A recommendation method based on generative adversarial network

Abstract

The present invention provides a recommendation system method based on a generative adversarial network, the generative adversarial network includes a first-layer generator, a second-layer generator and a discriminator, the method includes: constructing a data set required for the recommendation system through pre-training; constructing pre-training data after obtaining the pre-training results; creating an internal information flow network based on the first-layer generator and the second-layer generator; creating an external information flow network based on the results of the internal information flow network and the discriminator; and setting training parameters and using the pre-training data and the training parameters to perform iterative training on the internal information flow network and the external information flow network. In addition, the present invention also provides a computer-readable storage medium. Through the present invention, the limitations of generative adversarial networks for recommendation systems can be solved, and the effectiveness of recommendation systems can be significantly improved.

Description

Technical Field

[0001] The present invention relates to a recommendation system / information retrieval, and more particularly, to a recommendation system method based on a generative adversarial network. Background Art

[0002] A typical recommendation system contains two roles, namely users and items. The recommendation system calculates the user's preference for items, so that a series of items can be filtered out and provided to the user. In this process, appropriate results are sampled for specific users by capturing item node information. For some recommendation system tasks, existing related models have achieved good results. For example, BPR ([1] Steffen Rendle, Christoph Freudenthaler, Zeno Gantner and Lars Schmidt-Thieme. 2012. BPR: Bayesian personalized ranking from implicit feedback. In UAI.) is used to form BPR-Opt, which is derived from the maximum a posteriori estimator for optimal personalized ranking; NCF ([2] Xiangnan He, Lizi Liao, Hanwang Zhang, Liqiang Nie, Xia Hu and Tat-Seng Chua. 2017. Neural collaborative filtering. In WWW. 173–182.) combines matrix decomposition and multi-layer perceptron, enabling it to extract features from low and high dimensions, thereby obtaining good recommendation results.

[0003] The framework of the generative adversarial network model can be divided into two parts: the generator and the discriminator. The generator continuously fits the real data distribution and generates fake data to deceive the discriminator. The discriminator learns to distinguish between fake data and real data through repeated adversarial training of the generator. With the explosive development of recommendation systems and generative adversarial networks, more and more researchers are applying generative adversarial networks to the tasks of recommendation systems. Due to the particularity of recommendation systems, when applying generative adversarial networks to handle recommendation system tasks, the generator does not generate new things as usual, but processes information flow in the form of generating hard negative samples. The generator does not generate new things, but samples existing items for the discriminator, and the discriminator will subsequently perform secondary scoring on the sampled items. Some existing systems have applied generative adversarial networks to recommendation systems. For example, PDGAN ([3]Wu Q, Liu Y and Miao C.2019.PD-GAN:Adversarial Learning for Personalized Diversity-Promoting Recommendation.In IJCAI.3870-3876.) was proposed to better capture users’ personal preferences for the diversity of individual items and a group of items; and IRGAN ([4]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) was inspired by the application of generative adversarial networks in the image field. models.InSIGIR.515–524.) The adversarial network is applied to multiple semi-supervised IR tasks, which can correctly handle implicit feedback. For the recommendation system task, IRGAN implements classic collaborative filtering as one of the scoring functions of user preferences, which keeps the model clean and creates a special "sampling" of adversarial networks in the recommendation field. However, in IRGAN, since the training results of the generator and the discriminator are not equivalent, the improvement of the generator is much lower than that of the discriminator. In addition, since the sample generation of the generator is close to random, the effect of the generator is very poor in the early stage of training ([5]Wu Q, Burges CJC, Svore KM, et al. 2010. Adapting boosting for information retrieval measures. In Information Retrieval. 13(3): 254-270.).

[0004] Accordingly, there is a need in the art for improved generative adversarial network based recommender systems. Summary of the invention

[0005] This Summary is provided to introduce some concepts in a simplified form that are further described below in the Detailed Description. This Summary is not intended to identify key features or essential features of the claimed subject matter, nor is it intended to be used as an aid in determining the scope of the claimed subject matter.

[0006] In view of the defects in the prior art described above, an object of the present invention is to provide a recommendation system method based on a generative adversarial network to solve the limitations of the generative adversarial network in the recommendation system.

