KR20200046189A - Method and system for collaborative filtering based on generative adversarial networks - Google Patents

Abstract

Disclosed are a method and a system for collaborative filtering based on a generative adversarial network (GAN). The method for collaborative filtering based on a GAN comprises the steps of: training a generative model and a discriminative model of a GAN in a real-valued vector-wise manner on an item; and recommending a personalized item list to a user by using the trained generative model and discriminative model.

Description

METHOD AND SYSTEM FOR COLLABORATIVE FILTERING BASED ON GENERATIVE ADVERSARIAL NETWORKS

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

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

For example, in Korean Patent Registration No. 10-1030653 (registration date April 14, 2011), the similarity between the recommended user and other users is corrected by weighting the relationship between the recommended user and each user's preference information entropy of another user. Thus, a collaborative filtering recommendation technique with improved accuracy of preference prediction has been disclosed.

CF aims to provide a list of personalized items to the user by collecting a record of the user's previous preferences. The user's preference history is generally expressed in the form of a sparse matrix, where columns correspond to users, rows correspond to items, and values correspond to user preferences for items. Here, the user's preference (or feedback) may be provided on a scale of 1 to 5 (explicit), or may be provided in a unary manner such as purchase or click (implicit).

Among many CF methods, a model-based CF method that builds a user's past preference pattern learning model for an item 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 a user and an item's latent features.

Recently, CF methods based on deep neural networks have recently attracted attention. Due to DNN's ability to discover nonlinear interactions of potential factors between users and items and approximate arbitrary sequential functions, DNN-based CF models have strong performance for both rating prediction and top-N recommendations.

GAN is one of the DNNs that learn to mimic the distribution of a given data. GAN consists of two models: one is the creation model (hereinafter referred to as 'G') and the other is the discrimination 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 rather than 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 actual data from fake data. As a result of this hostile process, G learns at least how to generate realistic data such as photos that appear to be authentic to the human eye.

The above-mentioned GAN is most actively researched in the field of image generation, but research applied to the recommended system field is insignificant.

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

A new direction of vector-wise hostile training used in GAN-based collaborative filtering methods can be suggested.

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

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

According to another aspect, given a user-specific condition vector and a random noise vector 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 may be trained to distinguish the purchased vector from the actual purchase vector of the user.

According to another aspect, the training step includes masking the generated purchase vector, which is the output of the generated model, using an n-dimensional indicator vector specifying whether the generated model 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 by the user-specific condition vector and the random noise vector, and the discrimination model has its own input and has a probability that the input is actual data. It is a neural network that outputs a single scalar value, and the training step comprises the generation model and the discrimination through stochastic gradient descent using minibatch and back-propagation. You can train the model.

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

According to another aspect, the training step may train the generation model so that a value of the negative item approaches 0 (zero) after the user estimates an item that is not purchased as a negative item.

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

According to another aspect, the step of training uses the item purchased by the user and the negative item as input values of the discriminant model, and outputs the negative item with respect to the loss slope of the discriminant model again to the generating model The generation model may be trained through a method of passing to.

According to another aspect, the step of training includes using the negative item for zero-reconstruction normalization to minimize the objective function of the generation model, and inputting the user-purchased item and the negative item into the discriminant model. The generation model can 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 for the loss slope of the discrimination model is combined.

In combination with a computer, a computer program stored in a computer-readable recording medium is provided to execute the collaborative filtering method based on the GAN on the computer.

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

A computer system, comprising: a memory; And at least one processor coupled to the memory and configured to execute computer readable instructions contained in the memory, wherein the at least one processor is a real-valued vector-wise for an item. A computer characterized by training a generative model and a discriminative model of a GAN (generative adversarial network), and recommending a personalized item list to the user using the trained generating model and the discriminative model. Provide a system.

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

1 shows an example of a training process for generating a discrete item index.
2 is a visualization of the learning trend of IRGAN
Figure 3 is a visualization of the learning trend of GraphGAN.
FIG. 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) in an embodiment of the present invention.
6 shows an example of a learning algorithm for CFGAN_ZP in one embodiment of the present invention.
7 is a block diagram illustrating an example of an internal configuration of a computer system in 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 preferences and product preferences based on the ratings left by the user, and then predicts the rating or the top N items to be purchased for products that the user has not yet purchased.

