A product selection decision method based on dynamic consumption trend prediction and related products

Abstract

The application discloses a product selection decision method based on dynamic consumption trend prediction and related products. First, the global interactive users of the target brand are obtained to obtain a plurality of target users, and the multi-modal behavior data set of each target user in the target time period is obtained to construct a user portrait library including a plurality of user semantic vectors. The user semantic vectors in the time sliding window are obtained to obtain a plurality of target semantic vectors, and the target semantic vectors are subjected to time sequence dynamic semantic clustering analysis to obtain a plurality of current semantic clustering clusters. The number of target users corresponding to each current semantic clustering cluster is identified, and the trend score of each current semantic clustering cluster is calculated in combination with a trend scoring formula. Finally, the target clustering cluster is screened based on the trend score of each current semantic clustering cluster, and a product selection report is generated in combination with the similarity between the to-be-selected SKU and the target clustering cluster. The application can more sensitively capture the real needs of consumers and dynamically predict future consumption trends.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, specifically to a product selection decision-making method and related products based on dynamic consumer trend prediction. Background Technology

[0002] Traditional Customer Relationship Management (CRM) systems primarily focus on analyzing historical data, merely reviewing "what was sold in the past." By collecting and analyzing past sales data, they extract basic operational insights such as best-selling products and peak sales periods. However, these systems lack forward-looking assessments of market demand and cannot deeply interpret consumers' current real needs.

[0003] Traditional CRM systems can only clearly present past sales results, but they lack the ability to predict future market trends and capture dynamic changes in consumer demand. Modern consumers' product needs iterate rapidly and their preferences are diverse. If companies cannot accurately capture these real-time changing demand signals, new product development will become detached from market reality. This will not only make it difficult to create trend-aligned blockbuster products, but may also lead to sluggish sales due to a mismatch between new product positioning and market demand, resulting in ineffective investment in R&D and production resources.

[0004] Therefore, how to capture consumers' real needs and dynamically predict consumption trends is a technical problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] To address the aforementioned issues, this application provides a product selection decision-making method and related products based on dynamic consumption trend prediction, which can capture consumers' real needs and dynamically predict consumption trends.

[0006] The embodiments of this application disclose the following technical solutions:

[0007] A product selection decision-making method based on dynamic consumer trend prediction, the method comprising:

[0008] The system acquires multiple target users by obtaining the full-domain interactive users of the target brand, and obtains a multimodal behavior dataset for each target user within a target time period. The multimodal behavior dataset includes at least one behavior record. Each behavior record includes inventory quantity unit (SKU) attributes, search keywords, customer service dialogue records, product reviews, or return reasons generated based on different user interaction behaviors.

[0009] A user profile database is constructed based on the multimodal behavior dataset corresponding to each target user. The user profile database includes multiple user semantic vectors used to describe the consumption characteristics and product demand characteristics of each target user. Each user semantic vector corresponds to a behavior record. The user semantic vector is marked with a time stamp, which is the generation time of the behavior record corresponding to the user semantic vector.

[0010] Multiple target semantic vectors are obtained by acquiring user semantic vectors whose timestamps are within a time sliding window. Then, a density-based clustering algorithm is used to perform temporal dynamic semantic clustering analysis on the multiple target semantic vectors to obtain multiple current semantic clusters.

[0011] Identify the number of target users corresponding to each current semantic cluster, and calculate the trend score for each current semantic cluster based on the trend scoring formula and the number of target users corresponding to each current semantic cluster; the trend score is used to evaluate the development potential and explosive growth of the consumption trend corresponding to the current semantic cluster.

[0012] The target clusters are selected based on the trend scores of each current semantic cluster, and a product selection report is generated by combining the similarity between the candidate SKU and the target clusters.

[0013] In one possible implementation, the step of using a density-based clustering algorithm to perform temporal dynamic semantic clustering analysis on the multiple target semantic vectors to obtain multiple current semantic clusters includes:

[0014] When the clustering execution phase is the first temporal clustering, a density-based clustering algorithm is directly used to cluster the multiple target semantic vectors to obtain multiple current semantic clusters;

[0015] When the clustering execution stage is non-first-time sequential clustering, the cluster center of each historical semantic cluster obtained most recently is used as the initial center point, and the multiple target semantic vectors are clustered based on each initial center point to obtain multiple current semantic clusters.

[0016] In one possible implementation, constructing a user profile library based on the multimodal behavior dataset corresponding to each target user includes:

[0017] For each target user, at least one first semantic reasoning prompt word is generated based on at least one behavior record corresponding to the target user and a first preset instruction. The at least one first semantic reasoning prompt word is then input into a large model for analysis to obtain at least one first semantic tag set. Each first semantic reasoning prompt word corresponds to a first semantic tag set. The first semantic tag set is used to describe the user's consumption characteristics and product demand characteristics. Each first semantic tag set includes at least the user's consumption intention, user's purchase scenario, user's consumption style, and product pain points.

[0018] Vectorize the first semantic tag set corresponding to each target user to obtain at least one user semantic vector;

[0019] All user semantic vectors are integrated, stored, and time-annotated to obtain the user profile library.

[0020] In one possible implementation, the method further includes:

[0021] When the target user corresponds to at least two first semantic tag sets, the first semantic tag sets with at least N duplicate tags are merged to obtain a merged tag set, and the original duplicate first semantic tag sets are deleted; N is a positive integer;

[0022] The step of vectorizing the first semantic tag set corresponding to each of the target users to obtain at least one user semantic vector includes:

[0023] Vectorize the first semantic tag set and / or the merged tag set corresponding to each target user to obtain at least one user semantic vector.

[0024] In one possible implementation, the trend scoring formula is as follows:

[0025] Strend=α×log(Vcurrent)+β×(Vcurrent-Vpast) / Vpast;

[0026] Where Strend is the trend score; α is the size weight; Vcurrent is the number of users in the current semantic cluster; β is the growth weight; and Vpast is the number of users in the most recently obtained historical semantic cluster corresponding to the current semantic cluster.

[0027] In one possible implementation, the step of filtering target clusters based on the trend scores of each current semantic cluster and generating a product selection report by combining the similarity between the candidate SKU and the target cluster includes:

[0028] Screen the current semantic clustering clusters according to the screening rules and the trend scores of each current semantic clustering cluster to obtain at least one of the target clustering clusters;

[0029] Obtain the product semantic vectors of each candidate SKU of the target brand party, and calculate the similarity between each product semantic vector and the clustering center of the target clustering cluster;

[0030] Generate the product selection report based on the similarity.

[0031] In a possible implementation, one similarity corresponds to one candidate SKU;

[0032] The generating the product selection report based on the similarity includes:

[0033] If the similarity is greater than or equal to the first similarity threshold, generate a production recommendation for the candidate SKU corresponding to the similarity; the production recommendation is used to recommend direct production of the candidate SKU and recommend increasing the first batch of stocking quantity;

[0034] If the similarity is less than the first similarity threshold and greater than or equal to the second similarity threshold, generate an optimization and adjustment recommendation; the optimization and adjustment recommendation is used to recommend that the R & D department modify the design for the missing pain points or only conduct a small batch of product testing;

[0035] If the similarity is less than the second similarity threshold, generate an elimination recommendation; the elimination recommendation is used to recommend directly eliminating or not putting into production the candidate SKU;

[0036] Integrate the production recommendation, the optimization and adjustment recommendation and the elimination recommendation to generate the product selection report;

[0037] Wherein, the first similarity threshold is greater than the second similarity threshold.

