A method for predicting fashion trends of middle-aged and elderly clothing based on social graph data

Abstract

This invention discloses a method for predicting fashion trends in clothing for middle-aged and elderly people based on social media image and text data. Specifically, it relates to the field of clothing market data analysis and fashion trend prediction technology. The method includes acquiring social media image and text records of clothing for middle-aged and elderly people before the prediction benchmark date, forming an image and text time sequence according to account identifiers and posting times, extracting clothing structural boundaries and wearing actions that cause boundary displacements, and integrating these with product image and text records to generate a clothing sample sequence. For clothing regions with consistent texture fingerprints in the clothing sample sequence, a posture displacement field is generated based on human body key points. A clothing displacement field is generated using a neural deformation pyramid registration algorithm. The posture displacement field is subtracted from the clothing displacement field, retaining boundary displacements whose directions match the wearing actions and pass through the action positions. A temporary correction is made by extracting the same direction of the clothing structure after eliminating human posture displacements, and candidate style features are predicted by combining the inherent structural transformation relationships in product images and text.

Description

Technical Field

[0001] This invention relates to the field of apparel market data analysis and fashion trend prediction technology, and more specifically, to a method for predicting fashion trends of middle-aged and elderly people's clothing based on social graphic data. Background Technology

[0002] Fashion trend prediction focuses on styles, colors, fabrics, or combinations that have already gained a certain level of popularity. When processing data, edge computing nodes are generally used to first identify clothing categories and visual elements from social media posts, and then combine the release time, interactive feedback, and user profiles to predict the subsequent trend of popularity. When analyzing youthful dressing styles for middle-aged and elderly users, much of the content on social media platforms does not showcase the original patterns of ready-made garments. Instead, it reflects temporary solutions implemented by users in real-life situations, such as rolling up sleeves, layering innerwear, wearing belts, changing fastening methods, or covering up parts of the garment to address issues like overly long cuffs, low necklines, ill-fitting waistlines, or hems that restrict movement or are inconvenient to open or close. Furthermore, the predicted results need to be generated before the products are mass-produced and new products are launched. Edge computing is used to pre-consolidate similar dressing variations scattered across different accounts and scenarios. Existing identification and prediction processes typically categorize the overall appearance after temporary processing into new clothing categories, or weaken the impact of local processing as differences in posture, obstruction, shooting angle, and matching. In actual operation, it can be observed that when the same structural position is repeatedly modified in the same direction by multiple middle-aged and elderly users, the popularity of the original style continues to rise, while the corresponding improvements in sleeve length, neckline shape, waist cinching, or hem structure fail to be output as candidate popular features for the next stage. The technical problem this application aims to solve is: how to use edge computing to identify multiple users' unidirectional temporary corrections to the same clothing structure position from social media posts about middle-aged and elderly clothing, and to convert these unidirectional temporary corrections into candidate style features for predicting subsequent fashion trends. Summary of the Invention

[0003] To overcome the aforementioned deficiencies of the prior art, embodiments of the present invention provide a method for predicting the fashion trends of middle-aged and elderly clothing based on social graphic data. This method extracts a temporary correction of the clothing structure in the same direction after eliminating human posture displacement, and combines the inherent structural transformation relationship in the product images and text to predict the features of candidate styles, thereby solving the problems mentioned in the background art.

[0004] To achieve the above objectives, the present invention provides the following technical solution: a method for predicting fashion trends in clothing for middle-aged and elderly people based on social media image and text data, comprising: S1. Obtain social media image and text records of middle-aged and elderly clothing before the prediction benchmark date, form an image and text time sequence according to account identifier and posting time, extract the clothing structural boundary and wearing actions that cause boundary displacement, and integrate the product image and text records to generate a clothing sample sequence. S2. For clothing regions with consistent texture fingerprints in the clothing sample sequence, generate a pose displacement field based on human key points, generate a clothing displacement field through the neural deformation pyramid registration algorithm, subtract the pose displacement field from the clothing displacement field, retain the boundary displacements whose directions conform to the wearing action and pass through the action position, and generate a temporary correction record containing the start position and end position. S3. Perform cross-industry data mining on temporary correction records and subsequent product image and text records, and register the termination position to the product clothing area. When the termination position falls into the product structure boundary and the product text does not record the wearing action, generate historical style conversion records, and establish a daily style conversion network with the temporary correction record pointing to the inherent structure of the product and the number of conversion accounts as the edge weight. S4. Construct a daily Laplace matrix using the Laplace dynamic network change point detection algorithm, subtract adjacent date matrices and perform eigenvalue decomposition, collect the product of the absolute value of the projection of the feature vector and the absolute value of the feature value according to the inherent structure of the product, write the date corresponding to the first digit of the collected value as the market transformation change point, and generate style transformation relationship; S5. Extract temporary correction records that have not been converted into the inherent structure of the product before the forecast benchmark date as candidate style features. Calculate the historical conversion ratio based on the style conversion relationship. Multiply the number of corresponding accounts by the historical conversion ratio and then subtract the number of products with the same structure to obtain the market gap value. Output the prediction results of the fashion trend of middle-aged and elderly clothing in descending order of market gap value.

[0005] In a preferred embodiment, S1 includes: S11. Using the account identifier as the aggregation key, read the social image and text records of middle-aged and elderly clothing before the prediction benchmark date. After hashing the image content, perform a bitwise XOR operation with the hash of the accompanying text content to obtain the content verification value. Records with the same content verification value and the same publication time are grouped as duplicates and the record with the first writing order in the group is retained. Then, the image and text order number is written according to the publication time as the retained record, and the account image and text time sequence is output. S12. Read social image and text records one by one in the account image and text time sequence, cut out the clothing area based on the closed outer contour of the human body key points, calculate the gray difference of adjacent pixels along the scan line in the clothing area and connect the pixels whose difference sign changes from positive to negative or from negative to positive as the clothing structure boundary, and then search for action words in the three adjacent words in the word segmentation sequence of the caption based on the clothing part words hit by the clothing structure boundary, and output the boundary action record. S13. After splicing the clothing category code and clothing part code in the boundary action record with the image and text sequence number, perform a hash operation to obtain the market collection key. Read the product records whose release time is earlier than the prediction base date and whose hash result matches the market collection key from the product image and text records. Continuate the product release time to the account image and text sequence and output the clothing sample sequence.

[0006] In a preferred embodiment, S2 includes: S21. Read the source and target clothing regions with consistent texture fingerprints in the clothing sample sequence. Initialize the affine parameters of the triangular mesh with the coordinates of the human body key points. Enumerate candidate parameters in the one-dimensional integer neighborhood. Generate the attitude cost value by adding the sum of squared residuals of the key point reprojection to the sum of squared affine differences of adjacent meshes. Select the candidate parameter with the first position in ascending order of attitude cost value. Output the attitude displacement field when the affine parameters no longer change after one round of enumeration.

[0007] In a preferred embodiment, S2 further includes: S22. Using the deformed source clothing region and target clothing region as inputs to the neural deformation pyramid registration algorithm, generate the number of pyramid layers according to the number of binary bits of the long side of the bounding rectangle of the source clothing region. In each layer, the sampling points of the structural boundary are used as deformation neurons. The displacement of the neurons in this layer is initialized by the displacement interpolation of the previous layer. Candidate displacements are enumerated in the integer pixel neighborhood where the Manhattan distance does not exceed one. The registration cost is generated according to the texture Hamming distance, the bidirectional chamfer distance and the sum of squares of the displacement difference between adjacent neurons. The candidate displacement with the first position in ascending order of the registration cost is selected. The clothing displacement field is output when the sampling interval is one pixel and the displacement of all neurons no longer changes. S23. Subtract the attitude displacement field from the clothing displacement field to generate the structural residual displacement field. Read the structural residual displacement from the hit pixels of the wearing action on the structural boundary along the boundary chain. Retain the residual displacement chain that has a positive dot product with the wearing action direction and is registered back to the hit pixel in the reverse direction. When a residual displacement chain conflict occurs, retain the residual displacement chain that is first in ascending order of registration value. Write the first and last ends of the residual displacement chain along with the account identifier into the temporary correction record.

[0008] In a preferred embodiment, S3 includes: S31. Based on the temporary correction record, read the temporary correction release time, extract the product image and text records whose release time is greater than the temporary correction release time and less than the prediction base date, and write the clothing category code into the high-level market key when performing cross-industry data mining. Shift the high-level market key to the left by the part code width and write the structural part code, then shift the grid code width to the left and write the termination position grid number to obtain the corrected market key. Write the product boundary grid number into each product image and text record according to the same shift rule and generate the product market key. Filter the product image and text records whose product market key is equal to the corrected market key and output the market acceptance candidate set. S32. For the product image and text records in the market acceptance candidate set, take the temporary correction termination position as the projection starting point, enumerate candidate projection points along the boundary normal of the product clothing area, calculate the Hamming distance from the temporary correction texture fingerprint to the neighboring texture of the candidate projection point to generate the first cost, calculate the chamfer distance from the candidate projection point to the product structure boundary to generate the second cost, calculate the absolute value of the integer difference between the grid number of the candidate projection point and the grid number of the termination position to generate the third cost, and generate the product registration cost by adding the first cost, the second cost and the third cost, take the candidate projection point with the product registration cost sorting number as one, and output the product projection position when the candidate projection points with the sorting number as one in two adjacent rounds are consistent.

[0009] In a preferred embodiment, S3 further includes: S33. After obtaining the boundary hit position by bitwise ANDing between the pixel where the product projection position is located and the pixel of the product structure boundary, compare the wearing action word by word in the product text segmentation sequence and accumulate the number of action hits. When the number of action hits is zero, write the action disappearance position one. When the number of action hits is not zero, write the action disappearance position zero. When the boundary hit position multiplied by the action disappearance position equals one, write the product structure boundary as the inherent structure of the product and generate a historical style conversion record. S34. Collect historical style conversion records by natural day, write the temporary correction record number into the edge starting point, and write the product's inherent structure number into the edge ending point. Perform deduplication and count the account identifiers corresponding to the same edge starting point and the same edge ending point within the same natural day and write them as edge weights to establish a daily style conversion network.

