis algorithm PageRank, and generating initial values of all nodes; (2) inputting a keyword to generate an alternative OS; and (3) inputting the natural number 1, and generating a DS-rooted tree finally containing a node by a k LASP algorithm according to the obtained OS. The experimental results show that the experimental effects obtained by the method are significant.

Description

A kind of keyword retrieval method based on multiformity and proportionality

Technical field

The invention belongs to Data Mining, relate to a kind of keyword retrieval method based on multiformity and proportionality.

Background technology

Along with the development of the Internet, search engine brings huge as a kind of novel Network retrieval technology to user Convenient.But due to developing rapidly of network in recent years, significantly increasing occurs in the quantity of information of the Internet, and big data are as one Individual emerging field is flooded with life, and this allows for user when facing this substantial amounts of information, and search engine possibly cannot be recommended Go out information diversified, that arrange based on keyword retrieval by significance level.One that solves this problem has latent profit very much Method be to provide an arranging system, it can to return l bar important information, (wherein l be according to the key word that user is given Natural number), and arrange by multiformity and proportionality.

This technology introduces tuple-set (ObjectSummaries is abbreviated as OS), and it is to comprise the data of key word The set of the information tuple based on key word generated in storehouse.One OS can be with key word as root, adjacent with key word Node is the tree structure of its descendant nodes.In order to generate OS, one is intended to have about inquiry data subject (Data Subjects, is abbreviated as DS) relation of information, this relation is abbreviated as R^DS, it is i.e. the root of tree structure；Another need with R^DSThe relation of link, namely generates R^DSDescendants.For each R^DSFor can form a DS ideograph, namely G^DS.This technology is constantly to carry out beta pruning optimization according to the OS generated finally to draw important information.

One complete OS may have thousands of bar tuple information, all enumerate out by these information and not only can disappear Consume the more time, and it is also extremely difficult that user chooses information wherein useful for oneself, so selecting Choose the tuple information that l bar is most useful；Natural number l to input, (refers to step by using k-LASP algorithm in whole OS 3.3) l bar important information (i.e. size-l OS) is obtained, if the quiescent value that light uses PageRank or ValueRank to calculate is come Return information, then may make a plurality of similar information repeat, so in order to enable this l bar information going up to greatest extent Present to the more diversified information of user, allow users to more fully understand information, introduce multiformity (Dsize-l) and ratio The method of characteristic (Psize-l) two kinds balance information importance.This method can not only greatly reduce the consumption of time, improves The efficiency of return information, and disclosure satisfy that user to search information diversified demand, optimize to a certain extent based on The search of key word.

Summary of the invention

The present invention provides a kind of keyword retrieval method based on multiformity and proportionality, the key being inputted user Word and natural number l, then according to the linking relationship between key word and each tuple information, use algorithm to return to user's l bar Comprehensively tuple information based on key word.

A kind of keyword retrieval method based on multiformity and proportionality, the steps include:

Step one: inspired by link analysis technology PageRank, design static off-line sequence evaluation score, generate all The initial value of node；

Step 1.1: collect and disposal data collection, builds data relationship.At this moment definition directed graph G (V, E), wherein V (v₁,...,v_n) it is node (summit) collection, node on behalf various information here, E is the set of representative edge (arc), E={ < v_i, v_j>|v_i,v_j∈ V}, < v_i,v_j> represent from v_iTo v_jA limit (arc), i.e. v_iInformation can be linked to v_j；

Step 1.2:r is a vector (queue of the evaluation score of each the page), the most each node v_iAll exist Corresponding r_i, then the evaluation score of iterative computation vector r is carried out by below equation:

$r=dAr+(1-d)\frac{e}{\left|V\right|}---\left(1\right)$

Wherein d is the damped coefficient of (0,1), and this coefficient ensure that and obtains more accurate result, and general value is 0.85；A is a n*n matrix, and n represents number of vertices, if wherein existing from v_iTo v_jLimit (arc), then(O(v_j) Represent v_jOut-degree), be otherwise 0, say, that if there being three nodes, then A is a 3*3 matrix, v₀To v₁And v₂Have limit and v₁To v₂There is limit, thenAnd A₂₁=1, remaining is all 0；E=[1....1]^T；| V | is number of vertices.

To sum up, iterative computation goes out the evaluation score of each node in data set, and at this moment this value is referred to as overall situation weights (globalimportance is abbreviated as gi), i.e. gi (v_i) represent v_iThe initial value of node.