[0007] According to a first aspect of the present invention, a recommendation system method based on a generative adversarial network is provided, wherein the generative adversarial network includes a first-layer generator, a second-layer generator and a discriminator, and the method may include: constructing a data set required for the recommendation system through pre-training; constructing pre-training data after obtaining the pre-training results; creating an internal information flow network based on the first-layer generator and the second-layer generator; creating an external information flow network based on the results of the internal information flow network and the discriminator; and setting training parameters and using the pre-training data and the training parameters to perform iterative training on the internal information flow network and the external information flow network.

[0008] In an embodiment of the first aspect, pre-training may include: obtaining relationship features between a single user and an item.

[0009] In an embodiment of the first aspect, creating the internal information flow network may include: initializing a first layer generator and a second layer generator; and improving performance of the second layer generator by maximizing a cross entropy loss.

[0010] In an embodiment of the first aspect, creating an internal information flow network may include: sampling negative samples of items by a first layer generator; and inputting item features into a two-layer neural network after sampling, wherein the first layer of the neural network is activated by a nonlinear function Tanh, and the second layer of the neural network is activated by a nonlinear function Sigmoid.

[0011] In an embodiment of the first aspect, creating the external information flow network may include: initializing and setting the internal information flow and the discriminator; and improving the performance of the discriminator by maximizing the cross entropy loss.

[0012] In an embodiment of the first aspect, creating the external information flow network may include: inputting the internal information flow result into a two-layer linear neural network, wherein the first layer of the linear neural network is activated by a nonlinear function Tanh, and the second layer of the linear neural network is activated by a nonlinear function Sigmoid.

[0013] In an embodiment of the first aspect, the iterative training may include: selecting a minimum-maximum game for optimization to iteratively train the first-layer generator and the second-layer generator, wherein each training process calls the internal information flow network.

[0014] In an embodiment of the first aspect, the iterative training may include: selecting a minimum-maximum game for optimization to iteratively train the second-layer generator and the discriminator, wherein each training process calls an external information flow network.

[0015] In an embodiment of the first aspect, iterative training may include: the first layer generator teaches the screened negative samples to the second layer generator in the form of a probability distribution; the second layer generator gives the first layer generator a feedback reward; the second layer generator generates screened negative sample pairs to the discriminator; and the discriminator gives the second layer generator a discriminative feedback, wherein the second layer generator and the discriminator make progress together from learning the real negative sample pairs with the discriminative feedback.

[0016] According to a second aspect of the present invention, there is provided a computer readable storage medium storing a computer program which, when executed by a processor, implements the method according to the present invention.

[0017] Compared with the prior art, the present invention has the following significant advantages:

[0018] (1) This paper proposes a novel adversarial network framework, which enhances the adversarial network by adding two types of information flow processes to improve the accuracy of the recommendation system. The key feature of the model is that the two-level generator network is applicable to the two types of information flows and plays a role in distinguishing and generating.

[0019] (2) The present invention focuses on the relationship features between a single user and an item rather than the node features during the pre-training stage. These relationship features can be directly input into the adversarial network as the user-item connection, thereby allowing the two-layer adversarial network to perform more comprehensive item screening for users.

[0020] (3) The model proposed in this paper utilizes the insightful internal information flow LambdaRank, which provides more reasonable sampling; the external information flow uses RankNet, which plays an important role in the external information flow.

[0021] These and other features and advantages will become apparent by reading the following detailed description and by reference to the associated drawings.It is to be understood that the foregoing general description and the following detailed description are illustrative only and are not restrictive of the aspects of what is claimed. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] In order to understand the manner in which the above features of the present invention are used in detail, the above briefly summarized contents can be described in more detail with reference to various embodiments, some of which are shown in the accompanying drawings. However, it should be noted that the accompanying drawings only show some typical aspects of the present invention and should not be considered to limit its scope, because the description may allow for other equally effective aspects.

[0023] Figure 1 A flowchart illustrating a generative adversarial network-based recommendation system method according to an embodiment of the present invention is provided.

[0024] Figure 2 A schematic diagram of enhanced information flow generative adversarial network training according to an embodiment of the present invention is illustrated.

