KR101601820B1 - Method and program for similar user index - Google Patents

Abstract

A similar user index method is provided in a recommendation system according to one aspect of the present invention. The similar user index method in the recommendation system according to an embodiment of the present invention includes a step of generating a social network in a recommender data set, a step of selecting a seed using the generated social network, a step of clustering , The clustered cluster is readjusted, and the cluster is indexed into a pseudo-user cluster. The present invention can improve the accuracy and reliability of recommendation items by applying a similar user clustering technique using a social network analysis technique.

Description

[0001] METHOD AND PROGRAM FOR SIMILAR USER INDEX [0002]

The present invention relates to a pseudo-user index method and program, and more particularly to a method and program for a similar user index based on a social clustering.

Recently, the recommendation system has been developed in the field of electronic commerce due to the explosive increase of web contents. The purpose of the recommendation system is to recommend an item or item to suit the user's taste, and to assist the user in selecting or purchasing items for selection. The personalized recommendation system is very important in the e-commerce field where the choice is diverse and customer's taste is important.

Large e-commerce sites such as Amazon and Netflix have successfully adopted a recommendation system that delivers automatically generated personalized referral information to their customers. Collaborative filtering (CF) recommendation system was the first and was the most successful method in the research or utility field. In the recommendation system, the collaborative filtering approach is to predict user 's preference for product or service by learning user' s past history data. That is, the recommendation system algorithm applying almost all the collaborative filtering methods generally considers various data sets such as evaluation values or objects for products experienced by users. For example, Amazon has recorded over 30 million users and millions of products and has been using it for collaborative filtering-based recommendation systems.

Many of the recommendation technologies have evolved over the past decades, but most of them have been made up of small datasets and were not realistic overall. Recent data sets have exploded and many recommendation systems have been difficult to handle larger data sets due to performance and scalability issues.

Therefore, it is required to handle a large number of data sets and to construct high quality recommendation and quick recommendation products.

The background art of the present invention is disclosed in Korean Patent Laid-Open Publication No. 10-2009-0130774 (published on December 24, 2009, a user recommendation method and a recording medium on which a program for this is recorded).

It is an object of the present invention to provide a pseudo-user index method and program that effectively improve the performance of the recommendation system in consideration of both accuracy and scalability of the recommendation system.

It is another object of the present invention to provide a pseudo-user index method and program that consider explosively increasing data size and do not deteriorate the performance of the entire recommendation system as the data size increases.

The objects of the present invention are not limited to the above-mentioned objects, and other objects not mentioned can be clearly understood from the following description.

In order to achieve the above object, a similar user index method in a recommendation system according to one aspect of the present invention is provided.

The similar user index method in the recommendation system according to an embodiment of the present invention includes a step of generating a social network in a recommender data set, a step of selecting a seed using the generated social network, a step of clustering , The clustered cluster is readjusted, and the cluster is indexed into a pseudo-user cluster.

In addition, a computer program is provided in which a similar user index method is executed in a recommendation system according to another aspect of the present invention.

A computer program for implementing the pseudo-user index method according to an embodiment of the present invention includes a step of generating a social network in a recommender data set, a step of selecting a seed using the generated social network, A step in which the clustered cluster is readjusted, and a step in which the cluster is indexed into a pseudo-user cluster.

The present invention is a new indexing method that overcomes the problems of existing recommendation algorithms to improve the performance of a large-capacity recommendation system, which is demanded more in the age of large-capacity data.

The present invention applies a social network analysis technique to improve the accuracy and reliability of recommendation items.

The present invention can improve the accuracy of recommendation items as well as the accuracy of clustering by applying a pseudo-user clustering technique.

By using the pseudo user cluster index, the present invention can improve the performance of the entire recommendation system more effectively by greatly reducing the computation time for similar user extraction in the recommendation technique.

INDUSTRIAL APPLICABILITY The present invention can improve the performance of the entire recommendation system without being greatly affected by the amount of data which is explosively increased due to the configuration of the index through the index based on the pseudo-user clustering.

1 and 2 are views for explaining a pseudo-user index method according to an embodiment of the present invention.
3 is a diagram for explaining a seed extracting method according to an embodiment of the present invention.
FIG. 4 illustrates a clustering method according to an embodiment of the present invention. FIG.
5 and 6 are views for explaining a cluster rebalance method according to an embodiment of the present invention.
FIG. 7 is a view for explaining a similar-user index method according to an embodiment of the present invention; FIG.
8 and 9 are views for explaining the effect of the similar user index method according to an embodiment of the present invention.

Hereinafter, embodiments of the present invention will be described in detail with reference to the accompanying drawings so that those skilled in the art can easily carry out the present invention. The present invention may, however, be embodied in many different forms and should not be construed as limited to the embodiments set forth herein. Also, when a part is referred to as "including " an element, it is to be understood that it may include other elements as well, without departing from the other elements unless specifically stated otherwise.

