Method and device for conditional probability-based collaborative filtering using high-dimensional binary data

Abstract

The present invention relates to a method and device for conditional probability-based collaborative filtering using high-dimensional binary data, wherein according to whether a first customer having a purchase record of an object product has purchased the object product, a conditional probability calculated from purchase data of a second customer having no purchase record of the object product and the first customer and assigned to a purchase probability value or a non-purchase probability value is accumulatively summed for each combination of the first customer for the second customer, and a recommendation score can be calculated by selectively applying the accumulatively summed result.

Description

Method and device for conditional probability-based collaborative filtering using high-dimensional binary data

[0001] The present invention relates to a technology for calculating a recommendation score indicating the purchase probability of a product, and more specifically, to a technology for calculating a recommendation score for a product for which there is no purchase record.

[0002] Advances in communication technology have made it easier to collect purchase data, including product purchase histories. Technologies that recommend products based on this data are emerging. Recommendation technologies utilizing purchase data can be broadly categorized into those that utilize learning models and those that utilize specific mathematical formulas.

[0003] Existing methods using learning models train a learning model to output a user's purchase probability based on past purchase data, then provide a recommendation score. While this approach performs well for existing users with established purchase data, it suffers from a "cold start" problem, which degrades performance for new users or products without a purchase history.

[0004] Many methods for calculating recommendation scores using Pearson correlation, which considers relationships with other users, have been disclosed, utilizing specific mathematical formulas. These methods utilize purchase data for existing products, which includes purchases for both existing and new users, and purchase data for the target product, which only includes purchases from existing users. Based on the Pearson correlation between existing and new users, they can calculate a recommendation score, which represents the probability that a new user will purchase the target product. This score represents the probability that the new user will purchase the target product. Therefore, they mitigate the cold start problem of existing methods using learning models, and can achieve high performance even for target products with insufficient purchase data.

[0005] However, existing methods using Pearson correlation have limitations, such as requiring both variables (existing and new users) to be continuous to identify user relationships. Furthermore, if sufficient purchase data for existing products, including purchases by existing and new users, is not available, the lack of data for correlation analysis can lead to reduced accuracy in calculating recommendation scores.

[0006] The purpose of the present invention is to solve the above problem, and to provide a method and device for conditional probability-based collaborative filtering using high-dimensional binary data capable of calculating a recommendation score for a customer who has no purchase history of a target product.

[0007] The purpose of the present invention is not limited to the purposes mentioned above, and other purposes not mentioned can be clearly understood from the description below.

[0008] A method for conditional probability-based collaborative filtering using high-dimensional binary data according to one aspect of the present invention for achieving the above-described object relates to a method performed by a computing device, comprising: a step of calculating a first value, which is the number of existing products purchased by the first customer and the second customer, a second value, which is the number of existing products purchased by the first customer and not purchased by the second customer, a third value, which is the number of existing products purchased by the second customer and not purchased by the first customer, and a fourth value, which is the number of existing products not purchased by the first customer and not purchased by the second customer, a conditional probability calculation step, which calculates a conditional probability that the second customer will not purchase when the first customer purchases by dividing the second value by the sum of the first value and the second value, and a conditional probability calculation step, which calculates a conditional probability that the second customer will not purchase when the first customer does not purchase by dividing the fourth value by the sum of the third value and the fourth value, and a conditional probability calculation step, which calculates a conditional probability that the second customer will not purchase when the first customer does not purchase by dividing the fourth value by the sum of the third value and the fourth value, for each combination of the first customer and the second customer. In this case, the step of assigning a conditional probability that a second customer will not purchase when a first customer purchases a product to a first non-purchase probability value, and assigning a conditional probability that a second customer will not purchase a product when the first customer does not purchase a product to a second non-purchase probability value, and the step of calculating a recommendation score by selectively adding up the cumulative sum of the first non-purchase probability value and the cumulative sum of the second non-purchase probability value assigned to each combination of the first customer and the second customer according to a preset indicator variable are included.

[0009] According to another aspect of the present invention, a device for collaborative filtering based on conditional probability using high-dimensional binary data comprises a memory for storing commands and executing the commands stored in the memory, using purchase data for existing products in which purchases by a first customer and a second customer are recorded, to calculate a first value, which is the number of existing products purchased by the first customer and the second customer, a second value, which is the number of existing products purchased by the first customer and not purchased by the second customer, a third value, which is the number of existing products purchased by the second customer and not purchased by the first customer, and a fourth value, which is the number of existing products not purchased by the first customer and the second customer, and to calculate a conditional probability that the second customer will not purchase when the first customer purchases by dividing the second value by the sum of the first value and the second value, and to calculate a conditional probability that the second customer will not purchase when the first customer does not purchase by dividing the fourth value by the sum of the third value and the fourth value, and to calculate a conditional probability that the second customer will not purchase when the first customer does not purchase, and to calculate a conditional probability that the first customer will not purchase when the first customer does not purchase when the first customer does not purchase, for each combination of the first customer and the second customer, if the first customer purchases the target product, the first customer A processor is included that assigns a conditional probability that a second customer will not purchase a product to a first non-purchase probability value when a purchase is made, assigns a conditional probability that a second customer will not purchase a product when the first customer does not purchase a product to a second non-purchase probability value when the first customer does not purchase a product, and calculates a recommendation score by selectively adding up the cumulative sum of the first non-purchase probability value and the cumulative sum of the second non-purchase probability value assigned to each combination of the first customer and the second customer according to a preset indicator variable.

