KR20050005591A - A System of Web Information Prediction Dynamic Recommendation in Internet Environment and A Methode Thereof - Google Patents

Abstract

PURPOSE: A system and a method for dynamically predicting/recommending web information based on the Internet are provided to predict/offer the web information most needed from each user without depending on a feed-back technique and a user profile. CONSTITUTION: A preprocessor(10) processes only the information needed for constructing a model to a proper form by extracting click stream information of the users from log data of a web server and refining/filtering the log data. A predictor(31) makes a web page preference predicting model for the user with the ESVM(Ensemble Support Vector Machine) using the data processed in the preprocessor as learning data and applying kernel functions, and predicts user's preference of each page by using test data. A clustering part(32) clusters the users by using the HSOM(Hybrid Self-Organizing feature Maps). A recommender(40) recommends the web information to a new user based on a result predicted by the predictor and clustered by the clustering part. A feed-back part(50) feeds bask satisfaction of the user for the recommended web information.

Description

A System of Web Information Prediction Dynamic Recommendation in Internet Environment and A Method Thereof}

The present invention relates to a dynamic web information recommendation system and method based on the Internet, and more particularly, to predict web information required by each user without relying on explicit feedback techniques and user profiles of the web information service. The present invention relates to a dynamic web information power converter and a method that can be provided.

The popularization of the Internet has led to an exponential increase in the amount of information and netizens have been able to obtain information scattered around the world at very low cost, as well as the development of the Internet has changed the overall social structure. Most of human activities depended on the Internet so that they could not live. It is remarkable that Internet users spend most of their time searching for information in this environment, and the rest of the time is spent in the wrong place and not in the user's area of interest.

However, despite such a situation, the conventional information recommendation system has not yet achieved satisfactory results in terms of information accuracy and speed. The reason is that most conventional web information recommendation systems obtain necessary information by relying on explicit feedback, which is information obtained directly from the user by using user's profile, content, product, and the like. In the space, web users are cumbersome and do not want to reveal their own feelings, and they do not want to disclose or accurately describe their information. This is because the satisfaction is not explicitly fed back to the web server.

Therefore, it is very difficult to obtain useful information from the existing web information recommendation system. Therefore, it is difficult to obtain a stable response from all users, which causes the scarcity of the data, and it is not a discrete data such as rating information. There has been a problem that there is no way to deal with the continuity data that appears in the form of feedback information.

Accordingly, the present invention provides a dynamic web information recommendation system capable of predicting and providing the web information that each user needs most without relying on explicit feedback techniques and user profiles of the Internet-based web information service. For the purpose of

In addition, the present invention solves the scarcity problem of web log data by designing and using an Ensemble Support Vector Machine (hereinafter referred to as "ESVM"), and Hybrid Self-Organizing feature map (Hybrid Self-Organizing feature) Dynamic web information recommendation system that performs clustering of web users using "Maps" (hereinafter referred to as "HSOM") to perform differentiated web information services using representative characteristics of the corresponding clusters most similar to those of new users. The purpose is to provide.

1 is a configuration diagram of a dynamic web information recommendation system according to the present invention

2 is a flowchart illustrating a dynamic web information recommendation method according to the present invention.

3 is a conceptual diagram of an ESVM model

4 is a structure of web data having rare characteristics

5 is a structural diagram of a web page prediction model

6 is a flowchart for determining an optimal function

7 is a flowchart illustrating a process of obtaining an optimal abnormal plane.

8 is a flowchart illustrating an initialization process of the HSOM.

9 is a flowchart illustrating a process of determining a victory node of HSOM.

10 is a flowchart showing an update procedure for the weight distribution of the HSOM.

* Explanation of symbols for main parts of the drawings

10: preprocessor, 20: knowledge base

30: learning unit, 31: prediction unit

32: clustering department, 40 recommendation

50: feedback unit, 100: web information recommendation system

Dynamic web information recommendation system according to the present invention for achieving the above object is to extract the click stream information of the user from the log data of the web server, and then to refine and filter the extracted log data is only appropriate information necessary for building the model A preprocessing unit for processing into a form; By using the data processed by the preprocessor as training data and using the ensemble support vector machine (ESVM) to which a plurality of kernel functions are applied, a web page preference prediction model is created for the user, and the user for each page using the test data. A prediction unit for predicting the preference of the; A clustering unit for clustering users using hybrid self-organizing shape maps (HSOMs) with complete information from which the scarcity is removed from the prediction unit; A recommendation unit predicted by the prediction unit and recommending web information to a new user based on the clustering result in the clustering unit; And a feedback unit for feeding back a user's satisfaction with the recommended web information.

