KR20200103367A - Hybrid recommendation system and method - Google Patents

Abstract

According to the present invention, disclosed are a hybrid recommendation system and a method thereof. The present invention can provide an effective telemarketing service by selecting and providing customers with high contact potential, and has the advantage of improving telemarketing performance by increasing a subscription rate of products by recommending and providing an individualized product to a selected customer with high contact potential.

Description

Hybrid recommendation system and method {HYBRID RECOMMENDATION SYSTEM AND METHOD}

The present invention relates to a hybrid recommendation system and method, and more particularly, to a hybrid recommendation system and method for selecting customers with high contact potential to improve telemarketing performance and recommending individualized products to the selected customers. .

Telemarketing is one of the most widely used recommendation services for the purpose of marketing campaigns, and the recommendation system for telemarketing is becoming an important competency for improving business performance.

Recently, research on implicit data targeting customer patterns has been conducted in distribution management companies, and research on collaborative filtering considering distributed processing as in a P2P environment is also being conducted, which is differentiated from conventional recommendation systems. The system is being developed.

Initially, a recommendation algorithm was constructed by finding meaningful patterns or rules in a vast amount of data through automated or semi-automated means using data mining techniques.

In addition, various techniques from statistical techniques to artificial intelligence techniques are used for the search and analysis process, and through these techniques, various purposes such as Classification, Estimation, Prediction, and Clustering. Is achieving.

However, in the conventional recommendation system, a method such as collaborative filtering that predicts items with a high possibility to purchase in the future based on a preference for an existing purchase transaction or product has been mainly used.

In such an algorithm, sufficient preference information must be provided in order to present a reliable recommendation to the user. However, since users or items newly added to the recommendation system do not have sufficient information, it is generally impossible to obtain satisfactory recommendation items. Problems such as scarcity and scalability arise.

Due to this problem, reckless telemarketing attempts increase fatigue of customers who do not want to contact and cause dissatisfaction.

Document 1. Korean Patent Publication No. 10-0544985 (Name of invention: Personal telemarketing/telephone consultation system and method) Document 2. Korean Patent Publication No. 10-1797712 (Name of invention: Telemarketing standard script dynamic configuration system)

In order to solve this problem, an object of the present invention is to provide a hybrid recommendation system and method for selecting customers with high contact potential and recommending individualized products to the selected customers in order to improve telemarketing performance.

In order to achieve the above object, the hybrid recommendation system according to an embodiment of the present invention generates a data set based on at least one of customer master information and telemarketing transaction information, and extracts a preset contact probability from the generated data set. A contact probability extraction model unit that learns the contact prospect extraction model by applying it to a model, and provides a list of prospective contact customers including the selected customer information when a customer with a high probability of contact is selected through the learning of the contact prospect extraction model. Include.

In addition, the hybrid recommendation system according to an embodiment of the present invention generates a preset data set for analysis for each model based on at least one of the customer's master information and telemarketing transaction information, and the generated data set for each model Is applied to a preset product recommendation model to learn the product recommendation model, predicts a recommended product for each transaction through learning of the product recommendation model, and compares the predicted recommended product information with preset minimum recommendation reference information, and the The recommended products having the minimum recommendation criterion information or less may further include a product recommendation model that is filtered (or removed) and provided.

In addition, the product recommendation model unit is characterized in that it provides a recommended product list obtained by calculating a ranking of recommended products according to the filtering result.

In addition, the contact prospect extraction model unit generates a data set using telemarketing transaction information including customer master information, telemarketing history, point history, and subscription history to learn the contact prospect extraction model. ; A model learning unit for learning the contact probability extraction model by applying the generated data set to a preset contact probability extraction model; And when the learning of the contact probability extraction model is completed, a contact success prediction score is calculated by applying a preset test set to the learning-completed contact probability extraction model, and the contact probability is calculated in the order of the highest predicted connection success score. It characterized in that it comprises a contact potential customer calculation unit that selects high customers and provides a list of potential contact customers including customer information.

In addition, the data set generator includes a training set database including a data set for learning a contact probability extraction model; And a test set database including a data set for evaluating performance by applying the learned contact probability extraction model.

In addition, the contact probability extraction model of the model learning unit is characterized by learning through any one of a decision tree, an artificial neural network, a support vector machine, a logistic regression, a random forest, and a gradient boosting.

