Hierarchical contrastive learning method and hierarchical contrastive learning device for sequential recommendation

Abstract

The present specification relates to a deep learning model training method for sequential recommendation. The hierarchical contrastive learning method for sequential recommendation comprises the steps of: generating first enhanced sequence data (S1 u1) in which sequence data input from one user is enhanced by applying a first conversion operation to the sequence data; generating second enhanced sequence data (S1 u2) in which the sequence data is enhanced by applying, to the sequence data, a second conversion operation equal to or different from the first conversion operation; transmitting the S1 u1 and the S1 u2 to a hierarchical contrastive learning model including at least one learning block, and calculating one or more loss values; and training the hierarchical contrastive learning model on the basis of a total loss value obtained by summing the loss values.

Description

A hierarchical contrastive learning method and a hierarchical contrastive learning device for sequential recommendation.

[0001] The embodiments relate to a hierarchical contrastive learning method and a hierarchical contrastive learning device for sequential recommendation.

[0002]

[0003] Sequential recommendation can address the problem of preference shift by predicting the next item based on a user's previous behavior. Recently, a useful approach using contrastive learning has emerged, demonstrating its effectiveness in recommending items under sparse user-item interactions. Furthermore, the effectiveness of combining various augmentation methods has been demonstrated across diverse domains, particularly in computer vision.

[0004] Various algorithms and technologies can be used for sequential recommendations. These algorithms can generate recommendations by considering various factors, such as user history, item metadata, collaborative filtering, deep learning, and reinforcement learning. Examples include sequence-to-sequence models, recurrent neural networks (RNNs), long short-term memory networks (LSTMs), and reinforcement learning algorithms.

[0005] However, previous research on augmentation within the contrastive learning framework utilized for sequential recommendation has focused only on limited conditions and simple structures.

[0006] Prior art documents related to the present invention include Korean Patent Publication No. 10-2014-0100595 (August 18, 2014).

[0007]

[0008] There is a need for a method or device that can improve the performance of a recommendation system suffering from the sparse interaction problem by taking advantage of the currently utilized data augmentation operation and utilizing a hierarchical augmentation operation and a learning model with a hierarchical structure.

[0009]

[0010] A hierarchical contrastive learning method according to one embodiment applies a first transformation operation to sequence data input from a user to augment data, thereby generating first augmented sequence data ( ), a step of generating second augmented sequence data by applying a second transformation operation that is the same as or different from the first transformation operation to the sequence data, and augmenting the data ( ) generating step, the above and above The method may include transmitting the data to a hierarchical contrastive learning model including at least one learning block, calculating at least one loss value, and learning the hierarchical contrastive learning model based on a total loss value obtained by summing the loss values.

[0011] Additionally, the hierarchical contrastive learning method may further include a step of counting the number of times the first transformation operation is applied and the number of times the second transformation operation is applied as one.

[0012] Additionally, at least one loss value can be calculated based on the count number.

[0013] Additionally, the hierarchical contrastive learning method may include a step of determining the number of times the sequence data passes through the learning block based on the count number and training a hierarchical contrastive learning model.

[0014] In addition, the hierarchical contrastive learning method further includes, after the step of calculating the at least one loss value, a step of calculating a total loss value by assigning a weight to the at least one loss value, and the step of learning the hierarchical contrastive learning model updates the weight by backpropagation based on the total loss value, and can learn the hierarchical contrastive learning model.

[0015] A sequential recommendation method according to one embodiment is a method for sequentially recommending user preferred items using a deep learning model, comprising: a step of receiving sequence data from a user; a step of predicting the next user preferred item using a learned hierarchical contrastive learning model; and a step of recommending the predicted preferred item to the user, wherein the hierarchical contrastive learning model includes at least one learning block and applies a first transformation operation to sequence data obtained from one user to enhance data, and comprises first augmented sequence data ( ) and apply a second transformation operation that is the same as or different from the first transformation operation to the sequence data to create second augmented sequence data ( ) and create the above and above It can be set to transmit at least one learning block to the learning block to train the learning block.

