Recommendation method and system based on item-ranking promotion

Abstract

An item ranking promotion-based recommendation method and a system thereof are disclosed. An item ranking promotion-based recommendation method performed by a recommendation system according to one embodiment may include: a step of training a recommendation model to recommend items with additional value to a user while maintaining recommendation accuracy by utilizing items used by the user and items not used by the user; and a step of recommending items with additional value through the trained recommendation model.

Description

Technology Field

[0001] The following description concerns item recommendation technology. Background Technology

[0003] The use of recommendation systems is becoming increasingly important for various online business platforms (e.g., Netflix, Amazon). By accurately recommending items that users are likely to prefer, these systems enhance user satisfaction, thereby attracting new users or preventing existing users from switching to competing platforms. As competition among platforms intensifies, each platform is paying greater attention to securing and providing items that differentiate it from competitors. Examples of such items include exclusive products contracted exclusively with specific suppliers by a platform, or content produced and provided in-house by an OTT platform (e.g., Netflix Originals, Amazon Prime Originals). The more widely these items become known to users, the easier it is for a platform to showcase its distinctiveness or uniqueness to a large audience, which significantly aids in acquiring new users and preventing existing ones from leaving. In other words, from a platform's perspective, these items can be regarded as capable of creating additional value (i.e., generating additional revenue) compared to other items.

[0004] In this regard, recommending items of additional value to many users can be beneficial to the platform in terms of increasing revenue. However, conventional technology has not considered the problem of recommending as many items of additional value as possible while maintaining overall recommendation accuracy. The problem to be solved

[0006] It is possible to provide a recommendation system and method that can recommend items with additional value to many users while maintaining a high level of overall recommendation accuracy. means of solving the problem

[0008] A recommendation method based on item ranking promotion performed by a recommendation system may include: a step of training a recommendation model to recommend items with additional value to the user while maintaining recommendation accuracy by using items the user has used and items the user has not used; and a step of recommending items with additional value through the trained recommendation model.

[0009] The above learning step may include the step of constructing a loss function to maintain recommendation accuracy in the recommendation model and a loss function to recommend items with additional value to the user.

[0010] The above learning step may include a step of training a recommendation model in a direction that can accurately predict the user's preferences using items the user has used.

[0011] The above learning step includes the step of obtaining a prediction score between 0 and 1 for items used by the user through the learning of the recommendation model, and the closer the obtained prediction score is to 1, the more the loss of the loss function configured to maintain recommendation accuracy may be minimized.

[0012] The above learning step includes a step of training a model using items used by the user and additional valuable items not used by the user, in a direction in which additional valuable items not used by the user are promoted to a ranking as high as that of items used by the user, and as the ranking of additional valuable items not used by the user is promoted, the loss of the loss function can be minimized.

[0013] The loss function for maintaining the above recommendation accuracy uses cross-entropy loss, and the loss function for recommending items with additional value to the user can use list-specific loss among rank learning-based loss functions.

[0014] The above learning step may include a step of filtering the user from a set of candidate users for item ranking promotion when, through a separate recommendation model that has been pre-trained, an item of additional value among the items not used by the user is predicted to be an item of lower than the user's preset rank.

[0015] The above learning step may include a step of exploring the rankings within a target list of additionally valuable items that the user has not used through a dynamic random promotion strategy.

[0016] The above learning step may include a step of sorting items used by the user and valuable items not used by the user based on the prediction score obtained through the learning model.

[0017] The above learning step may include a step of randomly promoting valuable items that each user has not used to a rank higher than the current rank of the valuable items that the user has not used.

[0018] The above learning step may include a step of placing items used by a user who was in a rank corresponding to a promoted position, due to the promotion of valuable items that the user has not used, into a lower rank.

[0019] The above-mentioned items with additional value may include exclusive items provided by the platform through an exclusive contract with a specific supplier, or items produced and provided by the platform itself.

[0020] To execute an item promotion-based recommendation method on the recommendation system, a computer program stored on a non-transient computer-readable recording medium may be included.

