System and method for searching users based on behavior vectors

Abstract

A behavior vector-based user search system and a method thereof are disclosed. A system according to one embodiment of the present invention includes: a tokenization unit that embeds the value, type, and observation period of a behavior tag, respectively, and sums them to convert them into cell-unit behavior tokens; a behavior encoder that contextualizes the behavior tokens by an attention operation that learns within-user and between-user relationships and compresses them into a single user-unit behavior vector; a query encoder that converts a natural language request into a structured query and a query vector; a spatial alignment unit that aligns the query vector and the behavior vector in the same embedding space by contrastive learning of positive and negative learning pairs constructed from a campaign execution log; a vector search unit that searches for candidate users based on similarity in the aligned space; and a factual verification unit that looks up the behavior tags of candidate users and selects candidates that meet the request conditions.

Description

Technology Field

[0001] The present invention relates to a behavior vector-based user search system and method, and more specifically, to a behavior vector-based user search system and method that embeds the values, types, and observation periods of behavior tags collected for each user, respectively, sums them, converts them into cell-unit behavior tokens, compresses the behavior tokens into a single user-unit behavior vector by contextualizing them through an attention operation that learns the relationships between behavior tokens within the same user and the behavior pattern relationships between different users, converts a natural language request for a target user into a structured query and a query vector, aligns the query vector and the behavior vector in the same embedding space through contrastive learning of positive / negative learning pairs automatically configured from the execution logs of an advertising campaign, and searches for and selects users that match the natural language request through similarity search and factual verification by looking up behavior tags in the aligned space. Background Technology

[0002] Data Management Platforms (DMPs), which collect and analyze user behavior data in online and mobile environments for use in advertising and marketing, are widely used. Since the performance of advertising campaigns depends on how accurately and quickly target user sets aligned with campaign objectives are extracted, the importance of technology that searches for users meeting specific criteria based on marketer requests is growing.

[0003] Conventional target user extraction typically relies on rule-based segment definition. For instance, this method involves humans setting conditional rules, such as "users who have used automotive applications within the last 30 days," and data engineers converting these into database queries to extract a set of users that meet those criteria. This approach consumes time and cost because manual interpretation is involved in converting a marketer's request into a conditional query; furthermore, it can only extract users who conform to predefined rules, making flexible searching based on similar behavioral patterns impossible.

[0004] Meanwhile, as language models that learn the structure and context of language from large-scale texts have advanced, there have been attempts to utilize these models for the interpretation of user data. However, since user behavioral data differs from language and representational spaces, using publicly available language models pre-trained on text directly fails to accurately represent behavioral patterns. Representing behavioral data requires an approach that involves directly learning behavior-specific representations from scratch.

[0005] Furthermore, analysis based on individual tags or single events cannot yield representations that comprehensively reflect the user's intent, interests, purchasing stage, and conversion potential. For example, individual behaviors such as using a car application, visiting a dealership, and searching for lease products are each weak signals; however, when they appear together, they form a strong meaning indicating that the user is highly likely to switch to a car lease. To capture these complex signals, a representation system that learns the relationships between behaviors is required. While techniques for embedding user behavior sequences using recurrent neural networks are known, a system for tokenizing heterogeneous behavioral data—which differs in value scale, tag types, and observation periods—into a unified structure, and a means to mitigate learning bias caused by the skewed distribution inherent in numerical data, have not yet been presented.

[0006] Furthermore, since marketers' natural language requests and users' behavior vectors reside in different embedding spaces, direct comparison and searching are impossible. Although techniques for mapping natural language queries and search targets to the same embedding space are known, learning the alignment between the two spaces requires a large volume of labels regarding the relevance between the query and the user; if these labels are generated manually by humans, labeling costs are incurred and hypothesis bias is introduced.

[0007] Furthermore, large models with high representation quality involve significant computational and inference costs, making it difficult to rapidly generate and update behavior vectors for tens of millions of users. While knowledge distillation techniques—where a small model mimics the knowledge learned by a large model, or a structure where a student model learns from the output of a so-called teacher model—are known, conventional knowledge distillation primarily focuses on label distillation, which mimics predictive labels. Consequently, the structure of the vector space—specifically the distance relationships between similar users—is not preserved, and criteria for verifying before deployment whether a lightweight model produces the same search results as the original model have not been presented.

[0008] Finally, searches based solely on vector similarity may include candidates whose behavior vectors are similar but who do not actually satisfy the requested behavior conditions. For example, users whose behavior patterns are similar to those of users interested in travel but whose actual travel-related behavior is not observed may be mixed into the search results; if such candidates are included in the target, campaign performance deteriorates. Furthermore, there is a need for a means to reliably generate large-scale targets in real-world environments where only a small number of seed users are available.

[0009] Therefore, a technology is required that directly learns behavior-specific representations to compress users into a single behavior vector, aligns natural language requests with the behavior vector space to enable searching without manual labeling, verifies the matching of retrieved candidates to actual behavior conditions, and enables periodic updates of behavior vectors for a large number of users without loss of representation quality. Prior art literature

[0010] Korean Patent Publication No. 10-0832729 (Published May 27, 2008) Korean Patent Publication No. 10-1161078 (Published July 13, 2012) The problem to be solved

[0011] The objective of the present invention, which is to solve the aforementioned problems, is to provide a behavior vector-based user search system that embeds the value, type, and observation period of each of a plurality of behavior tags collected for each user, sums them up, and converts them into cell-unit behavior tokens; compresses the behavior tokens into a single user-unit behavior vector by contextualizing them through an attention operation that learns the relationships between behavior tokens within the same user and the behavior pattern relationships between different users; converts natural language requests for target users into structured queries and query vectors; aligns the query vectors and behavior vectors in the same embedding space through contrastive learning of positive and negative learning pairs automatically constructed from responding and non-responding users in the execution logs of advertising campaigns; searches for candidate users based on similarity in the aligned space; and selects candidate users that meet the request conditions by querying the behavior tags of the searched candidate users.

[0012] Another objective of the present invention is to provide a behavior vector-based user search method performed in the behavior vector-based user search system described above.

[0013] Another objective of the present invention is to provide a computer-readable recording medium having a program recorded thereon for executing the above-described behavior vector-based user search method on a computer. means of solving the problem

[0014] To achieve the above-mentioned objective, a behavior vector-based user search system according to an embodiment of the present invention comprises: a tokenization unit that, for each of a plurality of behavior tags collected for each user, embeds the value of the behavior tag, the type of the behavior tag, and the observation period, respectively, and sums them to convert them into cell-unit behavior tokens; a behavior encoder that contextualizes the behavior tokens by an attention operation that learns the relationships between behavior tokens within the same user and the behavior pattern relationships between different users together, compresses the contexted behavior tokens into a single behavior vector at the user level, and is learned by a learning task that predicts the actual behavior of the user observed after the first period from the behavior tags up to the first period; a query encoder that converts a natural language request for a target user into a structured query including intentions and conditions, and encodes the converted structured query into a query vector; a spatial alignment unit that aligns the query vector and the behavior vector so as to be compared in the same embedding space by contrastive learning that configures users who responded and users who did not respond in the execution log of an advertising campaign into positive learning pairs and negative learning pairs, respectively; and the query vector and in the aligned embedding space It includes a vector search unit that searches for candidate users based on similarity, and a fact verification unit that checks whether the behavior tags of the searched candidate users meet the conditions of the natural language request and selects the matching candidate users.