In addition, the dynamic web information recommendation method according to the present invention for achieving the above object comprises the steps of obtaining the user click stream information from the log data of the web server; Purifying and filtering the extracted log data to extract only information necessary for building a model into a proper form; Creating a web page preference prediction model for the user by using the purified log information as a single instance by collecting the pages and time spent on the page as a user, and extracting an appropriate number of instances as the training data. ; Applying information obtained using the test data to an ensemble support vector machine (ESVM) model of an entire web page to predict a user's preference for each page; Calculating a preference in consideration of the average interest of each user and the average interest of each page in the prediction value for each page obtained through the above processes; Completing missing data of the rare web data to create a complete data structure; Using the complete data obtained in the above process and performing clustering for a user using a hybrid self-organizing shape map (HSOM) algorithm; Calculating a representative preference of users belonging to the cluster by calculating an average access time for the web pages of the users belonging to each cluster; And intelligentizing the knowledge base by receiving a feedback on the satisfaction of the result recommended by the user.

Hereinafter, the present invention will be described in more detail with reference to the drawings.

1 is a block diagram showing the configuration of a dynamic web information recommendation system according to the present invention. As shown in FIG. 1, the dynamic web information recommendation system 100 according to the present invention includes a preprocessor 10, a knowledge base 20, a learning unit 30, a recommendation unit 40, and a feedback unit 50. First, the preprocessing unit 10 extracts only the click stream information necessary for applying the first web log data to the recommendation system. The knowledge base unit 20 stores preference information updated through feedback from the user about the result predicted by the learner 30 and the result recommended for the new user. That is, the knowledge base unit 20 continues to be intelligent based on the results of the dynamic web information recommendation system learned through the new user.

In addition, the learning unit 30 includes a predicting unit 31 and a clustering unit 32. The predicting unit 31 uses the ESVM algorithm to process the information having the scarcity refined by the preprocessor 20 using an ESVM algorithm. Removes and predicts a preference, and the clustering unit 32 serves to cluster the complete information whose scarcity is removed from the predicting unit 31 by using the HSOM algorithm.

The recommender 40 recommends web information predicted from the learner 30 to a new user. Finally, the feedback unit 50 receives the final satisfaction with the web information recommended by the recommendation unit 40.

Next, a dynamic web information recommendation method according to the present invention will be described with reference to FIG. 2.

The dynamic web information recommendation system shown in FIG. 1 predicts and recommends information through a process of several steps. In the first step, the recommendation system obtains click stream information of users from log data of a web server (S100). In general, web log data has various forms depending on the type of server generating the log, the nature of the web site, and the nature of the subject providing the site. Therefore, the initial log data extracts only the information necessary for building the model through the preprocessing process in an appropriate form.

The log information refined in the above process is used as a single instance by collecting the page and time spent on the page, and extracting the appropriate number of instances as the training data to predict the web page preference for the user later. It is used to

The next step is to remove the scarcity of the web log data (S110). In general, among the numerous web pages in the web server, the number of web pages viewed by a user in a single connection is relatively small. Therefore, instances for individual users remain as missing data in a large number of cells. In order to solve such a problem of scarcity of web log data and to reduce learning time, the present invention predicts web information using ESVM. .

Since the ESVM can make a predictive model for all web pages individually, it is possible to model rare data that cannot be analyzed by the existing web information recommendation algorithm.

In this step, the information obtained from the test data is applied to the ESVM algorithm of the entire web page to predict the user's preference for each page, where the test data is selected from among the data not included in the training data in the purified log file. Refers to randomly extracted data, which is used later for performance evaluation.

In addition, in this step, the preference is calculated by considering the average interest of each user and the average interest of each page in the prediction values for the pages obtained through the above processes. Through this process, rare web data forms a complete data structure in which all missing cells are filled.

