KR101116026B1 - Collaborative filtering recommender system based on similarity measures using the origin moment of difference random variable - Google Patents

Abstract

The present invention proposes a new similarity measure for improving the performance of memory-based collaborative filtering using the similarity of users or similarity of items. This measure of similarity, RMS (Raw Moment-based Similarity), has a very low common preference under cold-start conditions that occur after a new user or a new item has appeared in a collaborative filtering technique. It is valid in case of inaccuracy and can be used universally in the field for determining similarity between two variables having linear characteristics.

Description

Collaborative filtering recommendation system based on similarity measure using origin moment of difference probability variable {COLLABORATIVE FILTERING RECOMMENDER SYSTEM BASED ON SIMILARITY MEASURES USING THE ORIGIN MOMENT OF DIFFERENCE RANDOM VARIABLE}

The present invention relates to a collaborative filtering recommendation system and a method of constructing a similarity table based on a new similarity measure for improving the performance of memory-based collaborative filtering using similarity of users or similarity of items.

The recommendation system, as one of personalization techniques, is to recommend content or information to the user that may be of interest to the user. In the past, rule-based algorithms, which were simply recommended according to user's preference information, prevailed, but as information technology developed, users began to understand user's behavior patterns by soft computing techniques in artificial intelligence and data mining techniques in database. Recommendation systems are evolving, such as analyzing the propensity of users and recommending information and content that users might prefer.

Collaborative filtering, one of the techniques for building a recommendation system, is known as the best recommendation methodology in existence. This collaborative filtering is a way to automatically predict users' interests based on taste information from many users. The fundamental assumption of the collaborative filtering approach is the premise that users' past trends will remain the same in the future. In other words, if user A prefers a specific style of content in the past, it is a technique that performs recommendation based on the hypothesis that it will be in the future. For example, a collaborative filtering or recommendation system for music may predict a user's preference for music using a partial list of users' likes (likes or dislikes). The system is not limited to specific user information, but uses information gathered from many users. This is different from the way it simply handles the average rating of each item's interests based on the number of votes. In other words, it is a technique to identify customers with similar patterns in preference and interest based on their preferences and expressions of interest. Products that have not yet purchased each other to customers with similar tastes are used to provide services in the form of cross-recommending or recommending related products according to the tastes or lifestyles of the classified customers.

As described above, the collaborative filtering may predict the user's preference for an item (a medium for recommendation such as information or content) by using the preference information obtained from many users. Assuming the user-item preference matrix as shown in Table 1 below in a system that can assign user preferences to eight levels of 1, 2, 3, 4, 5, 6, 7, and 8, the ultimate goal of collaborative filtering Is an accurate prediction of the part that has not yet been evaluated, that is, the part of '?' In the matrix below.

Table 1. Example of user-item preference matrix

The collaborative filtering recommendation system evaluates the performance by calculating the error between the predicted preference and the preference given by the actual user. The memory-based collaborative filtering method, one of the technical classifications of the collaborative filtering technique, is executed in three steps as follows.

1) Similarity calculation step

Similarity calculation is the most important step in a memory-based recommendation system. The user based approach calculates the similarity between the recommended user and all other users. The user to which the recommendation target user and other users have given preference in common is used. Table 2 below shows the preference data used to calculate the similarity between USER 3 and USER 4 for the purpose of predicting USER 3's preference for ITEM 4 (marked with '?'). In Table 2, users USER 3 and USER 4, which are subject to similarity calculation, can be represented by discrete variables X and Y, respectively, as X = {4, 2, 3} and Y = {8, 3, 4}. The two variables have a linear relationship.

Table 2. Example data for calculating similarity between users in a user-based approach

The item based approach calculates the similarity between the item to be recommended and all other items. The preferences of users who give preference to both items to be calculated for the similarity are used. Table 3 below shows the preference data used to calculate the similarity between ITEM 4 and ITEM 5 for the purpose of predicting USER 3's preference for ITEM 4 (marked with '?'). ITEM 4 and ITEM 5, which are subject to similarity calculation, can also be represented by discrete variables X and Y, respectively, as X = {5, 4, 5}, Y = {2, 3, 8}. There is a linear relationship.

