Clothing recommendation method and system based on user preferences

Abstract

The application relates to the computer technical field and discloses a clothing recommendation method and system based on user preferences, which comprises the following steps: constructing a multi-dimensional utility target of user preferences, the target comprising at least one target or a plurality of combinations in clothing style matching degree, diversity index and novelty index; time evolution modeling of user recommendation preferences is carried out, a dynamic weight structure associated with user behavior feedback is established; a recommendation utility function is constructed and multi-target constraints are introduced, the nested optimization structure comprising outer layer recommendation target constraints and inner layer preference evolution constraints; a continuous evolution process of user preference weights is controlled by using partial differential equations; and user short-term behavior feedback and long-term behavior trend data are fused in the recommendation process. In the application, the multi-target recommendation utility function is introduced, and the style matching degree, diversity and novelty targets of the user are jointly modeled in the nested optimization structure, so that the clothing recommendation effect of simultaneously satisfying personalized and exploratory requirements is achieved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of computer technology, and in particular to a clothing recommendation method and system based on user preferences. BACKGROUND

[0002] With the rapid development of e-commerce platforms and mobile applications, clothing recommendation systems have become an important tool for users to obtain personalized product information. Traditional clothing recommendation methods mostly use collaborative filtering, content matching or shallow feature modeling methods, mainly focusing on improving the relevance and matching degree of recommendations. However, such methods often ignore the overall performance of the recommendation results in terms of style diversity, novelty and user exploration needs, leading to homogenization of recommended content, single user experience, and lack of ability to stimulate user interest.

[0003] Some existing improved methods attempt to introduce multi-objective optimization ideas, considering multiple target factors such as maximum click rate and product diversity improvement during the recommendation process. However, most of these methods fail to effectively coordinate the conflict between different targets, and the optimization process lacks structural constraints, easily falling into local optimization or target imbalance problems. In addition, the dependency and non-independence between recommendation targets are not systematically modeled, making it difficult to form an optimization mechanism with overall control ability. SUMMARY

[0004] To address the above shortcomings, the present application provides a clothing recommendation method and system based on user preferences, aiming to improve the existing clothing recommendation methods that fail to simultaneously consider both the user's personalized interests and the diversity and novelty of the recommendation results, resulting in high repetition rates of recommended content, single styles, and inability to stimulate users' desire for continuous exploration.

[0005] In a first aspect, the present application provides the following technical solution, a clothing recommendation method based on user preferences, comprising the following steps:

[0006] Constructing a multi-dimensional utility target of user preferences, the target including at least one of clothing style matching degree, diversity index and novelty index or a combination of multiple;

[0007] Modeling the time evolution of user recommendation preferences and establishing a dynamic weight structure associated with user behavior feedback;

[0008] Constructing a recommendation utility function and introducing multi-objective constraints, using a nested optimization structure for solution, the nested structure including outer recommendation target constraints and inner preference evolution constraints;

[0009] Using partial differential equations to control the continuous evolution process of user preference weights;

[0010] Fusing user short-term behavior feedback and long-term behavior trend data in the recommendation process;

[0011] Based on the trend data, a recommended set is generated, and a clothing recommendation result meeting each target constraint condition is output.

[0012] Preferably, the multi-dimensional utility target of the user preference includes the following three:

[0013] Based on the matching degree score of the clothing image features and the user label features;

[0014] The difference measurement of the clothing samples in the recommended set in the style, category, and color dimensions;

[0015] The semantic or vector space difference measurement between the current recommended clothing and the user's recent browsing or purchase records.

[0016] Preferably, the evolution process of the preference weight satisfies the following conditions:

[0017] The weight changes continuously over time;

[0018] The weight change rate is determined by the influence strength of the current recommendation target on the result, the user behavior feedback signal, and the change trend of the weight itself;

[0019] The system iteratively updates the weight evolution trajectory by analytical derivation or numerical integration methods.

[0020] Preferably, the outer layer constraint of the nested optimization structure includes the recommendation number control and the recommendation delay constraint, and the inner layer constraint includes the user preference evolution stability control term and the regularization term. The user behavior feedback variable is introduced into the recommendation preference weight update function, and the variable includes the click behavior, the browsing time, the purchase behavior, and is modeled according to two time scales of short-term feedback and long-term feedback.

