E-commerce personalized recommendation service method based on genetic fuzzy clustering

Abstract

The invention relates to an e-commerce personalized recommendation service method based on genetic fuzzy clustering. Comprising the steps of dividing commodities according to the color richness from low to high, performing analysis to obtain a preference relationship between a user group and the color richness of the commodities, establishing a preference prediction model according to the preference relationship to obtain a preference feature vector of the user group, and obtaining a preference feature vector of the user group according to behavior data of the user group and the preference feature vector. The difference and similarity of different user groups in color richness are analyzed, the user groups are divided into multiple categories of users according to the color richness, the color preference characteristics of the users of each category are obtained, and according to the color preference characteristics and the commodities corresponding to the corresponding color richness, the commodities corresponding to the users of each category are obtained. Generating a commodity recommendation list based on color richness, and dynamically adjusting the recommendation list according to feedback information of each category of users on the recommendation list; the problem that the experience feeling is poor in the process of selecting the required commodity by the user can be effectively solved.

Description

Technical Field

[0001] The present invention relates to the field of information technology, and in particular to an e-commerce personalized recommendation service method based on genetic fuzzy clustering. Background Art

[0002] On e-commerce platforms, the number of colors of products has a significant impact on users' purchasing intention. Products with a rich number of colors are more likely to attract users who prefer multi-color products, but for users who prefer monochrome or low-color-rich products, too many colors may cause visual fatigue and confusion, reducing their willingness to buy.

[0003] Existing intelligent recommendation solutions for e-commerce products rarely consider the psychological implications of color on users, resulting in a poor user experience when selecting the products they need. Summary of the Invention

[0004] Based on this, it is necessary to provide an e-commerce personalized recommendation service method based on genetic fuzzy clustering to address the above technical problems.

[0005] This application provides an e-commerce personalized recommendation service method based on genetic fuzzy clustering, which includes:

[0006] Convert the acquired and loaded product images corresponding to each product into RGB color space, extract the colors of the product images, and classify the corresponding products into monochrome, low color richness, medium color richness, and high color richness according to the number of colors in the product images, from low to high in terms of color richness;

[0007] Based on the basic information data of each user group and the corresponding behavioral data of the user group, the preference relationship between the user group and the color richness of the product is analyzed, and a preference prediction model is established based on the preference relationship to obtain the preference feature vector of the user group; the user group includes users of different genders, occupations and ages; the behavioral data includes the user group's browsing history, purchase history and collection history of the product;

[0008] Based on the behavioral data and preference feature vectors of user groups, a fuzzy clustering algorithm is used to analyze the differences and similarities in color richness among different user groups, and a genetic algorithm is used to optimize the fuzzy clustering algorithm.

[0009] Based on the analysis results of the optimized fuzzy clustering algorithm, the user group is divided into multiple categories according to color richness, and the color preference characteristics of users in each category are obtained. Based on the color preference characteristics and the corresponding color richness corresponding products, a recommendation list of products based on color richness is generated;

[0010] Collect and dynamically adjust the recommendation list based on the feedback information of users in each category to optimize the color preference characteristics.

[0011] The above-mentioned e-commerce personalized recommendation service method based on genetic fuzzy clustering can first divide the products into multiple color richness categories by converting them into RGB color space, then obtain the behavioral data of each user group to obtain the preference relationship between user behavior and color richness, establish a prediction model based on the preference relationship, convert user preferences into feature vectors to obtain structured data, and then further divide the user groups according to categories through the fuzzy clustering algorithm optimized by the genetic algorithm, enhance the targetedness of user groups, and associate them with color richness categories, finally generate a recommendation list based on the association results, recommend products to user groups in a targeted manner, and update the recommendation list in real time through dynamic adjustment to maintain the timeliness of the recommendation list and improve the accuracy of the recommendation and user satisfaction, effectively solving the problem that the existing e-commerce product intelligent recommendation solutions rarely start from the perspective of the psychological implications of color on users, resulting in a poor user experience in the process of selecting the required products. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] In order to more clearly illustrate the technical solutions in the embodiments of the present application or related technologies, the following briefly introduces the drawings required for use in the embodiments or related technical descriptions. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0013] Figure 1 The figure is a flowchart of a method for personalized e-commerce recommendation service based on genetic fuzzy clustering in one embodiment. DETAILED DESCRIPTION

[0014] In order to make the purpose, technical solutions and advantages of this application more clear, the following further describes this application in detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain this application and are not intended to limit this application.

[0015] In the description of this application, the terms "first" and "second" are used for descriptive purposes only and should not be understood to indicate or imply relative importance or implicitly specify the number of the technical features indicated. Therefore, a feature defined as "first" or "second" may explicitly or implicitly include at least one of such features. In the description of this application, "plurality" means at least two, for example, two, three, etc., unless otherwise specifically defined.

[0016] In this application, unless otherwise specified or limited, the terms "initial," "connected," "connect," "fixed," and the like should be understood in a broad sense. For example, they may refer to fixed connection, detachable connection, or integration; mechanical connection or electrical connection; direct connection or indirect connection through an intermediate medium; internal communication between two elements or interaction between two elements, unless otherwise specified. Those skilled in the art will understand the specific meanings of the above terms in this application based on the specific circumstances.

[0017] It should be noted that when an element is considered to be "connected" to another element, it can be directly connected to the other element or connected to the other element through an intervening element. In addition, the "connection" in the following embodiments should be understood as "electrical connection", "communication connection", etc., if there is transmission of electrical signals or data between the connected objects.

[0018] Users of different occupations, genders, and ages have different preferences for color quantity. Therefore, how to recommend appropriate color quantity products based on users' personal characteristics has become a technical challenge that e-commerce platforms need to address. Building a correlation model between gender, occupation, age, and product color quantity preferences based on the collection of basic user information and behavioral data is key to achieving personalized color quantity product recommendations. However, during the modeling process, how to select an appropriate algorithm and optimize its performance to achieve efficient classification of users' color quantity preferences is a technical challenge that needs to be addressed. Furthermore, after the model is established, how to use historical data and feedback to verify and adjust the model to ensure the accuracy of recommendations is also an issue that requires further exploration. Furthermore, in the actual recommendation process, how to design a reasonable color quantity grading scheme and provide corresponding product recommendations based on the grading is also a technical issue worth considering.

[0019] In an exemplary embodiment, Figure 1 As shown, a method for e-commerce personalized recommendation service based on genetic fuzzy clustering is provided, including the following steps 202 to 210. Among them:

[0020] In step 202, the obtained and loaded product images corresponding to the products are converted into RGB color space, the colors of the product images are extracted using a fuzzy clustering algorithm, and the corresponding products are divided into monochrome, low color richness, medium color richness, and high color richness according to the number of colors in the product images.

[0021] Specifically, first load the product image and convert it into RGB color space. According to business needs and data analysis, set the threshold for color richness classification. Select the threshold based on product type and color preference factors to ensure the rationality and practicality of the classification results.

[0022] In an exemplary embodiment, the steps of extracting the colors of product images using a fuzzy clustering algorithm and classifying the corresponding products into monochrome, low color richness, medium color richness, and high color richness according to the number of colors in the product images, from low to high color richness, include:

[0023] The converted RGB color space is processed using a fuzzy clustering algorithm; the fuzzy clustering algorithm is a k-means fuzzy clustering algorithm; the data points in the RGB color space are assigned to K cluster centers in an iterative manner, and the distance from each data point to the cluster center is adjusted to be the minimum;

[0024] Cluster the pixels of the product image into several representative colors to obtain the color distribution of the product image with respect to the representative colors;

[0025] Determine that when the number of colors in the color distribution is below a first preset threshold, the color richness of the product is monochrome; when the number of colors in the color distribution exceeds the first preset threshold and is below a second preset threshold, the color richness of the product is low color richness; when the number of colors in the color distribution exceeds the second preset threshold and is below a third preset threshold, the color richness of the product is medium color richness; when the number of colors in the color distribution exceeds the third preset threshold, the color richness of the product is high color richness; wherein the first preset threshold, the second preset threshold and the third preset threshold are all set according to pre-stored business needs and business data.