Step 2: input key word generates alternative OS；

Step 2.1: input key word (i.e. DS), system generation one is (i.e. R with DS summit as root node^DS), with energy and R^DS The tree that relation is descendants of link, i.e. OS.In order to distinguish each unit group node v in OS during generating OS_iImportant Property, is by this tuple overall weights in data base by a local weight (local importance is abbreviated as li) (gi) and this tuple in OS and R^DSAffinity (Affinity is abbreviated as Af) two parts determined；

Step 2.2: in generating OS, G^DSIn with R^DSThe relation having higher affinity will be added in OS, R_iTo R^DS's Affinity Af (R_i) by below equation iterative computation:

$Af\left(R_i\right)=\underset{j}{\&Sigma;}{w}_j{m}_j\&CenterDot;Af\left(R_{Parent}\right)---\left(2\right)$

Wherein j is a scope, and this scope is index set (m₁,m₂,...,m_n) and its corresponding weights set (w₁,w₂,...,w_n), four indexs of consideration here: index m₁For R_iTo R^DSDistance, the namely distance between two relations The least, affinity is the highest；Index m₂For the relative radix of relation, namely R_iWith R_PatentIn the average unit that is connected of each tuple The quantity of group；Index m₃Anti-phase to radix for relation, i.e. R_PatentWith R_iIn a tuple be connected par；Index m₄ For R_iThe connectedness of pattern, i.e. R_iThe quantity of the link in graph of a relation.Af(R_Parent) refer to R_iFather's node and R^DS's Affinity, initial value is 1, i.e. R^DSThe affinity of itself is 1.The fraction range of index is [0,1], and the summation of corresponding weights is 1 (the corresponding weights of aforementioned four index are all 0.25).And in the generation of OS, the affinity of all relation nodes all should Higher than marginal value θ；

Step 2.3: the formula of the importance Im (S) calculating alternative size-l OS S is:

$\mathrm{Im}\left(S\right)=\underset{n_i\&Element;S}{\&Sigma;}\mathrm{Im}(OS,R_i)---\left(3\right)$

Wherein Im (OS, R_i) it is OS interior joint R_iLi value, Im (OS, R_i) can be calculated by below equation:

Im(OS,R_i)=Im (R_i)·Af(R_i) (4)

Wherein, Im (R_i) it is R_iGl value, Af (R_i) it is R_iTo R^DSAffinity.

To sum up calculate Im value according to the key word of input, generate alternative OS.

Step 3: input natural number l generates final containing l according to OS k-LASP algorithm (referring to step 3.3) obtained The tree with DS as root of individual node.Three factors will be considered in this step: multiformity amount of attenuation (dv), proportionality increment (pv) and quiescent value (li), they are respectively in connection with drawing a last mark (i.e. dw, pw) the most at last.

Step 3.1: multiformity (Dsize-l)

In order to avoid repeating of the too high analog information of importance, should select to export the diversified information of l bar, so Provide the computational methods of a following multiformity amount of attenuation:

$dv\left(v_i\right)=1-\frac{z\left(g\right(v_i))-1}{l-1}---\left(5\right)$

Wherein, g (v_i) refer to and v_iSimilar first group node；z(g(v_i))-1 refer in size-l OS and v_iNode phase As unit group node summation；z(g(v_i)) refer to g (v_i) number of times in size-l OS to be occurred in.dv(v_i) codomain be [0,1].Definition dv [z] is that node occurs that in size-l OS the multiformity of z time weakens value, and example makes l=10, and " Marry " goes out Existing 2 times, i.e. z=2, then

Then, the multiformity weights that a node static value in Dsize-lOS is combined with multiformity weakening value are by such as Lower formula calculates:

dw(v_i)=li (v_i)·dv(v_i) (6)

To sum up, providing OS and l, generating a Dsize-l OS needs to meet following condition:

1) the tuple number in Dsize-lOS is l (l≤| OS |)；

2) this l node all must be connected with root node；

3) each node v_iThere is corresponding multiformity weights i.e. dw (v_i)；

4) a Dsize-l OS collect to be divided into

$\mathrm{Im}\left(DSl\right)=\underset{v_i\&Element;DSl}{\&Sigma;}dw\left(v_i\right)---\left(7\right)$

Step 3.2: proportionality (Psize-l)

In view of in an OS, a first group node may occur in multiple times, but these nodes may have more weak quiet State value, and their frequency has contact important with DS, thus, proportionality in actual size-l OS to be obtained Increment size can be calculated by equation below:

$pq\left(v_i\right)=\frac{fr\left(g\right(v_i))}{\mathrm{\&alpha;}\&CenterDot;z\left(g\right(v_i))+1}---\left(8\right)$

Wherein, fr (g (v_i)) it is g (v_i) occur in the number of times in OS；z(g(v_i)) refer to g (v_i) size-l to be occurred in Number of times in OS；α is a constant that can adjust ratio, typically takes α=2.

