KR102203252B1 - Method and system for collaborative filtering based on generative adversarial networks - Google Patents

Abstract

A method and system for collaborative filtering based on a generative adversarial network (GAN) are disclosed. The GAN-based collaborative filtering method includes: training a generative model and a discriminative model of the GAN in real-valued vector-wise for an item; And recommending a personalized item list to a user using the trained generation model and the discrimination model.

Description

Method and system for collaborative filtering based on generative adversarial neural networks {METHOD AND SYSTEM FOR COLLABORATIVE FILTERING BASED ON GENERATIVE ADVERSARIAL NETWORKS}

The following description relates to a collaborative filtering technology based on a generative adversarial network (GAN).

Collaborative filtering (hereinafter CF) is one of the most widely used methods in recommendation systems.

For example, Korean Patent Publication No. 10-1030653 (registration date April 14, 2011) corrects the similarity between the recommendation target user and other users by using the relationship between the recommendation target user and the entropy of each user preference information of another user as a weight. By doing so, a collaborative filtering recommendation technology in which the accuracy of prediction of preference is improved is disclosed.

The CF aims to provide the user with a personalized list of items by collecting the user's previous preference records. In general, the user's preference history is expressed in the form of a sparse matrix in which a column corresponds to a user, a row corresponds to an item, and a value corresponds to the user's preference for the item. Here, the user's preference (or feedback) may be provided with a rating of 1 to 5 (explicit), or may be provided in an unary manner such as purchase or click (implicit).

Among many CF methods, the model-based CF method, which builds a model for learning patterns of users' past preferences for items, is known to be the most successful. Some of the most widely used model-based CF methods are based on matrix factorization (MF), which learns linear interactions between the user and the latent features of an item.

Recently, deep neural network-based CF methods have recently attracted attention. Due to the DNN's ability to discover nonlinear interactions of potential factors between users and items and approximate arbitrary continuous functions, the DNN-based CF model has strong performance in both grade prediction and top-N recommendation work.

GAN is one of the DNNs that learn to mimic the distribution of the given data. The GAN consists of two models: one is the generative model (hereinafter referred to as'G') and the other is the discriminant model (hereinafter referred to as'D'). During training, G generates realistic (but fake) data and passes it to D. D evaluates the likelihood of the data being ground truth, not G. The training process is a minimax game between G and D: G tries to increase the error rate (i.e., fool) while D tries to accurately distinguish the actual data from the fake data. As a result of this hostile process, G learns how to generate realistic data, such as photographs, that at least appear to be authentic to the human eye.

The above-described GAN is being studied most actively in the field of image generation, but the research applied to the field of recommendation system is insignificant.

It is possible to provide a GAN-based collaborative filtering framework that provides high accuracy for recommendations.

A new direction of vector-wise hostile training used in the GAN-based collaborative filtering method can be proposed.

In a collaborative filtering method based on a generative adversarial network (GAN), training a generative model and a discriminative model of the GAN in real-valued vector-wise for an item ; And it provides a GAN-based collaborative filtering method comprising the step of recommending a personalized item list to the user by using the trained generation model and the discrimination model.

According to one aspect, the generation model generates a purchase vector composed of real values, and the discrimination model is whether a given purchase vector is an actual purchase vector corresponding to ground truth or a purchase vector generated from the generation model. Can be identified.

According to another aspect, when a user-specific condition vector and a random noise vector are given as real value vectors, the generation model generates an n-dimensional purchase vector, and the discrimination model is adjusted by the user-specific condition vector to generate the It can be trained to distinguish the obtained purchase vector from the actual purchase vector of the corresponding user.

According to another aspect, the training comprises: masking the generated purchase vector, which is an output of the generated model, using an n-dimensional indicator vector specifying whether the generated model has purchased each item by a user. It may include.

According to another aspect, the generation model is a neural network that outputs an n-dimensional purchase vector based on the user-specific condition vector and the random noise vector, and the discrimination model has its own input and determines the probability that the corresponding input is actual data. It is a neural network that outputs a single scalar value, and the training step includes the generation model and the discrimination through stochastic gradient descent using mini-batch and back-propagation. You can train the model.

According to another aspect, the purchase vector generated by the generation model includes a user's prediction preference score for all items in a data set, and the recommending step includes the top N number of prediction preference scores in the data set. You can select an item and recommend it.

According to another aspect, in the training step, after estimating an item that the user has not purchased as a negative item, the generated model may be trained so that the value of the negative item approaches zero.

According to another aspect, the training may train the generation model through a method of minimizing an objective function of the generated model by using the negative item for zero-reconstruction normalization.

According to another aspect, the training comprises using the item purchased by the user and the negative item as input values of the discrimination model, and outputting the negative item with respect to the loss slope of the discrimination model again in the generating model. The generation model can be trained through a method of passing to.

According to another aspect, the training includes a method of minimizing an objective function of the generated model by using the negative item for zero-reconstruction normalization, and inputting the item purchased by the user and the negative item to the discrimination model The generation model may be trained through a hybrid method in which a method of using as a value and transmitting the output of the negative item back to the generation model with respect to the loss slope of the discrimination model is combined.

It is combined with a computer to provide a computer program stored in a computer-readable recording medium in order to execute the GAN-based collaborative filtering method on the computer.

It provides a computer-readable recording medium, characterized in that a program for executing the GAN-based collaborative filtering method on a computer is recorded.

A computer system, comprising: a memory; And at least one processor connected to the memory and configured to execute computer readable instructions contained in the memory, wherein the at least one processor comprises: real-valued vector-wise for the item A computer, characterized in that it trains a generative model and a discriminative model of a generative adversarial network (GAN), and recommends a personalized item list to a user using the trained generation model and discriminant model. System.

