KR101985603B1 - Recommendation method based on tripartite graph - Google Patents

Abstract

The present invention relates to a recommendation method based on a tripartite graph and, more specifically, to a recommendation method based on a tripartite graph improving the accuracy of recommendation by manipulating weights of nodes and trunk lines in a tripartite graph. The present invention may utilize hidden information not easily revealed by including the structural characteristics obtained by constructing a tripartite graph while using a conventional tripartite graph in a recommendation process.

Description

{RECOMMENDATION METHOD BASED ON TRIPARTITE GRAPH}

The present invention relates to a three-minute graph-based recommendation method, and more particularly to a three-group graph based on a three-group data set consisting of a first group, a second group and a third group having three sets of x, y, I) For each of the first groups, a first weighting matrix is obtained by expressing each constituent element as a node, which is an association weight with a second group of constituent elements, as an xy matrix, Normalizing; Ii) dividing the information of the first group first node by p and distributing p to the second group of nodes using the ratio of the first node to the second group of nodes connected to the first node; Iii) obtaining a second weighting matrix represented by a yxz matrix having weights of edges, which are associations with nodes of a third group connected to each node of the second group, and normalizing the second weighting matrix by a column unit; Iv) distributing and summing p distributed in step ii) to the third group node using the normalized second weight matrix obtained in iii); V) distributing p to the first group of nodes by performing the processes of iii) to iv) for the first group in the third group, dividing the value distributed to each node by p, Obtaining a distribution ratio; Vi) sequentially performing the steps of ii) to v) for the remaining nodes of the first group to obtain a final distribution ratio of each node, expressing the final distribution ratios by xxx matrix and obtaining a first power matrix; Ⅶ) obtaining a modified first modified weighting matrix by redistributing the weight of each node according to the distribution ratio expressed in the first power matrix with respect to the first weighting matrix; (I) performing the steps (i) to (iii) for the second group and the third group to obtain respective modified weighting matrices.

In recent years, as the mobile environment evolves so rapidly that it can not be compared with the previous years, in the advertising market, there are also display posts, in-app ads, text ads, In recent years, various types of advertisements such as location-based advertisement, augmented reality advertisement, and compensation advertisement have appeared, and interest in mobile advertisement is increasing. However, current mobile advertising has various problems such as below profitability due to low advertising price, doubts about advertiser's mobile advertising effectiveness, and consumer's rejection of mobile advertisement.

First, the present mobile advertisement is displayed simply / indiscriminately, rather than providing information of the product when the consumer desires. For example, if you read news articles via mobile, you can see that ads that are not of interest at all appear indiscreet at the top, bottom or even in the middle of the article. Therefore, these ads are not only effective in advertising itself, but also cause consumers to feel rejection.

Second, from the indiscreet advertisement, [1] suggested personalized recommendation advertisement method 'Retargeting' when browsing website through browsing history of user's web site, Of course, is being applied to mobile. (Need to refer to existing patent or non-patent literature), but it is merely a re-advertisement or related advertisement for items that the user was interested in in the past. In other words, we can not be sure that the user is interested in the product at this point. In addition, it is impossible for the consumer to show the relevant advertisement / information at the point of contact with the actual interest goods.

Third, in the case of the existing digital signage, in order to transmit the related product or the supplementary information to the user like the digital signage contents, the relevant information must be directly exposed in the digital signage contents. However, this method may interfere with the experience of digital signage content consumption, which may cause a negative feeling to consumers, and may also deteriorate the quality of digital signage content itself. However, in order to receive the contents desired by the user, the user has to know the desired information and the contents in advance, directly input the related search keyword into the screen and the character input device of the smartphone, and then select desired contents among the various contents , Most smartphone users find it difficult to input accurate search keywords because they feel annoyed.