[0021] The recommendation system may include a model training unit that trains a recommendation model to recommend items with additional value to the user while maintaining recommendation accuracy by utilizing items the user has used and items the user has not used; and an item recommendation unit that recommends items with additional value through the trained recommendation model. Effects of the invention

[0023] It can improve the exposure rate and recommendation accuracy of items with additional value. Consequently, it allows companies to showcase their differentiation and uniqueness to users on online business platforms, thereby acquiring new users or preventing the churn of existing ones. Brief explanation of the drawing

[0025] FIG. 1 is a diagram illustrating an item ranking promotion-based recommendation operation in one embodiment. FIG. 2 is a block diagram illustrating the configuration of a recommendation system in one embodiment. FIG. 3 is a flowchart illustrating a recommendation method based on item ranking promotion in one embodiment. FIGS. 4 to 8 are graphs illustrating experimental results of an item ranking promotion-based recommendation method in one embodiment. Specific details for implementing the invention

[0026] Hereinafter, embodiments will be described in detail with reference to the attached drawings.

[0028] FIG. 1 is a diagram illustrating an item ranking promotion-based recommendation operation in one embodiment.

[0029] In the embodiments, the problem of recommending items with additional value to as many users as possible (Objective 2) while maintaining a high level of overall recommendation accuracy (Objective 1) is defined as an item-ranking promotion (IRP) problem, and items with additional value are defined as RP items in the sense that they are targets of the IRP.

[0030] First, let U, T, and V represent the total users, total items, and m items with additional value (i.e., V ⊂ T), respectively. First, for a user u ∈ U, let the set of all items observed / unobserved by u be, respectively Represented as, the set of unused items of u with additional value It is represented as, and also, for items with an additional value of v ∈ V, the set of unused users It is represented as , which represents the set of candidate users who can recommend v. In addition, the set of the top N items recommended for u is For u, the set of related items (i.e., ground truth) It is represented as follows. Finally, the problem is formulated so that the recommendation system satisfies both of the following two goals.

[0031] Objective 1 is for u ∈ U In item ∈ It maintains a high level of recommendation accuracy, including...

[0032] Objective 2 is to recommend v to users who have not used v as much as possible for v ∈ V (i.e., users of (includes v)

[0033] To achieve Goal 1, the recommendation system must be able to find items the user might prefer by learning from items previously used by the user through a recommendation model, which is the same as the general goal of recommendation systems, and to achieve Goal 2, the recommendation model must have items ∈ By predicting u's preference for to be sufficiently high You must learn so that you can include it.

[0034] To this end, the recommendation system proposes a new learning method for a new u that can train a recommendation model to achieve both goals.

[0035] Mathematical formula 1:

[0036]

[0037] Here, represents the loss required to maintain the overall accuracy of the model, and is an item ∈ Represents a loss to increase the recommendation of, is an item ∈ Weights for controlling the recommendation level (0≤ It represents ≤1). As ≥ increases, the recommendation model items ∈ It is trained to increase the recommendations of. Likewise, As ∈ becomes smaller, the recommendation model items ∈ It is trained to reduce the recommendation of.

[0038] As in mathematical formula 1 and While determining the balance between the two goals by combining them (i.e., while increasing losses too much), It ensures that the recommendation model is trained (to avoid optimizing). Therefore, the recommendation system can obtain recommendation results that satisfy both goals through a model trained in this balanced way. In addition, the proposed training method is orthogonal to all recommendation models (i.e., model agnostic).

[0039] The recommendation system is the loss function of Goal 1 You can train a recommendation model using this. As a general goal of recommendation systems, it focuses on training a model to accurately predict u's preference for unused items. At this time, It can use the commonly used cross-entropy loss.

[0040] Mathematical formula 2:

[0041]

[0042] Here, Q represents the model to be trained, and Q(t) is the predicted score for item t It is represented as. That is, the item ∈ used by Q / Unused items ∈ As u’s score for each is predicted to be close to 1 / 0 is minimized. Therefore, the recommendation model Goal 1 can be achieved by learning to search for items that may be preferred (i.e., items predicted to have a score close to 1).

