KR102154411B1 - A recommendation system for product purchase using collaborative filtering algorism and method thereof - Google Patents

Abstract

The present invention relates to a product purchase-related recommendation system and method using a collaborative filter algorithm, the step of storing product information for a plurality of products, and for existing customers who have purchased at least one of the plurality of products for a predetermined period. Storing existing customer information, generating purchase data based on whether or not the existing customer has purchased each product for each of the plurality of products and the frequency of purchase, and a result of applying a positive or negative Pearson correlation to the purchase data And calculating a recommendation score corresponding to a purchase possibility of a new customer for a predetermined product.
Accordingly, when calculating a recommendation score for a given product, it is possible to improve recommendation performance by considering the contribution ratio of data to each Pearson correlation when the existing customer purchased the product and when not.

Description

A recommendation system and method related to product purchase using a collaborative filter algorithm {A RECOMMENDATION SYSTEM FOR PRODUCT PURCHASE USING COLLABORATIVE FILTERING ALGORISM AND METHOD THEREOF}

The present invention relates to a product purchase-related recommendation system and method in which recommendation performance is improved by considering both the case where the existing customer purchased the product and the case not purchased when calculating a recommendation score for a predetermined product using a collaborative filter algorithm. will be.

The recommendation system has become an element that determines the competitiveness of service providers in online businesses such as search, advertisement marketing, e-commerce, and multimedia content distribution. The personalized targeted recommendation system can significantly affect corporate performance by shortening the time it takes to reach an item that a customer wants online, and inducing consumption by providing one-on-one marketing.

Various recommendation techniques have been proposed so far, but one of the most successful recommendation techniques among them is Collaborative Filtering (CF). It is a technique that analyzes the purchase propensity of multiple customers on a large scale through their respective past purchase records, and recommends an item to a specific customer by referring to the purchase record purchased by another customer having a high correlation in the purchase propensity with a specific customer. Many successful e-commerce companies are recommending items to customers using this collaborative filtering recommendation technique.

However, since the collaborative filtering recommendation method recommends based on past purchase records, the recommendation performance is degraded even when the data to derive the correlation is sparse due to insufficient purchase records compared to the number of items, and the recommendation score is considered. In general, there is a problem in that the accuracy of the prediction result for the actual product purchase possibility may be deteriorated because only the case where the product is purchased by another customer is considered.

KR 10-1567684 B1 KR 10-1567217 B1

An object of the present invention is to solve the above problems, and when calculating a recommendation score for a predetermined product, a collaborative filter capable of improving recommendation performance by considering both the case where the existing customer purchased the product and the case not purchased The object of the present invention is to provide a system and method for recommending product purchases using algorithms.

A method for recommending product purchase according to an aspect of the present invention for achieving the above object includes storing product information for a plurality of products, and for existing customers who have purchased at least one of the plurality of products for a predetermined period. Storing existing customer information for each of the plurality of products, generating purchase data based on whether the existing customer has purchased each product and the frequency of purchase for each of the plurality of products, and applying a positive or negative Pearson correlation to the purchase data And calculating a recommendation score corresponding to a purchase possibility of a new customer for a predetermined product based on the result.

In addition, a product purchase-related recommendation system according to another aspect of the present invention is a storage for storing product information for a plurality of products and existing customer information for an existing customer who has purchased at least one of the plurality of products for a predetermined period. Based on the wealth, a purchase data generation unit that generates purchase data based on whether the existing customer purchases each product for each of the plurality of products and the purchase frequency, and a result of applying a positive or negative Pearson correlation to the purchase data It characterized in that it comprises a recommendation score calculation unit that calculates a recommendation score corresponding to the purchase possibility of a new customer for a predetermined product.

According to the present invention, when calculating a recommendation score for a predetermined product, the effect of improving the recommendation performance by considering the contribution ratio of the data for each Pearson correlation when the existing customer purchased the product and when not. have.

In addition, according to the present invention, it is universally applicable when calculating recommendation scores for various types of products through a simple algorithm that selects the optimal combination according to the expected purchase possibility of new customers based on the purchase status of existing customers and the purchase frequency. There is an advantage.

1 is a block diagram showing the configuration of a product purchase-related recommendation system using a collaborative filter algorithm according to an embodiment of the present invention.
2 is a diagram for explaining a process of calculating a recommendation score by the recommendation score calculation unit of FIG. 1,
3 is a flow chart showing a method for recommending product purchases using a collaborative filter algorithm according to an embodiment of the present invention.

Specific matters, including the problems to be solved, means for solving the problems, and effects of the invention as described above, are included in the following examples and drawings. Advantages and features of the present invention, and a method of achieving the same will become apparent with reference to the embodiments described later in detail together with the accompanying drawings. The same reference numerals refer to the same components throughout the specification.

1 is a block diagram showing the configuration of a product purchase-related recommendation system using a collaborative filter algorithm according to an embodiment of the present invention, and FIG. 2 is a diagram illustrating a process of calculating a recommendation score by the recommendation score calculation unit of FIG. 1 It is a drawing to do.

Hereinafter, a system for recommending a product purchase using a collaborative filter algorithm according to an embodiment of the present invention will be described with reference to the aforementioned drawings.

Referring to FIG. 1, a product purchase-related recommendation system according to an embodiment of the present invention largely includes a storage unit 100, a purchase data generation unit 300, and a recommendation score calculation unit 500.

The storage unit 100 may be provided in a predetermined database format, product information 101 for a plurality of products, and an existing customer for an existing customer i who has purchased at least one of the plurality of products for a predetermined period Each of the information 102 is stored.