[0021] Preferably, the method further includes: to judge the independence between different recommendation targets, the recommendation target function satisfies the orthogonality condition in the function space, and the similarity or correlation between different targets is constrained by the system to be close to zero;

[0022] After the independence judgment, the reference trajectory function is constructed, and the deviation amplitude and its derivative square between the user preference weight and the reference trajectory are integrated. When the integral value exceeds the stability threshold, the system initializes or suppresses the weight state.

[0023] Preferably, the method further includes: using the alternating direction multiplier method for solving, first iterating in the recommendation target and the preference evolution sub-problems respectively, then coupling the variables, and finally converging to the solution satisfying the optimality and the constraint consistency.

[0024] Preferably, the output of the recommendation set satisfies the optimality necessary condition, including the first-order variation of the recommendation utility function being zero, and the KKT condition of the existence and uniqueness of the optimal solution.

[0025] In a second aspect, the present application provides the following technical solutions, a clothing recommendation system based on user preferences, the system comprises:

[0026] A user behavior acquisition module is configured to acquire historical behavior data of a user.

[0027] A preference modeling module is configured to construct a multi-objective preference model of the user.

[0028] An optimization solving module is configured to establish a nested structure recommendation optimization function and perform numerical solving.

[0029] A preference weight evolution module is configured to control dynamic changes of the preference weight based on user behavior feedback.

[0030] A recommendation generation module is configured to output a clothing recommendation set satisfying user interests and having diversity and novelty according to trend data.

[0031] In a third aspect, the present application provides the following technical solutions, a computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the clothing recommendation method based on user preferences.

[0032] In a fourth aspect, the present application provides the following technical solutions, a readable storage medium having a computer program stored thereon, wherein the computer program is executable by a processor to implement the clothing recommendation method based on user preferences.

[0033] The present application has the following beneficial effects:

[0034] 1. In the present application, a multi-objective recommendation utility function is introduced, and the style matching degree, diversity and novelty of the user are jointly modeled in the nested optimization structure, so that the global optimal trade-off among the multi-dimensional performance indicators of the recommendation result is achieved, and the clothing recommendation effect of simultaneously satisfying the personalized and exploratory requirements is achieved.

[0035] 2. In the present application, a preference weight evolution model based on user behavior feedback driving is constructed, and a partial differential equation is used to control the dynamic change process, so that the continuous tracking of the user interest evolution over time is achieved, and the high responsiveness and long-term adaptability of the recommendation system to the user behavior changes are achieved.

[0036] 3、The KKT condition and the first-order variational derivation mechanism are introduced in the recommended optimization problem in the application, the recommended result meets the optimality and the feasibility constraint, the mathematical rigor and the explainability of the recommended output are achieved, and the system convergence and stability are significantly enhanced. BRIEF DESCRIPTION OF DRAWINGS

[0037] Figure 1 A method flowchart of a clothing recommendation method based on user preferences is provided in the application.

[0038] Figure 2 A module relationship architecture diagram of a clothing recommendation system based on user preferences is provided in the application. DETAILED DESCRIPTION

[0039] The technical solutions in the embodiments of the application will be described clearly and completely in combination with the drawings. Obviously, the described embodiments are only part of the embodiments of the application, rather than all the embodiments. Based on the embodiments in the application, all other embodiments obtained by those skilled in the art without creative labor fall within the protection scope of the application.

[0040] Embodiment one

[0041] Reference Figure 1 In the first embodiment of the application, the application provides a clothing recommendation method based on user preferences, including the following steps:

[0042] A multi-dimensional utility target of user preferences is constructed, and the target includes at least one target or a combination of multiple targets in clothing style matching degree, diversity index and novelty index;

[0043] The time evolution modeling of user recommendation preferences is performed, and a dynamic weight structure associated with user behavior feedback is established;

[0044] A recommendation utility function is constructed and multi-target constraints are introduced, and a nested optimization structure is used for solving, and the nested structure includes outer recommendation target constraints and inner preference evolution constraints;

[0045] The method further includes: to judge the independence between different recommendation targets, the recommendation target function meets the orthogonality condition in the function space, and the similarity or correlation between different targets is constrained by the system to be close to zero;

[0046] After the independence is judged, a reference trajectory function is constructed, and the deviation amplitude and its derivative square between the user preference weight and the reference trajectory are integrated, when the integral value exceeds the stable threshold, the system initializes or inhibits the weight state;

[0047] The partial differential equation is used to control the continuous evolution process of the user preference weight;

[0048] The method further comprises: solving by using an alternating direction multiplier method, first iterating in the recommendation target and the preference evolution sub-problems respectively, then coupling variables, and finally converging to a solution satisfying the consistency of optimality and constraints;

[0049] In the recommendation process, user short-term behavior feedback and long-term behavior trend data are fused;

[0050] Based on the trend data, a recommendation set is generated, and a clothing recommendation result meeting the constraint conditions of each target is output.