In addition, the product recommendation model unit generates a set model-specific data set using telemarketing transaction information including customer master information, telemarketing history, point history, and subscription history to learn a preset product recommendation model. A data set generator; A recommendation model generation unit for learning the product recommendation model by applying the generated model-specific data set to a preset product recommendation model; When the learning of the product recommendation model is completed, the product recommendation model on which the learning has been completed is applied to a preset test set to predict a recommended product for each transaction, and the predicted recommended product information is a preset recommendation score, a recommendation rating, or a support level. A model application prediction unit comparing recommended products with less than or equal to the minimum recommendation reference information by comparing with at least one minimum recommendation reference information among reliability; And a recommendation ranking calculator that calculates a ranking of recommended products according to the filtering result and provides a list of recommended products.

In addition, the model-specific data set generator generates a data set for each model converted into a data frame, a rating matrix, and a transaction form according to a random forest model, a collaborative filtering model, and a correlation analysis model.

In addition, the recommendation model generation unit is characterized in that it generates any one of a random forest model, a collaborative filtering model, and a correlation analysis model as a recommendation model.

In addition, the recommendation ranking calculation unit may calculate a ranking of recommended products by comparing the filtered recommended products with ranking calculation reference information preset for each of a random forest model, a collaboration filtering model, and a correlation analysis model.

In addition, the ranking calculation criterion information set for each model is characterized in that a random forest model is a recommendation score score, a collaborative filtering model is a product recommendation rating, and a correlation analysis model is a relation rule.

In addition, the hybrid recommendation method according to an embodiment of the present invention includes a) master information of the customer, telemarketing history, point history, and subscription history in order for the contact prospect extraction model unit 100 to learn the contact prospect extraction model. Generating a data set by using at least one of telemarketing transaction information; And b) the contact prospect extraction model unit applies the generated data set to a preset contact prospect extraction model to learn the contact prospect extraction model, and selects customers with high contact prospects through learning of the contact prospect extraction model. And outputting a prospective contact customer list including the selected customer information.

In addition, the hybrid recommendation method according to an embodiment of the present invention is c) among telemarketing transaction information including master information of the customer, telemarketing history, point history, and subscription history in order for the product recommendation model to learn the product recommendation model. Generating a preset data set for analysis for each model by using at least one; And d) the product recommendation model unit learns the product recommendation model by applying the generated data set for each model to a preset product recommendation model, predicts the recommended product for each transaction through learning of the product recommendation model, and predicts the product recommendation model. Comparing the recommended product information with preset minimum recommendation criteria information, filtering recommended products below the minimum recommendation criteria information, and outputting a recommended product list obtained by calculating a ranking of the recommended products according to the filtering result. It features.

In addition, in step a), a training data set for learning a contact prospect extraction model and a test data set for evaluating the performance of the contact prospect extraction model are generated by applying the learned contact prospect extraction model. It features.

In addition, the step b) includes: b-1) learning the contact probability extraction model by applying the data set generated in step a) by the contact probability extraction model unit to a preset contact probability extraction model; b-2) when the learning of the contact prospect extraction model is completed, calculating a contact success prediction score by applying a preset test set to the learning completed contact prospect extraction model; And b-3) selecting customers with high probability of contact in the order of the calculated access success prediction score, and generating a list of potential contact customers including customer information.

In addition, the contact probability extraction model is characterized by learning through any one of a decision tree, an artificial neural network, a support vector machine, a logistic regression, a random forest, and a gradient boosting.

In addition, the data set for analysis for each model in step c) is characterized in that it is generated by converting it into a data frame, a rating matrix, and a transaction form according to a random forest model, a cooperative filtering model, and a correlation analysis model.

In addition, step d) includes: d-1) learning the product recommendation model by applying the model-specific data set generated in step c) by the product recommendation model unit to a preset product recommendation model; d-2) when learning of the product recommendation model is completed, predicting a recommended product for each transaction by applying the learned product recommendation model to a preset test set; d-3) comparing the predicted recommended product information with at least one of a preset recommendation score, a recommendation rating, a support level, or a reliability level, and filtering recommended products below the minimum recommendation reference information; And d-4) calculating a ranking of recommended products according to the filtering result and outputting a list of recommended products.

In addition, the filtered recommended product in step d-4) is characterized in that the ranking of the recommended product is calculated by comparing it with predetermined ranking calculation reference information for each of a random forest model, a collaborative filtering model, and a correlation analysis model.

