KR102142126B1 - Hierarchical Category Cluster Based Shopping Basket Associated Recommendation Method - Google Patents

Abstract

The present invention relates to a shopping cart product recommendation method based on a hierarchical category cluster, wherein the number of products belonging to each hierarchical category for each hierarchical category cluster is measured, and a category hierarchy in which the measured number of products is equal to or greater than a preset reference value is determined. Steps, applying a machine learning method for each determined category hierarchy, calculating support and reliability of products belonging to each category hierarchy, and determining product pairs having a high mutual purchase relationship based on the calculated support and reliability information It characterized in that it includes.
According to the present invention, it is expected that not only the convenience of product registration by the seller, the convenience of purchasing the product by the buyer, but also the increase in sales of the product trading shopping mall can be expected, thereby contributing to the activation and development of the e-commerce industry.

Description

Hierarchical Category Cluster Based Shopping Basket Associated Recommendation Method

The present invention induces to classify the products of sellers who want to sell products into the optimal category within the hierarchical category, and based on this hierarchical category, the products with high possibility of purchase together with the products in the shopping cart by the user are selected. It's about how to make recommendations efficiently.

With the development of the Internet and mobile communication, both B2B and B2C transactions are increasingly being used for online shopping malls.

Since online shopping malls handle a large number of products, these products are classified into hierarchical categories so that users can register products and search for products based on categories.

First of all, from a product seller's point of view, a seller who wants to sell a product accesses the online shopping mall site and manually registers the category of the product he wants to sell, or finds and registers a suitable category by referring to the category classification criteria of the online shopping mall site. It has a shape. However, when sellers do not understand the category classification system well, product categories are frequently registered incorrectly or overlapped with a category different from a previously registered category.

In addition, from the viewpoint of product buyers, when purchasing various products in a shopping mall, it is convenient to put the desired products in the shopping cart and then make payments at once, so the shopping cart function is frequently used.

Finally, from the viewpoint of a shopping mall site, a product recommendation service is provided to increase sales while providing convenience to users by providing an optimal purchasing decision environment for customers. Such a product recommendation service is a service that encourages the customer to purchase a recommended product by recommending an additional purchase product suitable for the customer in relation to the product when a customer purchases a specific product or puts it in a shopping cart.

In general, such a product recommendation service is a method of recommending products that a user who purchases a specific product a lot at the same time, and is most often used in many shopping malls or product transaction sites.

However, the conventional product recommendation service recommends products in terms of product versus product, that is, in units of individual products, so the difference in probability is large depending on the sales quantity of the product. Disadvantages arise.

In particular, a shopping mall that handles various product lines has a hierarchical category structure in order to effectively classify a number of products, but the conventional product recommendation service does not reflect the characteristics of the hierarchical category at all.

Accordingly, there is a demand for a method of enabling sellers to easily recognize and register their own product categories and to increase the success rate of product recommendation by using the characteristics of hierarchical categories.

1. Korean Patent Registration No. 00801662 (Name of invention: Product recommendation management system and its recommendation method) 2. Korean Patent Publication No. 2017-0119653 (Name of invention: Product purchase recommendation method)

Accordingly, the present invention was created in view of the above circumstances, and the main object of the present invention is to improve the accuracy of product category classification by effectively classifying hierarchical categories for specific products based on machine learning, and By providing a product recommendation service, a more suitable product is recommended to the buyer, so that the convenience of purchasing the product and the sales improvement of the shopping mall company can be obtained at the same time.

