CN102122291A - Blog friend recommendation method based on tree log pattern analysis - Google Patents

Abstract

The invention discloses a blog friend recommendation method based on tree-shaped log pattern analysis. Using the offline mining method, through the analysis of the server log, the visitor's access record to the blog page is extracted, and the access log tree rooted at the blog to be recommended is further constructed through grouping, sorting, loopback and other technologies. The access log tree is frequently mined to find the frequent subtrees that meet the preset requirements, and the nodes in the frequent subtrees are used as candidate blog friends, and the recommendation degree is calculated according to the set formula, and the ones with the highest scores are recommended . The algorithm is different from the traditional algorithm based on frequent item mining or frequent sequence mining. In view of the unique parallel link relationship and indirect access characteristics of the blog circle, the frequent tree structure mining method is used to fully explore and extract potential access links between blogs. , and recommended to visiting users, which improves the user experience, and is an efficient and practical blog recommendation method.

Description

A blog friend recommendation method based on tree log pattern analysis

technical field

The invention relates to data analysis technology of blog server logs and mining technology of frequent access patterns, in particular to a blog friend recommendation method based on tree log pattern analysis.

Background technique

With the continuous development of Internet technology, blogs are no longer simply a platform for publishing personal articles and information. After adding various interactive functions such as messages, attention, friends, etc., users will gradually form a blog circle. . The blogosphere includes friends, potential friends (bloggers who have not yet joined the friend list or friends of friends), other bloggers with similar interests, and so on. In a typical web2.0 application like a blog, the key to the success of the system is to establish a social relationship with like-minded users. Therefore, friend recommendation for blogs has become the main function of the blog system. The blog friend recommendation application discovers the potential correlation between blog users through the user's visit behavior to the blog, and tries to suggest that the blog converts people who may have common interests into friend relationships according to the correlation.

The blog circle is a complex tree or graph structure, and there are already some blog-oriented friend recommendation systems. They generally make recommendations based on the established friendship between blogs and the number of visits recorded by the server. These recommendation methods are based on frequent item mining or frequent sequence mining, and have the following deficiencies and shortcomings: 1) They do not consider the unique parallel link relationship between blogs and indirect access features; 2) did not consider the logical relationship between blog pages hidden in the order of users' access to pages; 3) did not fully consider the hierarchical relationship and depth relationship of the website organizational structure.

Contents of the invention

Aiming at the rich user behavior information and page organization information implied in blog server logs, the purpose of the present invention is to provide a blog friend recommendation method based on tree log pattern analysis, which is aimed at blog logs and based on tree structure mining Blog recommendation method.

The technical scheme that the present invention solves its technical problem adopts is:

The steps taken by this method are as follows:

1) Analyze the original log, extract valid information, and create a session table in the database to record the user's access path;

2) For the blog to be recommended, find the users who have visited the blog to be recommended in the database, according to the user's access log, go back and forth, and build an access log tree rooted at the blog to be recommended;

3) Do frequent recursive unordered tree mining on the constructed access log tree to find frequent subtrees that meet the preset requirements;

4) Take the nodes in the frequent subtree as candidate blog friends, calculate the recommendation degree according to the set formula, and recommend the ones with the highest scores.

2, a kind of blog friend recommending method based on the analysis of tree-shaped log pattern according to claim 1, it is characterized in that: in the described step 1), analyze original log, extract effective information, extract in the server exactly with log parser Log, get the access record within a time slice, remove the redundant information in the user request, convert it into an access triplet <visitor, access time, access blog> and save it in the session table, the selection of the time slice size is based on blog access The amount and performance of the computer running the mining algorithm. If the visitor is a registered user, the user name is identified as "visitor", and if the visitor is an anonymous user, the user IP is identified as "visitor".