[0026] During implementation, you can use an image processing library such as OpenCV to load product images. You can then use the cvtColor function in OpenCV to convert the image from the BGR color space (or other color space) to the RGB color space. Then, based on business needs and data analysis, you can set thresholds for color richness classification. For example, the first preset threshold can be set to 5, the second preset threshold can be set to 10, and the third preset threshold can be set to 20. A value greater than 20 is defined as high color richness. The first, second, and third preset thresholds can be adjusted based on actual conditions.

[0027] Furthermore, during the color clustering process, the K-means algorithm can be used to cluster pixels in the RGB color space by setting the number of cluster centers K to 10, the number of iterations to 100, and the initial center point selection method to random selection. The clustering results can be obtained by calculating the number of pixels in each cluster center to obtain several representative colors in the image.

[0028] Specifically, the number of representative colors may be set to the number of primary colors of the seven-color light.

[0029] To further evaluate the classification results, 100 product images can be randomly sampled and manually judged by multiple professional designers. The results are then compared with the algorithm's classification results. Metrics such as accuracy and recall are calculated. If accuracy exceeds 90% and recall exceeds 95%, the color richness classification method is considered feasible.

[0030] In step 204, based on the basic information data of each user group and the corresponding behavioral data of the user group, an association rule mining algorithm is used to analyze the preference relationship between the user group and the color richness of the product, and a preference prediction model is established based on the preference relationship to obtain the preference feature vector of the user group; the user group includes users of different genders, occupations and ages; the behavioral data includes the user group's browsing history, purchase history and collection history for the product.

[0031] Specifically, based on the collected basic information data, including users with the same gender, age, and occupation, user groups are formed, and these data are cleaned and preprocessed, and converted into structured feature vectors to obtain user features. The user's behavioral data on the e-commerce platform is obtained, including browsing history, purchase history, and collection history. Feature selection is performed on user characteristics and behavioral data, and features that have an impact on color preference prediction are selected as preferred user features to improve the effectiveness of the model. For users in each user group, the number and duration of their interactive behaviors on products with different color richness are counted as the preference measurement of the user's corresponding color richness. The metric values ​​of these preference measurements are normalized to obtain the user's preference distribution for different color richness. Then, an association rule mining algorithm, such as the Apriori algorithm or the FP-growth algorithm, is used to perform correlation analysis on the user feature vector and color preference data to discover the association rules between different feature combinations and color preferences. Based on the mined association rules, a user color richness preference prediction model is constructed to obtain the preference feature vector of the user group.

[0032] In an exemplary embodiment, the steps of using an association rule mining algorithm to analyze and obtain the preference relationship between user groups and the color richness of products, and establishing a preference prediction model based on the preference relationship to obtain the preference feature vector of the user group include:

[0033] Convert basic information data into structured feature vectors to obtain user features;

[0034] Perform feature selection on user features based on the behavioral data of the user group to obtain preferred user features associated with the behavioral data;

[0035] We counted the interactions of different user groups on products with different color richness levels to measure their preference for color richness. Interactions included the number and duration of browsing, purchasing, and adding products to favorites.

[0036] Each preference metric is normalized to obtain the preference distribution of different user groups for color richness;

[0037] Association rule mining algorithms are used to analyze the association between user preference features and preference relationships, and to obtain association rules between different feature combinations and color preferences; different feature combinations include color combinations with different color richness; association rule mining algorithms include Apriori algorithm or FP-growth algorithm;

[0038] Based on the association rules between different feature combinations and color preferences, a preference prediction model is constructed to obtain the preference feature vector of the user group. The preference prediction model is then optimized using collaborative filtering and matrix decomposition algorithms.

[0039] The preference prediction model is evaluated using the cross-validation method, and the accuracy, recall rate, and F1 index of the preference prediction model are calculated. The parameters of the preference prediction model are adjusted and feature selection of user features is performed based on the evaluation results.

[0040] In an exemplary embodiment, the steps of optimizing the preference prediction model using a collaborative filtering algorithm and a matrix decomposition algorithm include:

[0041] A collaborative filtering algorithm is used to construct the interaction matrix between user groups and products, and a matrix decomposition algorithm is used to calculate the interaction matrix to obtain a similarity matrix. Based on the similarity matrix, similar users of the target user in the set user group are found, and the preference threshold of similar users for each product corresponding to the color richness is calculated as the preference prediction value of the target user. The preference prediction model is trained based on the basic information data of the target user and the preference prediction value of the target user.

[0042] In an exemplary embodiment, the steps of using an association rule mining algorithm to analyze and obtain the preference relationship between user groups and the color richness of products, and establishing a preference prediction model based on the preference relationship to obtain the preference feature vector of the user group further include:

[0043] When the target user is a new user, similar users are found based on the new user's basic information data through the preference prediction model, and the average preference prediction value of similar users is calculated and used as the initial preference prediction value of the new user;

[0044] When a new user interacts, the preference prediction value is updated based on the new user's behavioral data; the color richness of the product is introduced, and an initial recommendation list is generated for the new user through a recommendation method based on the content of the product to promote the new user's interactive behavior.

[0045] Specifically, collaborative filtering is used to construct a user-item interaction matrix, calculate the similarity matrix between users or items, find the K users most similar to the target user based on the similarity matrix, and calculate the average preference value of these users for items of each color richness as the preference prediction value of the target user.

[0046] Furthermore, for new users, we use their basic information data and association rules to identify the user group with the most similar feature combinations. We then use the average preference of this group as the initial preference prediction for the new user. As the new user engages in certain interactions, we gradually incorporate their behavioral data and update the preference prediction results.

[0047] Furthermore, by incorporating product color, quantity, and content features, a content-based recommendation method was used to generate an initial recommendation list for new users, promoting interaction between them and the system. The prediction model was evaluated and validated using cross-validation to calculate the model's accuracy, recall, and F1 score. Based on the evaluation results, model parameters and feature selection were adjusted to improve prediction performance. The optimized prediction model was applied to actual product recommendations and personalized display scenarios, recommending appropriate products based on users' color preferences, thereby enhancing their shopping experience and satisfaction.

[0048] In the specific implementation process, basic user information, such as gender, age, and occupation, can be collected through a user registration form. The data is then cleaned and preprocessed using the Python pandas library, converting it into structured feature vectors to obtain user features. Simultaneously, user behavior data on the platform, such as browsing history, purchase history, and favorites, can be obtained from the database. Feature selection algorithms, such as the chi-square test and mutual information, are used to select the top-N features that have a significant impact on color preference prediction as preferred user features. For each user, the number and duration of their interactions with products of varying color richness (e.g., monochrome, low color richness, medium color richness, and high color richness) can be counted. These metrics are then mapped to intervals using Min-Max normalization to obtain the distribution of user preferences for different color richness levels. Next, the Apriori algorithm is used to mine association rules between the user feature vectors and color preference data, setting the minimum support to 0.01, the minimum confidence to 0.5, and the maximum item set length to 3. Association rules between different feature combinations and color preferences are then derived. Based on the association rules, a user color richness preference prediction model can be constructed to obtain the preference feature vector of the user group.

[0049] It can be understood that since association rules are obtained by correlation analysis of preference user features and preference relationships, and the preference prediction model is established based on association rules, the output of the preference prediction model can be a preference feature vector with structured features for quantitative analysis.

[0050] Furthermore, the UserCF method from the collaborative filtering algorithm can be used to first construct a user-item (user group and item) interaction matrix, where the matrix elements represent user preference metrics for items. Then, cosine similarity is used to calculate the similarity matrix between users. For the target user, the top-K most similar users (K is 20) are found, and the average preference value of these users for items of various color richness levels is calculated as the target user's preference prediction value.

[0051] In one embodiment, for a new user, association rules can be used to identify the top-M user groups (M is 5) with the most similar feature combinations based on their basic information. The average preferences of these groups for products with different color richness are used as the initial preference prediction for the new user. As the new user interacts, their behavioral data is gradually integrated to update the preference prediction results. Content features such as product color and style can also be incorporated to generate a personalized initial recommendation list for the new user using a content-based recommendation algorithm.