According to an aspect of the present invention for achieving the above object, in a method for providing a shopping cart product recommendation service based on a hierarchical category cluster in a product transaction server that is connected to a seller terminal and a buyer terminal through a network to mediate product transactions, A product registration step in which a product registration process is performed for a product that is requested to be registered from a seller terminal in the product transaction server, and a product for determining a shopping cart recommended product for the purchaser terminal using category information for the product previously registered in the product transaction server And a recommendation step, and the product registration step includes a first step of receiving product name text information of a product to be registered from the seller terminal in a product transaction server, and a noun among product name texts by analyzing the morpheme of the input product name text. And a second step of extracting English words, a third step of extracting category information having high correlation with the extracted nouns and English words by driving a machine learning-based category prediction model, transmitting the extracted category information to a seller terminal, and And a fourth step of determining the category information selected through the seller terminal as the category of the product, and the product recommendation step is the number of products belonging to each hierarchical category for each hierarchical category cluster in the product transaction server. The eleventh step of measuring a category hierarchy in which the measured number of products is greater than or equal to a preset reference value, and a twelfth step of calculating the support and reliability of products belonging to each category hierarchy by applying a machine learning method for each determined category hierarchy. And, when a buyer puts a specific product in the shopping cart, based on the calculated support and reliability information, a recommended product with a high probability of purchasing the product in the shopping cart is determined, but the recommended product belongs to a different category from the product in the shopping cart. There is provided a shopping cart product recommendation method based on a hierarchical category cluster, comprising: a thirteenth step of determining a product.

delete

delete

delete

delete

In addition, the machine learning-based category prediction model quantifies product name text data input in the process of registering a product in a product transaction server by a seller, and information on the category, number of occurrences, and score for each word from the machine learning process. May be stored as predictive model information.

According to the present invention, it is expected that not only the convenience of product registration by the seller, the convenience of product purchase by the buyer, but also the increase in sales of the product trading shopping mall can be expected, thereby contributing to the activation and development of the e-commerce industry.

1 is an overall configuration diagram of a shopping cart product recommendation system based on a hierarchical category cluster.
2 is a block diagram showing a detailed configuration of the product transaction server of FIG. 1.
3 shows an example of a method for analyzing a morpheme of a Korean sentence.
4 is a flowchart illustrating a process of performing a method for classifying and recommending product categories based on a hierarchical category cluster.
5 is a table showing the result of calculating transaction-based support.
6 is a table showing the number of items by category.
7 is a flowchart in which data training and analysis are performed in an association analysis method.
8 is a table showing test data comparison results tested on actual data.
9 shows a shopping cart product recommendation result when the number of items is set to 500 and the support level is set to 0.005.
10 shows the result of recommending shopping cart products when the number of items is set to 100 and the support level is set to 0.01.

Since the description of the present invention is merely an embodiment for structural or functional description, the scope of the present invention should not be construed as being limited by the embodiments described in the text. That is, since the embodiments can be variously changed and have various forms, the scope of the present invention should be understood to include equivalents capable of realizing the technical idea. In addition, since the object or effect presented in the present invention does not mean that a specific embodiment should include all or only such effects, the scope of the present invention should not be understood as being limited thereby.

Hereinafter, a system and method for recommending shopping cart products based on a hierarchical category cluster according to the present invention will be described with reference to the accompanying drawings.

1 is an overall configuration diagram of a shopping cart product recommendation system based on a hierarchical category cluster, and FIG. 2 is a block diagram showing a detailed configuration of the product transaction server of FIG. 1.

As shown in FIG. 1, a shopping cart product recommendation system based on a hierarchical category cluster according to the present invention includes a seller terminal 10, a buyer terminal 20, and a product transaction server 30.

The seller terminal 10 is a terminal provided in a sales company that wants to sell products through the product transaction server 30, and connects to the product transaction server 30 to register products to be sold by the company.

The purchaser terminal 30 is a terminal used by a purchaser who wants to purchase a product through the product transaction server 30 and includes various terminals capable of communication such as a PC, a laptop computer, a tablet PC, and a smartphone.

The product transaction server 30 is a computer provided in a company that is connected to a plurality of seller terminals 10 and buyer terminals 30 through a wired or wireless network to mediate product transactions. As shown in FIG. 2, the product transaction server 30 may largely include a category classifier 100, a shopping cart product recommender 200, a database unit 300, and a communication unit 400. In the block diagram of FIG. 2, only the main configurations related to the present invention are disclosed.

The communication unit 400 is connected to a plurality of seller terminals 10 and buyer terminals 20 through a network to exchange data.

The category classifier 100 includes a text preprocessor 110 and a learning-based category predictor 120.

The text preprocessor 110 analyzes and extracts meaningful information by processing the product name input from the purchaser terminal 30 through the communication unit 400, and filtering with the morpheme analysis module 111 Includes module 112.