According to embodiments of the present invention, by using real-valued vector-wise hostile training for a GAN-based collaborative filtering method, it is possible to effectively solve problems that occur when using hostile training and By taking full advantage of this, you can increase the accuracy of recommendations for collaborative filtering.

1 shows an example of a training process for generating a discrete item index.
Figure 2 is a visualization of the learning trend of IRGAN
3 is a visualization of the learning trend of GraphGAN.
4 shows the ratio of contradictory label items among items generated by the generation model G as hostile training progresses.
5 shows an overview of a GAN-based CF framework (CFGAN) according to an embodiment of the present invention.
6 shows an example of a learning algorithm for CFGAN_ZP according to an embodiment of the present invention.
7 is a block diagram illustrating an example of an internal configuration of a computer system according to an embodiment of the present invention.

Hereinafter, embodiments of the present invention will be described in detail with reference to the accompanying drawings.

Collaborative filtering (CF) is the most representative of several approaches to the recommendation algorithm. Collaborative filtering analyzes the user's taste and taste of products based on the user's rating, and then predicts the ratings to be given to products that the user has not yet purchased or the top N items worth purchasing.

The present invention proposes a new GAN-based collaborative filtering (CF) framework that provides higher accuracy for recommendations. First, the fundamental problem of the existing method for CF is identified, and the problem of the existing method is quantitatively highlighted through various experiments. Next, a new direction of vector-wise hostile training is proposed, and a GAN-based CF framework (CFGAN) is proposed based on this direction. In addition, we identify unique challenges (problems) that arise when vector-level hostile training is adopted for CF, and propose three CF methods that can solve these problems.

The reason why the benefits of hostile training in existing CFs have not been fully exploited stems from the way that G samples a single discrete item index. A single discrete item index is completely different from the G of the original GAN, which creates a vector (eg, image) with elements of consecutive values. Since the item index is discrete and finite, it is highly likely that G will sample exactly the same data as the actual data (i.e., the user's actual selection). Therefore, as training proceeds, D becomes confused, and thus the grade degrades because contradicting labels for the same item index are trained at a large rate. When the same item comes from G, this same item may sometimes be marked as "fake", and other cases from actual data may be marked as "real". As a result, D sends the wrong signal to G, reducing the likelihood that G will sample the actual data item. In other words, instead of improving each other's abilities, the accuracy of G and D worsens. After several iterations are performed, the accuracy of recommendation by D does not increase any more, but rather declines. This is due to the high percentage of items that are labeled contradictoryly. This prevents G from improving accuracy in CFs where competition between G and D does not help each other.

In order to solve the above problem, the present invention proposes a new GAN-based CF framework. In the framework according to the present invention, if a user is given, G attempts to generate a plausible purchase vector composed of real (real) value elements rather than sampling a single item index that the user may be interested in and can purchase. Similarly, D attempts to distinguish between the actual purchase vector obtained from the measured data and the generated purchase vector. This approach of generating and discriminating real vectors eliminates the confusion of D that arises in GraphGAN and IRGAN due to items with contradictory labels, since it is almost impossible to create a real vector that is exactly the same as the vector in which G belongs to the measured data. prevent. This property makes D closer to the actual purchase vector in the process of repeating the generated vector of G by guiding G through back propagation and continuously improving it. This GAN-based CF framework induces successful adversarial training to achieve high-accuracy recommendations.

When switching the target from discrete item index generation to real value vector generation for GAN-based CF, we identify unique challenges that naturally appear due to the nature of implicit feedback data. Unlike the image data considered by the original GAN (represented as a high-density vector composed of pixels with multiple bit depths), the measured data of the CF is a sparse single value vector. Here, when the user-item interaction is observed (ie, the user purchases the item), the element (element) is 1. G attempts to create a purchase vector that is similar to the actual data, but can easily get away by finding a useful and easy solution: that is, it creates a purchase vector with all elements of 1 without taking into account the user's relative preferences.

In relation to this problem, we propose three CF methods (ie, CFGAN-ZR, CFGAN-PM and CFGAN-ZP) developed in the framework according to the present invention. Each method successfully solves the problem in its own way. The commonalities of these methods are as follows: (1) samples the elements that were not observed; (2) We estimate that these methods will respond to negative feedback (ie, the values of these methods are not missing from the purchase vector and are zero); (3) It is possible to request G to generate an output close to 0 from the sampled element and to generate a 1 from the element corresponding to the user's actual preference (i.e., measured data). For this goal, CFGAN-ZR uses zero reconstruction normalization, and CFGAN-PM considers partial masking; CFGAN-ZP is a hybrid that uses both zero reconstruction normalization and partial masking. In addition, the framework according to the present invention formulates each of these methods according to a user-oriented and item-oriented approach.

In summary, the main contents of the present invention are as follows:

(1) Check the limitations of the existing GAN-based CF method and emphasize it through preliminary experiments.

(2) In order to improve the recommendation accuracy of CF, we propose a new direction that the GAN-based CF method should follow in order to make the most of the benefits of hostile training, namely, hostile training in units of actual values.

(3) We propose a new direction GAN-based CF framework. That is, it presents the direction of hostile training in units of actual values.

(4) We propose three CF methods, CFGAN-ZR, CFGAN-PM, and CFGAN-ZP realized in a GAN-based CF framework. CF methods successfully solve the problem that arises when using vector-wise hostile training for CF.

(5) Validate the validity of the proposed direction for vector unit hostile training, and prove the superiority of the CF method compared to the state-of-the-art top-N recommendation.

Hereinafter, the preliminary work on the GAN-based CF method will be described.

