Personalized costume design recommendation system and method based on AI and big data

Abstract

The invention discloses a personalized costume design recommendation system and method based on AI and big data, and relates to the technical field of personalized recommendation data processing, the system comprises an information acquisition module, a feature analysis module, a preference analysis module and a design recommendation module; the information acquisition module is used for acquiring user multi-modal information; the feature analysis module is used for extracting and fusing user preference features from the user multi-modal information and calculating a user multi-modal interest expression vector; the preference analysis module is used for carrying out joint modeling by combining a Transform network and a GRU structure based on the user multi-modal interest expression vector, and generating a user multi-dimensional dynamic preference vector through fusion of a gating mechanism; and the design recommendation module is used for outputting a personalized costume design recommendation result through a reinforcement learning model in combination with the multi-dimensional dynamic preference vector of the user.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of personalized recommendation data processing, and particularly relates to an AI and big data based personalized clothing design recommendation system and method. BACKGROUND

[0002] With the rapid development of information technology, the deep integration of big data and artificial intelligence technology has promoted the significant improvement of the data processing and analysis capability of personalized recommendation systems. In the field of recommendation systems, especially for the recommendation of commodities such as clothing with rich visual and style attributes, the core of data processing lies in how to effectively extract, integrate and utilize various data related to user preferences and commodity attributes. Currently, the data processing process of the recommendation system usually involves user behavior log analysis, commodity attribute structured processing and possible user portrait construction.

[0003] However, in the aspect of data processing of clothing recommendation, the existing technology still faces many challenges, and the main shortcomings are reflected in the globality and dynamics of data processing. When processing user data, the existing method often has the problem of insufficient modal fusion. User preference information is usually scattered in multiple modalities such as images (such as uploaded outfit photos), texts (such as comments, search records) and behaviors (such as click streams, purchase history). There are significant differences in feature dimensions, representation methods and semantic depth among various modal data. Traditional data processing procedures either use simple splicing or rely on specific modalities to dominate, making it difficult to achieve deep fusion and complementarity of cross-modal information, resulting in information redundancy or loss of key information in the generated user interest expression vector, which cannot fully and accurately depict user preferences. The existing data processing procedure has limited ability to model the dynamics of user preferences. The behavior sequence and interest expression of the user evolve over time, but most systems lack fine modeling and dynamic updating mechanisms for time series information during the data processing stage. For example, behavior data is often processed statically or only aggregated in a simple time series, failing to effectively capture short-term fluctuations and long-term trends in the behavior sequence, resulting in a lack of timeliness in the generated user portrait and the data processing result cannot reflect the current real interest state of the user. The objective function of data processing is often limited to short-term interaction indicators, lacking consideration of long-term user value. In the feature engineering and model training of the existing data processing procedure, the immediate feedback such as maximizing click rate or conversion rate is usually taken as the target, ignoring the long-term chain effects that may be produced by the recommendation behavior, such as the user's continuous exploration of series commodities and the improvement of brand awareness. This short-sighted data processing target leads to a recommendation strategy learned by the model that may not be optimal, which cannot effectively support recommendation decisions for long-term user value. SUMMARY

[0004] In view of the above existing problems, the present application is proposed.

[0005] Therefore, the present invention provides a personalized clothing design recommendation system based on AI and big data to solve the problems of insufficient multimodal user data fusion in existing clothing recommendation data processing methods, which leads to one-sided expression of user interests, and how to achieve closed-loop self-learning data processing for long-term value.

[0006] In order to solve the above technical problems, the present invention provides the following technical solutions:

[0007] In the first aspect, the present invention provides a personalized clothing design recommendation system based on AI and big data, which includes an information acquisition module, a feature analysis module, a preference analysis module, and a design recommendation module; the information acquisition module is used to obtain user multimodal information; the feature analysis module is used to extract and fuse user preference features from user multimodal information and calculate the user multimodal interest expression vector; the preference analysis module is used to perform joint modeling based on the user multimodal interest expression vector, jointly with the Transformer network and the GRU structure, and generate a user multidimensional dynamic preference vector through a gating mechanism; the design recommendation module is used to combine the user multidimensional dynamic preference vector and output a personalized clothing design recommendation result through a reinforcement learning model; the calculation of the user multimodal interest expression vector includes extracting global image features from the image modality and local texture features. Two independent networks are used to learn the global semantic information and local texture features of the image respectively, and the fusion coefficients of the image features are combined for weighted merger to form an accurate expression of the user's image preference. For the text modality, the BERT model that has been semantically fine-tuned in the clothing field is used to embed the text. Combined with the attention mechanism for emotional keywords, the weights of the text vectors are adjusted to make the emotional information and semantic content reflect the user's true preferences. The behavioral modality models the timestamp sequence of user behavior and uses the gated recurrent unit (GRU) to process the time series features to capture the dynamic trend of user behavior over time. The image modality feature encoding output, text modality feature encoding output, and behavioral modality feature encoding output are vectorized and mapped to the user's multimodal interest expression vector through a normalized mapping function.