Table 3. Example data for calculating similarity between items in an item-based approach

By measuring the similarity between the variables X and Y, we can see the similarity between the user and the user or between the items. Measures for measuring similarity include cosine similarity widely used in the field of information retrieval, Pearson product rate correlation coefficient and Spearman rank correlation coefficient, which are correlation coefficient based methods.

Equation (1) below represents cosine similarity, equation (2) represents the Spearman rank correlation coefficient, and equation (3) represents the Pearson product rate correlation coefficient.

(1)

(One)

(2)

(2)

(3)

(3)

Where X and Y are variables having linear characteristics, and Xi and Xi are i th values of variables X and Y having linear characteristics.

Whether there is a common preference between user-based or item-based approaches, similarity calculation is impossible and cannot be used for score prediction. Depending on the approach, a similarity table of items or users can be created to drastically reduce real-time recommendation time. 1 and 2 show examples of the similarity table when the total number of users is five or the total number of items is five. The update of the similarity table may be automatically performed at a time period determined by the system administrator.

2) Proximity neighbor formation step

A step of forming a neighboring neighborhood according to the similarity calculated in the step 1). Proximity neighborhood means data to be used for the prediction of preference. There is a method of considering all the data whose similarity is calculated, a Top-N method of selecting the top N, and a method of considering data above a certain similarity threshold. In the user-based approach, users are selected for score prediction, and in the item-based approach, items are used for score prediction.

Too few or too many neighboring neighbors can lead to poor performance of the recommender system. If you experiment with setting the number of neighbors to 1 and increasing it, it will gradually drop after performing at its best. For this reason, the Top-N method, which is fast in computation time and shows good performance, is used and is sometimes referred to as the K-nearest neighborhood (KNN) method.

3) Prediction Prediction Step

The similarity between the recommended user and the neighbors is weighted to predict the preference of items that do not occur to the recommended user.

Equation (4) below represents the prediction of the user based approach, and equation (5) represents the prediction of the item based approach.

(4)

(4)

(5)

(5)

Where u is the recommendation user, i is the recommendation item, r _u is the preference of items given by the recommendation user u, w _{u , v} are the similarities between user u and v, r _{v ,} i For r, r _v is the preference of user v, r _i is the preference of item i given by the user, w _{i , j} is the similarity between item i and j, r _{u , j} is the preference for user j's item j, r _j is the item j's preference given by the user.

Since the preference prediction score for a particular item is high, it is expected that the recommendation target user's preference for the item is high, so that the process of user-based collaborative filtering is completed by sorting the prediction score in descending order and recommending it to the user.

The following describes the problem of this existing similarity measure.

First, the cold start condition means that prediction performance is deteriorated when a new user or item is created in the database and the user preference data is extremely small. It is one of the conditions to be solved to improve the user preference prediction performance in the collaborative filtering technique.

The existing cosine similarity and correlation coefficient-based similarity measurement method causes a decrease in prediction performance in the preference graphs shown in FIGS. 3, 4, 5, and 6. 3, 4, 5, and 6 show a preference form frequently occurring in the cold start condition. When the non-cold start condition (a large number of common preferences) is used, FIGS. 3, 4, 5, and 6 are shown. Since it is rare to see a preference graph in the form of, a cosine similarity and a correlation-based method can guarantee a certain performance. However, as new users or new items will continue to be added to the database, it is clear that specialized similarity measures are necessary for cold start conditions.

3, 4, 5, and 6, the X axis represents an item, the Y axis represents a preference, u1 means user 1, and u2 means user 2. That is, the graph shows the preferences given to the four items by the user u1 and u2. What follows is described as a user-based approach, but the item-based approach is the same.