[0010] In a preferred embodiment, S4 includes: S41. Based on the daily style transformation network, read the edge weight of each natural day, write the temporary correction record number into the in-point dictionary, write the product inherent structure number into the out-point dictionary, use the concatenation order of the in-point dictionary and the out-point dictionary as the matrix row and column order, write the same edge weight into the in-point to out-point position and the out-point to in-point position, write zero for the missing position, and generate the same node adjacency matrix. S42. Accumulate the edge weights row by row in the adjacency matrix of the same node, write the accumulated row value to the diagonal position in the same order, and write the non-diagonal position to the opposite number of the edge weight at the same position in the adjacency matrix of the same node, to generate the daily Laplace matrix.

[0011] In a preferred embodiment, S4 further includes: S43. Using the Laplace dynamic network change point detection algorithm, read the Laplace matrix of two adjacent days in chronological order. Subtract the previous day's matrix from the matrix of the next day to generate a difference matrix. Perform eigenvalue decomposition on the difference matrix. Multiply the absolute value of the component of each eigenvector at the commodity inherent structure node by the absolute value of the same-order eigenvalue and accumulate it to the commodity inherent structure node. Output the daily commodity structure change table. S44. In the daily commodity structure change table, sort the dates according to the inherent structure nodes of the commodities. First, sort them in descending order by the aggregation value. If the aggregation values ​​are the same, sort them by the date that comes first. Write the date with the sort number as the market transformation change point. Read the historical style transformation records before the market transformation change point. Divide the edge weight of the temporary correction record within the same clothing category pointing to the inherent structure of the commodity by the total edge weight within the same clothing category to generate the style transformation relationship.

[0012] In a preferred embodiment, S5 includes: S51. Based on the temporary correction record before the prediction benchmark date, shift the clothing category code to the left by the part code width and write it into the structural part code, then shift the direction code width to the left and write it into the displacement direction code to obtain the candidate structure code. Query the number of records in the product image and text records that match the product inherent structure code and the candidate structure code. When the number of records is zero, write the temporary correction record as the candidate style feature. S52. Based on market conversion change points, read the historical edge weights in the style conversion relationship that are consistent with the entry point code and the candidate structure code. Divide each historical edge weight by the absolute value of the date difference between the corresponding natural day and the market conversion change point plus one, and sum them up to obtain the historical conversion volume. Then divide the historical conversion volume by the number of historical temporary correction accounts that are consistent with the entry point code and the candidate structure code to obtain the historical conversion ratio.

[0013] In a preferred embodiment, S5 further includes: S53. The number of currently adopted accounts is obtained by deduplicating the temporary correction records corresponding to the candidate style features according to the account identifier. The predicted demand is obtained by multiplying the current number of adopted accounts by the historical conversion ratio. Then, the number of products whose inherent structure code in the product image and text records is consistent with the candidate structure code is read and written as the market supply. S54. Obtain the market gap value by subtracting the market supply from the predicted demand. Sort the candidate style features in descending order of market gap value. When the market gap values ​​are the same, sort them in ascending order by the market transformation change point date corresponding to the candidate style features. Output the prediction result of the fashion trend of middle-aged and elderly clothing.

[0014] A system for predicting fashion trends in clothing for middle-aged and elderly people based on social media image and text data includes a data collection and sequencing module, a correction and identification module, a conversion and network building module, a change point analysis module, and a trend output module. The data collection and sequencing module is used to acquire social media image and text records of clothing for middle-aged and elderly people before the prediction benchmark date, form image and text time sequence according to account identifier and posting time, extract clothing structural boundaries and wearing actions that cause boundary displacement, and integrate product image and text records to generate clothing sample sequence; The correction and recognition module is used to generate a posture displacement field based on human key points for clothing areas with consistent texture fingerprints in the clothing sample sequence. It generates a clothing displacement field through the neural deformation pyramid registration algorithm, subtracts the posture displacement field from the clothing displacement field, retains the boundary displacements whose directions conform to the wearing action and pass through the action position, and generates a temporary correction record containing the start position and end position. The conversion network module is used to perform cross-industry data mining on temporary correction records and subsequent product image and text records. It registers the termination position to the product clothing area. When the termination position falls into the product structure boundary and the product text does not record the wearing action, it generates historical style conversion records and establishes a daily style conversion network with the temporary correction record pointing to the product's inherent structure and the number of conversion accounts as edge weights. The change point analysis module is used to construct a daily Laplace matrix through the Laplace dynamic network change point detection algorithm, subtract adjacent date matrices and perform eigenvalue decomposition, collect the product of the absolute value of the projection of the feature vector and the absolute value of the feature value according to the inherent structure of the product, write the date corresponding to the first digit of the collected value as the market transformation change point, and generate style transformation relationship; The trend output module is used to extract temporary correction records that have not been converted into the inherent structure of the product before the prediction benchmark date as candidate style features. Based on the style conversion relationship, the historical conversion ratio is calculated. The number of corresponding accounts is multiplied by the historical conversion ratio and then the number of products with the same structure is subtracted to obtain the market gap value. The prediction results of the fashion trend of middle-aged and elderly clothing are output in descending order of market gap value.

[0015] A storage device, comprising: A processor, and memory connected to the processor; memory is used to store computer programs.

[0016] A storage medium, comprising: The storage medium stores a computer program, which, when executed by a processor, performs the method.

[0017] The technical effects and advantages of this invention are as follows: This solution reduces the interference of human posture changes on temporary correction recognition by deducting posture displacement and verifying the direction of wearing action, while retaining the boundary displacement corresponding to the actual adjustment of clothing structure. By using edge computing nodes to run the neural deformation pyramid, the clothing structure boundary is registered layer by layer, so that the clothing boundary position has higher consistency under different shooting angles and partial occlusion. By linking the temporary corrections output by edge computing nodes with the inherent structure of subsequent products through cross-industry data mining, local structural changes in scattered graphics and text can form a traceable structural transformation record. Based on the Laplace spectrum changes of the daily style transformation network, the structural transformation and change points are located, so that the identification of structural changes in the graphic and text sequence changes from quantitative fluctuations to changes in network connection relationships. By combining historical conversion rates and the number of similar products, the market gap value is calculated, and the ranking of candidate style features is based on unified coding and recalculated values, thereby improving the consistency of the output results. Attached Figure Description

[0018] Figure 1 This is a flowchart outlining the method steps of the present invention. Detailed Implementation

[0019] 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.

[0020] Refer to the instruction manual appendix Figure 1 The present invention provides a method for predicting fashion trends in clothing for middle-aged and elderly people based on social media image and text data, comprising: S1. Obtain social media image and text records of middle-aged and elderly clothing before the prediction benchmark date, form an image and text time sequence according to account identifier and posting time, extract the clothing structural boundary and wearing actions that cause boundary displacement, and integrate the product image and text records to generate a clothing sample sequence. This implementation method is used to generate a clothing sample sequence that can simultaneously characterize social dressing behavior and commodity market supply before the prediction benchmark date. The prediction benchmark date is uniformly set as 00:00 on the prediction benchmark date. Records published before 00:00 on the prediction benchmark date are included in the processing scope of this implementation method. Social image and text records of middle-aged and elderly clothing should at least include record number, de-identified account identifier, publication time, writing order, image content, and caption content. Commodity image and text records should at least include commodity record number, commodity publication time, commodity image, commodity caption, and commodity writing order. Multi-image content is split into image and text sub-records according to image order. Multi-person images are split into wearer sub-records according to human body key point bounding boxes. Multiple garments are generated into clothing regions according to clothing category, so that one clothing region corresponds to one subsequent boundary action recognition object. The implementation process includes the following steps: S11 is used to eliminate the interference of repeated posts within the same account on market collection samples and to provide a stable time sequence within the account for subsequent boundary action recognition; the account identifier is used as the aggregation key to read the social image and text records of middle-aged and elderly clothing before the prediction benchmark date. First, a fixed-length perceptual hash is performed on the image content to obtain the image content hash value. Then, a same-bit wide text hash is performed on the accompanying text to obtain the accompanying text hash value. Finally, the image content hash value and the accompanying text hash value are XORed bit by bit to generate the content verification value. Records with the same content verification value and the same publication time under the same account identifier are written into the duplicate group. Within the duplicate group, the record with the first writing priority value is retained. Records with different content verification values ​​but the same publication time are arranged according to the writing priority. All retained records are written into the image and text priority number in ascending order of publication time. The account image and text sequence is output for S12 to read item by item. When the caption content is missing, the hash value of the caption content is written to a zero value of the same width. When the image content is missing, the corresponding social image and text record will not be included in the duplicate group generation and subsequent clothing area cropping. When the account identifier is missing, a temporary account identifier is generated by hashing the record number and written to the account missing marker. S12 is used to establish a calculable binding relationship between clothing structure boundaries and wearing actions from the time sequence of account images and texts, so that subsequent temporary correction recognition can locate specific clothing parts; the social image and text records are read item by item in the time sequence of account images and texts, and the clothing area is cut according to the closed circumscribed contour of the human body key points. The upper clothing area is enclosed by the shoulder key points, elbow key points, wrist key points and hip key points, and the lower clothing area is enclosed by the hip key points, knee key points and ankle key points. The gray level difference between adjacent pixels in the clothing area is calculated according to the row scan line and column scan line respectively. Pixels whose gray level difference sign changes from positive to negative or from negative to positive are written as boundary candidate pixels. Boundary candidate pixels with the same coordinates or a coordinate difference of one in adjacent scan lines are connected to form the clothing structure boundary; The clothing structure boundary is mapped to the clothing part code based on the local position of the human body key points. The clothing part words are read from the clothing part code. Then, with the word position of the clothing part word in the matching word segmentation sequence as the center, the action words in three word positions forward and backward are read. After the action words are matched, the wearing action and action direction code are written. The boundary action record containing the image and text sequence number, clothing category code, clothing part code, clothing structure boundary, wearing action and action direction code is output for S13 to read. When key points of the human body are missing, a key point missing marker is written and boundary action recognition of the current clothing area is stopped. When the text is missing or the action word is not matched, the clothing action is written with a null value. The boundary action record with the clothing action written with a null value is only used as the source clothing area where the clothing action does not appear for subsequent registration reading. S13 is used to connect the boundary action records formed on the social dressing side with the product image and text records formed on the product market side into the same market collection sample, so as to avoid the product records being unable to match due to the lack of account image and text sequence number; the market primary key is generated by performing segmented hashing on the clothing category code and clothing part code in the boundary action record, and then the time sequence check segment is generated by performing sequence hashing on the image and text sequence number. The market primary key and the time sequence check segment are concatenated to obtain the market collection key. Read product records whose release time is earlier than 00:00 on the prediction base date from the product image and text records. After extracting the clothing area of ​​the product image, generate the clothing category code and clothing part code. Generate the product market primary key according to the same segmented hash rule. When the product market primary key is consistent with the market primary key in the market collection key, write it into the hit product record. Then, in ascending order of product release time, continue the hit product records to the account image and text time sequence and output the clothing sample sequence for S2 to read. When no product record is matched, the clothing sample sequence retains the account image and text time sequence and writes a marker indicating that the product subsequence is empty. When multiple matched product records have the same product release time, they are arranged according to the order in which the products are written. When the clothing area of ​​the product image cannot be extracted, the corresponding product record is not written into the clothing sample sequence. Through the above processing, the changes in clothing within accounts in social media posts are organized into ordered samples. Clothing structure boundaries and wearing actions are bound to boundary action records. Product post records are then connected to the social media post sequence with clothing category and clothing parts as the market key, thus forming a clothing sample sequence that combines social clothing behavior and product market supply. In practical application: Account A publishes two pictures and texts containing the same top before the prediction benchmark date. The caption of the second picture is "Rolling up the cuffs makes it look neater". S11 deletes the duplicate post of the same picture and text from the same account and writes the picture and text sequence number. S12 identifies the code of the clothing part corresponding to the cuff from the edge of the top cuff and writes "rolling up" as the wearing action. S13 then accesses the product records of the same clothing category and the same cuff part from the product picture and text records. Subsequently, S2 can compare the transformation relationship between the temporary change of the cuff of account A and the cuff structure of the product.