Here, the product information 101 includes information on product names and product types of each of a plurality of products, and the existing customer information 102 includes information on a customer name and a product name corresponding to the product purchased by the customer. can do.

At this time, the product corresponding to the product information 101 is not limited to a tangible object that can be owned by purchasing the product, and an intangible value that can be viewed or operated through a purchase act of paying a predetermined ticket value. It may also include cultural products (eg, movies, performances, concerts, etc.).

The purchase data generation unit 300 generates purchase data D based on whether the existing customer i purchases each product (Wp) and the purchase frequency (Fp) for each of the plurality of products previously stored in the storage unit 100 .

Here, the purchase data generation unit 300 includes a first predetermined number in which each of a plurality of products previously stored in the storage unit 100 is used as one column and each of a plurality of customers previously stored in the storage unit 100 is used as a row. After the matrix M1 is configured, first purchase data D _{1 in} which a status value in a binary form of whether a predetermined customer purchases each product may be matched with corresponding row and column positions may be generated.

At this time, the status value is a 1-bit binary format, and when the customer corresponding to the row purchases the product corresponding to the column, the status value is 1, which is a 1-bit binary number. It may have a status value of 0, which is a 1-bit binary number different from the status value when purchased.

Table 1 below shows an example of first purchase data D ₁ generated in a predetermined matrix configuration corresponding to the aforementioned first matrix M1.

Here, the first matrix M1 is composed of a total of n rows and a total of m columns when the total number of customer information previously stored in the storage unit 100 is'n' and the total number of product information is'm'. It can be configured in the form of an'n×m' matrix.

For example, referring to Table 1, in the case of'customer 1'in the first purchase data (D), as'product 1'is purchased and'product 2'is not purchased, the first row corresponding to'product 1' The position of the first column may have a status value of "1", and the position of the second column of the first row corresponding to "Product 2" may have a status value of "0".

In addition, in the case of'customer n'in the first purchase data (D), as'product 2'is non-purchased and'product m'is purchased, the position of the second column of the nth row corresponding to'product 2'is "0". It has a status value of ", and may have a status value of "1" at a position of the m-th column of the n-th row corresponding to'product m'.

Here, the purchase data generation unit 300 sets each state value according to whether or not the existing customer i has purchased a predetermined product as one column, and each state value according to whether or not a new customer a has purchased the product. After configuring a second matrix M2 with a status value as one row, when the existing customer i and the new customer a each have a predetermined status value, the number of purchases (Fp) corresponding thereto Second purchase data D ₂ matching the corresponding row and column positions may be generated.

At this time, the status value is a 1-bit binary number format, and the purchase data generation unit 300 includes a first status value ("0" or "1") of whether a new customer a corresponding to a predetermined row has been purchased. ) And a second status value ("0" or "1") of whether the existing customer (i) has purchased or not corresponding to a predetermined column, and at a position corresponding to any one specific row and column, each row and column Binary bits according to the first combination in which the corresponding first state value and the second state value are sequentially listed, that is, a combination of the first state value and the second state value, when neither existing customer nor new customer has purchased "00", which is a row, "01", which is a binary bit string according to the second combination, when the existing customer does not purchase and the new customer purchases, and the third combination is when existing customers purchase and new customers do not. The second purchase data may be generated by matching the purchase frequency corresponding to the binary bit string "10" and the binary bit string "11" according to the fourth combination when both existing and new customers have purchased.

Table 2 below shows an example of second purchase data D generated in a predetermined matrix structure corresponding to the above-described second matrix M2.

For example, referring to Table 2, in the second purchase data D, the first state value when the'new customer (a)' does not purchase a predetermined product is '0 (non-purchase)', and the'existing customer (a)' (i) When '1 (purchase)' is the second status value when the product is purchased, each of the first status value ('0') and the second status value ('1') The frequency for the corresponding row and column is "10".

In addition, the first status value when the'new customer (a)' purchases the product is '1 (purchase)', and the second status value when the'existing customer (i)' purchases the product When is '1 (purchase)', the frequency of the row and column corresponding to each of the first state value ('1') and the second state value ('1') is "11".

In this case, the sum of the status values when the'existing customer (i)' purchased the product regardless of whether the'new customer (a)' has purchased the product is the sum of "10" and "11" Can have

The recommendation score calculation unit 500 is based on the result of applying a positive or negative Pearson correlation to the purchase data D generated by the purchase data generation unit 300, the possibility of purchasing a new customer a for a predetermined product. A recommendation score (P _ak ) corresponding to is calculated.

Here, the recommendation score calculation unit 500 is the amount when the existing customer (i) purchases the product (v _ij = 1) and the existing customer (i) does not purchase the product (v _ij = 0). A recommendation score (P _ak ) can be calculated by reflecting all of the similar weight values (w(a,i), hereinafter referred to as'w') based on the result of applying the Pearson correlation of and the negative Pearson correlation.

At this time, the Pearson correlation corresponds to a kind of statistical analysis method used to test a linear association between two continuous variables, for example, a correlation coefficient between "-1" to "+1" is derived, the correlation If the coefficient has a positive (+) value close to "+1", it is determined that there is a positive linear relationship between the two variables, and the correlation coefficient has a negative (-) value close to "-1" It can be determined that there is a negative linear relationship between the two variables.

