Apparatus and method of supporting e-commerce retrieval

Abstract

The present invention relates to an e-commerce search support device and method. According to one embodiment of the present invention, the device may include: a learning processing unit that learns a hyperparameter (HP) extraction unit that extracts a hyperparameter configured in a transformation function that converts learning data based on user interaction data for e-commerce into a search request for said e-commerce; an HP extraction unit that extracts a hyperparameter for evaluation from said learned hyperparameter based on a plurality of search performance indicators for a search result based on said learned hyperparameter; and an HP evaluation unit that evaluates a search result derived by configuring said extracted hyperparameter in the transformation of a search request through said transformation function of test data based on said interaction data, and selects and stores an optimized hyperparameter from said extracted hyperparameter.

Description

Technology Field

[0001] The present invention relates to an e-commerce search support device and method, and more specifically, to a technology that improves e-commerce search performance by optimizing hyperparameters (HP) for an e-commerce search service through learning via Bayesian optimization and sampling on Click Through Rate (CTR) and Click Through Conversion Rate (CTCVR), which correspond to indicators of user engagement and purchase intent. Background Technology

[0002] E-commerce search optimization is evolving across a wider range, and modern search frameworks integrate advanced quality across a diverse range of metrics that reflect user engagement and business goals.

[0003] Search results can be adjusted to meet goals using features such as sales volume and document query relevance.

[0004] Methods based on conventional technology generally focus on the Click Through Rate (CTR), which serves as a measure of participation.

[0005] However, focusing solely on CTR can lead to missing true purchase intent, and there is a significant gap between user interest and actual conversion.

[0006] Product search engines within conventional platforms play a pivotal role by acting as intermediaries that match consumer intentions with product listings.

[0007] E-commerce search engines go beyond simply interpreting and searching for related products; they serve as a key channel between digital suppliers and consumers, shape the user experience, and influence engagement and purchasing behavior across the online market.

[0008] To improve both consumer engagement and supplier profitability, e-commerce search optimization has rapidly evolved from simple keyword matching into a sophisticated multi-functional framework that integrates relevance, user engagement, and business-oriented metrics.

[0009] Conventional search systems have relied heavily on keyword matching algorithms, but this has limitations in identifying direct lexical overlap between search keywords and product lists.

[0010] However, as introduced in several recent studies, groundbreaking advancements in Large Language Models (LLMs) or neural multimodal search models have enabled systems to process dense searches by measuring the semantic similarity of query document vectors.

[0011] Most prior research focuses on the relevance scores of hybrid searches that utilize both density and relevance, but these scores are generally obtained by interpreting user interaction logs. However, they do not imply user preferences, interests, or engagement, which are important in e-commerce.

[0012] E-commerce search optimization has evolved from traditional keyword-based search methods into an advanced framework that integrates relevance, engagement, and various metrics critical to the business.

[0013] Conventional technologies have focused on maximizing the click-through rate (CTR) as an engagement metric for search optimization, but this is often insufficient for predicting actual conversions, which are the most important factor in e-commerce.

[0014] To address these issues, the Click-to-Convert Rate (CTCVR) has established itself as an essential metric for identifying purchase intent.

[0015] However, there are difficulties in training CTCVR alone due to sample selection bias and data sparsity, as the training data derived from clicks does not generalize well to a wider space of all exposures and convergence data is limited, which poses a risk of overfitting to items.

[0016] In conventional technologies such as high-density vector search and ranking through learning, existing search or ranking methods are generally optimized for a single goal, which can lead to mismatches between relevance, user engagement, and conversion.

[0017] Focusing on such a single goal limits adaptation to rapidly changing e-commerce trends and degrades performance in the online environment. Prior art literature

[0018] Korean Registered Patent No. 10-2109995, "Method and system for ranking search results and method and system for optimizing search result ranking" Korean Published Patent No. 10-2023-0174445, "Method and system for automatic optimization of search ad bidding" Korean Published Patent No. 10-2021-0060958, "Method for providing product search results of a meta shopping mall" The problem to be solved

[0019] The present invention aims to improve e-commerce search performance by optimizing hyperparameters (HP) for e-commerce search services through learning via Bayesian optimization and sampling on Click Through Rate (CTR) and Click Through Conversion Rate (CTCVR), which correspond to indicators of user engagement and purchase intent.

[0020] The present invention aims to improve the learning speed of hyperparameter selection optimization by including a meta-configuration voting method and a cumulative learning approach, and to improve the search result ranking processing performance and e-commerce search performance based on the improved learning speed.

