A multi-source data-based sales strategy intelligent optimization system and method

Abstract

This invention discloses an intelligent sales strategy optimization system and method based on multi-source data, belonging to the field of sales strategy optimization technology. The invention acquires and preprocesses outfit data from mainstream video platforms, extracts visual style features using computer vision algorithms, establishes association mappings by combining outfit text style keywords, constructs a style feature system, and categorizes outfits to form popular style clusters. It then digitally models in-store clothing, aligns with the style feature system dimensions to build an outfit library, and uses quantitative features of popular clusters as conditions to retrieve and filter outfits with matching scores through similarity algorithms, generating an associated list and targeting it across multiple channels. The invention also statistically analyzes outfit clicks and purchases during the promotion period, calculates the click-to-purchase ratio, constructs an outfit association topology, and analyzes its features; finally, it dynamically adjusts inventory configuration based on the click-to-purchase ratio, topology features, and existing inventory.

Description

Technical Field

[0001] This invention relates to the field of sales strategy optimization technology, specifically a sales strategy intelligent optimization system and method based on multi-source data. Background Technology

[0002] With the rapid development of the apparel e-commerce industry and the widespread adoption of short video platforms, consumer fashion preferences are significantly influenced by popular online trends. The apparel sales market is characterized by rapid trend iteration, personalized demand, and a high demand for precise marketing. Under these trends, optimizing sales strategies and managing inventory have become core competitive advantages for apparel companies.

[0003] Existing methods for optimizing apparel sales strategies have several limitations: First, trend capture relies on human experience, often using subjective judgment to select popular styles, lacking multi-source data support, and failing to accurately and in real-time uncover trends on video platforms, leading to a disconnect between promoted apparel and market demand. Second, style analysis is limited in scope, relying solely on visual feature recognition or extracting only text keywords, without establishing a mapping between visual and textual features, resulting in low style matching accuracy and poor selection of related apparel. Third, inventory adjustment mechanisms are rigid, often relying on historical sales data to formulate replenishment or clearance strategies, without considering dynamic data such as post-promotion click-through rates, apparel combinations, and cross-selling relationships, easily leading to stockouts of best-selling items and stockpiling of slow-moving items. Summary of the Invention

[0004] The purpose of this invention is to provide a sales strategy intelligent optimization system and method based on multi-source data to solve the problems raised in the prior art.

[0005] To achieve the above objectives, the present invention provides the following technical solution: Firstly, this application provides a sales strategy intelligent optimization method based on multi-source data, comprising the following steps: Obtain fashion content data from mainstream video platforms and preprocess it; Computer vision algorithms are used to analyze clothing content data and extract visual style features; style keywords are extracted from clothing content data and associated with visual style features to establish a style feature system; based on the style feature system, clothing is classified to form popular style clusters; Digital modeling is performed on all clothing in the store, and a unified dimension is established with the style feature system to form a clothing library in the store. Using the quantitative features of each popular style cluster as search conditions, the clothing feature library in the store is searched through similarity algorithm to select clothing with a style matching degree of a set threshold that meets the popular style features, thus forming a list of related clothing for each style. Targeted advertising across multiple channels based on a list of related clothing items; calculate the click-through rate (CTR) and purchase volume of each related clothing item during the promotion period; construct a clothing association topology and analyze the characteristics of the topology relationships; dynamically adjust inventory based on the CTR, topology relationship characteristics, and existing inventory levels.

[0006] In conjunction with the first aspect, in the first embodiment of the first aspect of this application, the step of obtaining and preprocessing the fashion content data from mainstream video platforms includes: The data on clothing content includes visual data, text data, and interactive data. A fixed data collection window of approximately one calendar month is used. For platforms providing open data interfaces, clothing content data is collected in batches after applying for the corresponding permissions based on the interface call specifications. For platforms without open interfaces or with incomplete interface data coverage, compliant web crawling technology is used for data collection. The crawler program follows the platform's robots protocol, limits the crawling frequency, and only crawls publicly accessible clothing content data. The clothing content data undergoes data cleaning and data format standardization.

[0007] In conjunction with the first aspect, in the second embodiment of the first aspect of this application, the step of using computer vision algorithms to analyze clothing content data and extract visual style features includes: For the visual data in the clothing content data, the inter-frame difference method is used to select the keyframes that display the most complete clothing display, are unobstructed, and have clear poses; the extracted keyframes are cropped to retain the clothing area and remove redundant background elements; the cropped images are then enhanced. For the inter-frame difference method, specifically, the standardized outfit video data from the preprocessing stage is retrieved to unify video resolution and frame rate, eliminating differences in video formats across different platforms. Noise reduction processing is applied to the video to filter out image noise and compression distortion. All image frames of the outfit video are extracted sequentially to generate an ordered frame sequence. A reasonable frame extraction interval is set based on the video duration to avoid redundant calculations while preserving complete information about the outfit display process. Adjacent frames in the frame sequence are compared pixel-by-pixel to identify the areas and degrees of change between frames. Focusing on the core areas of the clothing display, minor background changes and subtle movements of the figures are ignored to determine whether the frames are in a stable outfit display state.