[0008] As a preferred solution of the personalized clothing design recommendation system based on AI and big data described in the present invention, the acquisition of user multimodal information includes acquiring image modal information, text modal information and behavioral modal information.

[0009] As a preferred solution of the personalized clothing design recommendation system based on AI and big data described in the present invention, wherein: the calculation of the user's multimodal interest expression vector includes extracting the global semantic vector of the image through the ViT network of the image modal information, extracting the local texture features of the image through the CNN network and performing weighted fusion to obtain the image modal feature coding output; the text modal information outputs the semantic vector through BERT, and combines the attention weight distribution and linear projection weight matrix of the text emotional keywords to obtain the text modal feature coding output; the behavioral modal information is processed through the time vector embedding function, the timestamp is embedded and then input into the GRU to obtain the behavioral modal feature coding output; the image modal feature coding output, the text modal feature coding output and the behavioral modal feature coding output are spliced ​​and input into the normalization function to calculate the user's multimodal interest expression vector.

[0010] As a preferred solution of the personalized clothing design recommendation system based on AI and big data described in the present invention, the joint modeling includes cross-modal attention fusion of user multimodal interest expression vectors to construct a context fusion map; inputting the context fusion map into the Transformer structure to extract the user's long-term preference vector; and performing temporal modeling of user behavior based on the GRU structure to output the user's short-term interest vector.

[0011] As a preferred solution of the personalized clothing design recommendation system based on AI and big data described in the present invention, the generation of the user's multi-dimensional dynamic preference vector includes weighted fusion of the user's long-term preference vector extracted by the Transformer structure and the user's short-term interest vector output by the GRU structure through a gating mechanism.

[0012] As a preferred solution of the personalized clothing design recommendation system based on AI and big data described in the present invention, the output of personalized clothing design recommendation results includes inputting the user's multi-dimensional dynamic preference vector and the retrieval feature vector, style feature representation, trend vector and contextual environment information of each sample in the full set of candidate clothing samples into the recommendation function, and analyzing to obtain the optimal clothing design candidate sample.

[0013] As a preferred solution of the personalized clothing design recommendation system based on AI and big data described in the present invention, the output of personalized clothing design recommendation results also includes taking candidate clothing samples as recommendation actions in sequence; constructing a reward value based on the user's feedback behavior after each recommendation; using the current state, recommended action and reward value to update the Q-value function, and the Q-value function is iterated by maximizing future expected benefits; fusing contextual environment information with candidate clothing sample features to form an adaptation score, and using it together with dynamic preference correlation as the basis for recommendation sorting; selecting the candidate clothing sample with the largest expected value and the highest adaptation as the recommendation output, and continuously updating the recommendation strategy.

[0014] In the second aspect, the present invention provides a personalized clothing design recommendation method based on AI and big data, including obtaining user multimodal information; extracting and fusing user preference features from the user multimodal information, and calculating the user multimodal interest expression vector; based on the user multimodal interest expression vector, jointly modeling with the Transformer network and the GRU structure, and generating the user's multidimensional dynamic preference vector through a gating mechanism; combining the user's multidimensional dynamic preference vector, and outputting the personalized clothing design recommendation result through a reinforcement learning model.

[0015] In a third aspect, the present invention provides a computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: when the computer program is executed by the processor, any step of the personalized clothing design recommendation method based on AI and big data as described in the first aspect of the present invention is implemented.

[0016] In a fourth aspect, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein: when the computer program is executed by a processor, it implements any step of the personalized clothing design recommendation method based on AI and big data as described in the first aspect of the present invention.