)의 제1 차분값(dv₁)과, 기존 고객(i)의 복수 개의 상품 중 어느 하나의 특정 상품의 구매 여부에 대한 제2 상태값(v_ij)과 상품별 구매 여부에 대한 제2 평균 상태값(

)의 제2 차분값(dv₂)을 산출한 후, 복수 개의 상품 각각에 대하여 상기 제1 차분값(dv₁)과 상기 제2 차분값(dv₂)을 곱해서 합한 값을 분자로 하고, 복수 개의 상품 각각에 대한 상기 제1 차분값(dv₁)의 제곱합과 상기 제2 차분값(dv₂)의 제곱합의 곱의 제곱근값을 분모로 하여 유사가중치값(w)을 산출할 수 있다.Here, the recommendation score calculation unit 500 calculates the above-described similar weight value w, the first state value (v _aj) for whether to purchase any one specific product among the plurality of products of the new customer (a). ) And the first average state value (

) A first difference value (dv ₁₎ and a second state value (v 2 average state of whether _ij) and Product Purchase on any one of whether to buy a particular item of the plurality of items of an existing customer (i) the value(

) A second difference value (dv ₂₎ then calculated, and the for each of a plurality of product sum by multiplying said first difference value (dv ₁₎ and the second difference value (dv ₂₎ values in the molecule, a plurality of The similar weight value w may be calculated using a square root value of a product of the sum of squares of the first difference value dv ₁ and the sum of squares of the second difference value dv ₂ for each of the products as a denominator.

In this regard, referring to FIG. 2, when the total number of product information 101 previously stored in the storage unit 100 is'm' and the total number of existing customer information 102 is'n','i The xj' area is matrix data for the status value corresponding to the j-th product purchased by the existing customer (i), and the'a×j' area is the status value corresponding to the j-th product purchased by the new customer (a). May be matrix data for.

In this case, j may correspond to any one of natural numbers from 1 to m, and i may correspond to any one of natural numbers from 1 to n.

In this case,'v _ij ' is a value corresponding to _the'i ×j' area and indicates the status value of whether or not the product corresponding to the j-th order of the total m products of the existing customer (i) has been purchased, _{and'v aj} 'Is a value corresponding to the'a×j' area, and may indicate a status value of whether or not a product corresponding to the j-th order of the total m products of the new customer a is purchased.

), 제2 상태값(v_ij) 및 제2 평균 상태값(

)을 이용하여 수식으로 나타낸 것이다.Equation 1 below represents the above-described similar weight value (w) as a first state value (v _aj ) and a first average state value (

), the second state value (v _ij ), and the second average state value (

It is expressed as an equation using ).

)과 제2 평균 상태값(

)을 각각 수식으로 나타낸 것이다.In addition, Equation 2 and Equation 3 below are the first average state value (

) And the second average state value (

) Is represented by an equation.

Here,'m' denotes the total number of products,'v _aj ' denotes a first state value for whether a product corresponding to the j-th order of the total m products of the new customer (a) has been purchased, and'v _ij ' represents a second state value of whether or not a product corresponding to the j-th order is purchased among a total of m products of the existing customer i.

Table 3 below summarizes the types of four rule operations corresponding to combinations according to the signs of the above-described second state value (v _ij ) and similar weight value (w).

In this case, when the four-rule operation is'add(+)', the similar weight value (w) corresponding to the combination is added when calculating the recommended score (P _ak ), and the four-law operation is'subtract. In the case of (-))', the similar weight value (w) corresponding to the combination can be subtracted when calculating the recommendation score (P _ak ).

For example, referring to Table 3, when the second state value (v _ij ) is "1", that is, when an existing customer (i) purchases the j-th product, if the first similar weight value (w) is If the value is positive (w> 0), when calculating the recommendation score (P _ak ), the first similar weight value (w) is summed (+), and the second similar weight value (w) is negative (w < If it has a value of 0), the second similar weight value w may be subtracted (-) when calculating the recommendation score (P _ak ).

In addition, when the second state value (v _ij ) is "0", that is, when the existing customer (i) does not purchase the j-th product, if the first similar weight value (w) is positive (w> 0) If it has a value of, the first similar weight value (w) is subtracted (-) when calculating the recommendation score (P _ak ), and conversely, the second similar weight value (w) is negative (w <0). If so, the second similarity weight value w may be summed (+) when calculating the recommendation score _Pak .

Specifically, the recommendation score calculation unit 500 includes a first similar weight value (w(a, i) _1, or less) based on a result of applying a positive Pearson correlation when an existing customer i purchases a predetermined product. w ₁ ) and a second similar weight value based on the result of applying negative Pearson correlation when the existing customer (i) purchased the product (w(a,i) _2, hereinafter referred to as w ₂ ) And, a third similar weight value (w(a,i) _3, hereinafter referred to as w ₃ ) based on the result of applying a positive Pearson correlation when the existing customer (i) did not purchase the product, and the existing customer (i) After calculating the fourth similar weight value (w(a,i) _4, hereinafter referred to as w ₄ ) based on the result of applying the negative Pearson correlation when this product is not purchased _, 1 Sum or subtract depending on whether each of the similar weight value (w ₁ ), the second similar weight value (w ₂ ), the third similar weight value (w ₃ ), and the fourth similar weight value (w ₄ ) is selected. A recommendation score (P _ak ) may be calculated based on one result.

Equation 4 below is a summary of the formula for the recommendation score (P _ak ) in consideration of the Pearson correlation and whether the existing customer (i) and the new customer (a) are purchased, and Equation 5 is a predetermined gain constant ' The equation for k _a 'is expressed using the similar weight value (w).