[0008] A frame difference threshold is set, and frames with differences below the threshold are selected as candidate keyframes. These frames correspond to segments in the video where the character's posture is stable, the clothing has no obvious displacement, and the core display area remains unchanged. Transitional frames with differences above the threshold are removed. The candidate keyframes are deduplicated, and consecutively similar stable frames are merged. Representative frames from each stable clothing display segment are selected according to the video timeline, and the keyframes with the most complete clothing display, no obstructions, and clear posture are selected.

[0009] Based on the enhanced image, basic visual features are extracted; a feature concatenation algorithm is used to fuse the extracted basic visual features into a unified high-dimensional visual feature vector, and feature redundancy removal and dimensionality reduction are performed to obtain visual style features.

[0010] In conjunction with the first aspect, in a third embodiment of the first aspect of this application, the extraction of basic visual features based on the enhanced image includes: A convolutional neural network (CNN) model is used to perform preliminary analysis on the enhanced image. The Canny algorithm is used to extract the overall outline of the clothing, and morphological operations are combined to optimize the outline boundary to obtain outline features. Based on the outline features, the basic pattern of the clothing is determined, and the pattern features are output. The image is converted from RGB space to HSV space using a color space conversion algorithm. The K-Means clustering algorithm is used to cluster the image pixels by color, select the dominant color and its corresponding proportion, and mark the color distribution position to form a color feature set. High-order features of the image are extracted based on a residual network. The YOLO algorithm is used to locate and identify decorative elements on the clothing. Each decorative element is labeled with features to form a detail feature set. The texture analysis algorithm is used to extract the surface texture features of the clothing. For images containing multiple garments, a target segmentation algorithm is used to segment different categories of clothing into regions. Based on the segmentation results, the matching relationship between the garments is analyzed to extract matching combination features.

[0011] In conjunction with the first aspect, in the fourth embodiment of the first aspect of this application, the step of extracting style keywords from clothing content data and associating them with visual style features to establish a style feature system includes: Text data is extracted from the clothing content data. The text data is segmented using the jieba word segmentation algorithm to remove meaningless stop words. The TF-IDF algorithm is used to calculate the word weights and filter out words that are strongly related to the clothing style to form style keywords. The mapping rules are constructed as follows: for each style keyword, the corresponding visual style feature dimension and feature value are matched; for each combination of visual style features, the corresponding style keyword is matched; the correlation between style keywords and visual feature combinations is calculated using Euclidean distance to quantify the mapping confidence; for mapping relationships with a confidence level lower than a set threshold, the mapping is corrected by combining interactive data from the outfit content data, invalid mappings are eliminated, and a mapping relationship table is constructed; all mapping relationship tables are integrated to form an associated mapping relationship library. Based on the association mapping database, a style feature system framework is constructed, in which the first-level dimension is visual style features and style keywords, and the second-level dimension is specific feature values; the analytic hierarchy process is used to assign values ​​to the features at each level in the system, and the weights of visual style features and style keywords are adjusted based on the mapping confidence.

[0012] In conjunction with the first aspect, in the fifth embodiment of the first aspect of this application, the step of classifying clothing based on a style feature system to form a popular style cluster includes: Based on the style feature system, visual style features and style keywords of each outfit are extracted and fused into a comprehensive feature vector through feature splicing. After standardization to eliminate magnitude differences, PCA is used to reduce dimensionality and remove redundancy to form a core dataset. The K-Means algorithm is selected, and the parameters are initialized in combination with the core dataset, and the iteration termination condition is set. Based on the feature vector similarity calculation, outfits with similar styles are grouped into one category, and iterative optimization is performed according to the termination condition until the cluster center is stable, forming a popular style cluster.

[0013] In conjunction with the first aspect, in the sixth embodiment of the first aspect of this application, the step of digitally modeling all clothing in the store and establishing a unified dimension with the style feature system to form an in-store clothing library includes: Collect visual style features and style keywords of all clothing in the store, transform them into structured data, and form a feature dataset of clothing; calibrate the feature dataset of clothing against the style feature system, correct any deviations, and achieve complete alignment between clothing features and the system; integrate all feature datasets of clothing to establish an in-store clothing database with a search index.

[0014] In conjunction with the first aspect, in the seventh embodiment of the first aspect of this application, the construction of the clothing association topology and the analysis of topological relationship characteristics include: Using promoted clothing and related clothing in the store as nodes, the conversion rate and style matching degree of the node attributes are labeled; using matching relationship, cross-selling relationship, and style relationship as edges, the weight of the edges is quantified; based on the definition of nodes and edges, graph theory algorithm is used to build an undirected weighted topology graph, integrate the related data between clothing, and form a clothing related topology structure; calculate the node degree, clustering coefficient and centrality, locate key nodes and key related edges and combine them as topological relationship features.

[0015] In conjunction with the first aspect, in the eighth embodiment of the first aspect of this application, the dynamic adjustment of inventory based on click-to-buy ratio, topological relationship characteristics, and existing inventory levels includes: Using a weighting algorithm, corresponding weights are assigned to click-to-buy ratio, topological key node influence, and edge association strength based on business scenarios. Based on the weight results, a threshold determination algorithm is used, combined with the inventory safety line and inventory turnover rate, to divide the critical ranges for replenishment, reduction of stock, maintenance of inventory, and optimization of allocation, forming quantitative adjustment rules. Inventory is dynamically adjusted based on the quantitative adjustment rules.