[0038] In a possible implementation, the obtaining the product semantic vectors of each candidate SKU of the target brand party includes:

[0039] Obtain the core attributes of each candidate SKU, and convert the core attributes of each candidate SKU into a text description to obtain a structured product description; the core attributes are used to describe the style, version, material, function and applicable scenario of the candidate SKU;

[0040] For each structured product description, generate a second semantic inference prompt word based on the structured product description and a second preset instruction, and input the second semantic inference prompt word into a large model for analysis to obtain a second semantic label set; the second semantic label set is used to describe the core product features of the candidate SKU, and each second semantic label set at least includes style, version, material, function and applicable scenario;

[0041] Each second semantic tag set is vectorized to obtain multiple product semantic vectors.

[0042] A product selection decision-making device based on dynamic consumer trend prediction, the device comprising:

[0043] The first acquisition unit is used to acquire multiple target users from the target brand's full-domain interactive users, and to acquire a multimodal behavior dataset for each target user within a target time period; the multimodal behavior dataset includes at least one behavior record; each behavior record includes inventory quantity unit SKU attribute, search keywords, and customer service dialogue records generated based on different user interaction behaviors, as well as product reviews or return reasons;

[0044] The user profile database construction unit is used to construct a user profile database based on the multimodal behavior dataset corresponding to each target user. The user profile database includes multiple user semantic vectors that describe the consumption characteristics and product demand characteristics of each target user. Each user semantic vector corresponds to a behavior record. The user semantic vector is marked with a time stamp, which is the generation time of the behavior record corresponding to the user semantic vector.

[0045] The second acquisition unit is used to acquire user semantic vectors whose time markers are within a time sliding window to obtain multiple target semantic vectors;

[0046] Clustering unit, used to perform temporal dynamic semantic clustering analysis on the multiple target semantic vectors using a density-based clustering algorithm to obtain multiple current semantic clusters;

[0047] The user count identification unit is used to identify the number of users of the target user corresponding to each current semantic cluster;

[0048] The scoring unit is used to calculate the trend score of each current semantic cluster based on the trend scoring formula and the number of target users corresponding to each current semantic cluster; the trend score is used to evaluate the development potential and explosive growth of the consumption trend corresponding to the current semantic cluster.

[0049] The cluster filtering unit is used to filter target clusters based on the trend scores of each current semantic cluster;

[0050] The first report generation unit is used to generate a product selection report by combining the similarity between the candidate SKUs and the target cluster.

[0051] A product selection decision-making device based on dynamic consumer trend prediction includes: a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the product selection decision-making method based on dynamic consumer trend prediction as described above.

[0052] Compared with the prior art, this application has the following beneficial effects:

[0053] This application provides a product selection decision-making method and related products based on dynamic consumption trend prediction. Specifically, when implementing the product selection decision-making method based on dynamic consumption trend prediction provided in this application, the following steps are taken: First, the target brand's full-domain interactive target users and the multimodal behavior dataset of each target user within a specified time period can be obtained. Then, a user profile library is constructed based on the multimodal behavior dataset of each target user. The profile library includes multiple user semantic vectors with time stamps. Subsequently, user semantic vectors with time stamps within a time sliding window are extracted to obtain target semantic vectors. A density-based clustering algorithm is used to perform time-series dynamic semantic clustering analysis on these target semantic vectors to obtain multiple current semantic clusters. Next, the number of target users corresponding to each current semantic cluster is identified. The trend score of the current semantic cluster is calculated by combining the trend scoring formula and the number of target users corresponding to the current semantic cluster, thereby evaluating the development potential and explosive degree of the consumption trend corresponding to each cluster. Finally, target clusters are selected from the current semantic clusters based on the trend score, and a product selection report is generated by combining the similarity between the candidate SKU and the target cluster, providing data support for the brand's product selection decision-making.

[0054] This application, by collecting multimodal behavioral datasets of target brand users across all interactive domains, enables a more comprehensive and in-depth understanding of consumers' actual needs. Simultaneously, employing density-based clustering algorithms for temporal dynamic semantic clustering analysis allows for the identification of user groups with similar demand characteristics, enabling the timely discovery of new or emerging consumption trends. Furthermore, by calculating cluster trend scores, the development potential and explosive growth of each consumption trend are scientifically assessed, thereby accurately predicting market trends and formulating scientifically sound product selection decisions. This not only enhances the insight into consumer needs but also helps businesses seize market opportunities. Attached Figure Description

[0055] To more clearly illustrate the technical solutions in this embodiment or the prior art, the drawings used in the description of the embodiment or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0056] Figure 1A flowchart illustrating a product selection decision-making method based on dynamic consumer trend prediction, provided for an embodiment of this application;

[0057] Figure 2 A flowchart illustrating a method for constructing a user profile database, as provided in an embodiment of this application;

[0058] Figure 3 A flowchart illustrating a product selection report generation method provided in this application embodiment;

[0059] Figure 4 A flowchart illustrating a product semantic vector generation method provided in this application embodiment;

[0060] Figure 5 This is a schematic diagram of a product selection decision-making device based on dynamic consumption trend prediction, provided as an embodiment of this application. Detailed Implementation

[0061] To facilitate understanding of the technical solutions provided in the embodiments of this application, the background technology involved in the embodiments of this application will be described below.

[0062] Traditional CRM systems are limited to statistical analysis and review of historical transaction data, only clearly presenting predetermined results such as "what was sold in the past, which product categories sold well, and the average order value," failing to capture consumers' true consumption needs and potential purchasing motivations through data appearances. Traditional CRM systems can only rely on structured tags such as gender, age, and spending amount to build user profiles. This cannot interpret the implicit intentions behind unstructured behaviors such as user searches, customer service inquiries, and product reviews, nor can it perceive changes in market consumption trends. Essentially, it can only answer "what were the sales results," but cannot answer "why consumers bought and what they still want."

[0063] If a brand's new product development and selection decisions are based solely on such historical data, it is highly susceptible to falling into the trap of "lagging behind": on the one hand, it can only replicate past best-selling products, making it difficult to tap into emerging market demands; on the other hand, blindly testing products based on experience fails to match current consumer preferences, leading to a disconnect between new products and market trends. Ultimately, this results in missing the market opportunity to capitalize on high-growth, latent consumption trends and create blockbuster new products, and may also lead to sluggish sales due to inaccurate product positioning, resulting in a waste of multiple resources in R&D, production, and inventory, making it difficult to establish a differentiated product advantage in market competition.

[0064] To address this issue, this application provides a product selection decision-making method and related products based on dynamic consumer trend prediction. First, it acquires multiple target users from the target brand's full-domain interactive users and collects multimodal behavior datasets for each target user within a specific time period. Then, a user profile library is constructed based on each target user's multimodal behavior dataset. Next, by acquiring user semantic vectors within a time sliding window, dynamic temporal clustering analysis is performed on these semantic vectors, using a density-based clustering algorithm to divide the user semantic vectors into multiple current semantic clusters. Based on this, the number of users in each semantic cluster is identified, and a trend score formula is applied to calculate the trend score for each cluster. The trend score reflects the consumption trend development potential and explosive growth of each cluster, providing a basis for product selection decisions. Finally, based on the trend scores of each cluster, the most promising target clusters are selected, and combined with the similarity between the candidate SKUs and the target clusters, a precise product selection report is generated to guide subsequent product selection and market layout. This application not only captures dynamic consumer needs in real time but also proactively predicts future consumption trends, effectively avoiding the risks of blind new product development and market disconnect.

[0065] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.

[0066] See Figure 1 The figure is a flowchart of a product selection decision-making method based on dynamic consumer trend prediction provided in an embodiment of this application. Figure 1 As shown, the product selection decision-making method based on dynamic consumption trend prediction may include steps S101-S105:

[0067] S101: Obtain the target brand's full-domain interactive users to obtain multiple target users, and obtain the multimodal behavior dataset of each target user within the target time period.