Model-based CF

및

각각은 m명의 사용자와 n개 아이템 세트를 나타낸다. 여기서, I_u를 u로부터 암시적 피드백을 수신하는(즉, 구매) 아이템 세트로 정의한다. 아이템에 대한 사용자의 암시적 피드백은 희소 행렬

에 의해 표현된다. 이때, 원소 r_ui은 i가 I_u에 포함된 경우 1이며 다른 경우에는 비어 있다(즉, 관찰되지 않음). 사용자 u의 구매 벡터와 아이템 I 각각을 r_u와 r_i로 나타낸다.

And

Each represents m users and n item sets. Here, I _u is defined as a set of items that receive implicit feedback from u (ie, purchase). The user's implicit feedback on the item is sparse.

Is represented by In this case, the element r _ui is 1 if i is included in I _u , and is empty in other cases (ie, not observed). User u's purchase vector and item I are denoted by r _u and r _i respectively.

로 나타냄,

)에 대한 u의 선호도 지수(u가 i를 구매하고자 하는 정도를 반영)를 예측하고, u에 대하여 가장 높은 예측 지수를 가지는 아이템의 서브 세트를 추천한다. 모델 기반 방법의 관점으로부터, CF는

와 같이 공식화하여 나타낼 수 있는 모델을 사용하여

를 계산한다. 여기서, f는 모델을 나타낸다. 즉,

에 의해 파라미터화된 함수이며, 주어진 사용자 u 및 아이템 i의 쌍을 선호 지수로 매핑한다. 모델 파라미터 세트

는

을 트레이닝 데이터로 이용한 목적 함수를 최적화함으로써 학습될 수 있다.The goal of CF is j(

Represented by,

) For u (reflecting the degree to which u wants to purchase i), and recommends a subset of items with the highest prediction index for u. From a model-based method perspective, CF is

Using a model that can be formulated and represented as

Calculate Here, f represents the model. In other words,

It is a function parameterized by and maps a given pair of user u and item i to a preference index. Model parameter set

Is

Can be learned by optimizing the objective function using as training data.

및

의 오리지널 값인 r_ui 사이의 평균 제곱 에러를 최소화시키는 것을 목적으로 한다. 반면에, 쌍 단위 방법은 더 선호되는(즉, 구매된) 아이템과 덜 선호되는 아이템(즉, 구매되지 않은) 사이의 예측된 선호도 차이를 최대화시킨다.Existing model-based CF methods can be classified as follows according to the type of objective function used to train the model: point wise and pair wise. Point unit method

And

The aim is to minimize the mean square error between the original values of r _ui . On the other hand, the pairwise method maximizes the predicted preference difference between a more preferred (ie, purchased) item and a less preferred (ie, not purchased) item.

GAN Based CF( GAN -based CF)

Recently, generative adversarial neural networks (GANs), which propose a new method of learning machine learning models, are being studied. GAN is a technology that makes G generate plausible data by training G through a competitive structure between the generative model (G) and the discriminant model (D). In other words, through the competitive process including G and D, G is learned to capture the distribution of measured data, so that plausible synthetic data that is not different from the measured data can be generated. Officially, G and D use the objective function V to play the following two-player minimal-maximization game (Equation 1).

[Equation 1]

및

는 각각 G 및 D의 모델 파라미터를 나타낸다. x는 데이터 분포

으로부터의 실측 자료인 반면,

는 전형적으로

에 의해 유도된 모델 분포

로부터의 합성 데이터이다. z는 랜덤 노이즈 벡터이고,

는 입력이 자체 입력이 실측 자료일 추정된 확률을 나타낸다. 일반적으로, G와 D는 모두 신경망이며 동일한 목적 함수를 최소화하고 최대화하면서 각각의 모델 파라미터를 최적화하는 데 반복된다.here,

And

Represents the model parameters of G and D, respectively. x is the data distribution

While this is the actual data from

Is typically

Model distribution induced by

Synthetic data from z is the random noise vector,

Represents the estimated probability that the input is the actual data. In general, both G and D are neural networks and iterate to optimize each model parameter while minimizing and maximizing the same objective function.

The most successful field of GAN is image generation, but it has great potential to achieve more promising results in other fields, such as WaveGAN for music generation and SegGAN for sentence generation. Recently, methods of applying GAN to CF are being studied in the field of recommendation systems, and in order to train the CF model, a hostile training method is applied instead of optimizing a point-wise or pair-wise objective function. And try to get satisfactory accuracy in recommendations.

For example, as an example of GAN applied to the recommendation system, IRGAN (Jun Wang, Lantao Yu, Weinan Zhang, Yu Gong, Yinghui Xu, Benyou Wang, Peng Zhang, and Dell Zhang. 2017. Irgan: A minimax game for unifying generative and discriminative) Information retrieval models.) based on the following principles.

In IRGAN, G attempts to create (or sample) an index of items related to a given user, and D attempts to distinguish the user's measured data item from items that G has created comprehensively. More specifically, G is implemented by the softmax function of Equation 2.

[Equation 2]

는 주어진 u 및 인덱스 k에서 아이템을 샘플링하는 G의 조건부 확률 분포이고,

는 u가

를 구입할 확률을 반영하는 스코어링 함수이다.here,

Is the conditional probability distribution of G sampling the item at a given u and index k,

Is u

It is a scoring function that reflects the probability of purchasing.

를 사용하여, 주어진 사용자 u의 실측 자료에 속하는 각 아이템 i의 확률을 추정한다(수학식 3).On the other hand, D is the sigmoid function of the scoring function

By using, the probability of each item i belonging to the actual measured data of a given user u is estimated (Equation 3).

[Equation 3]

는 i와 u의 관련성을 반영한다. 스코어링 함수

와

는 모두 행렬 인수 분해 모델, 공식적으로,

와

에 기반한다. 이때, d_u와 d_i는 각각 D에 속하는 잠재(latent) 요인(factor)이다. 그리고, g_u와 g_i는 G에 속하는 잠재 요인이다.

