KR20170060828A - User's Contents Access History based Recommendation Method - Google Patents

Abstract

The present invention relates to a hybrid recommendation system combining a user preference prediction technique and a clustering technique, and a method of developing in consideration of user's item access history, more specifically, Among them, we use item-based clustering using the preference value of users' content, so that users who enjoy similar contents at a similar time can enjoy similar contents as recently accessed contents, It is about how to recommend different things.

Description

A recommendation method based on a user's content access history {User's Contents Access History based Recommendation Method}

The present invention relates to a method for recommending unexpected discoveries and useful content to a user. In more detail, the content is grouped based on ratings or access counts of contents that the users have accessed. It is assumed that the contents in the same group are relatively similar to the contents in the other group. Then, after judging which group the contents belong to for the contents that the user has recently accessed, it is a method for recommending contents existing in other groups except the groups to the user.

Also, in order to recommend useful contents from users' perspective, the content is selected in a high preference order by using the preference prediction technique developed so far. Representative preference prediction techniques include item / user based collaborative filtering and alternating least squares with weighted-lambda-regularization (ALS-WR).

The preference prediction technique has been developed with a focus on recommending appropriate content to the user by using the preference tendency and content characteristics of the user's contents. For example, in a content-based recommendation method, a similar movie that a user has not yet seen is recommended by using a keyword or a genre of a movie that the user has enjoyed in the usual way.

The content - based recommendation method mainly borrows methods developed and utilized in the field of information retrieval. And user-based collaborative filtering is a way to identify other users with a similar tendency to a particular user, to recommend similar movies that a similar user has judged useful and a particular user has not yet seen. Similarly, item-based collaborative filtering is a preference prediction technique based on calculating similarities between items (movies).

The main purpose of the recommendation systems developed so far is to predict users' expected preferences as accurately as possible. For this purpose, an implicit preference such as an explicit preference such as a rating of a content or a visit frequency is utilized, and information such as a profile of a user and contents is utilized.

In recent years, however, since the user can easily browse the desired information on the Internet, most of the contents that the user enjoys in the normal life are grasped. Therefore, contents exposed by traditional recommendation system are not easy to attract appeal to users. As a result, the topic of 'Serendipity' has recently attracted attention.

Most systems use a method of recommending the rest of the content to users, except for the top few popular content among the content submitted by the general recommendation system, in order to provide the serendipity recommendation. However, this method excludes only popular content in the beginning, but does not consider the tendency of users to enjoy content recently. Take, for example, the kind of food. Users who enjoy fried rice will not be able to feel serendipity because they recommend cheese fried rice which is not popular among kinds of fried rice.

This user likes fried rice, so they have already grasped various kinds of fried rice and it is likely that the preference is fixed. On the other hand, if you recommend something other than fried rice, which is generally popular, it is likely that you will feel serendipity because you have a chance to eat something new in a long time.

As we have seen so far, the proposed serendipity recommendation does not consider what content users enjoy recently. In other words, long-tail recommendation was the main method within the user's preferred content type. On the other hand, this patent is not only a recommendation for a long-tail, but also a recommendation of serendipity even when recommending popular contents in consideration of recent content consumption tendency of users.

In general, the recommendation system predicting the user's preference for the content that the user does not touch, and exposes the content that the user prefers, depending on the rating base or access frequency of the user's content.

In addition, the method of collecting various personal information of users and predicting the contents that the user prefers and exposing them to the users is a recommended recommendation system so far.

However, the above method does not consider factors that may cause new interest to the users, but it is controversial because it infringes on privacy because it utilizes personal information.

The recommendation system of the present invention is a hybrid system that combines a general preference prediction technique and a clustering technique, and provides a method of providing unexpected enjoyment and usefulness to users at the same time and not infringing privacy of an individual at all. In addition, it relates to a method of developing a fully automated recommendation system as a non-supervised learning, rather than establishing a recommendation system as supervised learning during machine learning.

The method includes the steps of: storing preferences of contents accessed by users in a DB; calculating an expected preference of contents that each user does not access; clustering contents using preferences of contents of users; Determining which clusters the contents that have recently been accessed by the users belong to: determining whether each user belongs to or belongs to a group of contents recently accessed by the user with respect to the contents having high expected preferences; And exposing the output contents to the user on the basis of the processing procedure up to.