[0015] The above tokenization unit normalizes the values ​​of numeric behavior tags into a uniform distribution by quantile transformation and then embeds them to mitigate attention bias caused by the skewness of the value distribution.

[0016] The above behavior encoder outputs the behavior vector after L2 normalization, and the vector search unit searches for the candidate user based on cosine similarity.

[0017] The vector search unit searches for the candidate user based on similarity with the reference vector, using the average of the behavior vectors of the seed user set as the reference vector.

[0018] The above vector search unit returns top candidate users within a preset response time by using an approximate nearest neighbor index loaded with the behavior vectors of all users.

[0019] The above query encoder is trained by applying supervised fine-tuning based on low-order adaptation (LoRA) to a pre-trained language model, and the above structured query includes at least one of an intent, category, behavior condition, exclusion condition, target size, and duration condition.

[0020] The contrast learning of the spatial matching unit above replaces a separate manual label generation process based on a contrast loss that increases the similarity of the positive learning pairs and decreases the similarity of the negative learning pairs for a single query vector.

[0021] The above system further includes a performance verification unit that verifies the prediction performance of the behavior encoder by sequential evaluation in which the period used for learning and the period used for evaluation are temporally separated.

[0022] The above system further includes a distillation learning unit that, using the behavior encoder as a reference model, trains a serving model of a lightweight structure in which at least one of the number of attention layers, hidden dimension, and number of parameters is reduced compared to the reference model by representation distillation.

[0023] The above expression distillation uses the behavior vector generated by the reference model for the same behavior tag input as the target vector, and trains the serving model to reduce the cosine similarity and distance difference between the behavior vector output by the serving model and the target vector.

[0024] The distillation learning unit verifies whether the search ranking of the serving model is preserved based on the top K overlap and similarity distributions between the similar user search results by the reference model and the similar user search results by the serving model, and deploys the serving model that passes the verification to the operating environment.

[0025] The distributed serving model is used for serving that generates and updates the behavior vectors of all users at a preset period, and the reference model is used for generating the target vector of the expression distillation.

[0026] To achieve the above-mentioned purpose, a behavior vector-based user search method according to an embodiment of the present invention comprises: a step in which a tokenization unit embeds the value of a behavior tag, the type of a behavior tag, and the observation period for each of a plurality of behavior tags collected for each user, respectively, and sums them to convert them into cell-unit behavior tokens; a step in which a behavior encoder contextualizes the behavior tokens by an attention operation that learns the relationships between behavior tokens within the same user and the behavior pattern relationships between different users together, and compresses the contextualized behavior tokens into a single behavior vector at the user level; a step in which a query encoder converts a natural language request for a target user into a structured query including intent and conditions, and encodes the converted structured query into a query vector; a step in which a spatial alignment unit aligns the query vector and the behavior vector so that they are compared in the same embedding space by contrastive learning in which a user who responded and a user who did not respond in the execution log of an advertising campaign are respectively configured as positive learning pairs and negative learning pairs; a step in which a vector search unit searches for candidate users based on similarity with the query vector in the aligned embedding space; and a factual verification unit [for] the searched candidate user The method includes a step of querying behavior tags to check whether the conditions of the natural language request are met and selecting matching candidate users, wherein the behavior encoder is trained by a learning task that predicts the actual behavior of the user observed after the first period from behavior tags up to the first period.

[0027] The contrast learning in the above alignment step is based on a contrast loss that increases the similarity of the positive learning pairs and decreases the similarity of the negative learning pairs for a single query vector.

[0028] The above method further includes the step of a distillation learning unit training a lightweight serving model by expression distillation using the behavior encoder as a reference model, verifying whether search rankings are preserved based on the top K overlaps of similar user search results between the reference model and the serving model, and deploying the serving model that passed the verification to an operating environment.

[0029] To achieve the above-mentioned objective, a behavior vector-based user search method according to one embodiment of the present invention may be stored and recorded on a computer-readable recording medium on which a program for execution on a computer is recorded. Effects of the invention

[0030] According to the behavior vector-based user search system and method of the present invention, since the definition and search of behavior-based target users are possible solely through natural language requests from marketers, rapid and flexible target extraction can be realized without manual query conversion by data engineers.

[0031] According to the behavior vector-based user search system and method of the present invention, by cell-unit tokenization that embedding and summing the value, type, and observation period of behavior tags, respectively, and by attention operations that learn dual relationships within the same user and between users, it is possible to generate a behavior representation that captures the complex meaning when multiple behaviors appear together, even if the signal is weak with only a single behavior.

[0032] According to the behavior vector-based user search system and method of the present invention, users who responded and users who did not respond in the execution log of an advertising campaign are automatically configured into positive and negative learning pairs, thereby enabling the learning of alignment between the natural language query space and the behavior vector space without the need for separate manual label generation.

[0033] According to the behavior vector-based user search system and method of the present invention, by querying the behavior tags of candidate users obtained through vector similarity search and verifying whether they match actual behavior conditions, the mixing of candidates with similar behavior vectors but no actual behavior is blocked, and large-scale targets can be stably generated with only a small number of seed users.

[0034] According to the behavior vector-based user search system and method of the present invention, a serving model with a lightweight structure is trained by representation distillation using a behavior encoder as a reference model, and only serving models that have passed spatial preservation verification based on the top K overlaps of similar user search results are deployed to the operating environment, thereby enabling the efficient generation and updating of the behavior vectors of all users at a preset period while maintaining the continuity of representation quality and search ranking.

[0035] According to the behavior vector-based user search system and method of the present invention, the predictive performance of a behavior encoder is verified by sequential evaluation in which the period used for learning and the period used for evaluation are temporally separated, thereby preventing future information from leaking into the learning process and ensuring reliable search quality in an actual operating environment. Brief explanation of the drawing

[0036] FIG. 1 is a block diagram showing each configuration of a behavior vector-based user search system according to one embodiment of the present invention. FIG. 2 is a block diagram showing each configuration of a distillation learning unit according to one embodiment of the present invention. FIG. 3 is a flowchart illustrating the flow of a behavior vector-based user search method according to one embodiment of the present invention. Specific details for implementing the invention

[0037] Some embodiments of the present invention will be described in detail below with reference to the exemplary drawings. It should be noted that in assigning reference numerals to the components of each drawing, the same components are given the same reference numeral whenever possible, even if they are shown in different drawings.

[0038] Furthermore, in describing the embodiments of the present invention, if it is determined that a detailed description of related known configurations or functions would hinder understanding of the embodiments of the present invention, such detailed description is omitted.