[0043] The recommendation system is the loss function of Objective 2 A recommendation model can be trained using . The recommendation system uses items ∈ used by u Items as high as A recommendation model can be trained to predict the score of u for . Specifically, the recommendation system trains for all items ∈ ∪ When sorted based on prediction scores obtained by a recommendation model, item ∈ This item ∈ The goal is to achieve a high ranking. To accomplish this, recommendation systems use a ranking learning method and train models to learn the rank differences between items using list-specific loss. ListMLE is formulated as follows:

[0044] Mathematical formula 3:

[0045]

[0046] Here, represents a list of items sorted in the order of u's actual preferences (i.e., actual list), and Is It represents the item located at rank i within, and n is It indicates the size of. The loss per list is items with a pre-set rank or higher (top rank) in the actual list (e.g., Q( ), Q( The predicted score for )) is for items with a pre-set rank or lower (lower rank) (e.g., Q( ), Q( The higher the predicted score for )), the smaller it becomes.

[0047] Some items ∈ This item ∈ It is assumed that the list with a higher rank is 'intentionally' used as the actual list. In this case, according to Equation 3, minimizing the loss is item ∈ Item with a higher score ∈ It is equivalent to training a model to predict. Based on this intuition, Equation 3 is redefined as follows: Formalizes.

[0048] Mathematical formula 4:

[0049]

[0050] Here, is an item ∈ ∪ Represents a target list composed of, where some items ∈ is an item ∈ It ranks higher. If you use as an actual measurement list, (For example, Q( ), Q( ), here , ∈ The higher the Q-based prediction score for top-ranked items in ), the It becomes smaller. As a result A recommendation model trained through for item ∈ for u It is highly likely that [it] will be recommended, and accordingly, Goal 2 can be achieved.

[0051] The recommendation system defines the training plan of the recommendation model as Item Ranking Promotion (IRP), where items with additional value , that is, the target of the IRP is considered as an RP item.

[0052] Recommendation systems can propose two strategies to address two issues: whether the IRP generated during the problem-solving process will be effective for all users, and where to place (position) the RP items within the list for learning during the IRP.

[0053] Recommendation systems can filter candidate users. Basically, RP item v (i.e., user ∈ Users who have not used ) are IRP candidates for v. However, performing IRP on candidates who are unlikely to prefer v degrades the overall recommendation accuracy of the recommendation model. To prevent this degradation in recommendation accuracy, we identify candidate users and Excluding The system is configured as follows. In other words, the recommendation system can exclude a user from the set of candidate users for IRP through a separate, pre-trained recommendation model if the RP item is predicted to be a low-ranking item among the items the user has not yet used. By identifying candidate users who predicted RP items as low-ranking and excluding (filtering) them from the set of candidate users for IRP, it becomes possible to determine the set of candidate users for each RP-item where IRP can be effectively applied.

[0054] Mathematical formula 5:

[0055]

[0056] Here, u obtained from a separate pre-trained model (i.e., Indicates the rank of v among all unused items by ), represents the proportion of items to be excluded. The set of candidate users for whom v's IRP can be effectively performed through this strategy. By determining it, accuracy is hardly lost even with IRP.

[0057] The recommendation system can perform dynamic random promotion. According to Strategy 1, the RP item for u is In items v containing u (e.g., v1, v2, where u ∈ , u ∈ It is determined by ). However, there may be trade-offs depending on where RP items are positioned in the target list for training. For example, if RP items are placed at the top rank, the probability of them being recommended to user u increases; however, this results in the overall recommendation accuracy for user u being lower than when IRP is not performed, making it impossible to satisfy Goal 1. Conversely, if RP items are placed at the bottom rank, the probability of them being recommended to user u decreases, making it impossible to satisfy Goal 1. To avoid these two extreme cases, the recommendation system can propose a dynamic promotion strategy to allow the recommendation model to find the rank of RP items within the target list. In other words, the dynamic promotion strategy dynamically reassigns items from their current rank to a higher rank within the target list while the recommendation model is learning. Specifically, for an RP item of rank i (>1), rank j is randomly selected from ranks 1 through i, and v is reassigned for rank j. That is, it is promoted to rank (ij).