[0016] Additionally, the hierarchical contrastive learning model may be set to count the number of times the first transformation operation is applied and the number of times the second transformation operation is applied as one.

[0017] Additionally, the hierarchical contrastive learning model can be set to learn by determining the number of times the sequence data passes through the learning block based on the count number.

[0018] In another embodiment, a computer-readable recording medium comprises one or more non-transitory computer-readable media storing one or more instructions, wherein the one or more instructions executable by one or more processors apply a first transformation operation to sequence data input from one user to create first augmented sequence data ( ) and apply a second transformation operation that is the same as or different from the first transformation operation to the sequence data to create second augmented sequence data ( ) and create the above and above The hierarchical contrastive learning model can be trained by transmitting the hierarchical contrastive learning model including at least one learning block.

[0019] A hierarchical contrastive learning device according to another embodiment comprises a memory configured to store a plurality of instructions, a hierarchical data augmentation model that performs data augmentation on sequence data input from one user, at least one learning block, a hierarchical contrastive learning model that is trained using data transmitted from the hierarchical data augmentation model, and a processor functionally coupled with the memory, wherein the processor, when the plurality of instructions are executed, applies a first transformation operation to the sequence data to augment the data (first augmented data) ) and apply a second transformation operation that is the same as or different from the first transformation operation to the sequence data to create second augmented data ( ) and create the above and above It can be configured to transmit to a hierarchical contrastive learning model including at least one learning block, calculate at least one loss value, and train the hierarchical contrastive learning model based on a total loss value obtained by summing the at least one loss value.

[0020] Additionally, the hierarchical contrastive learning model may include a first learning block in which data is learned, a second learning block in which data that has passed through the first learning block is learned, and a third learning block in which data that has passed through the second learning block is learned.

[0021] In addition, the processor, Applying a third transformation operation different from the first transformation operation, A fourth conversion operation different from the second conversion operation can be applied.

[0022] Additionally, the first learning block may include an encoder, the second learning block may include a transformer and a position-wise feed-forward network (PFFN), and the third learning block may include a transformer and a PFFN.

[0023] In addition, the processor may count the number of times the first transformation operation is applied and the number of times the second transformation operation is applied as one number, count the number of times the third transformation operation is applied and the number of times the fourth transformation operation is applied as one number, and determine whether the sequence data passes the first learning block, the second learning block, or the third learning block based on the counted number.

[0024] Additionally, the processor may be configured to produce at least one loss value based on the count number.

[0025] Additionally, the processor may be configured to learn by calculating a total loss value by weighting the at least one loss value and updating the weight by backpropagation based on the total loss value.

[0026]

[0027] According to the embodiment, the newly developed augmentation operation can be easily added and applied to the learning model.

[0028] According to an embodiment, even in situations where data is small or insufficient, the learning model can learn from various data through augmented computing and demonstrate excellent performance, thereby improving recommendation performance for new platforms or newly registered users.

[0029]

[0030] Figure 1 is a conceptual diagram showing the overall framework of the hierarchical contrastive learning device of the embodiment.

[0031] Figure 2 is a flowchart of the hierarchical contrastive learning method of the embodiment.

[0032] Figure 3 is a conceptual diagram illustrating the hierarchical contrastive learning device of the embodiment in more detail.

[0033] Figure 4 is a graph explaining that when the contrastive learning method of the embodiment is applied, recommendation performance is improved regardless of the sequence length of the data.

[0034] Figure 5 is a block diagram of a hierarchical contrastive learning device of an embodiment.