The morpheme analysis module 111 performs natural language processing through morpheme analysis for each part of speech. For analysis in text, analyze it by dividing it into a form of morpheme, which is the smallest unit that has meaning. In the case of English, morphemes are divided based on white space, and if exceptions are handled in only a few cases such as Can't, Didn't, etc., it is easy to analyze, but in the case of Korean, it is more difficult than English sentence analysis because investigations and endings are attached. have.

KoNLPy is a collection of Python packages for Korean information processing. KoNLPy is an open source software and includes various analyzers (Hannanum, Kkma, Komoran, Mecab, Twitter) inside. Therefore, more accurate morpheme analysis results can be obtained by applying the corresponding analyzer according to the characteristics of the input value.

Among these, Twitter analyzer can use Twitter because it is easy and easy to distinguish parts of speech. In general, product information registered in online shopping malls has the characteristic of additionally writing product descriptions, so a Twitter analyzer is used to remove the nouns and other parts of speech except English to extract the key nouns in the product name.

3 shows an example of a method for analyzing a morpheme of a Korean sentence. As shown in FIG. 3, the text "I am eating rice now" consists of several parts of speech. If a Twitter analyzer is used, each part of speech can be classified as shown in FIG. 3.

The filtering module 112 is for removing words without meaning in product category classification in the text preprocessing process before learning. This is implemented for the purpose of improving the accuracy of category recommendation by removing meaningless words in advance.

The learning-based category predictor 120 predicts product category classification through a machine learning technique using a supervised learning model. Machine learning is largely divided into a supervised learning model, an unsupervised learning model, and a reinforcement learning model. A supervised learning model is suitable to apply a machine learning technique to automatic recommendation of category information for large-scale products sold in online shopping malls. The supervised learning model is a method of learning learning data using a machine learning algorithm and making predictions on new data based on this.

In the present invention, as one of the machine learning classification algorithms, a naive Bayes classifier used for spam filtering or document classification using keyword search is used. The Naive Bayes classifier is a probabilistic classification method that applies Bayes Theorem to conditional probabilities and classifies input data by assuming independence of the probability that each element constituting a document or data appears.

A process of classifying and recommending product categories based on hierarchical category clusters using the category classifier 100 will be described with reference to FIG. 4 as follows.

The seller accesses the product transaction server 30 through the seller terminal 10 and inputs a product name to be sold (S400). For example, a seller may enter "military backpack soldier bag travel bag mountaineering bag" as a product name to be registered. As described above, there may be various types of product name input, such as inputting multiple product names into the city or including additional descriptions such as color or material in the product name.

When a seller inputs a product name through the seller terminal 10, the input data is received by the communication unit 500 of the product transaction server 30 through a network, and this data is transmitted to the morpheme analysis module 111. The morpheme analysis module 111 performs morpheme analysis on the product name data (S410), and only nouns and English words are extracted from the input text (S420). In the above input example, when only nouns or English are extracted through morpheme analysis, words such as "military", "backpack", "military", "bag", "travel", and "climbing" will be extracted.

Thereafter, the filtering module 112 determines whether there is a meaningless word in relation to the product name, and if there is a meaningless word, it filters (removes) it (S430). In the above input, all words are meaningful in relation to product classification, so the process goes to the next step without removing words.

When the noun words are extracted, the category prediction model is driven (S440). The category prediction model is performed based on the prediction model information stored in the prediction model DB 320. The previously stored product names for each category are used as learning data, and at this time, the product name (text information) data input in the process of inputting/registering the product that the seller wants to sell in the online shopping mall is converted into a number. When learning is completed, categories, words, number of occurrences, and scores are stored as predictive model information in the predictive model DB 320.

Finally, categories with high relevance are extracted through the category prediction model, and the top three categories are provided to the seller terminal as recommended categories (S450). When the seller selects one of the three recommended categories displayed on the seller terminal 10, the product name and category information corresponding thereto are registered in the category classification DB 310. Through this, there is an advantage in that a process of finding a product category as in the prior art can be omitted, and more accurate category classification is possible.