[0039] In addition, terms such as first, second, A, B, (a), (b), etc., may be used when describing the components of the embodiments of the present invention. These terms are used merely to distinguish the components from other components, and the essence, order, or sequence of the components is not limited by the terms used.

[0040] In this specification, the singular form includes the plural form unless specifically stated otherwise in the text. As used in this specification, "comprising" and / or "comprising" does not exclude the presence or addition of one or more other components in addition to the mentioned components.

[0042] First, the key terms used in this specification are defined.

[0043] In this specification, a user refers to a consumer whose behavioral data is collected in online and mobile environments, and the user may be identified by a synthetic consumer identifier (SCI Universal ID; hereinafter also referred to as 'SUID'). The synthetic consumer identifier is a user identifier for combining and integrating user-specific behavioral data collected from multiple domains based on a single user. Here, a domain is a distinction of the source where the behavioral data originates; for example, commerce, app usage, ad response, location, TV viewing, etc., each correspond to a single domain.

[0044] User behavior data can be managed using a hierarchical tag system. A tag that records the user's behavior facts exactly as they are without interpretation is called a behavior fact tag (Tier 1 tag); a continuous score derived from the behavior fact tag to represent the interpretation of the user's current state is called a behavior state score (Tier 2 tag); and a probability value representing the likelihood of the user performing a specific behavior within a certain future period is called a behavior probability (Tier 3 tag). In this specification, the term "behavior tag" is used as a higher-level concept encompassing behavior fact tags, behavior state scores, and behavior probabilities. The "type of behavior tag" refers to the distinction of which tier an individual behavior tag belongs to among the aforementioned tiers, that is, the distinction of which type it corresponds to among behavior fact tags, behavior state scores, and behavior probabilities. The "value" of a behavior tag refers to the numerical or category value recorded in the corresponding tag, and the "observation period" refers to the period division of the behavior observation from which the value of the corresponding tag was calculated (e.g., 1 day, 7 days, 30 days, 90 days).

[0045] The behavior vector (z_ID) is a single user-unit vector that compresses and represents a user's diverse behavioral history; it is a learned representation in which users with similar behavioral patterns are placed close together in the vector space, while users with different behavioral contexts are placed far apart.

[0046] Seed users refer to a small set of users that serve as the basis for a search. For example, users who have completed conversions, such as purchases, in a specific past advertising campaign can be used as the seed user set.

[0047] The reference model refers to a behavior encoder trained with a focus on representation quality, and the serving model refers to a lightweight model trained by representation distillation to mimic the behavior vector representations generated by the reference model.

[0048] Hereinafter, the present invention will be described in more detail with reference to the attached drawings.

[0050] FIG. 1 is a block diagram showing each configuration of a behavior vector-based user search system according to one embodiment of the present invention.

[0051] Referring to FIG. 1, the behavior vector-based user search system (100) according to the present invention includes a tokenization unit (110), a behavior encoder (120), a query encoder (130), a spatial matching unit (140), a vector search unit (150), a factual verification unit (160), a distillation learning unit (170), and a performance verification unit (180).

[0052] The behavior tag storage (10) is an external element that stores multiple behavior tags collected per user and provides behavior tags to the tokenization unit (110). The natural language request (20) is a request in the form of natural language for a target user entered by a user such as a marketer, and is input into the query encoder (130). The advertising campaign execution log (30) is a log in which user-specific response status (e.g., response such as conversion, click, or no response) in past advertising campaigns is recorded, and serves as a source of learning pairs used for contrastive learning in the spatial matching unit (140). That is, the responding users and non-responding users recorded in the advertising campaign execution log (30) are configured as positive learning pairs and negative learning pairs, respectively, in the spatial matching unit (140).

[0053] To summarize the overall data flow, the behavior tags in the behavior tag storage (10) are converted into behavior tokens by the tokenization unit (110), compressed into user-unit behavior vectors by the behavior encoder (120), and loaded into the index of the vector search unit (150). The natural language request (20) is encoded into a query vector by the query encoder (130), sorted to be comparable in the same embedding space as the behavior vector by the spatial matching unit (140), the vector search unit (150) searches for candidate users in the sorted space, and the factual verification unit (160) checks whether the candidate users match the actual behavior conditions and selects and outputs the matching candidate users. Thus, the structure in which the outputs of two encoding paths, the path of behavior data (110, 120) and the path of natural language query (130), meet in the same embedding space at the spatial matching unit (140) is one of the core structures of the present invention.

[0055] The tokenization unit (110) embeds the value of the behavior tag, the type of the behavior tag, and the observation period for each of the multiple behavior tags collected per user, respectively, and sums them to convert them into cell-unit behavior tokens.

[0056] Specifically, user-specific behavior data can be managed in cell units where the user and the behavior tag intersect, and the tokenization unit (110) converts the tag value of each cell into a single independent token. At this time, the tokenization unit (110) generates an embedding for the value of the behavior tag, an embedding for the type of the behavior tag (e.g., whether it is a behavior fact tag, behavior state score, or behavior probability), and an embedding for the observation period (e.g., 1 day, 7 days, 30 days, 90 days), and sums the generated embeddings to form a single behavior token. As a result, heterogeneous behavior data with different scales of values, types of tags, and observation periods are uniformly represented in a token space of the same dimension, and subsequent attention operations can learn the relationships between behavior tags of different types and periods.

[0057] In one embodiment, the tokenization unit (110) can mitigate attention bias caused by the skewed distribution of values ​​by excluding cells with a value of 0 among numeric behavior tags from the quantile transformation target and processing them as separate zero mask tokens, and normalizing non-zero values ​​into a uniform distribution by quantile transformation and then embedding them. Since a significant number of behavior fact tags (e.g., about 85%) are numeric, such as usage time and number of purchases, and most of their values ​​have a skewed distribution clustered around 0, if the original values ​​are embedded as is, attention may be biased toward some features that have extreme values. The tokenization unit (110) mitigates this bias by processing cells with a value of 0 among numeric behavior tags as separate zero mask tokens, and normalizing non-zero values ​​into a uniform distribution by quantile transformation and then embedding them.

[0059] The behavior encoder (120) contextualizes the behavior tokens by an attention operation that learns the relationships between behavior tokens within the same user and the relationships between behavior patterns between different users, and compresses the contextualized behavior tokens into a single behavior vector at the user level.

[0060] Specifically, the attention operation of the behavior encoder (120) has a dual relationship attention (dual attention) structure that simultaneously learns the relationships between behavior tags within the same user and the behavior patterns between similar users within the same behavior tag. Through this, even if a single behavior is a weak signal, the meaning can be captured when multiple behaviors appear together. For example, a pattern in which the use of a car application, a visit to a dealership, and a search for a lease product appear simultaneously is contextualized into a complex meaning of the possibility of switching to a car lease, which is not captured by each individual behavior. In addition, by learning the behavior pattern relationships between different users together, the behavioral expression of a specific user is reinforced by the patterns of other users who have similar behavioral histories.