[0021] S2. For clothing regions with consistent texture fingerprints in the clothing sample sequence, generate a pose displacement field based on human key points, generate a clothing displacement field through the neural deformation pyramid registration algorithm, subtract the pose displacement field from the clothing displacement field, retain the boundary displacements whose directions conform to the wearing action and pass through the action position, and generate a temporary correction record containing the start position and end position. This implementation distinguishes between changes in human posture and temporary corrections to clothing structures actively generated by the user from clothing sample sequences. The process first calculates the posture displacement field caused by body movements using human keypoints, then calculates the overall deformation of the clothing region using a neural deformation pyramid registration algorithm. Subsequently, the posture displacement field is subtracted from the clothing displacement field, and the actual temporary correction is selected based on the direction of the wearing action. The source clothing region is the clothing region in the same account's image-text sequence that was published earlier but did not match the corresponding wearing action; the target clothing region is the clothing region that was published later but matched the corresponding wearing action. When multiple candidate source clothing regions exist, candidate source clothing regions with consistent texture fingerprints are first retained, then sorted by publication time distance from nearest to farthest, and the candidate source clothing region with the sort number one is read. The texture fingerprint is generated by binary texture encoding within the neighborhood of the clothing structure boundary. If human keypoints, source clothing regions, or target clothing regions are missing, the current clothing region is not included in the temporary correction calculation. This implementation process includes the following steps: S21 is used to establish the base deformation corresponding to changes in human posture, so that the subsequent neural deformation pyramid registration algorithm will not mistakenly write the clothing displacement caused by raising the arm, turning around, or bending over as a temporary correction; read the source clothing region and target clothing region with consistent texture fingerprints in the clothing sample sequence, and read the corresponding human key point pairing coordinates of the two, and generate triangular meshes according to the order of human key points corresponding to the clothing category. Taking the upper clothing region as an example, the shoulder key points, elbow key points, wrist key points and hip key points are formed into triangular meshes in a fixed order; Each triangular mesh initializes its affine parameters based on the correspondence between the key point coordinates of the source clothing area and the key point coordinates of the target clothing area. The affine parameters are represented by integer pixels. Candidate parameters are enumerated in a one-dimensional integer neighborhood consisting of the current affine parameter minus one, the current affine parameter, and the current affine parameter plus one. For each set of candidate parameters, calculate the sum of squared coordinate differences between the key points of the source clothing region after affine transformation and the corresponding key points of the target clothing region to obtain the sum of squared reprojection residuals of the key points. Then calculate the sum of squared differences between the candidate affine parameters of adjacent triangular meshes. Add the sum of squared reprojection residuals of the key points to the sum of squared differences of the affine differences of adjacent meshes to generate the attitude value. Select the candidate parameter with the first position in the ascending order of attitude value to update the affine parameters of the triangular mesh. When the attitude values ​​are the same, retain the candidate parameter with the earlier lexicographical order of the affine parameters. When all the affine parameters of the triangular mesh are exactly the same as those of the previous round after one round of enumeration, the pixel displacement generated by the affine transformation of each triangular mesh is written into the pose displacement field and read by S22. If the number of human key points with the same name cannot form the triangular mesh corresponding to the clothing category, the pose calculation failure flag is written and the subsequent processing of the current source clothing region and target clothing region is stopped. S22 is used to calculate the actual deformation of the clothing structure layer by layer in the image space after eliminating the influence of the attitude basis, so that the boundary displacement can converge from coarse to fine to the pixel level. The source clothing region and the target clothing region after the attitude displacement field deformation are used as inputs to the neural deformation pyramid registration algorithm. The number of pixels on the long side of the bounding rectangle of the source clothing region is read. The number of pixels on the long side is counted by binary right shift until the shift result is zero, and the number of pyramid layers is obtained. The sampling interval of the highest layer is set to the pixel interval corresponding to the highest binary bit. Then, the sampling interval is halved layer by layer until the sampling interval is one pixel. Among them, the neural deformation pyramid registration algorithm adopts a layer-by-layer deformation neural network structure. Each layer reads the structural boundary sampling points from the source clothing structure boundary according to the current sampling interval, writes each structural boundary sampling point as a deformation neuron, and writes two adjacent structural boundary sampling points in the boundary chain as a neuron connection. Each deformation neuron stores row displacement parameters and column displacement parameters. The row displacement parameters and column displacement parameters of the first layer are both written to zero. The row displacement parameters and column displacement parameters of the next layer are obtained by bilinear interpolation of the displacement field of the previous layer at the current sampling point coordinates. The algorithm input is the source clothing region, target clothing region, source clothing structure boundary, target clothing structure boundary and texture fingerprint after deformation by the attitude displacement field. The algorithm output is the clothing displacement field corresponding to each pixel of the source clothing structure boundary. Each layer reads structural boundary sampling points from the clothing structure boundary of the source clothing area according to the current sampling interval, and writes the structural boundary sampling points as deformation neurons. The displacement of the first layer deformation neurons is written as zero, and the displacement of the next layer deformation neurons is obtained by performing bilinear interpolation on the current sampling position by the displacement field of the previous layer. For each deformable neuron, enumerate candidate displacements within an integer pixel neighborhood of the current displacement with a Manhattan distance not exceeding one. Delete candidate displacements that fall outside the clothing area after the displacement. Calculate the Hamming distance between the texture fingerprint of the candidate position's neighborhood and the corresponding texture fingerprint of the target clothing area, the bidirectional chamfer distance from the candidate position to the boundary of the target clothing structure, and the sum of squares of the differences between candidate displacements of adjacent deformable neurons. Add these three items directly to generate the registration value. Select the candidate displacement with the highest registration value in ascending order to update the deformable neuron displacement. If the registration values ​​are the same, retain the first candidate displacement in ascending order of row and column number. In this process, the parameters of each deformation neural network layer are updated iteratively within the layer. For the current deformation neuron, the row displacement parameters and column displacement parameters from the previous round are read. Row displacement parameters minus one, row displacement parameters, and row displacement parameters plus one are used as candidate row displacements. Column displacement parameters minus one, column displacement parameters, and column displacement parameters plus one are used as candidate column displacements. These combinations yield nine candidate two-dimensional displacements. When the coordinates of the sampling point after the application of a candidate two-dimensional displacement fall outside the outer contour of the source clothing region, the candidate two-dimensional displacement is deleted, and the remaining candidate two-dimensional displacements are used in the loss calculation. For each remaining candidate two-dimensional displacement, the coordinates of the structural boundary sampling point are added bit by bit to the candidate two-dimensional displacement to obtain the candidate position; the three-by-three neighboring pixels are read with the candidate position as the center and the candidate neighborhood texture fingerprint is generated; the candidate neighborhood texture fingerprint is XORed bit by bit with the target neighborhood texture fingerprint of the same coordinate three-by-three neighboring region in the target clothing area, and the number of single values ​​in the XOR result is accumulated to obtain the texture Hamming distance. Search for the target clothing structure boundary pixels along the row and column directions from the candidate position, record the Manhattan distance from the candidate position to the nearest target clothing structure boundary pixel, then search for the source clothing structure boundary pixels in reverse from the nearest target clothing structure boundary pixel and record the reverse Manhattan distance, add the two Manhattan distances to get the bidirectional chamfer distance; Read the candidate 2D displacements of deformable neurons adjacent to the current deformable neuron in the boundary chain, calculate the squared displacement difference in the row direction and the squared displacement difference in the column direction respectively, and add them together to obtain the sum of squared displacement differences between adjacent neurons; directly add the texture Hamming distance, the bidirectional chamfer distance and the sum of squared displacement differences between adjacent neurons to obtain the registration cost of the current candidate 2D displacement. When the sampling interval is reduced to one pixel and the displacement of all deformable neurons is consistent with the previous round after one round of update, the clothing displacement field is output and read by S23. If the structural boundary in the target clothing region is empty, the current source clothing region and the target clothing region will not output the clothing displacement field. After calculating the registration cost of all candidate two-dimensional displacements of the current deformed neuron, the candidate two-dimensional displacements are arranged in ascending order of registration cost. If the registration cost is the same, the candidate two-dimensional displacements are first arranged in ascending order of row displacement, and then in ascending order of column displacement. The first candidate two-dimensional displacement is read and the row displacement and column displacement parameters of the current deformed neuron are updated. After each update cycle, the row and column displacement parameters of the current cycle and the previous cycle are compared one by one. If all are equal, an intra-layer convergence flag is written. When the sampling interval of the current layer is greater than one pixel, bilinear interpolation is performed on the row and column displacement parameters of all deformable neurons in the current layer, and these parameters are used as the initial parameters for the deformable neurons in the next layer. When the sampling interval of the current layer is equal to one pixel and the intra-layer convergence flag is valid, the row and column displacement parameters of all deformable neurons are written into the clothing displacement field in the order of the source clothing structure boundary pixels. S23 is used to extract the posture displacement from the clothing displacement field and extract the boundary displacement chain that is mutually verified with the wearing action, so that the temporary correction record only corresponds to the structural changes actively formed by the user; the displacement of each structural boundary pixel in the clothing displacement field is subtracted from the displacement of the same structural boundary pixel in the posture displacement field to generate the structural residual displacement field; then, according to the boundary action record output by S12, the wearing action, action direction code and clothing part code are read, the clothing part code is mapped to the pixel set on the clothing structural boundary, and the structural boundary pixel corresponding to the action word position is written as the wearing action hit pixel. Starting from the pixel hit by the clothing action, read the structural residual displacement on the boundary chain along the increasing and decreasing pixel index directions in the local coordinates of the clothing. Take the dot product of the structural residual displacement with the unit direction vector corresponding to the action direction code, retain the structural residual displacement with positive dot product, and connect the retained structural residual displacement according to the boundary chain to form a residual displacement chain. Perform reverse registration on the remaining displacement chain, that is, subtract the remaining displacement of the corresponding structure from the position of the tail of the remaining displacement chain and then calculate back to the source clothing area. If the calculated pixel is consistent with the pixel coordinate of the wearing action, the remaining displacement chain is retained; if the calculated pixel is inconsistent with the pixel coordinate of the wearing action, the remaining displacement chain is deleted. Two remaining displacement chains are considered to be in conflict if they occupy the same structural boundary pixel or have the same tail position. In case of conflict, the remaining displacement chain corresponding to the first and second position of the registration value in S22 is retained. The first end of the retained remaining displacement chain is written as the temporary correction start position, and the last end of the retained remaining displacement chain is written as the temporary correction end position. The account identifier, clothing category code, structural part code, wearing action, action direction code, temporary correction release time, source clothing area number, target clothing area number, texture fingerprint, and registration value are written into the temporary correction record. If the wearing action is null, the action direction code is missing, or there is no structural remaining displacement with a positive dot product, no temporary correction record is generated for the current target clothing area. Through the above processing, the human body posture change is first calculated as the posture displacement field, and the pixel-level deformation of the clothing area is then calculated as the clothing displacement field. After subtracting the posture displacement field from the clothing displacement field, the structural residual displacement field is obtained. The structural residual displacement field is then verified by the direction of the wearing action, the pixels hit by the wearing action, and the consistency of the reverse registration, so that the temporary correction record can be distinguished from the boundary offset caused by ordinary posture changes, shooting angle changes, and occlusion. In practical application: In the first image of the shirt in account A, the cuffs hang naturally. In the second image of the shirt, the caption reads "Rolling up the cuffs makes it look neater." S21 calculates the posture displacement field caused by raising the arm based on the key points of the shoulder, elbow, wrist, and hip. S22 calculates the clothing displacement field of the cuff boundary between the two shirt areas after posture correction. S23 subtracts the posture displacement field from the clothing displacement field and retains the remaining displacement chain that is consistent with the "roll up" direction and can return to the cuff hit pixel. Finally, the original position of the cuff and the position after rolling up are written into the temporary correction record.