The next step is a user clustering step (S120). In order to recommend web information to new users using the complete data obtained in the above process, clustering of all users is required. This is because these clusters recommend web information about new users.

Next, the step of calculating the representative preference of each cluster (S130). In the present invention, clustering is performed using HSOM, and the clustered results are used to obtain a representative preference. That is, the average access time for the web pages of the users belonging to each cluster is calculated, and the result is the representative preference of the users belonging to the cluster. Using this information, the similarity with each cluster is calculated for the new user and the nearest cluster is assigned.

Euclidean distance is used as a measure to calculate the similarity between the new user and each cluster. In the user clustering step, a page having a high priority is recommended using the representative preference of the corresponding cluster allocated to the new user.

In this step, as an evaluation method for the entire system, the measurement of the error degree between the given preference and the predicted preference as the minimum square error is also performed at the same time. In other words, a certain number of high priority pages and low pages are randomly extracted and used to evaluate the performance of the system by measuring how they reacted to the pages.

The information obtained through the above process is stored in the knowledge base unit, and the knowledge base unit is continuously intelligent based on the result of the learning by the new web user through the dynamic web information recommendation system (S140).

Next, the prediction unit 31 using the ESVM will be described in more detail.

The conventional SVM model proposed by Vapnik is a supervised learning algorithm with two models: support vector classification (SVC), which is a classification model, and support vector regression (SVR), which is a prediction model. Use one kernel function to determine. However, when the performance of the conventional model is poor, the conventional model needs to learn again by using another kernel function.

Therefore, the present invention is based on the fact that it is desirable to determine the best kernel function through a single learning using a plurality of kernel functions during the learning process, using ESVM in advance to a given data using a plurality of kernel functions in advance. Provide a way to learn about This method uses an ensemble structure in which the best kernel function is determined in the process of constructing the prediction model using the given training data. This ESVM method compares the predicted SVM with one kernel function. It is characterized by higher accuracy.

On the other hand, learning using ESVM requires more time than the number of kernel functions applied when using multiple kernel functions than multiple kernel functions. For example, if ESVM builds a model of five kernel functions, it will take five times more learning time than SVM. However, in this case, the dynamic web information power conversion unit cannot be fast predictive power. Therefore, in order to solve the problem of prediction time, the bootstrap technique is applied in the present invention. In other words, the process of selecting the optimal kernel function using multiple kernel functions does not use all of the data, but makes a sample of the number of kernel functions by resampling from the entire training data. Each sample created at this time can be seen as the same sample. This is because the bootstrap resampling technique was used. In addition, because of the bootstrap technique, each sample can represent the entire data but is very small compared to the size of the entire data.

As a result, since the dynamic web information power converter according to the present invention performs learning using only very small sample data, the learning time for obtaining the kernel function has little effect on the learning time performed to construct the entire ESVM model. do. Of course, it takes more learning time than SVM, but the percentage of total learning time is very small.

3 illustrates the concept of an ESVM model, in which a thick thick line in the center becomes an ESVM, and the shape of the thick thick line is determined by an optimal kernel function performed in the ESVM. If the kernel function is non-linear and non-linear, the ESVM in the center of FIG. 3 appears as a curve.

In the present invention, the goal of ESVM design is to obtain a central thick straight line, that is, a hyperplane, as shown in FIG. 3. The optimal hyperplane is a classifier that maximizes the width (1 / w) with each individual object. (classifier) Kernel functions apply various function types to learn using sample data and then select the function that shows the best results. That is, the kernel function of the result with the smallest MSE value is determined.

Meanwhile, FIG. 4 illustrates a refined web data structure to which ESVM is applied. Thus, the structure of the refined web log data to which ESVM is applied has a scarcity structure.

In FIG. 4, a column represents each page existing in the web server, and a row represents a session ID. Each cell represents the duration time each user spent on the page.

As shown in FIG. 4, the characteristic of the click stream information of the web log data is that the data has a very sparse structure. The number of pages a customer sees on the web site is very small compared to the total number of pages on the web site, so the cell values for the remaining pages are all empty. Therefore, it is very difficult to generate predictive model through analysis of such rare web data with general data mining algorithm. Therefore, to solve this problem, we propose and apply an ESVM model.