[0021] The present invention aims to improve e-commerce search performance through balanced hyperparameter optimization that meets both user satisfaction and revenue goals. means of solving the problem

[0022] According to one embodiment of the present invention, the apparatus may include a learning processing unit that learns a hyperparameter (HP) that is configured in a transformation function that converts learning data based on user interaction data for e-commerce into a search request for said e-commerce; an HP extraction unit that extracts a hyperparameter for evaluation from said hyperparameter based on a plurality of search performance indicators for a search result based on said hyperparameter; and an HP evaluation unit that evaluates a search result derived by configuring said hyperparameter in the transformation of a search request through said transformation function of test data based on said interaction data, and selects and stores an optimized hyperparameter from said hyperparameter.

[0023] The above learning processing unit can learn the hyperparameters by applying at least one feature among a score predefined in statistics, a vocabulary matching score, an embedding-based similarity score, an embedding-based user preference score, an item quality score including item sales, views, number of comments, or delivery method, an inverted vocabulary index, and an item representative vector to the transformation function.

[0024] The above learning processing unit may apply a query keyword or query category to the vocabulary matching score in at least one feature and apply the inverted vocabulary index.

[0025] The above HP evaluation unit can evaluate positive or negative hyperparameters by learning at least one of the click-through rate (CTR), the click-through cart-adding rate (CTCAR), and the post-click cart adding rate (CAR), for the click-through conversion rate (CTCVR) that considers the number of purchases for the number of search impressions.

[0026] The above HP evaluation unit can filter the evaluated positive or negative hyperparameters with respect to the number of search impressions, the number of clicks, and the number of cart additions.

[0027] The above HP extraction unit can apply the evaluated positive hyperparameter or negative hyperparameter to any one of the plurality of search performance indicators for a search result based on the learned hyperparameter, and extract a hyperparameter in which the search performance exceeds the reference performance as the evaluated positive hyperparameter or negative hyperparameter is exceeded.

[0028] The above HP extraction unit can perform sampling of the extracted hyperparameters for evaluation or cumulative learning.

[0029] The above HP evaluation unit can extract the test data from the interaction data and evaluate the extracted hyperparameters to evaluate whether there is overlap between the training data and the test data.

[0030] The above HP evaluation unit can select the optimized hyperparameters based on a bagging method including a voting method for the extracted hyperparameters.

[0031] According to one embodiment of the present invention, an e-commerce search support method may include: a step of learning an HP extraction unit in a learning processing unit, which extracts a hyperparameter (HP) configured in a transformation function that transforms learning data based on user interaction data for e-commerce into a search request for said e-commerce; a step in the HP extraction unit, which extracts a hyperparameter for evaluation from said hyperparameter based on a plurality of search performance indicators for a search result based on said hyperparameter; and a step in the HP evaluation unit, which evaluates a search result derived by configuring said hyperparameter in the transformation of a search request through said transformation function of test data based on said interaction data, and selects and stores an optimized hyperparameter from said hyperparameter.

[0032] The step of training an HP extraction unit that extracts a hyperparameter (HP) configured in a transformation function that converts training data based on user interaction data for the above e-commerce into a search request for the above e-commerce may include the step of training the hyperparameter by applying at least one feature among a score predefined in statistics, a vocabulary matching score, an embedding-based similarity score, an embedding-based user preference score, an item quality score including item sales, views, comment count, or delivery method, an inverted vocabulary index, and an item representative vector to the transformation function.

[0033] The step of training an HP extraction unit that extracts the hyperparameters by applying the above at least one feature to the above transformation function may include the step of applying a query keyword or query category to the vocabulary matching score in the above at least one feature and applying the above inverted vocabulary index.

[0034] The step of evaluating search results derived by configuring the extracted hyperparameters in the transformation of a search request through the transformation function of test data based on the above interaction data, and selecting and saving optimized hyperparameters from the extracted hyperparameters, may include the step of evaluating positive hyperparameters or negative hyperparameters by learning at least one of the click-through rate (CTR) considering the number of clicks for the number of search impressions, the click-through cart-adding rate (CTCAR) considering the number of cart additions for the number of search impressions, and the post-click cart adding rate (CAR) considering the number of cart additions for the number of clicks, for the click-through conversion rate (CTCVR) considering the number of purchases for the number of search impressions, and the step of filtering the evaluated positive hyperparameters or negative hyperparameters for the number of search impressions, the number of clicks, and the number of cart additions.