[0016] Specifically, the priorities of indicators are determined based on the current business scenario. For example, during the peak season for apparel sales, the focus is on conversion efficiency, which can increase the weight of click-to-buy ratio; during the new product promotion period, the focus is on cross-selling, which can increase the weight of edge association strength; and during the inventory clearance period, the focus is on core traffic-driving products, which can increase the weight of the influence of key nodes in the topology. Using a weight allocation algorithm, corresponding weights are assigned to the three indicators according to their priorities, so that the sum of the weights is 1.

[0017] Historical sales data is selected for backtesting. The assigned weights are substituted into historical scenarios, and the consistency between the weight calculation results and actual sales performance is compared. If discrepancies exist, the weight ratios of each indicator are adjusted based on business experience. The verification process is repeated until the weight results accurately reflect the actual impact of the indicators on inventory adjustments, thus forming a weight allocation scheme. The inventory safety line is the minimum inventory level set to avoid stockouts, and inventory turnover rate is a key indicator for measuring inventory turnover efficiency. The weighted scores of each garment, real-time inventory data, inventory safety line threshold, and historical inventory turnover rate data are integrated to form the basic dataset for threshold determination.

[0018] Based on weighted scores and combined with inventory safety line and inventory turnover rate, a threshold determination algorithm divides the inventory into four critical intervals: First, the interval with high weighted score, inventory below the safety line, and high turnover rate corresponds to replenishment operations; second, the interval with low weighted score, inventory above the safety line, and low turnover rate corresponds to reducing replenishment operations; third, the interval with medium weighted score, inventory within the safety line range, and stable turnover rate corresponds to maintaining inventory operations; and fourth, the interval with medium weighted score, strong related sales, and unbalanced combined inventory ratio corresponds to optimizing the ratio.

[0019] For the four critical intervals, clear quantitative adjustment rules were formulated: the replenishment rule clearly states that the basis for determining the replenishment volume is the market demand potential corresponding to the weighted score; the reduction rule clearly states that the reference standard for the reduction ratio is the degree of inverse correlation between inventory turnover rate and weighted score; the inventory maintenance rule clearly states the inventory monitoring frequency and fluctuation tolerance range; and the optimized allocation rule clearly states the inventory ratio adjustment ratio of each individual product in the combination to ensure the supply and demand balance of the combination.

[0020] Real-time data for each garment item is input into a weighting scheme and critical interval rules to match the corresponding inventory operation type. For core garments meeting replenishment criteria, a replenishment process is initiated; for slow-moving garments meeting reduced stock criteria, a clearance sale or reduced order process is initiated; and for combination garments meeting optimization criteria, the inventory percentage of each item is adjusted, achieving dynamic inventory adjustment.

[0021] Secondly, this application provides a sales strategy intelligent optimization system based on multi-source data, including: Outfit Data Acquisition Module: Includes: Outfit data acquisition unit acquires outfit content data from mainstream video platforms and performs preprocessing; Popular Style Cluster Construction Module: Includes: Visual Style Feature Extraction Unit, which uses computer vision algorithms to analyze clothing content data and extract visual style features; Style Feature System Establishment Unit, which extracts style keywords from clothing content data and associates them with visual style features to establish a style feature system; Popular Style Cluster Construction Unit, which classifies clothing based on the style feature system to form popular style clusters. The associated clothing list generation module includes: an in-store clothing library construction unit that digitally models all clothing in the store and establishes a unified dimension with the style feature system to form an in-store clothing library; and an associated clothing list generation unit that uses the quantitative features of each popular style cluster as search conditions, searches the in-store clothing feature library through a similarity algorithm, and selects clothing with a style matching degree of up to a set threshold with the popular style features to form an associated clothing list corresponding to each style. The dynamic inventory adjustment module includes: a targeted delivery unit that performs multi-channel targeted delivery based on a list of related clothing items; a click-to-buy ratio calculation unit that calculates the click-to-buy ratio by statistically analyzing the click-to-buy volume and purchase volume of each related clothing item during the promotion period; a topology relationship feature analysis unit that constructs the clothing association topology structure and analyzes the topology relationship features; and a dynamic inventory adjustment unit that dynamically adjusts the inventory based on the click-to-buy ratio, topology relationship features, and the existing inventory level.

[0022] Compared with the prior art, the beneficial effects of the present invention are: 1. This invention obtains multi-source fashion data from mainstream video platforms, and after preprocessing, it uses technical means to mine trends, rather than relying on subjective human judgment. With the support of multi-source data and algorithm analysis, it can accurately capture real-time popular fashion trends on video platforms, so that the promoted clothing is highly consistent with market demand, and completely solves the problems of lagging and large deviations in the capture of trends in existing technologies.

[0023] 2. This invention constructs a two-dimensional analysis system of visual features and text keywords. It extracts visual style features through computer vision algorithms and mines style keywords, establishes a correlation mapping between the two to form a complete style feature system, and then forms popular clusters based on the system classification. Compared with single-dimensional analysis, it greatly improves the accuracy of style matching and ensures more accurate and better results in the screening of related clothing.