Here, w(a,i) ₁ is the first similar weight value, w(a,i) ₂ is the second similar weight value, w(a,i) ₃ is the third similar weight value, and w(a) ,i) ₄ is the fourth pseudo-weighted value.

In addition, l ₁ indicates whether the first similar weight value is selected, l ₂ indicates whether or not the second similar weight value is selected, l ₃ indicates whether or not the third similar weight value is selected, and l ₄ is It indicates whether or not the fourth similar weight value is selected, and the above-described l ₁ to l ₄ may have a value of '1' when all are selected, and a value of '0' when not selected.

At this time, in Equation 4, the first similar weight value (w ₁ ), the second similar weight value (w ₂ ), the third similar weight value (w ₃ ), and the fourth similar weight value (w ₄ ) The four rule operation codes ('+'or'-') may be based on Table 3 above.

In this case, the first similar weight value (w ₁ ) and the second similar weight value (w ₂ ) corresponding to the case where the existing customer (i) purchases the j-th product are positive (w> 0) and negative values, respectively. As it has a value of (w <0), the first similarity weight value (w) is summed (+) when calculating the recommendation score (P _ak ), and the second similarity weight value (w) is the recommendation score (P _ak ) It can be subtracted (-) when calculating.

In addition, when the existing customer (i) does not purchase the j-th product, the corresponding third similarity weight value (w ₃ ) and the fourth similarity weight value (w ₄ ) are positive (w> 0) and negative values, respectively. As it has a value of (w <0), the third similarity weight value (w ₃ ) is subtracted (-) when calculating the recommendation score (P _ak ), and the fourth similarity weight value (w ₄ ) is the recommendation score (P _{When ak} ) is calculated, it can be summed (+).

On the other hand, the recommendation score calculation unit 500 is the purchase frequency according to the first similar weight value (w ₁ ) and the corresponding status values (v _ij ,v _aj ) of each of the existing customers (i) and new customers (a) The first value multiplied by the first contribution factor based on (Fp), the second similar weight value (w ₂ ), and the corresponding status values (v _ij ,v _aj ) of the existing customers (i) and new customers (a) The second value multiplied by the second contribution factor based on the purchase frequency (Fp) according to ), the third similar weight value (w ₃ ), and the corresponding status value of each of the existing customers (i) and new customers (a) ( The third value multiplied by the third contribution factor based on the purchase frequency (Fp) according to v _ij ,v _aj ), the fourth similar weight value (w ₄ ) and corresponding existing customers (i) and new customers (a) After calculating a fourth value obtained by multiplying the fourth contribution coefficient based on the purchase frequency Fp according to each state value (v _ij ,v _aj ), the first value, the second value, the third value, and A recommendation score _Pak may be calculated based on a result of adding or subtracting each of the fourth values according to whether or not each of the fourth values is selected.

Equation 6 below is a summary of the formula for the recommendation score (P _ak ) in consideration of the Pearson correlation and the contribution according to the purchase or not and the purchase frequency of the existing customer (i) and the new customer (a).

In addition, Equations 7 to 10 below are a summary of equations for each of the first value, the second value, the third value, and the fourth value described above.

은 제1 기여도 계수를 나타내고,

은 제2 기여도 계수를 나타내고,

은 제3 기여도 계수를 나타내고,

은 제4 기여도 계수를 나타낸다.here,

Represents the first contribution coefficient,

Represents the second contribution coefficient,

Represents the third contribution coefficient,

Represents the fourth contribution factor.

At this time, "00" is the number of purchases corresponding to "00", which is a binary bit string according to the first combination, that is, the status value of the existing customer (i) is '0 (non-purchase)' and the status value of the new customer (a) is In the case of '0 (non-purchase)', it indicates the number of purchases for (0,0), which is the corresponding location, and "11" is the number of purchases corresponding to "11", which is a binary bit string according to the fourth combination, that is, existing customers When the status value of (i) is '1 (purchase)' and the status value of new customer (a) is '1 (purchase)', it indicates the frequency of purchase for the location (1,1), and "10 "Is the purchase frequency corresponding to "10", which is a binary bit string according to the third combination, that is, the status value of the existing customer (i) is '1 (purchase)' and the status value of the new customer (a) is '0 (non-purchase) )'represents the number of purchases for the location (1,0), which corresponds to the case of the second combination, and "01" is the number of purchases corresponding to the binary bit string "01" according to the second combination, that is, the In the case where the status value is '0 (non-purchase)' and the status value of the new customer (a) is '1 (purchase)', it may indicate the frequency of purchase for the corresponding location (0,1).

Here, as described above, w(a,i) ₁ is the first similar weight value, w(a,i) ₂ is the second similar weight value, and w(a,i) ₃ is the third similar weight value. And w(a,i) ₄ is the fourth similar weight value.

In addition, l ₁ indicates whether the sum of the first similar weight values is selected, l ₂ indicates whether or not the sum of the second similar weight values is selected, and l ₃ indicates whether or not the sum of the third similar weight values is selected. l ₄ indicates whether or not the sum of the fourth similar weight values is selected, and l ₁ to l ₄ described above may have a value of '1' when all are selected and a value of '0' when not selected.

Hereinafter, an embodiment of calculating a recommendation score using the device for analyzing data related to product purchase using a collaborative filtering algorithm according to the present invention will be described as follows.

Here, the apparatus for analyzing data related to product purchase according to the present invention may include all of the above-described components, that is, the storage unit 100, the purchase data generation unit 300, and the recommendation score calculation unit 500.