Then the proportionality weights that a node static value in Psize-l OS is combined with proportionality increment size by Equation below calculates:

pw(v_i)=li (v_i)·pq(v_i) (9)

To sum up, providing OS and l, generating a Psize-l OS needs to meet following condition:

1) the tuple number in Psize-l OS is l (l≤| OS |)

2) this l node all must be connected with root node

3) each node v_iThere is corresponding multiformity weights i.e. pw (v_i)

4) a Psize-l OS collect to be divided into

$\mathrm{Im}\left(PSl\right)=\underset{v_i\&Element;PSl}{\&Sigma;}pw\left(v_i\right)---\left(10\right)$

Step 3.3: generate the final tree with DS as root containing l node with k-LASP algorithm；

The maximum average value path of k-LASP (k-Largest Averaged Score Path) i.e. k node is (namely The meansigma methods of k node weights on one paths), in this step, dw and pw primary system is referred to as weight w；Each in OS Individual node v_iThere is a weight w (v_i), corresponding v_iWith its ancestor node (number n, n=max (k 1, actual (tube) length Degree)) average weight be defined asDuring generating OS, need a Hash table, use HFr table Showing, HFr includes three parts, and one is by v_iIn i as the numbering of node of graph, two is v_iNumber of times fr (the v occurred in OS_i), Three is v_iNumber of times z (the v occurred in size-l OS_i)；In order to preferably manage OS interior joint and corresponding AP value, set up one Queue W preserves these information, and in this queue, the order of node is successively decreased arrangement by corresponding AP value.

K-LASP algorithm generates the process of size-l OS:

1) OS is generated, including building HFr, calculating AP (v_i) and generate W

2) if | size-l | < l, 3 are turned), otherwise turn 11)

3)p_iRepresent the node having maximum AP value in current W to the path of root node,

By p_iIn the front individual node of l-| size-l | join in size-l OS

4) if | size-l | < l, 5 are turned), otherwise turn 10)

5) by selected p_iIn the individual node of l-| size-l | remove from OS and W

6) for p_iDescendant nodes (number n, n=max (k 1, the physical length)) v of each node_jDo and update as follows:

AP (v is updated in OS and W_j) value

7) for p_iIn each node g (v), if g (v) is at HFr, turn 8), otherwise turn 10)；

8)HFr(g(v)).z++

9) the node n for the g of making (n)=g (v) each in OS does and updates as follows:

For each node n in each subtree with node n as root_iDo:

In OS and W, AP (n is updated by HFr (g (v)) .z value_i) value

9) 2 are turned)

11) size-lOS is returned

The l bar tuple information that will retrieve that the size-lOS now returned is the most required.

Through the results show, the experiment effect that this method obtains is notable.

Accompanying drawing explanation

Fig. 1: system flow chart.

Fig. 2: system results figure.

Fig. 3: DBLP database schema figure.

The G generated according to author's tuple information in Fig. 4: DBLP^DS。

Fig. 5: k-LASP (k=2) algorithm implementation example figure.

Detailed description of the invention

Below in conjunction with relevant drawings the present invention explained and illustrate:

The data set that the present invention uses is DBLP data base, and DBLP is that the interior achievement to research of computer realm is with author The integrated database system of one computer english literature of core chronologically lists the scientific achievement of author, including the world The paper that periodical and meeting etc. are published.Its database schema figure is shown in accompanying drawing 3.

Step one: inspired by link analysis technology PageRank, design static off-line sequence evaluation score, generate all The initial value of node；

The initial value of each node of data set is calculated according to formula (1).

Step 2: input key word generates alternative OS

The present invention uses DBLP data base, and the searching keyword of input is " Michalis Faloutsos ", according to formula (4) quiescent value of each node band affinity, i.e. Im are calculated.The G generated according to author's tuple information in DBLP^DSSee accompanying drawing 4, bracket In numerical value represent be and R^DSCohesion (marginal value θ=0.7 of selected node).Based on G^DSThe portion generated according to Im value Divide alternative OS as shown in the table:

Part OS that tuple generates inquired about for " Michalis Faloutsos " by table 1

Step 3: input natural number l according to what the OS k-LASP algorithm generation obtained finally contained l node with DS is The tree of root.

