KR102205324B1 - Merchant Recommend Method using Double DNN - Google Patents

Abstract

The present invention relates to a method for recommending affiliated stores using a double DNN, and more particularly, in a financial company dealing with customers such as a credit card company or a bank, in order to understand the lifestyle of a customer using a credit card, It is for predicting and recommending the next visited affiliate store that is expected to use the card. Banks and credit card companies can provide suitable promotions and business opportunities to customers through such predictions, thereby increasing sales. To this end, in the present invention, the first neural network calculates the probability vector of the affiliated stores at the point t that the customer will visit next by using the customer vector and the affiliated store vector for the customer, and the affiliated store vector is the point immediately before the point t. A first step of calculating the average of the affiliate store vector at the point t-1 and the affiliate store vector at the point t-2 immediately before the point t-1; A second step of calculating, by a second neural network, a probability vector of affiliated stores that the customer will visit at the time t-1 by using the customer vector and the affiliated store vector at the time t-2; And a third step of calculating a joint probability distribution using the probability vector of the affiliated stores calculated by the first neural network and the probability vector of the affiliated stores calculated by the second neural network. It is preferable to include.

Description

Merchant Recommend Method using Double DNN}

The present invention relates to a method for recommending affiliated stores using a double DNN, and more particularly, in a financial company dealing with customers such as a credit card company or a bank, in order to understand the lifestyle of a customer using a credit card, It is for predicting and recommending the next visited affiliate store that is expected to use the card. Banks and credit card companies can provide suitable promotions and business opportunities to customers through such predictions, thereby increasing sales.

For credit card companies or banks that issue credit cards or debit cards to customers and use sales commission as a source of income, if you can predict and recommend the customer's next credit card use, that is, the next affiliate store, you will be able to grasp the customer's lifestyle. In addition, credit card companies and banks can capture various promotions and business opportunities suitable for customers through this, which can lead to increased sales.

In this way, the system for predicting the next affiliate store, etc. is a kind of recommendation system. The recommendation system is one of the most successful and widely used application programs for computer systems in the field of machine learning operated by computer systems. In particular, the Collaborative Filtering algorithm is one of the main algorithms used in the recommendation system and predicts the next item to be purchased according to the purchase order. Most of the latest models for collaborative filtering algorithms operate based on a deep neural network (DNN) installed in a computer system.

과 타겟 값

사이의 손실을 최소화함으로써 수행된다. Among the recommendation systems based on deep neural networks (DNN), there is Neural Collaborative Filtering (NCF). FIG. 1 shows the NCF model. As shown in FIG. 1, NCF recommends products using three deep neural networks. The first neural network used by NCF is a Generalized Matrix Factorization using an identity function as an activation function. GMF). The second neural network is Multi-Layer Perceptron (MLP), which uses ReLU as an activation function. The role of these two neural networks is to transform the customer or product into an embedding vector. Each transformed embedding vector is combined and fed into the input of the final neural network, NeuMF, called Neural Matrix Factorization (NeuMF). The last neural network, NeuMF, predicts the product the user will purchase, and the training is the predicted value as follows.

And target value

This is done by minimizing the loss between.

는 타겟값의 집합을 나타내며,

는 예측값의 집합을 나타낸다.

는 학습 인스턴스

의 가중치를 나타내는 하이퍼 매개 변수이다.here

Represents a set of target values,

Represents a set of predicted values.

Is a learning instance

Is a hyper parameter that indicates the weight of

Another recommendation system is Deep Neural Neworks for YouTube Recommendations. The YouTube Recommender System aims to recommend videos that users will watch next after they watch them. 2 is a DNN (Deep Neural Network) model of the YouTube Recommender System. As shown in Fig. 2, the YouTube Recommender System consists of two deep neural networks. One is a candidate generation neural network in charge of generating candidates, and the other is a ranking generation neural network used to rank the generated candidates. The candidate-generated neural network receives the user's video viewing history as input and finds millions of videos that the user can watch. The ranking neural network is trained using more features that can describe the relationship between the video and the user. The input of the ranking generating neural network is the output of the candidate generating neural network. That is, the ranking generation neural network ranks hundreds of videos out of the millions of videos generated by the candidate generation neural network. As described above, the present invention is to successfully grasp the relationship between the customer and the product in order to recommend the next product to the customer. YouTube Recommender System also calculates the probability of the next product, and uses the following equation for this.