Accordingly, a method of providing a personalized advertisement by reflecting the search condition of each user has been developed. The recommendation system is generally divided into three types: a collaborative filtering recommendation system, a content-based filtering recommendation system, and a hybrid recommendation system. The above three recommendation systems analyze past patterns and compare them to the current patterns. If a recommendation system is implemented through this method, it is recommended to make a recommendation through a simple comparison between elements of each user or item. Such a simple comparison does not make good use of the deep relationship between elements. In a typical recommendation system, in order to reflect deeply hidden pattern information in the recommendation, it is necessary to use a large amount of data sets that have been formalized. Korean Patent Laid-Open Publication No. 2004-104648 discloses a personalized content recommendation apparatus comprising: a situation information analyzing unit for analyzing situation information and extracting a recommended content group according to a current situation from among recommended target contents; A profile analyzer for analyzing user profile information and extracting a recommended content group according to a user profile from among the contents extracted by the situation information analyzer; A use history analyzing unit for analyzing usage history information and extracting a recommended content group based on usage history information among the contents extracted by the profile analyzing unit; A recommendation analyzing unit for analyzing the content information recommended by the human network on the social network and extracting content from the content extracted by the profile analyzing unit in accordance with recommendation information of the social network network; And a popularity analyzing unit for extracting popular contents according to the number of views among the recommended contents, wherein the usage history analyzing unit, the human network recommendation analyzing unit, and the personalized contents extracted from the popularity analyzing unit are sorted by date, , A recommendation order, and a popularity order. Also, Korean Patent Laid-Open Publication No. 8-88635 discloses a method for creating, managing and realizing a bucket list (a list of things that an individual wants to do) via a personal terminal 100 to a user on a social network service (SNS) And providing a platform based on a bucket list to receive advertisements to receive advertisements and to generate revenue by securing users who utilize the platform.

Recently, a graph based recommendation system, especially a three minute graph based recommendation system, has been proposed. Chen, B. [2] proposed a three-minute graph-based system for recommending moving image data, and Shams, B. [3] We propose a graph - based recommendation system. Jindal, H. [4] proposed a recommendation system that combines graph information and cluster result information. Korean Patent Laid-Open Publication No. 10-2017-0011963 discloses an item recommendation method comprising the steps of: collecting history information composed of a plurality of items generated by sample users; Pre-processing the history information using a regular rule for the items; Generating a collaborative filtering based graph from the preprocessed history information; Determining a content-based similarity of the history information; Modeling the graph by applying the content-based similarity to the graph; Determining a preference for each of a plurality of items for a target user from the modeled graph; And recommending announcement information for the target user based on the preference.

However, the above-mentioned researches do not use the information obtained through the graph structure, but merely connect the information that can not be connected with each other like the numerical data and the binary data through the graph structure.

Korean Patent Publication No. 10-2017-0011963

1.Shang, M., Fu, Y., Chen, D., "Personal Recommendation Using Weighted BiPartite Graph Projection", Apperceiving Computing and Intelligence Analysis, 2008. ICACIA 2008. International Conference on, Dec. 2008, pp. 198-202. 2. Chen, B., Wang, J., Huang, Q., Mei, T., "Personalized Video Recommendation Through Tripartite Graph Propagation", Proceedings of ACM Multimedia, 3. Shams, B., Haratizadeh, S., "Graph-based collaborative ranking", Expert Systems with Applications, vol. 67, Jan. 2017, pp. 59-70. 4. Jindal, H., Anjali, "Graph based Recommendation System in Social Networks", International Journal of Computer Applications, vol. 133, Mar. 2015, pp. 36-40. 5. Yifan, Hu., Yehuda, K., Chris, V., " Collaborative Filtering for Implicit Feedback Datasets ", Data Mining, 2008. ICDM '08. Eighth IEEE International Conference, Dec. 2008.

Therefore, the technical problem to be solved by the present invention is to provide a three-minute graph-based recommendation method which can utilize hidden information that is not easily revealed by including a structural characteristic obtained by constructing a three-minute graph using a conventional three- .