와

는 아이템 바이어스 조건이다.here,

Reflects the relationship between i and u. Scoring function

Wow

Are all matrix factorization models, formally,

Wow

Is based on. In this case, d _u and d _i are latent factors belonging to D, respectively. And, g _u and g _i are latent factors belonging to G.

Wow

Is the item bias condition.

은 수학식 4와같이 공식화된다:Finally, the objective function of IRGAN

Is formulated as in Equation 4:

[Equation 4]

D can be solved by stochastic slope elevation, but G cannot receive slope from D because sampling of i is discrete. Therefore, IRGAN uses policy gradient-based reinforcement learning to update G.

Although not necessarily a pure CF method, there are other research findings that focus on the GAN-based recommendation system, especially the visual aspect of the item. GAN can be used for recommendation work, but additional visual information of the item is previously required.

Hereinafter, limitations of the existing method are presented together with a preliminary result set, and then a new training method for GAN-based CF will be described.

Despite the possibility that GAN can succeed in CF, there is still room for improvement. The limitations of the "Discrete Item Index Generation" method are explained and we want to overcome these limitations to improve the accuracy of the CF.

1 shows an example of a training process for generating a discrete item index.

Referring to Fig. 1, in the initial step (a), G samples items (almost) randomly, but improves its own performance according to D's guiding (b). As training progresses further (c), G frequently samples items exactly the same as the measured data.

From D's point of view, the same item index (e.g. i ₃ ) is labeled "fake" (in G) and labeled "real" (i.e. from measured data), making D significantly confusing, As a result, D declines. This means that D can transmit an erroneous signal to G through the policy slope, and G can lower the likelihood of sampling an item of measured data, resulting in a deterioration in the quality of G.

In order to understand the degree of degradation of G due to the aforementioned limitations, we conduct preliminary experiments on the actual data set.

As an experiment example, MovieLens 100K, Watcha data is used. The previously known product recommendation model and the model of the present invention are trained using training data, respectively, and the top N recommendation accuracy of each model is calculated and compared using evaluation data. The scale for calculating the accuracy was the widely used precision and MRR, and the results for the top five (@5) were calculated.

2 is a visualization of the learning tendency of IRGAN, Figure 3 is a visualization of the learning trend of GraphGAN. Each graph of FIGS. 2 and 3 reports the accuracy of the top five recommendations. 4 shows the proportion of items that are labeled contradictoryly among items generated by G as hostile training proceeds.

As represented in Figures 2 and 3, the training seems to work well in the early stages. G samples experimental items, pushes D to its own limit, and D guides G. However, after this step, the recommendation accuracy by D does not increase after a sudden drop. This is mainly because there are many contradictory label items in the training data (see FIG. 4). Eventually, G can no longer perform as the reward from D can no longer give useful feedback to G, and even performance degrades. As a result, in this situation, the competition process between G and D cannot create a synergy effect when providing a high-quality recommendation list.

The present invention is intended to fully enjoy the benefits of hostile training for item recommendation tasks, how to properly train G and D so that G consistently pushes D to its own limit without pushing "well beyond" its own limit. In order to solve whether D can consistently give only beneficial feedback to G, "vector unit training" is proposed.

In the present invention, instead of a single item index, G and D perform generation and discrimination of a vector unit composed of real values. In other words, in the present invention, G creates a real value vector, and D distinguishes between G and an actual vector value coming from the measured data. More specifically, given a user, G attempts to generate a plausible purchase vector while D attempts to accurately determine whether the given purchase vector is real or a fake generated by G. For vector-wise training, G can produce a purchase vector that is very similar (but not the same) as the actual purchase, but it is unlikely to generate a purchase vector that is exactly the same as the actual purchase. This makes D rarely cheated by G, allowing D to consistently break out of its limits in a series of steps due to G's training data for which it is qualified. Likewise, D can continue to provide useful training signals to G that show how G needs to slightly change its own output in order to create a more realistic fake purchase vector via backpropagation. The vector unit training proposed in the present invention can successfully perform hostile training that performs high-quality recommendation by fully utilizing the advantages of GAN.

Hereinafter, a GAN-based CF framework (CFGAN) according to the present invention will be described.

5 shows an overview of a GAN-based CF framework (CFGAN) according to an embodiment of the present invention.

The CFGAN of the present invention is based on the conditional GAN framework.

를 생성한다. 구매 벡터

는 희소 벡터가 될 것으로 예상되며, 여기서 u의 구매 아이템에 해당하는 모든 요소

(

)는 희망적으로 1이다. 마찬가지로, c_u에 의해 조절된 D는 생성된 구매 벡터를 u의 실제 벡터와 구별하도록 트레이닝 된다. 공식적으로,

로 표시되는 D의 목적 함수는 수학식 5와 같다.Referring to FIG. 5, both G and D are user conditions, and the model parameters are learned while considering individual settings of each user. Given a user-specific condition vector c _u and a random noise vector z ² , CFGAN's G is an n-dimensional purchase vector

Create Purchase vector

Is expected to be a sparse vector, where all elements corresponding to the purchase item in u

(

) Is hopefully 1. Similarly, D adjusted by c _u is trained to distinguish the generated purchase vector from the actual vector of u. officially,

The objective function of D represented by is the same as in Equation 5.

[Equation 5]

Likewise, G is equal to Equation 6.

[Equation 6]

는 u가 각 아이템

를 구입했는지 여부를 특정하는 n-차원 표시기 벡터이며,

는 element-wise 곱셈을 나타낸다.In Equation 5 and Equation 6

Is u for each item

Is an n-dimensional indicator vector specifying whether or not

Denotes element-wise multiplication.