In addition, the ranking calculation criterion information set for each model is characterized in that a random forest model is a recommendation score score, a collaborative filtering model is a product recommendation rating, and a correlation analysis model is a relation rule.

The present invention is a method of randomly extracting target customers through the use of an approach that selects customers suitable for telemarketing by predicting the possibility of contact without randomly selecting customers for telemarketing when attempting telemarketing, and In contrast, there is an advantage of being able to effectively select valid customers.

In addition, the present invention has the advantage of providing an effective telemarketing service by selecting and providing customers with high contact potential.

In addition, the present invention has the advantage of improving telemarketing performance by increasing the subscription rate of products by recommending and providing individualized products to selected customers with high contact potential.

In addition, the present invention has the advantage of solving the cold start and scarcity problems of the existing recommendation algorithm by using content information such as customer master information and telemarketing history.

1 is a block diagram schematically showing a hybrid recommendation system according to an embodiment of the present invention.
2 is a block diagram showing a contact probability extraction model part of the hybrid recommendation system according to FIG. 1.
3 is a block diagram showing a data set generation unit of a contact probability extraction model unit according to FIG. 2.
4 is a block diagram showing a product recommendation model of the hybrid recommendation system according to FIG. 1.
5 is a flowchart illustrating a recommendation process using a hybrid recommendation system according to an embodiment of the present invention.
6 is a flow chart showing a data set generation process of a recommendation process using the hybrid recommendation system according to FIG. 5.
7 is a flowchart illustrating a process of selecting potential contact customers in a recommendation process using the hybrid recommendation system according to FIG. 5.
8 is a flow chart showing another data set generation process in the recommendation process using the hybrid recommendation system according to FIG. 5.
9 is a flowchart showing a personalized recommendation product selection process in the recommendation process using the hybrid recommendation system according to FIG. 5.

Hereinafter, preferred embodiments of a hybrid recommendation system and method according to an embodiment of the present invention will be described in detail with reference to the accompanying drawings.

In this specification, the expression that a certain part "includes" a certain component does not exclude other components, but means that other components may be further included.

In addition, terms such as "... unit", "... group", and "... module" mean units that process at least one function or operation, which can be classified into hardware, software, or a combination of the two.

1 is a block diagram schematically showing a hybrid recommendation system according to an embodiment of the present invention, FIG. 2 is a block diagram showing a contact probability extraction model unit of the hybrid recommendation system according to FIG. 1, and FIG. 3 is FIG. 4 is a block diagram showing a data set generation unit of the contact probability extraction model unit according to the present invention, and FIG. 4 is a block diagram showing a product recommendation model unit of the hybrid recommendation system according to FIG. 1.

1 to 4, the hybrid recommendation system according to an embodiment of the present invention is a configuration in which a model for predicting a customer with a high contact probability and a model for recommending a personalized product are combined, and a contact probability extraction model unit It is configured to include (100) and a product recommendation model unit (200).

The contact prospect extraction model unit 100 generates a data set based on at least one of customer master information and telemarketing transaction information, and applies the generated data set to a preset contact prospect extraction model to extract the contact prospect. When a customer with a high probability of contact is selected through learning of the contact probability extraction model, and a list of potential contact customers including the selected customer information is provided, a data set generation unit 110 and model learning It is configured to include a unit 120 and a contact potential customer calculation unit 130.

The data set generation unit 110 collects master information including the customer's age, gender, and name, and telemarketing transaction information including telemarketing history, point history, and subscription history in order to learn a contact prospect extraction model, When data collection is complete, a data set is created.

In addition, the data set may remove unnecessary variables by feature selection in order to improve the speed of the recommendation system and the accuracy of prediction.

In other words, transaction information with unsuccessful subscription can be excluded from the telemarketing history, and only transaction information with successful subscription can be collected. Preferably, the customer's master information and point history are used as independent variables, and the contact success of the telemarketing history Data set is created by using whether or not as a dependent variable.

In addition, the data set generation unit 110 generates a training data set for learning the contact probability extraction model, stores it in the training set database 111, and applies the learned contact probability extraction model to evaluate performance. A test data set is generated, divided and stored in the test set database 112.

The training data set is used for learning a contact prospect extraction model, and the test data set is used for evaluating performance by applying to the learning completed contact prospect extraction model, and the training data set and The test data set may be divided into a certain ratio, for example, a ratio of training data set 7: test data set 3, but is not limited thereto and may be selectively adjusted.