과

의 연관 관계가 적어 다음 동영상, 상품 또는 가맹점의 추천결과에 안 좋은 영향을 끼칠 수 있다. 가맹점 추천의 경우를 예를 들자면 고객이

에서 이용한 가맹점이 평상시 이용하는 패턴에 의한 가맹점이 아니라 타인을 위한 선물을 사기 위하여 이용한 가맹점인 경우

에서 이용한 가맹점정보는 노이즈가 되고, 이로 인하여 가맹점의 추천결과는 전혀 엉뚱하게 나올 수가 있는 것이다.However, in events that occur sequentially, for example, watching a video, there may be noise. In other words

and

The relationship between the following videos, products, or affiliates may adversely affect the recommendation results. In the case of affiliate recommendation, for example, a customer

If the affiliated store used in is an affiliated store used to buy a gift for another person, not an affiliated store based on the pattern used normally

The affiliated store information used in is noise, and the affiliated store's recommendation result may come out completely wrong.

Paul Covington, Jay Adams and Emre Sargin. “Neural Networks for YouTube Recommendations” In RecSys'16 Proceedings of the 10th ACM Conference on Recommender Systems, September 2016. pp.191-198 Xiangnam He, Lizi Liao, Hanwang Zhang, Liqiang Nie, Xia Hu, Tat-Seng Chua. “Collaborative Filtering” In WWW'17 Proceedings of the 26th International Conference on World Wide Web, April 2017. pp.173-182. Helene Massam, Jinnan Liu and Adrian Dobra. “Conjugate Prior for Discrete Hierarchical Log-linear models” In The Annals of Statistics Vol.37, No.6A, 2009. pp.3431-3467. W. Edwards Deming and Frederick F. Stephan. “A Least Squares adjustment of a sampled frequency table when the expected marginal totals are known” In The Annals of Mathematical Statistics. Vol. 35, 1940. pp. 615-630. Hao Wu, Yue Ning, Prithwish Chakraborty, Jilles Vreeken, Nikolaj Tatti, Naren Ramakrishnan. “Realistic Synthetic Population Datasets” arXiv:1602.06844v3 [cs.DB] 25 Feb 2016. Agresti, Alan. “Data Analysis” Third Edition. Wiley. 2012. Suvash Sedhain, Aditya Menon, Scott Sanner, Lexing Xie. “Autoencoders Meet Collaborative” In WWW'15 Companion Proceedings of the 24th International Conference on World Wide Web, May 2015. pp.111-112. Aaron van den Oord, Sander Dieleman, Benjamin Schrauwen. “Content-based music recommendation” In NIPS'13 Proceedings of the 26th International Conference on Neural Information Processing Systems. Vol. 2, December 2013. pp.2643-2651. Hao Wang, Naiyan Wang and Dit-Yan Yeung. “Deep Learning for Recommender Systems” In KDD'15 Proceedings of the 21st ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, August 2015. pp.1235-1244. Tomas Mikolov, Kai Chen, Greg Corrado and Jeffrey Dean. “Estimation of Word Representations in Vector Space” arXiv:1301.3781v3 [cs.CL] 7 Sep 2013. Tomas Mikolov, Ilya Sutskever, Kai Chen, Greg Corrado and Jeffrey Dean. “Representations of Words and Phrases and their Compositionality” In NIPS'13 Proceedings of the 26th International Conference on Neural Information Processing Systems Vol. 2, December 2013. pp. 3111-3119.

In the present invention, a DoubleDNN model is proposed to overcome the situation resulting from the above-described problem. In other words, in order to solve the problem that there may be noise in events that occur sequentially, the present invention uses DoubleDNN to predict and recommend affiliate stores that customers will visit next, but aims to present a recommendation model that minimizes the effect of noise. To

The technical problem to be achieved by the present invention is not limited to the above-mentioned technical problem, and other technical problems that are not mentioned can be clearly understood by those of ordinary skill in the technical field to which the present invention belongs from the following description. will be.