[0010] According to the present invention, the performance of calculating a recommendation score for a customer without a purchase history of a target product is improved in a situation where purchase data is insufficient by using conditional probability.

[0011] The present invention assigns a purchase or non-purchase probability value based on a conditional probability calculated from purchase data according to whether a customer has purchased a target product, and comprehensively considers the purchase or non-purchase probability value based on whether other users have purchased the target product based on user input, thereby providing superior performance compared to existing recommendation score calculation methods.

[0012] The effects of the present invention are not limited to the effects mentioned above, and other effects not mentioned will be clearly understood by those skilled in the art from the description of the claims.

[0013] Figure 1 is an exemplary diagram for explaining purchase data used in embodiments of the present invention.

[0014] FIG. 2 is a flowchart of a method for conditional probability-based collaborative filtering using high-dimensional binary data according to one embodiment of the present invention.

[0015] FIG. 3 is a block diagram of a device for conditional probability-based collaborative filtering using high-dimensional binary data according to another embodiment of the present invention.

[0016] A method for conditional probability-based collaborative filtering using high-dimensional binary data according to one aspect of the present invention for achieving the above-described object relates to a method performed by a computing device, comprising: a step of calculating a first value, which is the number of existing products purchased by the first customer and the second customer, a second value, which is the number of existing products purchased by the first customer and not purchased by the second customer, a third value, which is the number of existing products purchased by the second customer and not purchased by the first customer, and a fourth value, which is the number of existing products not purchased by the first customer and not purchased by the second customer, a conditional probability calculation step, which calculates a conditional probability that the second customer will not purchase when the first customer purchases by dividing the second value by the sum of the first value and the second value, and a conditional probability calculation step, which calculates a conditional probability that the second customer will not purchase when the first customer does not purchase by dividing the fourth value by the sum of the third value and the fourth value, and a conditional probability calculation step, which calculates a conditional probability that the second customer will not purchase when the first customer does not purchase by dividing the fourth value by the sum of the third value and the fourth value, for each combination of the first customer and the second customer. In this case, the step of assigning a conditional probability that a second customer will not purchase when a first customer purchases a product to a first non-purchase probability value, and assigning a conditional probability that a second customer will not purchase a product when the first customer does not purchase a product to a second non-purchase probability value, and the step of calculating a recommendation score by selectively adding up the cumulative sum of the first non-purchase probability value and the cumulative sum of the second non-purchase probability value assigned to each combination of the first customer and the second customer according to a preset indicator variable are included.

[0017] According to another aspect of the present invention, a device for collaborative filtering based on conditional probability using high-dimensional binary data comprises a memory for storing commands and executing the commands stored in the memory, using purchase data for existing products in which purchases by a first customer and a second customer are recorded, to calculate a first value, which is the number of existing products purchased by the first customer and the second customer, a second value, which is the number of existing products purchased by the first customer and not purchased by the second customer, a third value, which is the number of existing products purchased by the second customer and not purchased by the first customer, and a fourth value, which is the number of existing products not purchased by the first customer and the second customer, and to calculate a conditional probability that the second customer will not purchase when the first customer purchases by dividing the second value by the sum of the first value and the second value, and to calculate a conditional probability that the second customer will not purchase when the first customer does not purchase by dividing the fourth value by the sum of the third value and the fourth value, and to calculate a conditional probability that the second customer will not purchase when the first customer does not purchase, and to calculate a conditional probability that the first customer will not purchase when the first customer does not purchase when the first customer does not purchase, for each combination of the first customer and the second customer, if the first customer purchases the target product, the first customer A processor is included that assigns a conditional probability that a second customer will not purchase a product to a first non-purchase probability value when a purchase is made, assigns a conditional probability that a second customer will not purchase a product when the first customer does not purchase a product to a second non-purchase probability value when the first customer does not purchase a product, and calculates a recommendation score by selectively adding up the cumulative sum of the first non-purchase probability value and the cumulative sum of the second non-purchase probability value assigned to each combination of the first customer and the second customer according to a preset indicator variable.

[0018] The advantages and features of the present invention, and the methods for achieving them, will become clearer with reference to the embodiments described in detail below together with the accompanying drawings. However, the present invention is not limited to the embodiments disclosed below, but may be implemented in various different forms. These embodiments are provided only to ensure that the disclosure of the present invention is complete and to fully inform those skilled in the art of the invention of the scope of the invention, and the present invention is defined only by the description of the claims. Meanwhile, the advantages and features of the present invention, and the methods for achieving them will become clearer with reference to the embodiments described in detail below together with the accompanying drawings. However, the present invention is not limited to the embodiments disclosed below, but may be implemented in various different forms. These embodiments are provided only to ensure that the disclosure of the present invention is complete and to fully inform those skilled in the art of the invention of the scope of the invention, and the present invention is defined only by the description of the claims. Meanwhile, the terminology used in this specification is for the purpose of describing the embodiments and is not intended to limit the present invention. In this specification, the singular includes the plural unless specifically stated otherwise.

[0019] The present invention relates to a technology for calculating a recommendation score, which represents the probability that a second customer will purchase a target product for which only the purchase by the first customer is recorded, by using a conditional probability calculated from purchase data for existing products for which the purchase by the first customer and the second customer is recorded.