[0051] Specifically, the judgment of target independence is made through the following content: specifically, to judge the independence between different recommendation targets, the present embodiment defines the recommendation target function in the function space, and introduces the orthogonality condition as the judgment standard. The core target function in the recommendation system, such as matching , diversity , and novelty , should theoretically reflect complementary dimensions. The system evaluates the degree of independence of the target functions in the function space by calculating the inner product or correlation index between the target functions. If the standardized output of each target function in the current sample space is denoted as , the similarity between the targets can be expressed as:

[0052] ;

[0053] The system sets a threshold value , and forces the constraint Sim , that is, the recommendation targets are close to orthogonal in output performance, and the strategy directions provided by each sub-target do not overlap, avoiding the system from misjudging that the value of the same type of recommendation is repeatedly reinforced by multiple targets;

[0054] When the target independence is satisfied, the system further introduces a reference trajectory function to capture the natural evolution path of the user preference in multiple rounds of interaction. The trajectory function can be constructed based on the moving average or low-order fitting function of the weight change in the historical interaction window, reflecting the smooth trend of the expected preference structure. The deviation between the current preference weight trajectory and the reference trajectory is continuously monitored. To measure the deviation intensity, the system calculates the integral distance between the two, including the trajectory difference itself and its first derivative term:

[0055] ;

[0056] wherein is a derivative deviation adjustment factor that balances the influence of deviation amplitude and change rate. The integral result reflects whether the current strategy evolution deviates from the historical experience track, and is particularly suitable for identifying sudden strategy adjustment or feedback disturbance. When the integral result exceeds a preset stability threshold , the system considers that the strategy is at risk of instability. At this time, the weight state control mechanism is triggered, including the following two ways:

[0057] The first way resets the current preference weight to the current value or the average stable value of the reference trajectory, which is used to fall back to the controlled state;

[0058] The second way temporarily freezes the update step of the current weight component, or applies a high damping coefficient to suppress its change amplitude, which is used to protect and repair the recommended strategy in cases such as inconsistent multi-objective feedback information, external abnormal interference, and short-term user behavior fluctuations. Finally, through target orthogonality judgment, preference trajectory monitoring, and integral deviation constraint techniques, effective control of the structural independence and dynamic evolution stability of the multi-objective recommendation strategy is achieved;

[0059] Solving by alternating direction multiplier method, the following content:

[0060] In solving the multi-objective recommendation weight optimization problem, the alternating direction multiplier method is used as the core optimization algorithm framework to realize efficient joint solution of the two sub-problems of recommendation target and user preference evolution. The original optimization problem is a joint target problem with multiple constraint terms, which can be formalized as:

[0061]

[0062] Where:

[0063] : recommendation loss based on multi-objective functions (such as matching, novelty, etc.);

[0064] : trajectory deviation and regularization terms for constraining preference weight evolution process;

[0065] : including weight normalization, non-negativity, orthogonality approximation, trajectory deviation control, etc.

[0066] To reduce the coupling complexity of the overall problem, auxiliary variables are introduced, and the original problem is divided into two sub-problems:

[0067] Sub-problem 1: recommendation target minimization: taking as the optimization variable, focusing on the performance of the recommendation target;

[0068] Sub-problem 2: preference evolution constraint: taking as the optimization variable, responsible for controlling the trajectory stability and regularity, and introducing the constraint After that, the original problem is transformed into an alternately solvable form; ​

[0069] The solution process of the alternating direction multiplier method is as follows, let the Lagrange multiplier be , and the penalty factor be :1.