[0068] In order to capture consumers' consumption behavior and potential needs in a comprehensive and multi-dimensional way, we first identified the target brand's full-domain interactive users as the analysis objects, determined multiple target users, and collected multimodal behavior datasets of each target user within the target time period.

[0069] The multimodal behavior dataset contains at least one user behavior record, and each record integrates core information generated from different user interactions. Specifically, it includes four main categories: StockKeepingUnit (SKU) attributes, search keywords, customer service conversation records, and product reviews or return reasons. For example, among the users of a certain outdoor equipment brand, one user generated multiple types of interactions over the past eight months. Their multimodal behavior dataset would include the SKU attributes of the hiking boots purchased (such as shoe size 42, waterproof material, mid-top style, etc.), records of searching for keywords like "lightweight camping tent" and "carbon fiber trekking poles" on the brand's official website, conversations with online customer service regarding "hiking boot anti-slip performance" and "camping gear storage tips," product reviews of a purchased camping mat stating "thick enough but breathable is average," and a return reason given due to "hiking boots running small." These different types of interaction information would be comprehensively collected into the user's multimodal behavior dataset, forming a complete record of behavioral data.

[0070] In one possible implementation, the target time period is a range from one month to one year prior to the current date, that is, it covers all user interaction behaviors in the most recent month and the past year. The specific duration of the target time period can be flexibly selected within this range according to the actual analysis needs.

[0071] It should be noted that "omni-channel interactive users" refers to consumers accumulated within the target brand's private domain system, as well as authorized members or potential consumers who have interacted with the target brand on mainstream e-commerce platforms (including browsing, adding to cart, purchasing, customer service inquiries, etc.).

[0072] S102: Construct a user profile database based on the multimodal behavior dataset corresponding to each target user.

[0073] To transform discrete multimodal user behavior data into quantifiable and analyzable standardized features, and to accurately characterize the consumption and product demand features of each target user, a dedicated user profile library can be constructed based on the multimodal behavior dataset corresponding to each target user. This user profile library contains multiple user semantic vectors describing the consumption and demand features of each target user, following a matching rule where one user semantic vector corresponds to one behavior record. Furthermore, each user semantic vector is annotated with a corresponding timestamp, which is consistent with the actual generation time of the behavior record corresponding to the vector. This preserves the temporal dimension of user behavior, laying the foundation for subsequent time-series trend analysis.

[0074] For example, suppose a target user of a beauty brand searches for "long-lasting foundation". The system first generates a corresponding first semantic reasoning prompt based on the behavior record and the first preset instruction. After inputting the prompt into the large model, a semantic tag set containing consumption intention, purchase scenario, consumption style and product pain point is obtained. For example, if the tag set clearly indicates that the user's consumption intention is "to purchase long-lasting foundation", the purchase scenario is "daily commuting", the consumption style is "pursuing high cost performance", and the product pain point is "rejecting foundation that is prone to caking and fading", then the first semantic tag set is vectorized to generate the corresponding user semantic vector, and the specific time when the search behavior occurred is marked on the vector.

[0075] S103: Obtain user semantic vectors with time stamps within a time sliding window to obtain multiple target semantic vectors, and use a density-based clustering algorithm to perform temporal dynamic semantic clustering analysis on the multiple target semantic vectors to obtain multiple current semantic clusters.

[0076] To focus on analyzing recent user consumption demand characteristics, accurately identify user groups with similar needs, and dynamically capture emerging or changing consumption trends, user semantic vectors with timestamps within a preset time sliding window can be extracted from a pre-constructed user profile database and integrated to obtain multiple target semantic vectors. This time sliding window slides forward at a fixed step size to form continuous window slices, enabling time-series tracking of user needs. Simultaneously, density-based clustering algorithms (such as DBSCAN or Hdbscan) are used to perform time-series dynamic semantic clustering analysis on the target semantic vectors within each window slice. User groups are aggregated based on the semantic similarity between vectors, ultimately resulting in multiple current semantic clusters. Each cluster represents a user community with similar consumption needs and behavioral characteristics. Furthermore, by comparing the cluster density and size changes of adjacent window slices, the development trend of corresponding consumption trends can be monitored.

[0077] For example, suppose a clothing brand sets the time sliding window to the past 7 days with a sliding step of 1 day. Then, the period from the 1st to the 7th will form the first window slice, and the period from the 2nd to the 8th will form the next window slice. The system will continuously generate consecutive window slices according to this rule. Taking the window slice from the 1st to the 7th as an example, the semantic vectors of all users within this time period are extracted from the user profile database as target semantic vectors. These vectors cover the demand features corresponding to the user's search, consultation, purchase, and evaluation behaviors within these 7 days. After clustering analysis of these target semantic vectors through a density-based clustering algorithm, multiple current semantic clusters will be formed. For example, one cluster contains a large number of vectors containing semantic features such as "Y2K retro tops", "sequin embellishments", and "spicy girl style outfits", representing a similar demand user group for "Y2K outfits". Another cluster aggregates vectors containing semantic features such as "old-money style knitwear", "solid color straight-leg pants", and "simple commuting outfits", becoming a demand user community for "old-money style commuting". Subsequently, by comparing the user scale and vector density of the "Y2K fashion" cluster in window slices 2-8, it can be determined whether this consumption trend is in a boom or decline phase. The clusters in each window slice can clearly reflect the core consumption needs of similar users at present.

[0078] S104: Identify the number of target users corresponding to each current semantic cluster, and calculate the trend score of each current semantic cluster based on the trend scoring formula and the number of target users corresponding to each current semantic cluster.

[0079] To scientifically and quantitatively assess the development potential and explosive growth of consumption trends corresponding to each current semantic cluster, and to provide a reference for subsequent product selection decisions, we can first identify the number of target users corresponding to each current semantic cluster. Then, using a specific trend scoring formula, we can substitute the user data for each cluster to calculate the corresponding trend score. This score can intuitively reflect the current scale and growth trend of various consumption trends, distinguishing between mainstream consumer groups and high-growth emerging implicit consumption trends. The trend scoring formula is as follows:

[0080] Strend=α×log(Vcurrent)+β×(Vcurrent-Vpast) / Vpast;

[0081] Where Strend is the trend score; α is the size weight; Vcurrent is the number of users in the current semantic cluster; β is the growth weight; and Vpast is the number of users in the most recently obtained historical semantic cluster corresponding to the current semantic cluster.

[0082] It should be noted that when a completely new cluster is found (i.e., it does not exist in the previous period, Vpast=0), Vpast in the formula can be replaced with Vpast+∈ (∈ is usually a very small constant with a value of 1).

[0083] It should also be noted that α (scale weight) and β are configurable hyperparameters, and their sum is usually 1. In practical applications, if the brand aims to find a stable core market, it will increase α (e.g., α=0.7, β=0.3); if the brand aims to tap into an early-stage, high-growth blue ocean market, it will increase β (e.g., α=0.3, β=0.7).

[0084] For example, suppose a clothing brand calculates a trend score based on clustering results over a 7-day time window. If the brand wants to tap into an early, high-growth blue ocean market, the weights are configured as α=0.3 and β=0.7: First, for the "Y2K fashion" cluster, if the current number of users Vcurrent=500 and the historical number of users in the previous period Vpast=200, substituting these values ​​into the formula, the trend score for this cluster will emphasize its high-growth characteristics, resulting in a significantly higher score. Second, for the newly emerging "outdoor lightweight cycling" cluster, since there is no historical data Vpast=0, Vpast is replaced with 0+1=1. If the current number of users Vcurrent=100, the formula will calculate the trend score based on its current size and the high growth rate from zero, accurately identifying this emerging blue ocean trend. If the brand wants to solidify its core user base, it can adjust the weights to α=0.7 and β=0.3. In this case, mature clusters such as "classic commuter outfits" with a current user count of Vcurrent=2000 and a historical user count of Vpast=1800 will have a higher trend score due to their larger scale weight, which aligns with the brand's decision-making needs for steady development.