[0025] Figure 3 A block diagram illustrating an example of a hardware implementation for a computing device according to one embodiment of the present invention. DETAILED DESCRIPTION

[0026] The present invention will be described in detail below in conjunction with the accompanying drawings, and the features of the present invention will be further revealed in the following specific description.

[0027] Recommendation systems can help users process large amounts of information, and generative adversarial networks (GANs) have shown great potential in recommendation systems. The present invention proposes a new type of generative adversarial network, the main idea of ​​which is to propose internal information, based on the information between the original generator and the discriminator to enhance the accuracy of data information distribution acquisition within the generator, and its main purpose is to solve the limitations of generative adversarial networks in recommendation systems, and to enhance the adversarial network by adding two types of information flow processes, thereby alleviating the problem of insufficient generator training. In this process, both the generator and the discriminator will be better trained. In addition, the key feature of the present invention is that its two-stage generator network is applicable to two types of information flows, and plays a role in distinguishing and generating.

[0028] In other words, in order to make the training more effective, the focus of the present invention is to avoid the initial random sampling process in the generator, thereby adding internal information between the generator and the discriminator. The actual meaning of the internal information is to improve the obvious defects of the external information flow. In the internal information, the generator generates hard sample pairs for the discriminator to train, and the discriminator feeds back the reward to the generator. Therefore, after improving the internal part of the generator, the internal effect of the generator is enhanced, which encourages the generator to sample better negative sample pairs and provide them to the discriminator.

[0029] The following combination Figure 1 and Figure 2 Let's explain in more detail the application of the novel generative adversarial network of the present invention in the recommendation system.

[0030] Figure 1 A flowchart of a generative adversarial network-based recommendation system method 100 according to an embodiment of the present invention is illustrated. Figure 2 A schematic diagram illustrating the training of an enhanced information flow generative adversarial network 200 according to an embodiment of the present invention is illustrated.

[0031] like Figure 2 As shown, the generative adversarial network 200 of the present invention may include a first layer generator G1 210, a second layer generator G2 220 and a discriminator 230, wherein an internal information flow network 240 may be created based on the first layer generator G1 210 and the second layer generator G2 220, and an external information flow network 250 may be created based on the result of the internal information flow network 240 and the discriminator 230.

[0032] In box 110, method 100 may include: constructing a dataset required for the recommendation system through pre-training. The dataset of the present invention is experimented based on the MovieLens (100k) dataset. The MovieLens dataset contains user ratings (1 to 5). It contains 943 users, 1683 items and 18 explicit categories, and the rating 5 of the MovieLens dataset is regarded as a positive feedback selection. The pre-training process is based on the existing MovieLens dataset, and based on the traditional recommendation system model, the parameters are changed to obtain the feature columns of the current dataset users for the items. In order to allow the generator to better capture the connection between the user and the item, pre-training is performed on the current dataset to construct a 54-dimensional feature vector to represent the feature connection between the user and the item. Each dimension of the 54-dimensional feature vector can be regarded as the user's rating of the item. Then, the training after inputting the 54-dimensional feature vector into the adversarial network is similar to a learning-to-rank system, and the loss function also conforms to the learning-to-rank paradigm. These features are input into the generator, which greatly improves the generator. This process can be understood as expanding the structure of the adversarial network to enhance the information flow process inside the generator. The present invention focuses on the relationship features between a single user and an item rather than the node features during the pre-training stage. These relationship features can be directly input into the adversarial network as the user-item connection, so that the two-layer adversarial network can perform more complete item screening for users.

[0033] In block 120, the method 100 may include: constructing pre-training data after obtaining the pre-training results. For example, after obtaining the pre-training results, integrate them into the MovieLens dataset, divide them into three learning subsets: a training set, a validation set, and a test set, and use a cross-validation method to divide the data into several parts; the format of the dataset meets the following requirements: each row is a user-item pair, the first column is the relevant label of the pair, the second column is the user ID (userID), the third column is the item ID (itemID), and the middle column is the feature column obtained through the above pre-training.

[0034] At block 130, the method 100 may include creating an internal information flow network based on the first layer generator and the second layer generator. In the present invention, the internal information flow refers to the information flow between the first layer generator and the second layer generator.