[0061] The behavior encoder (120) integrates contextualized behavior tokens on a user-by-user basis and compresses them into a single behavior vector (z_ID). In one embodiment, the behavior encoder (120) outputs the behavior vector by L2 normalization, thereby making it suitable for comparison between behavior vectors to be performed by cosine similarity. The dimension of the behavior vector can be set considering representation capacity and storage and retrieval efficiency, for example, it can be 1024 dimensions.

[0062] The behavior encoder (120) is trained by a learning task that predicts the actual behavior of the user observed after the first period from behavior tags up to the first period. For example, the behavior encoder (120) is repeatedly trained to predict next week's purchases, clicks, app usage, changes in interest, etc., based on behavior patterns up to this week, and in this process, rather than by explicitly defined rules, the user's potential intentions and behavioral stages are learned to be reflected in the behavior vector from the temporal changes in actual behavioral data. This time-based self-supervised learning task does not require separate manual labels, and allows the user's behavioral trajectory to be automatically accumulated as an expression defining the user during the prediction of the next behavior. That is, what behavioral flow the user is in and what behavior is likely to occur next are learned together in a single expression learning process.

[0064] The query encoder (130) converts a natural language request for a target user into a structured query including intents and conditions, and encodes the converted structured query into a query vector.

[0065] Specifically, the query encoder (130) receives a natural language request (20), such as a request from a marketer, such as "50,000 users with a high probability of converting to car lease," and converts the intent and conditions embedded in the request into a structured query. In one embodiment, the structured query includes at least one of an intent, a category, a behavior condition, an exclusion condition, a target size, and a time period condition, and may further include a priority. For example, for the request, the intent may be structured as a search for users capable of conversion (purchase), the category as cars, the behavior condition as the existence of car-related search / visit behaviors, and the target size as 50,000 people. The query encoder (130) encodes the converted structured query into a query vector and transmits it to the spatial matching unit (140).

[0066] In one embodiment, the query encoder (130) is trained by applying supervised fine-tuning based on Low-Rank Adaptation (LoRA) to a pre-trained language model. The query encoder (130) is a domain that directly processes natural language in the system, and while utilizing the language comprehension capabilities of a publicly available pre-trained language model, it secures domain-specific interpretation capabilities through supervised fine-tuning (SFT) based on training data consisting of pairs of natural language requests and corresponding structured correct queries. At this time, since only the low-rank adaptation parameters are adjusted without retraining the parameters of the entire language model, the ability specialized for request interpretation can be efficiently secured while maintaining the natural language comprehension capabilities of the existing language model.

[0067] In one embodiment, the map fine-tuning is adopted because the output of the query interpretation is not a free-form sentence, but a structured query in which the value of each item always changes within the same schema. For example, a query requesting users to lease a vehicle in a specific month and a query requesting users to lease a vehicle in a different month have different extraction periods even if they belong to the same category, and behavioral conditions valid only at that time must be reflected. Therefore, the query encoder (130) is trained not to memorize the correct answer, but to read the time and context information within the query sentence to fill in the value of each item of the structured query, and pairs of natural language requests and correct structured queries at different times and contexts can be constructed from the history of past advertising campaigns. In addition, in one embodiment, since the low-coefficient adaptation trains only the mapping to the structured query while preserving the language comprehension ability of the pre-trained language model, even when the category system of behavior tags is reorganized or a new domain is added, it is possible to respond by replacing only the low-coefficient adaptation parameters (adapters) without retraining the entire language model, and the damage to the existing language comprehension ability that may accompany full retraining is prevented, making operation and maintenance easy.

[0069] The spatial alignment unit (140) aligns the query vector and the behavior vector so that they are compared in the same embedding space by contrast learning, which configures the responding users and the non-responding users in the execution log of the advertising campaign into positive learning pairs and negative learning pairs, respectively.

[0070] The query vector generated by the query encoder (130) and the behavior vector generated by the behavior encoder (120) are inherently different outputs of different encoders and cannot be directly compared. The spatial alignment unit (140) automatically configures the behavior vectors of users who actually converted or responded to a specific query (campaign) from the advertising campaign execution log (30) as positive learning pairs for the corresponding query vector, and the behavior vectors of users who did not respond as negative learning pairs, and learns alignment through contrastive learning so that the query vector and the user behavior vectors that match the condition are placed in close positions in the same embedding space. As a result, natural language-based requests and actual user behavior data can be directly connected in the vector space.

[0071] In one embodiment, the contrast learning of the spatial alignment unit (140) replaces a separate manual label generation process based on a contrast loss that increases the similarity of the positive learning pairs and decreases the similarity of the negative learning pairs for a single query vector. That is, instead of humans manually generating labels for the relationship between the query and the user, actual data, such as the execution results of past campaigns, is automatically utilized as the labels for the learning pairs, thereby eliminating labeling costs and hypothesis bias. In one embodiment, the alignment may be performed in a two-tower structure in which a query-side encoder that generates query vectors and a behavior-side encoder that generates behavior vectors are placed as separate towers and the two towers are trained jointly, thereby allowing different types of inputs to be directly compared in the same coordinate system. For example, the InfoNCE objective function may be used as the contrast loss.

[0073] The vector search unit (150) searches for candidate users based on similarity with the query vector in the aligned embedding space.

[0074] In one embodiment, the vector search unit (150) searches for the candidate user based on cosine similarity. As described above, since the behavior encoder (120) outputs the behavior vector after L2 normalization, the proximity in the aligned embedding space can be consistently measured by cosine similarity.

[0075] In one embodiment, the vector search unit (150) can search for candidate users based on similarity with the reference vector, using the average of the behavior vectors of the seed user set as the reference vector. For example, when about 500 converts from a past specific campaign are given as the seed user set, the vector search unit (150) can construct the average vector of their behavior vectors as the reference vector (anchor) and search for users whose behavior patterns are similar to the reference vector, thereby stably expanding the candidate set to the size of tens of thousands from a small number of seeds.

[0076] In one embodiment, the vector search unit (150) returns top candidate users within a preset response time using an approximate nearest neighbor index loaded with the behavior vectors of all users. For example, the vector search unit (150) can return the top N candidate users closest to the query vector (or reference vector) within 300ms using an approximate nearest neighbor index of the Hierarchical Navigable Small World (HNSW) method loaded with the behavior vectors of about 34 million users.

[0078] The factual verification unit (160) checks the behavior tags of the searched candidate users to determine whether they meet the conditions of the natural language request and selects the candidate users that meet the conditions.

[0079] Searching based on vector similarity reflects the similarity of behavioral patterns, but even if the behavior vectors are similar, users for whom the requested behavioral conditions were not actually observed may be included in the candidates. The factual verification unit (160) queries the behavioral tags (behavioral fact tags, behavioral state scores, and behavioral probabilities) stored in the behavioral tag repository (10) for each searched candidate user and checks whether they meet the conditions of the structured query converted from the natural language request (e.g., behavioral conditions, exclusion conditions). For example, it queries the behavioral tags of candidate users searched for a request for "travel interest" to check whether travel-related behavior was actually observed, excludes candidate users who do not meet the request conditions, and selects and outputs only the matching candidate users. Through this factual verification, the fulfillment of actual behavioral conditions, which is not guaranteed by vector similarity alone, is ensured, thereby securing the reliability of the search results.