[0035]

[0036] In describing embodiments of the present invention, if a detailed description of a known technology related to the present invention is judged to unnecessarily obscure the gist of the present invention, the detailed description will be omitted. In addition, the terms described below are terms defined in consideration of their functions in the present invention, and this may vary depending on the intention or custom of the user or operator. Therefore, the definitions should be made based on the contents throughout this specification. The terms used in the detailed description are only for the purpose of describing embodiments of the present invention and should never be construed as limiting. Unless clearly used otherwise, the singular form includes the plural form. In this description, expressions such as "comprises" or "having" are intended to indicate certain features, numbers, steps, operations, elements, parts or combinations thereof, and should not be construed to exclude the presence or possibility of one or more other features, numbers, steps, operations, elements, parts or combinations thereof other than those described.

[0037] Terms containing ordinal numbers, such as "first" and "second," may be used to describe various components, but these components are not limited by these terms. These terms may only be used in a nominal sense to distinguish one component from another, and their ordinal meaning is determined not from the names but from the context of the description.

[0038] The term “and / or” is used to include any combination of the multiple items being referred to. For example, “and / or B” means all three cases, “”“”“and B”.

[0039] When it is said that a component is "connected" or "connected" to another component, it should be understood that it may be directly connected or connected to that other component, but there may also be other components in between.

[0040] Hereinafter, specific embodiments of the present invention will be described with reference to the drawings. The following detailed description is provided to facilitate a comprehensive understanding of the methods, devices, and / or objects described herein. However, these are merely examples and the present invention is not limited thereto.

[0041] Figure 1 is a conceptual diagram showing the overall framework of the hierarchical contrastive learning device of the embodiment.

[0042] Referring to FIG. 1, the hierarchical contrastive learning device of the embodiment may include a hierarchical data augmentation model (100) and a hierarchical contrastive learning model (200).

[0043] A hierarchical contrastive learning model (200) may include a first learning block (210), a second learning block (220), and a third learning block (230). For example, the first learning block may include an encoder, and the second and third learning blocks may include a transformer and a position-wise feed-forward network (PFFN).

[0044] The hierarchical data augmentation model (100) can hierarchically perform transformation operations on items arranged in chronological order. The hierarchical data augmentation model (100) can perform augmentation operations on sequence data input from a single user.

[0045] Augmentation operations on sequence data can be performed multiple times. The hierarchical data augmentation model (100) can perform multiple augmentation operations on a single sequence data.

[0046] The hierarchical data augmentation model (100) applies a first transformation operation to one sequence data to create the first augmented sequence data ( ) can be generated. In addition, the hierarchical data augmentation model (100) applies a second transformation operation, which is the same as or different from the first transformation operation, to one sequence data to generate second augmented sequence data (( ) can be created.

[0047] The hierarchical data augmentation model (100) is a first augmented sequence data ( ) can be subjected to another transformation operation. That is, a total of two or three transformation operations or augmentation operations can be applied to the sequence data.

[0048] The hierarchical data augmentation model (100) is a second augmented sequence data ( ) can be subjected to another transformation operation. That is, a total of two or three transformation operations or augmentation operations can be applied to the sequence data.

[0049] First augmented sequence data ( ) data with another transformation operation applied to it. , second augmented sequence data (( ) data with another transformation operation applied to it. ) can be said.

[0050] First augmented sequence data ( ) and the second augmented sequence data ( ) can be the same or different. That is, each augmentation operation is performed hierarchically, and the hierarchical data augmentation model (100) can generate various data depending on the number of augmentations applied. In addition, each data transformation operation can be randomly selected and can have the same or different forms.

[0051] The hierarchical contrastive learning model (200) can learn a model by converting data pairs obtained through the hierarchical data augmentation described above into representation vectors.

[0052] The hierarchical contrastive learning model (200) can learn by transmitting data only up to a specific learning block depending on the number of transformation operations or augmentation operations applied.

[0053] A pair of data, each of which has been subjected to a different data transformation, passes through the first learning block (210), and the first learning block (210) can generate an expression vector for the transformed pair of data.

[0054] A pair of data, each of which has been subjected to two different data transformations, passes through the second learning block (220), and the second learning block (220) can generate an expression vector for the pair of transformed data.

[0055] A pair of data, each of which has undergone three different data transformations, passes through to the third learning block (230), and the third learning block (230) can generate an expression vector for the transformed pair of data.