A preferred embodiment of the method for recommending an affiliate store using a double DNN according to the present invention is a method for recommending an affiliate store performed by a computer system having two deep neural networks (DNNs), wherein the computer system is the first neural network, and the customer to the customer The vector and the affiliated store vector are used to calculate the probability vector of the affiliated stores at the point t that the customer will visit next, and the affiliated store vector is the affiliated store vector at the point t-1 and the point t-1, which is the point immediately preceding the point t. A first step of calculating the average of the affiliated store vector at the point t-2, which is a point immediately before of A second step of calculating, by the computer system, a probability vector of affiliated stores that the customer will visit at the time t-1 by using the customer vector and the affiliated store vector at the time t-2 using the second neural network; And a third step of calculating, by the computer system, a combined probability distribution using a probability vector of affiliated stores calculated by the first neural network and a probability vector of affiliated stores calculated by the second neural network. It is preferable to use the affiliate store recommendation method using a double DNN, characterized in that it includes.

In addition to the above features, the probability vector of the affiliated stores calculated by the first neural network is as follows,

The probability vector of the affiliated stores calculated by the second neural network is as follows,

The combination probability distribution is preferably an affiliate store recommendation method using a double DNN, characterized in that:

는 j번째 고객의 고객벡터,

,

,

는 t, t-1 및 t-2시점에서 j번째 고객이 방문한 가맹점의 가맹점벡터이다.here

Is the customer vector of the jth customer,

,

,

Is the affiliate store vector of the affiliated store visited by the j-th customer at the time points t, t-1 and t-2.

In addition to the above-described features, the loss function used for the first and second neural networks uses a categorical cross-entropy, and in the categorical cross-entropy, the distance H between the model probability distribution q and the target probability distribution p is It is also preferable to use a method for recommending an affiliate store using a double DNN, which is characterized in that the measurement is performed together.

In addition, in addition to the above-described features, the model probability distribution q is calculated in the Softmax layer, and the Softmax function is preferably used as an affiliate store recommendation method using a double DNN, characterized in that:

은 출력 계층의

번째 output이다. here

Is the output layer

Is the second output.

In the present invention, since a method for predicting and recommending an affiliated store to be visited next by customers is provided, there is an effect that a credit card company or the like can provide appropriate promotions and business opportunities to customers. In addition, in the present invention, since doubleDNN is used to predict and recommend an affiliate store to be visited next by a customer, even when there is noise in the data, there is an effect of canceling the noise and making accurate prediction and recommendation.

1 shows a Neural Collaborative Filtering (NCF) model of a DNN-based recommendation system according to the prior art.
2 shows a DNN model of a YouTube Recommender System among DNN-based recommendation systems according to the prior art.
3 shows a dual DNN model in which an affiliate store recommendation method using a dual DNN according to the present invention is performed.
4 is a diagram illustrating a model of a first neural network among affiliated store recommendation models using a double DNN according to the present invention.
FIG. 5 shows a model of a second neural network among the affiliate store recommendation models using a double DNN according to the present invention.
6 is a configuration diagram of a contingency table used in an experiment of an affiliate store recommendation method using a double DNN according to the present invention.
7 is a performance comparison between the DoubleDNN model according to the present invention and other DNN models, showing the results for Precisions@5.
8 is a performance comparison between the DoubleDNN model according to the present invention and other DNN models, showing the result of Recall@5.
9 is a performance comparison between the DoubleDNN model according to the present invention and other DNN models, showing results for F1-Score@5.
10 is a performance comparison between the DoubleDNN model according to the present invention and other DNN models, showing the results for NDCG@5.

Hereinafter, the present invention will be described in detail so that the above-described objects and features become clear, and accordingly, a person of ordinary skill in the technical field to which the present invention pertains will be able to easily implement the technical idea of the present invention. In addition, in describing the present invention, when it is determined that a detailed description of the known technology may unnecessarily obscure the subject matter of the present invention, a detailed description thereof is provided as it is already well known in the technical field among known technologies related to the present invention. I will omit it.

In addition, the terms used in the present invention have selected general terms that are currently widely used as far as possible, but in certain cases, there are terms arbitrarily selected by the applicant, and in this case, the meaning of the terms has been described in detail in the description of the corresponding invention. It should be noted that the present invention should be understood as the meaning of the term, not the name of. The terms used in the description of the embodiments are only used to describe specific embodiments, and are not intended to limit the embodiments. Singular expressions include plural expressions unless the context clearly indicates otherwise.

The embodiments may be changed in various forms and may have various additional embodiments, in which specific embodiments are indicated in the drawings and related detailed descriptions are described. However, this is not intended to limit the embodiments to a specific form, and it should be understood that all changes, equivalents, or substitutes included in the spirit and scope of the embodiments are included.

