CN102262661B - Web page access forecasting method based on k-order hybrid Markov model - Google Patents

Abstract

The invention discloses a Web page access forecasting method based on a k-order hybrid Markov model, and the method comprises the following steps of: firstly, gathering and handling Web server access log data, identifying a client and users and eliminating insignificant access data; then identifying user sessions to construct a Web log database; select log data from the database according to forecasting targets, and organizing (k+1) tuples by taking a session as a unit, wherein the (k+1) tuples are used for training the k-order hybrid Markov model; learning and calibrating a parameter set of the k-order hybrid Markov model by using an expectation-maximization algorithm; and identifying sessions according to page access operations of target users and applying the model for forecasting Web pages to be accessed by the users at the next step. The method provided by the invention can be used for recommending pages needed to be accessed by the users so as to reduce the delay of page access and optimize user experience; and in addition, from the point of a Web server, the organizational structure of the Web pages can be improved, the result sequence of a search engine is guided, and the page cache mechanism can be improved, thereby enhancing the quality of service.

Description

A kind of Web page access Forecasting Methodology based on k rank mixing Markov model

Technical field

The present invention relates to the personalized forecasting techniques of Web page access, especially for the Internet era Web server information more and more huger and complicated, the situation that user's visit capacity is increasing; Need to determine next step page that may access of user according to user's accessing characteristic, help the user faster and better find needed information; Assist simultaneously the Web server page of cache user needs in advance, improve the page link structure, thereby improve server access efficient.

Background technology

Be accompanied by the fast development of internet, WWW (World Wide Web) has become a worldwide hypermedia acquisition of information platform, people more and more depend on Web and obtain various information, have also become the part of people's daily life and browse Web.A complete Web uses the http protocol that comprises Web server, client browser, completes to communicate by letter between client and server, HTML (Hypertext Markup Language) HTML and the uniform resource position mark URL that is used for describing the Web page.Use the continuous expansion of scale along with Web, the page that comprises on Web server is more and more, and the institutional framework of the Web page also becomes increasingly complex.The user is easy in the ocean of the information that gets lost in.How to allow the user find own needed information faster and betterly; Allow Web server provide personalized service to promote service quality according to user's use habit, become the problem that present Web application need to solve.

Summary of the invention

Technical matters to be solved by this invention is to provide a kind of method that access characteristic historical according to the user predicted web access, access situation that simultaneously can be current according to the user, prediction or next step page that will access (set) of recommendation user.

For achieving the above object, the present invention adopts following step:

1) at first collect and arrange Web server access log data; For each Visitor Logs in daily record, identify customer end browser and user; Get rid of insignificant visit data, as auto-programmings such as robot and reptiles to the Visitor Logs of Web server, non-page access record etc., according to the record accessing operation o=＜u that extracts, x, t 〉, wherein u represents that user, x represent the Web page, t representation page access time;

2) then identify user conversation S, be used for setting up the Web log database, deposit is used for the historical data of Web page access prediction;

3) choose and organize daily record data according to target of prediction from database, arrange and tissue (k+1) tuple-set by session;

4) set up k rank mixing Markov model, and adopt greatest hope algorithm (Expectation-Maximization) to train this k rank mixing Markov model, then the based on data collection

The parameter set of study and calibration k rank mixing Markov model;

5) accessing operation of based target user to the Web page identified nearest user conversation, the Web page of next step access of k rank mixing Markov model predictive user after application training.

Above-mentioned steps 2) in, the process of identification user conversation is: make session S={o ₁, o ₂..., o _l, make all operations o in session S _iThe execution user identical; Then according to last accessing operation o _i=＜u, x _i, t _iThe time t that occurs _iAnd the page x that accesses _i, judge next operation o _i+ 1=＜u, x _i+1, t _i+1Whether belonging to same session S, this is judged based on following three conditions:

Page x _i+1By page x _iQuote, i.e. x _i+1URL be contained in page x _iIn;

Page x _i+1By the x that removes that was accessed in session S _iOutside other webpage t _jQuote, as x _j(j＜i, corresponding accessing operation o _jAnd t ∈ S), _jAnd t _i+1Mistiming less than session threshold value (as 30 minutes);

Page x _i+1Do not quoted by the page of being accessed in session S, but last operation t _iAnd t _i+1Mistiming less than page threshold value (as 5 minutes).