3. A blog friend recommendation method based on tree-shaped log pattern analysis according to claim 1, characterized in that: in said step 2), for the blog to be recommended, find out in the database the blog that has been visited and to be recommended According to the user's access log, go back and forth, and build an access log tree rooted at the blog to be recommended, that is, according to the organizational structure information of the website, for the blog to be recommended, find out the users who have visited the blog in the session table The user and the user's first visit to the blog, for each searched visitor, extract the records of other blogs visited by the searched visitor after visiting the blog to be recommended; the tree structure generator uses each The visitor constructs the access log tree as a unit, and each blog visited by the visitor corresponds to a node, and each node contains access triplet information, and the parent-child node relationship is formed according to the time sequence of continuous access requests; for the generated loopback , delete the latest edge in terms of access time, and the resulting access log tree has three characteristics: first, the access log tree has the same root node, which is the blog to be recommended; second, all access log trees do not have labels The same sibling nodes; Third, the access log tree is unordered, that is, the child nodes of each node are unordered.

4. A method for recommending blog friends based on tree-shaped log pattern analysis according to claim 1, characterized in that: in the step 3), frequent recursive unordered tree mining is performed on the constructed access log tree to find out The frequent subtree that meets the preset requirements is to record all the access log trees as t1, t2...tn, select the appropriate minimum support degree minsupÎ(0,1), and use the frequent subtree miner to mine. The specific steps are as follows :

The first step is to traverse t1, t2...tn, and classify the nodes with the same "visit blog" in the "visit triplet" as the same node, and count the number of times fre1 each node appears in the access log tree, for fre1>minsup* The node of n is recorded as the frequent subtree EQ1;

The second step is to expand EQ1, connect the two nodes in EQ1 to form a parent-child relationship, form a tree containing 2 nodes, and use them as candidate subtrees to count the occurrence of candidate subtrees in all access log trees The number of times fre2, for the candidate subtree of fre2>minsup*n, is recorded as the frequent subtree EQ2;

The third step, starting from EQ2, for the rightmost path of each tree, do enumeration expansion, expand one node at a time, find all possible candidate subtrees, and count the trees with occurrence times frei>minsup*n, record For the new frequent subtree EQi, do a similar recursive operation, and continuously increase the number of frequent subtree nodes that are mined until there is no matching candidate subtree.

5. A kind of blog friend recommendation method based on tree-shaped log pattern analysis according to claim 1, characterized in that: in the step 4), the node in the frequent subtree obtained by digging is used as a candidate blog friend, according to the setting The recommended formula is used to calculate the recommendation degree, and the ones with the highest scores are selected for recommendation, that is, for the frequent subtrees with more than 3 nodes, they are sorted according to the frequency of occurrence fre from large to small, and each frequent subtree is taken out in turn, and done The operation is as follows: According to breadth-first traversal, starting from the second layer of the tree, calculate the recommendation R of each node, the formula is as follows:

Parameter description: fre is the frequency of frequent subtrees; T indicates whether there is a direct page link, if it exists, then T is 1, if it does not exist, then T is 0; d is the depth of the node, and the depth of the root node is recorded as 0; W _k is the weight parameter of each layer, and the default is 1; B _k is the number of branches of each layer, that is, the number of sibling nodes under the same parent node; after calculating the recommendation degree of all candidate nodes, select the one with the highest score as needed There are several nodes, and the blogs corresponding to the nodes are recommended as blog friends.

The beneficial effects that the present invention has are:

According to the visitor's visit behavior to the blog and the structural characteristics of the blog site, combined with the existing data mining technology, the frequent visit pattern of the tree structure is excavated for the server's visit log. The service provider of the blog researches and analyzes the user's access behavior based on the frequent access patterns excavated, recommends blog friends to the user, and improves the user experience; at the same time, it can also assist the website architect to better organize the website structure and increase the user's access rate to the blog .

Description of drawings

Figure 1 is an overall structure diagram of a blog friend recommendation method based on tree log pattern analysis.

Figure 2 is the access session and its index.

Fig. 3 is an access log tree constructed according to the session of visitor1 in Fig. 2 .