[0017] The beneficial effects of the present invention are as follows: through the cross-modal attention mechanism and the joint modeling of GRU and Transformer, the long-term and short-term preferences of users can be accurately extracted, and weighted fusion is achieved through the gating mechanism, so that the user preferences are more dynamically and meticulously reflected in the recommendation, with higher adaptability and real-time performance. The design of the recommendation module is combined with the reinforcement learning model, and the self-learning and evolution of the personalized recommendation strategy can be achieved through the optimization of Q-value updates and context adaptation scores. Not only can it make real-time adjustments based on user feedback, but it can also dynamically respond to environmental factors such as weather, equipment, and geographic location to provide more personalized recommendation results that meet the user's current needs. There have been significant improvements in the precise characterization of user interests, the dynamic optimization of recommendation strategies, and the ability to adapt to the environment, which has improved the accuracy of recommendations and user satisfaction, and ultimately brought about a better personalized experience and commercial benefits. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0019] Figure 1 Schematic diagram of a personalized clothing design recommendation system based on AI and big data.

[0020] Figure 2 Flowchart of a recommended method for personalized clothing design based on AI and big data. DETAILED DESCRIPTION

[0021] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the specific embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0022] In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention may also be implemented in other ways different from those described herein. Those skilled in the art may make similar generalizations without violating the connotation of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0023] Secondly, the term "one embodiment" or "embodiment" herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in various places throughout this specification does not necessarily refer to the same embodiment, nor does it refer to a separate or selective embodiment that is mutually exclusive of other embodiments.

[0024] Reference Figure 1 This embodiment provides a personalized clothing design recommendation system based on AI and big data, including:

[0025] Information acquisition module, feature analysis module, preference analysis module, and design recommendation module.

[0026] The information acquisition module is used to obtain user multimodal information.

[0027] It should also be noted that obtaining user multimodal information includes obtaining image modal information, text modal information, and behavioral modal information.

[0028] It should also be noted that in a personalized recommendation system, it is difficult to fully express the user's preferences by relying solely on a certain modal input (such as click or search). Therefore, three types of heterogeneous data, namely image, text and behavior, are introduced, including image modal information I: clothing images browsed, collected and purchased by users; text modal information T: text content of user searches, comments and descriptive inputs; and behavioral modal information B, t: user temporal behavior of click, stay and purchase frequency.

[0029] The feature analysis module is used to extract and fuse user preference features from user multimodal information and calculate the user multimodal interest expression vector.

[0030] Furthermore, the calculation of the user's multimodal interest expression vector includes extracting the global semantic vector of the image through the ViT network of the image modal information, extracting the local texture features of the image through the CNN network and performing weighted fusion to obtain the image modal feature encoding output; the text modal information outputs the semantic vector through BERT, and combines the attention weight distribution and linear projection weight matrix of the text emotional keywords to obtain the text modal feature encoding output; the behavioral modal information is processed through the time vector embedding function, the timestamp embedding is processed and input into the GRU to obtain the behavioral modal feature encoding output; the image modal feature encoding output, the text modal feature encoding output and the behavioral modal feature encoding output are spliced ​​and input into the normalization function to calculate the user's multimodal interest expression vector.

[0031] It should also be noted that a preferred solution for calculating the user's multimodal interest expression vector specifically includes uniformly mapping the vector into a shared feature space after encoding through a dedicated sub-network, and calculating the user's multimodal interest expression vector, which is expressed as:

[0032]

[0033] M1(I)=α1·ViT(I)+(1-α1)·CNN(I)

[0034] M2(T)=BERT tuned (T)+W e ·SA(T)

[0035] M3(B,t)=GRU(Time2Vec(B,t))

[0036] Among them, U multi Represents the user's multimodal interest expression vector, which is a unified expression after the fusion of the three data modal features of image, text and behavior. Ψ is a normalized mapping function to ensure that the output vectors of different modalities are unified to the same dimension. M1(I), M2(T), and M3(B,t) are the encoding outputs of image modal features, text modal features, and behavioral modal features, respectively. t represents the timestamp sequence of user behavior, which records the time node of each user interaction operation. Represents the vector concatenation operation, which combines the vectors output by the three modalities in dimension to form a higher-dimensional user interest representation. α1 represents the image feature fusion coefficient, which is used to control the feature ratio of image transformer, ViT and convolutional neural network (CNN) in image encoding. ViT(I) represents the global image semantic vector extracted by the image transformer network for the image modality, emphasizing the overall style and high-level attributes of the image. CNN(I) represents the local texture features extracted by the convolutional neural network from the image modality. BERT tuned(T) represents the embedding output of the BERT model for text modality after semantic fine-tuning in the clothing field, which understands the clothing attributes and context expressed by users. SA(T) represents the attention weight distribution of text sentiment keywords, and W e The linear projection weight matrix representing the emotional attention vector is used to fuse the emotional vector with the original text semantics. GRU represents the gated recurrent unit, which captures the trends and patterns of user behavior over time. Time2Vec represents the conversion of timestamps in behavioral modalities into periodic vector features that can participate in neural network modeling.