If satisfy one of above-mentioned condition, decision o _i+1Belong to session S; Otherwise operation o _i+1Open a new session.

Above-mentioned steps 4) in, k rank mixing Markov model is by k state-transition matrix { Λ ₁, Λ ₂... Λ _kAnd a weight vector A={ α ₁, α ₂..., α _kForm.Make that in Web server, the page adds up to n, transition matrix Λ _jA n * n matrix, its element λ _jAfter (x, y) representation page x is accessed, page y accessed probability, i.e. conditional probability p (x after an interval j page in same session _k+1| x _K-j+1), page x wherein _K-j+1And x _k+1Be equal to respectively page x and y.

All n in k state-transition matrix in k rank mixing Markov model ²K element, and the weights of the k in weight vector A, the parameter set that component model need to be trained.By step 4), the data-oriented collection

Adopt greatest hope Algorithm Learning and these parameters of calibration.At first calculate the initial value of each parameter, in weight vector A, the initial value of element is:

${\mathrm{\α}}_j=\frac{1}{k}(1\≤j\≤k)$

For state-transition matrix Λ _j, element λ _jThe initial value of (x, y) is for (to make data set

In (k+1) tuple quantity be m, tuple X _i=＜x _{I, 1}, x _{I, 2}..., x _{I, k}, x _{I, k+1}):

Wherein δ () function is defined as:

$\mathrm{\δ}\left(b\right)=\left\{\begin{array}{cc}1& \mathrm{b\; is\; true}\\ 0& \mathrm{b\; is\; false}\end{array}\right.$

New parameter value when application greatest hope algorithm need to calculate next iteration according to the parameter current value is labeled as respectively α ' _jAnd λ ' _j(x, y).For calculating new parameter value, introduce hidden variable collection Θ={ θ ₁, θ ₂..., θ _m, hidden variable θ wherein _iRepresent tuple X _iThe random character of self.Make that j is the page number of interval access, θ _i(j) computing formula is as follows:

${\mathrm{\θ}}_i\left(j\right)=\frac{{\mathrm{\α}}_j\·{\mathrm{\λ}}_j(x_{i,k-j+1},x_{i,k+1})}{{\mathrm{\Σ}}_{l=1}^k{a}_l\·{\mathrm{\λ}}_l(x_{i,k-l+1},x_{i,k+1})}$

According to θ _i(j), can estimate that the new value of each parameter is as follows:

${\mathrm{\α}}_j^{\′}=\frac{1}{m}\underset{i=1}{\overset{m}{\mathrm{\Σ}}}{\mathrm{\θ}}_i\left(j\right)$

Respectively substitution previous round parameter value and new parameter value, likelihood function that can the calculation training collection.Can determine by judging whether likelihood function restrains whether k rank mixing Markov model trains complete.

Above-mentioned steps 5) process of using the Web page of next step access of k rank mixing Markov model predictive user in is: at first identify targeted customer's current sessions, obtain k the page＜x of user's connected reference recently in this session ₁, x ₂..., x _k.Then to the arbitrary page y on Web server, design conditions probable value p (y|x ₁, x ₂..., x _k):

$p\left(y\right|x_1,x_2,...,x_k)=\underset{j=1}{\overset{k}{\mathrm{\Σ}}}{\mathrm{\α}}_j\·{\mathrm{\λ}}_j(x_{k-j+1},y)$

Front T the page (for example T=10) conduct of selecting at last to have most probable value is looked ahead or is recommended user's the page.