[0052] Furthermore, the model was evaluated using 5-fold cross-validation, resulting in average accuracy of 85%, average recall of 78%, and an average F1 score of 0.82. The optimized model was applied to actual product recommendation scenarios, recommending the top N products (N being 10) that best match a user's color preferences, improving their shopping experience and satisfaction.

[0053] Step 206: Based on the behavioral data and preference feature vectors of the user groups, a fuzzy clustering algorithm is used to analyze the differences and similarities in color richness among different user groups, and a genetic algorithm is used to optimize the fuzzy clustering algorithm. By simulating natural selection and genetic evolution mechanisms, the clustering center and membership function of the fuzzy clustering algorithm are continuously iterated and optimized.

[0054] In an exemplary embodiment, the steps of analyzing the differences and similarities in color richness among different user groups using a fuzzy clustering algorithm based on the behavior data and preference feature vectors of the user groups include:

[0055] Perform data cleaning and preprocessing on the behavioral data. Preprocessing includes extracting the product ID, behavior type, and behavior time, and converting them into a structured format. The behavior type includes the action corresponding to the behavior data, such as browsing, purchasing, or adding to favorites.

[0056] Obtain the color richness of the corresponding product based on the product ID; cluster the pixels in the product image corresponding to the product ID into several representative colors using the K-means fuzzy clustering algorithm; calculate the first color richness of the representative colors based on the number and weight of the representative colors; extract color-related keywords from the product attribute metadata; obtain the second color richness of the representative colors based on the number and weight of the keywords; perform a weighted fusion of the first color richness and the second color richness to obtain the final color richness of the product;

[0057] The behaviors of users in the user group are sorted in chronological order. A sliding window method is used with a certain time span as the window. The number of actions taken by users in the user group on products with different color richness in each window is counted to obtain the color preference distribution of users in the user group in different time periods.

[0058] If the user is a new user in the user group, the average of the preference thresholds of the users in the user group corresponding to the new user, or the average of the preference thresholds of the new user's similar users found by the preference prediction model, is used as the preference threshold of the new user;

[0059] The color preference distribution of users in the user group is represented as a preference feature vector of length N, and the user features are converted into a numerical user feature vector using the One-hot encoding or Word2Vec method. Each element in the preference feature vector represents the user's preference for the corresponding color.

[0060] The fuzzy C-means algorithm is used to perform cluster analysis on the user's preference feature vectors, and users with similar preference feature vectors are divided into the same user category;

[0061] By adjusting the parameters of the fuzzy C-means algorithm, the fuzzy C-means algorithm is optimized until the clustering performance indicators meet the expectations; the parameters of the fuzzy C-means algorithm include the number of clusters and the membership threshold; the clustering performance indicators include the silhouette coefficient and the Davies-Bouldin index;

[0062] Based on the results of cluster analysis, the color preference characteristics of different user groups are analyzed to explore the differences and similarities in color preferences of different user groups.

[0063] Specifically, by analyzing users' historical browsing, purchasing, and collection behavior data, combined with the user's preference feature vectors for products with different color richness, fuzzy clustering can be used to analyze the differences and similarities in color preferences among different groups.

[0064] Furthermore, the user's historical browsing and purchasing behavior data on the platform can be collected, the data can be cleaned and preprocessed, key fields such as user ID, product ID, behavior type, behavior time, etc. can be extracted, and the data can be converted into a structured format.

[0065] Furthermore, the color richness information for each product is obtained based on the product ID. Product images are analyzed using the K-means clustering algorithm, with the image pixels clustered into representative primary colors. A color richness score is calculated based on the number and distribution of colors. Color-related keywords are extracted from product attribute metadata, and the number and weight of keywords are counted to obtain a color richness score. The color richness scores obtained from the two methods are weighted and combined to obtain the final product color richness. User behavior data is sorted chronologically. Using a sliding window method, with a certain time span as the window, the number of times users browse, purchase, or collect products of different color richness within each window is counted, resulting in the distribution of users' color preferences within different time periods (behavior time).

[0066] Furthermore, for users in the new user group, they are divided into similar user groups based on their basic attributes. The group's average preference distribution is used as their initial preference, or a collaborative filtering algorithm is used to estimate and fill in missing preference data based on the preference distribution of similar users. The user's color preference distribution is represented as a vector of length N, with each element representing the user's preference for the corresponding color richness. Other user characteristics, including age, gender, and occupation, are then converted into numerical user feature vectors using one-hot encoding or Word2Vec methods. The color preference vector and attribute feature vector are concatenated or weighted fused to construct the final user color preference feature vector.

[0067] The fuzzy C-means clustering algorithm, a fuzzy clustering algorithm, was used to perform cluster analysis on preference feature vectors, grouping users with similar preferences into the same group. The fuzzy C-means clustering algorithm minimizes the distance between the cluster center and the sample points it contains by minimizing the objective function, while allowing a sample point to belong to multiple clusters. This algorithm is suitable for processing fuzzy and uncertain data such as user preferences. By adjusting the parameters of the fuzzy C-means clustering algorithm, including the number of clusters and the membership threshold, the clustering effect is continuously optimized until the desired clustering performance indicators, including the silhouette coefficient and the Davies-Bouldin index, are achieved. Based on the clustering results, the color preference characteristics of different user groups are analyzed to explore the differences and similarities in color preferences among these groups.

[0068] During implementation, various user behavior data on the platform, such as browsing, clicking, adding to favorites, and purchasing, can be collected through tracking. This data can then be cleaned and preprocessed using big data processing frameworks like Hadoop, extracting key fields such as user ID (basic information), product ID, behavior type, and timestamp. This data is then stored in NoSQL databases like HBase for easy query and analysis. For product color richness, computer vision libraries like OpenCV can be used to cluster pixels in product images using the K-means clustering algorithm. The number of clusters, K, and the number of iterations are set to 16, resulting in 16 primary colors. The standard deviation of the number and distribution of colors is then calculated and normalized to obtain a color richness score. Furthermore, word segmentation tools like Jieba can be used to segment product attribute metadata, extract color-related keywords, and calculate the TF-IDF value as the keyword weight. This is then normalized to obtain another color richness score. The two scores are weighted and merged in a 7:3 ratio to obtain the final product color richness score.

[0069] Specifically, for analyzing user color preferences, a 30-day sliding window can be used to count the number and duration of each user's interactions with products of varying color richness within the window. Min-Max normalization is then used to convert these interactions into a preference score vector. For new users or users with less than 10 interactions per month, the UserCF collaborative filtering algorithm can be used to identify the top 50 most similar users based on attributes such as age and gender, and their average preference vectors can be used to populate the vector. Finally, user attribute features are converted into 0-1 vectors using one-hot encoding and concatenated with the color preference vector to obtain the final user feature vector with a dimension of 128. For cluster analysis, the fuzzy C-means clustering algorithm is used, with the number of clusters C = 20, the fuzzification coefficient m = 2, and a maximum number of iterations of 100. Cluster centers and membership matrices are obtained by minimizing the objective function. Clustering effectiveness is evaluated using the silhouette coefficient and the Determined Bias Index (DBI) index, with parameters adjusted continuously until the silhouette coefficient is greater than 0.7 and the DBI index is less than 1.5. Clustering results show significant differences in color preferences among different user groups. For example, male users prefer low-saturation products, while female users prefer high-saturation products. Young users prefer bright and colorful products, while middle-aged and elderly users prefer calm and elegant products. Based on these insights, personalized product recommendations and marketing content can be made for different groups, improving the platform's transaction conversion rate and user stickiness.