[0056] The hierarchical contrastive learning model (200) can count a pair of transformation operations as one number of times. If a pair of transformation operations is one time, the hierarchical contrastive learning model (200) can pass the data only up to the first learning block (210). If a pair of transformation operations is two times, the data can pass only up to the second learning block (220). If a pair of transformation operations is three times, the data can pass up to the third learning block (230).

[0057] According to one embodiment, data to which a first transformation operation is applied is sequence data. It can be said that the data is a second transformation operation that is the same as or different from the first transformation operation applied to the sequence data. It can be said that.

[0058] According to one embodiment Data to which a third transformation operation, different from the first transformation operation, is applied It can be said, Data that has been subjected to a fourth transformation operation that is different from the second transformation operation It can be said that.

[0059] That is, as described above All transformation operations applied to may be different transformations. In addition, the above-described The transformation operations applied to may all be different transformations. Each of the first transformation operation, the second transformation operation, the third transformation operation, and the fourth transformation operation described above may all be different transformation operations.

[0060] Each of the first conversion operation, the second conversion operation, the third conversion operation, and the fourth conversion operation can be randomly selected.

[0061] The hierarchical contrastive learning model (200) can learn the expression vector generated in each learning block and produce a contrastive loss, i.e., each loss value.

[0062] The first learning block extracts the data from which the expression vector is extracted. It can be said that. The first and second learning blocks extract the data from which the expression vectors are generated. It can be said that.

[0063] The first learning block extracts the data from which the expression vector is extracted. It can be said that. The first and second learning blocks extract the data from which the expression vectors are generated. It can be said that.

[0064] Additionally, the hierarchical contrastive learning model (200) can produce a total loss value as shown in the following mathematical expression 1.

[0065]

[0066] Here, λ represents the weight parameter, is the total loss value, is the loss value of data that passed through the first learning block (210), is the loss value of data that passed through the second learning block (220). refers to the loss value of data that passed through the third learning block (230).

[0067] The hierarchical data augmentation model (100) and the hierarchical contrastive learning model (200) of the embodiment may be referred to as a hierarchical contrastive learning device (Hierarchical Contrastive Learning with Multiple Augmentation for Sequential Recommendation, HCLRec). Furthermore, a method or device for performing sequential recommendation using the hierarchical contrastive learning device of the embodiment may be referred to as a hierarchical contrastive learning method and device for sequential recommendation.

[0068] Figure 2 is a flowchart of the hierarchical contrastive learning method of the embodiment.

[0069] Referring to FIG. 2, the hierarchical contrastive learning method of the embodiment applies a first transformation operation to the input sequence data to augment the data into first augmented sequence data ( ) generating step (S110), applying a second transformation operation that is the same as or different from the first transformation operation to the sequence data to augment the data to generate second augmented sequence data ( ) may include a step of generating (S120), a step of counting the number of times the first conversion operation is applied and the second conversion operation is applied as one (S130), and a step of determining the number of times the sequence data is learned according to the count number (S140).

[0070] In the contrastive learning method of the embodiment, data to which a pair of transformation operations are applied can be passed to the first learning block (210), data to which two pairs of transformation operations are applied can be passed to the second learning block (220), and data to which three pairs of transformation operations are applied can be passed to the third learning block (230).

[0071] Additionally, a pair of transformation operations may be a first transformation operation and a second transformation operation that is the same as or different from the first transformation operation.

[0072] Additionally, the two pairs of transformation operations can be a first transformation operation, a second transformation operation that is the same as or different from the first transformation operation, a third transformation operation that is different from the first transformation operation, and a fourth transformation operation that is different from the second transformation operation, for a total of four transformation operations.

[0073] Additionally, the three pairs of transformation operations may include the first transformation operation, the third transformation operation described above, a fifth transformation operation that is different from all of the first transformation operation, the third transformation operation, and a sixth transformation operation that is different from all of the second transformation operation, the fourth transformation operation, and a total of six transformation operations.