[0035] The step of extracting hyperparameters for evaluation or cumulative learning from the learned hyperparameters based on any one of the plurality of search performance indicators for the search results based on the learned hyperparameters may include the step of extracting the search performance indicator for the search results based on the learned hyperparameters and the evaluated positive hyperparameter or negative hyperparameter as the next hyperparameter according to a sampling method that is pre-learned or is being learned, and the step of performing sampling of the extracted hyperparameters for the evaluation or cumulative learning.

[0036] The step of evaluating the search results derived by configuring the extracted hyperparameters in the transformation of the search request through the transformation function of the test data based on the above interaction data, and selecting and storing the optimized hyperparameters from the extracted hyperparameters, may include the step of selecting the optimized hyperparameters based on a bagging method including a voting method for the extracted hyperparameters. Effects of the invention

[0037] The present invention can improve e-commerce search performance by optimizing hyperparameters (HP) for e-commerce search services through learning via Bayesian optimization and sampling on Click Through Rate (CTR) and Click Through Conversion Rate (CTCVR), which correspond to indicators of user engagement and purchase intention.

[0038] The present invention can improve the learning speed of hyperparameter selection optimization by including a meta-configuration voting method and a cumulative learning approach, and improve search result ranking processing performance and e-commerce search performance based on the improved learning speed.

[0039] The present invention can improve e-commerce search performance through balanced hyperparameter optimization that meets both user satisfaction and profit goals. Brief explanation of the drawing

[0040] FIG. 1 is a drawing illustrating an e-commerce search support device according to an embodiment of the present invention. FIG. 2 is a diagram illustrating a hyperparameter optimization structure based on an e-commerce search support device according to an embodiment of the present invention. FIG. 3 is a diagram illustrating the configuration of applying a conversion function according to a change in hyperparameters of an e-commerce search support device according to an embodiment of the present invention. FIG. 4 is a diagram illustrating a hyperparameter optimization algorithm of an e-commerce search support device according to an embodiment of the present invention. FIG. 5 is a diagram illustrating the analysis of search results in an e-commerce search support device according to an embodiment of the present invention. FIG. 6 is a diagram illustrating an e-commerce search support method according to an embodiment of the present invention. Specific details for implementing the invention

[0041] Specific structural or functional descriptions of embodiments according to the concept of the present invention disclosed herein are provided merely for the purpose of explaining embodiments according to the concept of the present invention, and embodiments according to the concept of the present invention may be implemented in various forms and are not limited to the embodiments described herein.

[0042] Embodiments according to the concept of the present invention may be subject to various modifications and may take various forms; therefore, embodiments are illustrated in the drawings and described in detail in this specification. However, this is not intended to limit the embodiments according to the concept of the present invention to specific disclosed forms, and includes modifications, equivalents, or substitutions that fall within the spirit and scope of the present invention.

[0043] Terms such as "first" or "second" may be used to describe various components, but said components should not be limited by said terms. For the sole purpose of distinguishing one component from another, for example, without departing from the scope of rights according to the concept of the present invention, the first component may be named the second component, and similarly, the second component may be named the first component.

[0044] When it is stated that one component is "connected" or "connected" to another component, it should be understood that while it may be directly connected or connected to that other component, there may also be other components in between. Conversely, when it is stated that one component is "directly connected" or "directly connected" to another component, it should be understood that there are no other components in between. Expressions describing the relationships between components, such as "between," "exactly between," or "directly adjacent to," should be interpreted in the same way.

[0045] The terms used herein are merely for describing specific embodiments and are not intended to limit the invention. Singular expressions include plural expressions unless the context clearly indicates otherwise. In this specification, terms such as “comprising” or “having” are intended to specify the existence of the described features, numbers, stages, actions, components, parts, or combinations thereof, and should be understood as not precluding the existence or addition of one or more other features, numbers, stages, actions, components, parts, or combinations thereof.

[0046] Unless otherwise defined, all terms used herein, including technical or scientific terms, have the same meaning as generally understood by those skilled in the art to which the present invention pertains. Terms such as those defined in commonly used dictionaries should be interpreted as having a meaning consistent with their meaning in the context of the relevant technology, and should not be interpreted in an ideal or overly formal sense unless explicitly defined in this specification.

[0047] Hereinafter, embodiments will be described in detail with reference to the attached drawings. However, the scope of the patent application is not limited or restricted by these embodiments. Identical reference numerals in each drawing indicate identical components.

[0049] FIG. 1 is a drawing illustrating an e-commerce search support device according to an embodiment of the present invention.