[0022] S3. Perform cross-industry data mining on temporary correction records and subsequent product image and text records, and register the termination position to the product clothing area. When the termination position falls into the product structure boundary and the product text does not record the wearing action, generate historical style conversion records, and establish a daily style conversion network with the temporary correction record pointing to the inherent structure of the product and the number of conversion accounts as the edge weight. This implementation method is used to perform cross-industry data mining on temporary correction records confirmed on the social side and subsequent product image and text records on the product side. The core processing is to first use the market key to filter out product image and text records that have a sequential relationship in both time and structure, then project the temporary correction termination position onto the product clothing area, and finally use the product structure boundary hit and the disappearance of product caption actions to jointly confirm the inherent structure of the product; the prediction base date is uniformly taken as 0:00 on the prediction base date, and subsequent product image and text records are those whose release time is greater than the temporary correction release time but less than 0:00 on the prediction base date; the temporary correction record includes at least the temporary correction record number, account identifier, temporary correction release time, clothing category code, structural part code, termination position grid number, temporary correction termination position, texture fingerprint, and wearing action, and the product image and text record includes at least the product record number, product release time, product image, product caption, and product writing order; the implementation process includes the following steps: S31 is used to establish retrieval constraints for the same structural position between the social clothing side and the commodity market side, reducing irrelevant commodity image and text records from entering subsequent projection calculations; based on the temporary correction record, the temporary correction release time is read, and the commodity image and text records with the release time greater than the temporary correction release time and less than the prediction base date zero hour are extracted from the commodity image and text records. The maximum code value in the clothing category code table, structural part code table and grid code table is read, and binary right shift counting is performed on the maximum code value until the shift result is zero, to obtain the clothing category code width, part code width and grid code width; When performing cross-industry data mining, the clothing category code is shifted left by the width of the part code and then written into the structural part code. The grid code is then shifted left by the width of the grid code and then written into the ending position grid number to obtain the corrected market key. For each subsequent product image and text record, the product clothing area and product structural boundary are first extracted from the product image. Then, for each boundary pixel on the product structural boundary, a product boundary grid number is generated according to the row priority order of the clothing local coordinates. The product market key is generated one by one according to the same shift rule as the corrected market key. When any product market key equals the corrected market key, the corresponding subsequent product image and text record is written into the market acceptance candidate set and read by S32. When the encoded value exceeds the corresponding bit width, an encoding overflow flag is written and the current record processing is stopped. When the market acceptance candidate set is empty, no historical style conversion record is generated. S32 is used to map the temporary correction termination position to the verifiable pixel position in the product and clothing area, so that whether the product side structure accepts the social side temporary correction can be constrained by both image distance and texture consistency. For the product image and text records in the market acceptance candidate set, the center of the same numbered grid in the product and clothing area is first located using the termination position grid number, and the center of the same numbered grid is used as the projection starting point. Then, the product structure boundary pixel and adjacent boundary pixel closest to the projection starting point are read, and the vertical direction of the line connecting the two is calculated as the boundary normal of the product and clothing area. Candidate projection points are enumerated pixel by pixel along the positive and negative directions of the boundary normal. The enumeration stops when the range reaches the outer contour of the product clothing area. For each candidate projection point, the Hamming distance from the temporary modified texture fingerprint to the neighboring texture fingerprint of the candidate projection point is calculated and written as the first cost. The chamfer distance from the candidate projection point to the boundary of the product structure is calculated and written as the second cost. The absolute value of the integer difference between the grid number of the candidate projection point and the grid number of the termination position is calculated and written as the third cost. The first cost, the second cost and the third cost are directly added together to obtain the product registration cost. Candidate projection points are sorted in ascending order of registration value. When the registration values ​​are the same, they are sorted in ascending order of pixel distance from the candidate projection point to the projection start point. The candidate projection point with the sort number 1 is selected. If the candidate projection points with the sort number 1 are the same in two adjacent rounds, the candidate projection point with the sort number 1 is written as the product projection position and read by S33. When the candidate projection points are in a loop, the candidate projection point with the sort number 1 of the registration value in the loop is read as the product projection position. When the product structure boundary is empty, the corresponding product image and text record is deleted. S33 is used to confirm whether the product image and text record has transformed the temporary correction into the structure of the product itself, rather than continuing to display the temporary changes caused by the user's wearing actions; after the product projection position is output, the product projection position is rounded to the pixel coordinates of the product image, and the pixel where the product projection position is located is ANDed with the binary mask of the product structure boundary to obtain the boundary hit bit. A boundary hit bit of 1 indicates that the product projection position falls into the product structure boundary, and a boundary hit bit of 0 indicates that the product projection position does not fall into the product structure boundary. Read the product description and generate a word segmentation sequence of the product description. Compare the wearing actions in the temporary correction record and the synonym action words recorded by the action word list word by word. Each hit is counted as one action hit count. When the product description is missing, the action hit count is recorded as zero. When the action hit count is zero, write the action disappearance bit 1. When the action hit count is not zero, write the action disappearance bit 0. When the product of the boundary hit bit and the action disappearance bit equals one, the product structure boundary is written as the product's inherent structure, and the product's inherent structure number is generated by the clothing category code, structural part code, product boundary grid number, and action direction code in a fixed shift order. At the same time, a historical style conversion record is generated for S34 to read. When the product of the boundary hit bit and the action disappearance bit does not equal one, the current product image and text record is only retained as an unconverted verification record and is not written into the historical style conversion record. S34 is used to summarize the transformation results of a single temporary modification into the inherent structure of a product into a daily market relationship, so that the subsequent Laplace dynamic network change point detection algorithm can read the transformation structure of the same date; convert the product release time in the historical style transformation record into a natural day according to a unified time zone, and collect the historical style transformation record by natural day, write the temporary modification record number into the edge start point, write the product inherent structure number into the edge end point, perform deduplication counting on the account identifiers corresponding to the same edge start point and the same edge end point within the same natural day, and write the number of deduplicated accounts as the edge weight; When the same product image and text record is published repeatedly, it is first deduplicated by product record number, and then counted for deduplication by account identifier. When the product publication time is missing, the warehouse entry time corresponding to the product writing order is read as the product publication time and written as the time replacement mark. When the daily style conversion network output is used, the natural day, edge start point, edge end point, edge weight, product registration value and product inherent structure number are retained for S4 to construct the same node adjacency matrix and the daily Laplace matrix. Through the above processing, the temporary correction record first completes cross-industry candidate shrinking using the correction market key and the product market key, then determines the product projection position through product registration value, subsequently confirms the product's inherent structure using boundary hit positions and action disappearance positions, and finally generates a daily style transformation network with edge weights according to the natural day, thereby transforming the scattered social clothing corrections into product structure inheritance relationships that can be used for market analysis. In practical application: Account A forms a temporary correction record of "cuffs rolled up" in social images and text. S31 uses the garment category code, cuff part code, and cuff termination position grid number to filter out subsequent product image and text records. S32 projects the cuff termination position onto the product cuff area. S33, when the product cuff boundary is hit and the product text does not contain "rolled up sleeves" or a synonymous action word, the product cuff boundary is written as the product's inherent structure. S34 then points the temporary correction record number corresponding to Account A to the product cuff inherent structure number and writes it into the daily edge weight, forming the input for subsequent market transformation change point detection.