[0037] It should also be noted that the joint extraction of image semantics based on the visual Transformer and CNN structure, the text modeling method combining BERT with the emotional attention mechanism, and the behavioral modeling process of time embedding GRU. By vector concatenating and normalizing the modal feature encoding, a user interest expression vector with a unified structure is formed. The semantic connection between visual preferences (style / texture), semantic preferences (description / comment), and temporal preferences (behavior sequence) is clearly established; semantic-level preference understanding is achieved without relying on predefined feature engineering, which can adapt to complex factors in real applications such as user potential interest migration, natural language style variation, and behavioral rhythm changes; it effectively overcomes the problem of insufficient adaptability of traditional recommendation systems in complex scenarios.

[0038] The preference analysis module is used to perform joint modeling based on the user's multimodal interest expression vector, combined with the Transformer network and GRU structure, and generate the user's multi-dimensional dynamic preference vector through gating mechanism fusion.

[0039] Furthermore, joint modeling includes cross-modal attention fusion of user multimodal interest expression vectors to construct a context fusion map; inputting the context fusion map into the Transformer structure to extract the user's long-term preference vector; and performing temporal modeling of user behavior based on the GRU structure to output the user's short-term interest vector.

[0040] It should also be noted that a preferred method for constructing a contextual fusion graph specifically includes inputting the output user multimodal interest expression vector into the cross-modal fusion module (CMAM), and combining long-term and short-term behavior patterns to jointly model the user's dynamic interest curve using Transformer and GRU. First, the contextual fusion graph is constructed through the cross-modal attention mechanism, and the unified feature vector is calculated, which is expressed as:

[0041]

[0042] Among them, Z represents the unified feature vector after fusion through the three-modal attention mechanism, m is the modality index variable, ranging from 1 to 3, representing image modality information, text modality information and behavioral modality information, respectively, and is used to calculate the respective attention contributions of different modalities, β m Represents the attention weight coefficient of the mth modality, Attn m (.) means extracting context-related features from the mth modality, Q m The query vector (Query) of the mth modality reflects the semantic target direction of the current input, K m The key vector (Key) of the mth modality represents the context clues of all comparable contents, V m The value vector (Value) of the mth modality represents the representation content used to generate the output in the modality. The Query, Key, and Value vectors of all modalities are generated from the user multimodal interest expression vector U multi The extracted substructure.

[0043] It should also be noted that generating a user's multi-dimensional dynamic preference vector includes weighted fusion of the user's long-term preference vector extracted by the Transformer structure and the user's short-term interest vector output by the GRU structure through a gating mechanism.

[0044] It should also be noted that the context fusion graph is input into different path modeling structures to obtain the user's long-term preference vector H long and user short-term interest vector H short , expressed as:

[0045] H long =Transforfmer(Z)

[0046] H short =GRU(Time2Vec(Z t-n ,...,Z t ))

[0047] Among them, Transformer (.) is a neural network with a multi-layer self-attention structure, which is used to learn the semantic correlation between multiple historical moments and extract long-term preferences. GRU (.) is a gated recurrent unit structure, which encodes behavioral trajectories with time sequence and extracts the trend of local short-term interest changes. t Represents the unified eigenvector at time point t, Z t-n Represents the unified feature vector of the nth time window that is traced back to form the historical trajectory sequence of user preferences. Finally, GatinaNework is used to automatically adjust the user's long-term preference vector H long and user short-term interest vector H shortThe weights are integrated to generate the user's multi-dimensional dynamic preference vector, which is expressed as:

[0048] P dyn =Ω(ρ·H long +(1-ρ)·H short )

[0049] Among them, P dyn It is represented as a user's multi-dimensional dynamic preference vector, which can be deconstructed into feature dimensions such as style, color, season, and scene. It serves as the core driving vector of the recommendation system. Ω is a normalized mapping that outputs the dynamic user interest state. ρ represents the dynamic gating weight coefficient, which is used to determine the weighted fusion ratio of long-term preferences and short-term interests.