[0050] FIG. 1 illustrates the components of an e-commerce search support device according to an embodiment of the present invention.

[0051] Referring to FIG. 1, an e-commerce search support device (100) according to one embodiment of the present invention includes a learning processing unit (110), an HP extraction unit (120), and an HP evaluation unit (130), and further includes a cumulative learning processing unit.

[0052] An e-commerce search support device (100) according to one embodiment of the present invention implements a technology that supports e-commerce search by optimizing a hyperparameter (HP) used together with a transformation function used for e-commerce search.

[0053] According to one embodiment of the present invention, the learning processing unit (110) can learn an HP extraction unit that extracts hyperparameters configured in a conversion function that converts learning data based on user interaction data for e-commerce into a search request for e-commerce.

[0054] The learning processing unit (110) can learn an HP extraction unit that extracts hyperparameters by applying at least one feature among a score predefined in statistics, a vocabulary matching score, an embedding-based similarity score, an embedding-based user preference score, an item quality score including item sales, views, comments, cart, cards contained in the item, number of shares, CTR, CTCVR score, etc., an inverted vocabulary index, and an item representative vector to a transformation function.

[0055] For example, in relation to the lexical match score, bm 25 algorithms such as lexical match may be included.

[0056] The learning processing unit (110) applies a query keyword or query category to the vocabulary matching score, applies an inverted vocabulary index, and applies a query representative vector to the embedding-based similarity score. For example, if only vector search is applied, the part can operate even if it is not included.

[0057] The learning processing unit (110) applies an item representative vector, and the query representative vector and user representative vector are applied to the embedding-based user preference score, and the item representative vector can be applied.

[0058] HP evaluation unit (130) can evaluate positive hyperparameters or negative hyperparameters by learning at least one of the click-through rate (CTR), the click-through cart-adding rate (CTCAR), and the post-click cart adding rate (CAR) for the click-through conversion rate (CTCVR) that considers the number of purchases for the number of search impressions.

[0059] CTCVR can correspond to the number of clicks that lead to purchase conversions in search, and CTR can represent the frequency with which an item is displayed in search.

[0060] CTCAR can represent the frequency of adding to the cart from a search, and CAR can represent the frequency of adding to the cart after a click.

[0061] CTCVR is the primary goal for learning, and CTR, CTCAR, and CAR can be used as secondary goals.

[0062] HP evaluation unit (130) can store positive or negative hyperparameters based on Bayesian smoothing of search exposure count, click count and cart addition count.

[0063] The stored values ​​corresponding to positive or negative hyperparameters can be used for better extraction of hyperparameters in the next HP extraction section, subsequent meta-evaluation processes, or cumulative learning.

[0064] An HP extraction unit (120) according to one embodiment of the present invention can extract hyperparameters for evaluation or cumulative learning from learned hyperparameters based on a search performance indicator for a search result based on learned hyperparameters.

[0065] The HP extraction unit (120) can extract the search performance indicator for the search result based on the learned hyperparameters and the evaluated positive hyperparameters or negative hyperparameters as the next hyperparameters according to the sampling method that is pre-learned or is being learned.

[0066] The HP extraction unit (120) can perform sampling for evaluation or cumulative learning of the extracted hyperparameters.

[0067] According to one embodiment of the present invention, the HP evaluation unit (130) can evaluate the search results derived by configuring hyperparameters extracted from the transformation of a search request through a transformation function of test data based on interaction data, and select or save optimized hyperparameters from the extracted hyperparameters.

[0068] The HP evaluation unit (130) can evaluate whether there is overlap between the training data and the test data by extracting test data from the interaction data and evaluating the extracted hyperparameters.

[0069] The HP evaluation unit (130) can select optimized hyperparameters based on a bagging method including a voting method for the extracted hyperparameters.

[0070] According to one embodiment of the present invention, the cumulative learning processing unit can cumulatively learn the learning results for selected and stored hyperparameters and remove duplicate hyperparameters from the selected and stored hyperparameters.

[0071] The cumulative learning processing unit initializes a new learning cycle through the dataset in cumulative learning instead of initializing the observation dataset, and through this setting, allows the sampler to focus on unexplored areas to minimize duplicate samples and facilitate exploration within the palace space.

[0072] The cumulative learning processing unit reduces computational redundancy by assigning priorities and dynamically adapts to changes in user behavior and market trends.

[0073] The cumulative learning processing unit establishes an adaptive feedback mechanism to enhance the robustness of hyperparameter optimization in an evolving e-commerce environment.