[0023] S4. Construct a daily Laplace matrix using the Laplace dynamic network change point detection algorithm, subtract adjacent date matrices and perform eigenvalue decomposition, collect the product of the absolute value of the projection of the feature vector and the absolute value of the feature value according to the inherent structure of the product, write the date corresponding to the first digit of the collected value as the market transformation change point, and generate style transformation relationship; This implementation method converts the daily style transformation network into a Laplace dynamic network that reflects changes in market structure, and locates the dates of concentrated transformations in the inherent structure of goods by analyzing the changes in the network spectrum between adjacent natural days. The daily style transformation network is output by S34 and includes at least the natural day, edge start point, edge end point, edge weight, inherent structure number of the goods, and clothing category code. The edge start point corresponds to the temporary correction record number, and the edge end point corresponds to the inherent structure number of the goods. The natural days use a unified time zone, and natural days without historical style transformation records still generate an all-zero matrix to ensure that adjacent natural days can be compared day by day. This implementation process includes the following steps: S41 is used to write the daily style transformation network of different natural days into the same node space to avoid the matrix being unable to be subtracted due to the addition or removal of nodes between dates; based on the daily style transformation network, read the temporary correction record number and the product inherent structure number that have appeared in all natural days, first write the temporary correction record number into the in-point dictionary in ascending order of value, then write the product inherent structure number into the out-point dictionary in ascending order of value, and arrange the in-point dictionary before the out-point dictionary to form the same node dictionary; For each natural day, generate a square matrix according to the order of the same node dictionary, read the edge weight of the day in the daily style conversion network, write the edge weight into the row corresponding to the edge start and the column corresponding to the edge end, and write the same edge weight into the row corresponding to the edge end and the column corresponding to the edge start. Write zero in the matrix position that is not hit by the edge weight, generate the same node adjacency matrix for that natural day and let S42 read it. If there is no historical style conversion record for a certain natural day, then generate an all-zero adjacency matrix of the same nodes according to the same node dictionary. If there are duplicate edge start points and edge end points for the same natural day, then first accumulate the edge weights and then write them into the matrix. The directional relationship continues to be stored in the historical style conversion record and style conversion relationship. S42 is used to transform the adjacency matrix of the same node into a daily Laplace matrix, so that the connection strength of the market conversion relationship and the node aggregation structure can be included in the subsequent spectral decomposition. The edge weights are accumulated row by row in the adjacency matrix of the same node, and the row accumulation value of each row is written into the same diagonal position to obtain the degree matrix. Then, the adjacency matrix of the same node is subtracted from the degree matrix to obtain the daily Laplace matrix. The diagonal position of the daily Laplace matrix represents the total number of connections of the corresponding node on that day, and the off-diagonal position represents the negative number of the edge weights at the same position in the adjacency matrix of the same node. The daily Laplace matrix is ​​written into the Laplace matrix sequence according to the natural day and is read by S43. Nodes with a row cumulative value of zero are retained in the daily Laplace matrix and written with a zero-degree node mark. The matrix dimension is based on the length of the dictionary of the same node and is not reduced due to the absence of nodes on the current day. S43 is used to extract the network change amount converging towards the inherent structure of the product between adjacent natural days through the Laplace dynamic network change point detection algorithm, instead of only counting the change in the number of releases; read the Laplace matrix of two adjacent days in the Laplace matrix sequence in date order, first confirm that the dimensions of the two day matrices are consistent according to the same node dictionary, and then subtract the Laplace matrix of the previous day from the Laplace matrix of the following day to generate the difference matrix; Eigenvalue decomposition is performed on the difference matrix. The eigenvalues ​​are arranged in ascending order of numerical value. The absolute value of the component of each eigenvector at the commodity inherent structure node is read. The absolute value of the component is multiplied by the absolute value of the eigenvalue in the same order and then accumulated to the corresponding commodity inherent structure node to obtain the commodity structure aggregation value corresponding to the natural day. Since the eigenvectors have variable signs, the absolute values ​​of their components are uniformly taken. When multiple eigenvalues ​​are the same, resulting in non-unique order of eigenvectors, the absolute values ​​of the components within the eigenvector group corresponding to the same eigenvalue are first accumulated and then multiplied by the absolute value of the corresponding eigenvalue. The output is a daily commodity structure change table containing the natural day, the commodity inherent structure number, and the commodity structure aggregation value for S44 to read. When the difference matrix is ​​a zero matrix, the commodity structure aggregation value of each commodity inherent structure node is written as zero. S44 is used to transform the daily commodity structure change table into change points and style conversion relationships for market analysis, so that subsequent candidate style features can read historical conversion ratios. In the daily commodity structure change table, the inherent structure number of the commodity is used as the grouping key to read the natural day and the commodity structure aggregation value respectively. For natural days under the same inherent structure number of the commodity, they are first sorted in descending order by the commodity structure aggregation value. If the commodity structure aggregation values ​​are the same, they are sorted in ascending order by the natural day. The natural day with the sorted number one is written as the market conversion change point corresponding to the inherent structure number of the commodity. Then, historical style conversion records with dates earlier than the market conversion change point are read. Using the same clothing category code as the statistical range, the edge weight of the temporary correction record number pointing to the inherent structure number of the product is accumulated as the conversion edge weight. The edge weight of all historical style conversion records within the same clothing category code is accumulated as the total edge weight of the category. The style conversion relationship is generated by dividing the conversion edge weight by the total edge weight of the category. When the total edge weight of the category is zero, no corresponding style conversion relationship is generated. When there is no non-zero product structure aggregation value at the market conversion change point, a no-change point mark is written. The style conversion relationship must include at least the temporary correction record number, the product inherent structure number, the clothing category code, the market conversion change point, the conversion edge weight, the total edge weight of the category, and the historical conversion ratio, for S5 to read when calculating the market gap value. Through the above processing, the daily style conversion network is first unified into the same node space, then converted into a daily Laplace matrix. Subsequently, the market structure changes on the inherent structure nodes of the product are extracted by the difference spectrum decomposition of the Laplace matrix of adjacent natural days. Finally, the market conversion change points and style conversion relationships corresponding to each product's inherent structure are formed. This processing enables subsequent predictions to not rely on the release volume of a single product, but to judge the market formation time point based on the network changes that converge towards the inherent structure of the product after temporary correction. In practical application: Multiple accounts generate temporary correction records of "cuffs rolled up" on different dates. Subsequently, the inherent short cuff structure begins to appear in the product images and text. S41 writes the temporary correction record number and the inherent short cuff structure number into the adjacency matrix of the same node. S42 generates a daily Laplace matrix. S43 obtains the natural day with the prominent product structure aggregation value on the inherent short cuff structure node. S44 writes this natural day as the market conversion change point of the inherent short cuff structure and generates the style conversion relationship of "temporary cuff correction to inherent short cuff structure" based on the historical edge weights before the change point.