Step 3.1: multiformity (Dsize-l)

As a example by author, l=10, following table is made to provide quiescent value li according to author's tuple information, according to formula (5), (6) The result (successively decreasing arrangement by dw [1] value) gone out:

Table 2 multiformity based on author information tuple weights

The weights calculating gained according to this table are seen, when C.Faloutsos and M.Mitzenmacher respectively occurs once Weights are respectively 1.8 and 1.4, but when the weights when C.Faloutsos occurs three times occur one time with M.Mitzenmacher Weights be equal be all 1.4, thus result can ensure that the multiformity of output tuple information, it is to avoid similar information Repeat.

Step 3.2: proportionality (Psize-l)

Or as a example by above-mentioned author, make l=10, following table provide quiescent value li according to author's tuple information and appearance Frequency fr, according to formula (8), the result (successively decreasing arrangement by pw [1] value) gone out of (9):

Table 3 proportionality based on author information tuple weights

The weights calculating gained according to this table are seen, quiescent value original for S.Krishnamurthy and C.Faloutsos is divided 0.6 and 1.8, between differ 1.2, but the frequency that S.Krishnamurthy is than C.Faloutsos many 25, recognize thus For having more important than C.Faloutsos, so can find out when they appear at Psize-l OS by S.Krishnamurthy The when of three times, their weights only differ from 0.1, and S.Krishnamurthy is higher than C.Faloutsos.One first group node, It may occur in multiple times, but these nodes may have a more weak quiescent value, and its frequency is actual to be obtained Having contact important with DS in size-l OS, thus result can ensure that this yuan of group node can have at Psize-l OS Individual more suitably position.

Step 3.3: generate the final tree with DS as root containing l node with k-LASP algorithm

As a example by k=2,2-LASP (2-LargestAveragedScore Path), i.e. the maximum average value of two nodes Path, in this step, is referred to as weight w by dw and pw primary system；Each node v in OS_iThere is a weight w (v_i), therewith Corresponding v_iIt is defined as AP (v with the average weight of its father node_i)；During generating OS, need a Hash table, use HFr Representing, HFr includes three parts, and one is by v_iIn i as the numbering of node of graph, two is v_iThe number of times fr occurred in OS (v_i), three is v_iNumber of times z (the v occurred in size-l OS_i)；In order to preferably manage OS interior joint and corresponding AP value, build A vertical queue W preserves these information, and in this queue, the order of node is successively decreased arrangement by corresponding AP value.

As a example by accompanying drawing 5, Fig. 5 (A) is initial OS, the l=5 of order input.

According to 2-LASP algorithm, first select the node i.e. n that in W, weights are the highest₁₁, select n₁₁To this path of root node p₁, path p₁In have 3 nodes, first these 3 nodes are added in size-l OS；According to algorithm the 4th step, 0 < 5, forward the 5th to Step, removes these three node from OS and W；Then the child nodes of these three node is updated AP (n_i) value:

According to algorithm the 7th step, HFr has node v_iSo forwarding the 8th step to by HFr (g (v)) .z++, thenFurther according to formulaUpdate AP (n_i) Value, asOther nodes are similar to.

To sum up complete to update for the first time, shown in result such as Fig. 5 (B), then carry out next update further according to above-mentioned algorithm, Shown in final updated result such as Fig. 5 (C).So obtain final size-5OS.

Also by searching keyword be as a example by " Michalis Faloutsos ", the size-15OS drawn see accompanying drawing 2 (A) and Dsize-15OS is shown in that accompanying drawing 2 (B), Psize-15OS are shown in accompanying drawing 2 (C).Through experimental investigation gained Dsize-15OS and Psize- 15OS more meets user's request.

