KR20220014499A - A taste network generation system based on user experience using CNN learning - Google Patents

Abstract

The present invention relates to a user experience-based taste network generation system using CNN learning. The user experience-based taste network generation system includes: a data collection part collecting service review data; a data processing part processing the collected data into a three-dimensional matrix form, while composed of two-dimensional images having a plurality of channels, and processing a pixel value of each of the images into a satisfaction score; a neural network analysis part instructing and analyzing a CNN neural network with the processed data, while inputting the two-dimensional images having the plurality of channels into the CNN neural network, and classifying the output of the network as a group; a network formation part classifying each user or dish as a group into the neural network, linking users or dishes in the same group to each other to form a network, and setting a relation between a user and a dish based on a pixel value of a two-dimensional image; and a service recommending part exploring for a dish along a link from a user requesting a recommendation (hereinafter referred to as "inquiring user"), through the network to recommend the discovered dish. Through the system, groups by taste are subdivided through the CNN neural network to predict a preference of a corresponding user from a taste relation with other users having a similar taste, and, as a result, a more reliable recommendation search can be carried out, and meaningful data applicable to various industries can be continuously processed and produced.

Description

A taste network generation system based on user experience using CNN learning }

The present invention relates to a user experience-based taste relationship network generation system using a CNN learning method that collects raw data generated from a large number of people with unspecified tastes and derives a taste group having the same preference in a specific part.

In particular, the present invention collects a large number of unspecified taste-related data and processes it into taste information, groups the processed information into related information using a CNN neural network, forms a relationship network based on the derived group group, and forms a relationship network. It relates to a user experience-based taste relationship network generation system using a CNN learning method that predicts users' preferences based on

In general, in order to find out about famous restaurants, cafes, pensions, bars, etc., they search cafes or blogs through the Internet, check the reviews of other users, and select a store to visit. However, there are many cases in which it is difficult to provide honest reviews of actual customers who have visited and visited because many of the articles published in cafes and blogs contain highly promotional content.

In addition, various platforms (Dining Code, Mango Plate, etc.) introduce users by recommending restaurants and recommending places suitable for users, and receive promotional expenses from business owners. However, there are many cases where these recommendations work well, but there are many cases where they do not. For example, in the case of Baek Jong-won's alley restaurant, which recently became the most fire, it was also often the subject of controversy.

The reason for this is that recommendations and promotions are made only with surrounding data in a situation where there is no personal preference standard. For this reason, many service providers and catering business operators have no choice but to focus on quantitative PR rather than qualitative PR, and the vicious cycle of continuously spending huge amounts of PR expenses continues.

In order to solve this problem, technologies for recommending items suitable for each consumer's individual taste have been proposed. As an example, a technology for selecting and recommending a common store preferred by customers with similar tastes by calculating a customer's store preference and correlation with other customers to match customers with high correlation has been proposed [Patent Document 1]. In addition, techniques for estimating and recommending webtoon readers' preference for unsubscribed webtoons based on their subscription histories [Patent Document 2], or a technology for recommending books as a preferred genre by gathering people with similar book tastes [Patent Document 3] are also presented. has become [Patent Document 2].

However, the prior arts group or filter using a mathematical model such as a Hamming weight method, a link sum value by a matrix, or a k-means clustering method. Therefore, since it is necessary to select a suitable mathematical model according to the recommendation target or data collection method, it is difficult to apply to a general model, and is not particularly suitable for the recommendation of the restaurant industry.

Therefore, there is a need for a technology that can present standards for each taste, subdivide groups for each taste, form a taste relationship with others with similar tastes, and provide information based on experience rather than publicity.

Korean Patent Publication No. 10-2005289 (2019.07.31. Announcement) Korean Patent Publication No. 10-1755653 (2017.07.11. Announcement) Korean Patent Publication No. 10-2020-0057209 (2020.05.26. Announcement)

An object of the present invention is to solve the above-described problems, by collecting an unspecified number of taste-related data, processing it into taste information, and grouping the processed information into relevant information using a CNN neural network, and To provide a system for generating a user experience-based preference relationship network using a CNN learning method that forms a relationship network based on group groups and predicts users' preferences based on the relationship network.

