KR20110031760A - Hybrid recommendation method and system for large scale data - Google Patents

Abstract

PURPOSE: A system and a method of hybrid recommending for high capacity data are provided to shorten a taking time for providing the recommendation result and improve accuracy of the recommendation result by providing a recommendation result by suing reduced data. CONSTITUTION: Similar neighbors having a similar purchase pattern with a target buyer and different neighbors having a different purchase pattern are selected(S110). By referring purchase history of the different neighbors and purchase history of the target buyer, a boundary function is obtained(S120). By referring the boundary function and a purchase history of the similar neighbors, a recommended item is selected(S130). The similar neighbors and different neighbors are selected by a calculation of a Pearson product-moment correlation coefficient.

Description

HYBRID RECOMMENDATION METHOD AND SYSTEM FOR LARGE SCALE DATA}

The present invention relates to a product recommendation method, and more particularly, to a hybrid recommendation method and system for a large amount of data.

With the development of information and communication technology, the amount of information is exploding. Accordingly, research on automated tools that can efficiently search large amounts of information is increasing. A recommendation system has been proposed to address these challenges.

The recommendation system recommends an item that satisfies the buyer's request from among a plurality of items (eg, a product or information) in response to the buyer's request. The methods used in the recommendation system include content-based filtering, collaborative filtering, and hybrid methods.

Content-based filtering refers to a product having an attribute value requested by the buyer by referring to the attribute value of the product. Content-based filtering has evolved from information searching and information filtering. Collaborative filtering is a method of finding and recommending products that a buyer may prefer by referring to purchase history information of similar customers. The hybrid method is a method of recommending a product by referring to the results of content-based filtering and collaborative filtering. The product recommendation service recommends a suitable product to the buyer using one of these methods.

In particular, the product recommendation service may be usefully used in electronic commerce. For example, an online bookstore like Amazon (Amazon.com) uses a recommendation system to analyze buyers' buying patterns and recommend books that the buyer might prefer. In addition, the recommendation system is being applied to various fields such as movie, news, and shopping mall product recommendation, and is gradually expanding the field.

However, the recommendation system is required to process a large amount of transaction data and item attribute data in order to provide a recommendation result. Transaction data is data about who bought what items. Item attribute data is data concerning the characteristics of an item. Transaction data accumulates as the transaction is performed.

As the amount of transaction data increases, the time taken to provide the recommendation results and the accuracy of the recommendation results decrease. If the recommendation results of the recommendation system are provided too late or the recommendation results do not accurately reflect the buyer's needs, the recommendation system will lose its utility. Therefore, it is required to improve the presentation time of the recommendation result and the accuracy of the recommendation result.

It is an object of the present invention to provide a recommendation method that can provide accurate recommendation results within a short time.

Hybrid recommendation method according to the present invention comprises the steps of selecting among the neighbors (Nearest neighbors) having a similar purchase pattern and a different purchase pattern (Farthest neighbors) among the plurality of buyers; Obtaining a boundary function by referring to the purchase history of the target buyer and the purchase history of the different neighbors; And selecting a recommendation item by referring to the purchase history of the similar neighbors and the boundary function.

In an embodiment, the similar neighbors and the different neighbors are selected by calculation of a Pearson product-moment correlation coefficient. Buyers with high Pearson product moment correlation coefficient values are preferentially chosen as the like neighbors. Buyers with low Pearson product moment correlation coefficient values are preferentially chosen as neighbors.

In another embodiment, the boundary function is configured to distinguish between an item purchased by the target buyer and an item purchased by the different neighbors. The recommended item is selected with reference to a collaboration filtering score and a distance to boundary (DTB) score. The collaboration filtering score is determined according to a correlation between the purchase pattern of the target buyer and the purchase pattern of the similar neighbors. The DTB score is determined according to the degree to which the purchased items of the similar neighbors are separated from the boundary function.

In another embodiment, the recommendation item is selected with reference to a sum of a value obtained by multiplying the collaboration filtering score and the DTB score by a weight adjustment factor. The weight adjustment coefficient may be varied for selective performance of content based filtering or collaborative filtering. When the weight adjustment coefficient has a value of 0, the weighting coefficient is performed by the collaborative filtering method. When the weight adjustment coefficient has an infinite value, the content adjustment filtering is performed.

Hybrid recommendation method according to the present invention comprises the steps of selecting among the neighbors (Nearest neighbors) having a similar behavior pattern and a different behavior pattern (farthest neighbors) among the plurality of users; Obtaining a boundary function by referring to the task history of the target user and the task history of the different neighbors; And selecting a recommendation item by referring to the work history of the similar neighbors and the boundary function.