Fig. 4 is a schematic diagram of a frequent subtree mining method.

FIG. 5 is a schematic diagram of a method for calculating a recommendation degree.

Detailed ways

The present invention will be further described below in conjunction with specific examples and accompanying drawings.

Through the blog log analysis method provided by the present invention, frequent access patterns can be extracted quickly and effectively, and potential blog friends are recommended to visiting users through an intelligent screening process. The overall structure diagram is shown in Figure 1, and the specific implementation Proceed as follows:

1) The log parser in Figure 1 parses the server logs within a period of time, deletes redundant information, and constructs a visit triple <visitor, visit time, visit blog> (triple<visitor, access_time, blog_url>), taking the Apache server log as an example, the specific process is as follows:

The logs recorded in the Apache server can be expressed in the following form:

117.24.255.86 - - [01/Jul/2010:18:01:25 +0800] "GET

http://B.blog.163.com HTTP/1.0" 200 1231

"117.24.255.230.1277794615926482" 46807 "Mozilla/4.0 (compatible; MSIE 6.0; Windows NT 5.1; SV1; .NET CLR 2.0.50727; .NET CLR 3.0.04506.30)"

This log record shows that the anonymous user with IP 117.24.255.86 visited the page http://A.blog.163.com at time 01/Jul/2010:18:01:25 +0800.

Therefore, the access triplet <117.24.255.86, 2010-7-1 18:01:25, blogA>, for a visitor who is a registered user, use the registration ID as the identifier of the visitor; for an anonymous user, use the IP as a distinction to create a temporary ID; for the visited page, In order to facilitate the next step of processing, the url address of the page can be simplified in combination with the organizational structure information of the website. For example, http://A.blog.163.com is simplified as blogA here, and the two must correspond one-to-one.

2) For the blog to be recommended, the tree structure generator in Figure 1 finds out the users who have visited the blog to be recommended in the database, according to the user's access log, removes the loop, and constructs the visit rooted at the blog to be recommended log tree, the specific steps are as follows:

In the first step, according to the organizational structure information of the website, for a blogA to be recommended, the grouping and sorting module of the tree structure generator finds out all users who have visited blogA and the user who visited blogA for the first time in the session table. Time (SQL query: select visitor, distinct access_time from triple where access_url = blogA). Assume that the time when user visitor0 visits blogA for the first time is access_time0, find all the pages of user visitor0 whose access time is after access_time0, and do the same operation for each user obtained from the query (SQL query: select visitor , access_time, access_url from triple where visitor = visitor0 and access_time > access_time0, after the above operations, you can get the user session information as shown in Figure 2

The second step is to construct the access log tree with the visitor as the unit for the records obtained by the query, and the formation of the parent-child node relationship is based on the time sequence of continuous access requests; for the generated loop, the loopback in the tree structure generator The module eliminates the loopback by deleting the edge with the latest time on the loopback. The access log tree generated according to the session of visitor1 in Figure 2 is shown in Figure 3. The generated access log tree has three characteristics: first, all access log trees have the same root node, which is the blog to be recommended; second, there are no nodes with the same label in all access log trees; third, the access The log tree is unordered, that is, the order of sibling nodes is not considered.

3) The frequent subtree excavator shown in Figure 1 performs frequent recursive unordered tree mining on the access log tree constructed in the previous step, and finds frequent subtrees that meet the preset requirements. The specific steps are as follows: the tree structure generator takes The access log tree numbers obtained in the previous step are respectively t1, t2,...tn.

The first step: the candidate subtree generation module in the tree structure generator traverses all the access log trees, and classifies the same nodes as the "visit blog" in the "visit triplet" as the same node, and the subtree frequency statistics module counts The position where each type of node appears in the access log tree and the total number fre1 (frequency) of the tree containing this type of node, for nodes with fre1>minsup*n, it is recorded as frequent subtree EQ1;

Step 2: Connect two nodes in EQ1 to form a parent-child relationship. As a candidate frequent subtree, count the number of times fre2 that the candidate frequent subtree appears in all logs. The specific steps are shown in Figure 4. Node A and node B both belong to EQ1, connect A and B, A is the parent node of B, and record the position of the last newly added node in the original tree (node B in Figure 4), for fre2>minsup* The candidate subtree of n is denoted as frequent subtree EQ2.