[0050] It should also be noted that by introducing a cross-modal attention fusion mechanism to construct a contextual fusion graph, the influence of each modality on user interest modeling in different situations is dynamically adjusted according to the weight. The long-term stable interests of users are then extracted through the Transformer structure, and the short-term behavioral trends are modeled by the GRU structure. Finally, the gating mechanism is used for fusion to output a dynamic preference representation for spatiotemporal changes. A time-sensitive, high-dimensional decomposition of a user's multi-dimensional interest expression structure is formed; the transformation of the user's interest state from static to dynamic is achieved. Compared with existing technologies that can only depict "current preferences", the present invention can predict "what they may like in the near future", effectively improving the adaptability of recommendations for new trend clothing and scene clothing (such as holiday and travel wear).

[0051] The design recommendation module is used to combine the user's multi-dimensional dynamic preference vector and output personalized clothing design recommendation results through the reinforcement learning model.

[0052] Furthermore, the output of personalized clothing design recommendation results includes inputting the user's multi-dimensional dynamic preference vector and the retrieval feature vector, style feature representation, trend vector and contextual environment information of each sample in the full set of candidate clothing samples into the recommendation function, and analyzing to obtain the optimal clothing design candidate sample.

[0053] It should be noted that outputting personalized clothing design recommendation results also includes taking candidate clothing samples as recommendation actions in turn; constructing a reward value based on the user's feedback behavior after each recommendation; using the current state, recommended action and reward value to update the Q-value function, and the Q-value function iterates by maximizing future expected benefits; fusing contextual environmental information with candidate clothing sample features to form an adaptation score, and using it together with dynamic preference correlation as the basis for recommendation sorting; selecting the candidate clothing sample with the largest expected value and the highest adaptation as the recommendation output, and continuously updating the recommendation strategy.

[0054] It should also be noted that a preferred solution for outputting personalized clothing design recommendation results through reinforcement learning model specifically includes constructing a recommendation function, for each candidate clothing sample Calculate the optimal clothing design candidate sample, expressed as:

[0055]

[0056] in, Indicates the optimal clothing design candidate sample that is finally recommended to the user, which is the recommended target selected and output in the current recommendation round. represents the full set of candidate clothing samples, which comes from a large-scale clothing design database or clothing knowledge graph. Represents the retrieval vector of the candidate clothing sample y, that is, the pre-encoded vector representation, which is used to calculate the similarity with the user preference vector. ∈ represents a small constant used to prevent division by zero errors and ensure the numerical stability of the calculation process. The value range of ∈ is 10 -6 to 10 -8 ζ is a hyperparameter, which represents the weight of the clothing style and fashion trend matching score in the final recommendation score. is the style feature representation of the candidate clothing sample y, F trend is the expression vector of the current overall popular trend, η(y,R t ) represents the feature vector R of clothing sample y in the current context t The adaptation score under is used to measure the influence of environmental factors on the recommended samples, and Q(s,a) represents the Q-value function.

[0057] The reward value is constructed based on the user's feedback behavior after each recommendation; the Q-value function is updated using the current state, recommended action, and reward value. The Q-value function iterates by maximizing the future expected return, which is expressed as:

[0058] Q(s,a)=E r [r+δ·max a' Q(s',a')]

[0059] Where s represents the current system state, including the user's dynamic preference vector, recent behavior label, and context (such as device, time, weather, etc.) information. a represents the current recommended action, specifically the action of pushing a candidate clothing sample y to the user. E r [.] represents the mathematical expectation of possible reward values, that is, the weighted sum of different user feedback behaviors. r represents the immediate reward, which quantifies the user's feedback on the current recommendation sample. For example, click gives 1 point, collection gives 2 points, and purchase gives 5 points. δ represents the discount factor, which controls the degree of influence of future rewards on the current recommendation strategy. max a'Q(s',a') represents the maximum future reward that can be obtained in state s', which is used to estimate the upper bound of long-term benefits and guide strategy convergence. s' represents the new state of the system after executing the recommended action a, that is, the updated behavior or interest of the user after viewing the recommended content. a' represents the set of all possible subsequent recommended actions in state s'.