[0085] S105: Filter target clusters based on the trend scores of each current semantic cluster, and generate a product selection report by combining the similarity between the candidate SKU and the target cluster.

[0086] The trend score of each current semantic cluster is used as the core screening criterion to select target clusters with high development potential and high growth rate. After semantic parsing of the selected SKUs, the semantic similarity between them and the target clusters is calculated. Finally, the similarity matching results are combined to generate a targeted product selection report, providing accurate data analysis support for brands' new product development, product layout and other product selection decisions.

[0087] Based on the content of S101-S105, the process first involves acquiring multimodal behavioral data of the target brand's interactive users across all domains, and then constructing a user profile library containing time-stamped user semantic vectors. Subsequently, target semantic vectors within a time-sliding window are extracted, and density clustering is used to obtain the current semantic clusters. A trend score is then calculated for each cluster using a trend scoring formula to assess the potential of consumer trends. Finally, high-potential target clusters are selected based on the trend scores, and a product selection report is generated by combining the similarity between candidate SKUs and target clusters, providing precise support for brand product selection decisions. This application, through multi-level data analysis and dynamic clustering, achieves accurate capture of consumer needs and prediction of future market trends, effectively avoiding the risks of blind new product development and market disconnect.

[0088] In one possible implementation, when using a density-based clustering algorithm to cluster multiple target semantic vectors to obtain multiple current semantic clusters, different clustering strategies are adopted according to the differences in the stage of clustering execution. This ensures the continuity and consistency of cluster identification during the temporal clustering process and avoids the problem of cluster ID jitter. Accordingly, step S103 uses a density-based clustering algorithm to perform temporal dynamic semantic clustering analysis on the multiple target semantic vectors to obtain multiple current semantic clusters, including:

[0089] If the clustering execution stage is the first time-series clustering, since there is no historical clustering reference data, a density-based clustering algorithm will be directly used to perform clustering analysis on the target semantic vector without initial centers, generating multiple current semantic clusters.

[0090] If the clustering execution phase is a non-first-time sequential clustering, the cluster centers of each historical semantic cluster obtained most recently will be retrieved first and used as the initial center points for this clustering. Then, the target semantic vector will be clustered based on these initial center points, and finally multiple current semantic clusters will be obtained.

[0091] For example, suppose a clothing brand conducts time-series dynamic clustering analysis using a 7-day sliding window and a 1-day step size. When clustering the target semantic vectors from number 1 to 7, this is the first time-series clustering. At this time, the DBSCAN density clustering algorithm is directly used to cluster all semantic vectors within this window, resulting in multiple current semantic clusters such as "Y2K fashion" and "old-fashion commuting," and the cluster center of each cluster is determined. When entering a non-first time-series clustering phase, analyzing the target semantic vectors from number 2 to 8, instead of re-clustering without a center, the cluster centers of all historical clusters such as "Y2K fashion" and "old-fashion commuting" obtained from number 1 to 7 are used as the initial center points for this clustering. Based on these initial points, density clustering is performed on the target semantic vectors from number 2 to 8 to generate the current semantic clusters within this window. This ensures that similar consumer trend clusters in the two windows can be matched, effectively solving the problem of cluster ID jitter and allowing the system to accurately track the continuous changes in various consumer trends.

[0092] See Figure 2 , Figure 2 This application provides a flowchart of a method for constructing a user profile database. Accordingly, step S102, which constructs the user profile database based on the multimodal behavior dataset corresponding to each target user, can be specifically implemented through steps S201-S203:

[0093] S201: For each target user, generate at least one first semantic reasoning prompt word based on at least one behavior record corresponding to the target user and a first preset instruction, and input the at least one first semantic reasoning prompt word into a large model for analysis to obtain at least one first semantic tag set.

[0094] To transform discrete, multimodal user behavior data into structured, high-dimensional consumer demand characteristics and accurately uncover users' deep-seated consumption motivations and preferences, for each target user, at least one first semantic inference prompt word can be generated based on at least one corresponding multimodal behavior record and pre-defined standardized instructions. This prompt word clarifies the inference direction and output format requirements of the large model. The generated first semantic inference prompt word is then input into the large model for deep semantic analysis and intent deduction, outputting at least one first semantic tag set, with a one-to-one correspondence between the first semantic inference prompt word and the first semantic tag set. This first semantic tag set comprehensively describes the user's consumption characteristics and product demand characteristics. Each tag set includes at least four core dimensions: user consumption intent, user purchase scenario, user consumption style, and product pain points. Extended dimension tags such as price sensitivity can also be added according to analytical needs.

[0095] For example, suppose a target user of an outdoor product brand is identified. First, their multimodal behavior records are integrated. Then, a first semantic inference prompt generated by a first preset instruction is used: "Based on the following user multimodal behavior data: [SKU attributes: black, XL size, windproof and waterproof, outdoor functional windproof jacket; search terms: 'Where to go this weekend,' 'Hiking shoe recommendations'; customer service dialogue: 'How is the windproof effect of this jacket? Is it suitable for outdoor hiking?'; return record: a jacket that is too thin], infer the user's deeper consumption intent, specific usage scenarios, and pain points. Please output in JSON format, which must include the following fields: intent, scene, pain_point." `style_preference` is used to input the prompt word into the large model. After semantic reasoning, the large model will output the corresponding first semantic tag set, for example: `{"intent":"Select windproof and warm outdoor clothing suitable for outdoor hiking","scene":"Weekend suburban outdoor hiking scene","pain_point":"Reject outdoor clothing with thin fabric and poor windproof and warmth","style_preference":"Outdoor functional style","price_sensitive":"Medium"}`. This tag set not only covers the four core dimensions of features, but also adds an extended tag for price sensitivity, clearly and structurally presenting the user's consumption characteristics and product demand characteristics.

[0096] S202: Vectorize the first semantic tag set corresponding to each target user to obtain at least one user semantic vector.

[0097] To transform structured semantic tag sets into quantified data that can be clustered and similarity calculated in vector space, and to meet the algorithmic requirements of subsequent time-series dynamic clustering and trend analysis, embedding models such as BGE-M3 can be used to vectorize the first semantic tag set corresponding to each target user. The model maps the text features in the tag set into high-dimensional numerical vectors, and generates a user semantic vector for each first semantic tag set. Finally, at least one user semantic vector that can represent the consumption characteristics and product demand characteristics of each target user is obtained.

[0098] For example, suppose a target user A of an outdoor brand has a first semantic tag set generated for their inquiry about windproof jackets as {"intent":"purchasing an outdoor hiking windproof jacket","scene":"weekend suburban hiking","pain_point":"thin fabric, not windproof","style_preference":"outdoor functional style","price_sensitive":"medium"}. After inputting this tag set into the BGE-M3 model, the model will convert the text features such as scene, style, and pain point into a string of high-dimensional numerical vectors, which is the user semantic vector corresponding to this behavior. If user A also has another first semantic tag set corresponding to their camping equipment search behavior, after processing by the same Embedding model, another user semantic vector will be generated, thereby realizing the quantitative representation of the user's multi-dimensional needs.