[0074] The first to sixth transformation operations can be set randomly.

[0075] Figure 3 is a conceptual diagram illustrating the hierarchical contrastive learning device of the embodiment in more detail.

[0076] Below, the transformation operation performed in the hierarchical data augmentation model (100) is described in more detail.

[0077] The augmented data set can be expressed by the following mathematical expression (2).

[0078]

[0079] Here, denotes the i-th augmentation operation, and denotes the total number of applicable operations.

[0080] The Removeone method can be used to hierarchically generate data with multiple operations applied. Removeone can generate the next data using a single augmentation operation from a set of methods excluding the previously used augmentation operation.

[0081] This method can be expressed by the following mathematical formula 3.

[0082]

[0083] Here, means sequence data to which m transformation operations have been applied. In addition, class represents a positive pair to which m transformation operations have been applied. Short sequence data is more sensitive to transformation operations and therefore must be processed in detail. Sequence data can be divided into short sequence data and relatively long sequence data. Different augmentation operations can be applied to the sequence depending on its length. In this case, the order of augmentation operations is not defined, and the augmentation operation can be randomly selected each time. In other words, if the order of augmentation operations is defined, the performance of sequential recommendation deteriorates, so a random augmentation operation is selected.

[0084] Below, the contrastive learning performed in the hierarchical contrastive learning model (200) is described in more detail.

[0085] Data pairs with m transformation operations applied class Is and can be expressed as . Data subjected to m transformation operations may exhibit nonlinear behavior. In this case, the encoder may have difficulty learning the data due to the nonlinear behavior. To solve this problem, a neural network block consisting of a transformer and a PFFN can be added.

[0086] The additional block can be expressed as mathematical expression 4 below.

[0087]

[0088] Here, SASRec stands for self-attentive sequential recommendation. While additional blocks can mitigate the problem of data learning due to nonlinear behavior, they increase memory complexity and require more time to build. Furthermore, the problem of sparse interactions becomes more severe as the model complexity increases. Therefore, it is more efficient to introduce a less complex model into sequential recommendation and train it through backpropagation.

[0089] Contrast loss can be calculated using the InfoNCE (info noise contrastive estimation) loss. By applying a transformation operation to the same sequence data, the resulting data can be considered positive pairs, while the remaining sequence data can be considered negative pairs.

[0090] Contrastive learning can be expressed in mathematical formula 5 as follows.

[0091]

[0092] Here, represents the temperature parameter, and B represents the batch size.

[0093] In an embodiment, the contrast loss value occurring in the existing sequential recommendation can be minimized in a relationship as shown in the following mathematical expression 6.

[0094]

[0095] is the main object for obtaining optimal parameters. Therefore, There is no need to assign weights.

[0096] Figure 4 is a graph explaining that when the contrastive learning method of the embodiment is applied, recommendation performance is improved regardless of the sequence length of the data.

[0097] Table 1 below shows the performance comparison between the existing framework and the Hierarchical Contrastive Learning with Multiple Augmentation for Sequential Recommendation (HCLRec) of the embodiment on four real-world benchmark data sets: Beauty, Sports, Toys, and Yelp.

[0098]

[0099] Referring to Table 1, it can be seen that the performance of HCLRec is improved compared to the existing frameworks CL4SRec, CoSeRec, ICLRec, and CBiT.

[0100] Additionally, referring to the graph shown in FIG. 4, the performance of the framework can be improved by applying the hierarchical data augmentation and hierarchical contrastive learning models of the embodiment.

[0101] Referring to Figure 4, it can be seen that HCLRec shows higher performance than existing frameworks regardless of whether the user uses the platform less or has short sequence data length.

[0102] Figure 5 is a block diagram of a hierarchical contrastive learning device of an embodiment.

[0103] Referring to FIG. 5, the hierarchical contrast learning device of the embodiment may include an interface unit (510), a processor (520), and a memory (530).