Next, in FIG. 2, the shopping cart product recommender 200 will be described in detail. Prior to describing the detailed configuration and operation of the shopping cart product recommender 200, the concepts of support and reliability will be first described.

Support is the ratio of a specific product appearing in the total number of transactions, and is calculated in the form of (number of specific items)/(number of total transactions). I can. 5 is a table showing the result of calculating the transaction-based support rating, and the total number of transactions is 9, of which, since the number of I1 products (items) appeared 6 times, the support of the I1 product was 6/9 = 66.66%. have.

Confidence is defined in the present invention as a probability of purchasing a specific portion B when purchasing a specific subset A, that is, a conditional probability. In Figure 5, the total number of transactions for purchasing I1 products in 9 total transactions is 6 (t100, t400, t500, t600, t800, t900), and among these 6 transactions, transactions for purchasing I5 products are 2 of t100 and t800. Since it is a case, the reliability is calculated as 2/6 = 1/3 = 33%.

The shopping cart product recommender 200 includes a category preprocessor 210 and an association analyzer 220.

The category preprocessor 210 is for determining a category hierarchy level to be analyzed by the association analyzer 220.

In order to provide a shopping cart recommendation function for a large number of items stored in online shopping malls, it is not efficient to perform a correlation analysis without category classification for items. The reason is that it takes a lot of time to derive the result value and for individual products. Rather, it is less meaningful to do association analysis. Therefore, in the present invention, a content-based recommendation method is applied by using category cluster information.

At this time, since it is not efficient to analyze all categories, it is desirable to measure the number of products registered in the online shopping mall and perform analysis on a certain number of categories or more. This is because valid data may not be extracted from a small category cluster during the machine learning process.

Referring to FIG. 6 showing the number of items for each category, some of the categories in which the first level classification, which is the highest level, belongs to 01 in the six-level category structure is illustrated. The total number of products in the category in which the first-level classification belongs to 01 is 3326, and the number of products in the category in which the second-level classification belongs to 01 as a sub-category cluster is 1676, and as the sub-category cluster, the third-level classification belongs to 01. It can be seen that the number of products in the category is 408.

If the number of products in the category to be analyzed is determined to be 500, the category with the highest number of products of 500 or more is a category (01_00_00_00_00_00) and its subcategory (01_01_00_00_00_00) with the highest level 1 category belonging to 01. Done.

The association analyzer 220 recommends a product that is highly likely to be purchased by the purchaser as a product related to the product in the shopping cart by the purchaser using machine learning for the category cluster determined by the category preprocessor 210.

In the present invention, as a machine learning method for association analysis, a Spark-based machine learning method may be used. As an in-memory system, Spark provides fast processing and operation speed for big data, supports R, Python, Java, and Scala languages, and has the advantage of having high scalability.

In addition, in the present invention, the FP-growth algorithm, which is one of the frequency pattern mining techniques, is used to implement shopping cart product recommendation based on spark. The FP-growth algorithm has the advantage of being fast because it scans the data only twice, so the support and reliability of each product were calculated using this algorithm. The FP-growth algorithm calculates the frequency of the whole scan during one scan, calculates the support level based on the calculated frequency, and uses it to sort in descending order of support for each transaction, and then uses the sorted transaction. Create a tree. As it descends along the created tree, the count value is recorded in each node, and when the count is completed for all transactions, the pattern is extracted by going back up again. Through this, the support and reliability of each product can be calculated.

7 is a flowchart illustrating data training and analysis in the association analysis method. First, the number of order records per category is measured (S700). And it performs the basic data filtering process. This may be performing a function of extracting only necessary data or only a specific category based on a certain period or date.

For efficient analysis, the depth of category analysis can be determined by setting the upper and lower limits. For example, if the upper limit line and the lower limit line are set to step 2 and step 4, respectively, in a category tree consisting of six steps, analysis can be performed on only the corresponding categories.

Then, each category layer is sequentially entered (S710), and the number of products in each category layer is calculated to determine a category layer satisfying the reference value (S720). In the above example, 500 is illustrated as a reference value, which can be determined as an appropriate value through simulation.