[0099] S203: Integrate and store all user semantic vectors along with time annotations to obtain the user profile library.

[0100] To create standardized user data assets that can support subsequent time-series dynamic clustering and trend analysis, this application integrates and centrally stores the user semantic vectors of all target users obtained through vectorization. At the same time, it accurately labels the corresponding behavior record generation time for each user semantic vector, so that each vector has a clear time dimension feature (i.e., time stamp). Finally, it constructs a structured, traceable, and time-series user profile library. This profile library can completely retain the consumption demand characteristics of users at different times, providing a data foundation for subsequent extraction of vectors by time dimension and analysis of changes in consumption trends.

[0101] It should be noted that the time dimension feature (i.e., time stamp) of a user's semantic vector is the generation time of the corresponding behavior record. For example, if a user searches for "lightweight camping tent" on March 10th, the first semantic inference prompt word generated based on this behavior record is analyzed by a large model to obtain the first semantic tag set, which is then vectorized by the Embedding model to obtain the corresponding user semantic vector. The time stamp of this vector is then marked as March 10th.

[0102] For example, after an outdoor brand completes semantic reasoning and vectorization of the behavior of its target users across all domains, it integrates the semantic vectors generated by user A searching for outdoor equipment on March 1st, inquiring about outdoor clothing on March 5th, user B purchasing hiking boots on March 2nd, reviewing camping mats on March 6th, and all other users' corresponding semantic vectors into a unified database. Each vector is labeled with its corresponding generation time, such as March 1st, March 5th, March 2nd, and March 6th. All time-stamped user semantic vectors together constitute the brand's user profile database. In subsequent analysis, vectors within any time range, such as the past 7 days or the past 30 days, can be directly extracted from this profile database to conduct dynamic clustering analysis and accurately track the temporal changes in user needs and consumption trends.

[0103] In one possible implementation, to reduce the volume of vector data in the user profile database, condense and highlight the core consumption needs of users, avoid feature redundancy caused by duplicate tag sets, and improve the overall efficiency of subsequent vectorization, clustering analysis, and trend calculation, the method also includes a tag set merging optimization step:

[0104] When at least two first semantic tag sets are generated for a single target user, a tag merging operation is performed on multiple first semantic tag sets containing at least N duplicate tags to integrate them into a unified merged tag set, and the original first semantic tag sets containing duplicate tags are deleted. Here, N is a positive integer that can be flexibly configured according to the actual needs and feature analysis precision.

[0105] For example, assuming N=3, the target users of an outdoor brand generate two sets of first semantic tags under different behaviors. Tag set 1 is {"intent":"selecting outdoor hiking gear","scene":"weekend suburban hiking","pain_point":"thin fabric, not windproof","style_preference":"outdoor functional style","price_sensitive":"medium"}, and tag set 2 is {"intent":"selecting outdoor hiking apparel","scene":"weekend suburban hiking","pain_point":"thin fabric, not windproof","style_preference":"outdoor functional style","price_sensitive"}. The tags ":"high"} are completely duplicated in the two tag sets: scene, pain_point, and style_preference, satisfying the merging condition of N=3. Therefore, the two tags are merged to generate the merged tag set {"intent":"Selecting outdoor hiking equipment and clothing","scene":"Weekend suburban hiking","pain_point":"Thin fabric, not windproof","style_preference":"Outdoor functional style","price_sensitive":"Medium to high"}. The original tag set 1 and tag set 2 are deleted from the data. This retains the core needs of users while integrating the details of needs under different behaviors, achieving a simplification and optimization of the tag set.

[0106] In one possible implementation, vectorizing the first semantic tag set corresponding to each target user to obtain at least one user semantic vector includes:

[0107] Vectorize the first semantic tag set and / or the merged tag set corresponding to each target user to obtain at least one user semantic vector.

[0108] See Figure 3 , Figure 3 The flowchart of a product selection report generation method provided in this application embodiment is as follows: Accordingly, step S105 filters target clusters based on the trend scores of each current semantic cluster, and generates a product selection report by combining the similarity between the candidate SKU and the target cluster. Specifically, this can be implemented through steps S301-S303:

[0109] S301: Combine the filtering rules and the trend scores of each current semantic cluster to filter each current semantic cluster and obtain at least one target cluster.

[0110] To accurately identify high-value consumer trend communities that align with a brand's market strategy, product development, and other practical business needs, and to define the core direction for subsequent precise matching of people and products and product selection decisions, a comprehensive screening can be conducted on all clusters using the trend scores of each current semantic cluster as the core quantitative basis, based on the pre-set screening rules of the brand. From this, at least one target cluster that meets the screening criteria can be identified. This type of cluster represents a consumer trend with development potential that matches the brand's positioning.

[0111] The selection rules can be flexibly customized by the brand based on its own market layout, product development and other development needs. The brand can set a specific trend score threshold and directly identify the current semantic clusters with trend scores greater than the threshold as target clusters. Alternatively, the brand can set a ranking quantification index and designate the current semantic clusters with the top M trend scores as target clusters. This will allow the brand to accurately select high-potential consumer trend communities that meet its development needs.

[0112] S302: Obtain the product semantic vector of each candidate SKU of the target brand, and calculate the similarity between each product semantic vector and the cluster center of the target cluster.

[0113] To achieve precise semantic matching between consumer demand and product characteristics, and to quantify the fit between candidate products and high-value consumption trends, providing a quantifiable basis for product selection decisions, we can first conduct multimodal semantic parsing and vectorization processing on all candidate SKUs from the target brand. This extracts core features such as product category, style, function, and applicable scenarios, generating a unique product semantic vector for each candidate SKU. Then, using a vector similarity algorithm, we calculate the semantic similarity between each product semantic vector and the cluster center of the target cluster. The similarity score directly represents the degree of fit between the product characteristics and the consumer demand trends represented by the target cluster.

[0114] For example, suppose an outdoor brand has identified a target cluster of "lightweight camping equipment" and determined its cluster center vector. The brand also sets 10 camping tents and 8 outdoor folding tables and chairs as candidate SKUs. First, multimodal analysis is performed on each candidate SKU. For instance, for an "ultra-lightweight portable folding camping tent (suitable for 2-3 people, rainproof and sunproof, lightweight design)," its core features are extracted and a corresponding product semantic vector is generated. Similarly, the vectorization of all 18 candidate SKUs is completed. Then, the cosine similarity between each product semantic vector and the cluster center vector of "lightweight camping equipment" is calculated. The similarity between two lightweight portable tents reaches 0.92 and 0.89 respectively, the similarity between a folding picnic table and a similarity of 0.85, while the similarity between several heavy-duty camping tents is only 0.4-0.5, clearly quantifying the degree of matching between different candidate SKUs and the target consumer trend.

[0115] S303: Generate the product selection report based on the similarity.

[0116] By combining the similarity calculation results of the semantic vectors of all candidate SKUs with the cluster centers of the target clusters, the suitability of each product can be quantitatively ranked and analyzed dimensionally. This integrates key information such as the core characteristics of high-potential consumer trends, the matching degree between products and trends, and suitability priorities to generate a structured product selection report. The report not only presents the matching degree ranking of candidate SKUs for different high-potential trends but also provides targeted product selection suggestions based on the development potential of the trends and the characteristics of the products. It clearly identifies core products that fit high-potential consumer trends, candidate products that can be optimized and upgraded, and products with low suitability that should be excluded.

[0117] In one possible implementation, one similarity level corresponds to one candidate SKU.

[0118] In one possible implementation, step S303 generates the product selection report based on the similarity, specifically including:

[0119] If the similarity is greater than or equal to the first similarity threshold, a production recommendation suggestion is generated for the candidate SKU corresponding to the similarity; the production recommendation suggestion is used to suggest that the candidate SKU be put into production directly and to suggest that the initial stock quantity be increased.