[0081] The distillation learning unit (170) uses the behavior encoder as a reference model and trains a serving model of a lightweight structure in which at least one of the number of attention layers, hidden dimensions, and number of parameters is reduced compared to the reference model by representation distillation.

[0083] FIG. 2 is a block diagram showing each configuration of a distillation learning unit according to one embodiment of the present invention.

[0084] Referring to FIG. 2, the distillation learning unit (170) includes a serving model (171), a space preservation verification unit (172), and a distribution unit (173).

[0085] Since the behavior encoder (120), which is trained with a focus on representation quality, has high computational and inference costs, a lightweight model is required for serving that rapidly generates and updates behavior vectors on a full user scale. To this end, in the distillation learning unit (170), the behavior encoder (120) generates a target vector as a reference model, and the same behavior tag input is provided to the serving model (171), and the serving model (171) is trained so that the behavior vector it outputs is close to the target vector of the reference model. That is, the representation distillation is to use the behavior vector generated by the reference model for the same behavior tag input as the target vector, and to train the serving model to reduce the cosine similarity and distance difference between the behavior vector output by the serving model and the target vector. At this time, since the vector representation itself is aligned as the target rather than simply mimicking the prediction label, the structure of the behavior representation space learned by the reference model is preserved in the serving model.

[0086] The serving model (171) is designed with a lightweight structure in which at least one of the number of attention layers, hidden dimension, and number of parameters is reduced compared to the reference model, so that the generation and updating of behavior vectors for a large number of users can be performed faster than the reference model.

[0087] In one embodiment, the distillation learning unit (170) verifies whether the search ranking of the serving model is preserved based on the top K overlap and similarity distribution between the similar user search results by the reference model and the similar user search results by the serving model, and distributes the serving model that has passed the verification to the operating environment. Specifically, the spatial preservation verification unit (172) compares the top K overlaps of the similar user search results of the reference model and the serving model, respectively, to determine whether users close to the reference model are similarly searched in the serving model, that is, whether the overlap ratio and cosine similarity distribution of the top K similar users satisfy the criteria. Since the verification is performed based on the search quality criterion of preserving the search ranking rather than simple vector error, changes in search results due to lightweighting are blocked before distribution. The distribution unit (173) distributes the serving model that has passed the verification to the operating environment, and the distributed serving model is used for index updating of the vector search unit (150), that is, for generating behavior vectors of all users and loading them into the approximate nearest neighbor index.

[0088] In one embodiment, the distributed serving model (171) is used for serving that generates and updates the behavior vectors of all users at a preset period, and the reference model is used for generating the target vectors of the representation distillation. That is, a dual operational structure is formed in which the reference model (behavior encoder (120)) focuses on securing representation quality and generating learning goals for the serving model, and the serving model (171) focuses on periodic vector update serving for a large number of users, thereby ensuring large-scale operational efficiency without loss of representation quality.

[0090] The performance verification unit (180) verifies the prediction performance of the behavior encoder (120) by sequential evaluation in which the period used for learning and the period used for evaluation are separated in time.

[0091] Since the training task of the behavior encoder (120) is to predict actual behavior after the first period from behavior tags up to the first period, if future information is leaked during training at the time of evaluation, performance in the actual operating environment is not guaranteed. The performance verification unit (180) provides reliable performance verification with time leakage blocked by separating the period used for training and the period used for evaluation in time, and by performing a sequential evaluation (walk-forward evaluation) in which, for example, a model trained with data up to a specific week is evaluated only with actual behavior data from the following week. Since the behavior encoder (120) that has passed the verification of the performance verification unit (180) is used as the reference model of the distillation learning unit (170), the baseline of the representation quality is guaranteed by the verification.

[0093] Hereinafter, as an organic operation of each of the above-described components, the batch processing and serving processing processes of the behavior vector-based user search system (100) according to one embodiment of the present invention are described.

[0094] Batch processing is performed at a preset interval (e.g., every morning) and is a process of generating and updating behavior vectors for all users to complete search preparation. First, the latest behavior tag snapshots of all users are extracted from the behavior tag repository (10), and the tokenization unit (110) performs cell-unit tokenization including quantile normalization of numeric behavior tags. Subsequently, behavior vectors are generated from the behavior tokens of all users. The entity performing the generation and updating of behavior vectors in this batch stage is the serving model (171) that has passed verification by the space preservation verification unit (172) and has been deployed to the operating environment by the distribution unit (173). That is, instead of the reference model (behavior encoder (120)) focused on representation quality being directly input into the daily vector generation for all users, the serving model (171), which has been distilled and verified to preserve the representation space of the reference model, takes charge of vector generation for batch processing, thereby ensuring the continuity of search quality and completing large-scale batches within a practical time. The generated behavior vector is L2 normalized and then loaded into the nearest neighbor index of the vector search unit (150) to complete the search preparation. In one embodiment, in batch processing, the behavior vector of all users is not partially updated with only the changes, but is entirely regenerated and replaced in its entirety at each period. Since the replacement is completed by the transition of the target referenced by the search index, the state in which new and old vectors are mixed during the replacement is not exposed to the search, and the entire behavior vector can be updated to the same reference point without interruption of service.

[0095] Serving processing is performed whenever a natural language request (20) is received. A query encoder (130) converts the natural language request into a structured query and encodes it into a query vector, and the query vector is processed to be comparable to the behavior vector in the same coordinate system according to the alignment learned by the spatial alignment unit (140). A vector search unit (150) searches for the candidate user closest to the query vector (or the reference vector based on the average of the seed user set) among the behavior vectors of all users loaded in the index in the aligned embedding space, and a factual verification unit (160) looks up the behavior tags of the searched candidate users to select only the candidate users that meet the request conditions. In this way, by utilizing the behavior vectors that the serving model (171) has pre-generated and loaded in batch processing for immediate search in serving processing, low-latency user search per request is realized.

[0096] Meanwhile, the reference model (behavior encoder (120)) is not directly input into the vector generation of batch processing, but is continuously used for the generation of target vectors for representation distillation and for quality control of the serving model (171) (comparison criteria for spatial preservation verification). That is, the aforementioned dual operation structure functions within the batch and serving process, and when the reference model is updated (retrained), the representation distillation and spatial preservation verification of the distillation learning unit (170) are performed again, and a new serving model is deployed, so that the improvement in representation quality is reflected in the operating environment without damaging the continuity of the search ranking.

[0098] FIG. 3 is a flowchart illustrating the flow of a behavior vector-based user search method according to one embodiment of the present invention.