When the analysis target category hierarchies are determined, association analysis is performed for each category hierarchy, and the analysis result is stored (S730). The association analysis calculates the support and reliability for each product using the FP-growth algorithm. Here, the support level is calculated as the number of appearances of each product relative to the total number of transactions, as described above, and the reliability may be calculated as a probability of purchasing a category B product when purchasing a category A product for each category to be analyzed.

Hereinafter, the results of a correlation analysis experiment will be described with reference to FIGS. 8 to 10.

8 is a table showing test data comparison results tested on actual data. This experiment was targeted at Company G, a B2B shopping mall operating company, and received information about 4,142 categories, 6,394 products, and 15,721 order records from Company G to provide support values and reference values for the number of products to determine the analysis target. While making the change, a shopping cart product association analysis was performed.

Referring to FIG. 8, a total of 3 experiments were performed. In one case, a total of 338 pairs of recommended products were calculated under the condition that the number of products was 500 or more and the support rating was 0.01 or more, and 216 out of 338 cases where two products belonged to different categories. I could see it was a dog.

In the second experiment, as a result of the experiment by lowering the support rating to 0.005, the number of recommended product pairs increased to 412, and 53 categories increased compared to the first experiment. In addition, it was found that 252 out of 412 products belonged to different categories. The result of recommending shopping cart products in the second experiment is shown in FIG. 9.

In the third experiment, as a result of lowering the number of products to more than 100, the number of recommended product pairs increased to 1108, and 761 categories increased compared to the first experiment. In addition, it was found that 556 out of 1108 cases where two products belong to different categories. In this experiment, the reliability was unified as 0.01 or higher. The result of recommending shopping cart products in the third experiment is shown in FIG. 10.

Considering that when a buyer purchases a product, there is a high possibility that the product belonging to the complementary product of the product is likely to be purchased at the same time when the buyer purchases a product, it is also possible to determine the recommendation target only when the two products belong to different categories. .

10: seller terminal 20: buyer terminal
30: product transaction server 100: category classifier
110: text preprocessor 111: morpheme analysis module
112: filtering module 120: learning-based category predictor
200: Cart product recommender 210: Category preprocessor
220: related analyzer 300: database unit
310: Category classification DB 320: Prediction model DB
330: User information DB

Claims (5)

In a method for providing a shopping cart product recommendation service based on a hierarchical category cluster in a product transaction server that is connected to a seller terminal and a buyer terminal through a network to mediate product transactions,
A product registration step of performing a category registration process for a product required for registration from a seller terminal in the product transaction server,
And a product recommendation step of determining a shopping cart recommended product for the purchaser terminal by using category information for the product previously registered in the product transaction server,
In the product registration step, a first step of receiving product name text information of a product to be registered from the seller terminal in a product transaction server, and a second step of extracting nouns and English words from the product name text by analyzing the morpheme of the input product name text. , A third step of extracting category information that is highly correlated with the extracted nouns and English words by driving a machine learning-based category prediction model, transmitting the extracted category information to a seller terminal, and category information selected through the seller terminal It is configured including a fourth step of determining as the category of the product,
The product recommendation step includes an eleventh step of measuring the number of products belonging to each hierarchical category in each hierarchical category cluster in a product transaction server, and determining a category hierarchy in which the measured number of products is equal to or greater than a preset reference value; and The 12th step of calculating the support and reliability of products belonging to each category by applying a machine learning method for each category hierarchy, and the product in the shopping cart based on the calculated support and reliability information when a buyer puts a specific product in the shopping cart A method for recommending shopping cart products based on hierarchical category clusters, characterized in that the method of recommending shopping cart products based on hierarchical category clusters, characterized in that the method of recommending shopping cart products based on hierarchical category clusters, characterized in that it comprises a thirteenth step of determining a recommended product having a high probability of purchasing together with, but determining the recommended product as a product belonging to a different category from the product in the shopping cart. delete delete delete The method of claim 1,
The machine learning-based category prediction model quantifies product name text data input in the process of registering a product in a product transaction server by a seller, and predicts the category, number of occurrences, and score information for each word from the machine learning process. A method for recommending shopping cart products based on hierarchical category clusters, characterized in that they are stored as model information.

Images