Table 4 below shows the number of customers, the number of products, the number of purchases, and the purchase rate corresponding to each area based on a plurality of areas shown in FIG. 2 by the recommendation score calculation unit 500.

In addition, in Table 5 below, a recommendation score (P _ak ) is calculated by applying the data based on Table 4 to Equation 6, but the first similar weight value (w ₁ ) based on a purchase by an existing customer (i) ) And the second similarity weight value (w ₂ ), the first recommendation score calculated by reflecting only the first similarity weight value (w ₁ ) and the second similarity weight value (w) based on the purchase by an existing customer (i) _{In addition to 2} ), the second recommendation score calculated by reflecting both the third similar weight value (w ₃ ) and the fourth similar weight value (w ₄ ) based on the case that the existing customer (i) did not purchase is calculated by product (N = 1,2,3,4,5).

Here, the first similar weight value (w ₁ ) is based on the result of applying a positive Pearson correlation when the existing customer (i) purchases a predetermined product, and the second similar weight value (w ₂ ) is the existing customer ( i) is based on the result of applying negative Pearson correlation when the product is purchased, and the third pseudo-weight value (w ₃ ) is positive Pearson correlation when the existing customer (i) does not purchase the product. It is based on a result of applying the relationship, and the fourth similar weight value w ₄ may be based on a result of applying a negative Pearson correlation when the existing customer i does not purchase the corresponding product.

For example, referring to Table 5 above, comparing the accuracy of the first recommendation and the accuracy of the second recommendation, when recommending one product, recommending two products, recommending three products, recommending four products. In one case, it can be seen that the second recommendation accuracy is higher than the first recommendation accuracy, so it can be seen that the recommendation performance is highest when both the case where the existing customer (i) purchased and the case not purchased is considered. .

3 is a flowchart showing a method of analyzing data related to product purchase using a collaborative filter algorithm according to an embodiment of the present invention.

Hereinafter, a method of analyzing product purchase-related data using a collaborative filter algorithm according to an embodiment of the present invention will be described with reference to FIG. 3 and the above-described drawings.

First, product information for a plurality of products is stored (S100).

Here, the product information 101 may include information on product names and product types of each of a plurality of products.

At this time, the product corresponding to the product information 101 is not limited to a tangible object that can be owned by purchasing the product, and an intangible value that can be viewed or operated through a purchase act of paying a predetermined ticket value. It may also include cultural products (eg, movies, performances, concerts, etc.).

Next, existing customer information about an existing customer who has purchased at least one of a plurality of products previously stored in step S100 for a predetermined period is stored (S200).

Here, the existing customer information 102 may include information on a customer name and a product name corresponding to a product purchased by the customer.

Next, purchase data D based on whether the existing customer i purchases each product (Wp) and the purchase frequency (Fp) for each of the plurality of products is generated (S300).

Here, the step of generating the purchase data (D) (S300) is a predetermined first matrix in which each of the plurality of products previously stored in step S100 is used as one column, and each of the plurality of customers previously stored in step S200 is used as one row. After configuring (M1), it may include a step (S320) of generating first purchase data (D ₁ ) in which a status value in the binary form of whether or not a predetermined customer purchases each product is matched with corresponding row and column positions. have.

At this time, the status value is a 1-bit binary format, and has a value of '1' when a customer corresponding to the row purchases a product corresponding to the column, and a value of '0' otherwise. I can.

In addition, in the step of generating purchase data (D) (S300), each state value according to whether or not the existing customer (i) has purchased a predetermined product as one column, and the purchase of a new customer (a) for the product When a predetermined second matrix (M2) in which each status value according to whether or not is set as one row is constructed, when the existing customer (i) and the new customer (a) each have a predetermined status value, the corresponding A step S340 of generating second purchase data D _{2 in} which the purchase frequency Fp is matched to corresponding row and column positions (S340) may be further included.

At this time, the status value is a 1-bit binary number format, and the first status value ("0" or "1") for the purchase of a new customer (a) corresponding to a predetermined row and an existing value corresponding to a predetermined column. It includes a second status value ("0" or "1") of whether the customer (i) has purchased or not, and the first status value corresponding to the row and column, respectively, at a position corresponding to any one specific row and column, and Values sequentially listing the second state values, for example, "00", "01", "10", and "11" may be matched.

Next, based on the result of applying a positive or negative Pearson correlation to the purchase data, a recommendation score (P _ak ) corresponding to the purchase possibility of a new customer for a predetermined product is calculated (S400).

Here, the step (S400) of calculating the recommendation score (P _ak ) is when the existing customer (i) purchases the product (v _ij = 1) and the existing customer (i) does not purchase the product (v _ij =0), each of the similar weight values (w(a,i), hereinafter referred to as'w') based on the result of applying the positive and negative Pearson correlations are reflected and the recommendation score (P _ak ) is calculated. Can be calculated.

At this time, the Pearson correlation corresponds to a kind of statistical analysis method used to test a linear association between two continuous variables, for example, a correlation coefficient between "-1" to "+1" is derived, the correlation If the coefficient has a positive (+) value close to "+1", it is determined that there is a positive linear relationship between the two variables, and the correlation coefficient has a negative (-) value close to "-1" It can be determined that there is a negative linear relationship between the two variables.