This model is constructed for each web page and is created as a predictive model based on the preferences for the remaining pages. In other words, it is a regression structure in which a specific page to be predicted becomes a target variable and the remaining web pages become explanatory variables.

5 visually illustrates the structure of such a web page prediction model.

In FIG. 5, the web page to be predicted is k , and a simple structure of two pages, i.e., page i and page j , is a web page which serves as an explanatory variable for the target variable.

Each point in FIG. 5 represents one user and the axes in the plane represent the page browsing time already passed by the user, and the vertical axis represents the browsing time for the page. Therefore, the figure predicts the browsing time for the page of the user according to the time of other pages except the page.

For example, in the rare web page data structure of FIG. 5, the preference prediction model for the i- th page may be expressed as Equation 1 below.

Affinity _{page i} = f (page ₁ ,…, page _i-1 , page _{i + 1} ,…, page _M )

Equation 1 can be used to predict the preference for the i th page. That is, the web page having the highest value is recommended to the user by calculating the preference for the web page in which the corresponding cell is empty because there is no access information for the specific user. In Equation 1, the function, f (·), is a kernel function obtained by applying the ensemble technique.

6 and [Algorithm 1] illustrate an ensemble technique and bootstrap resampling procedure for determining the optimal kernel function.

Algorithm 1

[ Algorithm: Voting_Kernel (Sel_ker [k])]

// determine the optimal parameter using multiple kernel functions

Choose optimal kernel K * such that min? Ω?

// 1-scatter smoothing, 2-bin, 3-running mean, 4-kernel smoother

// 5-equivalent kernel, 6-regression spline, 7-cubic smoothing spline

// bootstrap sampling

for (i = 1; i = <N; i ++)

if random_number <0.1

re_sampling // Apply Bootstrap's Resampling Technique

end

// determine optimal kernel function

for (k = 1; k = <7; k ++)

MSE [k] = risk (Sel_ker [k]) // Calculate Minimum Square Error of Hazard Function Values

if mse [k] = min;

voting // Select kernel function with smallest square error

end

As can be seen in Figure 6, the present invention various modifications to the kernel function in order to use the ensemble technique as a model construction strategy (S200). The kernel functions used in the present invention mainly include a simple scatter smoothing function, a bin function, a running-mean function, a kernel smoothers function, an equivalent kernels function, Regression spline function and cubic smoothing splines function are used. It can also be used as a spline family of additive functions that can be used as kernel functions for continuous data.

The process of determining an optimal parameter using the plurality of kernel functions described in FIG. 6 first undergoes a bootstrap resampling process to shorten the learning time (S210). After calculating the least square error of each risk function value (S220), the kernel function having the smallest least square error value is selected as a function to be applied to the ESVM (S230).

7 and [Algorithm 2] illustrate the process of applying an ensemble to a kernel function and obtaining an optimal ideal plane.

Algorithm 2

[ Algorithm: Find_Hyperplane (Input [i])]

// Algorithm for constructing an ideal plane using input data

Choose (1) an origin point <ω, x ₁ > + b = +1 , <ω, x ₁ > + b = -1

Choose (2) initial ω ₀ such that y _i <ω, x _i > + b ≥ 1 , for all i

// find the ideal anomaly plane

Calculate Support Vector through {x | <ω, x ₁ > + b } = 0

Selection if minimize subject to y _i <ω, x _i > + b ≥ 1 , for all i

then 'best hyper-plane found'

else choose optimization problem

determine the ω ^* with minimal margin

^{if (y *> 0) then} ξ i = (X - + X *) / 2 else ξ i * = (X + + X *) / 2

ω _h = ω _k-1 -Δω and ω _k '= ω _k / ∥ω _k ∥

if [y ^* f (x ^* )] <0 after update then go to Choose (1)

Repeat Calculate and Selection maximally min { N, l } times.