In order to achieve the above object, the present invention relates to a user experience-based taste relationship network generation system using a CNN learning method, comprising: a data collecting unit for collecting service review data; a data processing unit that processes the collected data in a three-dimensional matrix form, but consists of two-dimensional images having a plurality of channels, and processes the pixel value of each image to be a satisfaction score; A neural network analyzer that trains and analyzes a CNN neural network with processed data, wherein the input of the CNN neural network is a two-dimensional image having a plurality of channels, and the output is classified into groups; Classifying each user or food into groups through a neural network, forming a relationship network by connecting users or foods belonging to the same group with a link, and forming a relationship network that sets the relationship between users and food by pixel values of a two-dimensional image wealth; and a service recommendation unit for recommending the searched food by searching for a link from a user requesting a recommendation (hereinafter referred to as a querying user) using the relational network.

In addition, in the present invention, in a user experience-based taste relationship network generation system using a CNN learning method, one of the two axes of the two-dimensional image represents a food classification item for classifying food, and the other is an evaluation for classifying food evaluation Classification items are indicated, and each evaluation item is characterized in that it is an item by a combination of a food classification item and an evaluation classification item.

In addition, the present invention is characterized in that in the user experience-based taste relationship network generation system using the CNN learning method, the CNN neural network groups users or food by a plurality of fields and outputs them as a group by field.

In addition, the present invention is a user experience-based taste relationship network generation system using a CNN learning method, wherein the relationship network is composed of nodes and links, nodes are assigned to users or food, links connect nodes, and two nodes are When belonging to the same group, two nodes are connected to each other by a link, and the link has a field attribute corresponding to a field of the same group.

In addition, in the present invention, in the user experience-based taste relationship network generation system using the CNN learning method, the link between the user node and the food node is connected by the statistical score of the two-dimensional image of the evaluation item, and in proportion to the statistical score of the evaluation item It is characterized in that the weight is set.

As described above, according to the user experience-based taste relationship network generation system using the CNN learning method according to the present invention, the preference of the corresponding user is predicted from the taste relationship with others who have similar tastes by subdividing groups by taste with the CNN neural network. , a more reliable recommendation search can be performed, and the effect of continuous processing and production of meaningful data applicable to various industries is obtained.

1 is a diagram showing the configuration of an overall system for implementing the present invention.
2 is a block diagram of a configuration of a user experience-based preference relationship network generation system using a CNN learning method according to an embodiment of the present invention.
3 is a view showing the structure of three-dimensional data according to an embodiment of the present invention.
4 and 5 are exemplary configuration diagrams for 3D data according to an embodiment of the present invention.
6 is a graph showing the configuration of evaluation items of two-dimensional data according to an embodiment of the present invention.
7 is a block diagram of a CNN neural network according to an embodiment of the present invention.

Hereinafter, specific contents for carrying out the present invention will be described with reference to the drawings.

In addition, in demonstrating this invention, the same part is attached|subjected by the same code|symbol, and the repetition description is abbreviate|omitted.

First, the configuration of the entire system for carrying out the present invention will be described with reference to FIG. 1 .

As shown in Figure 1, the entire system for implementing the present invention is a user terminal 10 such as a smartphone used by a user, an order management app 20 installed in the user terminal 10, and a seller terminal ( 50), and an order management server 30 that processes or recommends an order. The order management server 30 is connected to each other through the user terminal 10, or the business terminal 50, and the network 80, such as the Internet.

First, the user terminal 10 is a mobile terminal used by a user, and is a terminal equipped with a normal computing function, such as a smart phone, a phablet, and a tablet PC. In particular, the user terminal 10 is a terminal in which an application or a mobile application (or an app, an application) can be installed and executed.