)의 제1 차분값(dv₁)과, 기존 고객(i)의 복수 개의 상품 중 어느 하나의 특정 상품의 구매 여부에 대한 제2 상태값(v_ij)과 상품별 구매 여부에 대한 제2 평균 상태값(

)의 제2 차분값(dv₂)을 산출한 후, 복수 개의 상품 각각에 대하여 상기 제1 차분값(dv₁)과 상기 제2 차분값(dv₂)을 곱해서 합한 값을 분자로 하고, 복수 개의 상품 각각에 대한 상기 제1 차분값(dv₁)의 제곱합과 상기 제2 차분값(dv₂)의 제곱합의 곱의 제곱근값을 분모로 하여 유사가중치값(w)을 산출할 수 있다.Here, in the step (S400) of calculating the recommendation score (P _ak ), in calculating the above-described similar weight value (w), whether or not to purchase any one specific product among a plurality of products of the new customer (a) The first state value (v _aj ) and the first average state value (

) A first difference value (dv ₁₎ and a second state value (v 2 average state of whether _ij) and Product Purchase on any one of whether to buy a particular item of the plurality of items of an existing customer (i) the value(

) A second difference value (dv ₂₎ then calculated, and the for each of a plurality of product sum by multiplying said first difference value (dv ₁₎ and the second difference value (dv ₂₎ values in the molecule, a plurality of The similar weight value w may be calculated using a square root value of a product of the sum of squares of the first difference value dv ₁ and the sum of squares of the second difference value dv ₂ for each of the products as a denominator.

In this regard, referring to FIG. 2, when the total number of product information 101 previously stored in step S100 and step S200 is'm' and the total number of existing customer information 102 is'n', ' The i×j' area is the matrix data for the status value corresponding to the j-th product purchased by the existing customer (i), and the'a×j' area is the status corresponding to the j-th product purchased by the new customer (a). It may be matrix data for values.

In this case, j may correspond to any one of natural numbers from 1 to m, and i may correspond to any one of natural numbers from 1 to n.

In this case,'v _ij ' is a value corresponding to _the'i ×j' area and indicates the status value of whether or not the product corresponding to the j-th order of the total m products of the existing customer (i) has been purchased, _{and'v aj} 'Is a value corresponding to the'a×j' area, and may indicate a status value of whether or not a product corresponding to the j-th order of the total m products of the new customer a is purchased.

Specifically, the step (S400) of calculating the recommendation score (P _ak ) is a first similar weight value (w(a)) based on the result of applying a positive Pearson correlation when an existing customer (i) purchases a predetermined product. ,i) _1, hereinafter referred to as w ₁ ) and a second similar weight value (w(a,i) _2, based on the result of applying negative Pearson correlation when an existing customer (i) purchased the product) _, Hereinafter referred to as w ₂ ) and a third similar weight value (w(a,i) _3, hereinafter referred to as w ₃ ) based on the result of applying a positive Pearson correlation when the existing customer (i) did not purchase the product. ) And the fourth similar weight value (w(a,i) _4, hereinafter referred to as w ₄ ) based on the result of applying the negative Pearson correlation when the existing customer (i) did not purchase the product, respectively. After calculation, each of the first similar weight value (w ₁ ), the second similar weight value (w ₂ ), the third similar weight value (w ₃ ), and the fourth similar weight value (w ₄ ) is selected. Depending on whether or not, a recommendation score (P _ak ) may be calculated based on the result of adding or subtracting.

In this regard, Equation 4 corresponding to the equation for the recommendation score (P _ak ) in consideration of the Pearson's correlation and whether the existing customer (i) and the new customer (a) are purchased, can be used, and Details are omitted as already described.

In this case, the first similar weight value (w ₁ ) and the second similar weight value (w ₂ ) corresponding to the case where the existing customer (i) purchases the j-th product are positive (w> 0) and negative values, respectively. As it has a value of (w <0), the first similarity weight value (w) is summed (+) when calculating the recommendation score (P _ak ), and the second similarity weight value (w) is the recommendation score (P _ak ) It can be subtracted (-) when calculating.

On the other hand, in the step (S400) of calculating the recommendation score (P _ak ), the first similar weight value (w ₁ ) and the corresponding state values (v _ij ,v) of the existing customers (i) and new customers (a) The first value (p ₁ ) multiplied by the first contribution coefficient based on the purchase frequency (Fp) according to _aj ), the second similar weight value (w ₂ ), and corresponding existing customers (i) and new customers (a) The second value (p ₂ ) multiplied by the second contribution factor based on the purchase frequency (Fp) according to each status value (v _ij ,v _aj ), the third similarity weight value (w ₃ ), and the corresponding existing customers The third value (p ₃ ) multiplied by the third contribution coefficient based on the purchase frequency (Fp) according to the status value (v _ij ,v _aj ) of each of the (i) and new customers (a), and the fourth similar weight value ( w ₄ ) and the fourth value (p ₄₎ multiplied by the fourth contribution coefficient based on the purchase frequency (Fp) according to the status values (v _ij ,v _aj ) of each of the existing customers (i) and new customers (a) corresponding thereto. ) Is calculated, and then summed or added according to the selection of each of the first value (p ₁ ), the second value (p ₂ ), the third value (p ₃ ), and the fourth value (p ₄ ). A recommendation score (P _ak ) may be calculated based on the subtraction result.

In this regard, Equation 6 corresponding to the formula for the recommendation score (P _ak ) in consideration of the Pearson correlation and the contribution according to the purchase status and purchase frequency of the existing customer (i) and the new customer (a) can be used, Detailed contents of Equation 6 will be omitted as already described.