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS Reference will now be made in detail to embodiments of the present invention, examples of which are illustrated in the accompanying drawings.

1 and 2 are views for explaining a similar user index method according to an embodiment of the present invention.

The pseudo-user index method according to the present invention indexes pseudo-user clusters for a method of recommending items to users with similar tastes in a recommendation system. Accordingly, it is apparent that the pseudo-user index method according to the present invention can be used in a computer for computation, a server for providing service, and can be implemented in a program form.

Referring to FIG. 1, a pseudo-user index method includes a step S110 of generating a social network in a data set of a recommender, a step S120 of selecting a seed using a social network, a step S130 of cluster- , Step S140 in which the cluster is readjusted, and step S150 in which the similar user cluster is indexed.

In step S110, a social network is created in the recommender data set.

Social network theory can be used for recommendation systems of items for people, and is distinguished in the first mode and the second mode. This mode relates to a distinct set of entities that have the same contribution. From the perspective of the recommendation system, the first mode may be considered as a user. The second mode can be regarded as an item. That is, people's evaluation patterns on items in the recommendation system lead to explicit social networks and also to network connectivity.

Suggested data sets include user sets U = {u1, u2, u3, ... , un}, metadata of items I = {i1, i2, ... , ik} and the set of states S = {s1, s2, ...} , sl}. Here, the situation set can be configured by, for example, the number of times the items are commonly experienced by the user, the number of times of association with another user, or the evaluation value for the item, but is not limited thereto.

Referring to FIG. 2, users U1 and U2 commonly experience items I1 and I2. Therefore, a social network as shown in the right picture of Fig. 2 is generated in the recommender data set.

In step S120, a social network is used to select a seed.

Here, the seed means a node located at the center of each cluster. Most existing clustering schemes perform clustering by randomly selecting initial seeds, which causes degradation of clustering accuracy and is not suitable for similar user clustering.

In order to construct the seed with more reliable user nodes, the influential nodes are extracted from the social network generated in the first step of the present invention. At this time, the critical node is determined according to the number of relationships with other user nodes and the number of common experiences.

We use the social network depicted in the right picture of Figure 2 to select the seeds of each cluster. The weight of the relationship between users in a social network means the frequency of common items experienced by users. The more common items experienced by a pair of users, the more similar users can be. Therefore, it is assumed that the important users node has more relationships with other users. That is, an important node can be considered as a seed of each similar user cluster. This is because the primary user node has a higher probability of being associated with any other user node. The seed extraction step will be described in more detail in FIG.

Clustered around the seed in step S130.

Here, clustering is a data mining technique that finds interesting patterns in a given database. The main idea of clustering is that given n pieces of data are divided into k clusters in m-dimensional space so that all the data pointed to in one cluster is more similar than the data in the other cluster. Clustering methods in recommendation systems are operated by identifying groups of users with similar preferences. When clustering is perfect, the performance of the recommendation system can be very good, because the size of the data that needs to be analyzed is much smaller.

The clustering method according to the present invention calculates a distance value between a seed and a user node of each extracted cluster so that the user node belongs to the cluster of the seed having the smallest distance value from the corresponding user node. This will be described in more detail in FIG.

In step S140, the set cluster is readjusted.

In the cluster rebalancing step, the clustering step is performed to rearrange the users belonging to each cluster to improve the accuracy of similar user clustering. In the cluster rebalancing step, a cluster rebalance algorithm applying Q, which is an accuracy evaluation value of network-based clustering, is applied. This will be described in more detail in FIG. 5 and FIG.

In step S150, the pseudo-user cluster is indexed.

The similar user cluster index step is a step for improving the performance of the recommendation system by effectively removing unnecessary data in the recommendation algorithm operation as well as improving the accuracy of the recommendation item. This will be described in more detail in FIG.

3 is a view for explaining a seed extracting method according to an embodiment of the present invention.

Referring to FIG. 3, in step S310, nodes of the top N (where N is a natural number, equal to the number of the seeds set in advance) having a high degree of centrality are extracted. Here, the importance may be the number of times that the nodes are related to each other.

For example, priority nodes U2 = 3, U3 = 4, U1 = 2, U7 = 2 are extracted first in FIG.

In step S320, among the extracted nodes, nodes having a large relation with other nodes are selected as seed candidates.

For example, in Fig. 2, U2 (3) and U3 (4) are selected.

In step S330, seeds are extracted in descending order of importance by summing the importance (the number of times the nodes are related to each other) of the neighbor nodes connected to the seed candidate nodes.

For example, in Fig. 2, U3 = 2 + 1 + 4 = 7 and U3 = 2 + 2 + 3 + 1 = 8.