In the description of various embodiments, expressions such as "first," "second," "first," or "second" may modify various elements of the embodiments, but the corresponding elements are not limited. For example, the expressions do not limit the order and/or importance of corresponding elements. The above expressions may be used to distinguish one component from another component.

Hereinafter, the present invention will be described with reference to the accompanying drawings. According to the present invention, the problem of recommending an affiliate store can be interpreted as a classification problem. Here, the label becomes the merchant the customer will visit next. The Softmax layer is used to calculate the probability of the output, which becomes the probability of the merchant to visit next. In the present invention, even if there is noise in the data, a computer system composed of two deep neural networks is used so that accurate prediction and recommendation can be made by canceling it.

To this end, it is preferable that the computer system first allow the first neural network (DNN-1), which is the first neural network, to calculate the probability of the affiliate stores that the customer will visit next for a given customer and affiliate store vector. That is, the first neural network (DNN-1) preferably performs the first step of calculating the probability vectors of the affiliated stores at the point t that the customer will visit next, using the customer vector and the affiliated store vector for the customer. . Here, the affiliated point vector is preferably calculated as an average of the affiliated point vector at the point t-1, which is a point immediately before the point t-1, and the affiliated point vector at point t-2, which is the point immediately before the point t-1. Therefore, it is preferable that the probability vectors of affiliated stores calculated by the first neural network (DNN-1) be as follows.

은 노이즈로 작용할 수 있기 때문에 본 발명에서는 두 번째 심층신경망을 이용하여 이를 상쇄시키게 된다. 따라서 두 번째인 신경망인 제2신경망(DNN-2)은, 상기 고객벡터와 상기 t-2시점의 가맹점벡터를 이용하여 상기 고객이 상기 t-1시점에 방문할 가맹점들의 확률벡터를 계산하는 제2단계를 수행하는 것이 바람직하다. 그리고 상기 제2신경망(DNN-2)이 계산한 가맹점들의 확률벡터는 아래와 같이 하는 것이 바람직하다.However, for the same reason as described above, the affiliated store at time t-1

Since silver can act as noise, in the present invention, the second deep neural network is used to cancel it. Therefore, the second neural network (DNN-2), which is the second neural network, uses the customer vector and the affiliate store vector at the time t-2 to calculate the probability vectors of the affiliate stores that the customer will visit at the time t-1. It is desirable to carry out step 2. In addition, the probability vector of the affiliated stores calculated by the second neural network (DNN-2) is preferably as follows.

과

의 결합 확률분포를 아래와 같이 계산하는 제3단계를 수행하는 것이 바람직하다.After the first and second steps are performed, a probability vector of affiliate stores calculated by the first neural network, which is an affiliate store probability vector that is an output of two neural networks, and a probability vector of affiliate stores calculated by the second neural network are used.

and

It is desirable to perform the third step of calculating the combined probability distribution of.

는 j번째 고객의 고객벡터,

,

,

는 t, t-1 및 t-2시점에서 j번째 고객이 방문한 가맹점의 가맹점벡터이다. 도 3은 상술한 상기 제1단계 내지 상기 제3단계가 수행되는 과정이 포함된, 본 발명에 의한 이중 DNN읕 이용한 가맹점 추천모델을 도시한 것이다. 도 3에서 보는 바와 같이 본 발명에 의한 DoubleDNN은 심층신경망이 2개(DNN-1, DNN-2) 이다. 즉 어떤 고객의 t-1시점과 t-2시점에 이용한 가맹점정보가 주어 졌을 때 t시점에서의 가맹점을 예측하는 모델이 왼쪽에 도시된 제1신경망(DNN-1)과, 어떤 고객의 t-2시점에 이용한 가맹점정보가 주어졌을 때 t-1시점에 이용할 것으로 예상되는 가맹점을 예측하는 제2신경망(DNN-2)을 포함한다. train data의 input은 word2vec model 사용해서 고객 embedding vector 및 가맹점 embedding vector를 생성하는 것이 바람직한데, 도 3에 표시된 mer2vec 및 cus2vec가 각각 이에 해당되는 부분이다. here

Is the customer vector of the jth customer,

,

,

Is the affiliate store vector of the affiliated store visited by the j-th customer at times t, t-1 and t-2. 3 is a diagram illustrating an affiliated store recommendation model using a double DNN according to the present invention, including the process of performing the above-described first to third steps. As shown in Figure 3, the DoubleDNN according to the present invention has two deep neural networks (DNN-1, DNN-2). In other words, given the affiliated store information used at the t-1 and t-2 points of a customer, the first neural network (DNN-1) shown on the left and the t- It includes a second neural network (DNN-2) that predicts the affiliate store expected to be used at the point t-1 when the affiliate store information used at the point 2 is given. For input of train data, it is preferable to generate a customer embedding vector and an affiliate store embedding vector using the word2vec model, and mer2vec and cus2vec shown in FIG. 3 are corresponding parts, respectively.