[0024] S5. Extract temporary correction records that have not been converted into the inherent structure of the product before the prediction benchmark date as candidate style features. Calculate the historical conversion ratio based on the style conversion relationship. Multiply the number of corresponding accounts by the historical conversion ratio and then subtract the number of products with the same structure to obtain the market gap value. Output the prediction results of the fashion trend of middle-aged and elderly clothing in descending order of the market gap value. This implementation method is used to extract current candidate style features from records that have undergone temporary corrections before the forecast benchmark date but have not yet been incorporated into the product structure, and to calculate the market gap value by combining the market transformation change points and style transformation relationships output by S4. During processing, the temporary correction records are first encoded as candidate structure codes, then the historical transformation ratio of similar corrections to the inherent product structure is read from the historical transformation relationships, and then a trend ranking is formed by combining the current account adoption quantity and the product market supply quantity. It should be noted that the product record query scope used in S51 to determine "not converted into the inherent product structure" is the product image and text record index that has established historical style transformation records with the current temporary correction record, while the product record query scope used in S53 to calculate the market supply quantity is the entire product image and text record database before the forecast benchmark date; the two query scopes are different. This implementation process includes the following steps: S51 is used to filter out current correction objects that have not yet been taken into account by the inherent structure of the commodity from the temporary correction records before the forecast base date, so that subsequent market gap calculations only target style changes that are not fully supplied by the market. Based on the temporary correction record before the prediction benchmark date, read the clothing category code, structural part code, displacement direction code, termination position grid number, temporary correction record number and account identifier. Using the part code width and direction code width obtained by right-shifting the maximum code value in the coding table in S31, shift the clothing category code to the left by the part code width and write it into the structural part code. Then shift the direction code width to the left and write it into the displacement direction code to obtain the candidate structure code. Use the termination position grid number as the position field of the candidate style feature. Then, query the product image and text record index that has established historical style conversion records with the current temporary correction record, count the number of records where the product's inherent structure code matches the candidate structure code. When the number of records is zero, write the temporary correction record number, account identifier, candidate structure code, termination position grid number, and corresponding clothing category code into the candidate style feature; when the displacement direction code is missing, the current temporary correction record does not generate a candidate structure code, and when the product image and text record index does not match, the number of records is recorded as zero. S52 is used to map historical market transformation experience to current candidate style features, so that the speed at which similar temporary modifications in the past transform into the inherent structure of the product can be used to predict demand calculations; based on the market transformation change points output by S44, the style transformation relationship is read, and the historical edge weights whose entry point codes are consistent with the candidate structure codes are retrieved in the style transformation relationship. The date difference between the natural day where the historical edge weight is located and the market transformation change point is taken as the absolute value of the integer difference of natural days, and then one is added to the date difference to obtain the time conversion denominator. The historical transformation amount is obtained by dividing each historical edge weight by the corresponding time conversion denominator and summing them up. Within the same date range, read historical temporary correction records that match the entry point code and candidate structure code. Perform deduplication and counting on the account identifiers under the same natural day and the same entry point code to obtain the number of historical temporary correction accounts. Then, divide the historical conversion volume by the number of historical temporary correction accounts to obtain the historical conversion ratio, and write the historical conversion ratio into the candidate style features. When the number of historical temporary correction accounts is zero, the historical conversion ratio is written to zero and marked as having no historical samples. When there is no market conversion change point, the date difference is not included in the calculation and the historical conversion ratio is written to zero. S53 is used to convert the current social adoption scale and the product supply scale into a comparable market volume, so that the modification popularity in social images and texts can be aligned with the structural supply in product images and texts; the temporary modification records corresponding to the candidate style features are counted by deduplication according to the account identifier to obtain the current number of adoption accounts, and the predicted demand is obtained by multiplying the current number of adoption accounts by the historical conversion ratio obtained in S52. Then, read the product image and text records that match the inherent structure code and candidate structure code from the full database of product image and text records before the forecast benchmark date. After deduplication by product record number, count the product quantity and write the product quantity as the market supply. Forecasted demand is allowed to be a decimal, while market supply is an integer. When there is no product with the same structure in the entire database of product images and text records, the market supply is written as zero. Temporary correction records with missing account identifiers are counted for deduplication based on the temporary account identifier and the missing account marker is retained. S54 is used to convert the predicted demand and market supply into trend prediction results, so that each candidate style feature corresponds to a sortable market analysis value; the market gap value is obtained by subtracting the market supply from the predicted demand, and the candidate structure code, the termination position grid number, the number of currently adopted accounts, the historical conversion ratio, the predicted demand, the market supply, the market gap value, and the sorting number are written into the trend prediction results of middle-aged and elderly clothing. All candidate style features are sorted in descending order of market gap value. When the market gap values ​​are the same, they are sorted in ascending order of the market transformation change point date corresponding to the candidate style feature. Candidate style features without market transformation change point are ranked after candidate style features with market transformation change point. When the market gap value is negative, the corresponding candidate style feature is not deleted. Instead, the corresponding candidate style feature is written into the supply sufficient flag and continues to participate in the sorting. When the candidate style feature does not exist, an empty result flag is output. Through the above processing, temporary correction records that have not yet formed an inherent structure to support the product before the forecast benchmark date are converted into candidate style features. The candidate style features are then combined with the historical conversion ratio, the current number of adopted accounts, and the market supply of the product to obtain the market gap value, thereby outputting the trend forecast results for the analysis of the middle-aged and elderly clothing market. In practical application: Multiple middle-aged and elderly accounts showed temporary correction records of "cuffs rolled up" before the prediction benchmark date. S51: The garment category code, cuff part code, and upward direction code were written as candidate structure codes, and it was confirmed that the current temporary correction record had not yet established a conversion with the inherent cuff structure of the product. S52: The conversion relationship of similar cuff corrections to inherent short cuff structures in history was read and the historical conversion ratio was calculated. S53: The predicted demand was obtained by multiplying the current number of adopted accounts by the historical conversion ratio, and the number of products with inherent short cuff structures was counted from the entire database of product image and text records. S54: The market gap value was obtained by subtracting the market supply from the predicted demand value. When the market gap value ranked higher, the prediction results showed that the inherent short cuff structure had a subsequent fashion trend in middle-aged and elderly clothing.

[0025] Furthermore, the present invention also includes a system for predicting fashion trends in clothing for middle-aged and elderly people based on social media image and text data, comprising a data collection and sequencing module, a correction and identification module, a conversion and network building module, a change point analysis module, and a trend output module: The data collection and sequencing module is used to acquire social media image and text records of clothing for middle-aged and elderly people before the prediction benchmark date, form image and text time sequence according to account identifier and posting time, extract clothing structural boundaries and wearing actions that cause boundary displacement, and integrate product image and text records to generate clothing sample sequence; The correction and recognition module is used to generate a posture displacement field based on human key points for clothing areas with consistent texture fingerprints in the clothing sample sequence. It generates a clothing displacement field through the neural deformation pyramid registration algorithm, subtracts the posture displacement field from the clothing displacement field, retains the boundary displacements whose directions conform to the wearing action and pass through the action position, and generates a temporary correction record containing the start position and end position. The conversion network module is used to perform cross-industry data mining on temporary correction records and subsequent product image and text records. It registers the termination position to the product clothing area. When the termination position falls into the product structure boundary and the product text does not record the wearing action, it generates historical style conversion records and establishes a daily style conversion network with the temporary correction record pointing to the product's inherent structure and the number of conversion accounts as edge weights. The change point analysis module is used to construct a daily Laplace matrix through the Laplace dynamic network change point detection algorithm, subtract adjacent date matrices and perform eigenvalue decomposition, collect the product of the absolute value of the projection of the feature vector and the absolute value of the feature value according to the inherent structure of the product, write the date corresponding to the first digit of the collected value as the market transformation change point, and generate style transformation relationship; The trend output module is used to extract temporary correction records that have not been converted into the inherent structure of the product before the prediction benchmark date as candidate style features. Based on the style conversion relationship, the historical conversion ratio is calculated. The number of corresponding accounts is multiplied by the historical conversion ratio and then the number of products with the same structure is subtracted to obtain the market gap value. The prediction results of the fashion trend of middle-aged and elderly clothing are output in descending order of market gap value.

[0026] A storage device, comprising: A processor, and memory connected to the processor; memory is used to store computer programs.

[0027] A storage medium, comprising: The storage medium stores a computer program, which, when executed by a processor, performs the method.