[0020] In particular, the present invention is characterized by providing a recommendation score with excellent performance by selectively adding up the purchase probability value of a second customer when a first customer purchases a target product, the non-purchase probability value of the second customer when the first customer purchases the target product, the purchase probability value of the second customer when the first customer does not purchase the target product, and the non-purchase probability value of the second customer when the first customer does not purchase the target product, based on an indicator variable, and adding up the non-purchase probability value of the second customer with a negative sign to produce a recommendation score.

[0021] Hereinafter, a method and device for conditional probability-based collaborative filtering using high-dimensional binary data according to embodiments of the present invention will be described in detail with reference to the attached drawings.

[0022] The method for conditional probability-based collaborative filtering using high-dimensional binary data according to one embodiment of the present invention illustrated in FIG. 2 is implemented as computer commands that are loaded into a processor or memory of a computing device capable of data processing (e.g., a desktop, a laptop, etc.) and perform designated functions, and can be executed by a device (10) for conditional probability-based collaborative filtering using high-dimensional binary data according to another embodiment of the present invention illustrated in FIG. 3.

[0023] For convenience of explanation, the drawing symbols are matched for functionally identical contents and duplicate explanations are avoided.

[0024] A device (10) for conditional probability-based collaborative filtering using high-dimensional binary data may include an input / output interface (11), a communication interface (12), a memory (13), and a processor (14).

[0025] The input / output interface (11) may be connected to an input / output device for receiving data from a user or outputting a data processing result, and may input data from a user or output a data processing result through the input / output device.

[0026] For example, input / output devices can include mouse, keyboard, display, etc.

[0027] The communication interface (12) communicates with an external electronic device through a network and may be capable of receiving data from an external electronic device or outputting processed data.

[0028] The memory (13) may include a computer-readable recording medium.

[0029] Memory (13) may store an operating system and preset computer instructions.

[0030] The memory (13) can store purchase data input through the input / output interface (11) or communication interface (12).

[0031] Purchase data may be composed of rows corresponding to customers and rows corresponding to products, and may be binary data with each matrix element value having one of two values ​​(1, 0) indicating whether the customer corresponding to each row purchased the product corresponding to each column.

[0032] Referring to Figure 1, the first customer ( ) is an existing product ( ) and target products ( ) refers to a customer who has a value recorded indicating whether or not he or she has made a purchase, and a second customer ( ) is an existing product ( ) is recorded only for the purchase or non-purchase value and the target product ( ) refers to customers whose purchase history is not recorded.

[0033] Here, the fact that a purchase is not recorded does not mean that the purchase data does not actually exist. In some embodiments, multiple ( ) and multiple customers ( ) in a situation where purchase data for a product exists, the user may arbitrarily select multiple ( )'s customers With the first customer of the name Distinguish them as second-party customers and multiple ( ) products The dog's existing products and You can calculate a recommendation score using purchase data divided by target product.

[0034] For example, plural( ) and multiple customers ( ) purchase data for the product of the first customer ( ) and second customer ( )'s existing products( ) matrix element value indicating whether to purchase or not ( , ) and the first purchase data and the first customer ( ) target product( ) matrix element value indicating whether to purchase or not ( ) can be used to calculate the recommendation score by dividing it into secondary purchase data.

[0035] The processor (14) may be capable of performing basic arithmetic, logic, and input / output operations, thereby executing computer instructions stored in the memory (13).

[0036] The processor (14) can use the first purchase data for existing products in which the purchases of the first and second customers are recorded by executing computer commands to calculate a first value, which is the number of existing products purchased by the first customer and the second customer, a second value, which is the number of existing products purchased by the first customer and not purchased by the second customer, a third value, which is the number of existing products purchased by the second customer and not purchased by the first customer, and a fourth value, which is the number of existing products not purchased by the first customer and not purchased by the second customer, for each combination of the first and second customers (S110).

[0037] The processor (14) counts the cases in which the matrix element values ​​for the first customer and the second customer in the column corresponding to each reference product for each combination of the first customers for each second customer in the first purchase data are all values ​​corresponding to purchase (e.g., 1), thereby calculating the first value (which is the number of existing products purchased by the first customer and the second customer) ) is calculated, and the matrix element value for the first customer in the column corresponding to each standard product is a value corresponding to purchase (e.g., 1) and the matrix element value for the second customer is a value corresponding to non-purchase (e.g., 0), and the second value (which is the number of existing products purchased by the first customer and not purchased by the second customer) is counted. ) is calculated, and the matrix element value for the first customer in the column corresponding to each standard product is a value corresponding to non-purchase (e.g., 0) and the matrix element value for the second customer is a value corresponding to purchase (e.g., 1), and the third value (which is the number of existing products that the first customer did not purchase and the second customer purchased) is counted. ) is calculated, and the fourth value (which is the number of existing products not purchased by the first and second customers) is counted by counting the cases where the matrix component values ​​for the first customer and the second customer in the column corresponding to each standard product are values ​​corresponding to non-purchase (e.g., 0). ) may be produced (S110).

[0038] And the processor (14) has a first value ( ), the second value ( ), the third value ( ), the fourth value ( ) can be used to calculate the conditional probability that a second customer will purchase an existing product when a first customer purchases it, the conditional probability that a second customer will not purchase when a first customer purchases it, the conditional probability that a second customer will purchase when a first customer does not purchase it, and the conditional probability that a second customer will not purchase when a first customer does not purchase it (S20).