[0070] In an exemplary embodiment, a genetic algorithm is used to optimize a fuzzy clustering algorithm. By simulating natural selection and genetic evolution mechanisms, the steps of continuously iteratively optimizing the cluster centers and membership functions of the fuzzy clustering algorithm include:

[0071] The initial parameters of the fuzzy clustering algorithm are encoded as genotypes in the genetic algorithm. The initial parameters of the fuzzy clustering algorithm include cluster centers and coefficients of the membership function. The cluster centers are encoded using real numbers, and each gene bit represents the coordinate value of a cluster center on the corresponding dimension. The membership function is encoded using floating-point numbers, and each gene bit identifies a parameter of the membership function. The length of the gene depends on the dimension of the cluster center and the number of parameters of the membership function.

[0072] Randomly generate an initial population with a preset number of individuals, each of which represents a set of possible clustering parameter combinations;

[0073] The parameters of the clustering algorithm are obtained according to the individual genotype decoding, and the clustering results are obtained based on the training samples and the parameters of the clustering algorithm;

[0074] The fitness function is used to evaluate the clustering performance of each individual, and the sum of squared errors of the clustering is introduced to measure the clustering error; the fitness function includes the silhouette coefficient, Davies-Bouldin index, Calinski-Harabasz index and Dunn index; the clustering performance includes evaluating the clustering of each individual including accuracy, compactness and separation indicators;

[0075] The fitness function is designed as a weighted combination of indicators, and the weights are adjusted according to actual needs. The larger the value of the fitness function, the better the clustering parameters corresponding to the individual;

[0076] The individuals are subjected to selection, crossover, and mutation genetic operations. The cluster centers are optimized through real-number coded arithmetic crossover and Gaussian mutation to ensure that the generated new cluster centers are still within the data space. The membership function is optimized using uniform crossover and uniform compilation to ensure that the generated membership coefficient is within [0, 1]. The selection operation adopts an elite retention strategy to prevent the loss of excellent individuals during the evolution process.

[0077] Newly generated offspring individuals are added to the population to replace some individuals with insufficient fitness, maintaining the diversity and evolutionary ability of the population; clustering parameters are continuously iterated and optimized until the preset number of iterations or fitness threshold is reached to obtain the optimal clustering parameters.

[0078] Specifically, the initial parameters of the fuzzy clustering algorithm, including cluster centers and membership function coefficients, can be encoded as genotypes in a genetic algorithm. Cluster centers are encoded using real numbers, with each genotype representing the coordinate value of a cluster center along a certain dimension. Membership functions are encoded using floating-point numbers, with each genotype representing a parameter of the membership function. The length of the genotype depends on the dimension of the cluster centers and the number of membership function parameters. An initial population is randomly generated, typically consisting of 50 to 100 individuals, with each individual representing a possible combination of clustering parameters. Based on the individual genotypes, they are decoded into clustering algorithm parameters, and the fuzzy clustering algorithm is run on the training samples to obtain clustering results.

[0079] A fitness function is used to evaluate the clustering performance of each individual. This combines clustering results with accuracy, compactness, and separability metrics, including the silhouette coefficient, Davies-Bouldin index, Calinski-Harabasz index, and Dunn index. The sum of squared errors is introduced to measure clustering error. The fitness function is designed as a weighted combination of metrics, with weights adjusted based on actual needs. A higher fitness function value indicates a better clustering parameter for that individual. Individuals in the population are subjected to genetic operations such as selection, crossover, and mutation.

[0080] Furthermore, for the optimization of cluster centers, real-number coded arithmetic crossover and Gaussian mutation are used to ensure that the generated new cluster centers remain within the data space. For the optimization of the membership function, uniform crossover and uniform mutation are used to ensure that the generated membership coefficient is within the range of [0, 1]. The selection operation adopts an elite retention strategy to prevent the loss of excellent individuals during the evolution process. The newly generated offspring individuals are added to the population to replace some individuals with insufficient fitness, maintaining the diversity and evolutionary ability of the population. The clustering parameters are continuously iterated and optimized until the preset number of iterations or fitness threshold is reached to obtain the optimal clustering parameter combination. Based on the optimal clustering parameters, the user data is fuzzy clustered to obtain the user's membership matrix.

[0081] Furthermore, for each user, we identify the cluster with the highest membership and assign it to the corresponding user group. For each group, we analyze its user attributes and behavioral preferences to build a targeted personalized recommendation model. When generating recommendation lists, we use the user's membership as a key personalization factor, adjusting the recommendation weights of different items to ensure that the recommendation results are more aligned with the user's actual preferences. This is then used in subsequent user group segmentation and personalized recommendation tasks.

[0082] In specific implementations, cluster centers can be represented as a 10-dimensional real vector, with each dimension corresponding to an attribute in the user feature space, such as age or gender. The membership function can be a Gaussian function, with two floating-point parameters representing the mean and variance. Assuming that users are divided into eight groups, the genotype length of each individual is 82. 100 individuals are randomly generated as the initial population, each representing a set of possible clustering parameters. For each individual, the cluster center and membership function parameters are decoded. The fuzzy C-means clustering algorithm is then run on the training data of 50,000 users, iterating 50 times to obtain the membership matrix. The fitness function is a weighted average of the silhouette coefficient, CH index, and Dunn index, with weights of 0.4, 0.3, and 0.3, respectively. Higher fitness indicates better clustering results.

[0083] Furthermore, in the genetic operation, ranked selection was used, with a probability of 0.8 for crossover, selecting parents from the top 20% of individuals in terms of fitness. Arithmetic crossover was used for cluster centers, and uniform crossover was used for membership functions, with a crossover probability of 0.7. In the mutation operation, Gaussian noise was added to each dimension of the cluster centers with a probability of 0.1, and each parameter of the membership function was randomly perturbed within a range with a probability of 0.05. The population was iterated 200 times, and the clustering parameters corresponding to the optimal individual were the final result. These parameters were applied to the full dataset of 10 million users, performing fuzzy clustering to determine each user's membership in eight groups. Users were then divided into groups with the highest membership. Analysis of user profiles, purchasing behavior, and interests across these groups revealed significant differences in age, gender, spending power, and category preferences. Based on this, a personalized recommendation model was constructed for each group, taking into account users' historical behavior and membership to generate a recommendation list.

[0084] In step 208, based on the analysis results of the optimized fuzzy clustering algorithm, the user group is divided into multiple categories of users according to color richness, and the color preference characteristics of users in each category are obtained; a product-based collaborative filtering algorithm is used to generate a recommendation list of products based on color richness according to the color preference characteristics and the corresponding products corresponding to the color richness.

[0085] In an exemplary embodiment, based on the analysis results of the optimized fuzzy clustering algorithm, the user group is divided into multiple categories of users according to color richness, and the step of obtaining the color preference characteristics of each category of users includes:

[0086] Apply the optimal clustering parameters to the preference feature vectors of users in the user group, and use the fuzzy C-means clustering algorithm to obtain the membership of each user to different clusters;

[0087] Based on the optimal clustering parameters, fuzzy clustering is performed on the basic information data of users in the user group to obtain the user membership matrix. Based on the membership matrix, each user is divided into the cluster with the largest membership, forming different color richness preference categories. The number of categories is obtained by the number of clusters in the optimal clustering parameters.

[0088] The average preference of each color richness preference category at different color richness levels was calculated to obtain the preference distribution characteristics of each color richness. The distribution of users in each color preference category in terms of age, gender, age, and income demographic attributes was also calculated. The differences in demographic attributes between different color richness preference categories were found through chi-square tests and t-tests to form demographic characteristic rules.

[0089] Association rule mining and frequent pattern mining are performed on the behavioral data of users in each category to obtain the key characteristics of the corresponding category of users in terms of category preference, price preference and activity preference, and form a color richness preference user portrait; the color richness preference category and the corresponding user feature rules are stored in the user classification table as the basis for subsequent personalized recommendations and marketing; when the user in the user group is a new user, the most similar color richness preference category is matched according to its color richness preference characteristics.

[0090] Specifically, we obtain fuzzy clustering parameters optimized by a genetic algorithm, including cluster centers and membership functions. We then apply these fuzzy clustering parameters to the user's color preference feature vector data and run a fuzzy C-means clustering algorithm to determine each user's membership in different clusters. Based on the membership matrix, we assign each user to the cluster with the highest membership, forming different color richness preference categories. The number of categories is determined by the number of clusters in the optimal clustering parameters.