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

The reason that the advantage of hostile training in the existing CF has not been fully utilized comes 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 produces a vector (eg, an image) with elements of consecutive values. Since the item index is discrete and finite, it is highly likely that G samples exactly the same data as the actual data (that is, the user's actual choice). Therefore, as training progresses, D becomes confused, and thus the grade is degraded because contradictory labels for the same item index are trained at a large rate. When the same item comes from G, this same item can sometimes be marked "fake", and in other cases from actual data it can be marked "real". As a result, D sends the wrong signal to G, which reduces the likelihood that G will sample the actual data item. In other words, instead of improving each other's ability, the accuracy of G and D is worse. After several iterations, the accuracy of the recommendation by D no longer increases, but rather plummets. This is due to the high proportion of contradictory labeled items. This prevents G from improving the accuracy in CFs where the competition between G and D does not help each other.

In order to solve the above problems, the present invention proposes a new GAN-based CF framework. In the framework according to the present invention, given a user, G attempts to generate a plausible purchase vector of real (real) value elements rather than sampling a single item index that the user may be interested in and purchase. Similarly, D attempts to distinguish the actual purchase vector from the actual data from the generated purchase vector. This approach to generating and discriminating real vectors is a confusion of D that occurs in GraphGAN and IRGAN due to items with contradictory labels, since it is almost impossible for G to generate a real vector that is exactly the same as the vector belonging to the actual data. prevent. This property allows D to continuously improve by guiding G through inverse propagation, thereby making the generated vector of G closer to the actual purchase vector in the iterative process. This GAN-based CF framework drives successful hostile training to achieve high-accuracy recommendations.

When shifting the goal from discrete item index generation to real-value vector generation for GAN-based CFs, we identify unique challenges that naturally appear due to the nature of the implicit feedback data. Unlike image data considered by the original GAN (represented by a high-density vector composed of pixels with multiple bit depths), the actual data of CF is a sparse single-valued vector. Here, the element (element) of the case where the corresponding user-item interaction is observed (that is, the user purchases the item) is 1. G attempts to generate a purchase vector similar to the actual data, but it can be easily escaped by finding a useless and easy solution: that is, it creates a purchase vector with all elements 1 without considering the user's relative preferences.

In connection with this problem, three CF methods (ie, CFGAN-ZR, CFGAN-PM and CFGAN-ZP) developed in the framework according to the present invention are proposed. Each method successfully solves the problem in a unique way. The commonalities of these methods are as follows: (1) Sample unobserved elements; (2) It is assumed that these methods will correspond to negative feedback (ie, the values of these methods are not omitted from the purchase vector and become zero); (3) You can request G to generate an output close to 0 from the sampled element and 1 from the element corresponding to the user's actual preference (ie, 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, other frameworks in accordance with the present invention formulate each of these methods according to a user-centric and item-centric approach.

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

(One) The limitations of the existing GAN-based CF method are confirmed and emphasized through preliminary experiments.

(2) In order to increase the recommended accuracy of CF, we propose a new direction that the GAN-based CF method should follow, ie, actual value vector unit hostile training, to maximize the benefits of hostile training.

(3) A new GAN-based CF framework is proposed. In other words, the direction of hostile training in actual value vectors is presented.

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

(5) The effectiveness of the proposed direction for vector-based hostile training is verified, and the superiority of the CF method is compared to the state-of-the-art top-N recommendation.

The following describes the preliminary work for the GAN based CF method.

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 a set of n items. Here, I _u is defined as a set of items that receive implicit feedback from u (ie, purchase). User's implicit feedback on items is a sparse matrix

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

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.CF's goal is j (

Represented by

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

Using a model that can be formulated and represented as

To calculate. Here, f represents a 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

The

It can be learned by optimizing the objective function using as the 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.The existing model-based CF method can be classified according to the type of the objective function used to train the model as follows: the point wise method and the pair wise method. Point unit method

And

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

GAN  Based CF ( GAN -based CF)

Recently, a generative hostile neural network (GAN) has been studied that proposes a new method of learning a machine learning model. GAN is a technique that trains G through the competition structure between the generation model (G) and the discrimination model (D) to make G generate plausible data. In other words, through a competition process involving G and D, G can be learned to capture the distribution of the measured data, and thus generate plausible synthetic data that does not differ in characteristics from the measured data. Officially, G and D use the objective function V to play the following two player minimum 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

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

While the actual data from

Is typically

Model distribution derived by

Synthetic data from. z is a random noise vector,

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

GAN's most successful field is image creation, 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 in the field of recommendation systems are being researched, and in order to train a CF model, instead of optimizing a point wise or pair wise objective function, a hostile training method is applied. And try to get satisfactory accuracy from the recommendation.

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 The principle based on information retrieval models.) is as follows.

In IRGAN, G attempts to create (or sample) the index of the item associated with a given user, and D attempts to distinguish the user's actual data item from the G-generated item. 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 given u and index k,

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

Using,, the probability of each item i belonging to the measured data of the 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

It is based on. At this time, d _u and d _i are latent factors belonging to D, respectively. And, g _u and g _i are potential factors belonging to G.