를 곱하여 G의 출력

를 마스크함으로써 G가 모방하려고 하는 실측 자료의 희소성을 처리한다는 것이다. 이렇게 함으로써 구매된 아이템들에 대한 G의 출력만이 G와 D의 학습에 기여할 수 있다. 즉, 구매되지 않은 아이템들에 대응하는 G의 출력 노드들을 드롭함으로써, (1) D는 드롭된 노드들을 무시하고, (2) G는 드롭된 노드에 관한 D로부터 손실의 기울기를 얻지 못한다. 이는 잠재적인 구매를 식별하는 데 나중에 사용될 수 있는 관찰된 사용자 아이템 구매 기록을 기반으로 모델을 작성한다는 점에서 포인트 단위 목적 함수를 사용하는 다른 CF들과 공통된다.The main difference between the CFGAN of the present invention and the existing GAN is that G

The output of G by multiplying

By masking, G handles the scarcity of the actual data that G is trying to imitate. By doing this, only the output of G for purchased items can contribute to the learning of G and D. That is, by dropping the output nodes of G corresponding to items not purchased, (1) D ignores the dropped nodes, and (2) G does not get the slope of the loss from D relative to the dropped node. This is common with other CFs that use a point-by-point objective function in that it builds a model based on the observed user item purchase history that can be used later to identify potential purchases.

와

로 파라미터화된 다층 신경망으로 G와 D를 구현한다. 특히, G는 입력

가

를 출력하는 L^G-레이어(L^G> = 2) 신경망이다. 여기서, {}는 두 벡터의 내부 연결을 나타낸다. D는

와 연결된

(즉, G로부터) 또는

(실측 자료로부터)의 입력 여부에 관계없는 입력을 가지는 L^D-레이어(L^D> = 2) 신경망이며, 이 신경망은 자체 입력이 G보다는 실측 자료로부터 오는 확률을 나타내는 단일 스칼라 값을 출력한다. G와 D를 트레이닝시키기 위해 미니배치(minibatch)와 역-전파(back-propagation)를 이용한 확률적 기울기 하강법(stochastic gradient descent)을 사용한다. 각각의 파라미터

와

를 번갈아 가며 업데이트하며, 하나가 업데이트되는 경우 다른 하나를 유지한다.In the present invention, G and D are each

Wow

We implement G and D with a multilayered neural network parameterized with. Specifically, G is the input

end

It is an L ^G -layer (L ^G > = 2) neural network that outputs. Here, {} represents the internal connection of two vectors. D is

Connected with

(I.e. from G) or

It is an L ^D -layer (L ^D > = 2) neural network that has an input regardless of whether it is input (from measured data), and this neural network outputs a single scalar value representing the probability that its own input comes from measured data rather than G. To train G and D, we use stochastic gradient descent using minibatch and back-propagation. Each parameter

Wow

It updates alternately, and keeps the other if one is updated.

가 공급되면 G는 조밀한 벡터

를 생성한다.

에는 데이터 세트의 모든 아이템들에 대한 예측 선호도 스코어가 포함될 수 있다.

의 j 번째 원소를 각 아이템 j(

)에 대한 사용자 u의 예측 선호도 스코어로 사용한다. 마지막으로, 예측된 스코어가 가장 높은 상위 N개 아이템을 선택하여 u로 추천한다.After hostile training is complete, z and

When is supplied, G is a dense vector

Create

May include predicted preference scores for all items in the data set.

J-th element of each item j(

) For user u's prediction preference score. Finally, the top N items with the highest predicted scores are selected and recommended as u.

The CF methods (user-oriented method and item-oriented method) realized on CFGAN are as follows.

The CFGAN of the present invention presents the direction of vector unit training and faces a unique problem that the existing GAN aiming at image generation does not face. In general, unlike image data expressed as a vector composed of pixels with multiple bit depths, the implicit feedback data for a user in recommendation is a very sparse single value vector, and the element of this vector is the corresponding user-item. 1 if the interaction (i.e. purchase) is observed, otherwise it is empty. G tries to trick D by creating a purchase vector similar to the actual data, but it will lead to a trivial solution that simply predicts all the outputs of the purchase vector as 1. This solution is easy to find, but not useful. This is because it cannot capture the relative preferences of users, which are the basic elements of CF.

)을 무작위로 선택하고, 해당 아이템을 음수(negative) 아이템으로 추정한 후, 해당 피드백이 누락되지 않았지만 0인 것을 나타낸다. 그런 다음, 음수 아이템의 값이 0에 가까워지도록, G가 사용자의 구매 벡터를 생성하도록 트레이닝한다. 이렇게 함으로써, G가 D를 기만하기 위해 구매된 아이템에 대해 1에 가까운 값을 생성하는 것에 초점을 맞추고, 음수 아이템에 대한 값을 산출하는 것을 고려할 수 있다. 결과적으로, 구매 벡터의 모든 요소에 대해 1을 단순 출력하는 G의 솔루션을 수용하는 것을 피하게 된다.The core idea shared by the different CF methods realized on the CFGAN of the present invention is as follows: In every training iteration (e.g., t), the percentage of items not purchased by each user (

) Is selected at random, and the corresponding item is estimated as a negative item, indicating that the corresponding feedback is not missing but is 0. Then, G is trained to generate the user's purchase vector so that the value of the negative item approaches zero. By doing so, we can focus on generating a value close to 1 for items purchased for G to deceive D, and consider calculating a value for negative items. As a result, we avoid accepting G's solution, which simply outputs 1 for every element of the purchase vector.