Claims (1)

1. a keyword retrieval method based on multiformity and proportionality, it is characterised in that: the enforcement step of the method is such as Under,

Step one: inspired by link analysis technology PageRank, design static off-line sequence evaluation score, generate all nodes Initial value；

Step 1.1: collect and disposal data collection, builds data relationship；At this moment definition directed graph G (V, E), wherein V (v₁,..., v_n) it is set of node, node on behalf various information here, E is the set of representative edge, E={ < v_i,v_j>|v_i,v_j∈ V}, < v_i,v_j > represent from v_iTo v_jLimit, i.e. a v_iInformation can be linked to v_j；

Step 1.2:r is a vector i.e. queue of the evaluation score of each the page, the most each node v_iAll exist corresponding r_i, then the evaluation score of iterative computation vector r is carried out by below equation:

$r=dAr+(1-d)\frac{e}{\left|V\right|}---\left(1\right)$

Wherein d is the damped coefficient of (0,1), and this coefficient ensure that and obtains more accurate result, and general value is 0.85； A is a n*n matrix, and n represents number of vertices, if wherein existing from v_iTo v_jLimit (arc), thenRepresent v_j Out-degree), be otherwise 0, say, that if there being three nodes, then A is a 3*3 matrix, v₀To v₁And v₂There are limit and v₁To v₂ There is limit, thenAnd A₂₁=1, remaining is all 0；E=[1....1]^T；| V | is number of vertices；

To sum up, iterative computation goes out the evaluation score of each node in data set, and at this moment this value is referred to as the overall situation weights, i.e. gi (v_i) represent v_iThe initial value of node；Overall situation weights global importance, is abbreviated as gi；

Step 2: input key word generates alternative OS；

Step 2.1: input key word (i.e. DS), system generation one is (i.e. R with DS summit as root node^DS), with energy and R^DSLink The tree that relation is descendants, i.e. OS；In order to distinguish each unit group node v in OS during generating OS_iImportance, will One local weight (local importance, be abbreviated as li) be by this tuple overall weights (gi) in data base and This tuple in OS and R^DSAffinity two parts determined；Affinity is Affinity, is abbreviated as Af；

Step 2.2: in generating OS, G^DSIn with R^DSThe relation having higher affinity will be added in OS, R_iTo R^DSAffine Degree Af (R_i) by below equation iterative computation:

$Af\left(R_i\right)=\underset{j}{\&Sigma;}{w}_j{m}_j\&CenterDot;Af\left(R_{Parent}\right)---\left(2\right)$

Wherein j is a scope, and this scope is index set (m₁,m₂,...,m_n) and its corresponding weights set (w₁, w₂,...,w_n), four indexs of consideration here: index m₁For R_iTo R^DSDistance, namely the distance between two relations is the least, Affinity is the highest；Index m₂For the relative radix of relation, namely R_iWith R_PatentIn the average tuple that is connected of each tuple Quantity；Index m₃Anti-phase to radix for relation, i.e. R_PatentWith R_iIn a tuple be connected par；Index m₄For R_i The connectedness of pattern, i.e. R_iThe quantity of the link in graph of a relation；Af(R_Parent) refer to R_iFather's node and R^DSAffine Degree, initial value is 1, i.e. R^DSThe affinity of itself is 1；The fraction range of index is [0,1], the summation of corresponding weights be 1 (on Stating four corresponding weights of index is all 0.25)；And in the generation of OS, the affinity of all relation nodes all should be higher than One marginal value θ；

Step 2.3: the formula of the importance Im (S) calculating alternative size-l OS S is:

$\mathrm{Im}\left(S\right)=\underset{n_i\&Element;S}{\&Sigma;}\mathrm{Im}(OS,R_i)---\left(3\right)$

Wherein Im (OS, R_i) it is OS interior joint R_iLi value, Im (OS, R_i) can be calculated by below equation:

Im(OS,R_i)=Im (R_i)·Af(R_i) (4)

Wherein, Im (R_i) it is R_iGl value, Af (R_i) it is R_iTo R^DSAffinity；

To sum up calculate Im value according to the key word of input, generate alternative OS；

Step 3: input natural number l generates final containing l joint according to OS k-LASP algorithm (referring to step 3.3) obtained The tree with DS as root of point；To consider three factors in this step: multiformity amount of attenuation (dv), proportionality increment (pv) and Quiescent value (li), they are respectively in connection with drawing a last mark (i.e. dw, pw) the most at last；

Step 3.1: multiformity (Dsize-l)

In order to avoid repeating of the too high analog information of importance, should select to export the diversified information of l bar, so providing The computational methods of one following multiformity amount of attenuation:

$dv\left(v_i\right)=1-\frac{z\left(g\right(v_i\left)\right)-1}{l-1}---\left(5\right)$