[0120] If the similarity is less than the first similarity threshold and greater than or equal to the second similarity threshold, optimization and adjustment suggestions are generated; the optimization and adjustment suggestions are used to suggest that the R&D department modify the design to address the missing pain points, or conduct only small-batch testing.

[0121] If the similarity is less than the second similarity threshold, a elimination suggestion is generated; the elimination suggestion is used to suggest that the candidate SKU be eliminated directly or not put into production for the time being.

[0122] The production recommendation, optimization and adjustment recommendation, and elimination recommendation are integrated to generate the product selection report;

[0123] Wherein, the first similarity threshold is greater than the second similarity threshold.

[0124] In one possible implementation, the size of the first similarity threshold can be set to, but is not limited to, 0.8, and the user can adjust the size of the first similarity threshold according to actual needs.

[0125] In one possible implementation, the size of the second similarity threshold can be set to, but is not limited to, 0.6, and the user can adjust the size of the second similarity threshold according to actual needs.

[0126] See Figure 4 , Figure 4This application provides a flowchart of a product semantic vector generation method. Accordingly, step S302, obtaining the product semantic vectors of each candidate SKU of the target brand, can be specifically implemented through steps S401-S403:

[0127] S401: Obtain the core attributes of each candidate SKU and convert the core attributes of each candidate SKU into a text description to obtain a structured product description.

[0128] To standardize and extract the core product features of candidate SKUs and form a unified and parsable product description text, laying the foundation for subsequent semantic reasoning and vectorization processing, we can first comprehensively extract the core attributes of each candidate SKU. These attributes focus on the key feature dimensions of the product, specifically including style, fit, material, function, and applicable scenarios. Then, the extracted multi-dimensional core attributes are transformed into standardized text descriptions, ultimately forming a structured product description, achieving a clear and unified presentation of product features.

[0129] For example, suppose we have a candidate SKU for outdoor apparel. We extract its core attributes as follows: style is outdoor functional, fit is loose hooded, material is windproof and waterproof fabric, function is breathable, wear-resistant and tear-resistant, and applicable scenarios are suburban hiking and camping. Then we transform these core attributes into the corresponding structured product description: "This product is an outdoor functional style, with a loose hooded fit. The main material is windproof and waterproof fabric, which has the functional characteristics of breathability, wear resistance and tear resistance, and is suitable for suburban hiking, outdoor camping and other usage scenarios."

[0130] S402: For each of the structured product descriptions, generate a second semantic reasoning prompt based on the structured product description and the first preset instruction, and input the second semantic reasoning prompt into the large model for analysis to obtain a second semantic tag set.

[0131] To transform standardized product text descriptions into structured, clearly defined product feature tags, and to meet the needs of subsequent vectorization and semantic similarity calculations, a second semantic inference prompt is generated for each candidate SKU's structured product description, combined with pre-defined domain-specific standardized instructions. These prompts clarify the analysis direction, output dimensions, and format requirements of the large model. The semantic inference prompts are then input into the large model for deep semantic parsing and feature extraction, ultimately outputting a second semantic tag set that accurately represents the core features of the product, achieving a structured and tagged expression of product features. The second semantic tag set describes the core product features of the candidate SKU, and each second semantic tag set includes at least style, fit, material, function, and applicable scenarios.

[0132] For example, suppose the structured product description of a candidate outdoor jacket is "This product is an outdoor functional style, with a loose hooded fit. The main material is windproof and waterproof waterproof fabric, which has the functional characteristics of breathability, abrasion resistance, and tear resistance, and is suitable for use in suburban hiking, outdoor camping, and other usage scenarios." The second semantic reasoning prompt generated by the second preset instruction is: "Based on the following product structured description, please extract the core features of this product and generate a structured tag set, which must include five core dimensions: style, fit, material, function, and applicable scenarios. Output in JSON format, with fields style, version, material, function, and scene respectively." After inputting this prompt into the large model, the corresponding second semantic tag set output after analysis is: {"style":"outdoor functional style","version":"loose hooded style","material":"windproof and waterproof waterproof fabric","function":["breathable","abrasion resistance","tear resistance"],"scene":["suburban hiking","outdoor camping"]}.

[0133] S403: Vectorize each second semantic tag set to obtain multiple product semantic vectors.

[0134] To achieve quantitative calculation and semantic matching of product features and consumer demand features in a unified vector space, embedding models such as BGE-M3 can be used to vectorize the second semantic tag set corresponding to each candidate SKU. The model maps the structured product text features in the tag set into high-dimensional numerical vectors. Each second semantic tag set generates a unique product semantic vector, and finally, the product semantic vector set of all candidate SKUs is obtained. This lays a quantitative data foundation for subsequent calculation of the semantic similarity between products and high-potential consumer trend clusters and for achieving accurate matching of people and goods.

[0135] See Figure 5 , Figure 5 This is a schematic diagram of a product selection decision-making device based on dynamic consumer trend prediction, provided as an embodiment of this application. Figure 5 As shown, the product selection decision-making device based on dynamic consumer trend prediction includes:

[0136] The first acquisition unit 501 is used to acquire multiple target users from the target brand's full-domain interactive users, and to acquire a multimodal behavior dataset for each target user within a target time period; the multimodal behavior dataset includes at least one behavior record; each behavior record includes inventory quantity unit SKU attribute, search keywords, and customer service dialogue records generated based on different user interaction behaviors, as well as product reviews or return reasons;

[0137] The user profile database construction unit 502 is used to construct a user profile database based on the multimodal behavior dataset corresponding to each target user; the user profile database includes multiple user semantic vectors for describing the consumption characteristics and product demand characteristics of each target user, and one user semantic vector corresponds to one behavior record; the user semantic vector is marked with a time stamp, and the time stamp is the generation time of the behavior record corresponding to the user semantic vector;

[0138] The second acquisition unit 503 is used to acquire user semantic vectors whose time markers are within a time sliding window to obtain multiple target semantic vectors.

[0139] Clustering unit 504 is used to perform temporal dynamic semantic clustering analysis on the multiple target semantic vectors using a density-based clustering algorithm to obtain multiple current semantic clusters;

[0140] User count identification unit 505 is used to identify the number of users of the target user corresponding to each current semantic clustering cluster;

[0141] Scoring unit 506 is used to calculate the trend score of each current semantic cluster based on the trend scoring formula and the number of target users corresponding to each current semantic cluster; the trend score is used to evaluate the development potential and explosive growth of the consumption trend corresponding to the current semantic cluster.

[0142] Clustering cluster filtering unit 507 is used to filter target clusters based on the trend scores of each current semantic cluster;

[0143] The first report generation unit 508 is used to generate a product selection report by combining the similarity between the candidate SKUs and the target cluster.

[0144] In one possible implementation, the clustering unit 504 is specifically used for:

[0145] When the clustering execution phase is the first temporal clustering, a density-based clustering algorithm is directly used to cluster the multiple target semantic vectors to obtain multiple current semantic clusters;

[0146] When the clustering execution stage is non-first-time sequential clustering, the cluster center of each historical semantic cluster obtained most recently is used as the initial center point, and the multiple target semantic vectors are clustered based on each initial center point to obtain multiple current semantic clusters.