Referring to FIG. 3, the preferences of two users are parallel. In this case, the calculation result of the cosine similarity outputs 1 indicating that the two users are completely identical. It can only be expressed as if the two users' preferences must match completely, such as u1 = {1, 1, 1, 1}, u2 = {1, 1, 1, 1}. However, when the similarity graphs of two users are parallel, cosine similarity outputs 1 unconditionally, which causes a problem in measuring similarity.

When the preferences of two users are parallel as shown in FIG. 3, the Pearson product rate correlation coefficient and the Spearman rank correlation coefficient, which are two similarity measures based on the correlation coefficient, have a standard deviation of 0, and thus a denominator of 0 in the above-described calculation. Similarity measurements are not possible. The inability to measure similarity makes the data useless because it cannot be included in the neighbors (memory-based collaborative filtering stage 2) and thus cannot be used for preference prediction.

Referring to FIG. 4, in the case where the preference graphs of the two users show a zigzag shape, correlation-based methods output -1 indicating that the propensity of the two users is opposite. It's obviously wrong to see two users in the same list of items having the opposite preferences, even though they have similar preferences.

In the graph shown in FIG. 4A and the graph shown in FIG. 4B, the correlation coefficient-based methods also output -1, but the cosine similarity is 0.8 for the left graph and 0.96 for the right graph. Outputs In both preference graphs, the cosine similarity shows different results even though the preference difference of the two users is the same, which is an unfair result.

Referring to FIG. 5, a preference is changed in a zigzag form similar to that of FIG. 4. The Pearson product ratio correlation coefficient outputs -0.904 in the graph shown in (a) of FIG. 5 and -0.816 in the graph shown in (b). Compared with FIG. 2, despite the stronger tendency of the opposite tendency than FIG. 2, the results show that the results are more and more similar.

6 illustrates a case where two users have a common preference. In this case, the cosine similarity outputs 1 regardless of the preference value, and the correlation coefficient-based methods cannot calculate the covariance and standard deviation because they cannot be obtained. That is, in the case of one common preference, the existing similarity measures show wrong results, thereby degrading the preference prediction performance.

The technical problem to be solved by the present invention is a similarity measure using the origin moment of the difference probability variable that can improve the preference prediction performance by complementing the problems caused by the cold start condition when the collaborative filtering recommendation system predicts the user's preferences. To provide a collaborative filtering recommendation system and similarity table construction method.

In the cold start condition, the number of common preferences for calculating the similarity of a user or an item is extremely small, and the existing similarity measure causes frequent problems as shown in FIGS. It becomes a factor.

In the present invention, the polynomial probability variable (D) consisting of the difference between the two variables is obtained using the equation shown in FIG. 9, and the similarity between the two variables is calculated using the equation shown in FIG. Measure the similarity exactly. In addition, the user having various different characteristics by adjusting the moment order r in the equation shown in FIG. 10 and changing the similarity distribution as shown in FIG. Provides flexibility to make it easy to apply to the preference matrix. In addition, even if various characteristics such as the preference range, the number of users, the number of items, and the like included in the user-preference matrix are changed, the similarity measure proposed in the present invention operates accordingly. Cold start condition, of course, even if not the case shown in Figures 1, 2, 3, 4, it can be confirmed through the experimental results of FIGS.

In the cold start condition, the number of common preferences for calculating the similarity of a user or an item is extremely small, and the existing similarity measure frequently causes problems as shown in FIGS. 1, 2, 3, and 4, and thus, the preference prediction performance It is a factor of deterioration.

Accordingly, the present invention obtains a polynomial probability variable consisting of the difference between the two variables, and calculates the moments of the two variables using the variable to measure the similarity more accurately in the cold start condition. In addition, by adjusting the moment order (r), the similarity can be differentiated under various conditions showing the same preference difference, and the similarity distribution can be adjusted so that it is easy to apply to a user-preference matrix having several different characteristics. Provide flexibility. In the case of the cold start problem, but not cold start condition, the performance was not inferior to the existing similarity measure.