[0060] It should also be noted that the purpose of the contextual adaptation function is to incorporate the user's actual contextual information (such as time, device, location, and weather) into the recommendation process, enabling the system to not only base recommendations on interests but also determine whether the recommendation is appropriate at this moment. Essentially, it is a contextual scorer, acting as a "filter" or "weighter" in recommendation decisions. By applying linear transformations and nonlinear mappings to the user's current context vector, combined with a learnable weighting structure, it outputs a continuous value representing the suitability of the recommended item in the current environment. For example, suppose a user is at 10 pm, using their phone, and the temperature is 30°C. The system originally intends to recommend a "padded down jacket." If the contextual adaptation function detects that: the time is late at night, the user prefers easy browsing over high-value conversions; the temperature is high, making a recommendation of winter clothing inappropriate; and the device is mobile, the user may not prefer complex visual styles. The contextual adaptation function outputs a low score close to 0, and the system can reduce the weight of the item in the ranking. Conversely, if the user is at 7 am in the north and the temperature is 5°C, the item will receive a higher suitability score, increasing its likelihood of being displayed. Most traditional systems ignore context or only use rule filtering (such as "block short sleeves in winter"); the present invention adopts a learnable context representation and automatically adjusts the weight of each context feature through an end-to-end network; it has continuous and fine-grained control capabilities to avoid information loss caused by hard filtering; it can adapt to complex multi-variable combination scenarios (such as "mobile phone + night + southern city"), which is extremely difficult to define in traditional recommendation logic.

[0061] The fitness score is calculated as:

[0062]

[0063] Among them, y represents the candidate clothing sample to be recommended, which comes from the candidate set in the system, R t The feature vector of the user's current context includes the access device type (such as mobile phone, PC), time period (such as daytime / nighttime), geographic location (such as city code), weather information (such as temperature, rainfall), holiday identifier, etc., γ T Represents a learnable contextual scoring bias vector, which is used to calibrate the scoring benchmark under different contexts. Different dimensions correspond to the default influence trend of different context variables. W r represents the learnable linear projection matrix, b rRepresents the bias term of the contextual scoring model, which is used to adjust the scoring baseline and model fitting ability.

[0064] It should also be noted that in order to make the contextual environment features quantified and applicable to neural network modeling, the system can embed various context variables in the following way:

[0065] Access device type v dev (such as mobile phones and PCs): Use one-hot encoding to construct classification vectors. For example, mobile phones are encoded as [1,0] and PCs are encoded as [0,1].

[0066] Time period v time (such as day and night): Divide into multiple time windows (such as morning, afternoon, and night) and perform one-hot encoding on each time window, such as morning (06:00–11:59): [1,0,0], afternoon (12:00–17:59): [0,1,0], and night (18:00–23:59): [0,0,1]. It can be expanded to 4 segments (including early morning) or use periodic embedding (such as Time2Vec) according to actual needs.

[0067] Geographic location loc (such as city code): Map the user location information to the city code, and then map it to a fixed-dimensional vector through the embedding layer, embedding matrix N represents the total number of city categories, d represents the output vector dimension corresponding to each city, represents the set of real numbers.

[0068] Weather Information weather (e.g., temperature, rainfall): Temperature is directly represented by a normalized value, and weather types such as rainfall are one-hot encoded as enumerations, such as [1,0,0] for sunny days and [0,1,0] for rainy days.

[0069] Holiday logo v holiday : Use a Boolean variable to mark whether the current time is a holiday, such as holiday = [1], weekday = [0].

[0070] The contextual features are finally concatenated into a unified contextual feature vector R t , the dimension is fixed and is used as input to model the context adaptation function, which is expressed as:

[0071] R t =[v dev ‖v time ‖v loc ‖v weather ‖v holiday ]

[0072] Among them, ‖ represents the vector concatenation operation.

[0073] It should also be noted that the reinforcement learning value function is used to estimate the total benefit that a certain recommendation behavior will ultimately bring. It not only considers whether the user "clicked on the current recommendation", but also considers whether this action leads the user to continue clicking, adding to a purchase, repurchasing, and other behaviors in the future. In other words, it evaluates "whether the current recommendation can form a behavioral chain." The Q-learning mechanism is used to combine immediate rewards with future reward predictions to update the strategy, so that the recommendation behavior changes from focusing only on the present to focusing on long-term effects. For example: User A receives a recommendation at 8 pm:

[0074] Recommendation A1: A popular item of the season is recommended. The user clicks on it but does not purchase it.

[0075] Recommendation A2: Recommend a regular item. The user clicks on it, saves it, returns the next day, and makes the purchase.