In order to accomplish the above object, the present invention provides a method for generating a three-group data set comprising three groups of data sets consisting of a first group, a second group and a third group having three sets of x, y, and z components, Expressing each element as a node by an xy matrix as a weight which is an association with an element of the second group to obtain a first weighting matrix and normalizing it by column; Ii) dividing the information of the first group first node by p and distributing p to the second group of nodes using the ratio of the first node to the second group of nodes connected to the first node; Iii) obtaining a second weighting matrix represented by a yxz matrix having weights of edges, which are associations with nodes of a third group connected to each node of the second group, and normalizing the second weighting matrix by a column unit; Iv) distributing and summing p distributed in step ii) to the third group node using the normalized second weight matrix obtained in iii); V) distributing p to the first group of nodes by performing the processes of iii) to iv) for the first group in the third group, dividing the value distributed to each node by p, Obtaining a distribution ratio; Vi) sequentially performing the steps of ii) to v) for the remaining nodes of the first group to obtain a final distribution ratio of each node, expressing the final distribution ratios by xxx matrix and obtaining a first power matrix; Ⅶ) obtaining a modified first modified weighting matrix by redistributing the weight of each node according to the distribution ratio expressed in the first power matrix with respect to the first weighting matrix; (I) performing the steps (i) to (iii) for the second group and the third group to obtain respective modified weighting matrices.

The present invention can utilize hidden information that is not easily revealed by including the structural characteristic obtained by constructing the three-minute graph using the conventional three-minute graph in the recommendation process.

1 is a schematic diagram for explaining an example of a recommendation method based on a three-minute graph according to the present invention;
FIG. 2 is a graph comparing the operation time of the recommendation method based on the three-minute graph according to the present invention and the operation method of the conventional three-minute graph
Figure 3 compares nDCG results of the recommendation algorithm based on the existing recommendation algorithm and the three-minute graph of the present invention

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS The present invention will now be described in detail with reference to the drawings attached hereto.

The recommendation method based on the three-minute graph of the present invention is based on the premise that a three-minute graph structure is constructed, and the weight changing process is extended by extending the previous research idea of Non-Patent Document 1. Since there is a three-dimensional graph structure, there are a total of three pieces of trunk information (weight value), and the recommended system is implemented through the process of manipulating the trunk information. At this time, for convenience of the operation process, the weight of the three-dimensional graph structure is used as a matrix.

In the present specification, 'graph' refers to a structure consisting of a node, which is a component constituting a set bound to one category, and a weight, which is an edge, which is a connection between the vertices. Means the vertex and trunk structure between the three sets of vertices constituting the vertices in the graph structure and having no weights in which the sets in the same set are the trunks. Also, 'power' is a ratio of how much information of one vertex is distributed to other vertices.

The ternary graph based recommendation method of the present invention is based on three groups of data sets consisting of a first group, a second group and a third group, which are three sets having x, y, and z components: i) Expressing each element as a node as a weight which is an association between the element of the second group and an xy matrix and obtaining a first weighting matrix and normalizing it by column; Ii) dividing the information of the first group first node by p and distributing p to the second group of nodes using the ratio of the first node to the second group of nodes connected to the first node; Iii) obtaining a second weighting matrix represented by a yxz matrix having weights of edges, which are associations with nodes of a third group connected to each node of the second group, and normalizing the second weighting matrix by a column unit; Iv) distributing and summing p distributed in step ii) to the third group node using the normalized second weight matrix obtained in iii); V) distributing p to the first group of nodes by performing the processes of iii) to iv) for the first group in the third group, dividing the value distributed to each node by p, Obtaining a distribution ratio; Vi) sequentially performing the steps of ii) to v) for the remaining nodes of the first group to obtain a final distribution ratio of each node, expressing the final distribution ratios by xxx matrix and obtaining a first power matrix; Ⅶ) obtaining a modified first modified weighting matrix by redistributing the weight of each node according to the distribution ratio expressed in the first power matrix with respect to the first weighting matrix; (I) performing the steps (i) to (iii) for the second group and the third group to obtain respective modified weighting matrices.

The ternary graph based recommendation method of the present invention is characterized in that it consists of three groups of data sets consisting of a first group, a second group and a third group which are three sets each having x, y and z components: i) And each element is represented as a node by an xy matrix as an weight which is an association between the element of the second group and an element of the second group to obtain a first weighting matrix and normalizing the first weighting matrix by a column unit. The process of manipulating the weights will be explained with a simple example. Suppose, for example, three groups with A = {a, b}, B = {1,2,3}, C = {x, y}. 1 is a schematic diagram for explaining an example of a recommendation method based on a three-minute graph according to the present invention. As shown in FIG. 1, first, the trunks between the constituent elements are represented by a matrix for the three sets A, B, and C. When the weights of the set A nodes and the set B nodes are given as Weight in Table 1, the weights of the nodes connected to each other are normalized. Consider node a in set A only. Since weights of node 1 and node 2 of set B connected to node a are 3 and 1, respectively, they become 0.75 and 0.25, respectively.