[0091] Furthermore, for each color richness preference category, the average preference of its users at different color richness levels was calculated to obtain the preference distribution characteristics of this category at each color richness level. For each color richness preference category, its user characteristics were further analyzed. The distribution of users in terms of demographic attributes such as age, gender, region, and income was calculated. Using chi-square tests and t-tests, significant differences between different categories in these attributes were identified to form demographic characteristic rules.

[0092] Furthermore, association rule mining and frequent pattern mining are performed on the browsing, collection, and purchasing behavior data of users in each category to obtain the key characteristics of category preferences, price preferences, and activity preferences of users in this category, forming a complete user profile with color richness preferences. The color richness preference categories and corresponding user feature rules are stored in the user classification table, which serves as an important basis for subsequent personalized recommendations and marketing. When a new user enters the system, based on their color richness preference characteristics, they are matched to the most similar preference category, quickly achieving user classification. Based on the color richness preference category to which the user belongs, products that match the category's preference characteristics are selected from the product library as candidate recommendation sets.

[0093] Furthermore, the matching degree is calculated by combining the similarity between the color richness characteristics of the product and the category preference distribution, as well as the category and price metadata of the product, to describe the degree of conformity with the characteristics of the user in that category. For users in different color richness preference categories, the number of purchases and the number of users who purchased products at each color richness level are counted, and the purchase probability (number of purchasing users / total number of users in the category) and purchase frequency (number of purchases / number of purchasing users) are calculated to obtain the association rules between color richness and purchasing behavior, which are used to guide the personalized recommendation strategy for products. When generating the recommendation list, for each candidate product, the cosine similarity between its color richness characteristics and the preference distribution of the category to which the user belongs is calculated as the color richness matching weight. At the same time, the comprehensive matching score is calculated by combining other product characteristics, including category and price, with the matching degree of user characteristics.

[0094] Finally, candidate products are ranked based on their comprehensive matching scores, with products with high matching scores at the top. A threshold for color richness weight is set, and products exceeding the threshold are further ranked to highlight the influence of color richness preferences. This ensures that recommendation results are more aligned with user color preferences, improving recommendation conversion rates and user satisfaction.

[0095] In practice, we can use optimized fuzzy clustering parameters, such as 10 cluster centers and corresponding Gaussian membership functions, to cluster the color preference feature vectors of one million users. By calculating each user's membership to each cluster center, we assign them to the cluster with the highest membership, forming 10 different color richness preference categories.

[0096] Furthermore, for each category, the average preference scores of its users at five levels of color richness were calculated, and a radar chart of the preference distribution was drawn. At the same time, an independence test was performed on the distribution of user attributes in each category. For example, it was found that the proportion of users aged 18-25 in the high color richness preference category was significantly higher than that in other categories (p<0.01). The Apriori algorithm was also used to mine association rules to obtain the purchasing patterns of users in each category in different price ranges and categories. For example, users with a high color richness preference were more likely to buy T-shirts and jackets priced between 100 and 200 yuan (support >5%, confidence >60%). Based on the user classification results, 1,000 products were selected from 100,000 products as the initial recommendation set. By calculating the JS divergence between the product color features and the user preference distribution, the 500 products with the highest matching degree were screened out. After comprehensively considering the matching degree of other features such as category and price, the top 50 products were selected to generate a personalized recommendation list. Finally, the differences in purchasing behavior of users in different clusters were also analyzed. For example, the average monthly purchase frequency of users with a high color richness preference was 1.5 times, which was twice that of users with a low color richness preference. The unit price of the pen was also 30% higher. This provides an important reference for subsequent marketing strategies and inventory management.

[0097] In an exemplary embodiment, using a collaborative filtering algorithm based on products, the steps of generating a recommendation list of products based on color richness according to color preference characteristics and corresponding products with corresponding color richness include:

[0098] The products in the product library are divided into corresponding product subsets according to their color richness, and candidate recommended products are selected from the product subsets with corresponding color richness according to each user's color richness preference category;

[0099] According to the user's color richness preference category, the corresponding user attribute value is set; the user attribute value is positively correlated with the number of colors in the color richness; for users with high color richness preference, products with color richness attribute values ​​greater than or equal to 4 are selected as candidates; for users with medium color richness preference, products with attribute values ​​between 2 and 4 are selected; for users with low color richness preference, products with attribute values ​​less than 2 are selected; and combined with the user's historical behavior data, weights are set that are positively correlated with the frequency of the user's browsing, purchasing, or collection records; and a product-based collaborative filtering algorithm is used to calculate and sort the similarity of candidate recommended products.

[0100] The similarity between products is determined by color richness and other similarity metrics, and the color richness difference is obtained based on the difference in color richness between the products. Other similarity metrics include whether there are browsing, purchasing, or favorite records of multiple users with the same user attribute values ​​between the products. Other similarity metrics are positively correlated with the frequency of browsing, purchasing, or favorite records of users with the same user attribute values. The difference in color richness between products is the difference in the number of colors. The color richness difference is mapped to the range of 0 to 1 using a mapping function to obtain color richness similarity. The mapping function includes a Gaussian kernel function. The color richness similarity is weighted and summed with other similarity metrics to obtain a comprehensive similarity.

[0101] The K-means clustering algorithm is used to cluster candidate recommended products according to color richness, and products with similar color richness are divided into the same cluster to form product subsets with different color richness; the candidate recommended products are sorted and divided into different color richness subsets to form a hierarchical recommendation list; the recommendation list is used to give priority to recommending products that match the user's color richness preference category; when generating the final color richness product recommendation list, the recommendation list is dynamically adjusted based on factors such as real-time product inventory, promotional activities, and new product launches to ensure the diversity and timeliness of the recommended products and continuously optimize the recommendation results.

[0102] Specifically, we obtain each user's basic information and color preference characteristics. Based on the fuzzy clustering results generated in the previous step, we determine the user's color quantity preference category. We then extract metadata for each product in the product library, focusing on the product's color richness attribute—the number of primary colors contained in the product image, as determined by the color clustering algorithm. We then divide the products in the library based on their color richness attribute, generating subsets of products with different color richness levels: low, medium, and high color richness.

[0103] Specifically, for each user, candidate recommended products are selected from a subset of products with the corresponding color richness level based on their color quantity preference category. Different screening thresholds are set based on the user's color richness preference category. For users with a high color richness preference, products with a color richness attribute value greater than or equal to 4 are selected as candidates; for users with a medium color richness preference, products with an attribute value between 2 and 4 are selected; and for users with a low color richness preference, products with an attribute value less than 2 are selected. At the same time, based on the user's historical behavior data, products that appear frequently in their browsing, collection, and purchase records are given a higher weight to increase their probability of selection.

[0104] Furthermore, for the candidate recommended product set, an item-based collaborative filtering algorithm is used to calculate and sort the similarity. When calculating the similarity between each two products, in addition to the factors of common purchase and common browsing, the similarity of the color richness attribute is also introduced. For two products A and B, their color richness attribute values ​​colorA and colorB are extracted respectively, and the absolute value of the difference between the two, abs(colorA-colorB), is calculated to obtain the color richness difference. The difference is mapped to the range of 0 to 1 to obtain the color richness similarity. The mapping function can use a Gaussian kernel function, such as exp(-abs(colorA-colorB)^2 / 2σ^2), where σ is a parameter that controls the similarity decay rate.

[0105] Furthermore, the color richness similarity is weighted and summed with other similarity metrics to obtain a comprehensive similarity. The candidate recommendation product set is clustered using the K-means clustering algorithm based on the product's color richness attribute value. The goal of clustering is to group products with similar attribute values ​​into the same cluster, forming subsets of products with different color richness levels. The number of clusters, K, can be set based on the size and color richness distribution of the candidate products, typically ranging from 3 to 5.