According to a first aspect of the present invention, a polynomial probability variable consisting of a difference between two variables based on a difference in common preferences of two users or two items is obtained, and a difference probability of measuring similarity by calculating moments of two variables using the same is calculated. A collaborative filtering recommendation system based on similarity measure using the origin moment of a variable.

A preference data receiving means for receiving a preference for an item within an integer number range defined by a system administrator input by a user;

Performance measurement means for measuring the performance of the system with respect to the preference received by the preference data receiving means, by calculating an error between the predicted preference and the item preference received from the user according to a predetermined procedure, according to a predetermined procedure;

A similarity measuring means for measuring user similarity or item similarity using a similarity measuring method based on input preferences, preferences input by a current user in the past, and preference data input by non-current users in the past,

Similar item search means for searching similar items with items that the user has not yet evaluated from the previously stored item similarity database by calculating similarity between each other for all pairs of items,

A similarity user search means for searching for users who have the same propensity as a user to recommend from a user similarity database;

(4)

(5)
(여기서, u는 추천 대상 사용자, i는 추천 대상 아이템, r_u는 추천 대상 사용자 u가 부여한 아이템들의 선호도, w_u,v는 사용자 u와 사용자 v의 유사도, r_v,i는 사용자 v의 아이템 i에 대한 선호도, r_v는 사용자 v의 선호도, r_i는 사용자들이 부여한 아이템 i의 선호도, w_i,j는 아이템 i와 j의 유사도, r_u,j는 사용자 u의 아이템 j에 대한 선호도, r_j는 사용자들이 부여한 아이템 j의 선호도이다.)The following equations (4) and (5) are selected by selecting one of a user based method based on a user searched by the user searching means and an item based method based on an item searched by the item searching means. Preference predictive means for predicting preference for items that the user has not yet rated,

(4)

(5)
Where u is the recommended user, i is the recommended item, r _u is the preference of the items given by the recommended user u, w _{u, v} is the similarity between user u and _v, and r _{v, i} is the item of user v the preference for i, r _v is the preference of user v, r _i is the preference of item i given by the user, w _{i, j} is the similarity of items i and j, r _{u, j} is the preference of user u's item j, r _j is the preference of item j given by the user.)

A recommendation item list generating means for generating a recommendation list by sorting the items in order from the highest preference by investigating the predicted preferences for the items which the user has not yet evaluated;

A similarity measure using the origin moment of the difference probability variable configured with the recommendation item list generated by the recommendation item list generating means and an item recommendation means for recommending to the user a predetermined number of items and an expected preference for the item. A collaborative filtering recommendation system based on is provided.

delete

The preferences for the received items are preferably stored in a public user-item preferences database where preferences for all other user's items are stored.

The performance measure is preferably calculated as the average of the errors each time a new preference is received from the user.

It is desirable for a system administrator to be able to limit the number of similar users by specifying specific thresholds.

In addition, according to the second aspect of the present invention, the similarity of two users to be measured similarity using the RMS to improve the performance of the collaborative filtering in the memory-based collaborative filtering recommendation system using the similarity of the user or the similarity of the items As a way to build a table,

Initializing both user B and user T variables to 0,

Increasing the user B by 1,

Setting the value of user T equal to the value of user B,

Increasing the user T by 1,

Extracting item ratings evaluated by both user B and user T;

Calculating the difference probability variable D,

Calculating 1/2, 2, 2/3, and third moments of the difference probability variable D,

Storing the calculation results in a user similarity table;

Determining whether the value of the user T is equal to the total number of users (N), and

A method of constructing a user similarity table using RMS is provided that includes determining whether the value of user B is equal to the total number of users (N).

In the method of constructing a user similarity table using the RMS, it is preferable to calculate the similarity between all users who have the preference data by using the preferences of items that the user commonly evaluates.

In the extracting the item preferences evaluated by both the user B and the user T, when the user B is 1 and the user T is 2, the extracted item preferences are {2, 3, 4} and {4, 5, 2}. It is preferable to extract.