[0039] The processor (14) has a first value ( ) and the second value ( ) as the second value ( ) and the conditional probability that the second customer (a) does not purchase when the first customer (i) purchases ( ) and the third value ( ) and the fourth value ( ) as the fourth value ( ) and the conditional probability that the second customer (a) will not purchase when the first customer (i) does not purchase ( ) can be produced (S20).

[0040] And, the processor (14) has a first value ( ) and the second value ( ) as the sum of the first value ( ) and the conditional probability that the second customer (a) will purchase when the first customer (i) purchases ( ) and the third value ( ) and the fourth value ( ) as the third value ( ) and the conditional probability that the second customer (a) will not purchase when the first customer (i) does not purchase ( ) can be produced (S20).

[0041] The processor (14) calculates the conditional probability that the second customer (a) will not purchase when the first customer (i) purchases, according to the following mathematical formula: ), the conditional probability that the second customer (a) will not purchase when the first customer (i) does not purchase ( ), the conditional probability that the second customer (a) will purchase when the first customer (i) purchases ( ), the conditional probability that the second customer (a) will not purchase when the first customer (i) does not purchase ( ) can be produced (S20).

[0042]

[0043]

[0044]

[0045]

[0046] Thereafter, the processor (14) can assign conditional probabilities or initial values ​​to purchase probability values ​​and non-purchase probability values ​​depending on whether the first customer purchased the target product for each combination of the first customer and the second customer (S130).

[0047] The processor (14) may assign a conditional probability that the second customer will not purchase the target product when the first customer purchases the target product to the first non-purchase probability value for each combination of the first customer and the second customer, and may assign a preset initial value (e.g., 0) to the first non-purchase probability value when the first customer does not purchase the target product (S130).

[0048] The processor (14) may assign a preset initial value (e.g., 0) to the second non-purchase probability value when the first customer purchases the target product for each combination of the first customer and the second customer, and may assign a conditional probability that the second customer will not purchase when the first customer does not purchase the target product to the second non-purchase probability value when the first customer does not purchase the target product (S130).

[0049] The processor (14) may assign a conditional probability that the second customer will purchase the target product when the first customer purchases the target product to the first purchase probability value for each combination of the first customer and the second customer, and may assign a preset initial value (e.g., 0) to the first purchase probability value when the first customer does not purchase the target product (S130).

[0050] The processor (14) may assign a preset initial value (e.g., 0) to the second purchase probability value when the first customer purchases the target product for each combination of the first customer and the second customer, and may assign a conditional probability that the second customer will purchase the target product when the first customer does not purchase the target product to the second purchase probability value when the first customer does not purchase the target product (S130).

[0051] The processor (14) calculates a first non-purchase probability value (k) based on whether the first customer (i) purchases the target product (k) for each combination of the first customer (i) and the second customer (a) according to the following mathematical formula. ), the second non-purchase probability value ( ), first purchase probability value ( ), second purchase probability value ( ) may be given a conditional probability or initial value (S130).

[0052]

[0053]

[0054]

[0055]

[0056] Here, is a binary variable indicating whether the first customer (i) purchases the target product (k). 1 means purchase, and 0 means non-purchase.

[0057] Thereafter, the processor (14) can calculate a recommendation score indicating the purchase probability of the second customer for the target product by adding a negative sign to the cumulative sum of the first non-purchase probability value and the cumulative sum of the second non-purchase probability value given for each combination of the first customer and the second customer and selectively adding them according to a preset indicator variable (S140).

[0058] The processor (14) multiplies a preset indicator variable for each of the first purchase possibility value, the second purchase possibility value, the first non-purchase possibility value, and the second non-purchase possibility value, and optionally adds them up, but adds a positive sign to the first purchase possibility value and the second purchase possibility value, and a negative sign to the first non-purchase possibility value and the second non-purchase possibility value, thereby calculating a recommendation score indicating a purchase probability of a second customer for a target product (S140).

[0059] The processor (14) calculates a recommendation score indicating the purchase probability of the second customer (a) for the target product (k) according to the following mathematical formula. can be produced (S140).

[0060]

[0061] Here, is the first indicator variable to determine whether to apply the first purchase probability value, is the second indicator variable to determine whether the first non-purchase probability value is applied, is the third indicator variable to determine whether to apply the second purchase probability value, is a fourth indicator variable for determining whether to apply the second non-purchase possibility value, and can be set to either 0 or 1 by the user.

[0062] The present invention selectively uses purchase probability values ​​and non-purchase probability values ​​calculated from conditional probabilities to calculate recommendation scores according to indicator variables input by a user, and can calculate recommendation scores for 16 cases depending on whether a first customer and a second customer make a purchase.

[0063] Specifically, according to the present invention, the following 16 cases of recommendation scores can be calculated depending on the combination of indicator variables.

[0064] 1. 1111

[0065] The recommendation score is calculated by subtracting the cumulative sum of the first non-purchase probability from the cumulative sum of the first purchase probability, adding the cumulative sum of the second purchase probability, and subtracting the cumulative sum of the second non-purchase probability. This means that when the first customer purchases the target product, the conditional probability that the second customer will purchase if the first customer purchases the target product, the conditional probability that the second customer will purchase if the first customer does not purchase if the first customer purchases the target product, the conditional probability that the second customer will not purchase if the first customer purchases the target product, and the conditional probability that the second customer will purchase if the first customer does not purchase the target product, are all considered.