[0104] The interface unit (510) can receive sequence data from a user. In addition, the interface unit (510) can present the user with the next sequence of user-preferred items predicted through the hierarchical contrastive learning model (200).

[0105] The interface unit (510) may be implemented as, for example, a touch screen, a touch pad, a key pad, a jog wheel, a jog switch, etc. In addition, the interface unit (510) may recognize the user's voice using AI (artificial intelligence) and receive sequence data from the user.

[0106] The processor (520) applies a first transformation operation or a second transformation operation to the acquired sequence data to generate first augmented sequence data ( ) or second augmented sequence data ) can be created.

[0107] The processor (520) and can be transmitted to the hierarchical contrastive learning model (200), and each loss value can be calculated.

[0108] The program executed by the processor (520) may include instructions for converting the input sentence into a feature through an encoder, generating a sequence composed of tokens having the same meaning but different expressions from the converted feature through a decoder, and generating a predicted text that is rewritten from the input text based on attention according to context.

[0109] The processor (520) Apply a third transformation operation that is different from the first transformation operation to the data that passed the first learning block, A fourth transformation operation, which is different from the second transformation operation, can be applied to data that has passed the first learning block.

[0110] Embodiments according to the present specification may be implemented by various means, for example, hardware, firmware, software, or a combination thereof. In the case of hardware implementation, an embodiment of the present specification may be implemented by one or more application specific integrated circuits (ASICs), digital signal processors (DSPs), digital signal processing devices (DSPDs), programmable logic devices (PLDs), field programmable gate arrays (FPGAs), processor controllers, microcontrollers, microprocessors, etc. In the case of firmware or software implementation, an embodiment of the present specification may be implemented in the form of a module, procedure, function, etc. that performs the capabilities or operations described above. Software code may be stored in the memory (530) and executed by the processor (520). The memory (530) may be located inside or outside the processor (520) and may exchange data with the processor (520) by various means already known.

[0111] Meanwhile, the embodiments of the present specification can be implemented as computer-readable codes on a computer-readable recording medium. Computer-readable recording media include all types of recording devices that store data that can be read by a computer system. Examples of computer-readable recording media include ROMs, RAMs, CD-ROMs, magnetic tapes, floppy disks, optical data storage devices, etc. In addition, the computer-readable recording media can be distributed across network-connected computer systems, so that the computer-readable codes can be stored and executed in a distributed manner. In addition, functional programs, codes, and code segments for implementing the embodiments can be easily inferred by programmers in the technical field to which the present specification pertains.

[0112] The foregoing description of the present invention is provided for illustrative purposes only. Those skilled in the art will readily appreciate that the present invention can be readily modified into other specific forms without altering the technical spirit or essential characteristics of the present invention. Therefore, the embodiments described above should be understood as illustrative in all respects and not restrictive.

[0113] The scope of the present invention is indicated by the claims described below rather than the detailed description above, and all changes or modifications derived from the meaning and scope of the claims and their equivalent concepts should be interpreted as being included in the scope of the present invention.

Claims

1. First augmented sequence data (augmented by applying the first transformation operation to sequence data input from one user) ) to create a step; Second augmented sequence data obtained by augmenting data by applying a second transformation operation that is the same as or different from the first transformation operation to the above sequence data ( ) to create a step; Above and above a step of transmitting the training data to a hierarchical contrastive learning model including at least one learning block and producing at least one loss value; and A hierarchical contrastive learning method for sequential recommendation, comprising a step of learning the hierarchical contrastive learning model based on a total loss value obtained by summing the above loss values.

2. In paragraph 1, A hierarchical contrastive learning method for sequential recommendation, further comprising a step of counting the number of times the first transformation operation is applied and the number of times the second transformation operation is applied as one.

3. In paragraph 2, A hierarchical contrastive learning method in which at least one loss value is calculated according to the above count number.

4. In paragraph 2, Based on the above count number A hierarchical contrastive learning method for sequential recommendation, comprising the step of determining the number of times the sequence data passes through the learning block and training a hierarchical contrastive learning model.