Accordingly, according to the present invention described above, when calculating a recommendation score for a predetermined product, when the existing customer purchases the product and when not, the recommendation performance is improved by considering the contribution ratio of the data to each Pearson correlation. There is an effect that can be made.

In addition, according to the present invention, through a simple algorithm that selects the optimal combination according to the expected purchase possibility of new customers based on the purchase status of existing customers and the frequency of purchase, it is universally applicable when calculating recommendation scores for various types of products. There is an advantage.

As described above, the present invention has been described in detail through preferred embodiments, but the present invention is not limited thereto and may be variously implemented within the scope of the claims.

In particular, the above-described content has been somewhat broadly described in terms of features and technical strengths of the present invention so that the claims of the invention to be described later can be better understood. Therefore, the concept and specific embodiments of the present invention described above are intended to serve similar purposes to the present invention. It should be recognized by those skilled in the art that it can be used immediately as a basis for designing or modifying other shapes for.

In addition, it is understood that the above-described embodiment is only one embodiment according to the present invention, and can be implemented in various modifications and altered forms within the scope of the technical idea of the present invention by those of ordinary skill in the art. You can understand. Accordingly, the disclosed embodiments should be considered from a descriptive point of view rather than a limiting point of view, and such various modifications and changes are also indicated in the claims of the present invention as falling within the scope of the technical idea of the present invention, and the scope equivalent thereto All differences within will be construed as being included in the present invention.

100: storage
300: purchase data generation unit
500: recommendation score calculation unit

Claims (10)