[0066] 2. 1110

[0067] This means that the recommendation score is calculated by subtracting the cumulative sum of the first non-purchase probability from the cumulative sum of the first purchase probability and adding the cumulative sum of the second purchase probability. This means that if the first customer purchases the target product, the conditional probability that the second customer will purchase it if the first customer purchases it, if the first customer purchases the target product, the conditional probability that the second customer will not purchase it if the first customer purchases it, and if the first customer does not purchase the target product, the conditional probability that the second customer will not purchase it if the first customer purchases it are considered.

[0068] 3. 1101

[0069] The recommendation score is calculated by subtracting the cumulative sum of the first non-purchase probability value from the cumulative sum of the first purchase probability value and then subtracting the cumulative sum of the second non-purchase probability value. This means that if the first customer purchases the target product, the conditional probability that the second customer will purchase it if the first customer purchases it, the conditional probability that the second customer will not purchase it if the first customer purchases it, and the conditional probability that the second customer will not purchase it if the first customer does not purchase the target product are considered.

[0070] 4. 1100

[0071] This means that the recommendation score is calculated by subtracting the cumulative sum of the first non-purchase probability values ​​from the cumulative sum of the first purchase probability values. This means that when the first customer purchases the target product, the conditional probability that the second customer will purchase it if the first customer purchases it, and the conditional probability that the second customer will not purchase it if the first customer purchases it, are considered.

[0072] 5. 1011

[0073] This means that the recommendation score is calculated by adding the cumulative sum of the second purchase probability value to the cumulative sum of the first purchase probability value and subtracting the cumulative sum of the second non-purchase probability value. This means that if the first customer purchases the target product, the conditional probability that the second customer will purchase if the first customer purchases it, if the first customer does not purchase the target product, the conditional probability that the second customer will purchase if the first customer does not purchase it, and if the first customer does not purchase the target product, the conditional probability that the second customer will not purchase if the first customer does not purchase it are considered.

[0074] 6. 1010

[0075] This means that the recommendation score is calculated by adding the cumulative sum of the second purchase probability values ​​to the cumulative sum of the first purchase probability values. This means that if the first customer purchases the target product, the conditional probability that the second customer will purchase it if the first customer purchases it, and if the first customer does not purchase the target product, the conditional probability that the second customer will purchase it if the first customer does not purchase it, are considered.

[0076] 7. 1001

[0077] This means that the recommendation score is calculated by subtracting the cumulative sum of the second non-purchase probability values ​​from the cumulative sum of the first purchase probability values. This means that if the first customer purchases the target product, the conditional probability that the second customer will purchase it if the first customer purchases it, and if the first customer does not purchase the target product, the conditional probability that the second customer will not purchase it if the first customer does not purchase it, are considered.

[0078] 8. 1000

[0079] This means that the recommendation score is calculated as the cumulative sum of the first purchase probability values, and only the conditional probability that the second customer will purchase the product when the first customer purchases the product is considered.

[0080] 9. 0111

[0081] This means that the recommendation score is calculated by subtracting the cumulative sum of the first non-purchase probability value from the preset initial value (0), adding the cumulative sum of the second purchase probability value, and subtracting the cumulative sum of the second non-purchase probability value. This means that when the first customer purchases the target product, the conditional probability that the second customer will not purchase when the first customer purchases it, when the first customer does not purchase the target product, the conditional probability that the second customer will purchase when the first customer does not purchase it, and when the first customer does not purchase the target product, the conditional probability that the second customer will not purchase when the first customer does not purchase it are considered.

[0082] 10. 0110

[0083] This means that the recommendation score is calculated by subtracting the cumulative sum of the first non-purchase probability value from the preset initial value (0) and adding the cumulative sum of the second purchase probability value. This means that when the first customer purchases the target product, the conditional probability that the second customer will not purchase the product when the first customer purchases it, and when the first customer does not purchase the target product, the conditional probability that the second customer will purchase the product when the first customer does not purchase it are considered.

[0084] 11. 0101

[0085] This means that the recommendation score is calculated by subtracting the cumulative sum of the first non-purchase probability value from the preset initial value (0) and then subtracting the cumulative sum of the second non-purchase probability value. This means that when the first customer purchases the target product, the conditional probability that the second customer will not purchase when the first customer purchases it, and when the first customer does not purchase the target product, the conditional probability that the second customer will not purchase when the first customer does not purchase it are considered.

[0086] 12. 0100

[0087] This means that the recommendation score is calculated by subtracting the cumulative sum of the first non-purchase probability value from the preset initial value (0). This means that when the first customer purchases the target product, only the conditional probability that the second customer will not purchase the product is considered.

[0088] 13. 0011

[0089] This means that the recommendation score is calculated by subtracting the cumulative sum of the second non-purchase probability values ​​from the cumulative sum of the second purchase probability values. This means that if the first customer does not purchase the target product, the conditional probability that the second customer will purchase if the first customer does not purchase it, and if the first customer does not purchase the target product, the conditional probability that the second customer will not purchase it if the first customer does not purchase it, are considered.

[0090] 14. 0010

[0091] This means that the recommendation score is calculated as the cumulative sum of the second purchase probability values, and if the first customer does not purchase the target product, only the conditional probability that the second customer will purchase the product if the first customer does not purchase it is considered.

[0092] 15. 0001

[0093] This means that the recommendation score is calculated by subtracting the cumulative sum of the second non-purchase probability value from the preset initial value (0). This means that when the first customer does not purchase the target product, only the conditional probability that the second customer will not purchase the product when the first customer does not purchase it is considered.