[0106] Specifically, when clustering, other important attributes of products can also be incorporated into the clustering attributes to improve the clustering effect. Based on the product similarity matrix and clustering results, the candidate recommended products are sorted and divided into different color richness subsets to form a hierarchical recommendation list. Priority is given to products in the color richness level that best matches user preferences. When generating the final color richness product recommendation list, the recommendation list is dynamically adjusted based on factors such as real-time product inventory, promotional activities, and new product launches to ensure the diversity and timeliness of the recommended products and continuously optimize the recommendation results.

[0107] For example, basic attributes such as age and gender, as well as color preference characteristics, can be obtained from user registration information and behavior logs. Based on users' browsing, clicking, adding to favorites, and purchasing behaviors on products with different color richness, their corresponding preference scores are calculated. Fuzzy clustering is then used to categorize users into three color richness preference categories: high, medium, and low.

[0108] For one million items, we extracted the RGB histograms of their images. Peak detection and threshold filtering were performed to determine the number of primary colors in each item. This value was normalized to a value between 0 and 1 as the color richness attribute. Based on this attribute value, we used the K-means clustering algorithm to cluster the items into five clusters, representing five color richness levels: very low, low, medium, high, and very high. For each user, we selected items with the corresponding color richness level based on their preferred category. We also assigned a weight of 1.5-2 times the user's historical behavior to the items.

[0109] Specifically, when calculating product similarity, the color richness attribute value of the product is extracted, the absolute value of the difference is calculated, and then it is mapped to the range of 0-1 through the Gaussian kernel function with σ=0.1 to obtain the color richness similarity. Then, combined with similarities such as co-purchase and co-browsing, the weighted sum is calculated in a ratio of 2:2:1 to obtain a comprehensive similarity matrix. The candidate products are further clustered, and products with similar color richness are clustered into 3-5 clusters. Within each cluster, they are sorted according to similarity to form a hierarchical recommendation list. Finally, the recommendation list must be dynamically adjusted, such as filtering out out-of-stock products and giving ranking boosts to products on promotion, to ensure the real-time and effectiveness of the recommendations.

[0110] Step 210 : Collect and dynamically adjust the recommendation list based on feedback information from users of each category on the recommendation list to optimize the color preference feature.

[0111] In an exemplary embodiment, the steps of collecting and dynamically adjusting the recommendation list to optimize the color preference characteristics based on the feedback information of users in each category on the recommendation list include:

[0112] Collect user feedback on historically recommended products; feedback includes behavioral data and explicit feedback; explicit feedback includes user ratings and comments on recommended products according to pre-set rules; extract features from feedback information to determine user preferences for products with different color richness; preference expressions are used to indicate click-through rate and purchase rate;

[0113] Compare the feedback information with the corresponding user's color richness preference category preference profile; the preference profile comparison is used to indicate the user's satisfaction with the recommended products in the preference category; the satisfaction includes click-through rate and purchase rate indicators;

[0114] In response to the click-through rate being lower than a preset click-through rate or the purchase rate being lower than a preset purchase rate, a heat distribution of recommended products is constructed to analyze whether there is insufficient recommendation of long-tail products or excessive recommendation of head products, and obtain an analysis result based on the heat distribution;

[0115] According to the analysis results based on the heat distribution, the recommendation list is dynamically adjusted, including: if the satisfaction level is higher than the first preset satisfaction level, the recommendation ratio of the corresponding recommended product is increased and it is recommended first, and the weight of the corresponding recommended product is increased during the similarity calculation process; if the satisfaction level is lower than the second preset satisfaction level, the recommendation ratio of the corresponding recommended product is reduced and the weight of the corresponding recommended product is reduced during the similarity calculation process.

[0116] Specifically, user feedback on historically recommended products is collected, including user click, browse, favorite, and purchase behavior data, as well as explicit user feedback on ratings and comments on recommended products. User feedback data is preprocessed and feature extracted to obtain user preferences for products with different color richness, including click-through rates and purchase rates for products with high color richness. User feedback features are compared with the user's original color richness preference profile to calculate the user's acceptance and satisfaction with the recommended products. By analyzing user behavioral feedback on recommended products, the overall click-through rate and purchase rate indicators of recommended products are calculated and compared with the platform's average level. If the indicators of recommended products are lower than the average level, it indicates that there are deficiencies in the recommendation strategy.

[0117] Furthermore, by constructing a popularity distribution for recommended products, we analyze whether there is underrecommendation of long-tail products or overrecommendation of top products, thereby identifying issues with the recommendation strategy. Based on user feedback, we optimize and adjust various aspects of the recommendation strategy. When selecting candidate products, we calculate the recommendation satisfaction level for each color richness level based on user feedback on products with different color richness levels. We increase the proportion of candidate products with high satisfaction levels, while decreasing the proportion of candidate products with low satisfaction levels. When calculating similarity, we increase the weight of product features with positive user feedback and decrease the weight of product features with negative feedback. The extent of weight adjustment can be determined based on the intensity and persistence of user feedback. When generating the recommendation list, we prioritize products with positive user feedback, while downgrading or filtering products with negative feedback. Furthermore, when clustering candidate products, in addition to color richness, we also incorporate new clustering features based on user feedback, including user satisfaction and recommendation success rate. We use the K-means algorithm for clustering, and calculate the silhouette coefficient to determine the optimal number of clusters. Based on user feedback, dynamically adjust cluster centers to group items with positive user feedback to form clusters with high satisfaction. Items with negative user feedback are removed or demoted, optimizing clustering results to better align with users' true preferences. Regularly update user color richness preference profiles, using long-term, stable user behavior feedback as the primary basis for preferences to reduce the impact of accidental, sudden behavior. Continuously optimize user color richness classifications to accurately reflect their true preferences. Conduct A / B testing to evaluate the effectiveness of the optimization of the recommendation strategy. Randomly divide users into an experimental group and a control group. The experimental group uses the optimized recommendation strategy, while the control group uses the original recommendation strategy.

[0118] Furthermore, over a certain period of time, the click-through rate, purchase rate, and satisfaction index of the two groups of users were compared. A significance test was performed to determine whether the optimized recommendation strategy resulted in a statistically significant improvement. A / B testing was used to quantify the effects of recommendation optimization and verify the effectiveness of the optimization measures. An evaluation and feedback mechanism for the recommendation strategy was established to continuously track user feedback on the recommendation results. Problems and deficiencies in the recommendations were promptly identified and adjustments and optimizations were made promptly. This formed a closed-loop improvement loop for the recommendation strategy, continuously improving the accuracy of recommendations and user satisfaction.

[0119] During implementation, tracking and logging can be used to collect user feedback on recommended products, including clicks, favorites, add-to-cart, and purchases, as well as explicit feedback such as ratings and comments. For a recommendation system with 10 million users and 1 million products, terabytes of user feedback data can be collected daily. Using big data processing platforms like Spark, this feedback data is cleaned, converted, and features extracted to derive user preference metrics for products at different color richness levels, such as click-through rate and conversion rate. These metrics are then correlated with users' existing color preference profiles, and user satisfaction with the recommended products is measured using metrics such as the Pearson correlation coefficient.

[0120] Furthermore, if the overall click-through rate of recommended products is more than 10% lower than the platform average, or if the top 1% of popular products account for more than 50% of recommended impressions, the recommendation strategy optimization process is triggered. The proportion of different color richness levels in the candidate product pool is dynamically adjusted, increasing the sampling volume for the highest satisfaction level by 20% and reducing the sampling volume for the lowest satisfaction level by 20%. In the ItemCF similarity calculation, the weights of user click and purchase behaviors are multiplied by 1.5 and 2.0, respectively, while the weights of user browsing and favorite behaviors are multiplied by 0.8 and 1.2. Newly added clustering features, such as satisfaction and success rate, are assigned weights of 0.3 and 0.5, respectively. KModes clustering is used to divide candidate products into 5-8 clusters. Product clusters with a satisfaction score greater than 0.7 and a recommendation success rate greater than 80% are ranked 20% higher when generating the recommendation list. Product clusters with a satisfaction score less than 0.3 or a recommendation success rate less than 30% are ranked 50% lower. When updating the user preference profile, the color features of products that have been clicked or collected for three consecutive days are given a weight of 1.5 times, and the color features of products that have been purchased in the last seven days are given a weight of 2 times, thereby increasing the weight of stable preference features.