[0028] Furthermore, the present invention also includes a prediction base date of 00:00 on March 1, 2026, and data sources including social media posts and images of clothing for middle-aged and elderly people that were publicly released before the prediction base date, and product posts and images in the clothing market database. The account identifiers in the social media posts are all de-identified account identifiers. The middle-aged and elderly attributes are derived from publicly available age group tags or content tags. The product posts are derived from product pages already published in the product market database. This implementation method uses the example of "after the cuffs of an upper garment are rolled up by multiple middle-aged and elderly wearers, a short cuff structure appears in subsequent products" to illustrate the complete processing process. First, obtain the top outfit photos and text records posted by the de-identified account A01 on January 5, 2026 and January 20, 2026. The text on January 20 included "Rolling up the cuffs makes it look neater." For two social media post records, calculate the image content hash value and the caption content hash value respectively. Then, XOR the image content hash value and the caption content hash value bitwise to obtain the content verification value. If there is a duplicate record with the same content verification value and the same posting time under the same account identifier, retain the record that is written first. After retaining the record and writing it into the post sequence number according to the posting time, the post sequence of the de-identified account A01 is formed. Subsequently, the upper garment area is enclosed based on the shoulder key points, elbow key points, wrist key points, and hip key points. The gray-level difference between adjacent pixels is calculated along the scan line within the upper garment area. Pixels where the gray-level difference sign changes to positive or negative are connected to form the cuff structure boundary. Then, using the clothing part word "cuff" corresponding to the cuff structure boundary as the search center, three words are read forward and backward in the matching text word segmentation sequence, and the action word "roll up" is matched. The action direction code is written as "towards the shoulder". The clothing category code is 3, the cuff structure code is 5, and the image and text sequence number is 2. The system generates a market primary key from the clothing category code and the structure code, and accesses the product image and text records belonging to the cuff structure of the upper garment before the prediction benchmark date from the product image and text records to obtain the clothing sample sequence. In the clothing sample sequence, the clothing area on January 5, 2026 is used as the source clothing area, and the clothing area on January 20, 2026 is used as the target clothing area. The texture fingerprints generated in the neighborhood of the cuff boundary of the two are consistent bit by bit. The system reads the same human key points from two images, constructs triangular meshes based on shoulder, elbow, wrist, and hip key points, and enumerates candidate parameters within the one-dimensional integer neighborhood of the affine parameters. In each round, it calculates the sum of squares of the reprojection residuals of the key points and the sum of squares of the affine differences of adjacent meshes, adds the two together to obtain the pose cost, and selects the candidate parameter with the ascending order number of the pose cost. After four rounds of enumeration, the affine parameters of all triangular meshes no longer change, and the pose displacement field is output. Subsequently, the source clothing region and the target clothing region after the attitude displacement field deformation are input into the neural deformation pyramid registration algorithm. The long side of the outer rectangle of the source clothing region is 240 pixels. After binary right shift counting, 8 layers of pyramid are obtained. Each layer uses the sampling point of the cuff structure boundary as the deformation neuron. Candidate displacements are enumerated in the integer pixel neighborhood with a Manhattan distance of no more than one. The registration cost is obtained by adding the texture Hamming distance, the bidirectional chamfer distance and the square of the displacement difference between adjacent neurons. When the sampling interval is reduced to one pixel and the displacement of all deformation neurons no longer changes after one round of update, the clothing displacement field is output. The system subtracts the attitude displacement field from the clothing displacement field to obtain the structural residual displacement field. It identifies that the cuff boundary moves from the local coordinate position (126, 340) to (126, 312). The dot product of this displacement direction and the "rolling up" action direction is positive. Furthermore, the reverse registration can return to the cuff hit pixel. Therefore, (126, 340) is written as the temporary correction start position and (126, 312) is written as the temporary correction end position, generating a temporary correction record R001. Next, cross-industry data mining was performed on the temporary correction record R001 and the subsequent product image and text records; the system read that the temporary correction was released on January 20, 2026, and extracted the product image and text records whose release time was later than January 20, 2026 but earlier than 00:00 on March 1, 2026. The width of the clothing category code, the width of the part code, and the width of the grid code are obtained by right-shifting the maximum code value of the corresponding coding table. In this embodiment, the width of the part code is 6 and the width of the grid code is 10. The ending position (126, 312) falls into grid number 184. The system generates the corrected market key 201912 according to the rule of "shifting the clothing category code to the left by the width of the part code and then writing the structural part code, then shifting the grid code width to the left and then writing the grid number of the ending position". Product image record P103 was published on February 15, 2026. The grid number of the cuff boundary in the product image is 184. The product market key 201912 is generated according to the same shift rule. Therefore, P103 enters the market acceptance candidate set. The system takes the center of the product area grid corresponding to grid number 184 as the projection starting point and enumerates candidate projection points along the normal of the product cuff boundary. For the candidate projection point with the sort number 1, the Hamming distance from the texture fingerprint to the candidate neighbor texture is temporarily corrected to 18, the chamfer distance from the candidate projection point to the product structure boundary is 11, the absolute value of the integer difference between the candidate projection point grid number and the termination position grid number is 0, and the product registration cost is 29. After the candidate projection points with the same number in two adjacent rounds of sorting are consistent, the product projection position is output; the pixel where the product projection position is located is bitwise ANDed with the pixel of the product structure boundary to obtain the boundary hit bit of 1. The word segmentation sequence of the product text does not hit "roll up" and synonymous action words, the action hit count is zero, the action disappearance bit is 1, the boundary hit bit and the action disappearance bit are multiplied by 1, so the cuff boundary in P103 is written as the inherent structure of the product G011, and the historical style conversion record H001 is generated; The system collects historical style conversion records by natural day, writes the temporary correction record number R001 into the edge start point, and writes the product's inherent structure number G011 into the edge end point. The account identifiers corresponding to the same edge start point and edge end point within the same natural day are deduplicated and counted, and then written as the edge weights to establish a daily style conversion network. The system then analyzes the daily style conversion network using the Laplace dynamic network change point detection algorithm; for the historical style conversion records from February 12, 2026 to February 16, 2026, multiple de-identified accounts formed a conversion edge of "temporary correction of cuff upward pointing to the inherent short cuff structure G011"; The system writes the temporary correction record number into the in-point dictionary in ascending order, writes the product's inherent structure number into the out-point dictionary in ascending order, and generates the same node adjacency matrix according to the concatenation order of the in-point dictionary and the out-point dictionary; the same edge weight is written to both the in-point to out-point position and the out-point to in-point position, and zero is written to the position where the edge is not hit. The degree matrix is ​​obtained by row-wise summation of the adjacency matrix of the same node for each natural day. Then, the daily Laplace matrix is ​​obtained by subtracting the adjacency matrix of the same node from the degree matrix. The daily Laplace matrix of adjacent natural days is formed by subtracting the previous day's matrix from the matrix of the next day. After eigenvalue decomposition, the system multiplies the absolute value of the component of each eigenvector at the commodity inherent structure node G011 by the absolute value of the same-order eigenvalue and sums them to obtain the commodity structure aggregation value of G011 for each natural day. Among them, the commodity structure aggregation value of G011 corresponding to February 16, 2026 is 7.50, and the sort number is one. The system writes February 16, 2026 as the market transformation change point of commodity inherent structure G011. Read the historical style conversion records before February 16, 2026, count the edge weights of temporary correction records pointing to G011 within the same clothing category and all edge weights within the same clothing category, and generate the style conversion relationship of "temporary correction of cuff upwards to inherent short cuff structure". Finally, in the temporary correction records before the prediction benchmark date, the system reads the "cuffs rolled up" temporary correction records that have not yet established historical style transformation records with the inherent structure of the product between February 20, 2026 and February 28, 2026, and generates candidate style features. The system shifts the upper garment category code 3 to the left by 6 code widths and writes it into the cuff structure code 5. Then, it shifts the direction code 3 to the left and writes the displacement direction code 2 "towards the shoulder", thus obtaining the candidate structure code 1578. The system also uses the termination position grid number 184 as the position field of the candidate style feature. For candidate structure code 1578, the system reads the historical edge weights with the same entry point code in the style conversion relationship: edge weight is 2 on February 12, 2026, edge weight is 4 on February 14, 2026, and edge weight is 3 on February 15, 2026; the market conversion change point is February 16, 2026. The three historical edge weights are divided by the absolute value of the date difference plus one, respectively, to obtain 0.40, 1.33 and 1.50, and the sum is 3.23 historical conversion amount. Within the same date range, the number of historical temporary correction accounts corresponding to entry point code 1578 is 8, and the historical conversion rate is 0.40; after deduplication, the account identifiers corresponding to the current candidate style features before the prediction benchmark date are 15, and the predicted demand is 15 multiplied by 0.40, resulting in 6.00; the number of product records with the same inherent structure code 1578 in the entire product image and text record database after deduplication by product record number is 2, the market supply is 2, and the market gap value is 6.00 minus 2, resulting in 4.00; The system outputs the trend prediction results of middle-aged and elderly clothing in descending order of market gap value. Among them, the "inherent short cuff structure with the cuff moving upward towards the shoulder and the grid number of the ending position being 184" is ranked as number one, indicating that this candidate style feature has a trend of entering the middle-aged and elderly clothing market after the prediction base date.

[0029] Working Principle: This solution first uses edge computing nodes to collect social media images and text related to clothing for middle-aged and elderly people before the prediction benchmark date. These images and texts are then organized into a timeline based on account and posting time. Clothing structural boundaries are extracted from the images, and wearing actions causing boundary changes are identified from the accompanying text. These are then integrated with product images and text to form a clothing sample sequence. Next, posture displacement is calculated using human keypoints, and the actual displacement of the clothing boundary is calculated using a neural deformation pyramid registration algorithm. The actual displacement is subtracted from the posture displacement to filter out temporary corrections consistent with the direction of the wearing action. Then, the temporary corrections output by the edge computing nodes are combined with subsequent product images and texts for cross-industry data mining to determine whether the user's temporary corrections have become part of the product's inherent structure, and a daily style transformation network is established. Finally, the Laplace dynamic network change point detection algorithm identifies the time points when temporary corrections are concentrated and transformed into product structures. The market gaps of candidate style features that are not yet fully covered by products are then calculated, and the prediction results for the fashion trends of clothing for middle-aged and elderly people are output. For example, when many middle-aged and elderly users post outfit photos on social media platforms, they roll up the cuffs of their shirts and express in the captions that the rolled-up cuffs make them look neater. The edge computing node will first exclude boundary movements caused by postures such as raising hands or turning around, and only retain temporary corrections to the cuff boundary that are indeed along the direction of the rolled-up cuffs. Then, it will search for product images and texts published after the prediction benchmark date and find that some products have already been made with short cuffs or shortened cuff structures, and the product captions no longer describe the action of rolling up the cuffs. It will then be considered that this temporary way of wearing clothes has been absorbed by the product structure. If a large number of accounts still show similar cuff corrections before the prediction benchmark date, and the number of products with the same structure is small, the solution will calculate a higher market gap value and rank the inherent short cuff structure at the top of the trend prediction results, indicating that this style has a future trend opportunity in the middle-aged and elderly clothing market.