The seed among the seed candidate nodes extracted in step S340 is arbitrarily extracted by the predetermined number of clusters.

4 is a diagram for explaining a clustering method according to an embodiment of the present invention.

Referring to FIG. 4, clusters are allocated to extraction seeds in step S410.

In step S420, the initial center value and the number of included nodes 1 are assigned to each cluster. Here, the initial center value is the position value of the extracted seed.

In step S430, the nodes are allocated to the clusters having the shortest distance from the seed using the Euclidean distance value based on Euclidean distance values for each of the nodes except the seed. Here, the distance is regarded as the degree of similarity between the user node and the seed.

Here, TN is the total number of items experienced by the user node or seed, and MN is the number of items that a pair of nodes have commonly experienced.

5 and 6 are views for explaining a cluster rebalance method according to an embodiment of the present invention.

In the cluster rebalancing step, the accuracy evaluation value Q is applied to rearrange the users belonging to each cluster in the configured clusters and to improve the accuracy of the similar user clustering. Here, Q, which is an accuracy evaluation value, has a higher accuracy with a value of 1, and the clustering is readjusted so that Q becomes a maximum value.

First, the first accuracy evaluation numerical value Q of each cluster is calculated.

Referring to FIG. 5, V1 and V2 are initially allocated to the first cluster, and V3, V4, and V5 are allocated to the second cluster.

When the Q value is calculated based on this,

Here, TR (e) denotes a portion of edges connected with user nodes in the closter, and a good cluster has a high value. a _k ² is the portion of the edges expected in the cluster when the edges are randomly distributed in the social network. ∥e∥ is the sum of matrix e elements.

Thereafter, two arbitrary nodes belonging to different clusters are extracted and swapped into the different clusters, and second accuracy evaluation numerical values Q of different clusters are calculated.

Referring to FIG. 6, after the V3 is readjusted to the first cluster, a second Q value is calculated.

Finally, when the first accuracy evaluation numerical value is compared with the second accuracy evaluation numerical value and the second accuracy evaluation numerical value is larger, the extracted nodes are swapped into different clusters and the cluster is readjusted.

For example, the second Q value-1.05476 is calculated, and then the second Q value -1.05476> the first Q value -1.81852, so that the cluster is readjusted like the second Q value.

As described above, two arbitrary nodes belonging to different clusters are extracted, the cluster IDs are continuously exchanged, the Q value is calculated, and the calculation is repeated so that the Q value has the maximum value.

7 is a view for explaining a similar user index method according to an embodiment of the present invention.

Referring to FIG. 7, the pseudo-user index method, which is the last step of the present invention, constitutes a final index with a cluster constituted in the pseudo-user clustering step.

Clusters have their own cluster IDs. The recommendation system can use such a cluster ID and define a hash function as shown in FIG. 7 to quickly access a similar user index. Given a user ID, users similar to this user ID can be retrieved through a hash function. A similar set of users retrieved represents a similar user cluster. A similar user index consists of a cluster ID dictionary and a list of user IDs. Here, the cluster ID dictionary records the cluster ID together with its own center value. In the defined hash function, the similarity degree between a given user ID and the center of each similar user cluster is calculated, and a given user outputs a cluster ID having the maximum value of similarity according to the calculated similarity.

8 and 9 are views for explaining the effect of the pseudo-user index method according to an embodiment of the present invention.

Referring to FIG. 8, the efficiency of the pseudo-user clustering-based indexing method for large recommendation systems is evaluated.

The data set of the recommended system arbitrarily generated for the efficiency evaluation is as follows.

designation Number of user IDs Data size (MB) D1 3706 1.4 D2 3580 0.97 D3 3430 0.85 D4 3310 0.78 D5 2940 0.71 D6 2610 0.67 D7 2470 0.58 D8 2310 0.51 D9 2005 0.48 D10 1850 0.41

Experiments were performed on the response time of the recommender system using the similar user index method according to the present invention and the general collaborative filtering method using the above-described data set.

Referring to FIG. 8, as a result of the experiment on the reaction time for the two recommendation systems, the similar user index method according to the present invention was executed without increasing the reaction time relatively as the data size increased.

Referring to FIG. 9, the clustering accuracy by K-means and the clustering accuracy according to the present invention are compared. The clustering method according to the present invention can be confirmed to be more accurate than clustering by K-means because the accuracy Q value is relatively less even if the data size is large.

The user authentication service method according to various embodiments of the present invention can be implemented in the form of program instructions that can be executed through various means such as servers.

Further, a program and an application for executing the user authentication service method according to the present invention may be installed in a computer means and recorded in a computer-readable medium. The computer readable medium may include program instructions, data files, data structures, and the like, alone or in combination.