At this time, the structure of the DB for storing preferences is composed of (user identifier: content identifier: content preference or content access frequency: content access time), and DB is periodically or non-periodically updated according to the characteristics and policy of the recommendation system And also continuously stores the above information in the DB.

In addition, all the methods for estimating the likelihood of contents that are not accessed by each user can be applied such as collaborative filtering, content-based filtering, and hybridization strategies. It can be selected according to the characteristics of the recommendation system.

Also, a method of clustering contents by using the preferences of users' contents can also be selected as a method of optimal clustering according to preference data characteristics. In this case, the clustering of contents based on the preference is intended to grasp the characteristics of the contents using the collective intelligence of the users rather than selecting suitable contents for the users in consideration of the characteristics of the contents, , And it is considered that the contents that do not have different attributes.

In addition, the step of determining to which cluster the contents that have recently been accessed by the users belongs is to classify the contents that the users have recently shown interest in, and the recommendation of contents in the cluster to which the contents recently accessed by the user belongs, , While conversely, content recommendations in other clusters can make users feel unexpected enjoyment, i.e., serendipity.

Also, in the step of determining whether or not each content item belongs to a group of contents recently accessed by each user for a content having a high estimated preference level for each user, as described above, a similar tendency to a content item that the user has recently been interested in Is a step for deciding whether to recommend the content or the content having a newer characteristic. At this time, similarity of contents is not determined by using various tag information of content, but is a result of utilizing collective intelligence using contents preference of users.

Also, in the step of exposing the recommended contents to the user on the basis of the processing procedure so far, it is possible to transmit the recommended content information through communication between the recommendation engine server and the client, communication between the recommendation engine server and another application server.

A recommendation method according to a user's content access history has various effects as follows.

First, the recommendation method according to the present invention can recommend contents different depending on the type of the user in consideration of the recent contents access history of the user. Similar content is recommended for users who enjoy similar contents in the usual way, and recommendation is possible considering users' tendency to consume the contents because users who do not recommend different contents are recommended.

Second, the recommendation method according to the present invention clusters content according to the preferences of users for each item. Accordingly, the contents are grouped into a plurality of clusters, and the contents in the same cluster are regarded as contents having a similar nature, and the contents in different clusters can be regarded as contents having a similar nature.

Third, the recommendation method according to the present invention reflects the latest trends of contents enjoyed by users. Because individual interests can change at any time, personalized content recommendations are possible by reflecting the latest trends in user content.

The recommendation system of the present invention as described above has the effect of recommending contents having usefulness and serendipity elements to users, and not invading privacy of an individual at all.

Brief Description of the Drawings Fig. 1 is a diagram showing a database structure for building a recommendation system of the present invention; Fig.
Figure 2 shows an example of how the recommendation process resides in memory in the database.
FIG. 3 is a diagram showing an overall sequence in which the present invention is utilized.
FIG. 4 is a flowchart showing the main recommendation operation procedure of the present invention in detail.
FIG. 5 is a diagram showing a recommendation operation procedure of the present invention as pseudo-code.

The present invention relates to a content recommendation method for a user, and more particularly, to a content recommendation method that allows a general recommendation algorithm and clustering, content recommendation using an individual's access history of contents to induce a user's interest, and feeling of serendipity . The present invention also relates to a method for recommending useful and fresh contents from a user's point of view by utilizing only the preference or access frequency of the contents accessed by the user without using the personal information of the users and the internal information of the contents.

The database structure for the recommendation system of the present invention and the overall operation procedure and recommendation system of the present invention will be described in detail with reference to the accompanying drawings.

1 shows a database structure for building a recommendation system of the present invention. The database consists of a total of four tuples, each consisting of a user identifier, a content identifier, a preference for the content the user has accessed, and the time the user accessed the content. The recommendation system according to the present invention does not need detailed information about the user and the contents, and only the time when the user accesses the specific contents and the preference for the contents are stored in the database.