The model learning unit 120 is a configuration for learning a contact probability extraction model based on a machine learning program (or machine learning algorithm), and a data set (for example, for training) generated by the data set generation unit 110 Data set) is applied to a preset contact probability extraction model to learn the contact probability extraction model.

In addition, the model learning unit 120 learns a contact probability extraction model through iterative experiments or learning, and the contact probability extraction model is among decision trees, artificial neural networks, support vector machines, logistic regression, random forest, and gradient boosting. Learn through either.

When the learning of the contact prospect extraction model is completed, the contact prospect calculation unit 130 applies a preset test set to the learning completed contact prospect extraction model, for example, between '0' and '1'. Calculate a contact success prediction score with a score.

In addition, the probable contact customer calculation unit 130 selects customers with a high probability of contact in the order of the calculated access success prediction score, and generates and provides a probable contact customer list including the selected customer information. Print.

The product recommendation model unit 200 generates a preset data set for analysis for each model based on at least one of a customer's master information and telemarketing transaction information, and applies the generated data set for each model to a preset product recommendation model. The product recommendation model is learned by applying the product recommendation model, and the recommended product information for each transaction is predicted through learning of the product recommendation model, and the predicted recommended product information is compared with preset minimum recommendation reference information, and a recommendation below the minimum recommendation reference information The product is a configuration that provides a list of recommended products obtained by filtering (or removing) the ranking of recommended products according to the filtering result, and includes a data set generation unit 210 for each model, a recommendation model generation unit 220, and a model application. It is configured to include a prediction unit 230 and a recommendation ranking calculation unit 240.

The data set generation unit 210 for each model collects telemarketing transaction information including customer master information, telemarketing history, point history, and subscription history in order to learn the set product recommendation model, and when data collection is complete Classify according to variables and generate a data set for each set product recommendation model.

In addition, the data set generation unit 210 for each model converts the collected data into a data frame form if the product recommendation model is a random forest model, and converts the collected data into a rating matrix form if it is a collaborative filtering model, and analyzes the association. If it is a model, it converts the collected data into a transaction form and creates a data set for each learning model.

In addition, since the subscription failure transaction is not required when learning the product recommendation model, the model-specific data set generation unit 210 may extract only successful subscription transactions and use them for model learning.

The recommendation model generation unit 220 learns the product recommendation model by applying the model-specific data set generated by the model-specific data set generation unit 210 to a preset product recommendation model.

The recommendation model generation unit 220 generates any one of a random forest model, a collaborative filtering model, and a correlation analysis model as a recommendation model.

The random forest model performs model learning based on a machine learning program, and in this case, a multi-class using one categorical variable and a multi-label using several dichotomous variables as a method to utilize a plurality of products as dependent variables. The model can be considered.

In the collaborative filtering model, a matrix for calculating similarity may be selected, and Pearson correlation, cosine similarity, Euclidean distance, and the like may be selected.

The association analysis model may select the priority of the association rule based on reliability, support, and improvement, and the performance of the product recommendation model may be determined according to the selected criteria.

When the learning of the product recommendation model is completed, the model application prediction unit 230 predicts a recommended product for each transaction by applying the learned product recommendation model to a preset test set.

At this time, in the random forest model, a recommendation score between '0' and '1' is derived for each recommended product, a recommendation grade is derived in the collaborative filtering model, and a correlation rule is generated in the association analysis model.

In addition, the model application prediction unit 230 compares the predicted recommended product information with at least one of a preset recommendation score, a recommendation grade, a support level, or a reliability level, and selects recommended products below the minimum recommendation reference information. Filter to remove.

In other words, in the case of recommendation based only on priority, the absolute degree of recommendation of each product is not known by considering only the relative ranking between products, and thus, ranking-based recommendation recommends products with a high ranking even if the recommendation score is low. As a result, there is a problem in that the recommendation success rate decreases and the reliability of the recommendation system itself occurs.

Therefore, the model application prediction unit 230 removes the minimum recommendation criterion information from the recommendation target when the absolute recommendation score falls below a predetermined reference value, and in the case of the random forest model, based on the recommendation score, the collaborative filtering model Is based on the recommendation grade, and the association analysis model can be determined based on support or reliability.