[0035] In one embodiment, the operation of box 130 may include: initializing the first layer generator and the second layer generator, and improving the performance of the second layer generator by maximizing the cross entropy loss. The first layer generator samples the negative sample of the item, and inputs the item features into the two-layer neural network after sampling the data set, wherein the first layer neural network is activated by the nonlinear function Tanh. When the second layer neural network is activated by the nonlinear function Sigmoid, for the feature matrix corresponding to the current user, the Sigmoid output result is multiple rows and one column, where the number of rows is equal to the number of rows of the feature matrix, and the number in each row represents the user's score for the item in that row. When the second layer neural network is activated by the nonlinear function Softmax, for the feature matrix corresponding to the current user, the Softmax output result is the probability distribution of the user's preference for the item.

[0036] Furthermore, the optimization goal during the internal information flow network training process is as follows (1):

[0037]

[0038] Among them i, u n , r refer to the item, the nth user and the user's score respectively, N refers to the number of user groups, G(i|u n ) refers to u n Under the condition, the second layer generator scores the current item. The probability that the first layer generator generates a sample for the second layer generator is p λ (i|u n ,r), It is subject to the basic correlation distribution of the data set p true (i|u n ,r) is the expected value of the calculation.

[0039] At block 140, the method 100 may include creating an external information flow network based on the result of the internal information flow network and the discriminator. In the present invention, the external information flow refers to the information flow between the generator and the discriminator.

[0040] In one embodiment, the operation of box 140 may include: initializing the internal information flow and the discriminator, and improving the performance of the discriminator by maximizing the cross entropy loss. The network architecture is as follows: according to the internal information flow result (the probability distribution of the user's preference for items), a two-layer linear neural network is input, the first linear layer is activated by the nonlinear function Tanh, and the second layer is activated by the nonlinear function Sigmoid; for a given user, the Sigmoid output is the user's score for all items. The final recommendation result is the user's ranking of the item scores.

[0041] Furthermore, the optimization goal during the external information flow network training process is as follows (2):

[0042]

[0043] Among them i, u n , r refer to the item, the nth user and the user's score respectively, N refers to the number of user groups, D(i|u n ) refers to u n Under the retrieval item preference, the discriminator scores the current item. The probability that the second-layer generator generates samples for the discriminator is p θ (i|u n ,r); is the probability of generating a sample p θ (i|u n ,r) is the expected value of the calculation.

[0044] At block 150 , method 100 may include setting training parameters and performing iterative training on the inner information flow network and the outer information flow network using pre-training data and the training parameters.

[0045] In one embodiment, the operation of box 150 may include: conducting experiments using the MovieLens (100k) dataset. The MovieLens dataset contains user ratings (1 to 5). It contains 943 users, 1683 items, and 18 explicit categories, and the rating 5 of the MovieLens dataset is regarded as a positive feedback selection. During the training process, the parameters of the neural network include the learning rate, the momentum term of the gradient algorithm, the batch parameter, the learning rate of the first layer generator for Stochastic Gradient Descent (SGD) is set to 1e-6, the learning rate of the second layer generator and the discriminator for Adam is set to 1e-5, the momentum term β1=0.9, β2=0.999, and there is no weight decay, and the batch parameter is set to 32.

[0046] Furthermore, the minimum-maximum game is selected for optimization to iteratively train the first-layer generator and the second-layer generator. The internal information flow network is called in each training process, and the training is iterated 200 times in total. In the actual training, since the sampling items given by the first-layer generator to the first-layer generator are not continuous, the following formula is used to optimize it:

[0047]

[0048] The core idea of ​​formula (3) is to transform the expected gradient into the expectation of the gradient, where K represents the total number of items that a specific user can choose.

[0049] Furthermore, the minimum-maximum game is selected for optimization to iteratively train the second-layer generator and discriminator. Each training process calls the external information flow network and iterates the training 200 times. The second-layer generator continuously generates false positive samples through the real positive samples sampled by the first-layer generator, and plays a game with the discriminator. The performance of the second-layer generator is improved through the following formula (4):

[0050]

[0051] The training of the external information flow can fully optimize the parameter θ, thereby improving the effect of the second-layer generator and discriminator.