The present invention arranges out user's history access record according to the web access daily record, the identification user conversation; Take correlation rule or the sequence pattern of session as unit Mining Web page access, the support of compute associations rule or sequence pattern and degree of confidence, the Web page sequence of having accessed according to the user calculates accessed (stablizing) probability of other pages; And then prediction or next step page that need to access of recommendation user.Existing method or lack the identification of user conversation and effective tissue; Perhaps use the low order Markov model to cause precision of prediction (as the Top-T degree of accuracy) too low; Perhaps use the high-order Markov model to cause that the model complexity is high and the data cover rate is low, in the situation that increase predicted time and complexity may reduce precision of prediction on the contrary.

The method is calculated simple, and the model complexity is low, has higher execution performance.Be with good expansibility simultaneously, easily combine with other Web method for digging such as user clustering analysis etc.Experimental data shows, the execution performance of the inventive method is equal to the low order Markov model, and precision of prediction is equivalent to the high-order Markov model.The inventive method also comprises the daily record data finishing techniques such as user conversation analysis, can effectively promote the effect of Web page access prediction.

Description of drawings

Fig. 1 is the overall framework of Web page access Forecasting Methodology;

Fig. 2 is the treatment scheme that arranges web access daily record identification user conversation;

Fig. 3 chooses and organizes the treatment scheme of daily record data according to target of prediction;

Fig. 4 is the flow process that adopts greatest hope Algorithm for Training k rank mixing Markov model;

Fig. 5 uses the flow process that k rank mixing Markov model is completed the Web Web page predicting;

Fig. 6 is the Comparison of experiment results of 2 rank mixing Markov models and complete 2 rank Markov models.

Embodiment

Figure 1 shows that the general technical framework of Web page access Forecasting Methodology, as seen, the input of the method is the Web page of Web server history log record and targeted customer's recent visit.The output of method is the Web page set of recommending or looking ahead for the targeted customer.Technological frame is divided into 5 modules: arrange Web page access daily record and identify the user; The identification user conversation is also set up the Web log database; Choose and organize daily record data according to target of prediction; Training k rank mixing Markov model; Operate next step accession page of prediction according to the targeted customer at last.

Web server and application program thereof be continuous recording user access log data in the course of the work, and different Web servers realizes that the mode of its log recording and data layout are slightly different.In the standard situation, for each HTTP request of user (or other agreements such as FTP, the inventive method is mainly for the Web page request), all can generate a log recording.The information that log recording comprises usually has: the resource of client ip address, request (page or URL), optional parameter, HTTP executing state, user agent's (browser and OS Type, version), referring domain (page), user Cookie etc.The Web logdata record access behavior of different user, to the reasonable analysis of daily record data with utilize the service quality that can effectively promote Web server.

Table 1 is depicted as an example of Web server (maya.cs.depaul.edu) log recording:

Table 1:

In table 1, IP address of record 1 expression is the page "/classes/cs589/papers.html " on the user access server (maya.cs.depaul.edu) of 1.2.3.4; The HTTP operation of adopting is " GET "; User browser (Mozilla) and operating system (Windows NT) information recording/is in Agent Domain; Referring domain represents that where the user from accessing current page, i.e. " http://dataminingresources.blogspot.com/ ".Record 2 reflections be same user, access resources "/classes/cs589/papers/cms-tai.pdf " again after having accessed " paper.html ", its referring domain is " paper.html ".Record 3 another users of expression pass through the Google search engine inquiry to the book server page, and referring domain " http://www.google.com/search? ... ", access resources "/classes/ds575/papers/hyperlink.pdf ".At last, record 4～6 reflection third party connected reference server resources, actual in the same page of access: "/classes/cs480/announce.html "; Wherein record 5 and 6 expressions obtaining 2 resource objects in the page.