In addition, according to the third aspect of the present invention, the similarity of two items to be measured for similarity by using RMS to improve the performance of the collaborative filtering in the memory-based collaborative filtering recommendation system using the similarity of the user or the similarity of the items As a way to build a table,

Initializing both the Item B and Item T variables to 0,

Increasing the item B by 1,

Setting the value of the item T equal to the value of the item B,

Increasing the item T by 1,

Extracting an item preference of the user who rated both the item B and the item T,

Calculating the difference probability variable D,

Calculating 1/2, 2, 2/3, and third moments of the difference probability variable D,

Storing the calculation results in an item similarity table;

Determining whether the value of the item T is equal to the total number N;

A method of constructing an item similarity table using RMS is provided that includes determining whether the value of the item B is equal to the total number of items (N).

In the method of constructing the item similarity table using the RMS, it is preferable to calculate the similarity between all users who have the preference data by using the preferences evaluated by one user on both items.

In the step of extracting the item preference of the user who rated both the item B and the item T, when the item B is 1 and the item T is 2, the item preferences of the extracted user are {1, 3} and {2, 2 } Is preferably extracted.

According to the present invention, a polynomial probability variable consisting of the differences between two variables is obtained, and the moments of the two variables are calculated using the same to measure the similarity more accurately in the cold start condition, and the moment order r is adjusted to adjust the same preference. There are features that allow similarity differentiation under different conditions, and the similarity distribution can be adjusted to provide flexibility for easy application to user-preference matrices with different characteristics.

Hereinafter, embodiments of the present invention will be described with reference to the accompanying drawings.

FIG. 7 shows the notation in the equations appearing in the description and drawings of the invention, and FIG. 8 shows a collaborative filtering recommendation system based on a similarity measure in accordance with an embodiment of the invention.

In FIG. 8, the preference data receiver 1 receives a preference for an item within an integer number range defined by a system administrator input by a user. The preferences for the received item are stored in the public user-item preferences database where the preferences for all other user's items are stored.

The performance measurer 2 measures the performance of the system by calculating an error between the previously predicted preference and the item preference received from the user according to the following procedure with respect to the preference received by the preference data receiver 1. The performance measure is calculated as the average of the errors each time a new preference is received from the user.

The RMS similarity measurement unit 3 measures the similarity proposed by the present invention by extracting the input preference, the preference input by the current user in the past, and the preference data input by people who are not the current user in the past from the public user-item preference database. RMS, which is a method, is used to measure user similarity (FIG. 14) and item similarity (FIG. 15). The measured item similarity and user similarity are stored separately.

The similar item search unit 4 searches for similar items (items to be used for preference prediction) that are not evaluated by the user from the RMS item similarity database 5, which calculates the similarity between each item pair and stores them in advance. . System administrators can specify specific thresholds to limit the number of similar users.

The similar user searcher 6 searches for users (users to be used for the preference prediction) that have a similar tendency to the user to be recommended from the RMS user similarity database 7. System administrators can specify specific thresholds to limit the number of similar users.

The preference prediction unit 8 selects one of a user-based method and an item-based method to predict a preference for an item that the user has not yet evaluated.

The recommendation item list generation unit 9 generates a recommendation list by sorting the items in descending order in the order of high predicted preference for items that the user has not yet evaluated.

The item recommendation unit 10 recommends the top n items and the expected preferences for the items to the user based on the recommended item list.

The present invention is a new measure of similarity to ameliorate the problem of degraded preference prediction performance in cold start conditions. As mentioned above, the preference graphs as shown in FIGS. 3, 4, 5, and 6 occur more frequently as well as when the common preference is relatively high, especially when the common preference is small. In other words, it occurs frequently in cold start conditions, which is a major factor in degrading the predictive performance of the collaborative filtering recommendation system. Accordingly, the present invention proposes a Raw Moment-based Similarity (RMS) using moments of random variables.