5. In paragraph 1, After the step of calculating at least one loss value, Further comprising a step of calculating a total loss value by weighting at least one loss value, A hierarchical contrastive learning method in which the step of learning the hierarchical contrastive learning model updates the weights by backpropagation based on the total loss value and learns the hierarchical contrastive learning model.

6. A method for sequentially recommending user-preferred items using a deep learning model. A step of receiving sequence data from a user; A step of predicting the next order of user preference items using the learned hierarchical contrastive learning model; and including a step of recommending the predicted preferred item to the user; The above hierarchical contrastive learning model includes at least one learning block, First augmented sequence data obtained from one user by applying a first transformation operation to the data ( ) and create Second augmented sequence data obtained by augmenting data by applying a second transformation operation that is the same as or different from the first transformation operation to the above sequence data ( ) and create Above and above A sequential recommendation method configured to train the learning block by transmitting the learning block to at least one of the learning blocks.

7. In paragraph 6, The above hierarchical contrastive learning model is, A sequential recommendation method in which the number of times the first transformation operation is applied and the number of times the second transformation operation is applied are counted as one number.

8. In paragraph 7, The above hierarchical contrastive learning model is, Based on the above count number A sequential recommendation method set to be learned by determining the number of times the above learning block passes through the above sequence data.

9. In one or more non-transitory computer-readable media storing one or more instructions, The one or more instructions executable by one or more processors, First augmented sequence data (which is data augmented by applying the first transformation operation to sequence data input from one user) ) and create Second augmented sequence data obtained by augmenting data by applying a second transformation operation that is the same as or different from the first transformation operation to the above sequence data ( ) and create Above and above A computer-readable recording medium for transmitting a hierarchical contrastive learning model including at least one learning block to train the hierarchical contrastive learning model.

10. Memory configured to store multiple instructions; A hierarchical data augmentation model that performs data augmentation on sequence data input from a single user; A hierarchical contrastive learning model comprising at least one learning block and trained using data transmitted from the hierarchical data augmentation model; and comprising a processor functionally coupled with the above memory, When the above processor executes the above plurality of instructions, First augmented data obtained by applying a first transformation operation to the above sequence data ( ) and create Second augmented data obtained by augmenting data by applying a second transformation operation that is the same as or different from the first transformation operation to the above sequence data ( ) and create Above and above transmits the training data to a hierarchical contrastive learning model including at least one learning block, and produces at least one loss value; A hierarchical contrastive learning device configured to learn the hierarchical contrastive learning model based on a total loss value obtained by summing at least one of the above loss values.

11. In paragraph 10, The hierarchical contrastive learning model is a hierarchical contrastive learning device including a first learning block in which data is learned, a second learning block in which data that has passed through the first learning block is learned, and a third learning block in which data that has passed through the second learning block is learned.

12. In paragraph 11, The above processor, Above Applying a third transformation operation that is different from the first transformation operation, Above A hierarchical contrastive learning device that applies a fourth transformation operation that is different from the second transformation operation.

13. In paragraph 11, The above first learning block includes an encoder, The second learning block includes a transformer and a position-wise feed-forward network (PFFN), The third learning block is a hierarchical contrastive learning device including a transformer and a PFFN.

14. In paragraph 12, The above processor, The number of times the first conversion operation is applied and the number of times the second conversion operation is applied are counted as one number, The number of times the third conversion operation is applied and the number of times the fourth conversion operation is applied are counted as one number, Based on the above count number, A hierarchical contrastive learning device that determines whether the above sequence data passes the first learning block, the second learning block, or the third learning block.

15. In paragraph 14, The above processor, A hierarchical contrastive learning device set to produce at least one loss value according to the above count number.

16. In paragraph 10, The above processor, Calculating a total loss value by weighting at least one of the above loss values, A hierarchical contrastive learning device set to learn by updating the weights through backpropagation based on the total loss value.

Images