The hybrid recommendation system according to the present invention includes a profile database including purchase history of buyers; And a recommendation system, wherein the recommendation system refers to the profile database to identify nearest neighbors having similar buying patterns as target buyers and different neighbors having different purchasing patterns from among a plurality of buyers. The user selects a boundary function by referring to the purchase history of the target buyer and the purchase history of the different neighbors, and selects a recommended item by referring to the purchase history of the similar neighbors and the boundary function.

The hybrid recommendation method according to the present invention provides a recommendation result using the reduced data. As a result, according to the present invention, the time taken to provide the recommendation result can be shortened and the accuracy of the recommendation result can be improved.

It is to be understood that both the foregoing general description and the following detailed description are exemplary, and that additional explanations of the claimed invention are provided. Reference numerals are shown in detail in preferred embodiments of the invention, examples of which are shown in the reference figures. In any case, like reference numerals are used in the description and the drawings to refer to the same or like parts.

In the following, a hybrid recommendation method is used as an example for explaining the features and functions of the present invention. However, one of ordinary skill in the art will readily appreciate the other advantages and performances of the present invention in accordance with the teachings herein, and the present invention may also be implemented or applied through other embodiments. In addition, the detailed description may be modified or changed according to aspects and applications without departing from the scope, technical spirit and other objects of the present invention.

One recommendation method is content-based filtering (CBF) and colaborative filtering (CF).

1 shows an item attribute table used in content-based filtering. Content-based filtering recommends items with attributes that match the buyer's purchasing pattern. Each item may be divided into a plurality of attributes.

Referring to FIG. 1, each of the plurality of items I ₁ to I _M has a plurality of attributes f ₁ to f _L. Each property has an attribute value between 0 and 1. For example, an attribute value of 0 means that the attribute has almost no characteristic, and an attribute value of 1 means that the attribute is completely possessed. In addition, if the attribute value is 0.5, it means that the characteristic of the attribute is about half.

Content-based filtering refers to the item attribute table to recommend items to the buyer that have attributes similar to those of the item the buyer requires. For example, suppose a buyer wants to download a wallpaper for a mobile terminal on a web page. Each background screen may be divided into a plurality of attributes. For example, the attributes may include color, saturation, brightness, and themes of the wallpaper. If the buyer selects a wallpaper about a particular celebrity, the content-based filtering system will recommend to the buyer other wallpapers based on the celebrity theme.

Content-based filtering has proven to be effective in finding items that meet the needs of buyers using techniques such as Boolean model, vector space model, probabilistic model, artificial neural network model, and fuzzy set model.

However, in content-based filtering, the item you are looking for must be in a machine-classifiable form (eg text), and if the word in the purchase record does not exactly match the word that appears in the item's content, There is a term mismatch problem that cannot be selected.

Collaborative filtering, on the other hand, predicts the preference of the target buyer by referring to taste information obtained from the plurality of buyers. The collaborative filtering system generates recommendation results using purchase information of buyers having a taste similar to the target buyer. The basic assumption of collaborative filtering is that buyers' past trends will remain the same in the future.

2 shows a transaction data table used in collaborative filtering. Referring to FIG. 2, a purchase variable y _ji indicating whether a plurality of items I ₁ to I _M has been purchased by a plurality of buyers c ₁ to c _N is displayed in the form of a table.

The purchase variable y _ji has a value of zero or one. 0 indicates that it has not been purchased and 1 indicates that it has been purchased. For example, if customer c ₁ has purchased item I ₁ , y ₁₁ has a value of 1. On the other hand, if the customer c ₁ has never purchased the item I ₂ , y ₂₁ has a value of zero.

The first thing to do in collaborative filtering is to find buyers with a taste similar to the target buyer. At this point, similarity weights may be obtained for all buyers to generate a recommendation result, but this is not preferable in terms of speed and accuracy. This is because it takes considerable time to analyze the purchase patterns of all buyers and generate the recommendation results. Therefore, it is desirable to generate recommendation results using buyers with similar preferences.

On the other hand, when generating a recommendation result considering only too few neighbors with similar preferences, there is a case that the lack of data cannot find an item similar to the item requested by the buyer. This is because an item similar to the item requested by the buyer may not be included in the neighbor's purchase record. Therefore, it is important to determine the appropriate number of neighbors so that the recommender system can make a quick and accurate prediction.