[0070] Initialization

[0071] Recommended target sub-problem (fixed ):

[0072] ;

[0073] Preference evolution sub-problem (fixed ):

[0074] ;

[0075] Lagrange multiplier update:

[0076] ;

[0077] In the iteration process, the optimization problem of each step is simpler in structure than the original problem, and has convexity or local convexity structure, so it is easy to solve. With the iteration advancing: the sub-variables gradually converge, the consistency constraints are gradually met; the Lagrange multiplier converges to the equilibrium point, and finally, the system converges to a set of solutions , which simultaneously satisfy the recommended target optimality and the evolution constraint consistency.

[0078] The multi-dimensional utility target of user preference includes the following three items:

[0079] The matching degree score based on the features of clothing images and user label features;

[0080] The difference measure of clothing samples in the recommended set in the style, category, and color dimensions;

[0081] The semantic or vector space difference measure between the current recommended clothing and the user's recent browsing or purchase records.

[0082] Specifically, first, the system extracts features from the user's static labels including gender, age, style preference, color preference, etc., and embeds them into vector representation through the feature coding module, denoted as . At the same time, the image information of all clothing samples to be recommended will extract the middle-level features through the visual neural network, and get the image feature vector . On this basis, the matching degree score function between the recommended set and the user label vector is defined as:

[0083] ;

[0084] The function represents the average similarity of each clothing in the recommended set and the user's static interest label in the semantic space, reflecting the degree of fit of the recommended content to the user's preference. The higher the score, the more the recommended set as a whole meets the user's explicit preference;

[0085] Define the difference function between any two clothes in the set in three dimensions including style, category, and color, respectively 、 、

[0086] Where , the three-dimensional difference is aggregated into the set diversity score function in a weighted combination way:

[0087]

[0088] Where, is the weight coefficient of style, category, and color, satisfying , which is used to control the contribution of each dimension to diversity. This function is used to measure the visual and semantic difference of the recommended set. The higher the score, the richer the structure of the set, which can reduce user visual fatigue and improve the freshness of the recommendation;

[0089] Finally, the difference between the current recommended set and the user's historical interest trajectory is measured. Let the user's recent interaction behavior set be , and each historical clothing sample has a corresponding image feature vector . The system defines the average semantic distance function between the recommended set and the historical set as:

[0090]

[0091] Where is the standard cosine similarity function. This score measures the semantic deviation between the current recommendation and past behavior. The higher the score, the greater the difference between the recommended clothing and the historical content in visual style, category characteristics, or color matching, i.e., the stronger the novelty. After integrating the above three objective functions, the system constructs a recommendation utility functional as a comprehensive evaluation index in a weighted manner for the final recommended set sorting and selection, defined as follows:

[0092]

[0093] Where, is the target weight vector of the system, satisfying the constraint ​​​​1, for controlling the relative importance among preference fitting, diversity, novelty. In specific implementation, the weight can be dynamically updated by user behavior characteristics, recent click density, repeated click rate, recent activity, etc. factors will jointly affect real-time adjustment.

[0094] The evolution process of preference weight satisfies the following conditions:

[0095] The weight changes continuously over time;

[0096] The weight change rate is determined by the influence strength of the current recommendation target on the result, user behavior feedback signal, and the change trend of the weight itself;

[0097] The system iteratively updates the weight evolution trajectory by analytical derivation or numerical integration method.

[0098] Specifically, by constructing a weight model that evolves over time, real-time adjustment of the recommendation target fusion structure is achieved. The weight is not only a static hyperparameter, but also a continuous variable with dynamic properties. Its change process is influenced by the recommendation result structure, user behavior feedback, and the historical state of the weight. In the recommendation process, the system will periodically calculate the performance of the recommendation set under multiple sub-goals;

[0099] The first factor is that the weight value at any time is a natural continuation of its previous state, there is no jump or reset phenomenon. When the system detects that the performance indicator of a sub-goal has a significant contribution to the overall recommendation quality, it will increase the weight corresponding to the goal. The degree of contribution is usually difficult to judge by experience, so the system needs to calculate the sensitivity of the recommendation utility function to each sub-goal. At this time, if the utility function is differentiable, introducing partial derivatives is necessary:

[0100] ;

[0101] Where, represents the current recommendation utility function, is the weight of the th recommendation target. This partial derivative represents the reaction strength of the recommendation performance to the current target weight. This partial derivative is not for gradient descent, but for sensing which part of the recommendation structure is playing a major role, i.e., how much trust to put in different targets;