[0024] 3. This invention breaks through the rigid model of relying on historical data, and combines the click-through rate after promotion, the clothing association topology, and the existing inventory level to drive inventory adjustment through dynamic data; it can accurately predict the demand for best-selling items and replenish them in time, avoid the backlog of slow-moving items, and achieve optimal allocation of inventory resources. Attached Figure Description

[0025] Figure 1 This is a schematic diagram illustrating the steps of an intelligent optimization method for sales strategies based on multi-source data according to the present invention. Figure 2 This is a system architecture diagram of a sales strategy intelligent optimization system based on multi-source data according to the present invention. Detailed Implementation

[0026] 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 the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0027] Example: Figures 1-2 As shown, the present invention provides a technical solution: like Figure 1 As shown, this application provides a sales strategy intelligent optimization method based on multi-source data, including the following steps: Step S100: Obtain fashion content data from mainstream video platforms and perform preprocessing; Specifically, the outfit content data includes visual data, text data, and interactive data. A fixed data collection window of approximately one calendar month is used. For platforms providing open data interfaces, outfit content data is collected in batches after applying for the corresponding permissions based on the interface call specifications. For platforms without open interfaces or with incomplete interface data coverage, compliant web crawling technology is used for data collection. The crawler program follows the platform's robots protocol, limits the crawling frequency, and only crawls publicly accessible outfit content data. The outfit content data undergoes data cleaning and data format standardization.

[0028] In one specific embodiment, this embodiment selects three major mainstream video platforms, namely, TikTok, XX Book, and X Station, and uses January 1 to January 31, 2026 as a fixed collection time window. The target is to collect visual data (short videos, cover images), text data (titles, copy, hashtags), and interaction data (play counts, likes, favorites, comments) from the clothing content data for subsequent style analysis and strategy optimization.

[0029] For platforms like TikTok and XX Books, which offer open data interfaces, permission applications were submitted according to the interface call specifications and approved before data collection of outfit content was completed in batches. TikTok collected 52,000 short videos and 52,000 corresponding cover images, while XX Books collected 38,000 outfit notes and 21,000 sets of text and image content. Text information and interaction data for each piece of content were captured simultaneously. The collection process took 48 hours, and the data integrity reached 98%.

[0030] For X site, which does not have an open interface, compliant crawling technology was used to crawl publicly available fashion content. The crawling program strictly followed the X site's robots protocol, limiting the crawling frequency to ≤100 videos per hour to avoid putting pressure on the platform's servers. Only publicly posted fashion videos were crawled, and no private data was collected. In the end, 15,000 valid videos were collected, and the completeness of the accompanying text and interactive data reached 95%.

[0031] A total of 105,000 pieces of clothing content data were cleaned. Duplicate data of 8,200 pieces were removed by comparing hash values. Low-quality content of 3,500 pieces with blurry images and no clear clothing display was filtered out, as well as illegal content of 2,300 pieces involving vulgarity and infringement. After cleaning, 91,000 pieces of valid data were retained.

[0032] Finally, the data format was standardized. Visual data was uniformly adjusted to 1080P resolution, MP4 video format, and PNG image format; text data was converted to UTF-8 encoding, and word segmentation and stop word removal preprocessing were completed; interactive data was normalized, and metrics such as play count and like count were uniformly mapped to the 0-1 range to form a structured dataset for subsequent algorithm analysis.

[0033] Step S200: Use computer vision algorithms to analyze the clothing content data and extract visual style features; extract style keywords from the clothing content data and associate them with visual style features to establish a style feature system; classify the clothing based on the style feature system to form popular style clusters; Specifically, for the visual data in the clothing content data, the inter-frame difference method is used to select the keyframes that display the most complete clothing display, are unobstructed, and have clear poses; the extracted keyframes are cropped to retain the clothing area and remove redundant background elements; the cropped images are then enhanced. Based on the enhanced image, basic visual features are extracted; a feature concatenation algorithm is used to fuse the extracted basic visual features into a unified high-dimensional visual feature vector, and feature redundancy removal and dimensionality reduction are performed to obtain visual style features.

[0034] Furthermore, the extraction of basic visual features based on the enhanced image includes: A convolutional neural network (CNN) model is used to perform preliminary analysis on the enhanced image. The Canny algorithm is used to extract the overall outline of the clothing, and morphological operations are combined to optimize the outline boundary to obtain outline features. Based on the outline features, the basic pattern of the clothing is determined, and the pattern features are output. The image is converted from RGB space to HSV space using a color space conversion algorithm. The K-Means clustering algorithm is used to cluster the image pixels by color, select the dominant color and its corresponding proportion, and mark the color distribution position to form a color feature set. High-order features of the image are extracted based on a residual network. The YOLO algorithm is used to locate and identify decorative elements on the clothing. Each decorative element is labeled with features to form a detail feature set. The texture analysis algorithm is used to extract the surface texture features of the clothing. For images containing multiple garments, a target segmentation algorithm is used to segment different categories of clothing into regions. Based on the segmentation results, the matching relationship between the garments is analyzed to extract matching combination features.