// approximate hyper-plane discovery

Go to Choose (1), or if maximal trials is attained,

'approximative hyper-plane found'

In FIG. 7, first, a reference point is selected (S300), and an initial value ω ₀ of a weight parameter representing a group interval is satisfied for all individuals (S310), and then an optimal ideal plane is selected. In order to find out, a support vector (hereinafter referred to as "SV") satisfying the following Equation 3 is obtained (S320).

y _i <ω, x _i > + b ≥ 1 , for all i

{x | <ω, x ₁ > + b } = 0

Then, the plane having the minimum margin while satisfying the excellent ideal plane condition represented by the following [Equation 2] for all the objects is regarded as the optimal ideal plane (S330). If not, however, the optimization problem is selected as shown in Equation 4 below to determine ω ^* having a minimum margin (S340).

, Where Q stands for ∑.

Then, a predetermined calculation is performed according to the size of y ^* value, or data is updated and step 1 is executed again (S350). This process is repeated for the smaller of N and l. After that, if the optimum abnormal plane is not found, the near superplane is selected (S360). The above procedure determines the optimal model of ESVM for one kernel function. In this procedure, the solution of the optimal equation is determined by Lagrange expansion. By repeating the above process, the SV calculated by the final ESVM model is obtained (S370), and through this, the visit time of the web page for the user can be predicted (S380).

Next, the clustering unit 32 using the HSOM will be described in more detail.

After the scarcity of the data is removed by using the ESVM in the prediction unit, clustering for each user is performed by using the HSOM. First, the design of the probability structure of HSOM is based on Gaussian distribution. For clustering using the given data, it is necessary to determine the concrete structure of the prior probability distribution, the likelihood function and the posterior probability distribution of HSOM. The reason why the Gaussian distribution is used in the present invention is that the data of the page access time of the web user follows the Gaussian distribution and facilitates the application of Bayesian inference. Therefore, before starting clustering, the prior probability distribution, the likelihood function, and the posterior probability distribution of the Gaussian distribution must be determined. In the present invention, the prior probability distribution regarding the weight is assumed to be a Gaussian distribution as shown in Equation 5 as an embodiment.

Equation 5 used as a hyperparameter of the initial weight distributionμHas a Gaussian distribution with mean 0 and variance 1.σ _w ² Is a known fixed value. This prior probability distribution is a Gaussian distribution, resulting in a conjugate prior probability distribution. sureσ _w ² You can define a hyperparameter for, usually using an inverse gamma distribution.

The likelihood function, which is the distribution of the training data x , is also assumed to be a Gaussian distribution as shown in [Equation 6].

In Equation 6, w has a prior probability distribution, and in general, σ _w ² is assumed to be a known fixed value. The posterior probability distribution is calculated by Bayes' theorem.

HSOM updates the posterior probability distribution for each learning resource. So we should consider the likelihood function for a single piece of data. [Equation 7] is a calculation formula for calculating the posterior distribution.

In the post hoc distribution process of Equation (7), the weights are averaged through the likelihood function of prior probability distribution and learning data. , The variance is It can be seen that it is updated to the Gaussian distribution. The posterior probability distribution of w in Equation 7 is to learn one data and become an updated prior probability distribution for new training data. In addition, it can be seen from the above equation that the variance of the posterior probability distribution is reduced compared to the prior probability distribution. As training progresses, as the model fits into the training data, the variation of the weight update decreases. However, to prevent the variance from becoming too small, if the variance becomes smaller than the allowable limit, there is no further reduction in variance.

8 is a flowchart illustrating an initialization process of HSOM, and [Algorithm 3] lists codes for performing initialization of HSOM.

Algorithm 3

Algorithm: Initialize_HSOM (Input [i])

// initialization of HSOM 1: initialization of network constants

Initialize network parameter; Input_layer (int i, int o, int init_neigh_size)

num_inputs = i, num_outputs = o // Initialize the input and output vectors

neighborhood_size = init_neigh_size // Initialize the neighbor radius of the shape map

weights = new float [num_inputs * num_outputs] // specify weights

outputs = new float [num_outputs] // specify outputs

// initialization of HSOM 1: initialization of probability distribution

Initialize bayesian distributions // Determine Bayesian Probability Distributions

Initialization of the weight vector, w _j (0) to have

probabilistic distribution, N (0, 1).

// the distribution of the initial weights is a Gaussian with mean and variance of 0 and 1, respectively

Determined by distribution

Initialization of the learning rate α (0), α (t) ∝t ^-α ; 0 <α <1 ,.