FIG. 9 is a diagram illustrating a method of generating a polynomial probability variable D that can be generated with a difference in common preference between two users or two items. D is used to calculate the RMS similarity proposed in the present invention. The polynomial probability variable (D) is based on the following multinomial random law in statistics. In the method proposed in the present invention, the number of events that can appear in the polynomial probability variable (D) is R + 1.

(6)

(6)

In a collaborative filtering recommendation system that can give integer preferences from 1 to 5, when X = {5, 3, 2} and Y = {3, 2, 4}, the polynomial probability variable ( D) will be as follows.

D = {2, 1, 2}

10 shows an equation for calculating the RMS similarity proposed in the present invention. The rth moment of the polynomial probability variable (D) generated by the equation shown in FIG. 9 is obtained and divided by the r power of the difference between the maximum and minimum preferences that can be input to the preference database, and the result ranges from 0 to 1 Normalized. 1 means that the variables X and Y are completely different, and 0 means they are exactly the same.

In order to be able to use it as a weight in predicting preferences, the more similar, the normalized value from 0 to 1 should converge to 1. Therefore, the inverse of the result is taken. FIG. 8 when X = {5, 3, 2} and Y = {3, 2, 4} in a collaborative filtering recommendation system capable of giving integer preferences from 1 to 5 (R is 5-1 = 4) By applying the polynomial probability variable (D) and applying it to the equation shown in Figure 10, RMS (X, Y) using the first moment (r = 1) can be obtained as in the following equation (7).

(7)

(7)

In addition, RMS (X, Y) using the secondary moment (r = 2) can be calculated | required as following Formula (8).

(8)

(8)

In addition, RMS (X, Y) using the square root moment can be obtained as shown in the following equation (9).

(9)

(9)

The method proposed in the present invention compensates for the problems of the existing similarity measures resulting from the situation as shown in FIGS. 3, 4, 5, and 6, and can determine the degree of similarity more intuitively and clearly in the cold start condition. . The reason is as shown in FIG. FIG. 11 shows how the similarity is expressed when the moment order r is set to the square root (1/2 square) of di, the square, the cube root of the square root (2/3 square), and the square.

In FIG. 11, two main features of the similarity measure proposed by the present invention are distinguished from the existing similarity measure. The first feature is found in the similarity change in three cases where the sum of differences in the table shown in FIG. Although the sum of the differences is equal to all 16, it shows how the differentiation can be achieved by changing the moment order r. Due to the high and low frequency of the large preference difference, it changes as indicated by the square of the graph shown in FIG.

The second feature is that the lower the moment order r is set, the similarity distribution is formed at the lower side, and the higher is set at the higher side. This flexibility of the algorithm proposed in the present invention is intended to derive high performance in the user-preference matrix, which has different characteristics (minimum and maximum values of preference, number of users, number of items, number of data of preference, etc.). It can be a means.

12 and 13 illustrate measurement results of Mean Absolute Error (MAE), which is a measure for evaluating preference prediction performance in collaborative filtering. The algorithms to be compared are COS (Cosine Similarity), PCC (Pearson Product Ratio Correlation), SRCC (Spearman Rank Correlation), and the data set uses MovieLens's data set which has been used in many previous studies. In the same way as used in many research papers, 30,000 samples were randomly sampled from a total of 100,000 preference data in [user-item-preference] format and used as test data (input data for experiments). The remaining 70,000 data were used as training data (data for constructing similarity tables).

Referring to FIG. 12, the similarity measures proposed in the present invention are mixed even though the cold start condition, which is frequently shown in FIGS. 3, 4, 5, and 6, is not a cold start condition. It can be seen that the performance is significantly improved over the existing methods. This shows that not only cold start conditions, but also similarity calculations of two variables having many common preferences lag behind existing similarity measures.

FIG. 13 shows the results of experiments by forcibly limiting the number of common preferences when calculating similarity. This is a result of selecting the nearest neighbor number that showed the best performance in FIG. 12 and artificially creating a cold start condition, and intuitively indicating how the RMS proposed in the present invention is stronger in the cold start problem than the existing similarity measure. The result is obvious. As shown in FIG. 13, it can be seen that the method proposed by the present invention is stronger in the cold start problem than the conventional method.