(Ii) dividing the information of the first group first node by p and distributing p to the second group of nodes by using the ratio of the distribution to the other nodes of the second group connected to the first node; Iii) obtaining a second weighting matrix represented by a yxz matrix having weights of edges, which are associations with nodes of a third group connected to each node of the second group, and normalizing the second weighting matrix by a column unit; Iv) distributing and summing p distributed in step ii) to the third group node using the normalized second weight matrix obtained in iii); V) distributing p to the first group of nodes by performing the processes of iii) to iv) for the first group in the third group, dividing the value distributed to each node by p, The distribution ratio is obtained. In the above example, the power of node a is 1, and this power is distributed to the set B nodes. When this power is decomposed into the normalized weight ratio obtained earlier, the power of 0.75 is distributed to node 1 and the power of 0.25 is distributed to node 2. [ Next, we normalize the weights of nodes 1 and 2 of the connected set C and then distribute the previously distributed power by the weighting ratio normalized to the next set C, and the power of the nodes a and b is 0.19 and 0.81, respectively, do. When this process is set again in the set B in the set C, the power of the node a is 0.36 and the power of the node b is 0.64.

Next, vi) sequentially performing the steps ii) to v) for the remaining nodes of the first group to obtain a final distribution ratio of each node, expressing it as an xxx matrix and obtaining a first power matrix; And (e) redistributing weights of the respective nodes according to the distribution ratios expressed in the first power matrix with respect to the first weight matrix, thereby obtaining a corrected first modified weight matrix. That is, in the above example, the power is obtained through the same process as the node a, and then the weight is redistributed as the ratio of the power obtained as the final result.

The recommendation method of the present invention includes the steps of (i) through (e)) for each of the second group and the third group to obtain respective modified weighting matrices. As described above, it is defined as 1 step that the above process is performed for all the nodes of the set, and the step is performed in the order of the node sets A, B, C. Finally, the weight information obtained through the above process is passed through the whole three-minute graph structure, so the information that can be obtained only by using the graph structure is reflected in the weight. Therefore, if a recommendation is made for a specific menu, a node having a higher weight is higher in a recommendation order than a node having a higher weight.

In addition, the present invention may further comprise the step of (i) performing the above steps (i) to (i) one more time. In the case of performing steps (i) to (i) above, the present invention is described as a one-cycle process. When the cycle continues, the matrix information changes, and the rate of change of the matrix converges to a certain level. The rate of change for the matrix A is defined as the numerical change to the updated matrix A 'after 1 cycle of the matrix A.

Table 2 summarizes the rate of change of each matrix per cycle. As shown in Table 2, the rate of change after 2 cycles converges to about 10 ^-7 . Since this value is very small even if it is considered simple, the weight information ranking does not change because the cycle proceeds further. Therefore, Each matrix can be said to have converged in two cycles.

The weight manipulation process described above was conducted with a three-minute graph consisting of a menu, a context word, and a set of restaurants. In this case, because of the nature of the matrix operation, a large amount of computation occurs due to the overlapping of the iterative queries, and the execution time of one cycle of the initial implementation model is about 12 hours. In order to increase the computation speed, we changed the code so that most operations are processed by the GPU using a GPU specialized for matrix operation. However, it is problematic that the amount of memory space required to implement the matrix transformation process is the same as the square of the largest node size in the graph. In the initial experiment graph, there are 50,000 restaurant nodes, and the required memory size is 50,000 X50,000 X 4 bytes = 10GB. However, since the memory size of the GPU used in the experiment is 8GB, . As a result, we were able to reduce the execution time of one cycle up to 2 hours, but we found that the bottleneck caused by moving information between CPU and GPU occupies most of the execution time, while the time of pure operation is very short. In order to further improve the execution time of one cycle, memory usage is reduced to less than 1 GB by directly modifying the matrix obtained after one operation of the node by modifying the procedure of weighting operation described in Section 4.1. Thus, we could change all of the operations to be handled by the GPU and reduce the execution time of one cycle to 2 minutes (Figure 2).