[0035] Furthermore, the text data in the clothing content data is extracted, and the text data is segmented by the jieba word segmentation algorithm to remove meaningless stop words. The TF-IDF algorithm is used to calculate the word weights and filter out words that are strongly related to the clothing style to form style keywords. The mapping rules are constructed as follows: for each style keyword, the corresponding visual style feature dimension and feature value are matched; for each combination of visual style features, the corresponding style keyword is matched; the correlation between style keywords and visual feature combinations is calculated using Euclidean distance to quantify the mapping confidence; for mapping relationships with a confidence level lower than a set threshold, the mapping is corrected by combining interactive data from the outfit content data, invalid mappings are eliminated, and a mapping relationship table is constructed; all mapping relationship tables are integrated to form an associated mapping relationship library. Based on the association mapping database, a style feature system framework is constructed, in which the first-level dimension is visual style features and style keywords, and the second-level dimension is specific feature values; the analytic hierarchy process is used to assign values ​​to the features at each level in the system, and the weights of visual style features and style keywords are adjusted based on the mapping confidence.

[0036] Furthermore, based on the style feature system, visual style features and style keywords for each outfit are extracted and fused into a comprehensive feature vector through feature splicing. After standardization to eliminate magnitude differences, PCA is used to reduce dimensionality and remove redundancy to form a core dataset. The K-Means algorithm is selected, and the parameters are initialized in conjunction with the core dataset, and the iteration termination condition is set. Based on the feature vector similarity calculation, outfits with similar styles are grouped into one category, and iterative optimization is performed according to the termination condition until the cluster center is stable, forming a popular style cluster.

[0037] In one specific embodiment, this embodiment focuses on processing visual and text data based on the 91,000 structured outfit data obtained in step S100, and gradually completes the extraction of visual style features, association mapping, system construction and cluster formation.

[0038] For 68,000 outfit video data, the inter-frame difference method was used to filter key frames. The inter-frame difference threshold was set to 0.3. Transitional frames and blurry frames were removed from the total of 724,000 frames, resulting in 182,000 key frames that completely displayed the outfits. The key frames were then intelligently cropped to remove redundant background elements and retain the core display area of ​​the clothing. The images were then enhanced through illumination equalization and noise reduction. The image clarity of the processed images was improved by 35% compared to the original frames, laying the foundation for feature extraction.

[0039] Based on the enhanced 182,000 frames of images, basic visual features were extracted: contour features were extracted using CNN and Canny algorithms to identify three types of garment styles: loose, slim, and fitted, with loose styles accounting for 42% and slim styles for 38%; after conversion to HSV space, K-Means clustering was used to select 3-5 main color tones, with earth tones, black, and white being the most frequent main colors, accounting for 32%, 28%, and 19%, respectively; decorative elements were identified using ResNet+YOLO algorithms, with floral, plaid, and embroidery having the highest proportions, at 25%, 18%, and 12%, respectively; materials were identified using LBP texture algorithms, with cotton accounting for 45% and knitwear for 23%; and multiple garments were segmented using MaskR-CNN, with layering and matching accounting for 23%, clarifying the matching relationships between individual items.

[0040] 89,000 text data were extracted, segmented by jieba, and stop words were removed. The TF-IDF algorithm was used to calculate the word weights and select 210 core style keywords. Among them, "commuter style", "sweet and cool style" and "retro style" ranked in the top three with weights of 0.82, 0.79 and 0.76 respectively.

[0041] After constructing the mapping rules, Euclidean distance was used to quantify the association confidence. A confidence threshold of 0.7 was set, and 320 sets of invalid mappings with low confidence were removed, resulting in a final table of 1280 valid mapping relationships with an average confidence of 0.85. The mapping was then corrected by combining interaction data, and 150 sets of fuzzy mappings for the top 10% of content in terms of likes were calibrated to improve mapping accuracy.

[0042] A style feature system was constructed, with the first-level dimension consisting of visual style features (including 5 sub-dimensions) and style keywords (210), and the second-level dimension covering 136 specific feature values. The weights were assigned using the analytic hierarchy process, with visual features having a weight of 0.55 and keywords having a weight of 0.45. The weights were adjusted based on confidence levels, with the weight of the visual features corresponding to "retro style" being increased by an additional 5%.

[0043] Visual and keyword features were integrated into a comprehensive vector, and after standardization, the dimensionality was reduced by PCA (from 256 dimensions to 64 dimensions) to form a core dataset of 87,000 entries. The K-Means algorithm was selected, with K=8 initialized and 25 iterations set as the termination condition. Finally, 8 popular style clusters were formed, of which the "retro earth tone commuter cluster" accounted for 22% and the "sweet and cool floral girl cluster" accounted for 18%, with the cluster center stability reaching 92%.

[0044] Step S300: Digitally model all clothing in the store, establish a unified dimension with the style feature system, and form a clothing library in the store; use the quantitative features of each popular style cluster as search conditions, search the clothing feature library in the store through similarity algorithm, and filter out clothing with a style matching degree of up to a set threshold with the popular style features to form a list of related clothing for each style. Specifically, the visual style features and style keywords of all clothing in the store are collected and transformed into structured data to form a feature dataset of clothing; the feature dataset of clothing is calibrated against the style feature system, and deviation data is corrected to achieve complete alignment between clothing features and the system; the feature datasets of all clothing are integrated to establish an in-store clothing database with a search index.