Step 3: Starting from EQ2, for the rightmost path of each tree, do enumeration expansion, expand one node at a time, find all possible candidate subtrees, and count the trees with occurrence times frei>minsup*n, record is the new frequent subtree EQi. As shown in Figure 4, firstly, the rightmost path is extended for node A, and a new node B is expanded, and the original node B can also be expanded, but only one node can be expanded at a time. Do similar recursive operations in this way, and continuously increase the number of nodes in the frequent subtrees that are mined until there is no matching candidate frequent subtree. In the process of mining the tree, in order to facilitate the recording of the tree, the string encoding of the tree is adopted. For example, in Figure 4, the encoding of tree t1 is ABC-1BD-1E-1-1B, and the character encoding of t2 is ABC-1DE-1- 1-1B, the encoding is based on the depth-first traversal order, and a -1 is inserted each time it goes back. According to this method, the tree and the string encoding are in one-to-one correspondence.

4) After digging out all the frequent subtrees, the candidate node recommender shown in Figure 1 sorts the frequent subtrees from large to small according to the occurrence frequency fre of the frequent subtrees, and takes out each frequent subtree in turn to do the following operations: according to the width first In order of traversal, starting from the second layer of the tree, the node recommendation calculation module calculates the recommendation R of each node, the formula is as follows:

为每层的分支数目，即同一父节点下兄弟节点数目。 Parameter description: fre is the frequency of frequent subtrees; T indicates whether there is a direct page link, if it exists, then T is 1, if it does not exist, then T is 0; d is the depth of the node, and the depth of the root node is recorded as 0; is the weight parameter of each layer, the default is 1;

is the number of branches of each layer, that is, the number of sibling nodes under the same parent node.

As shown in Figure 5, the frequent subtree ABC-1D-1-1B (string code) is excavated. This tree appears in both t1 and t2, so the frequency fre is 100%. Calculate the candidate subtree The recommendation degree R, the steps are as follows: For point A, it is in the first layer of the tree, so it is skipped. For node B in the second layer, if there is no direct link from A to B in the website structure, then T is 0, so R _B =0; if there is a direct link from A to B in the website structure, T=1, then R _B =1*1*1/2=0.5. For node C, if there is no direct link from node B to node C, then T=0, so R _C =0; if there is, then T=1, then R _C =1*1*(1/2) (1 /3) = 0.167. Node D is the same as C.

After calculating the recommendation degrees of all candidate nodes, select the blogs corresponding to several nodes with the highest scores as required as blog friends for recommendation. The nodes calculated according to Figure 5, assuming that there are direct links, according to the calculation, the recommendation degrees of node B and node E are both 0.5, so the blogs corresponding to these two nodes are recommended as blog friends first, and node C and node D The recommendation degree is 0.167, and their corresponding blogs are further recommended as blog friends if necessary.

Claims (5)

1. blog intimate recommend method of analyzing based on tree-like logging mode is characterized in that the step that this method adopts is as follows:

1) resolves original log, extract effective information, in database, create conversational list, be used for the access path of recording user;

2) at blog to be recommended, in database, find out the user of visit blog to be recommended excessively, the access log according to the user removes winding, and making up with blog to be recommended is the access log tree of root;

3) the access log tree that constructs is cooked frequent recurrence unordered tree and excavate, find out the frequent subtree that meets preset requirement;

4) the node in the frequent subtree as candidate's blog intimate, calculate by the formula degree of setting of recommendation, get the highest several of score value and recommend.