[0058] First, the recommendation system can sort the items used by the user and RP items based on the prediction scores of the recommendation model being trained. For each RP item, the recommendation system randomly promotes it to a rank higher than its current rank, and due to this promotion, the item originally in that rank can be placed in a lower rank. Through such dynamic random promotion, the recommendation system can determine the ranking of RP items for the user in a way that strikes a balance between increasing the recommendation of RP items and maintaining the overall recommendation accuracy of the model.

[0059] As shown in Fig. 1, for the user, a total of 5 used items (i.e., t a , t b , t c , t d , t e) and 3 RP items (i.e., v a , v b , v c Let us assume that there is ). First, the recommendation system can sort these 8 items based on the prediction scores obtained from the current recommendation model and construct a prediction list. Next, for each RP item v, the recommendation system can randomly promote v from its current rank within the prediction list. For example, v originally ranked 5th a is the 2nd priority randomly selected from 1st to 5th priority, v b It is promoted from the original 7th rank to the 3rd rank, which is randomly selected from ranks 1 through 7. Finally, by repeatedly performing the above steps You can train a recommendation model with a target list of RP items promoted from DRP while trying to minimize it. As a result, as long as this promotion of RP items does not lower the overall recommendation accuracy, the ranking of RP items increases as training progresses.

[0060] FIG. 2 is a block diagram illustrating the configuration of a recommendation system in one embodiment, and FIG. 3 is a flowchart illustrating an item ranking promotion-based recommendation method in one embodiment.

[0061] The processor of the recommendation system (100) may include a model learning unit (210) and an item recommendation unit (220). These components of the processor may be representations of different functions performed by the processor according to control commands provided by program code stored in the visualization analysis system. The processor and the components of the processor may control the recommendation system for steps (310 to 320) included in the item ranking promotion-based recommendation method of FIG. 3. At this time, the processor and the components of the processor may be implemented to execute instructions according to the code of the operating system included in the memory and the code of at least one program.

[0062] The processor can load program code stored in a file of a program for an item ranking promotion-based recommendation method into memory. For example, when a program is executed in a recommendation system, the processor can control the recommendation system to load program code from a file of a program into memory under the control of the operating system. At this time, the model learning unit (210) and the item recommendation unit (220) may each be different functional representations of the processor for executing commands of corresponding parts of the program code loaded into memory to execute subsequent steps (310 to 320).

[0063] In step (310), the model training unit (210) can train a recommendation model to recommend items with additional value to the user while maintaining recommendation accuracy by using items used by the user and items not used by the user. The model training unit (210) can configure a loss function to maintain recommendation accuracy and a loss function to recommend items with additional value to the user in the recommendation model. At this time, the loss function to maintain recommendation accuracy uses cross-entropy loss, and the loss function to recommend items with additional value to the user can use a list-by-list loss among rank-learning-based loss functions. The model training unit (210) can train the recommendation model in a direction that can accurately predict the user's preference using items used by the user. Through the training of the recommendation model, the model training unit (210) can obtain a prediction score between 0 and 1 for items used by the user. At this time, the closer the obtained prediction score is to 1, the more the loss of the loss function configured to maintain recommendation accuracy can be minimized. The model training unit (210) can train the model by using items used by the user and additional valuable items not used by the user, in a direction in which additional valuable items not used by the user are promoted to a ranking as high as that of items used by the user. At this time, the loss of the loss function can be minimized as the ranking of additional valuable items not used by the user is promoted. The model training unit (210) can filter the user from the set of candidate users for item ranking promotion if, through a separate recommendation model pre-trained, additional valuable items among the items not used by the user are predicted to be items with a ranking lower than or equal to the user's preset ranking. The model training unit (210) can explore the ranking within the target list of additional valuable items not used by the user through a dynamic random promotion strategy.The model learning unit (210) can sort the items used by the user and the valuable items not used by the user based on the prediction score obtained through the learning model. The model learning unit (210) can randomly promote each valuable item not used by the user to a rank higher than the current rank of the valuable item not used by the user. Due to the promotion of the valuable items not used by the user, the model learning unit (210) can place the items used by the user that were in the rank corresponding to the promoted position to a lower rank.