[0121] Furthermore, through regular monthly updates, the consistency between color-rich user segmentation and actual preferences can be increased from 85% to over 93%. A weekly A / B test of the recommendation strategy is conducted, randomly selecting 5% of users as the experimental group. Personalized recommendations are generated for them using the optimized strategy. The CTR, conversion rate, and other indicators of the experimental group are observed to see if they exceed the control group by more than 20%. Paired t-tests and Mann-Whitney U tests are performed to verify the significance of the improvement in the experimental group. Through continuous strategy optimization and effect evaluation, the average CGR has increased from 8% to 12%, and the NDCG has increased from 0.32 to 0.41, effectively improving recommendation effectiveness and user satisfaction.

[0122] It should be understood that, although the steps in the flowcharts of the above embodiments are shown in sequence as indicated by the arrows, these steps are not necessarily performed in the order indicated by the arrows. Unless otherwise specified herein, there is no strict order restriction on the execution of these steps, and these steps can be performed in other orders. Moreover, at least a portion of the steps in the flowcharts of the above embodiments may include multiple steps or multiple stages, and these steps or stages are not necessarily performed at the same time, but can be performed at different times. The execution order of these steps or stages is not necessarily to be performed in sequence, but can be performed in turn or alternately with other steps or at least a portion of steps or stages in other steps.

[0123] In an exemplary embodiment, the present application also provides a computer device including a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, the steps of a low-altitude monitoring method for drones as provided above in the present application are implemented.

[0124] In an exemplary embodiment, the present application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of a method for low-altitude monitoring of a drone as provided above in the present application.

[0125] Those skilled in the art will appreciate that all or part of the processes in the above-mentioned embodiment methods can be implemented by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, database or other media used in the embodiments provided in this application may include at least one of non-volatile and volatile memory. Non-volatile memory may include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory may include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM). The database involved in the various embodiments provided herein may include at least one of a relational database and a non-relational database. Non-relational databases may include, but are not limited to, distributed databases based on blockchains. The processor involved in the various embodiments provided herein may be, but are not limited to, a general-purpose processor, a central processing unit, a graphics processing unit, a digital signal processor, a programmable logic unit, a data processing logic unit based on quantum computing, and the like.

[0126] The technical features of the above embodiments can be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0127] The above-described embodiments merely represent several implementation methods of the present application. While the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the present application. It should be noted that a person of ordinary skill in the art may make various modifications and improvements without departing from the spirit of the present application, and these modifications and improvements fall within the scope of protection of the present application. Therefore, the scope of protection of the present application shall be determined by the appended claims.

Claims

1. A personalized recommendation service method for e-commerce based on genetic fuzzy clustering, characterized in that: The method comprises: Convert the acquired and loaded product images corresponding to the products into RGB color space, extract the colors of the product images, and classify the corresponding products into monochrome, low color richness, medium color richness, and high color richness according to the number of colors of the product images from low to high in terms of color richness; Based on the basic information data of each user group and the corresponding behavior data of the user group, the preference relationship between the user group and the color richness of the product is obtained, and a preference prediction model is established according to the preference relationship to obtain the preference feature vector of the user group; the user group includes users of different genders, occupations and ages; the behavior data includes the browsing records, purchase records and collection records of the user group for the product; According to the behavior data of the user groups and the preference feature vectors, a fuzzy clustering algorithm is used to analyze the differences and similarities in color richness of different user groups, and a genetic algorithm is used to optimize the fuzzy clustering algorithm; Based on the analysis results of the optimized fuzzy clustering algorithm, the user group is divided into multiple categories of users according to the color richness, and the color preference characteristics of each category of users are obtained; according to the color preference characteristics and the corresponding products corresponding to the color richness, a recommendation list of products based on color richness is generated; The recommendation list is collected and adjusted dynamically according to feedback information of users of each category on the recommendation list to optimize the color preference feature.

2. The method according to claim 1, characterized in that The step of extracting the color of the product image and classifying the corresponding product into monochrome, low color richness, medium color richness and high color richness according to the number of colors of the product image from low to high in terms of color richness includes: The converted RGB color space is processed by using a fuzzy clustering algorithm; the data points of the RGB color space are assigned to K cluster centers in an iterative manner, and the distance from each data point to the cluster center is adjusted to be the minimum; Clustering the pixels of the product image into a plurality of representative colors to obtain a color distribution of the product image with respect to the representative colors; Determine that when the number of colors in the color distribution is below a first preset threshold, the color richness of the product is monochrome; when the number of colors in the color distribution exceeds the first preset threshold and is below a second preset threshold, the color richness of the product is low color richness; when the number of colors in the color distribution exceeds the second preset threshold and is below a third preset threshold, the color richness of the product is medium color richness; when the number of colors in the color distribution exceeds the third preset threshold, the color richness of the product is high color richness; wherein the first preset threshold, the second preset threshold and the third preset threshold are all set according to pre-stored business needs and business data.

3. The method according to claim 1, characterized in that The steps of analyzing and obtaining the preference relationship between the user group and the color richness of the product, and establishing a preference prediction model according to the preference relationship to obtain the preference feature vector of the user group include: Converting the basic information data into a structured feature vector to obtain user features; Performing feature selection on the user features according to the behavior data of the user group to obtain preferred user features associated with the behavior data; The interactive behaviors of different user groups on the commodities of different color richness are counted respectively as the preference measurement of the user groups for color richness; the interactive behaviors include the number and duration of browsing records, purchase records and collection records of the commodities; Normalizing the preference metrics to obtain the preference distribution of different user groups for color richness; Performing correlation analysis on the preferred user features and the preference relationship to obtain association rules between different feature combinations and the color preferences; the different feature combinations include color combinations of color richness; constructing a preference prediction model based on the association rules between the different feature combinations and the color preferences to obtain the preference feature vector of the user group, and optimizing the preference prediction model using a collaborative filtering algorithm and a matrix decomposition algorithm; The accuracy, recall rate and F1 index of the preference prediction model are evaluated, and the parameters of the preference prediction model are adjusted and feature selection of user features is performed according to the evaluation results.

4. The method according to claim 3, characterized in that The steps of optimizing the preference prediction model by using collaborative filtering algorithm and matrix decomposition algorithm include: An interaction matrix between the user group and the commodities is constructed, and a matrix decomposition algorithm is used to calculate the interaction matrix to find similar users of the target user in the set user group, and the preference thresholds of the similar users for each commodity corresponding to the color richness are calculated as the preference prediction value of the target user; the preference prediction model is trained according to the basic information data of the target user and the preference prediction value of the target user.

5. The method according to claim 4, characterized in that The step of using an association rule mining algorithm to analyze and obtain the preference relationship between the user group and the color richness of the product, and establishing a preference prediction model based on the preference relationship to obtain the preference feature vector of the user group also includes: In the case where the target user is a new user, based on the basic information data of the new user, the similar users are found through the preference prediction model, and the average of the preference prediction values ​​of the similar users is calculated and used as the initial preference prediction value of the new user; When the new user generates interactive behavior, the preference prediction value is updated according to the behavioral data of the new user; the color richness of the product is introduced, and an initial recommendation list is generated for the new user through a recommendation method based on the content of the product to promote the interactive behavior of the new user.