는 t 번째 반복에서 샘플링된 음수 아이템 세트를 나타낸다. 구매 행렬

의 해당 항목은 누락되지 않은 것으로 간주되지만, 이 반복에서만 0으로 간주된다. S는

로부터 식별된 아이템의 비율(백분율)을 나타낸다. 예를 들어, S가 30으로 설정되면,

의 아이템들 중 30%가 선택된다.Formally,

Represents a set of negative items sampled in the t-th iteration. Purchase procession

Corresponding items of are considered not to be missing, but are considered zero only in this iteration. S is

Represents the percentage (percentage) of the items identified from. For example, if S is set to 30,

30% of the items are selected.

Depending on how the negative items identified in the CFGAN framework are used, three CF methods are proposed: a method based on zero-reconstructive normalization (CFGAN-ZR), and a method based on partial-masking (CFGAN-PM). And a hybrid method in which the above two are combined (CFGAN-ZP).

CFGAN - ZR

를 나타내며, 비 구매 아이템으로부터 선택된 음수 아이템의 비율(백분율)을 특정하는 파라미터인 S를 지시하기 위해

를 나타낸다. 수학식 6에서 정의된 G의 목적 함수에 더하여 수학식 7을 통해 재구성 손실 함수를 정규화 항으로 사용한다.CFGAN-ZR is used to indicate the selected negative item.

And to indicate S, a parameter that specifies the ratio (percentage) of negative items selected from non-purchased items

Represents. In addition to the objective function of G defined in Equation 6, the reconstruction loss function is used as a normalization term through Equation 7.

[Equation 7]

의 두 번째 항은

와

가 전체 방정식에서 ZR 항의 중요성을 제어하기 위한 조정 가능한 파라미터인 제로-재구성(즉, ZR) 정규화에 해당한다.

를 최소화함으로써 G는 관찰된 항목들에 대해 1에 가까운 스코어를 생성하고, 관찰되지 않은 항목들에 대해 낮은 값들(즉, 0에 가깝게)을 생성하여 G가 사소하고 무용한 솔루션에 도달하는 것을 방지하는데 집중할 수 있다.In Equation 7

The second term in

Wow

Corresponds to zero-reconstruction (i.e., ZR) normalization, which is an adjustable parameter to control the importance of the ZR term in the overall equation.

By minimizing G, G produces a score close to 1 for observed items and low values (i.e. close to 0) for items that are not observed, preventing G from reaching a trivial and useless solution. I can focus on doing it.

CFGAN -PM

와

를 사용하여 선택된 음수 아이템과 샘플링 파라미터 S를 각각 나타낸다. CFGAN-PM에서 비구매 아이템에 해당하는 모든 출력 노드를 반드시 드롭하지는 않지만,

에 해당하는 출력 노드는 그대로 유지한다. 이로써, D는 구매된 아이템뿐만 아니라 판별 작업의 음수 아이템으로부터의 입력 값을 활용한다. 또한, D로부터의 손실의 기울기에 대하여 선택된 음수 아이템에 대한 출력을 G로 다시 전달할 수 있다. 결과적으로, G는 구매 아이템에 대해 생성된 출력을 1에 가깝게 만들고, 선택된 음수 아이템에 대해 0에 더 가깝게 만든다.Symbol in CFGAN-PM

Wow

Denote the selected negative item and sampling parameter S, respectively. CFGAN-PM does not necessarily drop all output nodes corresponding to non-purchased items,

The output node corresponding to is kept as it is. Thus, D utilizes not only the purchased item, but also the input value from the negative item of the discrimination operation. Also, the output for the negative item selected for the slope of the loss from D can be passed back to G. As a result, G makes the output generated for the purchased item closer to 1 and closer to 0 for the selected negative item.

와 수학식 9의

에 적용된다.The above idea is in Equation 8

And of Equation 9

Applies to

[Equation 8]

[Equation 9]

는 n차원 지시자 벡터이며,

이면

이고, 그렇지 않으면

이다.In Equation 8 and Equation 9

Is an n-dimensional indicator vector,

Back side

And otherwise

to be.

CFGAN - ZP

및

는 각각 제로 재구성 및 부분 마스킹에 사용된다. 여기서, D의 목적 함수는 G의 목적 함수를 나타내는 수학식 10과 동일하다.CFGAN-ZP is a hybrid method that can take advantage of both methods by using both zero reconstruction normalization and partial masking. A set of selected negative items

And

Is used for zero reconstruction and partial masking, respectively. Here, the objective function of D is the same as Equation 10 representing the objective function of G.

[Equation 10]

The learning algorithm for CFGAN_ZP is the same as Algorithm 1 of FIG. 6. CFGAN_ZR and CFGAN_PM are regarded as special cases of CFGAN_ZP, CFGAN_ZR corresponds to CFGAN_ZP skipping line 5 of Algorithm 1, and CFGAN_PM corresponds to CFGAN_ZP skipping line 4 of Algorithm 1.

In the above, CFGAN is described in terms of a user-oriented method that considers each user as a training instance. Here, each training instance corresponds to a user, and G (and D) creates (and determines) a purchase vector of the user, but the training instance can also implement CFGAN in an item-oriented manner corresponding to an item other than the user.

Item-oriented CFGAN

G, conditioned by item-specific information, attempts to generate a plausible purchase vector of each item, i.e., an m-dimensional vector. Here, m is the number of users, which indicates how likely a user corresponding to each element of the vector is to purchase an item. Similarly, D, conditioned according to item-specific information, attempts to distinguish the purchase vector of the item's actual measurement data from the vector generated by G. In addition, the agent attempts to select a negative (negative) user for each item representing people who are unlikely to purchase the item. After training, G predicts a preference score for each given item i and all users. Finally, N items with the highest prediction score are recommended to each user.

Each of the CF methods (CFGAN_ZR, CFGAN_PM, and CFGAN_ZP) may be implemented using an item-oriented method as well as a user-oriented method.