Meanwhile, the loss function used for the first neural network (DNN-1) and the second neural network (DNN-2) uses a categorical cross-entropy, and in the categorical cross-entropy, a model probability distribution q and a target probability distribution p It is desirable to measure the distance H as follows.

In addition, the model probability distribution q is calculated in the Softmax layer, and the Softmax function is preferably calculated as follows.

은 출력 계층의

번째 output 이다.here

Is the output layer

Is the second output.

Meanwhile, FIGS. 4 and 5 are diagrams illustrating models of the first neural network (DNN-1) and the second neural network (DNN-2), respectively. First, FIG. 4 is a model of the first neural network (DNN-1), which is a model for predicting affiliate store information to be used at time t when a customer is given affiliate store information used at time t-2 and time t-1. As an input of the first neural network model, an average vector of affiliate store embedding vectors together with a customer embedding vector is entered. And Output becomes the index for the affiliate store at time t. In this case, it is preferable to use cross entropy for the loss function and Adam optimizer for the optimization method.

And first, as a model of the second neural network (DNN-2), FIG. 3 is a model for predicting the affiliate store information to be used at the time t-1 when a customer is given the affiliate store information used at the time t-2. As an input of the second neural network model, an affiliate store embedding vector is entered together with a customer embedding vector. And the output becomes the index for the affiliated store at the time t-1. In this case, it is preferable to use cross entropy for the loss function and Adam optimizer for the optimization method.

Hereinafter, the present invention will be described in more detail through examples. The embodiments described below are illustratively shown to aid understanding of the present invention, and it should be understood that the present invention may be variously modified and implemented differently from the embodiments described herein. As described above, it is obvious to those of ordinary skill in the art that various changes and modifications can be made within the scope and spirit of the present invention, and it is natural that such modifications and modifications belong to the appended claims.

<Example>

The excellence of the present invention was verified by conducting an experiment in which an affiliate store recommendation method using a double DNN according to the present invention was conducted using artificially generated credit card usage history data following the data distribution of BC Card.

1. Artificial data creation

In order to conduct an experiment to verify the excellence of the present invention, artificial data on credit card usage was generated using the contingency table of the original data and the Iterative Proportional Fitting (IPF) algorithm. A common method of representing data in N patterns is to use a contingency table. Accordingly, the contingency table is configured to show the frequency distribution of N patterns, and FIG. 6 is a configuration diagram of a contingency table used in an embodiment of the present invention. As shown in Fig. 6, four category attributes were created for each pattern, and each category attribute was composed of customer (12 groups), affiliated stores (15 groups), transaction date (12 categories), and amount (6 categories), and then N In the contingency table that displays the frequency distribution of dog patterns, customers were divided into 12 categories based on age and gender. And the franchisees were divided into 15 business categories. As for the amount, it was divided into 6 categories according to the size. The transaction date was divided into 12 categories on a monthly basis.

A logarithmic linear model used to search for the relationship such as independence or conditional independence between variables was used in data in the form of a contingency table consisting of frequencies observed according to the categories of variables without any response variable. The model for finding the distribution of data using the contingency table is a log-linear model. There are several algorithms for generating data using a log-linear model. In the present invention, artificial data is generated using an iterative proportional fitting (IPF) algorithm to create credit card usage history data that follows the data distribution of BC Card.

2. Data

Synthetic data of credit card usage history was used for model learning and verification. Detailed information on the data is shown in the table below. The affiliate store that each customer visited last was used as a verification set, and the rest of the data were used as a training set. The vector dimensions for customers and affiliated stores were 32, 64, and 128.

Number of customers Number of affiliated stores Number of data Credit card data 1 4,427 6,959 159,169

3. Evaluation index

는 고객 집합이며,

는 타겟 가맹점의 집합이다. 그리고

은 모델이 추천한 가맹점 집합이며,

는 고객 수이다.