FIG. 14 shows a construction flowchart of a user similarity table (database) using RMS proposed in the present invention in a user-based approach of memory-based collaborative filtering. In this figure, B_USER and T_USER represent two users to be measured for similarity, and N is the maximum number of users. It shows how to calculate the similarity among all users who have preference data. An example of the structure of the constructed user similarity table (database) is shown in Table 4 below.

Table 4

ITEM 1 ITEM 2 ITEM 3 ITEM 4 USER 1 One 2 3 4 USER 2 4 5 2 USER 3 One 2 3 USER 4 3 2 3 USER 5 5 4 2

Referring to Figure 14, the construction procedure of the user similarity table (database) using the RMS proposed in the present invention is as follows.

First, in step S1, both the B_USER variable and the T_USER variable are initialized to 0, and in step S2, the B_USER is increased by one.

Next, in step S3, the value of T_USER is set equal to the value of B_USER, and in step S4, T_USER is increased by one.

Subsequently, in step S5, the item preference evaluated by both B_USER and T_USER is extracted. In Table 4, when B_USER is 1 and T_USER is 2, the extracted item preferences are {2, 3, 4} and {4, 5, 2}.

Next, the difference probability variable D is calculated in step S6, and 1/2, 2, 2/3 and cubic moments of D are calculated in step S7.

Next, in step S8, the calculation results are stored in the user similarity table (DB) 20 in the following form. for example,

     {B_USER T_USER RMS (1/2) RMS (2) RMS (2/3) RMS (3)}

ex) 1 2 Result value Result value Result value

Next, in step S9, if the value of T_USER is equal to the total number of users N, the flow advances to the next step, and otherwise returns to step S4.

Next, in step S10, if the value of B_USER is equal to the total number of users N, the flow advances to the next step, otherwise, the flow returns to step S3.

In this way, a user similarity table (database) using RMS proposed in the present invention can be constructed.

15 shows a construction procedure of an item similarity table (database) using RMS proposed in the present invention. In this figure, B_ITEM and T_ITEM indicate two items to be measured for similarity, and N is the maximum number of items.

For the RMS calculation, the user-based approach uses the preferences of items that two users have rated in common, while the item-based approach differs in using the preferences that one user has rated on both items.

An example of the structure of the constructed item similarity table (databases) is shown in Table 5 below.

Table 5

ITEM 1 ITEM 2 ITEM 3 ITEM 4 USER 1 One 2 3 4 USER 2 4 5 2 USER 3 One 2 3 USER 4 3 2 3 USER 5 5 4 2

Referring to Figure 15, the construction order of the item similarity table (database) using the RMS proposed in the present invention is as follows.

First, in step S11, both the B_ITEM variable and the T_ITEM variable are initialized to 0, and in step S12, the B_ITEM is increased by one.

Next, in step S13, the value of T_ITEM is set equal to the value of B_ITEM, and in step S14, T_ITEM is increased by one.

Subsequently, in step S15, the item preferences of the user who evaluated both B_ITEM and T_ITEM are extracted. In Table 5, when B_ITEM is 1 and T_ITEM is 2, the item preferences of the extracted users are {1, 3} and {2, 2}.

Next, the difference probability variable D is calculated in step S16, and 1/2, 2, 2/3, and third moments of D are calculated in step S17.

Next, in step S18, the calculation results are stored in the item similarity table (DB) 30 in the following form. for example,

    {B_ITEM T_ITEM RMS (1/2) RMS (2) RMS (2/3) RMS (3)}

ex) 1 2 0.5 0.6 0.7 0.8

Next, in step S19, if the value of T_ITEM is equal to the total number of items N, the flow advances to the next step; otherwise, the flow returns to step S14.

Next, in step S20, if the value of B_ITEM is equal to the total number of items N, the process advances to the next step, otherwise, the process returns to step S13.