In addition, the minimum recommendation reference information for filtering recommended products is not limited to a specific index or value, and may be appropriately changed and set according to industry, product, situation, user experience, and the like.

In addition, in the case of the random forest model and the collaborative filtering model, the model application prediction unit 230 generates a prediction result with a test set and then filters the recommended products according to the minimum recommendation criterion, and in the case of the association analysis model, the association rule After filtering is applied to the test set.

When filtering of recommended products is completed by applying a product recommendation model to a test set in the model application prediction unit 230, the recommendation ranking calculation unit 240 calculates a ranking of the recommended products according to the filtering result.

That is, the recommendation ranking calculation unit 240, in which the product recommendation model is a random forest model, assigns a recommendation rank from a product with a high recommendation score, and the collaborative filtering model gives a recommendation rank based on the recommendation rating for each product, and The analysis model gives the personalized product recommendation ranking based on the related rules.

In addition, when a product predicted with the same ranking occurs, the recommendation ranking calculation unit 240 may determine a ranking according to a predetermined importance of the product (eg, subscription success rate, unit price, etc.).

In addition, when a list of recommended products is derived for each transaction of the test set, the recommendation ranking calculation unit 240 generates a list of recommended products according to the ranking of the derived recommended products, and outputs and provides the generated list of recommended products. .

Next, a hybrid recommendation method according to an embodiment of the present invention will be described with reference to FIGS. 1 to 9.

The hybrid recommendation method according to the present embodiment includes the step of generating a data set (S100), selecting a customer who is likely to be contacted (S200), generating a data set for analysis for each learning model (S300), and It is configured to include a step (S400) of selecting a recommended product.

In the step of generating the data set (S100), in order for the contact possibility extraction model unit 100 to learn the contact possibility extraction model, among telemarketing transaction information including the customer's master information, telemarketing history, point history, and subscription history, At least one is collected (S110), and when data collection is completed, a data set is generated by classifying according to a preset variable (S110).

In addition, in the step S110, a training data set for learning a contact prospect extraction model and a test data set for evaluating the performance of the contact prospect extraction model are generated by applying the learned contact prospect extraction model, The training data set and the test data set may be generated by dividing by a predetermined ratio, for example, a training data set 7: a test data set 3.

In the step of selecting the contact prospect (S200), the contact prospect extraction model unit 100 applies the data set generated in the step S100 to a preset contact prospect extraction model to learn the contact prospect extraction model, and the A customer with a high probability of contact is selected through learning of the contact probability extraction model, and a list of potential contact customers including the selected customer information is output.

That is, in the step S200, the contact probability extraction model unit 100 converts the data set generated in the step S100 through any one of a decision tree, an artificial neural network, a support vector machine, logistic regression, random forest, and gradient boosting. The contact probability extraction model is trained by applying it to the learning contact probability extraction model (S210).

When the learning of the contact prospect extraction model in step S210 is completed, the contact prospect extraction model unit 100 calculates a contact success prediction score by applying a preset test set to the learned contact prospect extraction model (S220). .

When the contact success prediction score of the step S220 is calculated, the contact possibility extraction model unit 100 selects customers with a high probability of contact in the order in which the calculated access success prediction score is high, and generates a list of potential contact customers including customer information. (S230) and output.

In the step of generating the data set for analysis for each learning model (S300), in order to learn the set product recommendation model, the telemarketing transaction information including the customer's master information, telemarketing history, point history, and subscription history is collected (S310). And, when data collection is completed, a data set corresponding to each product recommendation model is generated by classifying it according to variables (S320).

In addition, the data set for each model generated in step S320 converts the collected data into data frame type data if the product recommendation model is a random forest model, and converts the collected data into data in the form of a rating matrix if the collaborative filtering model is In the case of a correlation analysis model, a step of converting the collected data into transactional data (S330) is additionally performed, and a data set is generated for each learning model based on the data converted in step S330 (S340).

In the step of selecting the personalized recommendation product (S400), the product recommendation model unit 200 applies the generated model-specific data set to a preset product recommendation model to learn the product recommendation model, and the product recommendation model Predicts recommended products for each transaction through learning of, compares the predicted recommended product information with preset minimum recommendation criteria information, filters recommended products below the minimum recommendation criteria information, and ranks recommended products according to the filtering result The recommended product list calculated by is displayed.

That is, in step S400, the product recommendation model unit 200 learns the product recommendation model by applying the model-specific data set generated in step S300 to a preset product recommendation model (S410).