[0102] The second factor comes from user behavior feedback. If the system finds that the diversity in the recommendation result decreases, the user bounce rate increases and the dwell time becomes shorter, it believes that diversity has an impact on user satisfaction, and thus increases its weight. This feedback signal does not have a fixed expression and can be any measurable behavior trend, including:

[0103] The click proportion of a target corresponding item increases by a certain amount;

[0104] Difference of page average dwell time from historical baseline;

[0105] Fast skip rate change of certain recommended content, these signals are included in the system through change rate including sliding window average or difference value, for dynamic adjustment of target focus;

[0106] The third factor If a certain weight is in a rising state for a long time, it means that the target is continuously preferred; on the contrary, if it fluctuates sharply, it may be that the feedback signal is unstable or the user behavior noise is large, so a mechanism is introduced to adjust the trend of the weight itself, that is, when the system finds that a certain weight increases rapidly within two periods, but the feedback signal does not increase synchronously, it will actively reduce its increase speed to keep the strategy stable. At this time, if the system resources allow, or the recommendation accuracy requirement is higher, the recommendation engine will use numerical integration method to calculate the weight value, through the explicit Euler method, its iterative form is as follows:

[0107] ;

[0108] Wherein, represents the target change rate of the first , which is composed of the weights of the above three driving factors. Whether to use integral update depends on the balance between response speed and calculation cost in the scene. If it is a high-frequency fluctuation scene of user behavior including short video stream, numerical integration is more robust; if the recommendation update period is longer, fixed step iteration can be used.

[0109] The outer constraints of the nested optimization structure include recommendation quantity control and recommendation delay constraint, and the inner constraints include user preference evolution stability control term and regularization term. The user behavior feedback variables are introduced into the recommendation preference weight update function, including click behavior, browsing time, purchase behavior, and are modeled according to two time scales of short-term feedback and long-term feedback.

[0110] Specifically, the evolution function of preference weight is the kernel part of nested structure optimization, and the recommendation quantity and delay control constraints are located in the outer layer. In each recommendation period, the system first generates a candidate recommendation set based on user context and behavior input, and then fuses and sorts the set according to the current recommendation preference weight , and combines the target score. The recommendation output must satisfy two outer constraints:

[0111] One is the recommendation quantity control constraint, which is used to limit the size of the recommendation set to prevent information overload due to too much content or reduce attractiveness due to insufficient content;

[0112] Two is the recommendation delay constraint, that is, the generation and refresh of the system recommendation strategy must be completed within an acceptable time window to ensure the immediacy of user interface response. The delay constraint can be expressed as:

[0113] ;

[0114] where, denotes the time consumed by the current policy computation and ranking execution, is the maximum response time set by the platform service level agreement, after satisfying the above outer-layer constraints, the system enters the inner-layer optimization of the nested structure, which contains two key sub-constraints:

[0115] First, the stability control item of user preference evolution does not allow the weight to mutate without significant behavior signal support, so the system introduces a smoothing item or change penalty item to limit the amplitude of weight variation in continuous time:

[0116] ;

[0117] This item can be directly added to the objective function or used as a constraint boundary to suppress rapid oscillation and maintain the natural evolution of the preference curve. The second is the regularization term, which is mainly used to prevent the weight from deviating to a single target for a long time, which destroys the balance of the multi-objective recommendation structure. Commonly used are entropy regularization or balance constraints:

[0118] ;

[0119] The recommendation weight evolves continuously over time, and the system uses numerical integration methods to iteratively update it to ensure that the recommendation strategy converges gradually in the continuous time domain. After each update, the weight vector is normalized to keep its sum equal to 1 and prevent it from deviating from the target space.

[0120] The output of the recommendation set needs to satisfy the necessary conditions for optimality, including the first-order variation of the recommendation utility function being zero, and satisfying the KKT conditions for the existence and uniqueness of the optimal solution.