[0094] 16. 0000

[0095] This means that the recommendation score is calculated with a preset initial value (0), and the conditional probability that the second customer will purchase the target product if the first customer purchases it, the conditional probability that the second customer will not purchase the target product if the first customer purchases it, the conditional probability that the second customer will purchase the target product if the first customer does not purchase it, and the conditional probability that the second customer will not purchase the target product if the first customer does not purchase it, are all not considered.

[0096] That is, the present invention can produce 16 types of recommendation scores that consider all combinations of purchase or non-purchase possibilities according to indicator variables.

[0097]

[0098] In summary, according to the present invention, a recommendation score can be calculated even for customers who have no purchase history for the target product by using conditional probabilities calculated from purchase data for existing products.

[0099] In particular, the present invention has the advantage of improving recommendation performance by enabling the calculation of 16 recommendation scores by considering both the customer's purchase and non-purchase possibilities for each product, thereby checking the performance of each recommendation score and deriving an effective recommendation score for each data.

[0100] Below, an experiment was conducted to compare the performance of calculating a recommendation score according to an embodiment of the present invention with that of an existing recommendation score calculation technology.

[0101] In the conventional technology, random forest (RF), logistic regression using principal components (PCA+LR), and Pearson correlation (Pearson)-based recommendation score calculation technology are used, and in the embodiment of the present invention, indicator variables The recommendation score was calculated for each combination.

[0102] Using a high-dimensional grocery data set that records customers' purchases of grocery products, the first and second customers (new customers) are randomly selected to obtain the first purchase data of the first and second customers for the existing product, the second purchase data of the first customer for the target product, and the third purchase data of the second customer for the target product. Then, a recommendation score is calculated from the first and second purchase data, and based on the third purchase data, the indicator variables are compared with the existing technologies. The performance of calculating recommendation scores of embodiments of the present invention according to the combination was confirmed.

[0103] As a performance indicator, Top-n was used, which indicates whether the recommended products were actually purchased when n products were output in order of highest recommendation score.

[0104] The table below shows the accuracy of calculating the recommendation score according to the conventional technologies and the present invention for 100 existing customers.

[0105] Accuracy RFPCA+LRPearson Invention (1,0,1,1), (1,0,0,1) Top-10.19400.23300.17100.2700 Top-20.16000.17900.13950.2015 Top-30.14030.15100.12530.1657 Top-40.12530.13050.11130.1338 Top-50.11200.12100.09700. 1226Top-60.10100.10670.08800.1067Top-70.09170.09300.07900.0961Top-80.08230. 08230.07300.0863Top-90.07540.07320.06670.0789Top-100.07070.06590.06130.0723

[0106] Referring to the above table, in the embodiments of the present invention, when the recommendation score is calculated as (1,0,1,1) with indicator variables as (1,0,0,1), it can be confirmed that it shows higher accuracy than the prior art in the entire range from Top-1 to Top-10. Specifically, PCA+LR, which has the highest accuracy among the prior art techniques, has a Top-1 accuracy of 0.2330, which means that when recommendations are performed 100 times for each item, the number of times it recommends a product that was actually purchased is about 23, whereas when the indicator variables are (1,0,1,1) or (1,0,0,1), the Top-1 accuracy of the present invention is 0.2700, which means that when recommendations are performed 100 times for each item, the number of times it recommends a product that was actually purchased is 27, which shows superior performance than the prior art.

[0107] The table below shows the accuracy of calculating the recommendation score using 20 existing customers according to conventional technologies and the present invention.

[0108] Accuracy RFPCA+LRPearson Invention (0,0,0,1), (0,0,1,1) Top-10.16400.20000.07700.2700 Top-20.14800.16750.07950.1815 Top-30.13770.14000.08100.1657 Top-40.12680.11200.07680.1345 Top-50.11220.09080.06900. 1148Top-60.09530.07630.05920.1020Top-70.08260.06590.05160.0893Top-80.07260. 05790.04590.0784Top-90.06500.05180.04160.0698Top-100.05850.04700.03960.0634

[0109] Referring to Table 2 and Table 1, we can see that the Top-1 accuracy decreases from 0.1800 to 0.1640 for RF, from 0.2040 to 0.2000 for PCA+LR, and from 0.1370 to 0.0770 for Pearson, compared to when the number of existing customers is 100. In comparison, in the embodiment of the present invention where the indicator variables are (0,0,0,1) or (0,0,1,1), it can be confirmed that the Top-1 accuracy is 0.2700, which is the same as Table 1 with 100 existing customers. In addition, it can be confirmed that the accuracy of the present invention is lower from Top-2 to Top-10 than Table 1 with 100 existing customers, but while the accuracy of the Top-10 of the conventional technologies is lower by about 0.02 than in Table 1, the accuracy of the present invention is lower by about 0.01, which can be confirmed to be less of a drop in accuracy than the conventional technologies.

[0110] That is, it can be confirmed that the present invention alleviates the cold start problem when the number of existing customers is small compared to conventional technologies and purchase data is not sufficiently secured.

[0111] Therefore, according to the present invention, product recommendations can be made with superior performance compared to conventional technologies even in situations where purchase data for existing products is not sufficiently secured.

[0112]

[0113] In addition, an experiment was conducted to compare the performance of the present invention with the existing conditional probability-based recommendation score calculation technology (application number 10-2019-0038171) using high-dimensional product data.

[0114] Here, high-dimensional product data refers to data in which the number of products is significantly greater than the number of customers.