However, collaborative filtering has three disadvantages. First, an item that has never been purchased has a First-Rater Problem that is excluded from the recommendation from the beginning. Second, because most buyers only have a purchase record for some of all items, the purchase record shows a sparse problem. Third, the computation time of the nearest neighbor algorithm, which is mainly used in collaborative filtering, is exponentially proportional to the number of buyers and items. Therefore, in a trading environment where the number of buyers and items is large, this calculation time can be fatal. Finally, an approach that combines the advantages of content-based and collaborative filtering is needed to improve the quality of recommendations.

3 is a flowchart illustrating a hybrid recommendation method according to the present invention. Referring to FIG. 3, the hybrid recommendation method according to the present invention includes the steps of S110 to S130.

In step S110, the closest neighbors and the best neighbors are selected with reference to the transaction data. The closest neighbors are a collection of buyers with the most similar buying pattern to the target buyer. The best neighbors are a collection of buyers who show the different buying patterns from the target buyer. The size of the closest neighbors and the best neighbors may be adjusted as appropriate. For example, it is possible to set the number of closest neighbors and the best neighbors to 20 each out of a total of 100 buyers.

In this embodiment, a correlation coefficient is used as a method for finding the nearest neighbors and the best neighbors. The correlation coefficient is a measure of the strength of the association between two variables. A commonly used correlation coefficient is the Pearson product-moment correlation coefficient. In the Pearson product moment correlation coefficient, the strength of the association between two buyers C _T and C _i can be obtained by the following equation (1).

Where A is the number of items purchased by both buyers C _T , C _i , and B is an item purchased by first buyer C _T but not purchased by second buyer C _i . Is the number of items, C is the number of items purchased by the second buyer C _i but not purchased by the first buyer C _T , and D is the purchase by both buyers C _T , C _i The number of items that are not.

The more similar the purchase pattern among buyers, the larger the result value (sim (C _T , C _i )). Conversely, if buyers have different purchasing patterns, the resulting value will be small. As described above, the calculation of the similarity is simplified by using the Pearson product moment correlation coefficient to obtain the similarity of the buyers. This is because only the number of items is considered in the calculation using the Pearson product moment correlation coefficient. Thus, the time required for calculating the similarity is reduced. As time spent on similarity calculations decreases, time spent on providing recommendations results also decreases.

However, the scope of the present invention is not limited to this. In the present embodiment, although the Pearson product moment correlation coefficient is used, it will be apparent to those skilled in the art that any correlation coefficient may be used in the present invention.

In step S120, a boundary function is obtained. The boundary function determines a boundary that distinguishes an item that has already been purchased by the target buyer and an item that is already purchased by the neighbors of the best. That is, the boundary function distinguishes between items that are likely to be purchased by the target buyer and items that are less likely to be purchased.

To obtain the boundary function, the purchase history of the target buyer and the purchase history of the neighbors of the best neighbor are referenced. To this end, the purchase history of the target buyer and the purchase history of the best neighbors are first collected. At this time, the items already purchased by the target buyer among the purchase details of the neighbors of the best neighbor are excluded. The best neighbors are found using the correlation coefficient in step S110 described above.

4 is a graph for explaining a method of obtaining a boundary function. Referring to FIG. 4, the horizontal axis represents the first attribute (Feature 1) of the item, and the vertical axis represents the second attribute (Feature 2) of the item. For example, for a background screen of a mobile device, the first attribute may be color and the second attribute may be saturation. Although only two attributes are shown for convenience of description, as shown in FIG. 1, an item may include arbitrary attributes.

The squares in the graph represent items already purchased by the target buyer, and the circles represent items already purchased by the neighbors. f (x) is a boundary function. The boundary function distinguishes an area B containing items purchased by the target buyer from an area A containing items purchased by the neighbors.

The best neighbors show a different buying pattern than the target buyer. Thus, it is possible to obtain an accurate boundary function by only comparing the purchase history of the target buyer with the purchase history of the neighbors that are the best. This is because the purchase history of the target buyer and the purchase history of the neighbors can be easily separated. If a boundary function is obtained by referring to the purchase history of all buyers, this will inevitably cause the recommendation system to slow down.

In addition, items that are less likely to be purchased by the target buyer can be distinguished more precisely because only the purchase records of the neighbors with the best showing different purchase patterns from the target buyer are referenced.

One way to get the bounding function is to use a support vector machine (SVM). SVM is basically a method of classifying observations that fall into two categories. The goal of SVM is to find the optimal hyperplane that separates given data into two groups as far as possible.

Since there are generally many boundary functions that classify data belonging to different categories, SVM chooses the optimal boundary function among these many boundary functions. The distance until the boundary function meets the data is called the margin. The optimal classification hyperplane is the hyperplane that bisects the center of the largest margin. The data touching the marginal area is called 'Support Vector'.