Wherein, g (v_i) refer to and v_iSimilar first group node；z(g(v_i))-1 refer in size-l OS and v_iNode is similar The summation of unit's group node；z(g(v_i)) refer to g (v_i) number of times in size-l OS to be occurred in；dv(v_i) codomain be [0,1]； Definition dv [z] is that node occurs that in size-l OS the multiformity of z time weakens value, and example makes l=10, and " Marry " occurs 2 times, I.e. z=2, then

Then, the multiformity weights that a node static value in Dsize-lOS is combined with multiformity weakening value are by following public Formula calculates:

dw(v_i)=li (v_i)·dv(v_i) (6)

To sum up, providing OS and l, generating a Dsize-l OS needs to meet following condition:

Tuple number in 1.Dsize-lOS is l (l≤| OS |)；

2. this l node all must be connected with root node；

3. each node v_iThere is corresponding multiformity weights i.e. dw (v_i)；

4. a Dsize-l OS collect to be divided into

$\mathrm{Im}\left(DSl\right)=\underset{v_i\&Element;DSl}{\&Sigma;}dw\left(v_i\right)---\left(7\right)$

Step 3.2: proportionality (Psize-l)

In view of in an OS, a first group node may occur in multiple times, but these nodes may have more weak static state Value, and their frequency has contact important with DS in actual size-l OS to be obtained, thus, proportionality increases Value can be calculated by equation below:

$pq\left(v_i\right)=\frac{fr\left(g\left(v_i\right)\right)}{\mathrm{\&alpha;}\&CenterDot;z\left(g\left(v_i\right)\right)+1}---\left(8\right)$

Wherein, fr (g (v_i)) it is g (v_i) occur in the number of times in OS；z(g(v_i)) refer to g (v_i) in size-l OS to be occurred in Number of times；α is a constant that can adjust ratio, typically takes α=2；

Then the proportionality weights that a node static value in Psize-l OS is combined with proportionality increment size are by as follows Formula calculates:

pw(v_i)=li (v_i)·pq(v_i) (9)

To sum up, providing OS and l, generating a Psize-l OS needs to meet following condition:

Tuple number in 1.Psize-l OS is l (l≤| OS |)

2. this l node all must be connected with root node

3. each node v_iThere is corresponding multiformity weights i.e. pw (v_i)

4. a Psize-l OS collect to be divided into

$\mathrm{Im}\left(PSl\right)=\underset{v_i\&Element;PSl}{\&Sigma;}pw\left(v_i\right)---\left(10\right)$

Step 3.3: generate the final tree with DS as root containing l node with k-LASP algorithm；

(namely one, the maximum average value path of k-LASP (k-Largest Averaged Score Path) i.e. k node The meansigma methods of k node weights on path), in this step, dw and pw primary system is referred to as weight w；Each joint in OS Point v_iThere is a weight w (v_i), corresponding v_iWith its ancestor node (number n, n=max (k 1, physical length)) Average weight is defined asDuring generating OS, need a Hash table, represent with HFr, HFr bag Including three parts, one is by v_iIn i as the numbering of node of graph, two is v_iNumber of times fr (the v occurred in OS_i), three is v_i? Number of times z (the v occurred in size-l OS_i)；In order to preferably manage OS interior joint and corresponding AP value, set up a queue W Preserving these information, in this queue, the order of node is successively decreased arrangement by corresponding AP value；

K-LASP algorithm generates the process of size-l OS:

1) OS is generated, including building HFr, calculating AP (v_i) and generate W

2) if | size-l | < l, 3 are turned), otherwise turn 11)

3)p_iRepresent the node having maximum AP value in current W to the path of root node,

By p_iIn the front individual node of l-| size-l | join in size-l OS

4) if | size-l | < l, 5 are turned), otherwise turn 10)

5. by selected p_iIn the individual node of l-| size-l | remove from OS and W

6) for p_iThe descendant nodes v of each node_jDo and update as follows:

AP (v is updated in OS and W_j) value；The number of descendant nodes is n, n=max (k 1, physical length)；

7. for p_iIn each node g (v), if g (v) is at HFr, turn 8), otherwise turn 10)；

8.HFr(g(v)).z++

9) the node n for the g of making (n)=g (v) each in OS does and updates as follows:

For each node n in each subtree with node n as root_iDo:

In OS and W, AP (n is updated by HFr (g (v)) .z value_i) value

9. turns 2)

11) size-l OS is returned

The l bar tuple information that will retrieve that the size-l OS now returned is the most required；

Through the results show, the experiment effect that this method obtains is notable.