[0052] Furthermore, during the iterative training of the internal and external information flows, the first-layer generator teaches the filtered negative samples to the second-layer generator in the form of probability distribution, and the second-layer generator gives the first-layer generator a feedback reward. The second-layer generator generates filtered negative sample pairs to the discriminator, and the discriminator gives the second-layer generator a discriminative feedback. From the learning of the discriminative feedback for the real negative sample pairs, the second-layer generator and the discriminator make progress together.

[0053] In optional box 160, method 100 may include performing a multi-index validity test to verify the performance of the iteratively trained generative adversarial network. For example, after training the recommendation system model, three indicators are set to judge the accuracy of the recommendation results, namely NDCG@3, NDCG@5 and NDCG@10. The Normalized Discounted Cumulative Gain (NDCG) score evaluates the ranking performance by considering the position of the correct item, assigning higher scores to the top-ranked popular items to calculate the position of the popular items, that is, assigning higher scores to the top k items in the ranking table. When the adversarial network proposed in the present invention is applied to the recommendation system, the average performance improvement of NDCG is about 13.77% higher than the baseline. In addition, considering that the quality of the top-ranked items in real-world scenarios is important, the improvement of NDCG@3 is significant, reaching 30.98%, as shown in Table 1 below.

[0054]

[0055]

[0056] Table 1 Improvements of the model of the present invention relative to the existing model

[0057] Furthermore, the best results of the NDCG indicators of the two ablation experimental models are compared with the original model to verify the effectiveness of the adversarial network proposed in the present invention for the recommendation system, as shown in Table 2 below.

[0058]

[0059] Table 2 Effectiveness of the proposed model for recommendation system

[0060] Figure 3 A block diagram of an exemplary computing device 300 according to one embodiment of the present invention is shown, which is one example of a hardware device applicable to aspects of the present invention.

[0061] refer to Figure 3 , a computing device 300 will now be described, which is an example of a hardware device applicable to various aspects of the present invention. The computing device 300 can be any machine that can be configured to perform processing and / or calculations, and can be but is not limited to a workstation, a server, a desktop computer, a laptop computer, a tablet computer, a personal digital assistant, a smart phone, a vehicle-mounted computer, or any combination thereof. The various methods described above can be implemented in whole or in part by the computing device 300 or a similar device or system.

[0062] The computing device 300 may include components that may be connected or communicated via one or more interfaces and a bus 302. For example, the computing device 300 may include a bus 302, one or more processors 304, one or more input devices 306, and one or more output devices 308. The one or more processors 304 may be any type of processor and may include, but are not limited to, one or more general-purpose processors and / or one or more special-purpose processors (e.g., a dedicated processing chip). The input device 306 may be any type of device capable of inputting information to the computing device and may include, but are not limited to, a mouse, a keyboard, a touch screen, a microphone, and / or a remote controller. The output device 308 may be any type of device capable of presenting information and may include, but are not limited to, a display, a speaker, a video / audio output terminal, a vibrator, and / or a printer. The computing device 300 may also include or be connected to a non-transient storage device 310, which may be any storage device that is non-transient and capable of data storage, and may include, but is not limited to, a disk drive, an optical storage device, a solid-state memory, a floppy disk, a floppy disk, a hard disk, a tape or any other magnetic medium, an optical disk or any other optical medium, a ROM (read-only memory), a RAM (random access memory), a cache memory, and / or any memory chip or cassette, and / or any other medium from which a computer can read data, instructions, and / or code. The non-transient storage device 310 may be detachable from the interface. The non-transient storage device 310 may have data / instructions / code for implementing the above methods and steps. The computing device 300 may also include a communication device 312. The communication device 312 can be any type of device or system that can communicate with internal devices and / or with a network and can include but is not limited to a modem, a network card, an infrared communication device, a wireless communication device and / or a chipset, such as a Bluetooth device, an IEEE 1302.11 device, a WiFi device, a WiMax device, a cellular communication device and / or the like.

[0063] The computing device 300 may also include a working memory 314 , which may be any type of working memory capable of storing instructions and / or data that facilitate the operation of the processor 304 and may include, but is not limited to, a random access memory and / or a read-only storage device.