[0099] Referring to FIG. 3, a behavior vector-based user search method according to one embodiment of the present invention includes a cell-unit tokenization step of a behavior tag (S110), a contextualization and behavior vector compression step by dual relational attention (S120), a structured query transformation and query vector encoding step of a natural language request (S130), a campaign log learning pair-based contrastive learning space alignment step (S140), a similarity-based candidate user search step in the aligned space (S150), and a step of determining whether the candidate user meets the request conditions (S160).

[0101] In step S110, the tokenization unit (110) embeds the value of the behavior tag, the type of the behavior tag, and the observation period for each of the multiple behavior tags collected per user, respectively, and sums them to convert them into cell-unit behavior tokens. As described above, among the numeric behavior tags, cells with a value of 0 are excluded from the quantile transformation target and are processed as separate zero mask tokens, and non-zero values ​​can be normalized to a uniform distribution by quantile transformation and then embedded.

[0102] In step S120, the behavior encoder (120) contextualizes the behavior tokens by an attention operation that learns the relationships between behavior tokens within the same user and the relationships between behavior patterns between different users, and compresses the contextualized behavior tokens into a single behavior vector at the user level. At this time, the behavior encoder (120) is trained by a learning task that predicts the actual behavior of the user observed after the first period from behavior tags up to the first period. The compressed behavior vector can be L2 normalized and loaded into an index of the vector search unit (150).

[0103] In step S130, the query encoder (130) converts a natural language request for a target user into a structured query including intents and conditions, and encodes the converted structured query into a query vector.

[0104] In step S140, the spatial alignment unit (140) aligns the query vector and the behavior vector so that they are compared in the same embedding space by contrast learning, which forms the responding users and the non-responding users in the execution log of the advertising campaign into positive learning pairs and negative learning pairs, respectively. In one embodiment, the contrast learning of the alignment step is based on a contrast loss that increases the similarity of the positive learning pairs and decreases the similarity of the negative learning pairs for a single query vector.

[0105] In step S150, the vector search unit (150) searches for candidate users based on similarity with the query vector in the sorted embedding space. As described above, a search using the average of the behavior vectors of the seed user set as a reference vector and a return within a preset response time using an approximate nearest neighbor index may be applied.

[0106] In step S160, the factual verification unit (160) checks whether the behavior tags of the searched candidate users meet the conditions of the natural language request and selects the matching candidate users. As illustrated in FIG. 3, if the candidate's behavior tags meet the request conditions, the candidate user is selected, and if they do not meet, they are eliminated. The set of selected candidate users can be used to configure the target audience for an advertising campaign, etc.

[0107] In one embodiment, the method may further include the step of a distillation learning unit (170) training a lightweight structure serving model (171) by expression distillation using the behavior encoder (120) as a reference model, verifying whether to preserve search rankings based on the top K overlaps of similar user search results between the reference model and the serving model, and distributing the serving model that passed the verification to an operating environment. As described above, the distributed serving model generates and updates the behavior vectors of all users at a preset period during batch processing.

[0109] Hereinafter, the specific implementation scale, detailed steps of expression distillation, and verification system of the behavior vector-based user search system (100) according to one embodiment of the present invention will be described.

[0110] In one embodiment, the behavior vector-based user search system (100) manages approximately 14,162 behavior tags for approximately 34 million SUIDs, and batch processing is performed at dawn every day. Specifically, it may take approximately 30 minutes to extract the latest behavior tag snapshot, approximately 20 minutes for cell tokenization of the tokenization unit (110) (including zero mask processing of numeric behavior tags and quantile normalization, maintenance of values ​​of behavior state scores and behavior probabilities, and processing of missing values), approximately 2 to 3 hours to generate behavior vectors of all users by the distributed serving model (171), and approximately 40 minutes to load the index of the vector search unit (150) after L2 normalization. For each request, the serving process can be completed within a total of 300ms, with less than 100ms required for query conversion and encoding by the query encoder (130), less than 100ms for searching for top candidates by the vector search unit (150), and less than 50ms for verifying condition matching by the factual verification unit (160). The dimension of the behavior vector is, for example, 1024 dimensions, and the size of the seed user set is, for example, about 500 past convertors, from which a candidate set of tens of thousands can be expanded and selected.

[0111] In one embodiment, a publicly available lightweight language model (e.g., a Gemma family model) may be used as the pre-trained language model of the query encoder (130), and the ground truth structure of the training data used for supervised fine-tuning based on low-reference adaptation (LoRA) may include intent, category, behavior condition, exclusion condition, target scale (target scale), time condition, and priority. An InfoNCE objective function may be used as the contrast loss of the spatial matching unit (140), and users who actually converted or responded in past campaign logs are configured as positive learning pairs, and users with low relevance or who did not respond are configured as negative learning pairs.

[0112] In one embodiment, the expression distillation of the distillation learning unit (170) can be performed in the following five steps (Table 1).

[0114] step Input → Output explanation ① Generation of reference action vectors Behavioral data → Behavioral vector of the reference model The trained behavior encoder (120) is fixed as a reference model, and user behavior data is input to generate a target vector that the serving model (171) must imitate. This target vector is used as a correct answer expression that reflects the meaning of the user's behavior. ② Lightweight Model Input Alignment Identical behavior data → Input into serving model The serving model (171) is provided with the same behavior tag input as the reference model. The serving model (171) is designed as a lightweight structure in which at least one of the number of attention layers, hidden dimensions, and number of parameters is reduced. ③ Learning by imitating expressions Serving model output → Approaches target vector Learning is performed to reduce the cosine similarity and distance difference between the two vectors so that the behavior vector output by the serving model (171) approaches the target vector. Since it is an imitation of the representation space itself rather than an imitation of the prediction label, the structure of the behavior representation space of the reference model is preserved. ④ Verification of spatial preservation Serving model behavior vector → Search quality verification The space preservation verification unit (172) compares the search results of similar users for each of the reference model and the serving model to verify whether the overlap and cosine similarity distributions of the top K similar users satisfy the criteria, that is, whether the search ranking is preserved. ⑤ Serving Distribution Serving Model That Passed Verification → Operational Environment The distribution unit (173) distributes the serving model that has passed verification to the operating environment. The distributed serving model is used for serving that generates and updates the behavior vectors of all users at a preset period, and the reference model is used for generating the target vector for expression distillation (dual operation).

[0116] In one embodiment, the verification system of the performance verification unit (180) may be configured as follows (Table 2). All evaluations are performed in a sequential evaluation (walk-forward) manner in which the learning time and the evaluation time are separated to prevent leakage of future information, and for example, learning can be done with data up to a specific point in time and evaluation can be done only with actual behavior from that point in time up to 7 days later (T+7).