An order management app 20 and the like may be installed in the user terminal 10 . The user may use services such as order processing and order recommendation of the order management server 30 by using the order management app 20 installed in the user terminal 10 .

In addition, the order management app 20 is provided with a store selection function, a food ordering function, a payment function, and the like. That is, the order management app 20 according to the user's input, in conjunction with the order management server 30, selects a store, orders food, and performs payment.

In addition, the order management app 20 requests the user to write a review for the food service, and receives review data from the user. The order management app 20 collects the received review data and transmits it to the order management server 30 .

In this case, the review data is input as structured data such as evaluation target items, preference scores, and the like.

On the other hand, the user terminal 10 or the order management app 20 may be used by a general user in addition to the person ordering food. In particular, a general user or a non-payment user may register by inputting an evaluation or satisfaction with respect to food or a store (restaurant, business, etc.).

Next, the seller terminal 50 is a computing terminal used by the seller, and is a terminal equipped with a normal computing function, such as a PC, a notebook computer, a smart phone, a phablet, and a tablet PC.

In the seller terminal 50, a web browser or the like is installed or an order processing app is installed, and the administrator accesses the order management service provided by the order management server 30 through a web browser or an order processing app and provides an order management service. is available.

Next, the order management server 30, as a normal application server, performs tasks such as order reception, order processing, and food recommendation, and provides such an order management service. In particular, the order management server 30 interworks with the order management app 20 and the seller terminal 50 to process the corresponding order management service.

In particular, the order management server 30 collects each user's satisfaction data from the user or the order management app 20 , and processes the data (hereinafter, 3D taste data) in a three-dimensional matrix form.

In addition, the order management server 30 includes a CNN (Convolution Neural Network, convolutional neural network) neural network, and trains the neural network with data processed in the form of a three-dimensional matrix. The CNN neural network receives three-dimensional taste data and outputs a taste group.

In addition, the order management server 30 forms a relationship network based on the derived group group, and groups similar users according to their preferences.

And the order management server 30 searches for the user's taste group, extracts the food or store of the searched taste group, and recommends the extracted food or store.

On the other hand, the order management server 30 and the order management app 20 may be implemented according to a typical method of configuring a server and a client. That is, the functions of the entire system can be divided according to the performance of the client or the amount of communication between the server and the server. As an example, the order management app 20 performs only a simple interface function, and the order management server 30 may be configured to perform most tasks such as data processing, CNN-based analysis, and taste recommendation. As another example, the order management app 20 performs data input and data processing, and the like, and the order management server 30 may perform only an analysis operation using a CNN neural network. Hereinafter, a system for generating a preference relational network will be described, but it may be implemented in various forms according to a server-client configuration method.

Next, the configuration of the user experience-based preference relationship network generating system 300 using the CNN learning method according to an embodiment of the present invention will be described with reference to FIG. 3 .

The taste relationship network generation system 300 according to an embodiment of the present invention is configured as a client such as an order management app 20 and an order management server 30 and is implemented as a server-client system.

As shown in FIG. 3 , the order management system 300 according to an embodiment of the present invention includes a data collection unit 31 for collecting service review data, a data processing unit for processing the collected data in a three-dimensional matrix form ( 32), a neural network analysis unit 33 that trains and analyzes a CNN neural network with processed data, a relational network shaper 34 that forms a relational network into groups classified through a neural network, and a service that provides a recommendation service to a user It is composed of a recommendation unit (35). In addition, it is configured to further include a convolutional neural network (CNN) 38 and a relational network 39 .

First, the data collection unit 31 collects taste data (or review data) through a user's direct or indirect experience.

The user may register a service review after receiving the catering service directly, or may register a service review even if the catering service is not directly provided. The data collection unit 31 collects review data registered by the user as described above.

Specifically, in the case of a non-face-to-face order payment system for the catering industry, order payment is performed through a mobile application like a kiosk. A review service is provided so that users can enter and manage their experiences like reviews after placing an order and payment.