In this way, an item similarity table (database) using RMS proposed in the present invention can be constructed.

Similarity measure RMS proposed in the present invention can be applied to various other fields as well as a collaborative filtering recommendation system. In addition to information retrieval using statistical cosine similarity and correlation coefficient-based similarity measures, statistical analysis, etc. Applicable anywhere in the field where processing for determining similarity of variables is required.

1 illustrates a similarity table for reducing execution time of user-based collaborative filtering.

2 is a diagram illustrating a similarity table for reducing execution time of item-based collaborative filtering.

3 is a diagram illustrating a case in which common preferences are parallel.

4 is a diagram illustrating a case (1) when the common preference is in a zigzag form.

5 is a diagram illustrating a case (2) when the common preference is in a zigzag form.

6 illustrates a case where one common preference is one.

It is a figure explaining the radix method in this invention.

8 is a diagram illustrating a collaborative filtering recommendation system based on a similarity measure according to an embodiment of the present invention.

FIG. 9 is a diagram for explaining a method of obtaining a polynomial probability variable D generated as an absolute value of a difference between two variables having a linear relationship.

10 is a view for explaining the similarity measure using the moment of the random variable proposed in the present invention.

11 is a view showing an example of the similarity distribution change according to the moment order adjustment of the RMS proposed in the present invention.

FIG. 12 is a diagram illustrating a full-rating experiment of absolute mean error for a change in the number of neighboring neighbors.

FIG. 13 is a view showing a cold-start experiment for measuring an average number of common preference changes. FIG.

14 is a flowchart of building a user similarity database using RMS.

15 is a flowchart of building an item similarity database using RMS.

Claims (11)

A polynomial probability variable consisting of the difference of two variables based on the difference of common preferences of two users or two items is calculated. Based collaborative filtering recommendation system, A preference data receiving means for receiving a preference for an item within an integer number range defined by a system administrator input by a user; Performance measurement means for measuring the performance of the system with respect to the preference received by the preference data receiving means, by calculating an error between the predicted preference and the item preference received from the user according to a predetermined procedure, according to a predetermined procedure; A similarity measuring means for measuring user similarity or item similarity using a similarity measuring method based on input preferences, preferences input by a current user in the past, and preference data input by non-current users in the past, Similar item search means for searching similar items with items that the user has not yet evaluated from the previously stored item similarity database by calculating similarity between each other for all pairs of items, A similarity user search means for searching for users who have the same propensity as a user to recommend from a user similarity database; The following equations (4) and (5) are selected by selecting one of a user based method based on a user searched by the user searching means and an item based method based on an item searched by the item searching means. Preference predictive means for predicting preference for items that the user has not yet rated,

(4)

(5) Where u is the recommended user, i is the recommended item, r _u is the preference of the items given by the recommended user u, w _{u, v} is the similarity between user u and _v, and r _{v, i} is the item of user v the preference for i, r _v is the preference of user v, r _i is the preference of item i given by the user, w _{i, j} is the similarity of items i and j, r _{u, j} is the preference of user u's item j, r _j is the preference of item j given by the user.) A recommendation item list generating means for generating a recommendation list by sorting the items in order from the highest preference by investigating the predicted preferences for the items which the user has not yet evaluated; The origin moment of the difference probability variable, characterized in that it comprises an item recommending means for recommending to the user a predetermined number of items and an expected preference for the item based on the recommendation item list generated by the recommendation item list generating means. Collaborative filtering recommendation system based on similarity measure. delete The method of claim 1, wherein the preference for the received item is stored in a public user-item preference database in which preferences of all other users are stored in the similarity measure using the origin moment of the difference probability variable. Based collaborative filtering recommendation system. The collaborative filtering recommendation system of claim 1, wherein the performance measure is calculated as an average of errors each time a new preference is received from a user. The collaborative filtering recommendation system of claim 1, wherein the system administrator may limit the number of similar users by designating a specific threshold value. delete delete delete delete delete delete

Images