[0074] Accordingly, the present invention can improve e-commerce search performance by optimizing hyperparameters (HP) for e-commerce search services through learning via Bayesian optimization and sampling on Click Through Rate (CTR) and Click Through Conversion Rate (CTCVR), which correspond to indicators of user engagement and purchase intention.

[0076] FIG. 2 is a diagram illustrating a hyperparameter optimization structure based on an e-commerce search support device according to an embodiment of the present invention.

[0077] Referring to FIG. 2, a hyperparameter optimization structure (200) based on an e-commerce search support device according to an embodiment of the present invention is composed of a learning processing unit (210), an HP extraction unit (220), and an HP evaluation unit (230).

[0078] The learning processing unit (210) applies the learning data based on user interaction log data to a transformation function to extract search items as search results from the search engine through transformation with sample hyperparameters, and samples hyperparameters through evaluation.

[0079] The HP extraction unit (220) extracts hyperparameters with superior performance from the sampled hyperparameters and samples the extracted hyperparameters to prepare for an evaluation of the learning results.

[0080] After the hyperparameter optimization learning is completed, the HP extraction unit (220) selects a set of various candidate hyperparameter configurations based on the performance of the hyperparameters.

[0081] The HP extraction unit (220) can estimate online performance and extract hyperparameters to store additional meta-level evaluations for use in subsequent cumulative learning.

[0082] The HP evaluation unit (230) applies test data based on hyperparameters extracted from the HP extraction unit (220) and user interaction log data to a transformation function to derive search results, and evaluates the derived results through voting.

[0083] If hyperparameter configurations are selected simply based on training metrics, hyperparameter optimization is often biased toward the training data, which may lead to overfitting as it may not generalize well to production due to sample selection bias and data sparsity issues.

[0084] To address this, the HP evaluation unit (230) uses time-based and query-based data partitioning to evaluate the best-performing hyperparameter configuration for unseen queries and future periods to ensure there is no overlap with the training data.

[0085] The HP evaluation unit (230) evaluates the robustness of each hyperparameter across multiple metrics and uses a meta-configuration voting strategy to prioritize hyperparameters that show consistent and balanced performance across goals.

[0086] The approach of the HP evaluation department (230) can select parameters that are effectively generalized by combining the strengths of various configurations.

[0087] The hyperparameter optimization structure (200) can derive an optimal hyperparameter configuration that is balanced.

[0088] The hyperparameter optimization structure (200) focuses on improving relevance, ranking quality, and overall device search support performance.

[0089] The hyperparameter optimization structure (200) identifies a set of hyperparameters (h ∈ H) that simultaneously maximizes the performance of a search engine across multiple goals O = {O1, O2, . . . ,OM}.

[0090] Given an observational data set D0 = and a set of reference truth labels y = {y1, ( · ) ,y2, ( · ) , . . . ,ym, ( · )} for each objective, optimization iteratively approximates the optimal performance by sampling and evaluating hyperparameter configurations.

[0091] In other words, the search system can be evaluated by selecting a hyperparameter configuration ht in the search space H at each attempt t = 1, 2, . . . , N.

[0092] During optimization, various sampler methods such as random search, grid search, or Bayesian optimization (e.g., tree structure Parzen estimator (TPE) or Gaussian process (GP)) can be used to sample the hyperparameter configuration.

[0093] Accordingly, the present invention can improve the learning speed of optimization of hyperparameter selection by including a meta-configuration voting method and a cumulative learning approach, and improve the performance of search result ranking processing tasks and e-commerce search performance based on the improved learning speed.

[0095] FIG. 3 is a diagram illustrating the configuration of applying a conversion function according to a change in hyperparameters of an e-commerce search support device according to an embodiment of the present invention.

[0096] FIG. 3 illustrates a configuration for applying a transformation function according to a change in hyperparameters of an e-commerce search support device according to an embodiment of the present invention.

[0097] Referring to FIG. 3, a structure (300) is shown in which a query request to a search engine is configured in the conversion function of an e-commerce search support device according to an embodiment of the present invention, so that the conversion function can be adjusted according to a change in hyperparameters.

[0098] In the structure (300), ht is sampled from the hyperparameter configuration in each attempt and the search system evaluates it in the query set.

[0099] Configuration ht demonstrates its influence on the search process through transformation functions that adjust the search engine's behavior based on sampled hyperparameters.

[0100] The structure (300) may include changes to the ranking strategy, modifications to relevance weights, or adjustments to search heuristics.