[0121] Specifically, first, the recommendation set needs to satisfy the first-order variation of the recommendation utility function being zero to ensure that the current output is optimal under the multi-objective function. The recommendation utility function is usually a weighted combination of the objective functions, and its variation being zero means that under the current recommendation set, a small perturbation no longer brings utility improvement, thus constituting the first-order necessary condition for optimality;

[0122] Second, to further guarantee the existence and feasibility of the optimal solution, the system introduces the KKT condition as a constraint judgment standard, which includes the first-order derivative optimality, the original constraint feasibility including the recommendation quantity limit, the item legality, the weight normalization and non-negativity, the complementary relaxation and the non-negativity of the multiplier. Satisfying the KKT condition means that the current recommendation solution is a local optimal solution in the constraint domain;

[0123] Finally, in the solving process, the system adopts the alternating direction multiplier method for optimization, decomposes and iterates the recommended target and the preference weight evolution problem. After optimization of each sub-problem, the variables are coupled and the Lagrange multiplier is updated to realize the gradual convergence of the recommended variables to the optimal solution under the premise of meeting the constraint consistency.

[0124] Embodiment two:

[0125] With reference to Figure 2 In the second embodiment of the present application, the present application provides a clothing recommendation system based on user preferences, which comprises:

[0126] a user behavior collection module for collecting historical behavior data of the user;

[0127] a preference modeling module for constructing a multi-objective preference model of the user;

[0128] an optimization solving module for establishing a nested structure of a recommendation optimization function and performing numerical solving;

[0129] a preference weight evolution module for controlling the dynamic change of the preference weight based on user behavior feedback;

[0130] a recommendation generation module for outputting a set of clothing recommendations that meet the user's interest and have diversity and novelty based on trend data.

[0131] Specifically, the user behavior collection module is used to collect historical behavior data of the user, including click records, browsing time, collection, purchase behavior, and interactive information such as bounce rate. This module classifies and stores the behavior data into short-term and long-term, and extracts key behavior features as the basis for subsequent preference modeling and dynamic feedback update;

[0132] The preference modeling module receives the feature data output by the user behavior collection module, constructs a multi-objective preference model of the user, represents the user's preference tendency in multiple dimensions such as matching, diversity, novelty, and conversion potential, and outputs the result in a vectorized structure as an input feature of the recommendation optimization module;

[0133] The optimization solving module receives the multi-objective preference model provided by the preference modeling module, and establishes a nested structure of a recommendation optimization function in combination with the recommendation target and constraint conditions configured by the system. The alternating direction multiplier method is used to iteratively optimize the recommended target and the preference evolution sub-problem, complete the variable coupling, solve the recommended result that meets the first-order variation condition and the KKT optimality condition, and output it to the preference weight evolution module for weight adjustment;

[0134] The preference weight evolution module takes the current recommendation effect output by the optimization solving module and real-time behavior feedback of the user as input, dynamically adjusts the evolution track of the preference weight, the feedback behavior includes short-term click fluctuation and long-term conversion trend, and the updated preference weight result is returned to the optimization solving module through the introduction of a track stability control mechanism and an orthogonal constraint;

[0135] The recommendation generation module generates a clothing recommendation set meeting the user interest structure based on the final convergence result of the optimization solving module. The recommendation result takes into account matching, diversity and novelty, covers the richness of clothing in multiple dimensions such as quality, style and category, and improves the user satisfaction and exploratory experience of the recommendation as a whole.

[0136] Embodiment three

[0137] The third embodiment of the present application is based on the same inventive concept, and the present application provides a computer readable storage medium storing a computer program, which is executed by a processor to implement the steps of the clothing recommendation method based on user preferences of the above-mentioned embodiments.

[0138] Embodiment four

[0139] The fourth embodiment of the present application is based on the same inventive concept, and the present application provides a computer device, which includes a processor and a memory. The processor and the memory communicate with each other. The memory is used to store instructions, and the processor is used to execute the instructions in the memory to implement the clothing recommendation method based on user preferences of the above-mentioned embodiments.

[0140] It should be understood that parts of the present application can be realized by hardware, software, firmware or a combination thereof. In the above-mentioned embodiments, a plurality of steps or methods can be realized by software or firmware stored in a memory and executed by a suitable instruction execution system. For example, if realized by hardware, and as in another embodiment, it can be realized by any one or a combination of the following technologies known in the art: discrete logic circuit with logic gate circuit for implementing logic function on data signal, application specific integrated circuit with suitable combination logic gate circuit, programmable gate array (PGA), field programmable gate array (FPGA) and the like.