[0118] step Checklist characteristic explanation 1 Accuracy of ranking GAUC The average value obtained by separately calculating the ranking power of whether high-scoring users actually performed actions for each category. Since action ratios vary by category and mixing them causes distortion, the values ​​are calculated separately and then integrated. 2 Correction of probability PCOC The ratio of the sum of predicted probabilities to the actual number of actions. 1.0 indicates accurate correction, and a value greater than 1 indicates overprediction. 3 Monotonicity between hierarchies F <S<P Verify whether prediction accuracy increases in the order of behavior fact tags (Layer 1), behavior state scores (Layer 2), and behavior probabilities (Layer 3). 4 Random contrast effect Lift The value obtained by dividing the conversion rate of the top set selected by the model by the overall average conversion rate. For example, it can be set based on a minimum of 2.0. 5 Factuality of the screening results Verification of facts It queries the extracted user behavior tags to check for the existence of actual behaviors corresponding to the request conditions, and if there are no behaviors, it excludes them. 6 Blocking time leakage walk-forward T+7 By temporally separating the period used for learning from the period used for evaluation, it structurally guarantees that future information was not used for learning.

[0120] Hereinafter, various extended embodiments of the present invention are described. The extended embodiments described below are operations performed by each of the components described above or by two or more components working together, and can be combined and implemented together within a single behavior vector-based user search system (100).

[0121] In one embodiment, the behavior encoder (120) may be implemented or used in combination with a structure other than an attention operation-based encoder. For example, the behavior encoder (120) may be implemented as a sequence encoder based on a recurrent neural network (RNN) that processes the user's behavior history in chronological order, or it may be implemented as a multi-axis attention structure that separates and alternately applies an axis for learning the relationship between behavior tokens within the same user and an axis for learning the relationship between behavior patterns between different users. Even in the case of such an alternatively implemented behavior encoder (120), the system of taking cell-unit behavior tokens generated by the tokenization unit (110) as input, learning by a learning task that predicts actual behavior after the first period from behavior tags up to the first period, and outputting a single L2-normalized user-unit behavior vector is maintained identically, and it is identical to the above-described embodiment in that it serves as a reference model for the distillation learning unit (170) and is a subject of sequential evaluation for the performance verification unit (180).

[0122] In another embodiment, the behavior encoder (120) may perform a pre-training task before or in parallel with the time-based prediction learning task. Specifically, the behavior encoder (120) may be pre-trained by a masking and restoration task that masks some (e.g., about 15%) of the input behavior tokens and restores the values ​​of the masked behavior tokens, a contrastive learning task that forms positive pairs of sets of behavior tokens from different views configured from the behavior history of the same user, and Distributionally Robust Optimization (DRO) that learns to be robust against deviations in the data distribution between domains and categories, and then fine-trained by the time-based prediction learning task described above. In this case, the performance verification unit (180) may be used to select the learning configuration with the highest verification performance by performing sequential evaluations for each combination of learning tasks.

[0123] In another embodiment, the training data used for supervised fine-tuning of the query encoder (130) can be automatically generated by a separate pre-trained language model. Specifically, in the training phase of the query encoder (130), the first language model is configured to synthesize a natural language request for training and pair it with a corresponding structured query for the correct answer, and the synthesized natural language request-structured query pair can be used for supervised fine-tuning of the query encoder (130). This overcomes the limitations of the scale of training data built manually and ensures query interpretation capabilities even for new categories where actual marketer requests have not been sufficiently accumulated. Furthermore, the query vector encoded from the synthesized natural language request can be utilized as a query anchor for categories lacking campaign execution history in the contrastive learning of the spatial matching unit (140).

[0124] In another embodiment, the spatial matching unit (140) can perform multiple positive contrast learning using multiple positive learning pairs simultaneously for a single query vector. Specifically, the spatial matching unit (140) can simultaneously configure behavior vectors of multiple users who responded to the same campaign in the advertising campaign execution log (30) as positive learning pairs for a single query vector (anchor), and configure the contrast loss to increase the similarity of the multiple positive learning pairs together and decrease the similarity of the negative learning pairs. Due to the characteristics of the advertising campaign execution log (30), where multiple converters exist in a single campaign, multiple positive configuration increases the utilization efficiency of learning pairs and the stability of matching learning.

[0125] In another embodiment, the query encoder (130) and the behavior encoder (120) may be coupled to perform query interpretation that reflects the behavioral characteristics of the seed user set. Specifically, the query encoder (130) receives the behavior vector of the seed user set (e.g., the average vector of the behavior vectors of the seed users) generated by the behavior encoder (120) (or the distributed serving model (171)) as a soft prompt, that is, a prefix input in the form of a continuous vector, and processes it together with the natural language request to supplement the behavioral conditions of the structured query so as to reflect the actual behavioral characteristics of the seed user set. The query vector encoded from the supplemented structured query is used as a search criterion for the vector search unit (150) in the embedding space aligned by the spatial alignment unit (140), thereby allowing the unique behavioral patterns of the seed set, which are not specified by the natural language request alone, to be reflected in the search.

[0126] In another embodiment, the distillation learning unit (170) may perform a retraining loop based on the results of the spatial preservation verification. Specifically, if the top K overlap or similarity distributions in the verification of the spatial preservation verification unit (172) fall short of the standard, the distribution unit (173) suspends the distribution of the corresponding serving model, and the serving model (171) additionally performs representation imitation learning with adjusted learning conditions (e.g., learning rate, size or composition of the training data), and can be distributed to the operating environment by the distribution unit (173) only if it passes the re-verification. At this time, the verification of the spatial preservation verification unit (172) may be configured to prevent time leakage during the distillation and verification process by sharing the principle of temporally separating the period used for learning and the period used for evaluation, just like the sequential evaluation of the performance verification unit (180).

[0127] In another embodiment, the lightweighting of the serving model (171) may be performed in parallel with pruning or quantization in addition to expression distillation. In this case as well, the serving model (171) to which pruning or quantization has been applied is distributed by the distribution unit (173) on the condition that it passes the same verification based on the top K overlap and similarity distributions of the spatial preservation verification unit (172), so that a single quality criterion of search ranking preservation is maintained regardless of the type of lightweighting technique.

[0128] The aforementioned extended embodiments can be organically combined within a single system. For example, the behavior encoder (120), which is pre-trained with masking restoration, contrast learning, and distribution robust optimization, is supplied as a reference model to the distillation learning unit (170) regardless of its implementation structure (attention-based, recurrent neural network-based, or multi-axis attention-based), and the natural language request-correct answer structured query pairs synthesized by the first language model are commonly used for the supervised fine-tuning of the query encoder (130) and the query anchor configuration of the multiple positive contrast learning of the spatial matching unit (140), and the behavior vector generated and updated by the distributed serving model (171) in batch processing is reused together for the search of the vector search unit (150) and the soft prompt input configuration of the query encoder (130), and the verification gate of the spatial preservation verification unit (172) is commonly applied to the retraining loop and the pruning and quantization parallel lightweighting, while sharing the time separation evaluation principle of the performance verification unit (180).

[0130] In this specification, each component of the behavior vector-based user search system (100), namely the tokenization unit (110), behavior encoder (120), query encoder (130), spatial matching unit (140), vector search unit (150), factual verification unit (160), distillation learning unit (170), and performance verification unit (180), may be a processor that executes a series of processes stored in memory. Alternatively, it may operate as a software module driven and controlled by a processor. Furthermore, the processor may be a hardware device.