System implementation

7 is a block diagram illustrating an example of an internal configuration of a computer system according to an embodiment of the present invention. For example, a GAN-based CF system according to embodiments of the present invention may be implemented through the computer system 700 of FIG. 7. As shown in FIG. 7, the computer system 700 is a component for executing the above-described GAN-based CF method, and includes a processor 710, a memory 720, a permanent storage device 730, a bus 740, and input/output. An interface 750 and a network interface 760 may be included.

Processor 710 may include or be part of any device capable of processing a sequence of instructions as a component for a GAN-based CF. The processor 710 may include, for example, a processor and/or a digital processor in a computer processor, mobile device or other electronic device. The processor 710 may be included, for example, a server computing device, a server computer, a series of server computers, a server farm, a cloud computer, a content platform, and the like. The processor 710 may be connected to the memory 720 through the bus 740.

Memory 720 may include volatile memory, permanent, virtual or other memory for storing information used by or output by computer system 700. The memory 720 may include, for example, random access memory (RAM) and/or dynamic RAM (DRAM). Memory 720 may be used to store any information, such as state information of computer system 700. The memory 720 may also be used to store instructions of the computer system 700, including instructions for, for example, GAN-based CF. Computer system 700 may include one or more processors 710 as needed or where appropriate.

Bus 740 may include a communication infrastructure that enables interaction between various components of computer system 700. Bus 740 may carry data between components of computer system 700, for example, between processor 710 and memory 720, for example. Bus 740 may include wireless and/or wired communication media between components of computer system 700 and may include parallel, serial, or other topological arrangements.

Persistent storage device 730 is a component such as a memory or other persistent storage device as used by computer system 700 to store data for a predetermined extended period of time (e.g., compared to memory 720). Can include. Persistent storage device 730 may include non-volatile main memory as used by processor 710 in computer system 700. The persistent storage device 730 may include, for example, a flash memory, a hard disk, an optical disk, or other computer-readable medium.

The input/output interface 750 may include interfaces to a keyboard, a mouse, voice command input, a display, or other input or output device. Configuration commands and/or inputs for the GAN-based CF may be received through the input/output interface 750.

The network interface 760 may include one or more interfaces to networks such as a local area network or the Internet. The network interface 760 may include interfaces for wired or wireless connections. Configuration commands and/or input for a GAN-based CF may be received through the network interface 760.

Further, in other embodiments, the computer system 700 may include more components than the components of FIG. 7. However, there is no need to clearly show most of the prior art components. For example, the computer system 700 may be implemented to include at least some of the input/output devices connected to the input/output interface 750 described above, or a transceiver, a global positioning system (GPS) module, a camera, various sensors, Other components such as a database may be further included.

As described above, according to embodiments of the present invention, by using real-valued vector-wise hostile training for the GAN-based collaborative filtering method, it is possible to effectively solve the problem that occurs when hostile training is used, and hostile training. You can increase the recommendation accuracy of collaborative filtering by taking full advantage of the benefits of.

The apparatus described above may be implemented as a hardware component, a software component, and/or a combination of a hardware component and a software component. For example, the devices and components described in the embodiments are a processor, a controller, an arithmetic logic unit (ALU), a digital signal processor, a microcomputer, a field programmable gate array (FPGA), a programmable gate array (PLU). It may be implemented using one or more general purpose computers or special purpose computers, such as a logic unit), a microprocessor, or any other device capable of executing and responding to instructions. The processing device may execute an operating system (OS) and one or more software applications executed on the operating system. In addition, the processing device may access, store, manipulate, process, and generate data in response to the execution of software. For the convenience of understanding, although it is sometimes described that one processing device is used, one of ordinary skill in the art, the processing device is a plurality of processing elements and/or a plurality of types of processing elements. It can be seen that it may include. For example, the processing device may include a plurality of processors or one processor and one controller. In addition, other processing configurations are possible, such as a parallel processor.

The software may include a computer program, code, instructions, or a combination of one or more of these, configuring the processing unit to behave as desired or processed independently or collectively. You can command the device. Software and/or data may be embodyed in any type of machine, component, physical device, computer storage medium or device to be interpreted by the processing device or to provide instructions or data to the processing device. have. The software may be distributed over networked computer systems and stored or executed in a distributed manner. Software and data may be stored on one or more computer-readable recording media.

The method according to the embodiment may be implemented in the form of program instructions that can be executed through various computer means and recorded in a computer-readable medium. In this case, the medium may be one that continuously stores a program executable by a computer, or temporarily stores a program for execution or download. In addition, the medium may be a variety of recording means or storage means in a form in which a single or several pieces of hardware are combined, but is not limited to a medium directly connected to a computer system, but may be distributed on a network. Examples of media include magnetic media such as hard disks, floppy disks, and magnetic tapes, optical recording media such as CD-ROMs and DVDs, magnetic-optical media such as floptical disks, and And ROM, RAM, flash memory, and the like may be configured to store program instructions. In addition, examples of other media include an app store that distributes applications, a site that supplies or distributes various software, and a recording medium or a storage medium managed by a server.

As described above, although the embodiments have been described by the limited embodiments and drawings, various modifications and variations are possible from the above description by those of ordinary skill in the art. For example, the described techniques are performed in a different order from the described method, and/or components such as a system, structure, device, circuit, etc. described are combined or combined in a form different from the described method, or other components Alternatively, even if substituted or substituted by an equivalent, an appropriate result can be achieved.

Therefore, other implementations, other embodiments, and claims and equivalents fall within the scope of the claims to be described later.