[0064] In step (320), the item recommendation unit (220) can recommend items with additional value through a learned recommendation model. At this time, items with additional value may include exclusive items provided by the platform through an exclusive contract with a specific supplier, or items produced and provided by the platform itself.

[0065] FIGS. 4 to 8 are graphs illustrating experimental results of an item ranking promotion-based recommendation method in one embodiment.

[0066] Experiments may be conducted to evaluate the technical effectiveness of the recommendation model. Two real-world datasets widely used in recommendation systems, MovieLens 1M (ML1M) and Amazon Beauty (ABeauty), can be used as datasets. Here, user ratings are converted into binary values ​​of 1 / 0 for observed / unobserved ratings, respectively.

[0067] The evaluation protocol randomly selects items with a rating of 5 or higher after excluding the top 10% most popular items for RP items. Here, m is set to 1, 5, 10, and 20. For evaluation, the ratings can be randomly split in a 6:2:2 ratio for the training, validation, and test sets. The accuracy of top N recommendations can be measured using the commonly used Normalized Discounted Cumulative Gain (NDCG). Additionally, to estimate the increase in recommendations for RP items compared to the original model (i.e., baseline), a metric (i.e., ERgain@N) can be defined as follows.

[0068] Mathematical formula 6:

[0069]

[0070] Here, Q represents the baseline and Q' represents a model trained with a specific IRP method, such as the proposed learning method. ER@N represents the exposure rate of each item v, where user ∈ v represents the portion of recommended users. MF, NeuMF, and CDAE are used as standard models.

[0071] Since the IRP problem of increasing RP item recommendations while maintaining accuracy has never been solved before, a simple but competitive method is constructed as follows for comparison.

[0072] Promotion-by-Postprocessing (PP) identifies users predicted by the trained model to have RP items in a high rank after model training is complete. Then, for such users u, PP [provides] RP items RP items by placing them in some randomly selected ranks within Include in.

[0073] Additionally, the following state-of-the-art value-aware recommendation method can be adopted for comparison. PE-LTR trains a model with two losses in terms of accuracy and profit maximization, aiming for Pareto efficiency.

[0074] Referring to Figure 4, this is a graph comparing the NDCG@10 method with other @10 methods. RQ 1: These are the experimental results regarding whether our method achieves both objectives better than the competing method. Figure 4 shows the change in overall accuracy according to the increase in RP item recommendations for the three methods: PP, PE-LTR, and Ours; the purple dotted line represents the accuracy of the baseline without IRP applied. First, ER gain For most levels of @10, PP demonstrates superior accuracy compared to PELTR. This confirms that training a model aimed at maximizing each user's benefit, as in PE-LTR, may not be effective in IRP. Second, regarding our method, the model and ER gain Regardless of the @10 level, it universally and consistently outperforms both PP and PE-LTR in terms of NDCG@10. In other words, a learning method that dynamically promotes RP items during model training is more beneficial than PP, which performs IRP through post-processing. Third, ER gain @10 remains below a certain level (approximately 50), whereas the accuracy of our method is higher than or similar to the baseline. Furthermore, it demonstrates that it not only increases RP item recommendations but also appropriately recommends RP items to a segment of users likely to prefer those items.

[0075] Referring to Fig. 5, NDCG@10 and ER gain @10 This is a graph that changes according to (ML1M). RQ 2: (Strategy 1) These are the experimental results regarding how effective filtering candidate users is in maintaining accuracy. In Figure 5, the filtering ratio NDCG@10 and ER obtained by modifying gain Shows @10. As increases, the number of candidate users filtered decreases, so =100 represents the case without filtering, which indicates the lowest accuracy among all models. Specifically, for NeuMF NDCG@10 for =10, 30, 50, and 70 is approximately 15%, 17%, 16%, and 12% higher, respectively, than the case without filtering. Since it shows the highest accuracy among all models, in all other experiments I set it to 30.