[0030] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for predicting fashion trends in clothing for middle-aged and elderly people based on social media image and text data, characterized in that, include: S1. Obtain social media image and text records of middle-aged and elderly clothing before the prediction benchmark date, form an image and text time sequence according to account identifier and posting time, extract the clothing structural boundary and wearing actions that cause boundary displacement, and integrate the product image and text records to generate a clothing sample sequence. S2. For clothing regions with consistent texture fingerprints in the clothing sample sequence, generate a pose displacement field based on human key points, generate a clothing displacement field through the neural deformation pyramid registration algorithm, subtract the pose displacement field from the clothing displacement field, retain the boundary displacements whose directions conform to the wearing action and pass through the action position, and generate a temporary correction record containing the start position and end position. S3. Perform cross-industry data mining on temporary correction records and subsequent product image and text records, and register the termination position to the product clothing area. When the termination position falls into the product structure boundary and the product text does not record the wearing action, generate historical style conversion records, and establish a daily style conversion network with the temporary correction record pointing to the inherent structure of the product and the number of conversion accounts as the edge weight. S4. Construct a daily Laplace matrix using the Laplace dynamic network change point detection algorithm, subtract adjacent date matrices and perform eigenvalue decomposition, collect the product of the absolute value of the projection of the feature vector and the absolute value of the feature value according to the inherent structure of the product, write the date corresponding to the first digit of the collected value as the market transformation change point, and generate style transformation relationship; S5. Extract temporary correction records that have not been converted into the inherent structure of the product before the forecast benchmark date as candidate style features. Calculate the historical conversion ratio based on the style conversion relationship. Multiply the number of corresponding accounts by the historical conversion ratio and then subtract the number of products with the same structure to obtain the market gap value. Output the prediction results of the fashion trend of middle-aged and elderly clothing in descending order of market gap value.

2. The method for predicting fashion trends of middle-aged and elderly clothing based on social graphic data according to claim 1, characterized in that: S1 includes: S11. Using the account identifier as the aggregation key, read the social image and text records of middle-aged and elderly clothing before the prediction benchmark date. After hashing the image content, perform a bitwise XOR operation with the hash of the accompanying text content to obtain the content verification value. Records with the same content verification value and the same publication time are grouped as duplicates and the record with the first writing order in the group is retained. Then, the image and text order number is written according to the publication time as the retained record, and the account image and text time sequence is output. S12. Read social image and text records one by one in the account image and text time sequence, cut out the clothing area based on the closed outer contour of the human body key points, calculate the gray difference of adjacent pixels along the scan line in the clothing area and connect the pixels whose difference sign changes from positive to negative or from negative to positive as the clothing structure boundary, and then search for action words in the three adjacent words in the word segmentation sequence of the caption based on the clothing part words hit by the clothing structure boundary, and output the boundary action record. S13. After splicing the clothing category code and clothing part code in the boundary action record with the image and text sequence number, perform a hash operation to obtain the market collection key. Read the product records whose release time is earlier than the prediction base date and whose hash result matches the market collection key from the product image and text records. Continuate the product release time to the account image and text sequence and output the clothing sample sequence.

3. The method for predicting fashion trends in clothing for middle-aged and elderly people based on social media image and text data according to claim 2, characterized in that: S2 includes: S21. Read the source and target clothing regions with consistent texture fingerprints in the clothing sample sequence. Initialize the affine parameters of the triangular mesh with the coordinates of the human body key points. Enumerate candidate parameters in the one-dimensional integer neighborhood. Generate the attitude cost value by adding the sum of squared residuals of the key point reprojection to the sum of squared affine differences of adjacent meshes. Select the candidate parameter with the first position in ascending order of attitude cost value. Output the attitude displacement field when the affine parameters no longer change after one round of enumeration.

4. The method for predicting fashion trends of middle-aged and elderly clothing based on social graphic data according to claim 3, characterized in that: S2 also includes: S22. Using the deformed source clothing region and target clothing region as inputs to the neural deformation pyramid registration algorithm, generate the number of pyramid layers according to the number of binary bits of the long side of the bounding rectangle of the source clothing region. In each layer, the sampling points of the structural boundary are used as deformation neurons. The displacement of the neurons in this layer is initialized by the displacement interpolation of the previous layer. Candidate displacements are enumerated in the integer pixel neighborhood where the Manhattan distance does not exceed one. The registration cost is generated according to the texture Hamming distance, the bidirectional chamfer distance and the sum of squares of the displacement difference between adjacent neurons. The candidate displacement with the first position in ascending order of the registration cost is selected. The clothing displacement field is output when the sampling interval is one pixel and the displacement of all neurons no longer changes. S23. Subtract the attitude displacement field from the clothing displacement field to generate the structural residual displacement field. Read the structural residual displacement from the hit pixels of the wearing action on the structural boundary along the boundary chain. Retain the residual displacement chain that has a positive dot product with the wearing action direction and is registered back to the hit pixel in the reverse direction. When a residual displacement chain conflict occurs, retain the residual displacement chain that is first in ascending order of registration value. Write the first and last ends of the residual displacement chain along with the account identifier into the temporary correction record.

5. The method for predicting fashion trends in clothing for middle-aged and elderly people based on social graphic data according to claim 4, characterized in that: S3 includes: S31. Based on the temporary correction record, read the temporary correction release time, extract the product image and text records whose release time is greater than the temporary correction release time and less than the prediction base date, and write the clothing category code into the high-level market key when performing cross-industry data mining. Shift the high-level market key to the left by the part code width and write the structural part code, then shift the grid code width to the left and write the termination position grid number to obtain the corrected market key. Write the product boundary grid number into each product image and text record according to the same shift rule and generate the product market key. Filter the product image and text records whose product market key is equal to the corrected market key and output the market acceptance candidate set. S32. For the product image and text records in the market acceptance candidate set, take the temporary correction termination position as the projection starting point, enumerate candidate projection points along the boundary normal of the product clothing area, calculate the Hamming distance from the temporary correction texture fingerprint to the neighboring texture of the candidate projection point to generate the first cost, calculate the chamfer distance from the candidate projection point to the product structure boundary to generate the second cost, calculate the absolute value of the integer difference between the grid number of the candidate projection point and the grid number of the termination position to generate the third cost, and generate the product registration cost by adding the first cost, the second cost and the third cost, take the candidate projection point with the product registration cost sorting number as one, and output the product projection position when the candidate projection points with the sorting number as one in two adjacent rounds are consistent.

6. The method for predicting fashion trends of middle-aged and elderly clothing based on social graphic data according to claim 5, characterized in that: S3 also includes: S33. After obtaining the boundary hit position by bitwise ANDing between the pixel where the product projection position is located and the pixel of the product structure boundary, compare the wearing action word by word in the product text segmentation sequence and accumulate the number of action hits. When the number of action hits is zero, write the action disappearance position one. When the number of action hits is not zero, write the action disappearance position zero. When the boundary hit position multiplied by the action disappearance position equals one, write the product structure boundary as the inherent structure of the product and generate a historical style conversion record. S34. Collect historical style conversion records by natural day, write the temporary correction record number into the edge starting point, and write the product's inherent structure number into the edge ending point. Perform deduplication and count the account identifiers corresponding to the same edge starting point and the same edge ending point within the same natural day and write them as edge weights to establish a daily style conversion network.

7. The method for predicting fashion trends in clothing for middle-aged and elderly people based on social media image and text data according to claim 6, characterized in that: S4 includes: S41. Based on the daily style transformation network, read the edge weight of each natural day, write the temporary correction record number into the in-point dictionary, write the product inherent structure number into the out-point dictionary, use the concatenation order of the in-point dictionary and the out-point dictionary as the matrix row and column order, write the same edge weight into the in-point to out-point position and the out-point to in-point position, write zero for the missing position, and generate the same node adjacency matrix. S42. Accumulate the edge weights row by row in the adjacency matrix of the same node, write the accumulated row value to the diagonal position in the same order, and write the non-diagonal position to the opposite number of the edge weight at the same position in the adjacency matrix of the same node, to generate the daily Laplace matrix.

8. The method for predicting fashion trends of middle-aged and elderly clothing based on social graphic data according to claim 7, characterized in that: S4 also includes: S43. Using the Laplace dynamic network change point detection algorithm, read the Laplace matrix of two adjacent days in chronological order. Subtract the previous day's matrix from the matrix of the next day to generate a difference matrix. Perform eigenvalue decomposition on the difference matrix. Multiply the absolute value of the component of each eigenvector at the commodity inherent structure node by the absolute value of the same-order eigenvalue and accumulate it to the commodity inherent structure node. Output the daily commodity structure change table. S44. In the daily commodity structure change table, sort the dates according to the inherent structure nodes of the commodities. First, sort them in descending order by the aggregation value. If the aggregation values ​​are the same, sort them by the date that comes first. Write the date with the sort number as the market transformation change point. Read the historical style transformation records before the market transformation change point. Divide the edge weight of the temporary correction record within the same clothing category pointing to the inherent structure of the commodity by the total edge weight within the same clothing category to generate the style transformation relationship.

9. The method for predicting fashion trends of middle-aged and elderly clothing based on social graphic data according to claim 8, characterized in that: S5 includes: S51. Based on the temporary correction record before the prediction benchmark date, shift the clothing category code to the left by the part code width and write it into the structural part code, then shift the direction code width to the left and write it into the displacement direction code to obtain the candidate structure code. Query the number of records in the product image and text records that match the product inherent structure code and the candidate structure code. When the number of records is zero, write the temporary correction record as the candidate style feature. S52. Based on market conversion change points, read the historical edge weights in the style conversion relationship that are consistent with the entry point code and the candidate structure code. Divide each historical edge weight by the absolute value of the date difference between the corresponding natural day and the market conversion change point plus one, and sum them up to obtain the historical conversion volume. Then divide the historical conversion volume by the number of historical temporary correction accounts that are consistent with the entry point code and the candidate structure code to obtain the historical conversion ratio.

10. The method for predicting fashion trends of middle-aged and elderly clothing based on social graphic data according to claim 9, characterized in that: S5 also includes: S53. The number of currently adopted accounts is obtained by deduplicating the temporary correction records corresponding to the candidate style features according to the account identifier. The predicted demand is obtained by multiplying the current number of adopted accounts by the historical conversion ratio. Then, the number of products whose inherent structure code in the product image and text records is consistent with the candidate structure code is read and written as the market supply. S54. Obtain the market gap value by subtracting the market supply from the predicted demand. Sort the candidate style features in descending order of market gap value. When the market gap values ​​are the same, sort them in ascending order by the market transformation change point date corresponding to the candidate style features. Output the prediction result of the fashion trend of middle-aged and elderly clothing.

Images