[0064] Software components may be located in the working memory 314, including but not limited to an operating system 316, one or more application programs 318, drivers and / or other data and codes. Instructions for implementing the above methods and steps may be included in the one or more application programs 318, and the aforementioned various modules / units / components may be implemented by the processor 304 reading and executing the instructions of the one or more application programs 318.

[0065] In the description of the present invention, it should be understood that the terms “first”, “second” and “third” are only used for descriptive purposes and cannot be understood as indicating or implying relative importance.

[0066] It will be appreciated by those skilled in the art that various embodiments of the present invention may be provided as methods, devices, systems or computer program products. Therefore, the present invention may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present invention may take the form of a computer program product implemented on one or more computer-readable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) in which computer executable program code is stored.

[0067] The present invention is described with reference to the flowcharts and / or block diagrams of the methods, devices, systems and computer program products according to embodiments of the present invention. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor or other programmable data processing device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate a device for implementing the functions specified in one or more processes in the flowchart and / or one or more boxes in the block diagram.

[0068] Although various aspects of the present invention have been described so far with reference to the accompanying drawings, the above-mentioned methods, systems and devices are only examples, and the scope of the present invention is not limited to these aspects, but is limited only by the attached claims and their equivalents. Various components may be omitted or may be replaced by equivalent components. In addition, the steps may be implemented in an order different from the order described in the present invention. In addition, various components may be combined in various ways. It is also important that, as technology develops, many of the components described may be replaced by equivalent components that appear later. Various modifications to the present disclosure will be apparent to those skilled in the art, and the universal principles defined herein may be applied to other variations without departing from the scope of the present disclosure. Thus, the present disclosure is not limited to the examples and designs described herein, but should be granted the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A recommendation method based on a generative adversarial network, wherein the generative adversarial network comprises a first-layer generator, a second-layer generator and a discriminator. include: Build the data set required for the recommendation system through pre-training; Build pre-training data after obtaining the pre-training results; creating an internal information flow network based on the first layer generator and the second layer generator; Creating an external information flow network based on the result of the internal information flow network and the discriminator; as well as Setting training parameters and performing iterative training on the internal information flow network and the external information flow network using the pre-training data and the training parameters; Creating an internal information flow network includes: Initializing and setting the first layer generator and the second layer generator; and Improve the performance of the second layer generator by maximizing the cross entropy loss; Creating an internal information flow network also includes: The first layer generator samples negative samples of items; and After sampling, the item features are input into a two-layer neural network, where the first layer of the neural network is activated by the nonlinear function Tanh, and the second layer of the neural network is activated by the nonlinear function Sigmoid; Creating an external information flow network includes: Initialize and set the internal information flow and the discriminator; Improving the performance of the discriminator by maximizing the cross entropy loss; Creating an external information flow network also includes: The internal information flow result is input into a two-layer linear neural network, where the first layer of the linear neural network is activated by the nonlinear function Tanh, and the second layer of the linear neural network is activated by the nonlinear function Sigmoid; The internal information flow refers to the information flow between the first layer generator and the second layer generator; The external information flow of the external information flow network refers to the information flow between the internal information flow network and the discriminator.

2. The method of claim 1, wherein the pre-training include: Get the relationship features between a single user and an item.

3. The method of claim 1, wherein the iterative training include: Selecting the minimum maximum game for optimization to iteratively train the first layer generator and the second layer generator, Each training process calls the internal information flow network.

4. The method of claim 1, wherein the iterative training include: Select the minimum maximum game for optimization to iteratively train the second layer generator and the discriminator, Each training process calls the external information flow network.

5. The method of claim 1, wherein the iterative training include: The first layer generator teaches the second layer generator the screened negative samples in the form of probability distribution; The second layer generator gives the first layer generator a feedback reward; The second layer generator generates filtered negative sample pairs to the discriminator; and The discriminator gives a discriminative feedback to the second layer generator, wherein the second layer generator and the discriminator make progress together from learning the discriminative feedback for true negative sample pairs.

6. A computer-readable storage medium storing a computer program, wherein the computer program, when executed by a processor, implements the method according to any one of claims 1 to 5.

Images