Wow

Is the item bias condition.

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

Is formulated as Equation 4:

[Equation 4]

D can be solved with a stochastic gradient rise, but G cannot receive a gradient from D because the 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 findings that focus on the GAN-based recommendation system, especially the visual aspects of the item. GAN can be used for recommended work, but previously requires additional visual information of the item.

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

Despite the potential for GAN to succeed in CF, there is still room for improvement. Explain the limitations of the "Discrete Item Index Generation" method and overcome these limitations to improve CF accuracy.

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 (almost) randomly, but improves its performance according to the guiding of D (b). As training progresses further (c), G frequently samples the exact same item as the actual data.

From D's point of view, the same item index (e.g., i ₃ ) is labeled as "fake" (in G) and labeled as "real" (i.e., from actual data), making D significantly confusing, As a result, D will decline. This means that D can transmit an erroneous signal to G through the policy gradient, and G can lower the possibility of sampling items of actual data and cause G to deteriorate.

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

As an example of experiment, MovieLens 100K and Watcha data are used. The existing well-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. As the scale for calculating accuracy, widely used precision and MRR were used, and the results for the top 5 (@ 5) were calculated.

FIG. 2 visualizes the learning trend of IRGAN, and FIG. 3 visualizes the learning trend of GraphGAN. Each graph in FIGS. 2 and 3 reports the accuracy of the top 5 recommendations. FIG. 4 shows the proportion of contradictory labeled items among items generated by G as hostile training progresses.

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

The present invention is intended to fully enjoy the benefits of hostile training for item recommendation tasks, and how to properly train G and D so that G consistently pushes D to its own limit rather than "well beyond" its push. And "vector unit training" is proposed to solve whether D can consistently give only beneficial feedback to G.

In the present invention, G and D perform generation and determination of a vector unit composed of real values instead of a single item index. In other words, in the present invention, G generates a real-valued vector, and D distinguishes G from an actual vector value from actual data. More specifically, G is attempting to generate a plausible purchase vector while the user is given an attempt to accurately determine if the given purchase vector is real or fake generated by G. For per-vector training, G can produce a purchase vector that is very similar (but not identical) to the actual purchase, but is unlikely to produce a purchase vector that is exactly the same as the actual purchase. This makes D seldom deceived by G, allowing D to consistently escape its limits in a series of steps due to the qualified G's training data. Likewise, D can continue to provide useful training signals to G, showing how G needs to slightly alter its output to generate a more realistic fake purchase vector through back propagation. 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) in an embodiment of the present invention.

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 the personal setting of each user. Given a user-specific condition vector c _u and a random noise vector z ² , G of CFGAN is an n-dimensional purchase vector.

Produces Purchase vector

Is expected to be a sparse vector, where all elements corresponding to u's purchased items

(

) Is hopefully 1. Likewise, 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 is represented by Equation 5.

[Equation 5]

Similarly, G is represented by 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

U have each item

Is an n-dimensional indicator vector that specifies whether or not

Represents 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 CFGAN and conventional GAN of the present invention is G

Multiplied by the output of G

By masking, G handles the scarcity of the actual data to be copied. By doing so, 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 unpurchased items, (1) D ignores the dropped nodes, and (2) G does not obtain a slope of loss from D for the dropped node. This is common to other CFs that use point-by-point objective functions in that they build a model based on 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 respectively

Wow

G and D are implemented with a multi-layered neural network parameterized by. In particular, G is 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

(Ie from G) or

It is an L ^D -layer (L ^D > = 2) neural network with input regardless of whether or not (from actual data) is input, and this neural network outputs a single scalar value indicating the probability that its input comes from actual data rather than G. To train G and D, we use stochastic gradient descent using minibatch and back-propagation. Each parameter

Wow

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.Z and after hostile training

When supplied, G is a compact vector

Produces

May include a prediction preference score for all items in the data set.

The jth element of each item j (

) As the predicted preference score of user u. Finally, the top N items with the highest predicted scores are selected and recommended as u.

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

The CFGAN of the present invention presents a 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 represented by a vector of pixels with multi-bit depth, the implicit feedback data for the user in the recommendation is a very sparse single value vector, the elements of which are corresponding user-items 1 if the interaction (i.e. purchase) is observed, otherwise it is empty. G tries to deceive D by generating a purchase vector similar to the actual data, but it will lead to a trivial solution that simply predicts all outputs of the purchase vector as 1. This solution is easy to obtain, but not useful. This is because the relative preference of the user, which is a basic element of CF, cannot be captured.