[0076] Referring to Fig. 6, NDCG@10 and ER have different ranking methods. gain This is a graph representing @10 (ML1M). This is the experimental result regarding RQ 3: (Strategy 2) How effective is DRP in determining the rank of RP items? Figure 6 shows NDCG@10 and ER obtained in different ways regarding the rank of RP items within the target list. gain Shows @10. 'Top' promotes RP items to the highest rank within the target list, while 'Bottom' places RP items at the lowest rank they can be promoted to. 'Top' has the highest ER. gain While it shows @10 but the lowest NDCG@10, 'Bottom' shows the highest NDCG@10 but the lowest ER among all models. gain It shows @10. However, DRP shows NDCG@10 and ER gain It shows more balanced results at @10, which means that Strategy 2 can effectively determine the appropriate rank for RP items to be promoted.

[0077] Referring to Fig. 7, NDCG@10 and ER gain @10 This is a graph that changes according to (ML1M). RQ 4: This is the experimental result regarding how the recommendation results differ depending on... Figure 7 is NDCG@10 and ER obtained by modifying gain It shows @10, here =0 represents the baseline without IRP. As increases, NDCG@10 decreases, whereas ER gain Note the consistent increasing trend of @10. In the experiments for RQ 2, 3, and 5, ER was maintained appropriately while keeping NDCG@10. gain Since it increases @10, for MF and CDAE respectively Set to 0.04 and 0.02.

[0078] Referring to Fig. 8, ER gain This represents NDCG@10 according to the number of RP items m for each level of @10. These are the experimental results for RQ 5: How do recommendation results differ depending on m? Generally, it shows consistent accuracy comparable to the baseline regardless of m. In all previous experiments, m was set to 10.

[0079] In recommendation systems, the IRP problem was defined with two goals: maintaining high recommendation accuracy and increasing the recommendation of RP items to many users. Then, a new framework based on self-loss was proposed that uses a ranking learning method to train a recommendation model with the goal of achieving these two goals. Additionally, the framework was validated through extensive experiments in terms of the recommendation accuracy and exposure rate of RP items.

[0080] The device described above may be implemented as a hardware component, a software component, and / or a combination of a hardware component and a software component. For example, the device and components described in the embodiments may be implemented using one or more general-purpose or special-purpose computers, such as, for example, a processor, a controller, an arithmetic logic unit (ALU), a digital signal processor, a microcomputer, a field programmable gate array (FPGA), a programmable logic unit (PLU), a microprocessor, or any other device capable of executing and responding to instructions. The processing unit may execute an operating system (OS) and one or more software applications executed on said operating system. Additionally, the processing unit may access, store, manipulate, process, and generate data in response to the execution of the software. For ease of understanding, the processing unit may be described as being used as a single unit, but those skilled in the art will understand that the processing unit may include a plurality of processing elements and / or a plurality of types of processing elements. For example, the processing unit may include multiple processors or one processor and one controller. Additionally, other processing configurations, such as parallel processors, are also possible.

[0081] Software may include computer programs, code, instructions, or a combination of one or more of these, and may configure a processing unit to operate as desired or instruct the processing unit independently or collectively. Software and / or data may be embodied in any type of machine, component, physical device, virtual equipment, computer storage medium, or device so as to be interpreted by the processing unit or to provide instructions or data to the processing unit. Software may be distributed over networked computer systems and may be stored or executed in a distributed manner. Software and data may be stored on one or more computer-readable recording media.

[0082] 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. The computer-readable medium may include program instructions, data files, data structures, etc., either alone or in combination. The program instructions recorded on the medium may be those specifically designed and configured for the embodiment, or they may be those known and available to those skilled in the art of computer software. Examples of computer-readable recording media include magnetic media such as hard disks, floppy disks, and magnetic tapes; optical recording media such as CD-ROMs and DVDs; magneto-optical media such as floptical disks; and hardware devices specifically configured to store and execute program instructions, such as ROM, RAM, and flash memory. Examples of program instructions include machine code, such as that generated by a compiler, as well as high-level language code that can be executed by a computer using an interpreter, etc.