Figure 2 shows that and arrange the web access daily record, identify user conversation and lay in the treatment scheme in daily record storehouse.In arranging web access daily record process, at first for each log recording identification calling party.User's identification can be based on client ip address and browser, and this is a kind of rough recognition methods; Identification more accurately is based on the Cookie of client, and Cookie is created in client by server, is used for sign and follows the tracks of user's request.Then clear up invalid Visitor Logs, comprise the Visitor Logs of getting rid of non-page access record and being produced by auto-programmings such as reptiles.

According to the web access daily record after arranging, wherein each record can be defined as an accessing operation o=＜u, x, t 〉, wherein u represents that user, x represent the Web page, t representation page access time.The web access daily record is regarded as the list of accessing operation.Next by the user, list is sorted out, safeguarded its accessing operation list for each user.

By shown in Figure 2, need to identify user's session in each user list.Define a session S={o ₁, o ₂..., o _l, one group of page operation that representative of consumer access Web server is associated around the time.Existence dependency relationship before and after page operation in session can be used for access behavior and the custom of analysis user; It is separate that page operation between session can be thought.If ignore the identification of session, can produce a large amount of noises in Web Web page predicting process, cause predict the outcome inaccurate.

Identification user conversation S is according to accessing operation o _i=＜u, x _i, t _iThe time t that occurs _iAnd the page x that accesses _iAt first all operations o in session S _iThe execution user must be identical, and o _iSequence in chronological order.Then for the next operation of user o _i+1=＜u, x _i+1, t _i+1, decision o _i+1Whether belong to same session S based on following three conditions:

Page x _i+1By page x _iQuote, i.e. x _i+1URL be contained in page x _iIn;

Page x _i+1By the x that removes that was accessed in session S _iOutside other pages quote, as x _j(j＜i, corresponding accessing operation o _jAnd t ∈ S), _jAnd t _i+1Mistiming less than session threshold value (as 30 minutes);

Page x _i+1Do not quoted by the page of being accessed in session S, but last operation t _iAnd t _i+1Mistiming less than page threshold value (as 5 minutes).

If satisfy one of above-mentioned condition, decision o _i+1Belong to session S; Otherwise operation o _i+1Open a new session.Wherein session threshold value and page threshold value can be adjusted according to actual conditions.After the identification user conversation, set up the Web log database.Log database is by user, two hierarchical organizations of session, as the data basis of web access prediction.

Figure 3 shows that according to target of prediction and choose daily record data and be organized into the treatment scheme of (k+1) tuples list.At first choose daily record data from database according to target of prediction.Can obtain its historical visit data according to the targeted customer, if data volume is less, may affect the effect of Web page predicting.A kind of improving one's methods is all historical visit datas of choosing similar user.Can use clustering method attribute close (as age, sex, educational background, income level etc.) or the approaching user of access module are classified as a class.

After obtaining daily record data, selected constant k, k page operation before requiring to consider during representative prediction next page, i.e. the page of next step access of user depends on front k the page of having accessed.The k value is selected and need to according to the smallest size of user conversation, be done compromise between precision of prediction.Extract (k+1) tuple X=＜x take session as unit ₁, x ₂..., x _k, x _k+1.The page in each (k+1) tuple belongs to same session, comprises one group of page of user's connected reference in session.Belonging to the page that adjacent two (k+1) tuples of same session comprise can repeat, and for example next tuple of tuple X can be: X '=＜x ₂, x ₃..., x _k+1, x _k+2.All (k+1) tuples consist of a set

Be used for training k rank mixing Markov model.

Figure 4 shows that the treatment scheme that adopts greatest hope Algorithm for Training k rank mixing Markov model.At first mixing Markov model in k rank is by k state-transition matrix { Λ ₁, Λ ₂... Λ _kAnd a weight vector A={ α ₁, α ₂..., α _kForm.Make that in Web server, the page adds up to n, transition matrix Λ _jA n * n matrix, its element λ _jAfter (x, y) representation page x is accessed, page y accessed probability, i.e. conditional probability p (x after an interval j page in same session _k+1| x _K-j+1), page x wherein _K-j+1And x _k+1Be equal to respectively page x and y.