)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 key ideas shared by the various CF methods realized on the CFGAN of the present invention are as follows: the proportion of items not purchased by each user in all training iterations (e.g. t).

) Is randomly selected, and after estimating the item as a negative item, it indicates that the 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, it is possible to focus on generating values close to 1 for G purchased items to deceive D, and consider calculating values for negative items. As a result, we avoid accepting G's solution that 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,

Denotes a set of negative items sampled in the t-th iteration. Purchase matrix

The corresponding item of is considered not to be missing, but is considered 0 only in this iteration. S is

The percentage (percentage) of the item 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-reconstruction normalization (CFGAN-ZR), and a method based on partial-masking (CFGAN-PM). And a hybrid method (CFGAN-ZP) in which the above two are combined.

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 to indicate the item of the selected note

And indicates a parameter S specifying a ratio (percentage) of a negative item selected from non-purchased items.

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

[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 of

Wow

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

By minimizing, G produces a score close to 1 for observed items and low values (ie close to 0) for unobserved items to prevent G from reaching a trivial and useless solution. You can focus on 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

Each of the selected negative items and sampling parameter S is represented by using. CFGAN-PM does not necessarily drop all output nodes corresponding to non-purchased items,

The output node corresponding to is maintained. Thus, D utilizes input values from the purchased item as well as 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 purchase item closer to 1, and closer to 0 for the selected negative item.

와 수학식 9의

에 적용된다.The above idea is that of Equation 8

And 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 using both zero reconstruction normalization and partial masking. The set of negative items selected

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 in FIG. 6. CFGAN_ZR and CFGAN_PM are considered 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 in which each user is regarded as a training instance. Here, each training instance corresponds to the user and G (and D) generates (and discriminates) the user's purchase vector, but CFGAN can be implemented in an item-oriented manner in which the training instance corresponds to an item other than the user.

Item-oriented CFGAN

G conditioned by item-specific information attempts to produce a plausible purchase vector of each item, i.e. an m-dimensional vector. Here, m is the number of users, which indicates how likely the user corresponding to each element of the vector is to purchase the item. Likewise, D conditioned according to item-specific information attempts to distinguish the vector generated by G from the purchase vector of the item's actual data. The agent also attempts to select a negative (negative) user for each item representing those who are unlikely to purchase the item. After training, G predicts the preference score for each given item i and all users. Finally, N items with the highest prediction scores are recommended for 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 in 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 GAN-based CF method, the processor 710, the memory 720, the permanent storage device 730, the bus 740, input and output It may include an interface 750 and a network interface 760.

The processor 710 is a component for a GAN-based CF and may include or be a part of any device capable of processing a sequence of instructions. The processor 710 may include, for example, a computer processor, a processor in a mobile device or other electronic device, and / or a digital processor. The processor 710 may be included in, 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.

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

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

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

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

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

Further, in other embodiments, the computer system 700 may include more components than those in FIG. 7. However, there is no need to clearly show most prior art components. For example, the computer system 700 is 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, It may further include other components, such as a database.

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, problems that occur when using hostile training can be effectively solved and hostile training You can increase the recommendation accuracy of collaborative filtering by making the most of the benefits of.

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

The software may include a computer program, code, instruction, or a combination of one or more of these, and configure the processing device to operate as desired, or process independently or collectively You can command the device. Software and / or data may be embodied on 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 on networked computer systems, and stored or executed in a distributed manner. Software and data may be stored in 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 on a computer readable medium. At this time, the medium may be to continuously store a program executable on a computer or to temporarily store it for execution or download. In addition, the medium may be various recording means or storage means in the form of a combination of single or several hardware, and is not limited to a medium directly connected to a computer system, but may be distributed on a network. Examples of the media include magnetic media such as hard disks, floppy disks and magnetic tapes, optical recording media such as CD-ROMs and DVDs, and magneto-optical media such as floptical disks, And program instructions including ROM, RAM, flash memory, and the like. In addition, examples of other media include an application store for distributing applications, a site for distributing or distributing various software, and a recording medium or storage medium managed by a server.

As described above, although the embodiments have been described by a limited embodiment and drawings, those skilled in the art can make various modifications and variations from the above description. For example, the described techniques are performed in a different order than the described method, and / or the components of the described system, structure, device, circuit, etc. are combined or combined in a different form from the described method, or other components Alternatively, even if replaced or substituted by equivalents, appropriate results can be achieved.

Therefore, other implementations, other embodiments, and equivalents to the claims are also within the scope of the following claims.