In addition, program instructions recorded on a computer-readable medium may be those specially designed and constructed for the present invention or may be available to those skilled in the computer software arts. Examples of computer-readable media include magnetic media such as hard disks, floppy disks and magnetic tape; optical media such as CD-ROMs and DVDs; magnetic media such as floppy disks; Includes hardware devices specifically configured to store and execute program instructions such as magneto-optical media and ROM, RAM, flash memory, and the like. The above-mentioned medium may also be a transmission medium such as a light or metal wire, wave guide, etc., including a carrier wave for transmitting a signal designating a program command, a data structure and the like. Examples of program instructions include machine language code such as those produced by a compiler, as well as high-level language code that can be executed by a computer using an interpreter or the like. The hardware devices described above may be configured to operate as one or more software modules to perform the operations of the present invention, and vice versa.

The embodiments of the present invention have been described above. It will be understood by those skilled in the art that various changes in form and details may be made therein without departing from the spirit and scope of the invention as defined by the appended claims. Therefore, the disclosed embodiments should be considered in an illustrative rather than a restrictive sense. The scope of the present invention is defined by the appended claims rather than by the foregoing description, and all differences within the scope of equivalents thereof should be construed as being included in the present invention.

Claims (12)

delete delete delete In a similar user index method in a recommendation system,
Generating a social network in a recommender data set;
Selecting a seed using the generated social network;
Clustering the selected seeds;
The clustered cluster being readjusted; And
Wherein the cluster is indexed into a pseudo-user cluster,
Clustering about the selected seeds
Assigning an initial center value and a node to each cluster; And
Each of the nodes excluding the seeds is calculated by the following Eq. (2) based on Euclidian distance measurement and assigned to the cluster with the shortest distance,
&Quot; (2) "

Here, TN is the total number of items experienced by the user node or seed, and MN is the number of items that a pair of nodes have commonly experienced.
In a similar user index method in a recommendation system,
Generating a social network in a recommender data set;
Selecting a seed using the generated social network;
Clustering the selected seeds;
The clustered cluster being readjusted; And
Wherein the cluster is indexed into a pseudo-user cluster,
The step of the clustered cluster being readjusted
Calculating a first accuracy evaluation numerical value of each cluster;
Two arbitrary nodes belonging to different clusters are extracted and swapped into the different clusters, and a second accuracy evaluation numerical value of the different cluster is calculated; And
Wherein the first accuracy evaluation numerical value and the second accuracy evaluation numerical value are compared so that when the second accuracy evaluation numerical value is larger, the extracted nodes are swapped into the different clusters and the clusters are rearranged User index method.
In a similar user index method in a recommendation system,
Generating a social network in a recommender data set;
Selecting a seed using the generated social network;
Clustering the selected seeds;
The clustered cluster being readjusted; And
Wherein the cluster is indexed into a pseudo-user cluster,
Wherein the cluster is pseudo-user cluster indexed
A similarity degree between a given user ID and a center of each similar user cluster is calculated using a hash function, and a cluster ID is output according to the calculated similarity.
delete delete delete A computer program stored on a medium for performing pseudo-user indexing in a recommendation system,
Generating a social network in a recommender data set;
Selecting a seed using the generated social network;
Clustering the selected seeds;
The clustered cluster being readjusted; And
Wherein the cluster is indexed into a pseudo-user cluster,
Clustering about the selected seeds
Assigning an initial center value and a node to each cluster; And
Computing each of the nodes except the seed by the following Euclidean distance measure and assigning it to the cluster with the shortest distance.
&Quot; (3) "

Here, TN is the total number of items experienced by the user node or seed, and MN is the number of items that a pair of nodes have commonly experienced.
A computer program stored on a medium for performing pseudo-user indexing in a recommendation system,
Generating a social network in a recommender data set;
Selecting a seed using the generated social network;
Clustering the selected seeds;
The clustered cluster being readjusted; And
Wherein the cluster is indexed into a pseudo-user cluster,
The step of the clustered cluster being readjusted
Calculating a first accuracy evaluation numerical value of each cluster;
Two arbitrary nodes belonging to different clusters are extracted and swapped into the different clusters, and a second accuracy evaluation numerical value of the different cluster is calculated; And
Wherein the first accuracy evaluation numerical value and the second accuracy evaluation numerical value are compared so that when the second accuracy evaluation numerical value is larger, the extracted nodes are swapped into the different clusters and the clusters are rearranged A computer program stored on a medium for performing user indexing.
A computer program stored on a medium for performing pseudo-user indexing in a recommendation system,
Generating a social network in a recommender data set;
Selecting a seed using the generated social network;
Clustering the selected seeds;
The clustered cluster being readjusted; And
Wherein the cluster is indexed into a pseudo-user cluster,
Wherein the cluster is pseudo-user cluster indexed
Wherein the hash function is used to calculate a similarity between a given user ID and a center of each similar user cluster, and a cluster ID is output according to the calculated similarity.

Images