[0045] In one specific embodiment, this embodiment conducts digital modeling on 3,000 garments from offline and online stores, benchmarking the style feature system constructed in step S200 throughout the process to ensure dimensional consistency, ultimately forming a standardized in-store garment database and completing associated retrieval.

[0046] Visual style features and style keywords were collected for each garment. The visual features covered five sub-dimensions, including cut, color, material, and decorative elements. The text keywords were associated with the 210 core style words selected earlier. The collected information was transformed into a structured feature dataset, which included 840 loose-fitting garments, 660 earth-toned garments, and 1350 cotton garments. There were 375 garments with floral prints and 288 garments with plaid patterns.

[0047] The 3,000 clothing feature datasets were calibrated against the style feature system, correcting 86 data entries with issues such as pattern annotation deviation and color proportion statistical deviation, and supplementing 42 missing decorative element feature annotations, ensuring that all clothing features are fully aligned with the system's 136 secondary feature values, with a feature alignment accuracy of 99.2%.

[0048] Integrate all calibrated clothing feature data to establish a structured in-store clothing database, equipped with a multi-dimensional search index. The index directory is constructed by classifying and building according to style keywords, visual feature dimensions, and specific feature values. Each piece of clothing data corresponds to a unique search identifier and contains 28 feature fields. It supports fast retrieval by popular cluster quantitative features, and the database data storage efficiency is improved by 40% compared to ordinary documents.

[0049] Step S400: Conduct multi-channel targeted advertising based on the associated clothing list; count the clicks and purchases of each associated clothing item during the promotion period, and calculate the click-to-purchase ratio; construct the clothing association topology and analyze the topology relationship characteristics; dynamically adjust the inventory based on the click-to-purchase ratio, topology relationship characteristics, and existing inventory levels.

[0050] Specifically, nodes are the promoted clothing and related clothing in the store, with node attributes labeled with conversion rate and style matching degree; edges are the matching relationship, related sales relationship, and style relationship, with the weight of the edges quantified; based on the definition of nodes and edges, graph theory algorithms are used to build an undirected weighted topology graph, integrate the related data between clothing, and form a clothing related topology structure; node degree, clustering coefficient and centrality are calculated, key nodes and key related edges are located and combined as topological relationship features.

[0051] Furthermore, a weighting algorithm is used to assign corresponding weights to click-to-buy ratio, topological key node influence, and edge association strength based on business scenarios. Based on the weighting results, a threshold determination algorithm is used to divide the critical ranges for replenishment, reduction of stock, maintenance of inventory, and optimization of allocation by combining inventory safety line and inventory turnover rate, forming quantitative adjustment rules. Inventory is dynamically adjusted based on the quantitative adjustment rules.

[0052] In one specific embodiment, this embodiment selects 240 core garments based on the 8 style-related clothing lists generated by S300 for multi-channel targeted advertising. The advertising period is set from February 1 to February 28, 2026, covering three major channels: Taobao information flow, XX book recommendations, and Taobao direct traffic. The advertising proportions of each channel are 40%, 35%, and 25%, respectively. The advertising content is linked to the optimized display of the characteristics of the corresponding popular style clusters.

[0053] After the promotion period ended, the data for each related clothing item was analyzed: the total number of clicks for the 240 clothing items reached 186,000, and the total number of purchases was 1,674, with an overall click-to-purchase ratio of 9.0%. Among them, the "Retro Earth Tone Commuter Cluster" clothing had the highest click-to-purchase ratio at 12.3%; the "Sweet and Cool Floral Girl Cluster" was second at 10.5%; the click-to-purchase ratios for low-popularity cluster clothing were generally below 6%, showing significant differences.

[0054] A clothing association topology was constructed, with 240 featured clothing items and 480 related in-store clothing items as nodes, totaling 720 nodes. Node attributes were labeled with corresponding conversion rates (0.03-0.12) and style matching degrees (0.75-0.92). Three types of relationships were used as edges: matching relationships (0.3-0.8), related sales relationships (0.2-0.7), and style association relationships (0.4-0.9). An undirected weighted topology graph was constructed using graph theory algorithms. The calculated node degree ranged from 2-15, the clustering coefficient from 0.18-0.65, and the top 30 centrality nodes corresponded to core traffic-driving items. Finally, 35 key nodes and 58 key association edges were identified as topology relationship features.

[0055] Based on the peak sales season business scenario, a weighted allocation algorithm was used to assign values: click-through rate to purchase 0.45, topological key node influence 0.35, and edge association strength 0.2. An inventory safety line was set (50 pieces / item for core items, 30 pieces / item for general items), and an inventory turnover rate threshold of 1.2. A threshold determination algorithm was used to define critical intervals and form quantitative rules. Inventory was adjusted based on these rules: replenishment was provided for 12 high-weight key node garments, with a replenishment quantity of 20-35 pieces per item; inventory was reduced by 30%-50% for 8 low-conversion, weakly associated garments; and the allocation of 15 related garment sets was optimized. After the adjustments, the inventory turnover rate increased to 1.5, and the stockout rate decreased from 8.2% to 2.7%.