// determine the learning rate function

Initialization of the neighborhood function K (j, j ^* ) ,

K decreases as to increase | jj ^* |

// Determine neighbor radius function

Where i is the input vector, o is the output vector, and init_neigh_size is the neighbor radius.

As shown in FIG. 8 and the algorithm, the initialization process of the HSOM first initializes the weight vector (S400). At this time, the initial weight vector w _j (0) has a probability distribution. Next, the learning rate is initialized (S410). The initial learning rate α (0) must satisfy the conditions α (t) ∝ ^t-α and 0 <α <1 . In the next step, the neighbor function (or “neighbor radius function”) K (j, j ^* ) newly introduced in the present invention is initialized (S420). Neighbor function | jj ^* | As the value increases, the K value decreases.

This initial stage of HSOM must first consider the prior probability distribution of the weights. This distribution has prior knowledge of the weights from the structure of the model to be constructed. In general, the initial weight distribution is the prior probability distribution of the exponential family, as shown in Equation 8. distribution).

Here, Z _w (α) is a normalization factor as shown in Equation 9 below, and α is a hyper parameter.

In addition, the probability distribution E _w has the form as shown in [Equation 10].

In Equation 10, M is the total number of weights of the model. Specifying the prior probability distribution as an exponential group makes it easier to calculate the posterior probability distribution. Of course, the prior probability distribution can be determined by any distribution other than the exponential group.

In addition, the parameter of this prior probability distribution has a hyperparameter. For example, assuming that the weight vector is a random variable following a Gaussian distribution, this Gaussian distribution has different parameters for each component in the weight vector. That is, the i th weight ω _i follows a Gaussian distribution with mean μ _i and variance σ _i ² , where μ has a prior distribution of N (0, 1) , where variance is It is possible to determine with a known fixed value or consider an initial weight with a prior distribution of an inverse gamma distribution with position and shape parameters.

In order to remove the influence of different scales between the nodes in the p-dimensional input vector, normalization of each input node is required. Normalization of the i th input node x _i = (x _i1 ,…, x _ip ) is calculated by subtracting the total mean of each node from the values of each node and dividing this value by the total standard deviation of each node. That is, the standardized i th input node is expressed by Equation 11 below.

As a measure for determining the victory node in the shape map of the HSOM, the euclidean distance between the standardized input vector and the output node is used. The Euclidean distance between the weight of the i th input node and the j th output node can be calculated using Equation 12.

The determination of the victory node w for the i th input vector is made with the node having the smallest distance, dist (x _i ^normal , w _j ) . In HSOM, the region size of the winning node is initially determined to be quite wide. However, as learning progresses, the range gradually decreases. This decreases the range of true values of the parameters as learning progresses. Similarly, HSOM follows the method of determining the area size of this SOM.

For each learning, HSOM, which updates the weights of victory nodes and nodes within its neighborhood size, needs to consider the distribution of update distributions that are not considered in SOM. This is because, unlike the mean of the distribution, variance needs to be gradually reduced as the model is optimized through learning.

The variance of the post-probability distribution by the likelihood function of the weights and the prior probability distribution for the weights is smaller than the variance of the prior distribution by Bayes' theorem. Since this post-distribution becomes a pre-distribution of new learning data, the variance of the update distribution becomes small. As the model becomes intelligent, the variance becomes smaller, so consideration of the reduction of variance along with the learning rate of SOM is naturally made in HSOM.

Once the prior probability distribution and the likelihood function are determined, the Bayes' theorem gives the form of the posterior probability distribution of Equation (13).

_{Here, S (w) = βE D} + and _{αE w, Z s (α,} β) = ∫exp - a (-βE _D αE _w) dw. Solving S (w) again and summarizing [Equation 14].

The mean of the new weight distribution to be updated in Equation (14) is the expectation for the mean of the posterior probability distribution by the product of the prior probability distribution and the likelihood function of the learning data, which are considered beforehand regardless of the training data. Is calculated. In this case, the integral for calculating the expected value of the posterior probability distribution is obtained by the Monte Carlo method, among which the Metropolis algorithm is commonly used. In the markov chain defined by the Metropolis algorithm, the new state, θ ^{(t + 1),} becomes the candidate state of the specified proposal distribution generated from the previous state, θ ^(t) . . This candidate state is accepted or rejected by the probability values of the previous state and the original distribution.