In content-based filtering, an item's rating is determined by the distance from the SVM Decision Boundary to the item. Items with higher ratings are more likely to be purchased by the target buyer. Conversely, items with lower ratings are less likely to be purchased by the target buyer.

However, the scope of the present invention is not limited thereto. The present invention may be performed by various classifiers such as Linear Discriminant Analysis (LDA), Quadratic Discriminant Analysis (QDA), and K Nearest Neighbor (KNN) in addition to SVM.

Referring back to FIG. 3, in step S130, the recommended item is determined. The boundary function and the purchase history of the closest neighbors are used to determine the recommended item. The item is scored based on the collaboration filtering score and the distance to boundary (DTB) score. The item with the higher score will be recommended to the target buyer.

The collaborative filtering score is based on the correlation coefficient between the two customers. In order to obtain a collaborative filtering score, the purchase pattern of the target buyer and the purchase pattern of the closest neighbors are referred to. The higher the correlation coefficient between the two customers, the higher the collaborative filtering score. In other words, the more similar the purchase patterns of the two customers, the higher the collaborative filtering score.

The DTB score indicates how far the item is from the decision boundary. The further the item is from the decision boundary (ie toward the B area), the higher the DTB score. Conversely, the closer the item is from the decision boundary or in the opposite direction (ie, toward the A area), the lower the DTB score. Purchase items of the closest neighbors are used to obtain the DTB score. Among the purchased items of the closest neighbors, the item located away from the decision boundary in the direction of the B area will have a high DTB score.

5 is a graph for explaining a method of calculating a DTB score. Referring to FIG. 5, items marked with triangles mean items purchased by the closest neighbors. Of these items, the item furthest from the boundary function will have the highest DTB score. Thus, the item marked with a black triangle will have the highest DTB score.

Equation 2 below is a formula for obtaining a score for an item.

S ^CF is the collaborative filtering score and S ^DTB is the DTB score. λ is a coefficient for weight adjustment of the collaborative filtering score and the DTB score. The operating mode can be selected by adjusting the λ value. For example, when λ = 0, it operates with collaborative filtering, and when λ = ∞, it operates with content based filtering.

Therefore, it is possible to improve performance by adjusting (lambda) suitably. For example, if each item has unique characteristics, λ has a large value. This is because content-based filtering has good performance in this case. On the other hand, if the neighbors show a purchase pattern that is very similar to the target buyer, λ is small. This is because collaborative filtering has excellent performance in this case.

6 is a graph showing the accuracy of the recommended result according to the lambda value. The accuracy for each classifier (LDA, QDA, SVM, KNN) is shown in the form of a graph. Referring to FIG. 6, it can be seen that the highest accuracy is shown in the vicinity of?. However, the value of λ can be appropriately adjusted as necessary.

7 is a graph showing the accuracy of the hybrid recommendation method according to the present invention. Referring to FIG. 7, the accuracy for various classifiers (LDA, QDA, SVM, KNN) is shown graphically. For all classifiers (LDA, QDA, SVM, KNN) it can be seen that the accuracy in the hybrid recommendation method according to the present invention is the highest. Next, CF with reduced data has high accuracy. The collaborative filtering method using the reduced data is λ = 0 in the hybrid recommendation method.

In addition, CB with reduced data has higher accuracy than the simple hybrid method except for QDA. The content-based filtering method using the reduced data is a case where λ = ∞ in the hybrid recommendation method according to the present invention.

8 is a table showing the improvement in processing speed when applying the hybrid recommendation method according to the present invention. Referring to FIG. 8, the processing rates for the various classifiers (LDA, QDA, SVM, KNN) are shown in the table. For all classifiers (LDA, QDA, SVM, KNN), it can be seen that the processing speed is improved when only the reduced data is considered than when all the data are calculated. The reduced data here includes only the purchase history of the target buyer, the purchase history of the closest buyer, and the purchase history of the best buyer.

9 is a block diagram briefly illustrating a configuration of a hybrid recommendation system to which the present invention is applied. Referring to FIG. 9, the purchaser 101 may access the web server 102 through a network and receive a recommendation service. The web server 102 presents the buyer with web content including general recommendations. The network performs a communication connection between a plurality of to allow data communication between each other, such network includes the Internet, wireless communication network and the like.

The recommendation system 103 refers to the profile database 104 to recommend suitable items to the purchaser 101. Profile database 104 includes purchase history of buyers. The recommendation system 103 will divide the buyers in the profile database 104 into best neighbors and most like neighbors. The recommendation system 103 will find the boundary function with reference to the purchase history of the best neighbors. The recommendation system 103 will select the item to be recommended by referring to the purchase history of the closest neighbors.