At this time, the preference of the content can be an explicit preference such as the actual rating, and an implicit preference such as the number of times the user approached the content can also be used. For example, if a user accesses 3 or 4 times of content, respectively, the corresponding preference value can be calculated by storing 3 and 4 in the database, respectively. The time of accessing the content can be stored in any form as long as the value is comparable with the time when the user executes the application.

FIG. 2 shows an example of allocating the database illustrated in FIG. 1 on the recommendation system process memory. FIG. 2 shows the assumption that the total number of users stored in the database and the number of contents are respectively assumed. And the user's preference value for the content.

3 is a diagram showing how the recommendation system of the present invention can be utilized. When the user executes the application, the user identifier is transmitted to the corresponding application server. At this time, the application server is dependent on the user application.

For example, when a user runs the benple G application, the user identifier is sent to the benple G recommendation system server. And the user application is applicable to everything where the recommendation system is applicable. When the application server receives the user identifier, the application server calculates appropriate content to be recommended to the user and responds to the user application with data such as the corresponding content identifier or actual content. A more detailed process of running the referral program is illustrated in FIG.

FIG. 4 is a diagram showing how an application server calculates a recommended content for a user. First, the application server builds and maintains a database as illustrated in Figure 1. The data stored in the database can be stored in the memory of the application server, stored in the form of a file, or read by the application server process whenever a request comes in.

Then, the application server executes content-based clustering according to the data stored in the database. Content-based clustering is performed based on user's preference for contents, and clustering method can be appropriately selected according to recommended environment. It may also be possible to modify the information in the database for better clustering to implement content-based clustering.

After clustering, contents in the same cluster are regarded as being related to each other, and conversely, contents in different clusters are compared with contents in the same cluster. Clustering results can either reside in the server's memory or be saved as a file so that the file can be read into memory whenever needed. At this time, the file may be stored in a physical storage device such as a hard disk on the server, or may be distributed to a plurality of servers according to the file size.

Similar to the clustering process, we use the information in the database to execute a preference prediction function. The preference prediction function can be selected according to the recommendation situation, such as user-based collaborative filtering and item-based collaborative filtering. As in the case of clustering, the results calculated by the preference prediction function can be stored in the server's memory or stored in a file format. Thereafter, upon receiving the user identifier from the user application, the application server calculates the recommended content corresponding to the user identifier. The number of recommended contents is a predefined value. In an application showing a lot of recommendation results according to an application policy, a large number of recommended contents are set, and a small number of recommended contents are calculated in a non-recommended application.

After calculating the contents to be recommended, the user identifier is used to determine to which community the contents that the user has recently accessed belong to. In this case, "recent" can be viewed in terms of time or in terms of the number of recently accessed contents.

For example, content that has been accessed within a week can be defined as the content that the user has recently accessed, or the recently accessed five content can be defined as the content that the user has recently accessed. As before, the definition of this can be determined by the recommendation system and recommendation environment. Finally, the intersection of the contents in the community to which the user recently accessed the contents belongs and the contents calculated by the recommendation program is determined as the recommended contents for the user, or the non-overlapping contents are determined as the recommended contents for the user.

In the former case, similar contents are recommended in consideration of the user's latest preference tendency. In the latter case, it is recommended to users that content that is different from the content the user has recently accessed, but which is deemed useful by other users. By doing this, users who want content that they are interested in in the future can recommend similar content, and users who want a new one can recommend relatively fresher content.

This can be determined either directly by the user in the same menu as the settings in the application, or in the application server.

FIG. 5 shows a pseudocode showing in greater detail the contents depicted in FIG.

Claims (4)

A database construction method as shown in FIG. In order to consider the user's recent item access history, a method for constructing a database for considering recent access history, such as a content access time, a content access order, and a recent access time, is within the claims. As in Figure 2, item-based clustering is within the scope of the claims. As shown in FIG. 4, a method for recommending contents differently for each type of user. For users who enjoy similar contents in particular, similar items are recommended based on recent access history. In the opposite case, a method of recommending contents that have been recently accessed and contents existing in another cluster. As shown in FIG. 5, all the recommendation methods based on the preference prediction technique and the item-based clustering are included in the claims.

Images