In step S410, the product recommendation model may be any one of a random forest model, a collaborative filtering model, and a correlation analysis model, and the random forest model may perform model learning based on a machine learning program.

When the learning of the product recommendation model in step S410 is completed, the product recommendation model unit 200 predicts recommended products for each transaction by applying the learned product recommendation model to a preset test set (S420).

In step S420, a recommendation score between '0' and '1' is derived for each recommended product in the random forest model, a recommendation grade is derived in the collaborative filtering model, and a correlation rule is generated in the association analysis model.

In addition, the product recommendation model unit 200 compares the recommended product information predicted in step S420 with at least one of a preset recommendation score, a recommendation rating, a support level, or a reliability level, and The recommended products are filtered (S430) and removed.

In the step S430, the minimum recommendation criterion information is, in the case of a random forest model, a preset recommendation score as a minimum recommendation criterion, the collaborative filtering model uses a preset recommendation rating as a minimum recommendation criterion, and the association analysis model is a predetermined support level or Reliability is determined as the minimum recommendation criterion.

In addition, the minimum recommendation reference information for filtering recommended products is not limited to a specific index or value, and may be appropriately changed and set according to industry, product, situation, user experience, and the like.

When the filtering in step S430 is completed, the product recommendation model unit 200 compares the recommended product selected according to the filtering result with preset ranking calculation reference information to calculate the ranking of the recommended product (S440).

That is, the recommended product is information on the ranking calculation criteria set for each random forest model, a collaborative filtering model, and a correlation analysis model.For example, a random forest model is a recommendation score score, a collaboration filtering model is a product recommendation rating, and a correlation analysis model is a correlation rule. Compared with, the personalized recommended product ranking is calculated.

In addition, in the step S440, when a recommended product predicted with the same ranking occurs, the ranking may be determined according to the importance of the predetermined product (eg, subscription success rate, unit price, etc.).

When the calculation of the recommended product ranking in step S440 is completed, the product recommendation model unit 200 generates and outputs a list of recommended products based on the calculated recommended product ranking (S450).

On the other hand, in the present embodiment, the extraction of customers with a high contact probability in the process of generating a data set for selecting a customer for telemarketing has been described as an example, but is not limited thereto, and customers with low contact probability or randomly It can also be configured to extract.

In addition, in this embodiment, although the learning model of the product recommendation model unit 200 is configured for each model such as a random forest model, a collaborative filtering model, and an association analysis model, one learning model can be set and configured according to the recommended product. May be.

Therefore, conventional recommendation by selecting a telemarketing target customer through a hybrid recommendation model by combining the learning model of the contact probability extraction model unit 100 and the product recommendation model unit 200, and suggesting personalized recommended products to the target customer. Compared to the system, telemarketing performance can be greatly improved.

As described above, the present invention has been described with reference to preferred embodiments of the present invention, but those skilled in the art can variously modify and change the present invention within the scope not departing from the spirit and scope of the present invention described in the following claims. You will understand that you can do it.

In addition, reference numerals in the claims of the present invention are provided for clarity and convenience of description, and are not limited thereto. In the process of describing the embodiments, the thickness of the lines shown in the drawings, the size of components, etc. May be exaggerated for clarity and convenience of description, and the above-described terms are terms defined in consideration of functions in the present invention and may vary according to the intention or custom of users and operators. Is to be made based on the contents throughout this specification.

100: contact possibility extraction model part
110: data set generation unit
111: training set database
112: test set database
120: Model Learning Department
130: Contact prospective customer extraction unit
200: Product recommendation model section
210: Data set generation unit for each model
220: Recommendation model generation unit
230: model application prediction unit
240: recommendation ranking calculation unit

Claims (8)