는

가맹점에 대한 점수,

은

가맹점에 랭킹 점수이다. Using Precision, Recall, F1-score, and Normalized Discounted Cumulative Gain (NDCG) as evaluation indicators, the top-5 affiliates were recommended for each customer. The formula for each evaluation index is as follows, in the formulas below

Is the set of customers,

Is the set of target affiliate stores. And

Is the set of affiliate stores recommended by the model,

Is the number of customers.

Is

The score for the merchant,

silver

This is the ranking score for the affiliated store.

First, Precision is a measure of exactness, and Precision calculates the probability of accurately recommending affiliates based on all recommended affiliates. That is, Precision is a concept as follows.

Therefore, the precision of the top-5 affiliates is as follows.

Meanwhile, Recall is a measure of completeness. Therefore, Recall calculates the probability of the recommended affiliate store accurately based on the target affiliate store. That is, Recall is a concept as follows.

Therefore, the recall to the top-5 affiliated stores is as follows.

F1-Score uses precision and recall to calculate how good the recommended system is. Therefore, the F1-score for the top-5 affiliated stores is calculated as follows.

Since the sequence of recommended affiliates is also important, calculate how good the recommendation system is using NDCG (Normalized Discounted Cumulative Gain). The NDCG for the top-5 affiliates is calculated as follows.

4 experimental results

Experimental results are shown in FIGS. 7 to 10. 7 to 10 show that the DoubleDNN model, which is a model according to the present invention, has better performance than other DNN models, and FIG. 7 shows that the double DNN is superior to the DNN as a result of Precisions@5, and FIG. As a result of Recall@5, it can be seen that double DNN is superior to DNN. In addition, FIG. 9 shows that the double DNN is superior to the DNN as a result of F1-Score@55, and FIG. 10 shows that the double DNN is superior to the DNN as a result of NDCG@5. All of them show that DoubleDNN always outperforms DNN as the dimension of the affiliate store vector changes to 32, 64, and 128.

As described above, the present invention proposes a DoubleDNN model using two neural networks in order to remove noise from events that occur sequentially, such as visiting an affiliate store. From the experimental results, it was confirmed that the proposed model shows better performance than the latest DNN model. The evaluation indices changed according to the dimension of the affiliate store vector. It was found that if the dimension of the embedding vector increases, overfitting may occur and the performance of the recommendation model may be degraded.

Although the present invention has been described with various examples described above, the present invention is not necessarily limited to these examples, and various modifications may be made without departing from the spirit of the present invention. Accordingly, the examples disclosed in the present invention are not intended to limit the technical idea of the present invention, but to explain the technical idea, and the scope of the technical idea of the present invention is not limited by these examples. The scope of protection of the present invention should be interpreted by the claims below, and all technical ideas within the scope equivalent thereto should be construed as being included in the scope of the present invention.

Claims (4)

As an affiliate store recommendation method performed by a computer system having two deep neural networks (DNNs),
The computer system is the first neural network, and calculates the probability vector of the affiliated stores at the point t that the customer will visit next, using the customer vector and the affiliated store vector for the customer, wherein the affiliated store vector is the point immediately before the point t. a first step of calculating the average of the affiliate store vector at the time point t-1 and the affiliate store vector at the time point t-2, which is a point immediately before the time point t-1;
A second step of calculating, by the computer system, a probability vector of affiliated stores that the customer will visit at the time t-1 by using the customer vector and the affiliated store vector at the time t-2 using the second neural network; And
A third step of calculating, by the computer system, a combined probability distribution using a probability vector of affiliated stores calculated by the first neural network and a probability vector of affiliated stores calculated by the second neural network; Affiliate recommendation method using dual DNN, characterized in that including
The method of claim 1,
The probability vector of the affiliated stores calculated by the first neural network is as follows,

The probability vector of the affiliated stores calculated by the second neural network is as follows,

The combination probability distribution is a method for recommending an affiliate store using a double DNN, characterized in that:

here

Is the customer vector of the jth customer,

,

,

Is the affiliate store vector of the affiliated store visited by the j-th customer at times t, t-1 and t-2
The method of claim 1,
The loss function used for the first and second neural networks uses a categorical cross-entropy, and the distance H between the model probability distribution q and the target probability distribution p in the categorical cross-entropy is measured as follows. Affiliate recommendation method using double DNN

The method of claim 3,
The model probability distribution q is calculated in the Softmax layer, and the Softmax function is an affiliate store recommendation method using a double DNN, characterized in that:

here

Is the output layer

Th output

Images