[0056] like Figure 2 As shown, this application provides a sales strategy intelligent optimization system based on multi-source data, including: Outfit Data Acquisition Module: Includes: Outfit data acquisition unit acquires outfit content data from mainstream video platforms and performs preprocessing; Popular Style Cluster Construction Module: Includes: Visual Style Feature Extraction Unit, which uses computer vision algorithms to analyze clothing content data and extract visual style features; Style Feature System Establishment Unit, which extracts style keywords from clothing content data and associates them with visual style features to establish a style feature system; Popular Style Cluster Construction Unit, which classifies clothing based on the style feature system to form popular style clusters. The associated clothing list generation module includes: an in-store clothing library construction unit that digitally models all clothing in the store and establishes a unified dimension with the style feature system to form an in-store clothing library; and an associated clothing list generation unit that uses the quantitative features of each popular style cluster as search conditions, searches the in-store clothing feature library through a similarity algorithm, and selects clothing with a style matching degree of up to a set threshold with the popular style features to form an associated clothing list corresponding to each style. The dynamic inventory adjustment module includes: a targeted delivery unit that performs multi-channel targeted delivery based on a list of related clothing items; a click-to-buy ratio calculation unit that calculates the click-to-buy ratio by statistically analyzing the click-to-buy volume and purchase volume of each related clothing item during the promotion period; a topology relationship feature analysis unit that constructs the clothing association topology structure and analyzes the topology relationship features; and a dynamic inventory adjustment unit that dynamically adjusts the inventory based on the click-to-buy ratio, topology relationship features, and the existing inventory level.

[0057] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

Claims

1. A sales strategy intelligent optimization method based on multi-source data, characterized in that, Includes the following steps: Obtain fashion content data from mainstream video platforms and preprocess it; Computer vision algorithms are used to analyze clothing content data and extract visual style features; Extract style keywords from outfit content data and map them to visual style features to establish a style feature system; based on the style feature system, classify outfits to form popular style clusters; Digital modeling is performed on all clothing in the store, and a unified dimension is established with the style feature system to form a clothing library in the store. Using the quantitative features of each popular style cluster as search conditions, the clothing feature library in the store is searched through similarity algorithm to select clothing with a style matching degree of a set threshold that meets the popular style features, thus forming a list of related clothing for each style. Targeted advertising across multiple channels based on a list of related clothing items; statistical analysis of clicks and purchases for each related clothing item during the promotion period, calculating the click-to-purchase ratio; construction of a clothing association topology and analysis of topological relationship characteristics; Inventory is dynamically adjusted based on click-through rate, topological relationship characteristics, and existing inventory levels.

2. The intelligent optimization method for sales strategies based on multi-source data according to claim 1, characterized in that, The process of acquiring and preprocessing fashion content data from mainstream video platforms includes: The data on clothing content includes visual data, text data, and interactive data. A fixed data collection window of approximately one calendar month is used. For platforms providing open data interfaces, clothing content data is collected in batches after applying for the corresponding permissions based on the interface call specifications. For platforms without open interfaces or with incomplete interface data coverage, compliant web crawling technology is used for data collection. The crawler program follows the platform's robots protocol, limits the crawling frequency, and only crawls publicly accessible clothing content data. The clothing content data undergoes data cleaning and data format standardization.

3. The intelligent optimization method for sales strategies based on multi-source data according to claim 1, characterized in that, The process involves using computer vision algorithms to analyze clothing content data and extract visual style features, including: For the visual data in the clothing content data, the inter-frame difference method is used to select the keyframes that display the most complete clothing display, are unobstructed, and have clear poses; the extracted keyframes are cropped to retain the clothing area and remove redundant background elements; the cropped images are then enhanced. Based on the enhanced image, basic visual features are extracted; a feature concatenation algorithm is used to fuse the extracted basic visual features into a unified high-dimensional visual feature vector, and feature redundancy removal and dimensionality reduction are performed to obtain visual style features.

4. The intelligent optimization method for sales strategies based on multi-source data according to claim 3, characterized in that, The extraction of basic visual features based on the enhanced image includes: A convolutional neural network (CNN) model is used to perform preliminary analysis on the enhanced image. The Canny algorithm is used to extract the overall outline of the clothing, and morphological operations are combined to optimize the outline boundary to obtain outline features. Based on the outline features, the basic pattern of the clothing is determined, and the pattern features are output. The image is converted from RGB space to HSV space using a color space conversion algorithm. The K-Means clustering algorithm is used to cluster the image pixels by color, select the dominant color and its corresponding proportion, and mark the color distribution position to form a color feature set. High-order features of the image are extracted based on a residual network. The YOLO algorithm is used to locate and identify decorative elements on the clothing. Each decorative element is labeled with features to form a detail feature set. The texture analysis algorithm is used to extract the surface texture features of the clothing. For images containing multiple garments, a target segmentation algorithm is used to segment different categories of clothing into regions. Based on the segmentation results, the matching relationship between the garments is analyzed to extract matching combination features.