Create a data set based on at least one of the customer's master information and telemarketing transaction information, apply the generated data set to a preset contact prospect extraction model to learn the contact prospect extraction model, and the contact prospect extraction model When a customer with a high probability of contact is selected through learning, the contact probability extraction model unit 100 provides a list of prospective customers including the selected customer information; And
The product recommendation model is learned by generating a preset model-specific analysis data set based on at least one of the customer's master information and telemarketing transaction information, and applying the generated model-specific data set to a preset product recommendation model. And, by predicting a recommended product for each transaction through learning of the product recommendation model, comparing the predicted recommended product information with preset minimum recommendation criteria information, and filtering (or removing) recommended products below the minimum recommendation criteria information. Hybrid recommendation system including a product recommendation model unit 200 for providing a list of recommended products obtained by calculating a ranking of recommended products according to a filtering result.
The method of claim 1,
The contact probability extraction model unit 100 is a data set that generates a data set using telemarketing transaction information including customer master information, telemarketing history, point history, and subscription history in order to learn the contact probability extraction model. Generating unit 110;
A model learning unit 120 for learning the contact probability extraction model by applying the generated data set to a preset contact probability extraction model; And
When the learning of the contact probability extraction model is completed, a predetermined test set is applied to the learning-completed contact probability extraction model to calculate a contact success prediction score, and the contact probability is high in the order of the calculated connection success prediction score. A hybrid recommendation system comprising a prospective contact calculation unit 130 that selects a customer and provides a list of potential contact customers including customer information.
The method of claim 2,
The data set generator 110 includes a training set database 111 including a data set for learning a contact probability extraction model; And
A hybrid recommendation system comprising a test set database (112) including a data set for evaluating performance by applying the learned contact probability extraction model.
The method of claim 1,
The product recommendation model unit 200 generates a set model-specific data set using telemarketing transaction information including customer master information, telemarketing history, point history, and subscription history in order to learn a preset product recommendation model. A model-specific data set generation unit 210;
A recommendation model generation unit 220 for learning the product recommendation model by applying the generated model-specific data set to a preset product recommendation model;
When the learning of the product recommendation model is completed, the product recommendation model on which the learning has been completed is applied to a preset test set to predict a recommended product for each transaction, and the predicted recommended product information is a preset recommendation score, a recommendation rating, or a support level. A model application prediction unit (230) comparing recommended products with less than or equal to the minimum recommendation reference information by comparing with at least one minimum recommendation reference information among reliability; And
And a recommendation ranking calculation unit 240 for calculating a ranking of recommended products according to the filtering result and providing a list of recommended products.
a) In order for the contact prospect extraction model unit 100 to learn the contact prospect extraction model, a data set is generated using at least one of the customer's master information, telemarketing history, point history, and telemarketing transaction information including subscription history. Step to do;
b) The contact prospect extraction model unit 100 applies the generated data set to a preset contact prospect extraction model to learn the contact prospect extraction model, and the customer with a high contact prospect through learning the contact prospect extraction model Selecting and outputting a contact potential customer list including the selected customer information;
c) In order for the product recommendation model unit 200 to learn the product recommendation model, the customer's master information, telemarketing history, point history, and telemarketing transaction information, including subscription history, are used to analyze preset models by using at least one of Generating a data set for use; And
d) The product recommendation model unit 200 learns the product recommendation model by applying the generated model-specific data set to a preset product recommendation model, and predicts the recommended product for each transaction through learning of the product recommendation model. , Comparing the predicted recommended product information with preset minimum recommendation reference information, filtering recommended products less than the minimum recommendation reference information, and outputting a recommended product list obtained by calculating a ranking of recommended products according to the filtering result. Including hybrid recommendation method.
The method of claim 5,
In the step a), a training data set for learning a contact prospect extraction model and a test data set for evaluating the performance of the contact prospect extraction model are generated by applying the learned contact prospect extraction model. Hybrid recommendation method.
The method of claim 5,
The step b) includes: b-1) learning the contact probability extraction model by applying the data set generated in step a) to a preset contact probability extraction model by the contact probability extraction model unit 100;
b-2) when the learning of the contact prospect extraction model is completed, calculating a contact success prediction score by applying a preset test set to the learning completed contact prospect extraction model; And
b-3) selecting customers with a high probability of contact in the order of the calculated connection success prediction score, and generating a list of potential contact customers including customer information.
The method of claim 5,
Step d) includes: d-1) learning the product recommendation model by applying the model-specific data set generated in step c) to a preset product recommendation model by the product recommendation model unit 200;
d-2) when learning of the product recommendation model is completed, predicting a recommended product for each transaction by applying the learned product recommendation model to a preset test set;
d-3) comparing the predicted recommended product information with at least one of a preset recommendation score, a recommendation rating, a support level, or a reliability level, and filtering recommended products below the minimum recommendation reference information; And
d-4) calculating a ranking of recommended products according to the filtering result, and outputting a list of recommended products.

Images