[0101] The structure (300) applies the query (301) to the feature (302), applies the feature (303) to the feature (302), or applies the transformation function (304).

[0102] The transformation function (304) applies the feature (302) and feature (303) and hyperparameter (305) to derive a search result (307) through the search engine (306).

[0103] Specifically, the transformation function can be expressed as T(ht, q), which represents the search process transformed according to the hyperparameter configuration ht applied to the query q.

[0104] Depending on the sampled hyperparameters, this transformation can modify how items are ranked or filtered in search engines.

[0105] For each query q, the system generates a prediction (item rank), and the relevance / quality metrics of these items are evaluated using a goal, respectively.

[0106] The metric for the goal is the predicted ranking It can be calculated based on and compared with the reference truth label ym,q,(·), and It can be formalized as.

[0107] Then, performance evaluations for all queries are aggregated to generate performance metrics for each goal.

[0108] Aggregated metrics are used to update the optimization process, guiding it to find better hyperparameter configurations.

[0109] This process continues iteratively until convergence is reached, at which point the goal selects a hyperparameter configuration that optimizes the trade-offs between various objectives.

[0110] The structure (300) may be a new hyperparameter optimization framework specifically tailored for e-commerce search services focused on goals such as CTR and CTCVR, which are essential indicators of user engagement and purchase intent.

[0111] The structure (300) implements multi-purpose hyperparameter optimization in e-commerce search and ranking tasks.

[0112] FIG. 4 is a diagram illustrating a hyperparameter optimization algorithm of an e-commerce search support device according to an embodiment of the present invention.

[0113] FIG. 4 illustrates a hyperparameter optimization algorithm of an e-commerce search support device according to an embodiment of the present invention.

[0114] Referring to FIG. 4, the algorithm (400) represents a full summary algorithm using a TPE sampler.

[0115] According to the algorithm (400), first, the sampler is initialized, and then, an object corresponding to the ground truth is defined.

[0116] Next, if cumulative learning is in progress, perform cumulative learning; otherwise, initialize and sample the hyperparameters.

[0117] Next, the algorithm (400) uses a transform function to search for items and evaluates the search results (objects).

[0118] Next, the algorithm (400) aggregates the evaluation results, stores the results, updates the state based on the evaluation results, and updates the hyperparameters.

[0119] Next, the algorithm (400) calculates the expected improvement (EI) for the candidate hyperparameters, selects the hyperparameters, updates the dataset, and terminates.

[0121] FIG. 5 is a diagram illustrating the analysis of search results in an e-commerce search support device according to an embodiment of the present invention.

[0122] Figure 5 illustrates an e-commerce conversion funnel in item exposure based on e-commerce search result analysis.

[0123] Referring to FIG. 5, the e-commerce conversion funnel (500) includes a total potential search target space (501) in production, an inference space through search results (502), a click or wish space (503), a cart space (504), and a conversion space (505).

[0124] The data corresponding to the wish space (503) and the cart space (504) can be determined as an auxiliary training space of the transformation space (505) corresponding to the training space.

[0125] To enhance user engagement and sales in e-commerce while addressing sample selection bias and data sparsity, it is essential to mine user interaction logs, including clicks and conversions per impression.

[0126] Based on the e-commerce conversion funnel (500), the click-through conversion rate (CTCVR) that considers the number of purchases for the number of search impressions, the click-through rate (CTR) that considers the number of clicks for the number of search impressions, the click-through cart-adding rate (CTCAR) that considers the number of cart additions for the number of search impressions, and the post-click cart adding rate (CAR) that considers the number of cart additions for the number of clicks can be used.

[0127] CTCVR is the primary goal, and CTR, CTCAR, and CAR can be used as secondary goals to mitigate selection bias and scarcity.

[0128] The labels for these items are derived from goal-specific statistics and can be used as thresholds for positive / negative scores for the corresponding goal. Here, the threshold can be an evaluated result value of a past learned HP or a pre-set random variable.

[0129] Additionally, while naive statistical ratios can be used, these scores can be further refined by filtering based on the number of exposures and Bayesian smoothing for rare events.

[0130] Based on the metric calculations zt, m, and q, metrics are aggregated across the entire query, and hyperparameter optimization performs multi-goal tuning to improve engagement, conversion efficiency, and purchase intent, and can improve search performance in e-commerce.

[0132] FIG. 6 is a diagram illustrating an e-commerce search support method according to an embodiment of the present invention.