2. a kind of blog intimate recommend method of analyzing based on tree-like logging mode according to claim 1, it is characterized in that: resolve original log in the described step 1), extract effective information, be exactly with the daily record in the daily record resolver extraction server, obtain the Visitor Logs in the timeslice, remove the redundant information in user's request, change into visit tlv triple＜visitor, access time, accesses blog〉deposit in the conversational list, the selection of timeslice size is according to the performance of the computing machine of blog visit capacity and operation mining algorithm, the visitor is the registered user's, with user by name " visitor's " sign, the visitor is an anonymous, is the sign of " visitor " with User IP.

3. a kind of blog intimate recommend method of analyzing based on tree-like logging mode according to claim 1, it is characterized in that: described step 2) at blog to be recommended, in database, find out the user of visit blog to be recommended excessively, access log according to the user, remove winding, structure is the access log tree of root with blog to be recommended, be exactly institutional framework information according to the website, at blog to be recommended, in conversational list, find out the time that the user that visited this blog and user visit this blog for the first time, search the visitor who obtains at each, extract the record of searching other blog that the visitor that obtains visits after visit blog to be recommended; The tree structure maker is a unit structure access log tree with each visitor, the corresponding node of each blog of Accessor Access, each node comprises the visit triplet information, and the formation of father and son's node relationships is according to the temporal sequencing of connected reference request; For the winding that produces, delete limit at the latest on the access time, the access log tree of generation has three characteristics: the first, the access log tree has identical root node, is blog to be recommended; The second, there is not the identical brotgher of node of label in all access log trees; The 3rd, the access log tree is unordered, and promptly the child node of each node is unordered.

4. a kind of blog intimate recommend method of analyzing based on tree-like logging mode according to claim 1, it is characterized in that: in the described step 3) access log tree that constructs is cooked frequent recurrence unordered tree and excavate, find out the frequent subtree that meets preset requirement, exactly all access log trees are designated as t1 respectively, t2 ... tn selects suitable minimum support minsup (0,1), excavate with frequent subtree delver, concrete steps are as follows:

The first step, traversal t1, t2 ... tn is classified as same node point to " accesses blog " identical node in " visit tlv triple ", adds up the number of times fre1 that every kind of node occurs in access log tree, for fre1〉node of minsup*n, be designated as frequent subtree EQ1;

Second the step, EQ1 is done expansion, node among two EQ1 is done attended operation, constitute set membership, formation comprises the tree of 2 nodes, as candidate's subtree, count the occurrence number fre2 of candidate's subtree in all-access daily record tree, for fre2〉candidate's subtree of minsup*n, be designated as frequent subtree EQ2;

The 3rd the step, from EQ2, right wing footpath for every tree, do and enumerate expansion, node of each expansion is found out all possible candidate's subtree, counts occurrence number frei〉tree of minsup*n, be designated as new frequent subtree EQi, do similar recursive operation, constantly increase the interstitial content of the frequent subtree of excavating, till the candidate's subtree that does not meet.

5. a kind of blog intimate recommend method of analyzing based on tree-like logging mode according to claim 1, it is characterized in that: the node in the frequent subtree that in the described step 4) excavation is obtained is as candidate's blog intimate, calculate by the formula degree of setting of recommendation, getting the highest several of score value recommends, be exactly greater than 3 frequent subtree to the node number, fre sorts from big to small according to occurrence frequency, take out every frequent subtree successively, be done as follows: according to breadth first traversal, from the tree the 2nd layer, calculate the recommendation degree R of each node, formula is as follows:

Parameter declaration: fre is the frequency of frequent subtree; T represents whether there is direct page link, exists, and then T is 1, does not exist, and then T is 0; D is the degree of depth of this node, and the root node degree of depth is designated as 0; W _kBe every layer weight parameter, be defaulted as 1; B _kBe every layer number of branches, brotgher of node number under the promptly same father node; After calculating the recommendation degree of all both candidate nodes, as required, select the highest plurality of nodes of score value, get the blog of node correspondence and recommend as blog intimate.

Images