[0141] Finally, it should be noted that: the above-mentioned only for the preferred embodiments of the present application, and not for limiting the present application, although the present application has been described in detail with reference to the foregoing embodiments, for those skilled in the art, the technical solutions recorded in the foregoing embodiments can still be modified, or some technical features can be replaced by equivalent, any modification, equivalent replacement, improvement, etc. within the spirit and principles of the present application, should be included in the protection scope of the present application.

Claims

1. A clothing recommendation method based on user preferences, characterized in that, Includes the following steps: The user behavior collection module collects users' historical behavior data, classifies and stores the behavior data into short-term and long-term categories, and extracts key behavioral features as the basic input for subsequent preference modeling and dynamic feedback updates. The preference modeling module receives feature data output by the user behavior collection module and constructs a multi-objective preference model for the user. The multi-objective preference model includes a matching score based on clothing image features and user tag features, a measure of the differences between clothing samples in the recommendation set in terms of style, category, and color, and a measure of the semantic or vector space differences between the currently recommended clothing and the user's recent browsing or purchase records. The results are represented in a vectorized structure and used as input features for the recommendation optimization module. The optimization solution module receives the multi-objective preference model provided by the preference modeling module, and combines it with the recommended objectives and constraints configured by the system to establish a nested optimization function. It uses the alternating direction multiplier method to iteratively optimize the recommended objectives and preference evolution subproblems, completes variable coupling, and solves the recommendation results that satisfy the first-order variational conditions and KKT optimality conditions. At the same time, it outputs the results to the preference weight evolution module for weight adjustment. The preference weight evolution module takes the current recommendation effect output by the optimization solution module and the real-time user behavior feedback as input, and dynamically adjusts the evolution trajectory of the preference weight. The feedback behavior includes short-term click fluctuations and long-term conversion trends. By introducing a trajectory stability control mechanism and orthogonality constraints, the updated preference weight results are fed back to the optimization solution module. The recommendation generation module generates a set of clothing recommendations that meet the user's interest structure based on the final convergence result of the optimization solution module. The outer constraints of the nested optimization structure include recommendation quantity control and recommendation delay constraints, while the inner constraints include user preference evolution stability control terms and regularization terms. The recommendation preference weight update function introduces user behavior feedback variables, including click behavior, browsing duration, and purchase behavior, and models them separately according to two time scales: short-term feedback and long-term feedback.

2. The clothing recommendation method based on user preferences according to claim 1, characterized in that, The evolution of the preference weights satisfies the following conditions: The weights change continuously over time; The rate of change of weights is determined by the strength of the current recommendation target's influence on the results, user behavior feedback signals, and the trend of weight changes themselves. The system iteratively updates the weight evolution trajectory using analytical differentiation or numerical integration methods.

3. A clothing recommendation system based on user preferences, characterized in that, The system, used to implement the user preference-based clothing recommendation method as described in claim 1 or 2, comprises: The user behavior collection module is used to collect users' historical behavior data; it categorizes and stores the behavior data into short-term and long-term categories, and extracts key behavioral features as the basic input for subsequent preference modeling and dynamic feedback updates; The preference modeling module receives feature data output from the user behavior collection module and uses it to construct a multi-objective user preference model; the results are represented in a vectorized structure and used as input features for the recommendation optimization module. The optimization and solution module is used to establish a recommended optimization function with a nested optimization structure and perform numerical solutions. It receives a multi-objective preference model provided by the preference modeling module, and combines it with the recommended objectives and constraints configured by the system to establish a recommended optimization function with a nested optimization structure. It uses the alternating direction multiplier method to iteratively optimize the recommended objectives and preference evolution subproblems, completes variable coupling, and solves for the recommended results that satisfy the first-order variational conditions and KKT optimality conditions. At the same time, it outputs the results to the preference weight evolution module for weight adjustment. The preference weight evolution module is used to control the dynamic changes of preference weights based on user behavior feedback. It takes the current recommendation effect output by the optimization solution module and real-time user behavior feedback as input to dynamically adjust the evolution trajectory of preference weights. The recommendation generation module is used to output a set of clothing recommendations that meet user interests and are diverse and novel based on trend data; based on the final convergence result of the optimization solution module, a set of clothing recommendations that meet the user's interest structure is generated.

4. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements a clothing recommendation method based on user preferences as described in claim 1 or 2.

5. A readable storage medium, characterized in that, The readable storage medium stores a computer program that, when executed by a processor, implements a user preference-based clothing recommendation method as described in claim 1 or 2.

Images