Claims (20)

In the collaborative filtering method based on the GAN (generative adversarial network),
Training the generative model and discriminative model of the GAN in real-valued vector-wise on an item; And
Recommending a personalized item list to a user using the trained generation model and discrimination model
Collaborative filtering method based on GAN comprising a. According to claim 1,
The generation model generates a purchase vector consisting of real values,
The discrimination model determines whether a given purchase vector is an actual purchase vector corresponding to ground truth or a purchase vector generated from the generation model.
Collaborative filtering method based on GAN, characterized by. According to claim 1,
Given a user-specific condition vector and a random noise vector as real value vectors,
The generation model generates an n-dimensional purchase 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 user.
Collaborative filtering method based on GAN, characterized by. According to claim 3,
The training step,
Masking the generated purchase vector, which is the output of the generated model, using an n-dimensional indicator vector specifying whether the generated model purchased each item by a user
Collaborative filtering method based on GAN comprising a. According to claim 3,
The generation model is a neural network that outputs an n-dimensional purchase vector by the user specific condition vector and the random noise vector,
The discrimination model is a neural network that has its own input and outputs a single scalar value indicating the probability that the input is actual data.
The training step,
Training the generation model and the discrimination model through stochastic gradient descent using minibatch and back-propagation
Collaborative filtering method based on GAN, characterized by. According to claim 1,
The purchase vector generated in the generation model includes a user's predicted preference score for all items in the data set,
The recommended steps,
Selecting and recommending the top N items with high prediction preference scores from the data set
Collaborative filtering method based on GAN, characterized by. According to claim 1,
The training step,
Training the generation model so that the value of the negative item approaches 0 (zero) after estimating an item that the user has not purchased as a negative item.
Collaborative filtering method based on GAN, characterized by. The method of claim 7,
The training step,
Training the generated model through a method of minimizing the objective function of the generated model by using the negative item for zero-reconstruct normalization
Collaborative filtering method based on GAN, characterized by. The method of claim 7,
The training step,
Training the generation model through a method of using the purchased item and the negative item as input values of the discrimination model and passing the output of the negative item to the production model for the loss slope of the discrimination model. that
Collaborative filtering method based on GAN, characterized by. The method of claim 7,
The training step,
A method of minimizing the objective function of the generation model by using the negative item for zero-reconstruction normalization, and using the item purchased by the user and the negative item as input values of the discrimination model, and for the loss slope of the discrimination model. Training the generation model through a hybrid method in which a method of passing the output for the negative item back to the generation model is combined.
Collaborative filtering method based on GAN, characterized by. A computer program stored in a computer-readable recording medium in combination with a computer to execute the collaborative filtering method based on the GAN of any one of claims 1 to 10 in a computer. A computer-readable recording medium characterized in that a program for executing a cooperative filtering method based on the GAN of any one of claims 1 to 10 is recorded on a computer. In the 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,
We train the generative model and the discriminative model of the GAN GAN (generative adversarial network) in real-valued vector-wise for the item,
Using the trained generation model and discrimination model to recommend a personalized item list to the user
Computer system characterized by. The method of claim 13,
Given a user-specific condition vector and a random noise vector as real value vectors,
The generation model generates an n-dimensional purchase 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 user.
Computer system characterized by. The method of claim 14,
The generation model is a neural network that outputs an n-dimensional purchase vector by the user specific condition vector and the random noise vector,
The discrimination model is a neural network that has its own input and outputs a single scalar value indicating the probability that the input is actual data.
The at least one processor,
Training the generation model and the discrimination model through stochastic gradient descent using minibatch and back-propagation
Computer system characterized by. The method of claim 13,
The purchase vector generated in the generation model includes a user's predicted preference score for all items in the data set,
The at least one processor,
Selecting and recommending the top N items with high prediction preference scores from the data set
Computer system characterized by. The method of claim 13,
The at least one processor,
Training the generation model so that the value of the negative item approaches 0 (zero) after estimating an item that the user has not purchased as a negative item.
Computer system characterized by. The method of claim 17,
The at least one processor,
Training the generated model through a method of minimizing the objective function of the generated model by using the negative item for zero-reconstruct normalization
Computer system characterized by. The method of claim 17,
The at least one processor,
Training the generation model through a method of using the purchased item and the negative item as input values of the discrimination model and passing the output of the negative item to the production model for the loss slope of the discrimination model. that
Computer system characterized by. The method of claim 17,
The at least one processor,
A method of minimizing the objective function of the generation model by using the negative item for zero-reconstruction normalization, and using the item purchased by the user and the negative item as input values of the discrimination model, and for the loss slope of the discrimination model. Training the generation model through a hybrid method in which a method of passing the output for the negative item back to the generation model is combined.
Computer system characterized by.

Images