[0076] If we only look at the real-time clicks, the scores of the two are similar; however, through value function modeling, the system can learn:

[0077] Although the initial conversion rate of recommendation A2 is slow, there will be subsequent repeat purchases; the Q value corresponding to A2 has higher long-term benefits.

[0078] The system trains a preference strategy to prioritize A2-type products, thereby optimizing user lifetime value (LVL) rather than single clicks. Traditional systems are mostly static ranking and scoring models, such as collaborative filtering or shallow neural networks. This invention introduces the Q-value mechanism of reinforcement learning into clothing recommendations, enabling closed-loop self-learning of behavior, feedback, and strategy. It learns not only what users like but also what kinds of recommendations generate cascading feedback. It is particularly suitable for complex scenarios such as multi-round recommendations, series combination matching, and adjustment of recommendation strategies before and after holidays.

[0079] It should also be noted that a reinforcement learning Q-value mechanism and a contextual adaptation function are introduced into clothing design recommendations. The system constructs a recommendation scoring function encompassing four dimensions: retrieval similarity, style popularity matching, contextual adaptation, and the long-term benefits of behavioral feedback. The reinforcement learning value function is updated through user feedback behavior after each round of recommendations, enabling the self-evolution of the recommendation strategy. This constructs an intelligent recommendation mechanism that integrates immediate interest expression, environmental suitability, and long-term conversion potential. Compared to traditional recommendation systems that focus solely on immediate click conversions, the reinforcement learning of this invention can capture the user's long-term value behavior chain, achieving "future-oriented" recommendation decisions. At the same time, through the contextual adaptation function, the interference and incentives of environmental factors on recommendations are dynamically controlled, effectively avoiding the "inappropriate recommendations" problem of conventional systems. For example, it avoids recommending winter clothing in hot weather and avoids pushing complex patterned clothing at night, thereby improving user acceptance and recommendation hit rate.

[0080] Reference Figure 2, which is an embodiment of the present invention, further provides a personalized clothing design recommendation method based on AI and big data, comprising the following steps:

[0081] S1: Obtain user multimodal information.

[0082] S2: Extract and fuse user preference features from user multimodal information and calculate the user multimodal interest expression vector.

[0083] S3: Based on the user's multimodal interest expression vector, the Transformer network and GRU structure are jointly modeled and fused through a gating mechanism to generate a user's multi-dimensional dynamic preference vector.

[0084] S4: Combined with the user's multi-dimensional dynamic preference vector, the reinforcement learning model is used to output personalized clothing design recommendation results.

[0085] This embodiment also provides a computer device, which is suitable for the personalized clothing design recommendation method based on AI and big data, including: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute computer-executable instructions to implement the personalized clothing design recommendation method based on AI and big data proposed in the above embodiment.

[0086] The computer device may be a terminal, comprising a processor, a memory, a communication interface, a display screen and an input device connected via a system bus. The processor of the computer device is used to provide computing and control capabilities. The memory of the computer device comprises a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The communication interface of the computer device is used to communicate with an external terminal in a wired or wireless manner, and the wireless manner may be achieved through WIFI, an operator network, NFC (near field communication) or other technologies. The display screen of the computer device may be a liquid crystal display or an electronic ink display screen, and the input device of the computer device may be a touch layer covering the display screen, or a button, trackball or touchpad provided on the housing of the computer device, or an external keyboard, touchpad or mouse.

[0087] This embodiment also provides a storage medium having a computer program stored thereon, which, when executed by a processor, implements the personalized clothing design recommendation method based on AI and big data proposed in the above embodiment; the storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic memory, flash memory, disk or optical disk.

[0088] In summary, the present invention achieves personalized clothing design recommendations that are more accurate, dynamic, intelligent, and context-adaptive than traditional recommendation systems by deeply integrating multimodal user data, dynamically capturing changes in interests, introducing reinforcement learning to optimize long-term value, and considering real-time context, thereby more effectively improving user satisfaction and lifetime value.