All n in k state-transition matrix in k rank mixing Markov model ²K element, and the weights of the k in weight vector A, the parameter set that component model need to be trained.Corresponding with it, the complete needed number of parameters of k rank Markov model is

According to Fig. 4, the data-oriented collection Adopt greatest hope algorithm (Expectation-Maximization) to learn and calibrate these parameters.Calculate the initial value of each parameter, in weight vector A, the initial value of element is:

${\mathrm{\α}}_j=\frac{1}{k}(1\≤j\≤k)---\left(1\right)$

For state-transition matrix Λ _j, element λ _jThe initial value of (x, y) is for (to make data set

In (k+1) tuple quantity be m, tuple X _i=＜x _{I, 1}, x _{I, 2}..., x _{I, k}, x _{I, k+1}):

Wherein δ () function is defined as:

$\mathrm{\δ}\left(b\right)=\left\{\begin{array}{cc}1& \mathrm{b\; is\; true}\\ 0& \mathrm{b\; is\; false}\end{array}\right.---\left(3\right)$

Then calculate likelihood function

Representative under the parameter current collection, data set

There is the possible probability of (namely becoming sampled value).Computing formula is as follows:

P (X wherein _i| Λ ₁, Λ ₂... Λ _k, A) be illustrated under the parameter current collection, (k+1) tuple X _iBecome the probability of sampled value, computing formula is as follows:

$p\left(X_i\right|{\mathrm{\Λ}}_1,{\mathrm{\Λ}}_2,...{\mathrm{\Λ}}_k,A)=\underset{j=1}{\overset{k}{\mathrm{\Σ}}}{\mathrm{\α}}_j\·{\mathrm{\λ}}_j(x_{i,k-j+1},x_{i,k+1})---\left(5\right)$

New parameter value when next calculating next iteration according to the parameter current value is labeled as respectively α ' _jAnd λ ' _j(x, y).For calculating new parameter value, introduce hidden variable collection Θ={ θ ₁, θ ₂..., θ _m, hidden variable θ wherein _iRepresent tuple X _iThe random character of self.Make that j is the page number of interval access, θ _i(j) computing formula is as follows:

${\mathrm{\θ}}_i\left(j\right)=\frac{{\mathrm{\α}}_j\·{\mathrm{\λ}}_j(x_{i,k-j+1},x_{i,k+1})}{{\mathrm{\Σ}}_{l=1}^k{a}_l\·{\mathrm{\λ}}_l(x_{i,k-l+1},x_{i,k+1})}---\left(6\right)$

According to θ _i(j), can estimate that the new value of each parameter is as follows:

${\mathrm{\α}}_j^{\′}=\frac{1}{m}\underset{i=1}{\overset{m}{\mathrm{\Σ}}}{\mathrm{\θ}}_i\left(j\right)---\left(7\right)$

According to new parameter value, the likelihood function under new parameters sets is calculated in application of formula (4) and (5) again

Calculate the relative difference Δ of a same likelihood function of new likelihood function:

Judge that at last whether difference DELTA is less than set-point, as 10 ^-6, showing that if it is iteration tends towards stability (convergence), the training of k rank mixing Markov model finishes; Otherwise reenter the next round iteration, continue calculate new parameters sets and repeat above-mentioned computing according to the parameter current collection.

Figure 5 shows that and use the flow process that k rank mixing Markov model is completed the Web Web page predicting.At first identify the last session of targeted customer, obtain the page＜x of user's front k connected reference in this session ₁, x ₂..., x _k.Then to the arbitrary page y on Web server, design conditions probable value p (y|x ₁, x ₂..., x _k):

$p\left(y\right|x_1,x_2,...,x_k)=\underset{j=1}{\overset{k}{\mathrm{\Σ}}}{\mathrm{\α}}_j\·{\mathrm{\λ}}_j(x_{k-j+1},y)---\left(10\right)$

Front T the page (Top-T, for example T=10) conduct of selecting at last to have most probable value is looked ahead or is recommended user's the page.