In addition, it provides a service for creating a list of users even if they do not pay, and provides a review service to register the user's non-payment preference data in the platform.

On the other hand, review data is structured and composed of evaluation targets, evaluation items, preference (or satisfaction) scores, and the like. That is, the data collection method collects satisfaction with taste, cleanliness, price, etc. after the service experience by inputting the specified range of 1-10.

The evaluation target may be food, food type, company name (restaurant name), and the like. In addition, evaluation items may be subdivided into taste, cleanliness, price, and the like. In addition, satisfaction (or preference) is divided into a plurality of grades and is input as grades. As an example, the satisfaction level is set by dividing it into 10 grades such as 1-10.

Next, the data processing unit 32 processes the collected data in the form of a three-dimensional matrix.

On the other hand, as shown in FIG. 3 , the three-dimensional matrix structure consists of a three-dimensional image or three-dimensional image data, in particular, two-dimensional images having a plurality of channels. In this case, the two-dimensional image of one channel has a W×H resolution. In addition, the number of channels C is configured as L.

As shown in FIG. 4 , the satisfaction (or preference) of the evaluation target is configured as a two-dimensional image. That is, the position coordinate (x,y) corresponds to the evaluation item, and the pixel value of the coordinate (x,y) is set as the satisfaction score of the evaluation item.

As shown in FIG. 5 , evaluation items are composed of two-dimensional coordinates, and each coordinate axis may indicate a food classification (type, etc.) and evaluation classification (taste, etc.), respectively.

As an example, information on food classification (hereinafter referred to as food classification item) is contained on the x-axis. For example, food classification items are main ingredients, cooking methods, and the like. That is, food classification indicates items that classify and classify food.

In addition, as shown in FIG. 6 , it may be classified into large, medium, and small items on the x-axis (items of food classification). The major category indicates the main ingredient type, the middle category indicates the main ingredient type, and the small category indicates the cooking method.

In addition, the y-axis contains information on food evaluation classification (hereinafter referred to as evaluation classification item). For example, the evaluation classification item may be composed of spicy taste, terminal, quantity, service quality, cleanliness, and scenery around the store. Evaluation classification refers to items that classify and classify evaluation viewpoints.

At this time, before data conversion, in the case of the catering industry (restaurant business, resting restaurant business, etc.), the type of catering business (western food, Chinese food, Japanese food, etc.), main menu (meat, fish, etc.), cooking method (roasted, fried, etc.), taste A data table for defining tastes such as (sweet, salty, bitter, etc.) is defined and processed in the form of a three-dimensional data matrix.

That is, each evaluation item is composed of a combination of a food classification item (x coordinate) and an evaluation classification item (y coordinate).

In addition, a two-dimensional table is implemented based on the x and y axes, and the corresponding table is accumulated for each company to become three-dimensional data. In addition, each user has corresponding 3D data, and these 3D data are grouped through a CNN (Convolutional Neural Network) method. Accordingly, one two-dimensional image represents evaluation items corresponding to a specific user and a specific food (or restaurant).

Also, the satisfaction score can be normalized by pixel range. That is, since pixel values have 0 to 255, satisfaction scores 1 to 10 can be proportionally mapped to the pixel values 0 to 255 range. In particular, the satisfaction score further includes 0, which means that there is no satisfaction score or there is no corresponding item. Therefore, satisfaction scores 0 to 10 are proportionally mapped to the range of pixel values 0 to 255. That is, as an example, a satisfaction score of 0 is mapped to a pixel value of 0, a satisfaction score of 10 is mapped to a pixel value of 255, and the like.

On the other hand, pixels that are not mapped and set in the 2D image (when there is no evaluation item or when there is no satisfaction score) are all padded with 0. Preferably, it is processed by setting the corresponding pixel value to 0.

In addition, the channel (C) represents each combination of a user and food as a measurement object.