[0083] Although the embodiments have been described above with reference to limited examples and drawings, those skilled in the art can make various modifications and variations from the description above. For example, suitable results can be achieved even if the described techniques are performed in a different order than described, and / or the components of the described system, structure, device, circuit, etc. are combined or assembled in a form different from described, or replaced or substituted by other components or equivalents.

[0084] Therefore, other implementations, other embodiments, and equivalents to the claims also fall within the scope of the claims set forth below.

Claims

Claim 1 A recommendation method based on item ranking promotion performed by a recommendation system, comprising: a step of training a recommendation model to recommend items with additional value to a user while maintaining recommendation accuracy by utilizing items used by the user and items not used by the user; and a step of recommending items with additional value through the trained recommendation model, wherein the training step comprises: constructing a loss function to maintain recommendation accuracy and a loss function to recommend items with additional value to the user in the recommendation model; training the recommendation model in a direction that can accurately predict the user's preference using items used by the user; and training the recommendation model in a direction that utilizes items used by the user and items with additional value not used by the user so that items with additional value not used by the user are promoted to a ranking as high as that of items used by the user. Claim 2 delete Claim 3 delete Claim 4 A recommendation method comprising: the learning step in claim 1, wherein the learning step includes the step of obtaining a prediction score between 0 and 1 for items used by a user through the learning of the recommendation model, and wherein the closer the obtained prediction score is to 1, the more the loss of the loss function for maintaining the recommendation accuracy of the recommendation model is minimized. Claim 5 A recommendation method according to claim 1, characterized in that as the ranking of additionally valuable items that the user has not used is promoted, the loss of the loss function for recommending additionally valuable items of the recommendation model to the user is minimized. Claim 6 A recommendation method according to claim 1, wherein the loss function for maintaining the recommendation accuracy uses cross-entropy loss, and the loss function for recommending items with additional value to the user uses list-specific loss among rank learning-based loss functions. Claim 7 A recommendation method according to claim 1, wherein the learning step comprises the step of filtering a user from a set of candidate users for item ranking promotion when, through a separate pre-trained recommendation model, an item of additional value among the items not used by the user is predicted to be an item of lower than or equal to the user's preset rank. Claim 8 A recommendation method according to claim 1, wherein the learning step comprises a step of exploring the ranking within a target list of additionally valuable items that the user has not used through a dynamic random promotion strategy. Claim 9 In claim 8, the learning step comprises a recommendation method that includes a step of sorting items used by the user and valuable items not used by the user based on a prediction score obtained through a recommendation model. Claim 10 In claim 9, the learning step comprises a recommendation method including a step of randomly promoting valuable items that each user has not used to a rank higher than the current rank of the valuable items that the user has not used. Claim 11 In claim 10, the above learning step is a recommendation method that includes the step of placing items used by a user, which were positioned in a rank corresponding to a promoted position due to the promotion of valuable items not used by the user, into a lower rank. Claim 12 A recommendation method according to claim 1, characterized in that the item having additional value includes an exclusive item provided by the platform through an exclusive contract with a specific supplier or an item produced and provided by the platform itself. Claim 13 A computer program stored on a non-transient computer-readable recording medium that executes the item promotion-based recommendation method of any one of claims 1, 4 through 12 on the recommendation system. Claim 14 A recommendation system comprising: a model learning unit that learns a recommendation model to recommend items with additional value to a user while maintaining recommendation accuracy by utilizing items used by the user and items not used by the user; and an item recommendation unit that recommends items with additional value through the learned recommendation model, wherein the learning unit constructs a loss function to maintain recommendation accuracy and a loss function to recommend items with additional value to the user in the recommendation model, learns the recommendation model in a direction that can accurately predict the user's preferences using items used by the user, and learns the recommendation model in a direction that promotes items with additional value not used by the user to a ranking as high as that of items used by the user by utilizing items used by the user and items with additional value not used by the user.

Images