[0147] In one possible implementation, the portrait library construction unit 502 specifically includes:

[0148] The first semantic tag set generation unit is used to generate at least one first semantic reasoning prompt word for each target user based on at least one behavior record corresponding to the target user and a first preset instruction, and input the at least one first semantic reasoning prompt word into a large model for analysis to obtain at least one first semantic tag set; one first semantic reasoning prompt word corresponds to one first semantic tag set, and the first semantic tag set is used to describe the user's consumption characteristics and product demand characteristics. Each first semantic tag set includes at least the user's consumption intention, user's purchase scenario, user's consumption style, and product pain points.

[0149] The first vectorization unit is used to vectorize the first semantic tag set corresponding to each of the target users to obtain at least one user semantic vector;

[0150] An integrated storage unit is used to integrate and store all user semantic vectors and time-annotate them to obtain the user profile library.

[0151] In one possible implementation, the device further includes:

[0152] The merge and delete unit, when the target user corresponds to at least two first semantic tag sets, is used to merge first semantic tag sets with at least N duplicate tags to obtain a merged tag set, and delete the original duplicate first semantic tag sets; N is a positive integer.

[0153] In one possible implementation, the first vectorization unit is specifically used for:

[0154] Vectorize the first semantic tag set and / or the merged tag set corresponding to each target user to obtain at least one user semantic vector.

[0155] In one possible implementation, the trend scoring formula is as follows:

[0156] Strend=α×log(Vcurrent)+β×(Vcurrent-Vpast) / Vpast;

[0157] Where Strend is the trend score; α is the size weight; Vcurrent is the number of users in the current semantic cluster; β is the growth weight; and Vpast is the number of users in the most recently obtained historical semantic cluster corresponding to the current semantic cluster.

[0158] In one possible implementation, the cluster filtering unit 507 is specifically used for:

[0159] By combining the filtering rules and the trend scores of each current semantic cluster, at least one target cluster is obtained.

[0160] In one possible implementation, the first report generation unit 508 specifically includes:

[0161] The third acquisition unit is used to acquire the product semantic vectors of each candidate SKU of the target brand.

[0162] The similarity calculation unit is used to calculate the similarity between the semantic vector of each product and the cluster center of the target cluster.

[0163] The second report generation unit is used to generate the product selection report based on the similarity.

[0164] In one possible implementation, one similarity level corresponds to one candidate SKU.

[0165] In one possible implementation, the second report generation unit is specifically used for:

[0166] If the similarity is greater than or equal to the first similarity threshold, a production recommendation suggestion is generated for the candidate SKU corresponding to the similarity; the production recommendation suggestion is used to suggest that the candidate SKU be put into production directly and to suggest that the initial stock quantity be increased.

[0167] If the similarity is less than the first similarity threshold and greater than or equal to the second similarity threshold, optimization and adjustment suggestions are generated; the optimization and adjustment suggestions are used to suggest that the R&D department modify the design to address the missing pain points, or conduct only small-batch testing.

[0168] If the similarity is less than the second similarity threshold, a elimination suggestion is generated; the elimination suggestion is used to suggest that the candidate SKU be eliminated directly or not put into production for the time being.

[0169] The production recommendation, optimization and adjustment recommendation, and elimination recommendation are integrated to generate the product selection report;

[0170] Wherein, the first similarity threshold is greater than the second similarity threshold.

[0171] In one possible implementation, the third acquisition unit is specifically used for:

[0172] Obtain the core attributes of each candidate SKU and convert them into text descriptions to obtain structured product descriptions; the core attributes are used to describe the style, pattern, material, function, and applicable scenarios of the candidate SKUs;

[0173] For each structured product description, a second semantic reasoning prompt is generated based on the structured product description and the second preset instruction. The second semantic reasoning prompt is then input into a large model for analysis to obtain a second semantic tag set. The second semantic tag set is used to describe the core product features of the candidate SKU. Each second semantic tag set includes at least style, pattern, material, function, and applicable scenarios.

[0174] Each second semantic tag set is vectorized to obtain multiple product semantic vectors.

[0175] In addition, this application embodiment also provides a product selection decision device based on dynamic consumption trend prediction, including: a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the product selection decision method based on dynamic consumption trend prediction as described above.

[0176] This application's embodiments, by acquiring a multimodal behavioral dataset of all interactive users of the target brand, break through the limitations of traditional data dimensions, enabling comprehensive and in-depth mining and interpretation of consumers' true consumption needs and implicit motivations. Simultaneously, it employs a density-based clustering algorithm to conduct time-series dynamic semantic clustering analysis on user semantic vectors, accurately identifying user communities with similar demand characteristics. This allows for the keen capture of emerging and developing consumption trends, achieving real-time perception of changes in market demand. Based on this, a trend score is calculated for each current semantic cluster using a trend scoring formula, scientifically quantifying the development potential and explosive growth of each consumption trend, effectively distinguishing between mainstream consumption needs and high-growth implicit consumption trends. Finally, a product selection report is generated by combining the semantic similarity analysis of candidate SKUs and high-potential target clusters, deeply integrating consumption trend prediction with actual product selection decisions. This achieves both accurate and forward-looking prediction of market trends and enables enterprises to create product selection solutions that meet market demands, leading to scientific, reasonable, and targeted product selection decisions, significantly improving the matching degree between new product development and market demand.

[0177] The foregoing has provided a detailed description of a product selection decision-making method based on dynamic consumer trend prediction and related products provided in this application. The various embodiments in the specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since it corresponds to the method disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to in the method section. It should be noted that those skilled in the art can make several improvements and modifications to this application without departing from the principles of this application, and these improvements and modifications also fall within the protection scope of the claims of this application.

[0178] It should be understood that in this application, "at least one (item)" means one or more, and "more than" means two or more. "And / or" is used to describe the relationship between related objects, indicating that three relationships can exist. For example, "A and / or B" can represent three cases: only A exists, only B exists, and both A and B exist simultaneously, where A and B can be singular or plural. The character " / " generally indicates that the preceding and following related objects are in an "or" relationship. "At least one (item) of the following" or similar expressions refer to any combination of these items, including any combination of single or plural items. For example, at least one (item) of a, b, or c can represent: a, b, c, "a and b", "a and c", "b and c", or "a and b and c", where a, b, and c can be single or multiple.

[0179] It should also be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

Claims

1. A product selection decision-making method based on dynamic consumer trend prediction, characterized in that, The method includes: The system acquires multiple target users by obtaining the full-domain interactive users of the target brand, and obtains a multimodal behavior dataset for each target user within a target time period. The multimodal behavior dataset includes at least one behavior record. Each behavior record includes inventory quantity unit (SKU) attributes, search keywords, customer service dialogue records, product reviews, or return reasons generated based on different user interaction behaviors. A user profile database is constructed based on the multimodal behavior dataset corresponding to each target user. The user profile database includes multiple user semantic vectors used to describe the consumption characteristics and product demand characteristics of each target user. Each user semantic vector corresponds to a behavior record. The user semantic vector is marked with a time stamp, which is the generation time of the behavior record corresponding to the user semantic vector. Multiple target semantic vectors are obtained by acquiring user semantic vectors whose timestamps are within a time sliding window. Then, a density-based clustering algorithm is used to perform temporal dynamic semantic clustering analysis on the multiple target semantic vectors to obtain multiple current semantic clusters. Identify the number of target users corresponding to each current semantic cluster, and calculate the trend score for each current semantic cluster based on the trend scoring formula and the number of target users corresponding to each current semantic cluster; the trend score is used to evaluate the development potential and explosive growth of the consumption trend corresponding to the current semantic cluster. The target clusters are selected based on the trend scores of each current semantic cluster, and a product selection report is generated by combining the similarity between the candidate SKU and the target clusters. The trend scoring formula is as follows: Strend=α×log(Vcurrent)+β×(Vcurrent-Vpast) / Vpast; Where Strend is the trend score; α is the size weight; Vcurrent is the number of users in the current semantic cluster; β is the growth weight; and Vpast is the number of users in the most recently obtained historical semantic cluster corresponding to the current semantic cluster. The process of filtering target clusters based on the trend scores of each current semantic cluster and generating a product selection report by combining the similarity between the candidate SKU and the target cluster includes: The current semantic clusters are filtered by combining the filtering rules and the trend scores of each current semantic cluster to obtain at least one target cluster; Obtain the product semantic vectors of each candidate SKU of the target brand, and calculate the similarity between each product semantic vector and the cluster center of the target cluster. The product selection report is generated based on the similarity. One similarity score corresponds to one candidate SKU; The process of generating the product selection report based on the similarity includes: If the similarity is greater than or equal to the first similarity threshold, a production recommendation suggestion is generated for the candidate SKU corresponding to the similarity; the production recommendation suggestion is used to suggest that the candidate SKU be put into production directly and to suggest that the initial stock quantity be increased. If the similarity is less than the first similarity threshold and greater than or equal to the second similarity threshold, optimization and adjustment suggestions are generated; the optimization and adjustment suggestions are used to suggest that the R&D department modify the design to address the missing pain points, or conduct only small-batch testing. If the similarity is less than the second similarity threshold, a elimination suggestion is generated; the elimination suggestion is used to suggest that the candidate SKU be eliminated directly or not put into production for the time being. The production recommendation, optimization and adjustment recommendation, and elimination recommendation are integrated to generate the product selection report; Wherein, the first similarity threshold is greater than the second similarity threshold.