Next, the neural network analyzer 33 trains the CNN neural network on data processed in the form of a three-dimensional data matrix, and analyzes it using this.

That is, each individual's various taste data are classified and analyzed based on the taste data table defined initially. That is, the neural network is used to analyze which group the processed taste data of the corresponding user belongs to.

In the case of static analysis that does not use a neural network, when a person named ‘A’ visits ‘A’ and leaves review data, only ‘A’’s preference for ‘A’ can be known. However, since the system according to the present invention uses the group value derived from the analysis using a neural network, it is possible to grasp the preferences of several people with different tastes and many companies visited by those people.

In addition, when the data of ‘B’ is input, it is possible to determine whether the tastes that can be satisfied by ‘A’ or ‘B’ or ‘C’ through the group value of ‘B’ derived after analysis.

As shown in FIG. 7 , the input of the CNN neural network is processed 3D data, and the output is composed of a group of users. That is, the output of the neural network is output in the form of a specific value by passing the corresponding 3D data through the neural network layer composed of the hidden layer. In this case, the output specific value becomes a group group for a user or food (or restaurant) of the corresponding 3D data.

In particular, the group is grouped by a plurality of fields. In the field, the field to be grouped by taste, quantity, price, cleanliness, etc. is set in advance. And grouping is performed for users or food (restaurants) for each field.

On the other hand, the test target data for learning creates a group for labeling based on the tastes of small groups (ABROSS team members and experiential groups, etc.) for the experiment. In addition, an appropriate number of group sizes and neuron activation functions (ReLU, etc.) are set, thresholds and biases are set, and the learning model is tested to use a model with a high hit rate.

Therefore, after the CNN neural network is trained, if a two-dimensional image of an evaluation item for a specific food (restaurant) of a specific user is input to the corresponding CNN neural network, the field group (or group group) to which the combination of the user and the food belongs is output do.

Next, the relational network forming unit 34 forms a relational network based on the previously derived group group. That is, a two-dimensional image (evaluation item image) for each user or each food is input to the neural network, and a group of the corresponding user or food is output and classified, and a relational network is formed by the classified group.

The relational network represents a group relationship of users derived through a CNN neural network, a group relationship of foods (restaurants), and a preference relationship between users and food.

Specifically, the relational network consists of nodes and links. Nodes are assigned users or food, and links connect nodes. In this case, when two nodes belong to the same group, the two nodes are connected to each other by a link. Also, the link has a field attribute. That is, the field of the group to which two nodes belong identically is set as the field attribute of the corresponding link.

In addition, a relationship network is formed based on the group group based on the classified type (type), such as classification according to taste preference (taste field), classification according to atmosphere preference level (atmosphere field), and classification according to cleanliness preference (cleanliness field). .

It is possible to derive the same group in terms of 'specific taste', 'price', and 'quantity' of food, and recommend a company that meets the user's needs.

For example, a case where users A and B are classified into the same group as follows will be described.

A and B eat a lot. [ positive equal group ]

A and C like the spicy taste. [ Same group of spicy taste ]

B and C prefer lower prices. [ Same group with preferred price range ]

A and B like a clean place [same group for cleanliness]

In this case, nodes A and B are connected to each other by a link, and the field properties of the link have <quantity> and <cleanliness>. Also, nodes A and C are connected by a link having an attribute in the <spicy taste> field. In addition, nodes B and C are connected by a link with an attribute of the field <price>.

In this case, there is a high probability that user A will like the company preferred by user B, and A and B have a relationship between quantity and cleanliness.

The above example describes a relationship network representing a group relationship of users.

The group relationship of foods is also formed in the same form as the relationship network of the group relationship of the users. That is, when the three-dimensional data of foods is input into the CNN neural network and classified into groups by field, the foods belonging to the same field group are linked by a link. In this case, the field attribute of the link is set to the field of the corresponding group.