In order to construct a three-minute graph that is necessary in the present invention, a menu, a restaurant, and a condition set are used as three node sets of a three-dimensional graph as node information. The definition of a conditional language is a kind of context for recommendation. For example, "A family is going to have dinner for Christmas." , All meaningful morphemes except for menus and restaurants, such as family, Christmas, and evening, correspond to situation words. In order to extract node information, review of blog data was analyzed by morpheme. First, 86 types of menus were analyzed by analyzing the morpheme for the menu, and then 215 morphemes judged as meaningful words were analyzed by analyzing the remaining morphemes in order to extract the condition words. Information about the restaurant was used by the tag information set by the blog author. After gathering all the tag information and obtaining about 50,000 pieces of information, in order to remove the inaccurate information, only the matching restaurant information was compared with the actual restaurant information and 2069 restaurants information could be obtained. The weight information (trunk line) was constructed using the number of simultaneous occurrences of each morpheme (node).

Using the three-minute graph completed for the previous experiment, the evaluation was conducted through recommendation data up to 300 cycles. In this paper, we use Neighborhood based CF scheme proposed by Yifan and Hu [5] as a comparative technique. This technique is selected as a comparison technique because it is the most common collaborative filtering method of recommendation system. For the quantitative evaluation, the value for the evaluation was the Normalizing Discounted Cumulative Gain (nDCG).

이때 relevance 함수는

The relevance function

Respectively. Both the correct answer and the recommendation information use the Top 5 ranking information, and the evaluation of each of the three matrices of the three-minute graph was carried out. Each evaluation value was the average of recommendation results for three different inputs. In the case of the menu-to-language relation, the recommendation technique of [5] was used. In the evaluation of the menu-context relation, This did not work at all. This is because the menu-context relation information (86 x 215) used in the verification experiment of the present invention is small in finding the hidden information through the recommendation technique of [5], and the relation based on the information extracted from the irregular data called the social review data It is understood that the pattern captured in the recommendation technique proposed by the present invention can not be found. In this paper, by incorporating the structural information of the 3-minute graph into the calculation process by using only a small amount of irregular data, it is possible to find out deeply hidden pattern information which can not be obtained through the existing recommendation technique, It is experimentally shown that it can achieve better performance than the conventional technique.

The embodiments of the present invention described above should not be construed as limiting the technical idea of the present invention. The scope of protection of the present invention is limited only by the matters described in the claims, and those skilled in the art will be able to modify the technical idea of the present invention in various forms. Accordingly, such improvements and modifications will fall within the scope of the present invention as long as they are obvious to those skilled in the art.

Claims (2)

Executed on a programmable machine selectively activated or reconfigured by an electronic device or a computer program comprising at least a processor and a memory for storing instructions executed by the processor,
From the three group data set consisting of the first group, the second group and the third group, which are three sets having x, y, and z components,
I) For each of the first groups, a first weighting matrix is obtained by expressing each constituent element as a node, which is an association weight with a second group of constituent elements, as an xy matrix, Normalizing;
Ii) dividing the information of the first group first node by p and distributing p to the second group of nodes using the ratio of the first node to the second group of nodes connected to the first node;
Iii) obtaining a second weighting matrix represented by a yxz matrix having weights of edges, which are associations with nodes of a third group connected to each node of the second group, and normalizing the second weighting matrix by a column unit;
Iv) distributing and summing p distributed in step ii) to the third group node using the normalized second weight matrix obtained in iii);
V) distributing p to the first group of nodes by performing the processes of iii) to iv) for the first group in the third group, dividing the value distributed to each node by p, Obtaining a distribution ratio;
Vi) sequentially performing the steps of ii) to v) for the remaining nodes of the first group to obtain a final distribution ratio of each node, expressing the final distribution ratios by xxx matrix and obtaining a first power matrix;
Ⅶ) obtaining a modified first modified weighting matrix by redistributing the weight of each node according to the distribution ratio expressed in the first power matrix with respect to the first weighting matrix;
(I) performing the steps (i) to (iii) for the second group and the third group to obtain respective modified weighting matrices; The method according to claim 1,
Ⅸ) A three-minute graph-based recommendation method including the steps of performing steps i) through 1) one more time.

Images