Algorithm 4 below shows the process of determining the size of the HSOM layer.

Algorithm 4

Algorithm: HSOM_Size (Layer [k])

Determine initial size of feature maps // Determine dimension of shape maps

for (i = 0; i <= max; i ++)

choose layer_size,

input_layer_size [0]

feature_layer_size [i] // Determine dimension of input layer and shape map

end

Set layers // Determine hierarchy of shape map

current layer replaces as new feature_layer (layer_size [i])

// Determine hierarchy structure according to the change of dimension

for (i = 0; i <= max; i ++)

if (layer_ptr [i] == convergence_const)

// investigation of the convergence of dimensions

end

The process of determining the size of the HSOM layer consists of two hierarchical structures. First, the initial hierarchical structure of the HSOM is determined before learning using the given learning data. In this step, the input structure of the training data is initialized, and the hierarchical size of the shape map for the first learning from this data is determined. The structure of the input data once determined does not change until the end of HSOM's final learning, but the hierarchical size of the shape map increases sequentially during the learning, until the given conditions are met and the convergence of the optimal cluster results. Continue until it is done.

9, 10, and [Algorithm 5] show an update procedure for weight distribution of HSOM by using a victory node determination of a SOM model and a probabilistic distribution update strategy of weights.

Algorithm 5

Algorithm: Train_HSOM (Train [j])

// update the Bayesian probability distribution of HSOM

Determine the winner node // determine the winner node

// normalize input vector along Gaussian distribution

Normalization of input vector, Gaussian distribution with mean 0, variance 1

// determine hyper = parameter of weight

Choose the distribution of weights, w ～ f (θ)

// determine the node with the minimum Euclidean distance as the winner

Choose the winner node j ^* = arg max y , using Euclidean criteria.

Update of parameters // Update Parameter of Weight Distribution

w _j ^New = w _j ^Old + α (j) K (j, j *) (X-w _j ^Old ),

where, K (j, j ^* ) ; Neighborhood function

end

// iterate the parameter update task of the weight distribution

Replace old distribution by current.

// Determine Victory Node Of HSOM

Set Winner_index = 0,

maxval = -1000000 // Specify maximum number of iterations

Find the winner neuron // Determine Winning Node

for (j = 0; j <= num_outputs; j ++)

for (i = 0; i <num_inputs; i ++)

// node with minimum Euclidean distance becomes the winning node

winner [i, j] = arg min ∥ x (k)-w _j ∥

end

// update weight distribution

Set m = (int) alpha, delta = alpha-m

// create parameter from Gaussian distribution

while (count <data_size)

v = rand () / 32767.0, w = pow (y / m, delta) / (1+ (y / m-1) * delta)

end

// iterate until convergence

Repeat Until given criteria satisfaction.

As shown in FIG. 9 and the algorithm, the present invention first initializes an input vector (S500) and generates weights to be used for nodes of each shape map in an initial Gaussian distribution (S510). The initial predistribution uses a Gaussian distribution with a mean of 0 and a variance of 1. This is because the Euclidean distance is used to determine the triumph node in the shape map of HSOM, and the values of the input vectors are initialized such that the mean follows zero variance. Therefore, updating the weight from the beginning is updating the parameters of the distribution to which they belong. These values are updated by subsequent learning.

Next, the victory node of the HSOM is determined (S520). In this step, the maximum range of a given iterative learning is specified, and the node of the shape map whose final weight value is to be updated is determined.

10 shows a procedure for updating the distribution of weights. Calculation of the new weight (S600) is performed as shown in Equation 15 below.

w _j ^New = w _j ^Old + α (j) K (j, j *) (X-w _j ^Old )

Here, K (j, j ^* ) means a neighbor function. At this time, an update is made from the distribution of the hyperparameters for the parameters of the Gaussian distribution, which is the weight distribution (S610). The algorithm uses a gamma distribution as a distribution for the hyperparameters, because the Bayesian posterior probability becomes a conjugate distribution, making calculations easier.

As such, the dynamic web information power converter plays a very important role in the learning module in which the prediction unit using the ESVM and the clustering unit using the HSOM are organically combined.