In addition, the link between the user node and the food node is set by the two-dimensional image (or the pixel value of the image) of the evaluation item. That is, the two-dimensional image is a score obtained by evaluating the satisfaction score (pixel value) between the corresponding user node and the corresponding food node for each two-dimensional evaluation item. Therefore, if the overall score of the two-dimensional image is high, it indicates that the user prefers the corresponding food.

Therefore, if the statistical score of the two-dimensional evaluation item is greater than or equal to a predetermined reference value, a preference relationship is created by connecting the corresponding user node and the corresponding food node to each other.

Alternatively, a weight may be assigned to each link. The weight of the link is proportionally applied to the statistical score of the evaluation item. That is, statistical scores and weights have a proportional relationship.

Next, the service recommendation unit 35 provides a recommendation service to the user.

In other words, after forming a group among users with similar tastes (taste, atmosphere, cleanliness, etc.), a network of relationships is established so that the most similar users can continue to receive recommendations for services they have not experienced but subscribe to the good experiences of those users. service is provided.

Specifically, the service recommendation unit 35 recommends food using a relational network. That is, food is searched for and selected by following a link connected to a node of a user requesting a recommendation (hereinafter referred to as a query user). That is, it searches for other user nodes connected to the query user node, searches for foods connected with the searched user node and links, and extracts and recommends the searched food.

In addition, the service recommendation unit 35 may search for a food having the highest weight in consideration of the weight of the link between the user and the food when searching for the searched user's preferred food.

Also, the service recommendation unit 35 may search for other foods that are connected to each other by using a relational network between foods in the searched foods.

Also, when receiving the user's preference field, the service recommendation unit 35 may search by using a link in the preference field. For example, if a querying user prioritizes the "price" field, when navigating with a user-to-user link or a food-to-food link, the price field link searches for the linked user or food.

In the above, the invention made by the present inventors has been described in detail according to the above embodiments, but the present invention is not limited to the above embodiments, and various changes can be made without departing from the gist of the present invention.

10: user terminal 20: order management app
30: order management server 50: seller terminal
80: network

Claims (5)

In the user experience-based preference relationship network generation system using the CNN learning method,
Data collection unit for collecting service review data;
a data processing unit which processes the collected data in a three-dimensional matrix form, but consists of two-dimensional images having a plurality of channels, and processes the pixel value of each image to be a satisfaction score;
A neural network analyzer that trains and analyzes a CNN neural network with processed data, wherein the input of the CNN neural network is a two-dimensional image having a plurality of channels, and the output is classified into groups;
Classifying each user or food into groups through a neural network, forming a relationship network by connecting users or foods belonging to the same group with a link, and forming a relationship network that sets the relationship between users and food by pixel values of a two-dimensional image wealth; and,
Using the relationship network, a user experience-based preference relationship network generation system using a CNN learning method, characterized in that it includes a service recommendation unit that searches for a recommendation by following a link from a user requesting a recommendation (hereinafter, a query user), and recommends the searched food. .
According to claim 1,
One of the two axes of the two-dimensional image represents the food classification item that classifies food, the other represents the evaluation classification item that classifies the food evaluation, and each evaluation item is a food classification item and a combination of the evaluation classification item. User experience-based taste relationship network generation system using CNN learning method, characterized in that it is an item by
According to claim 1,
The CNN neural network groups users or food by a plurality of fields and outputs them as groups for each field.
4. The method of claim 3,
The relational network is composed of nodes and links, nodes are assigned to users or food, links connect nodes, and when two nodes belong to the same group, the two nodes are connected to each other by a link, and the link is of the same group. User experience-based preference relationship network generation system using CNN learning method, characterized in that it has field properties corresponding to the field.
5. The method of claim 4,
The link between the user node and the food node is connected by the statistical score of the two-dimensional image of the evaluation item, and the user experience-based taste relationship network generation system using the CNN learning method, characterized in that the weight is set in proportion to the statistical score of the evaluation item .

Images