[0115] High-dimensional product data has a value of 1 for purchase and 0 for non-purchase depending on whether each customer purchased the product. Referring to Figure 1, the structure of the high-dimensional product data used in the experiment is as shown in Table 3 below.

[0116] Number of purchases of target products by customer area Number of purchases by customer area Purchase ratio (%) 1st customer × existing product 2020 778 919.06 1st customer × target product 2050 11611.60 2nd customer × existing product 12120 7923 436.87 2nd customer × target product 12150 204633.82

[0117] Table 4 below shows the accuracy of each recommendation score produced when applying the above high-dimensional product data to the existing conditional probability-based recommendation score production technology.

[0118] Accuracy Top-10.83470.82650.39670.3884Top-20.86780.85950.50000.5000Top-30.85680.8 5400.52620.5234Top-40.83680.82850.53100.5413Top-50.76030.75210.53060.5372 Top-60.69280.69600.50410.5193Top-70.64110.66000.48640.4912Top-80.60020.62 090.47730.4773Top-90.56840.89600.46740.4683Top-100.55950.58020.45540.4620

[0119] Referring to Table 4, the existing conditional probability-based recommendation score calculation technology only produces four types of limited prediction values ​​(recommendation scores), and the prediction values ​​(recommendation scores) that show an accuracy exceeding 0.5 in the Top-1 to Top-10 overall are and It can be confirmed that there are only two. Table 5 below shows the accuracy of each recommendation score according to the combination of indicator variables when the high-dimensional product data above is applied to the recommendation score calculation technique according to the present invention.

[0120] 1234567891011Top-10.8260.3050.8260.6360.8340.8260.8340.8340.4950.2470.826Top-20.8590.2930.8590.6520.8630.8590.8630.8670.4580.2230.859Top-30.8530.3050.8510.6280.8620.8530.8620.8560.4600.2530.853Top-40.8280.2910.8360.6050.8380.8280.8380.8360.4310.2430.828Top-50.7520.2640.7600.5570.7700.7520.7630.7600.3850.2160.752Top-60.6950.2720.6900.5170.7020.6950.6960.6920.3730.2280.695Top-70.6590.2730.6450.4980.6420.6590.6390.6410.3700.2330.659Top-80.6200.2870.6070.4700.6010.6200.6020.6000.3710.2530.620Top-90.5950.2960.5770.4600.5670.5950.5630.5680.3640.2640.595Top-100.5800.3140.5570.4520.5510.5800.5550.5590.3660.2790.5801213141516Top-10.2470.8260.2470.8340.545Top-20.2230.8260.2230.8630.396Top-30.2530.8260.2530.8620.396Top-40.2430.8010.2430.8380.351Top-50.2160.7370.2160.7610.343Top-60.2240.6870.2310.7030.344Top-70.2320.6300.2430.6390.362Top-80.2440.5890.2640.5940.341Top-90.2540.5610.2750.5630.315Top-100.2660.5350.2860.5380.321

[0121] Referring to Table 5, the present invention can produce 16 types of prediction values ​​(recommendation scores). While existing conditional probability-based recommendation score production techniques can produce 4 types of recommendation scores by considering only a limited number of purchase and non-purchase possibilities, the present invention can produce 16 types of recommendation scores by considering all purchase or non-purchase possibilities based on indicator variables.

[0122] In addition, referring to Table 5, it can be confirmed that the present invention has a high accuracy exceeding 0.5 in the Top-1 to Top-10 overall in 1, 3, 5, 6, 7, 8, 11, 13, and 15.

[0123] While the existing conditional probability-based recommendation score calculation technology has high accuracy in only two cases, the present invention has high accuracy in nine cases, confirming that the present invention can provide a higher performance recommendation score than the existing conditional probability-based recommendation score calculation technology.

[0124] Although the specific accuracy figures may vary depending on the data applied, existing technologies can provide up to four predictions (recommendation scores), and existing technologies corresponds to when the indicator variable of the present invention is (1,0,0,0). Since the indicator variable of the present invention corresponds to (1,0,1,0), it can be inferred that the present invention can always provide a recommendation score that provides accuracy equivalent to that of existing technologies and is more diverse than existing technologies.

[0125] That is, the present invention can improve recommendation performance by providing a variety of recommendation scores with high accuracy compared to existing technologies by considering all cases according to combinations of all purchase and non-purchase possibilities.

[0126] Meanwhile, since the computer program instructions in which the steps of the flowchart are implemented can be stored in a computer-readable memory, the functions described in the steps of the flowchart can also be produced as a manufactured article that includes a command means for performing them.

[0127] Those skilled in the art will appreciate that the present invention can be implemented in other specific forms without altering its technical spirit or essential characteristics. Therefore, the embodiments described above should be understood as illustrative in all respects and not restrictive. The scope of the present invention is indicated by the claims below rather than the detailed description above, and all changes or modifications derived from the claims and their equivalents should be construed as being included within the scope of the present invention.

[0128] The present invention relates to a technology for calculating a recommendation score indicating the purchase probability of a product, and more specifically, to a technology for calculating a recommendation score for a product for which there is no purchase record.