[0133] FIG. 6 illustrates a hyperparameter optimization procedure for improving e-commerce search performance in an e-commerce search support method according to an embodiment of the present invention.

[0134] Referring to FIG. 6, in step (S601), an e-commerce search support method according to an embodiment of the present invention learns an HP extraction unit that extracts hyperparameters for converting training data into a search request.

[0135] That is, an e-commerce search support method according to one embodiment of the present invention can learn an HP extraction unit that extracts a hyperparameter (HP) configured in a conversion function that converts training data based on user interaction data for e-commerce into an e-commerce search request.

[0136] In step (S602), the e-commerce search support method according to an embodiment of the present invention extracts hyperparameters for evaluation from the learned hyperparameters.

[0137] That is, an e-commerce search support method according to one embodiment of the present invention can extract hyperparameters for evaluation from learned hyperparameters based on a plurality of search performance indicators for search results based on learned hyperparameters.

[0138] In step (S603), the e-commerce search support method according to an embodiment of the present invention evaluates the extracted hyperparameters and selects and stores the optimized hyperparameters.

[0139] That is, an e-commerce search support method according to one embodiment of the present invention can evaluate search results derived by configuring hyperparameters extracted from the transformation of a search request through a transformation function of test data based on interaction data, and select and store optimized hyperparameters from the extracted hyperparameters.

[0140] Therefore, the present invention can improve e-commerce search performance through balanced hyperparameter optimization that meets both user satisfaction and profit goals.

[0141] The present invention relates to an e-commerce search support device and method, which is a new multi-purpose hyperparameter optimization (HPO) framework adapted to the requirements of e-commerce search.

[0142] In particular, the present invention can optimize important performance indicators such as CTR, CTCAR, and CTCVR, which are important for improving user engagement and conversion in e-commerce systems, by optimizing hyperparameters in accordance with search support.

[0143] The present invention can effectively solve the unique complexity associated with simultaneously optimizing various competitive indicators in e-commerce.

[0144] The present invention demonstrates, through extensive evaluation, that an e-commerce search support device can achieve significant improvements in CTR, CTCVR, and related objectives by demonstrating the ability to achieve outstanding stability, convergence, and robustness in multi-objective optimization using sampling strategies of various optimization hyperparameters.

[0145] The present invention can implement multi-objective optimization for e-commerce by further enhancing the performance of hyperparameter optimization and providing measurable gains in relevance and user engagement.

[0146] The present invention can reduce sampling costs by developing advanced techniques such as meta-configuration voting and cumulative training, thereby increasing the adaptability of hyperparameter optimization and providing inspection changes between each cumulative step.

[0148] 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 array (FPA), 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 multiple processing elements and / or multiple types of processing elements. For example, the processing unit may include multiple processors or one processor and one controller. In addition, other processing configurations, such as parallel processors, are also possible.

[0149] 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 command the processing unit independently or collectively. Software and / or data may be permanently or temporarily embodied in any type of machine, component, physical device, virtual equipment, computer storage medium or device, or transmitted signal wave 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.

[0150] Although the embodiments have been described above with reference to the limited 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.

[0151] Therefore, other implementations, other embodiments, and equivalents to the claims also fall within the scope of the claims set forth below. Explanation of the symbols

[0152] 100: E-commerce search support device 110: Learning processing unit 120: HP Extraction Unit 130: HP Evaluation Unit