A storage unit that stores product information for a plurality of products and existing customer information for an existing customer who has purchased at least one of the plurality of products for a predetermined period, whether the existing customer has purchased each product for each of the plurality of products And a purchase data generation unit that generates purchase data based on the purchase frequency, and a first similar weight value based on a result of applying a positive Pearson correlation when the existing customer makes a purchase based on the purchase data, and the existing customer A second similarity weight value based on the result of applying a negative Pearson correlation in the case of purchase, a third similarity weight value based on the result of applying a positive Pearson correlation when the existing customer has not purchased, and the existing customer When the purchase is not made, a fourth similarity weight value is calculated based on a result of applying a negative Pearson correlation, and the first similarity weight value, the second similarity weight value, the third similarity weight value, and the fourth The present invention relates to a product purchase-related recommendation method using a collaborative filter algorithm using a system including a recommendation score calculation unit that calculates a recommendation score based on the result of adding or subtracting the similar weight value selected according to the selection of each similar weight value. ,
Storing product information for a plurality of products in the storage unit;
Storing existing customer information about an existing customer who has purchased at least one of the plurality of products for a predetermined period in the storage unit;
Consists of a number of rows corresponding to the customer information stored in the purchase data generation unit and a number of columns corresponding to each of the plurality of products, and a row corresponding to the customer according to whether each customer has purchased any one product. In the column corresponding to the product, the first purchase data having a 1-bit binary status value when purchased by an existing customer and a 1-bit binary status value different from the status value when the existing customer has not purchased Generating;
The purchase data generation unit consists of two rows corresponding to a first status value according to whether a new customer has purchased a predetermined product, and two columns corresponding to a second status value according to whether an existing customer has purchased the product. In the row and column positions corresponding to the first state value and the second state value, the combination of the first state value and the second state value, according to the first combination, in which neither the existing customer nor the new customer has purchased. Binary bit string, binary bit string according to the second combination when existing customers do not purchase and new customers purchase, binary bit string according to the third combination when existing customers purchase and new customers do not, existing customers Generating second purchase data obtained by matching a purchase frequency corresponding to each of the binary bit strings according to the fourth combination when both the and new customers have purchased;
A first similar weight value based on the result of applying a positive Pearson correlation when the existing customer purchases the product in the recommendation score calculation unit, and a negative Pearson correlation when the existing customer purchases the product. A second similar weight value based on the result of applying the product, a third similar weight value based on the result of applying a positive Pearson correlation when the existing customer does not purchase the product, and the existing customer does not purchase the product. Calculating fourth similar weight values based on a result of applying the negative Pearson's correlation in the case of not being the same;
A first value obtained by multiplying the first similarity weight value by the recommendation score calculation unit and a first contribution coefficient based on a purchase frequency according to the state values of the existing and new customers corresponding thereto, and the second similarity weight value And a second value obtained by multiplying a second contribution coefficient based on a purchase frequency according to the status value of each of the existing and new customers corresponding thereto, and the third similar weight value and the corresponding existing and new customers respectively A third value multiplied by a third contribution coefficient based on the purchase frequency according to the status value of, and the fourth similar weight value and a fourth contribution based on the purchase frequency according to the status values of the existing and new customers corresponding thereto. Calculating fourth values obtained by multiplying the coefficients, respectively; And
Calculating a recommendation score based on a result of adding or subtracting the first value, the second value, the third value, and the fourth value, respectively, in the recommendation score calculation unit; Including,
The first contribution coefficient is a purchase frequency corresponding to a binary bit string according to a fourth combination of the second purchase data, a purchase frequency corresponding to a binary bit string according to a fourth combination, and a purchase corresponding to a binary bit string according to the first combination. It has a value by dividing by the sum of the frequency,
The second contribution coefficient is a purchase frequency corresponding to a binary bit string according to a second combination of the second purchase data, a purchase frequency corresponding to a binary bit string according to a second combination, and a purchase corresponding to a binary bit string according to a third combination. It has a value by dividing by the sum of the frequency,
The third contribution coefficient is a purchase frequency corresponding to a binary bit string according to a third combination of the second purchase data, a purchase frequency corresponding to a binary bit string according to a third combination, and a purchase corresponding to a binary bit string according to the second combination. It has a value by dividing by the sum of the frequency,
The fourth contribution coefficient is a purchase frequency corresponding to the binary bit string according to the first combination of the second purchase data, the purchase frequency corresponding to the binary bit string according to the first combination and the purchase corresponding to the binary bit string according to the fourth combination A recommendation method related to product purchase using a collaborative filter algorithm, characterized in that it has a value obtained by dividing by a value obtained by adding a frequency number. delete delete delete The method of claim 1,
In the step of calculating each of the similar weight values, the recommendation score calculation unit,
Among the plurality of products of the existing customer, a first difference value between a first state value for whether a particular product is purchased by the new customer and a first average state value for whether to purchase each product. After calculating a second difference value between the second status value for the purchase of any one specific product and the second average status value for the purchase of each product,
A numerator is a value obtained by multiplying the first difference value and the second difference value for each of the plurality of products, and a sum of squares of the first difference value and the second difference value for each of the plurality of products. A method for recommending product purchase using a collaborative filter algorithm, characterized in that the similar weight value is calculated by using the square root value of the product as a denominator. A storage unit for storing product information on a plurality of products and existing customer information on existing customers who have purchased at least one of the plurality of products for a predetermined period;
It is composed of a number of rows corresponding to pre-stored customer information and a number of columns corresponding to each of the plurality of products, and corresponding to the product in a row corresponding to the customer according to whether each customer has purchased any one product. In the column, when an existing customer purchases, it has a 1-bit binary status value, and when an existing customer does not purchase, first purchase data having a 1-bit binary status value different from the status value is generated, and It consists of two rows corresponding to the first status value according to whether a new customer has purchased or not, and two columns corresponding to the second status value according to whether or not an existing customer has purchased the product, and a first status value and a second status value. At the position of the row and column corresponding to the status value, as a combination of the first status value and the second status value, a binary bit string according to the first combination when neither the existing customer nor the new customer has purchased. Binary bit string according to the 2nd combination when new customers have purchased it, binary bit string according to the 3rd combination when existing customers have purchased but not new customers, and when both existing and new customers have purchased A purchase data generation unit that generates second purchase data obtained by matching the purchase frequency corresponding to each binary bit string according to the four combinations; And
A first similarity weight value based on the result of applying a positive Pearson correlation when the existing customer makes a purchase based on the purchase data, and a first similar weight value based on the result of applying a negative Pearson correlation when the existing customer makes a purchase. 2 A similar weight value, a third similar weight value based on the result of applying a positive Pearson correlation when the existing customer did not purchase, and a result of applying a negative Pearson correlation when the existing customer did not purchase. A first value obtained by calculating each of the based fourth similarity weight values and multiplying the first similarity weight value and a first contribution coefficient based on a purchase frequency according to the state values of the existing and new customers corresponding thereto, and the A second value obtained by multiplying a second similar weight value and a second contribution coefficient based on the purchase frequency according to the status values of the existing and new customers corresponding thereto, and the third similar weight value and the corresponding existing customer And a third value multiplied by a third contribution coefficient based on a purchase frequency according to the status value of each of the new customers, the fourth similar weight value, and the number of purchase frequencies according to the status values of the existing and new customers corresponding thereto. Each of the fourth values multiplied by the fourth contribution coefficient based on is calculated, and is recommended based on the result of summation or subtraction depending on whether the first value, the second value, the third value, and the fourth value are selected. A recommendation score calculation unit that calculates a score; Including,
The first contribution coefficient is a purchase frequency corresponding to the binary bit string according to the fourth combination of the second purchase data, the purchase frequency corresponding to the binary bit string according to the fourth combination and the purchase corresponding to the binary bit string according to the first combination It has a value obtained by dividing by the sum of the frequency and the purchase frequency corresponding to "11",
The second contribution coefficient is a purchase frequency corresponding to a binary bit string according to a second combination of the second purchase data, a purchase frequency corresponding to a binary bit string according to a second combination, and a purchase corresponding to a binary bit string according to a third combination It has a value by dividing by the sum of the frequency,
The third contribution coefficient is a purchase frequency corresponding to a binary bit string according to a third combination of the second purchase data, a purchase frequency corresponding to a binary bit string according to a third combination, and a purchase corresponding to a binary bit string according to the second combination. It has a value by dividing by the sum of the frequency,
The fourth contribution factor is a purchase frequency corresponding to a binary bit string according to a fourth combination of the second purchase data, a purchase frequency corresponding to a binary bit string according to the fourth combination, and a purchase corresponding to the binary bit string according to the first combination. A product purchase-related recommendation system using a collaborative filter algorithm, characterized in that it has a value obtained by dividing by a value obtained by adding a frequency number. delete delete delete The method of claim 6,
The recommended score calculation unit,
Among the plurality of products of the existing customer, a first difference value between a first state value for whether a particular product is purchased by the new customer and a first average state value for whether to purchase each product. After calculating a second difference value between the second status value for the purchase of any one specific product and the second average status value for the purchase of each product,
A numerator is a value obtained by multiplying the first difference value and the second difference value for each of the plurality of products, and a sum of squares of the first difference value and the second difference value for each of the plurality of products. A product purchase-related recommendation system using a collaborative filter algorithm, characterized in that the similar weight value is calculated by using the square root of the product as a denominator.

Images