[0089] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. A personalized clothing design recommendation system based on AI and big data, characterized by: include, Information acquisition module, feature analysis module, preference analysis module, design recommendation module; The information acquisition module is used to obtain user multimodal information; The feature analysis module is used to extract and fuse user preference features from user multimodal information and calculate the user multimodal interest expression vector; The preference analysis module is used to perform joint modeling based on the user's multimodal interest expression vector, combined with the Transformer network and GRU structure, and generate a user's multi-dimensional dynamic preference vector through gating mechanism fusion; The design recommendation module is used to combine the user's multi-dimensional dynamic preference vector and output personalized clothing design recommendation results through a reinforcement learning model; The calculation of the user's multimodal interest expression vector includes extracting global image features and local texture features from image modal information, learning the global semantic information and local texture features of the image through two independent networks, and combining them with the fusion coefficient of the image features for weighted merging to form an accurate expression of the user's image preference. For text modal information, the BERT model, which has been fine-tuned for semantics in the clothing domain, is used to embed the text. In combination with the attention mechanism for emotional keywords, the weight of the text vector is adjusted to ensure that the emotional information and semantic content reflect the user's true preferences. Behavioral modal information is captured by modeling the timestamp sequence of user behavior and processing the time series features with a gated recurrent unit (GRU) to capture the dynamic trend of user behavior over time. The image modal feature encoding output, text modal feature encoding output, and behavioral modal feature encoding output are vector-concatenated and mapped to the user's multimodal interest expression vector through a normalized mapping function.

2. The personalized clothing design recommendation system based on AI and big data according to claim 1, characterized in that: The acquiring of user multimodal information includes acquiring image modal information, text modal information, and behavioral modal information.

3. The personalized clothing design recommendation system based on AI and big data according to claim 2, characterized in that: The calculation of the user's multimodal interest expression vector includes extracting the global semantic vector of the image through the ViT network of the image modal information, extracting the local texture features of the image through the CNN network and performing weighted fusion to obtain the image modal feature encoding output; The text modality information is output through BERT as a semantic vector, and the text modality feature encoding output is obtained by combining the attention weight distribution and linear projection weight matrix of the text sentiment keywords; The behavioral modality information is processed through the time vector embedding function, and then the timestamp embedding is input into the GRU to obtain the behavioral modality feature encoding output; The image modality feature encoding output, text modality feature encoding output, and behavior modality feature encoding output are spliced ​​and input into the normalization function to calculate the user's multimodal interest expression vector.

4. The personalized clothing design recommendation system based on AI and big data according to claim 3, characterized in that: The joint modeling includes cross-modal attention fusion of user multimodal interest expression vectors to construct a context fusion graph; Input the context fusion graph into the Transformer structure to extract the user's long-term preference vector; Based on the GRU structure, the user behavior is temporally modeled to output the user's short-term interest vector.

5. The personalized clothing design recommendation system based on AI and big data according to claim 4, characterized in that: Generating the user's multi-dimensional dynamic preference vector includes weighted fusion of the user's long-term preference vector extracted by the Transformer structure and the user's short-term interest vector output by the GRU structure through a gating mechanism.

6. The personalized clothing design recommendation system based on AI and big data according to claim 5, characterized in that: The output of personalized clothing design recommendation results includes inputting the user's multidimensional dynamic preference vector and the retrieval feature vector, style feature representation, trend vector and contextual environment information of each sample in the entire set of candidate clothing samples into the recommendation function, and analyzing to obtain the optimal clothing design candidate sample.

7. The personalized clothing design recommendation system based on AI and big data according to claim 6, characterized in that: Outputting the personalized clothing design recommendation result further includes taking the candidate clothing samples as recommended actions in sequence; Construct reward values ​​based on user feedback after each recommendation; The Q-value function is updated using the current state, recommended action, and reward value. The Q-value function is iterated by maximizing the expected future benefit. The contextual environment information is integrated with the features of the candidate clothing samples to form an adaptation score, which is then used together with the dynamic preference correlation as the basis for recommendation ranking. The candidate clothing samples with the largest expected value and the highest fitness are selected as the recommendation output, and the recommendation strategy is continuously updated.

8. A personalized clothing design recommendation method based on AI and big data, based on the personalized clothing design recommendation system based on AI and big data according to any one of claims 1 to 7, characterized in that: include, Obtain user multimodal information; Extract and fuse user preference features from user multimodal information and calculate the user multimodal interest expression vector; Based on the user's multimodal interest expression vector, the Transformer network and GRU structure are combined for joint modeling, and the user's multi-dimensional dynamic preference vector is generated through fusion through a gating mechanism; Combined with the user's multi-dimensional dynamic preference vector, the reinforcement learning model is used to output personalized clothing design recommendation results.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the personalized clothing design recommendation method based on AI and big data described in claim 8 are implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the personalized clothing design recommendation method based on AI and big data described in claim 8 are implemented.

Images