5. The intelligent optimization method for sales strategies based on multi-source data according to claim 1, characterized in that, The process of extracting style keywords from outfit content data and mapping them to visual style features to establish a style feature system includes: Text data is extracted from the clothing content data. The text data is segmented using the jieba word segmentation algorithm to remove meaningless stop words. The TF-IDF algorithm is used to calculate the word weights and filter out words that are strongly related to the clothing style to form style keywords. The mapping rules are constructed as follows: for each style keyword, the corresponding visual style feature dimension and feature value are matched; for each combination of visual style features, the corresponding style keyword is matched; the correlation between style keywords and visual feature combinations is calculated using Euclidean distance to quantify the mapping confidence; for mapping relationships with a confidence level lower than a set threshold, the mapping is corrected by combining interactive data from the outfit content data, invalid mappings are eliminated, and a mapping relationship table is constructed; all mapping relationship tables are integrated to form an associated mapping relationship library. Based on the association mapping database, a style feature system framework is constructed, in which the first-level dimension is visual style features and style keywords, and the second-level dimension is specific feature values; the analytic hierarchy process is used to assign values ​​to the features at each level in the system, and the weights of visual style features and style keywords are adjusted based on the mapping confidence.

6. The intelligent optimization method for sales strategies based on multi-source data according to claim 1, characterized in that, The style feature system categorizes outfits to form popular style clusters, including: Based on the style feature system, visual style features and style keywords of each outfit are extracted and fused into a comprehensive feature vector through feature splicing. After standardization to eliminate magnitude differences, PCA is used to reduce dimensionality and remove redundancy to form a core dataset. The K-Means algorithm is selected, and the parameters are initialized in combination with the core dataset, and the iteration termination condition is set. Based on the feature vector similarity calculation, outfits with similar styles are grouped into one category, and iterative optimization is performed according to the termination condition until the cluster center is stable, forming a popular style cluster.

7. The intelligent optimization method for sales strategies based on multi-source data according to claim 1, characterized in that, The process involves digitally modeling all clothing in the store and establishing a unified dimension with the style feature system to form a clothing library for the store, including: Collect visual style features and style keywords of all clothing in the store, transform them into structured data, and form a feature dataset of clothing; calibrate the feature dataset of clothing against the style feature system, correct any deviations, and achieve complete alignment between clothing features and the system; integrate all feature datasets of clothing to establish an in-store clothing database with a search index.

8. The intelligent optimization method for sales strategies based on multi-source data according to claim 1, characterized in that, The construction of the clothing association topology and the analysis of topological relationship characteristics include: Using promoted clothing and related clothing in the store as nodes, the conversion rate and style matching degree of the node attributes are labeled; using matching relationship, cross-selling relationship, and style relationship as edges, the weight of the edges is quantified; based on the definition of nodes and edges, graph theory algorithm is used to build an undirected weighted topology graph, integrate the related data between clothing, and form a clothing related topology structure; calculate the node degree, clustering coefficient and centrality, locate key nodes and key related edges and combine them as topological relationship features.

9. The intelligent optimization method for sales strategies based on multi-source data according to claim 1, characterized in that, The method of dynamically adjusting inventory based on click-through rate, topological relationship characteristics, and existing inventory levels includes: Using a weighting algorithm, corresponding weights are assigned to click-to-buy ratio, topological key node influence, and edge association strength based on business scenarios. Based on the weight results, a threshold determination algorithm is used, combined with the inventory safety line and inventory turnover rate, to divide the critical ranges for replenishment, reduction of stock, maintenance of inventory, and optimization of allocation, forming quantitative adjustment rules. Inventory is dynamically adjusted based on the quantitative adjustment rules.

10. A sales strategy intelligent optimization system based on multi-source data, using the sales strategy intelligent optimization system and method based on multi-source data as described in any one of claims 1-9, characterized in that, include: Outfit Data Acquisition Module: Includes: Outfit data acquisition unit acquires outfit content data from mainstream video platforms and performs preprocessing; Popular Style Cluster Construction Module: Includes: Visual Style Feature Extraction Unit, which uses computer vision algorithms to analyze clothing content data and extract visual style features; Style Feature System Establishment Unit, which extracts style keywords from clothing content data and associates them with visual style features to establish a style feature system; Popular Style Cluster Construction Unit, which classifies clothing based on the style feature system to form popular style clusters. The associated clothing list generation module includes: an in-store clothing library construction unit that digitally models all clothing in the store and establishes a unified dimension with the style feature system to form an in-store clothing library; and an associated clothing list generation unit that uses the quantitative features of each popular style cluster as search conditions, searches the in-store clothing feature library through a similarity algorithm, and selects clothing with a style matching degree of up to a set threshold with the popular style features to form an associated clothing list corresponding to each style. The dynamic inventory adjustment module includes: a targeted delivery unit that performs multi-channel targeted delivery based on a list of related clothing items; a click-to-buy ratio calculation unit that calculates the click-to-buy ratio by statistically analyzing the click-to-buy volume and purchase volume of each related clothing item during the promotion period; a topology relationship feature analysis unit that constructs the clothing association topology structure and analyzes the topology relationship features; and a dynamic inventory adjustment unit that dynamically adjusts the inventory based on the click-to-buy ratio, topology relationship features, and the existing inventory level.

Images