Through the above process, preference prediction information for a web page that can be finally used in the recommendation system for a new user can be obtained, and the information is stored in the base unit.

As described above, in the dynamic web information recommendation system according to the present invention, a series of processes for recommending suitable web information for each user is performed in real time, and in terms of accuracy, learning time for prediction, and convergence of the model, compared to the existing recommendation system. More excellent properties. In particular, it is proved through experiments that the learning time is inferior to the existing recommendation algorithm SVM, and the performance of the model is superior to the existing recommendation system. In addition, the experimental results show that the accuracy is superior to Pearson's correlation coefficient algorithm, which is currently used most.

Claims (9)

In the system for providing web information to the user in the Internet environment, A pre-processing unit extracting click stream information of users from log data of a web server, and then refining and filtering the extracted log data to process only information necessary for building a model into a proper form; By using the data processed by the preprocessor as training data and using the ensemble support vector machine (ESVM) to which a plurality of kernel functions are applied, a web page preference prediction model is created for the user, and the user for each page using the test data. A prediction unit for predicting the preference of the; A clustering unit for clustering users using hybrid self-organizing shape maps (HSOMs) with complete information from which the scarcity is removed from the prediction unit; A recommendation unit predicted by the prediction unit and recommending web information to a new user based on the clustering result in the clustering unit; And Dynamic web information recommendation system comprising a feedback unit for feeding back a user's satisfaction with the recommended web information. The method of claim 1, wherein the prediction unit further comprises a bootstrapping unit, which serves to resample a sample from the entire data, and a kernel determination unit that determines an optimal kernel function to shorten the learning time of the model. Features a dynamic web information recommendation system. In the method of providing web information to the user in the Internet environment, Obtaining click stream information of users from log data of a web server; Purifying and filtering the extracted log data to extract only information necessary for building a model in a suitable form (S110); Creating a web page preference prediction model for the user by using the purified log information as a single instance by collecting the pages and time spent on the page as a user, and extracting an appropriate number of instances as the training data. ; Applying information obtained using the test data to an ensemble support vector machine (ESVM) model of an entire web page to predict a user's preference for each page; Calculating a preference in consideration of the average interest of each user and the average interest of each page in the prediction value for each page obtained through the above processes; Completing missing data of the rare web data to create a complete data structure; Using the complete data obtained in the above process and performing clustering for a user using a hybrid self-organizing shape map (HSOM) algorithm; Calculating a representative preference of users belonging to the cluster by calculating an average access time for the web pages of the users belonging to each cluster; And Dynamic Web information recommendation method comprising the step of intelligentizing the knowledge base receiving feedback on the satisfaction of the results recommended by the user. The method of claim 3, wherein the predicting the preference comprises: Determining a kernel function; Determining an optimal parameter; Bootstrapping sampling step; Determining an optimal kernel function; Optimal out-of-plane determination; And a rare data building step, wherein the determining of the optimal kernel function comprises selecting one function among the plurality of kernel functions to minimize an error. 5. The kernel of claim 4, wherein the plurality of kernel functions include a scatter smoothing function, a binary function, a running mean function, a kernel smoother function, and an equivalent kernel. Dynamic web information recommendation method characterized in that it is a function, a regression spline function or a cubic smoothing spline function. The method of claim 4, wherein the determining the optimal out-of-plane Reference point selection step; A weight initialization step; A support vector calculation step; Dynamic web information recommendation method comprising the step of selecting the plane as an ideal plane, if the optimum abnormal plane is found. The method according to claim 6, wherein in the optimum abnormal plane selection step, Estimating a parameter using an updated support vector if an optimal anomaly is not found; Dynamic data recommendation method further comprising the step of updating the data. The method of claim 3, wherein the user clustering step, An initialization step comprising initializing a weight vector, initializing a learning rate, and initializing a neighbor function; A winner node determination step comprising initializing an input vector, selecting a weight distribution, and selecting a victory node; And Updating the parameter with a new weight and replacing the previous update with a new update distribution. The parameter updating step of the weight vector includes a Bayesian inference technique based on a probability distribution. A dynamic web information recommendation method characterized by ensuring convergence by applying. 10. The method of claim 8, wherein the determining of the victory node uses a Euclidean distance between an input vector and an output vector.