As described above, the concept of closest neighbor and best neighbor is used in the present invention to solve the scalability problem. As the most likely neighbors and the best use the concept of neighbors, only useful information is subject to calculation, thereby reducing the amount of data to be processed. As the amount of data to be processed decreases, the time taken to generate the recommendation results also decreases.

In addition, collaborative filtering techniques are used to solve the sparsity problem. The leanness problem can be solved by using the purchase pattern of closest neighbors with a purchase pattern similar to the target buyer. In addition, the accuracy of the recommendations may also increase.

In order to facilitate understanding of the present invention, embodiments related to product recommendation in electronic commerce have been described, but the scope of the present invention is not limited thereto. The present invention can be generally applied to working with a large amount of data. For example, it is possible to select the recommended item by referring to the work history of the neighbors having similar behavior patterns and the neighboring neighbors having different behavior patterns among the plurality of users. The user here may be a buyer of a product, as already described, or an individual searching for a specific piece of data from a large volume of data.

That is, when the target user wants to search for specific data, it is possible to select the recommendation data by referring to the data search history of different neighbors who have searched for data different from similar neighbors who have searched for similar data. In other words, the present invention can be applied to all activities for providing data suitable for a target user among large amounts of data.

It will be apparent to those skilled in the art that the structure of the present invention can be variously modified or changed without departing from the scope or spirit of the present invention. In view of the foregoing, it is intended that the present invention cover the modifications and variations of this invention provided they fall within the scope of the following claims and equivalents.

1 shows an item attribute table used in content-based filtering.

2 shows a transaction data table used in collaborative filtering.

3 is a flowchart illustrating a hybrid recommendation method according to the present invention.

4 is a graph for explaining a method of obtaining a boundary function.

5 is a graph for explaining a method of calculating a DTB score.

6 is a graph showing the accuracy of the recommended result according to the lambda value.

7 is a graph showing the accuracy of the hybrid recommendation method according to the present invention.

8 is a table showing the improvement in processing speed when applying the hybrid recommendation method according to the present invention.

9 is a block diagram briefly illustrating a configuration of a hybrid recommendation system to which the present invention is applied.

Claims (14)

Selecting among the plurality of buyers, neighbors having a similar purchase pattern as the target buyer and farthest neighbors having different purchase patterns; Obtaining a boundary function by referring to the purchase history of the target buyer and the purchase history of the different neighbors; And And selecting a recommendation item by referring to the purchase history of the similar neighbors and the boundary function. The method of claim 1, And the similar neighbors and the different neighbors are selected by calculation of a Pearson product-moment correlation coefficient. The method of claim 2, A hybrid recommendation method in which a buyer having a high Pearson product moment correlation coefficient value is preferentially selected as the similar neighbor. The method of claim 2, A hybrid recommendation method in which a buyer with a low Pearson product moment correlation coefficient value is preferentially chosen as the difference neighbor. The method of claim 1, The boundary function is determined to distinguish between the item purchased by the target buyer and the item purchased by the different neighbors. The method of claim 1, The recommendation item is a hybrid recommendation method is selected based on the collaboration filtering score and distance to boundary (DTB) score. The method of claim 6, The collaborative filtering score is determined based on a correlation between the purchase pattern of the target buyer and the purchase pattern of the similar neighbors. The method of claim 6, The DTB score is determined according to the degree to which the purchased items of the similar neighbors are separated from the boundary function. The method of claim 6, And the recommendation item is selected by referring to a sum of a value obtained by multiplying the collaboration filtering score and the DTB score by a weight adjustment factor. The method of claim 9, The weight adjustment coefficient may be variable for selective performance of content-based filtering or collaborative filtering. 11. The method of claim 10, The hybrid recommendation method performed by the collaborative filtering method when the weight adjustment coefficient has a value of zero. 11. The method of claim 10, The hybrid recommendation method performed by the content-based filtering method when the weight adjustment coefficient has an infinite value. Selecting among the plurality of users, neighbors having similar behavior patterns as those of the target user, and different neighbors having different behavior patterns; Obtaining a boundary function by referring to the task history of the target user and the task history of the different neighbors; And And selecting a recommendation item by referring to the work history of the similar neighbors and the boundary function. A profile database including purchase history of buyers; And Including a referral system, The recommendation system may select nearest neighbors having similar purchase patterns and target neighbors and different neighboring patterns having different purchase patterns from among a plurality of buyers by referring to the profile database. Hybrid recommendation system to obtain a boundary function by referring to the purchase history and the purchase history of the different neighbors, and to select the recommended item by referring to the purchase history of the similar neighbors and the boundary function.

Images