Claims (20)

In a collaborative filtering method based on a generative adversarial network (GAN) executed in a computer system,
The computer system includes at least one processor configured to execute computer readable instructions contained in a memory,
Collaborative filtering method based on the GAN,
Training, by the at least one processor, a generative model and a discriminative model of the GAN in a real-valued vector-wise for an item; And
Recommending, by the at least one processor, a personalized item list to a user using the trained generation model and the discrimination model
Including,
The generation model generates an n-dimensional purchase vector consisting of real values,
The discrimination model is trained to distinguish the generated purchase vector from an actual purchase vector corresponding to ground truth,
The training step,
Instead of a single item index, training is performed in a vector unit consisting of real values,
Masking the generated purchase vector, which is an output of the generation model, using an n-dimensional indicator vector specifying whether the generated model has purchased each item by a user
Collaborative filtering method based on GAN including a. The method of claim 1,
The discrimination model is to determine whether a given purchase vector is the actual purchase vector or a purchase vector generated from the generation model.
Collaborative filtering method based on GAN, characterized in that. The method of claim 1,
Given a user-specific condition vector and a random noise vector as a real value vector,
The discrimination model is adjusted by the user-specific condition vector and trained to distinguish the generated purchase vector from the actual purchase vector of the corresponding user.
Collaborative filtering method based on GAN, characterized in that. delete The method of claim 3,
The generation model is a neural network that outputs an n-dimensional purchase vector based on the user specific condition vector and the random noise vector,
The discriminant model is a neural network that has its own input and outputs a single scalar value indicating the probability that the input is actually measured data,
The training step,
Training the generation model and the discriminant model through stochastic gradient descent using mini-batch and back-propagation
Collaborative filtering method based on GAN, characterized in that. The method of claim 1,
The purchase vector generated by the generation model includes the predicted preference score of the user for all items in the data set,
The recommending step,
Selecting and recommending the top N items with the high prediction preference score from the data set
Collaborative filtering method based on GAN, characterized in that. The method of claim 1,
The training step,
After estimating an item not purchased by the user as a negative item, training the generation model so that the value of the negative item approaches 0 (zero)
Collaborative filtering method based on GAN, characterized in that. The method of claim 7,
The training step,
Training the generative model through a method of minimizing the objective function of the generative model by using the negative item for zero-reconfiguration normalization
Collaborative filtering method based on GAN, characterized in that. The method of claim 7,
The training step,
Training the generation model by using the item purchased by the user and the negative item as input values of the discrimination model and transmitting the output of the negative item back to the generating model with respect to the loss slope of the discrimination model that
Collaborative filtering method based on GAN, characterized in that. The method of claim 7,
The training step,
A method of minimizing the objective function of the generated model by using the negative item for zero-reconfiguration normalization, and using the item purchased by the user and the negative item as an input value of the discrimination model, and for the loss slope of the discrimination model Training the generation model through a hybrid method in which the method of transmitting the output of the negative item back to the generation model is combined
Collaborative filtering method based on GAN, characterized in that. A computer program combined with a computer and stored in a computer-readable recording medium for executing the GAN-based collaborative filtering method of any one of claims 1 to 3 and 5 to 10 on a computer. A computer-readable recording medium, characterized in that a program for executing the GAN-based collaborative filtering method according to any one of claims 1 to 3 and 5 to 10 is recorded on a computer. In a computer system,
Memory; And
At least one processor coupled to the memory and configured to execute computer readable instructions contained in the memory
Including,
The at least one processor,
Train a generative model and a discriminative model of a GAN Generative Adversarial Network (GAN) in real-valued vector-wise for an item,
Recommend a personalized item list to the user using the trained generation model and discrimination model,
The generation model generates an n-dimensional purchase vector consisting of real values,
The discrimination model is trained to distinguish the generated purchase vector from an actual purchase vector corresponding to ground truth,
The at least one processor,
Instead of a single item index, training is performed in a vector unit consisting of real values,
Masking the generated purchase vector, which is the output of the generation model, by using an n-dimensional indicator vector specifying whether the generation model has purchased each item by the user
Computer system, characterized in that. The method of claim 13,
Given a user-specific condition vector and a random noise vector as a real value vector,
The discrimination model is adjusted by the user-specific condition vector and trained to distinguish the generated purchase vector from the actual purchase vector of the corresponding user.
Computer system, characterized in that. The method of claim 14,
The generation model is a neural network that outputs an n-dimensional purchase vector based on the user specific condition vector and the random noise vector,
The discriminant model is a neural network that has its own input and outputs a single scalar value indicating the probability that the input is actually measured data,
The at least one processor,
Training the generation model and the discriminant model through stochastic gradient descent using mini-batch and back-propagation
Computer system, characterized in that. The method of claim 13,
The purchase vector generated by the generation model includes the predicted preference score of the user for all items in the data set,
The at least one processor,
Selecting and recommending the top N items with the high prediction preference score from the data set
Computer system, characterized in that. The method of claim 13,
The at least one processor,
After estimating an item not purchased by the user as a negative item, training the generation model so that the value of the negative item approaches 0 (zero)
Computer system, characterized in that. The method of claim 17,
The at least one processor,
Training the generative model through a method of minimizing the objective function of the generative model by using the negative item for zero-reconfiguration normalization
Computer system, characterized in that. The method of claim 17,
The at least one processor,
Training the generation model by using the item purchased by the user and the negative item as input values of the discrimination model and transmitting the output of the negative item back to the generating model with respect to the loss slope of the discrimination model that
Computer system, characterized in that. The method of claim 17,
The at least one processor,
A method of minimizing the objective function of the generated model by using the negative item for zero-reconfiguration normalization, and using the item purchased by the user and the negative item as an input value of the discrimination model, and for the loss slope of the discrimination model Training the generation model through a hybrid method in which the method of transmitting the output of the negative item back to the generation model is combined
Computer system, characterized in that.

Images