Claims

1. Regarding a method performed by a computing device, A step of calculating a first value, which is the number of existing products purchased by the first customer and the second customer, a second value, which is the number of existing products purchased by the first customer and not purchased by the second customer, a third value, which is the number of existing products purchased by the second customer and not purchased by the first customer, and a fourth value, which is the number of existing products not purchased by the first customer and not purchased by the second customer, by using purchase data for existing products for which purchase by the first customer and the second customer are recorded, for each combination of the first customer and the second customer; A conditional probability calculation step of calculating a conditional probability that a second customer will not purchase when a first customer purchases by dividing the second value by the sum of the first value and the second value, and calculating a conditional probability that a second customer will not purchase when a first customer does not purchase by dividing the fourth value by the sum of the third value and the fourth value; A step of assigning a conditional probability that the second customer will not purchase the target product when the first customer purchases the target product to a first non-purchase probability value for each combination of the first customer and the second customer, and assigning a conditional probability that the second customer will not purchase the target product when the first customer does not purchase the target product to a second non-purchase probability value for each combination of the first customer and the second customer; and A method for conditional probability-based collaborative filtering using high-dimensional binary data, comprising: a step of calculating a recommendation score by selectively adding the cumulative sum of the first non-purchase probability value and the cumulative sum of the second non-purchase probability value assigned to each combination of the first customer and the second customer according to a preset indicator variable.

2. In paragraph 1, The steps for calculating the above recommended score are Adding a negative sign to the cumulative sum of the first non-purchase probability value and the cumulative sum of the second non-purchase probability value and selectively adding them according to a preset indicator variable A method for conditional probability-based collaborative filtering using high-dimensional binary data.

3. In paragraph 1, The above conditional probability calculation step is Divide the first value by the sum of the first value and the second value to further calculate the conditional probability that the second customer will purchase when the first customer purchases, and divide the sum of the third value and the fourth value by the third value to further calculate the conditional probability that the second customer will purchase when the first customer does not purchase. A method for conditional probability-based collaborative filtering using high-dimensional binary data.

4. In paragraph 3, The above-mentioned granting steps are In the case where the first customer purchases the target product for each combination of the first customer and the second customer, the conditional probability that the second customer will purchase the product if the first customer purchases the product is added to the first purchase probability value, and in the case where the first customer does not purchase the target product, the conditional probability that the second customer will purchase the product if the first customer does not purchase the product is added to the second purchase probability value. A method for conditional probability-based collaborative filtering using high-dimensional binary data.

5. In paragraph 4, The steps for calculating the above recommended score are The recommendation score is calculated by multiplying the first purchase possibility value, the second purchase possibility value, the first non-purchase possibility value, and the second non-purchase possibility value by a preset indicator variable and optionally adding them, while adding a positive sign to the first purchase possibility value and the second purchase possibility value, and a negative sign to the first non-purchase possibility value and the second non-purchase possibility value. A method for conditional probability-based collaborative filtering using high-dimensional binary data.

6. Memory for storing commands; and By executing the commands stored in the above memory, the purchase data for existing products in which the purchases of the first and second customers are recorded is used to calculate, for each combination of the first customer and the second customer, a first value which is the number of existing products purchased by the first customer and the second customer, a second value which is the number of existing products purchased by the first customer and not purchased by the second customer, a third value which is the number of existing products purchased by the second customer and not purchased by the first customer, and a fourth value which is the number of existing products not purchased by the first and the second customer, and the second value is divided by the sum of the first value and the second value to calculate the conditional probability that the second customer will not purchase when the first customer purchases, and the fourth value is divided by the sum of the third value and the fourth value to calculate the conditional probability that the second customer will not purchase when the first customer does not purchase, and for each combination of the first customer and the second customer, if the first customer purchases the target product, the conditional probability that the second customer will not purchase when the first customer purchases is the first non-purchase probability value. A device for collaborative filtering based on conditional probability using high-dimensional binary data, comprising: a processor that assigns a conditional probability that a second customer will not purchase a target product when the first customer does not purchase the target product to a second non-purchase probability value, and selectively adds the cumulative sum of the first non-purchase probability value and the cumulative sum of the second non-purchase probability value assigned to each combination of the first customer and the second customer according to a preset indicator variable to produce a recommendation score; 7. In paragraph 6, The above processor Selectively adding the cumulative sum of the above first non-purchase probability value and the negative sign of the probability value of the above second non-purchase probability value according to a preset indicator variable. A device for conditional probability-based collaborative filtering using high-dimensional binary data.

8. In paragraph 6 The above processor Divide the first value by the sum of the first value and the second value to further calculate the conditional probability that the second customer will purchase when the first customer purchases, and divide the sum of the third value and the fourth value by the third value to further calculate the conditional probability that the second customer will purchase when the first customer does not purchase. A device for conditional probability-based collaborative filtering using high-dimensional binary data.

9. In paragraph 8, The above processor In the case where the first customer purchases the target product for each combination of the first customer and the second customer, the conditional probability that the second customer will purchase the product if the first customer purchases the product is added to the first purchase probability value, and in the case where the first customer does not purchase the target product, the conditional probability that the second customer will purchase the product if the first customer does not purchase the product is added to the second purchase probability value. A device for conditional probability-based collaborative filtering using high-dimensional binary data.

10. In paragraph 9, The above processor The recommendation score is calculated by multiplying the first purchase possibility value, the second purchase possibility value, the first non-purchase possibility value, and the second non-purchase possibility value by a preset indicator variable and optionally adding them, while adding a positive sign to the first purchase possibility value and the second purchase possibility value, and a negative sign to the first non-purchase possibility value and the second non-purchase possibility value. A device for conditional probability-based collaborative filtering using high-dimensional binary data.

Images