2. The method according to claim 1, characterized in that, The method employs a density-based clustering algorithm to perform temporal dynamic semantic clustering analysis on the multiple target semantic vectors to obtain multiple current semantic clusters, including: When the clustering execution phase is the first temporal clustering, a density-based clustering algorithm is directly used to cluster the multiple target semantic vectors to obtain multiple current semantic clusters; When the clustering execution stage is non-first-time sequential clustering, the cluster center of each historical semantic cluster obtained most recently is used as the initial center point, and the multiple target semantic vectors are clustered based on each initial center point to obtain multiple current semantic clusters.

3. The method according to claim 1, characterized in that, The construction of the user profile database based on the multimodal behavior dataset corresponding to each target user includes: For each target user, at least one first semantic reasoning prompt word is generated based on at least one behavior record corresponding to the target user and a first preset instruction. The at least one first semantic reasoning prompt word is then input into a large model for analysis to obtain at least one first semantic tag set. Each first semantic reasoning prompt word corresponds to a first semantic tag set. The first semantic tag set is used to describe the user's consumption characteristics and product demand characteristics. Each first semantic tag set includes at least the user's consumption intention, user's purchase scenario, user's consumption style, and product pain points. Vectorize the first semantic tag set corresponding to each target user to obtain at least one user semantic vector; All user semantic vectors are integrated, stored, and time-annotated to obtain the user profile library.

4. The method according to claim 3, characterized in that, The method further includes: When the target user corresponds to at least two first semantic tag sets, the first semantic tag sets with at least N duplicate tags are merged to obtain a merged tag set, and the original duplicate first semantic tag sets are deleted; N is a positive integer; The step of vectorizing the first semantic tag set corresponding to each of the target users to obtain at least one user semantic vector includes: Vectorize the first semantic tag set and / or the merged tag set corresponding to each target user to obtain at least one user semantic vector.

5. The method according to claim 1, characterized in that, The step of obtaining the product semantic vectors of each candidate SKU of the target brand includes: Obtain the core attributes of each candidate SKU and convert them into text descriptions to obtain structured product descriptions; the core attributes are used to describe the style, pattern, material, function, and applicable scenarios of the candidate SKUs; For each structured product description, a second semantic reasoning prompt is generated based on the structured product description and the second preset instruction. The second semantic reasoning prompt is then input into a large model for analysis to obtain a second semantic tag set. The second semantic tag set is used to describe the core product features of the candidate SKU. Each second semantic tag set includes at least style, pattern, material, function, and applicable scenarios. Each second semantic tag set is vectorized to obtain multiple product semantic vectors.

6. A product selection decision-making device based on dynamic consumer trend prediction, characterized in that, The device includes: The first acquisition unit is used to acquire multiple target users from the target brand's full-domain interactive users, and to acquire a multimodal behavior dataset for each target user within a target time period; the multimodal behavior dataset includes at least one behavior record; each behavior record includes SKU attributes, search keywords, and customer service dialogue records generated based on different user interaction behaviors, as well as product reviews or return reasons; The user profile database construction unit is used to construct a user profile database based on the multimodal behavior dataset corresponding to each target user. The user profile database includes multiple user semantic vectors that describe the consumption characteristics and product demand characteristics of each target user. Each user semantic vector corresponds to a behavior record. The user semantic vector is marked with a time stamp, which is the generation time of the behavior record corresponding to the user semantic vector. The second acquisition unit is used to acquire user semantic vectors whose time markers are within a time sliding window to obtain multiple target semantic vectors; Clustering unit, used to perform temporal dynamic semantic clustering analysis on the multiple target semantic vectors using a density-based clustering algorithm to obtain multiple current semantic clusters; The user count identification unit is used to identify the number of users of the target user corresponding to each current semantic cluster; The scoring unit is used to calculate the trend score of each current semantic cluster based on the trend scoring formula and the number of target users corresponding to each current semantic cluster; the trend score is used to evaluate the development potential and explosive growth of the consumption trend corresponding to the current semantic cluster. The cluster filtering unit is used to filter target clusters based on the trend scores of each current semantic cluster; The first report generation unit is used to generate a product selection report by combining the similarity between the candidate SKUs and the target cluster. The trend scoring formula is as follows: Strend=α×log(Vcurrent)+β×(Vcurrent-Vpast) / Vpast; Where Strend is the trend score; α is the size weight; Vcurrent is the number of users in the current semantic cluster; β is the growth weight; and Vpast is the number of users in the most recently obtained historical semantic cluster corresponding to the current semantic cluster. The cluster screening unit 507 is specifically used for: The current semantic clusters are filtered by combining the filtering rules and the trend scores of each current semantic cluster to obtain at least one target cluster; The first report generation unit 508 specifically includes: The third acquisition unit is used to acquire the product semantic vectors of each candidate SKU of the target brand. The similarity calculation unit is used to calculate the similarity between the semantic vector of each product and the cluster center of the target cluster. The second report generation unit is used to generate the product selection report based on the similarity. One similarity score corresponds to one candidate SKU; The second report generation unit is specifically used for: If the similarity is greater than or equal to the first similarity threshold, a production recommendation suggestion is generated for the candidate SKU corresponding to the similarity; the production recommendation suggestion is used to suggest that the candidate SKU be put into production directly and to suggest that the initial stock quantity be increased. If the similarity is less than the first similarity threshold and greater than or equal to the second similarity threshold, optimization and adjustment suggestions are generated; the optimization and adjustment suggestions are used to suggest that the R&D department modify the design to address the missing pain points, or conduct only small-batch testing. If the similarity is less than the second similarity threshold, a elimination suggestion is generated; the elimination suggestion is used to suggest that the candidate SKU be eliminated directly or not put into production for the time being. The production recommendation, optimization and adjustment recommendation, and elimination recommendation are integrated to generate the product selection report; Wherein, the first similarity threshold is greater than the second similarity threshold.

7. A product selection decision-making device based on dynamic consumer trend prediction, characterized in that, include: A memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the computer program, implements the product selection decision-making method based on dynamic consumer trend prediction as described in any one of claims 1-5.

Images