The inventive method is used k rank mixing Markov model, on the basis of identification user conversation, and k page access situation before the based target user, next step page that will access of predictive user.This method obtains active balance between precision of prediction and model complexity, can fully use on the one hand the historical visit data of Web server, completes prediction with less time and space cost; Basic suitable with the high-order Markov model aspect precision of prediction on the other hand, be with good expansibility simultaneously.

Figure 6 shows that the Comparison of experiment results of 2 rank (k=2) mixing Markov model and complete 2 rank Markov models.Experiment based on daily record data be MSWEB-DATA, comprise 285 pages, the user conversation quantity that can identify is 37211.Experiment performance index one relatively are the Top-T precision of predictions, are defined as the ratio that next step actual access page of user exists in front T the page that method is recommended; The 2nd, algorithm execution time is mainly the predicted time of module 5.For guaranteeing the robustness of experiment, the user conversation of random selection 80% is used for model training in experiment, and the user conversation of residue 20% is used for the check precision of prediction and calculates the execution time.Algorithm realized by Java language, and dominant frequency is 1.6GHz, in save as 2GB.Because session is selected with randomness, experiment repeats 10 times, and the average of calculation of performance indicators.

As seen from Figure 6, aspect the Top-T precision of prediction, the inventive method can obtain the prediction effect that is better than complete 2 rank Markov models, and when T=10, the inventive method can reach the precision of prediction more than 80% on the experimental data collection.Aspect algorithm execution time, the inventive method significantly is better than complete 2 rank Markov models, and needed predicted time is less than the latter's 1%.

The present invention is directed to the forecasting problem of Web page access, access characteristic historical according to the targeted customer and pattern are effectively predicted or recommend next step page that will access of user, the information that helps user quick location to obtain; Help simultaneously Web server effectively to organize page structure and page cache is provided, promoting service quality.By above-mentioned specific embodiment as seen, this method can realize more accurate Web Web page predicting with less model complexity.

Claims (4)

1. Web page access Forecasting Methodology based on k rank mixing Markov model is characterized in that comprising following steps:

1) at first collect and arrange Web server access log data, for each Visitor Logs in daily record, identify customer end browser and user; Get rid of insignificant visit data; According to each record accessing operation o=＜u that extracts, x, t 〉, wherein u represents that user, x represent the Web page, t representation page access time;

2) identification user conversation S is used for setting up the Web log database, and deposit is used for the historical data of Web page access prediction;

3) choose and organize daily record data according to target of prediction from database, arrange and tissue (k+1) tuple-set by session; Flow process is: at first select the user and obtain session data based on target of prediction; Then extract (k+1) tuple X=＜x take session as unit ₁, x ₂..., x _k, x _k+1, each (k+1) tuple belongs to same session, comprises one group of page of user's connected reference in session; The accession page of adjacent two (k+1) tuples allows to overlap; All (k+1) tuples consist of a data set at last

4) set up k rank mixing Markov model, and adopt this k rank mixing Markov model of greatest hope Algorithm for Training, then the based on data collection

The parameter set of study and calibration k rank mixing Markov model; Flow process is: k rank mixing Markov model is by k state-transition matrix { Λ ₁, Λ ₂... Λ _kAnd a weight vector A={ α ₁, α ₂..., α _kForm, make that in Web server, the page adds up to n, transition matrix Λ _jA n * n matrix, its element λ _jAfter (x, y) representation page x is accessed, page y accessed probability, i.e. conditional probability p (x after an interval j page in same session _k+1| x _K-j+1), page x wherein _K-j+1And x _k+1Be equal to respectively page x and y; All n in k state-transition matrix in k rank mixing Markov model ²K element, and the weights of the k in weight vector A, the parameter set that component model need to be trained;

At first data-oriented collection

Calculate the initial value of each parameter, in weight vector A, the initial value of element is:

${\mathrm{\α}}_j=\frac{1}{k},(1\≤j\≤k)$

For state-transition matrix Λ _j, make data set

In (k+1) tuple quantity be m, tuple X _i=＜x _{I, 1}, x _{I, 2}..., x _{I, k}, x _{I, k+1}, element λ _jThe initial value of (x, y) is:

Wherein δ () function is defined as:

$\mathrm{\δ}\left(b\right)=\left\{\begin{array}{cc}1& \mathrm{b\; is\; true}\\ 0& \mathrm{b\; is\; false}\end{array}\right.$

Then calculate the new parameter value of next iteration needs according to the parameter current value, be labeled as respectively α ' _jAnd λ ' _j(x, y) for calculating new parameter value, introduces hidden variable collection Θ={ θ ₁, θ ₂..., θ _m, hidden variable θ wherein _iRepresent tuple X _iThe random character of self; Make that j is the page number of interval access, θ _i(j) computing formula is as follows:

${\mathrm{\θ}}_i\left(j\right)=\frac{{\mathrm{\α}}_j\·{\mathrm{\λ}}_j(x_{i,k-j+1},x_{i,k+1})}{{\mathrm{\Σ}}_{l=1}^k{\mathrm{\α}}_l\·{\mathrm{\λ}}_l\left(x_{i,k-l+1,}{x}_{i,k+1}\right)}$

According to θ _i(j), can estimate that the new value of each parameter is as follows:

${\mathrm{\α}}_j^{\′}=\frac{1}{m}\underset{i=1}{\overset{m}{\mathrm{\Σ}}}{\mathrm{\θ}}_i\left(j\right)$

Distinguish at last substitution previous round parameter value and new parameter value, whether the likelihood function of calculation training collection restrains by judging likelihood function, determines whether k rank mixing Markov model trains complete;

5) accessing operation of based target user to the Web page identified nearest user conversation, the Web page of next step access of k rank mixing Markov model predictive user after application training.

2. the Web page access Forecasting Methodology based on k rank mixing Markov model according to claim 1, is characterized in that the described insignificant visit data of step 1) comprises the Visitor Logs of getting rid of non-page access record and being produced by auto-programmings such as reptiles.

3. the Web page access Forecasting Methodology based on k rank mixing Markov model according to claim 1 and 2, is characterized in that step 2) the process of identification user conversation be: make session S={o ₁, o ₂..., o _l, all operations o _iThe execution user identical, then according to last accessing operation o _i=＜u, x _i, t _iThe time t that occurs _iAnd the page x that accesses _i, judge next operation o _i+1=＜u, x _i+1, t _i+1Whether belonging to same session S, this is judged based on following three conditions:

Page x _i+1By page x _iQuote, i.e. x _i+1URL be contained in page x _iIn;

Page x _i+1By the x that removes that was accessed in session S _iOutside other webpage t _jQuote, and t _jAnd t _i+1Mistiming less than the session threshold value of setting;

Page x _i+1Do not quoted by the page of being accessed in session S, but last operation t _iAnd t _i+1Mistiming less than the page threshold value of setting;

If satisfy one of above-mentioned condition, decision o _i+1Belong to session S; Otherwise operation o _i+1Open a new session.

4. the Web page access Forecasting Methodology based on k rank mixing Markov model according to claim 1 and 2, is characterized in that step 5) in the process of the Web page of next step access of k rank mixing Markov model predictive user after application training be: the current sessions of at first identifying the targeted customer; Then obtain k the page:＜x of user's connected reference recently in current sessions ₁, x ₂..., x _k, next for all page y on Web server, design conditions probable value p (y|x ₁, x ₂..., x _k):

$p\left(y\right|x_1,x_2,...,x_k)=\underset{j=1}{\overset{k}{\mathrm{\Σ}}}{\mathrm{\α}}_j\·{\mathrm{\λ}}_j\left(x_{k-j+1,}y\right)$

Front T page conduct selecting at last to have most probable value looked ahead or recommended user's the page.

Images