Claims

Claim 1 An e-commerce search support device characterized by comprising: a learning processing unit that learns a hyperparameter (HP) extraction unit that extracts a hyperparameter configured in a transformation function that converts learning data based on user interaction data for e-commerce into a search request for said e-commerce; an HP extraction unit that extracts a hyperparameter for evaluation from said learned hyperparameter based on a plurality of search performance indicators for a search result based on said learned hyperparameter; and an HP evaluation unit that evaluates a search result derived by configuring said extracted hyperparameter in the transformation of a search request through said transformation function of test data based on said interaction data, and selects and stores an optimized hyperparameter from said extracted hyperparameter. Claim 2 An e-commerce search support device according to claim 1, wherein the learning processing unit learns the hyperparameters by applying at least one feature among a score predefined in statistics, a vocabulary matching score, an embedding-based similarity score, an embedding-based user preference score, an item quality score including item sales, views, comments, or delivery method, an inverted vocabulary index, and an item representative vector to the transformation function. Claim 3 An e-commerce search support device according to paragraph 2, wherein the learning processing unit applies a query keyword or query category to the vocabulary matching score in at least one feature and applies the inverted vocabulary index. Claim 4 An e-commerce search support device according to claim 1, wherein the HP evaluation unit evaluates positive hyperparameters or negative hyperparameters by learning at least one of a click-through rate (CTR) considering the number of clicks for search exposures, a click-through cart-adding rate (CTCAR) considering the number of cart additions for search exposures, and a post-click cart adding rate (CAR) considering the number of cart additions for clicks, with respect to a click-through conversion rate (CTCVR) considering the number of purchases for the number of search exposures. Claim 5 An e-commerce search support device according to claim 4, wherein the HP evaluation unit stores the evaluated positive hyperparameters or negative hyperparameters with respect to the number of search impressions, the number of clicks, and the number of cart additions. Claim 6 An e-commerce search support device according to claim 4, wherein the HP extraction unit extracts one of the plurality of search performance indicators for a search result based on the learned hyperparameters and the evaluated positive hyperparameter or negative hyperparameter as the next hyperparameter according to a sampling method that is pre-learned or is being learned. Claim 7 An e-commerce search support device according to claim 6, wherein the HP extraction unit performs sampling of the extracted hyperparameters for evaluation or cumulative learning. Claim 8 An e-commerce search support device according to claim 1, wherein the HP evaluation unit selects the optimized hyperparameters based on a bagging method including a voting method for the extracted hyperparameters. Claim 9 A method for supporting e-commerce search, characterized by comprising: a step of learning an HP extraction unit in a learning processing unit for extracting hyperparameters (HP) configured in a transformation function that transforms learning data based on user interaction data for e-commerce into a search request for said e-commerce; a step in which, in the HP extraction unit, a hyperparameter for evaluation is extracted from said learned hyperparameters based on a plurality of search performance indicators for a search result based on said learned hyperparameters; and a step in which, in an HP evaluation unit, a search result derived by configuring said extracted hyperparameters in the transformation of a search request through said transformation function of test data based on said interaction data is selected and stored from said extracted hyperparameters, and an optimized hyperparameter is selected and stored. Claim 10 In claim 9, the step of learning an HP extraction unit that extracts a hyperparameter (HP) configured in a transformation function that converts learning data based on user interaction data for the e-commerce into a search request for the e-commerce comprises the step of learning the hyperparameter by applying at least one feature among a score predefined in statistics, a vocabulary match score, an embedding-based similarity score, an embedding-based user preference score, an item quality score including item sales, views, comment count, or delivery method, an inverted vocabulary index, and an item representative vector to the transformation function. Claim 11 An e-commerce search support method according to claim 10, wherein the step of learning an HP extraction unit that extracts the hyperparameters by applying the at least one feature to the transformation function comprises the step of applying a query keyword or query category to the vocabulary matching score in the at least one feature and applying the inverted vocabulary index. Claim 12 In claim 9, the step of evaluating search results derived by configuring the extracted hyperparameters in the transformation of a search request through the transformation function of test data based on the interaction data, and selecting and storing optimized hyperparameters from the extracted hyperparameters, comprises: a step of evaluating positive hyperparameters or negative hyperparameters by learning at least one of a click-through rate (CTR) considering the number of clicks for the number of search impressions, a click-through cart-adding rate (CTCAR) considering the number of cart additions for the number of search impressions, and a post-click cart adding rate (CAR) considering the number of cart additions for the number of clicks, with respect to a click-through conversion rate (CTCVR) considering the number of purchases for the number of search impressions; and a step of filtering the evaluated positive hyperparameters or negative hyperparameters with respect to the number of search impressions, the number of clicks, and the number of cart additions. Claim 13 In claim 12, the step of extracting a hyperparameter for evaluation or cumulative learning from the learned hyperparameter based on any one of the plurality of search performance indicators for the search result based on the learned hyperparameter comprises: a step of extracting the next hyperparameter according to a sampling method that is pre-learned or is learning, the search performance indicator of any one of the plurality of search performance indicators for the search result based on the learned hyperparameter and the evaluated positive hyperparameter or negative hyperparameter; and a step of performing sampling of the extracted hyperparameter for evaluation or cumulative learning. Claim 14 An e-commerce search support method according to claim 9, wherein the step of evaluating the search results derived by configuring the extracted hyperparameters in the transformation of a search request through the transformation function of the test data based on the interaction data, and selecting and storing the optimized hyperparameters from the extracted hyperparameters, comprises the step of selecting the optimized hyperparameters based on a bagging method including a voting method for the extracted hyperparameters.

Images