6. The method according to claim 5, characterized in that The step of analyzing the differences and similarities in color richness of different user groups using a fuzzy clustering algorithm based on the behavior data of the user group and the preference feature vector comprises: Performing data cleaning and preprocessing operations on the behavior data; the preprocessing operations include extracting the product ID, behavior type and behavior time of the product, and converting them into a structured format; the behavior type includes the action corresponding to the behavior data, and the action includes browsing, purchasing or collecting; Obtaining the color richness of the corresponding product according to the product ID; clustering the pixel points in the product image corresponding to the product ID into several representative colors by a fuzzy clustering algorithm, calculating the first color richness of the representative color according to the number and weight of the representative colors, extracting color-related keywords from the attribute metadata of the product, and obtaining the second color richness of the representative color according to the number and weight of the keywords; performing weighted fusion of the first color richness and the second color richness to obtain the final color richness of the product; The behaviors of users in the user group are sorted in chronological order, and a sliding window method is used to take a certain time span as a window, and the number of actions of users in the user group on commodities with different color richness in each window is counted to obtain the color preference distribution of users in the user group in different time periods; In the case where the user is a new user in the user group, the average value of the preference thresholds of the users in the user group corresponding to the new user, or the average value of the preference thresholds of the users similar to the new user found by the preference prediction model is used as the preference threshold of the new user; The color preference distribution of users in the user group is represented as a preference feature vector of length N, and converted into a numerical user feature vector; wherein: each element in the preference feature vector represents the user's preference for the corresponding color; The fuzzy C-means algorithm is used to perform cluster analysis on the user's preference feature vectors, and users with similar preference feature vectors are divided into the same user category; By adjusting the parameters of the fuzzy C-means algorithm, the fuzzy C-means algorithm is optimized until the clustering performance index reaches the expected value; the parameters of the fuzzy C-means algorithm include the number of clusters and the membership threshold; the clustering performance index includes the silhouette coefficient and the Davies-Bouldin index; According to the results of cluster analysis, the color preference characteristics of different user groups are analyzed to explore the differences and similarities in color preferences of different user groups.

7. The method according to claim 1, characterized in that The step of optimizing the fuzzy clustering algorithm by using a genetic algorithm comprises: Encoding the initial parameters of the fuzzy clustering algorithm as genotypes in the genetic algorithm; the initial parameters of the fuzzy clustering algorithm include the coefficients of the cluster center and the membership function; the cluster center is encoded by real numbers, and each gene bit represents the coordinate value of a cluster center on the corresponding dimension; the membership function is encoded by floating point numbers, and each gene bit identifies a parameter of the membership function; the length of the gene depends on the dimension of the cluster center and the number of parameters of the membership function; Randomly generate an initial population, the number of individuals in the population is a preset number, and each individual represents a set of possible clustering parameter combinations; Obtaining parameters of a clustering algorithm according to the genotype decoding of the individual, and obtaining a clustering result based on the training sample and the parameters of the clustering algorithm; The clustering performance of each individual is evaluated by using a fitness function, and the clustering error sum of squares is introduced to measure the clustering error; the fitness function includes a silhouette coefficient, a Davies-Bouldin index, a Calinski-Harabasz index and a Dunn index; the clustering performance includes the evaluation of the clustering of each individual, including accuracy, compactness and separation indicators; The fitness function is designed as a weighted combination of indicators, and the weights are adjusted according to actual needs. The larger the value of the fitness function, the better the clustering parameters corresponding to the individual; The individuals are subjected to selection, crossover and mutation genetic operations; the cluster centers are optimized to ensure that the generated new cluster centers are still within the data space; the membership function is optimized by uniform crossover and uniform compilation to ensure that the generated membership coefficient is within [0, 1]; the selection operation adopts an elite retention strategy to prevent the loss of excellent individuals during the evolution process; The newly generated offspring individuals are added to the population to replace some individuals with insufficient fitness, so as to maintain the diversity and evolutionary ability of the population; the clustering parameters are continuously optimized by iteration until the preset number of iterations or fitness threshold is reached to obtain the optimal clustering parameters.

8. The method according to claim 7, characterized in that The step of dividing the user group into multiple categories of users according to the color richness based on the analysis result of the optimized fuzzy clustering algorithm and obtaining the color preference characteristics of each category of users includes: Applying the optimal clustering parameters to the preference feature vectors of users in the user group, and using the fuzzy C-means clustering algorithm to obtain the membership of each user to different clusters; According to the optimal clustering parameters, fuzzy clustering is performed on the basic information data of users in the user group to obtain the user's membership matrix, and each user is divided into a cluster with the largest membership according to the membership matrix to form different color richness preference categories, and the number of categories is obtained by the number of clusters in the optimal clustering parameters; The average preference of each color richness preference category at different color richness is counted to obtain the preference distribution characteristics at each color richness, and the distribution of users of each color preference category in terms of age, gender, age and income demographic attributes is counted. The differences of different color richness preference categories in the demographic attributes are found through chi-square test and t-test methods to form demographic characteristic rules; Association rule mining and frequent pattern mining are performed on the behavioral data of users in each category to obtain the key characteristics of the corresponding category of users in terms of category preference, price preference and activity preference, and form a color richness preference user portrait; the color richness preference category and the corresponding user feature rules are stored in the user classification table as a basis for subsequent personalized recommendations and marketing; when the user in the user group is a new user, the most similar color richness preference category is matched according to his color richness preference characteristics.

9. The method according to claim 7, characterized in that: The step of generating a recommendation list of commodities based on color richness according to the color preference characteristics and the commodities corresponding to the corresponding color richness comprises: The products in the product library are divided into corresponding product subsets according to the color richness, and candidate recommended products are selected from the product subsets of the corresponding color richness according to the color richness preference category of each user; According to the color richness preference category described by the user, the corresponding user attribute value is set; the user attribute value is positively correlated with the number of colors in the color richness; for users with high color richness preference, select products with color richness attribute values ​​greater than or equal to 4 as candidates; for users with medium color richness preference, select products with attribute values ​​between 2 and 4; for users with low color richness preference, select products with attribute values ​​less than 2; and in combination with the user's historical behavior data, set a weight that is positively correlated with the frequency of occurrence of the user's browsing, purchasing or collection records; use a collaborative filtering algorithm based on the products to perform similarity calculation and sorting to perform similarity calculation and sorting on the candidate recommended products; The similarity between the commodities is judged by color richness and other similarity metrics, and the color richness difference is obtained according to the difference in color richness between the commodities; the other similarity metrics include whether there are browsing, purchasing or collection records of multiple users with the same user attribute value between the commodities; and the other similarity metrics are positively correlated with the frequency of browsing, purchasing or collection records of users with the same user attribute value; the difference in color richness between the commodities is the difference in the number of colors; the color richness difference is mapped to the interval of 0 to 1 by a mapping function to obtain color richness similarity; the mapping function includes a Gaussian kernel function; the color richness similarity is weightedly summed with other similarity metrics to obtain a comprehensive similarity; The candidate recommended products are clustered according to color richness using a K-means clustering algorithm, and products with similar color richness are divided into the same cluster to form product subsets with different color richness; the candidate recommended products are sorted and divided into different color richness subsets to form a hierarchical recommendation list; the recommendation list is used to give priority to recommending products that match the user's color richness preference category.

10. The method according to claim 9, characterized in that The step of collecting and dynamically adjusting the recommendation list to optimize the color preference feature based on the feedback information of the users of each category on the recommendation list comprises: Collect user feedback information on historically recommended products; the feedback information includes behavioral data and explicit feedback; the explicit feedback includes the user's rating and comments on the recommended products according to the set rules; extract features from the feedback information to obtain the user's preference for products with different color richness; the preference expression is used to indicate click-through rate and purchase rate; Comparing the feedback information with the color richness preference category of the corresponding user for a preference portrait; the preference portrait comparison is used to indicate and calculate the user's satisfaction with the recommended products in the preference category; the satisfaction includes the click rate and the purchase rate indicators; In response to the click-through rate being lower than a preset click-through rate or the purchase rate being lower than a preset purchase rate, by constructing a heat distribution of recommended products, analyzing whether there is an under-recommendation of long-tail products or an over-recommendation of head products, and obtaining an analysis result based on the heat distribution; According to the analysis results based on the heat distribution, dynamically adjusting the recommendation list includes: if the satisfaction level is higher than a first preset satisfaction level, increasing the recommendation ratio of the corresponding recommended product and giving priority to the recommendation, and increasing the weight of the corresponding recommended product during the similarity calculation process; if the satisfaction level is lower than a second preset satisfaction level, reducing the recommendation ratio of the corresponding recommended product and reducing the weight of the corresponding recommended product during the similarity calculation process.

Images