[0131] Meanwhile, the method according to the present invention may be combined with a computer, which is hardware, and may be implemented as a program for executing the behavior vector-based user search method on a computer or stored on a computer-readable recording medium. That is, according to one embodiment of the present invention, a computer-readable recording medium on which a program for executing the above-described behavior vector-based user search method on a computer is recorded may be provided. The computer-readable recording medium may 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.

[0132] In this specification, the term "part" includes a unit realized by hardware, a unit realized by software, and a unit realized using both. Additionally, one unit may be realized using two or more hardware, and two or more units may be realized by one hardware.

[0133] The scope of protection of the present invention is not limited to the descriptions and expressions of the embodiments explicitly described above. Furthermore, the scope of protection of the present invention cannot be limited by obvious modifications or substitutions within the technical field to which the present invention pertains. Explanation of the symbols

[0134] 10: Behavior Tag Repository 20: Natural Language Request 30: Ad Campaign Execution Log 100: Behavior Vector-Based User Search System 110: Tokenization section 120: Behavior Encoder 130: Query Encoder 140: Spatial alignment 150: Vector Search Section 160: Fact-checking Department 170: Distillation Learning Department 171: Serving Model 172: Space Preservation Verification Department 173: Distribution Section 180: Performance Verification Department

Claims

Claim 1 A tokenization unit that, for each of the multiple behavior tags collected per user, embeds the value of the behavior tag, the type of the behavior tag, and the observation period respectively, sums them, and converts them into cell-unit behavior tokens; a behavior encoder that contextualizes the behavior tokens by an attention operation that learns the relationships between behavior tokens within the same user and the behavior pattern relationships between different users together, compresses the contextualized behavior tokens into a single behavior vector at the user level, and is trained by a learning task that predicts the actual behavior of the user observed after the first period from the behavior tags up to the first period; a query encoder that converts a natural language request for a target user into a structured query including intent and conditions, and encodes the converted structured query into a query vector; a spatial alignment unit that aligns the query vector and the behavior vector so that they are compared in the same embedding space by contrastive learning, which forms responding users and non-responding users in the execution log of an advertising campaign into positive learning pairs and negative learning pairs, respectively; and a vector search unit that searches for candidate users based on similarity with the query vector in the aligned embedding space. A behavior vector-based user search system characterized by including a fact verification unit that checks whether the behavior tags of the searched candidate users meet the conditions of the natural language request and selects the matching candidate users by querying the behavior tags of the searched candidate users. Claim 2 A behavior vector-based user search system according to claim 1, wherein the tokenization unit normalizes the value of a numeric behavior tag into a uniform distribution by quantile transformation and then embedding it to mitigate attention bias caused by the skewness of the value distribution. Claim 3 A behavior vector-based user search system according to claim 1, wherein the behavior encoder outputs the behavior vector after L2 normalization, and the vector search unit searches for the candidate user based on cosine similarity. Claim 4 A behavior vector-based user search system according to claim 1, wherein the vector search unit searches for the candidate user based on similarity with the reference vector, using the average of the behavior vectors of a seed user set as a reference vector. Claim 5 A behavior vector-based user search system according to claim 1, wherein the vector search unit returns a top candidate user within a preset response time using an approximate nearest neighbor index loaded with the behavior vectors of all users. Claim 6 A behavior vector-based user search system according to claim 1, wherein the query encoder is trained by applying low-reference adaptation (LoRA)-based supervised fine-tuning to a pre-trained language model, and the structured query includes at least one of an intent, category, behavior condition, exclusion condition, target scale, and period condition. Claim 7 A behavior vector-based user search system according to claim 1, wherein the contrast learning of the spatial matching unit replaces a separate manual label generation process based on a contrast loss that increases the similarity of the positive learning pairs and decreases the similarity of the negative learning pairs for a single query vector. Claim 8 A behavior vector-based user search system according to claim 1, wherein the system further comprises a performance verification unit that verifies the prediction performance of the behavior encoder by sequential evaluation in which the period used for learning and the period used for evaluation are temporally separated. Claim 9 The behavior vector-based user search system according to claim 1, wherein the system further comprises a distillation learning unit that learns a serving model of a lightweight structure in which at least one of the number of attention layers, hidden dimension, and number of parameters is reduced relative to the reference model, using the behavior encoder as a reference model, by expression distillation, and wherein the expression distillation is characterized by taking a behavior vector generated by the reference model for the same behavior tag input as a target vector and learning the serving model to reduce the cosine similarity and distance difference between the behavior vector output by the serving model and the target vector. Claim 10 A behavior vector-based user search system according to claim 9, wherein the distillation learning unit verifies whether the search ranking of the serving model is preserved based on the top K overlap and similarity distributions between the similar user search results by the reference model and the similar user search results by the serving model, and deploys the serving model that passed the verification to an operating environment. Claim 11 A behavior vector-based user search system according to claim 10, wherein the distributed serving model is used for serving to generate and update the behavior vector of all users at a preset period, and the reference model is used for generating the target vector of the expression distillation. Claim 12 A tokenization unit, for each of the multiple behavior tags collected per user, embeds the value of the behavior tag, the type of the behavior tag, and the observation period, respectively, and sums them to convert them into cell-unit behavior tokens; a behavior encoder, contextualizes the behavior tokens by an attention operation that learns the relationships between behavior tokens within the same user and the behavior pattern relationships between different users together, and compresses the contextualized behavior tokens into a single behavior vector at the user level; a query encoder, converts a natural language request for a target user into a structured query including intent and conditions, and encodes the converted structured query into a query vector; a spatial alignment unit, aligns the query vector and the behavior vector so that they are compared in the same embedding space by contrastive learning, which forms responding users and non-responding users in the execution log of an advertising campaign into positive learning pairs and negative learning pairs, respectively; a vector search unit, searches for candidate users based on similarity with the query vector in the aligned embedding space. A behavior vector-based user search method comprising a factual verification unit that includes the step of querying the behavior tags of the searched candidate users to check whether they meet the conditions of the natural language request and selecting the matching candidate users, wherein the behavior encoder is characterized by being trained by a learning task that predicts the actual behavior of the user observed after the first period from the behavior tags up to the first period. Claim 13 A behavior vector-based user search method according to claim 12, wherein the contrast learning of the alignment step is based on a contrast loss that increases the similarity of the positive learning pairs and decreases the similarity of the negative learning pairs for a single query vector. Claim 14 A behavior vector-based user search method according to claim 12, wherein the method further comprises the step of a distillation learning unit training a lightweight structure serving model by expression distillation using the behavior encoder as a reference model, verifying whether to preserve search rankings based on the top K overlap degrees of similar user search results between the reference model and the serving model, and deploying the serving model that passed the verification to an operating environment. Claim 15 A computer-readable recording medium having a program recorded thereon for executing a behavior vector-based user search method of any one of claims 12 to 14 on a computer.

Images