A wig commodity intelligent recommendation system and method based on hairstyle matching and user portrait

Abstract

The application discloses a wig commodity intelligent recommendation system and method based on hairstyle matching and user portrait. The method first constructs a hairstyle feature database and a wig commodity database, and extracts the features of the hairstyle and the wig commodity through a multi-modal deep neural network. Secondly, a hairstyle-commodity intelligent matching technology is adopted to calculate the comprehensive similarity of the wig commodity and the hairstyle, and generate a candidate commodity set. Then, a user portrait is constructed, including budget preference, brand preference, material preference and other dimensions, and various methods are used to estimate the user budget range. Next, the candidate commodities are multi-dimensionally layered, including price layering, material layering and brand layering. Finally, based on the user portrait and the commodity layering results, a personalized recommendation algorithm with dynamic weight adjustment is used to generate a wig commodity recommendation list. Through precise hairstyle matching, multi-dimensional commodity layering and user portrait construction, the application realizes personalized recommendation of wig commodities, and the recommendation click rate and purchase conversion rate are significantly improved.

Description

Technical Field

[0001] This invention relates to the field of e-commerce recommendation technology, specifically to an intelligent recommendation system and method for wig products based on hairstyle matching and user profile. Background Technology

[0002] As people's demand for personalized image increases, the wig product market continues to grow. When choosing a wig, consumers often hope that the hairstyle, color, and style of the wig will match their ideal hairstyle. However, existing wig product recommendation technologies have the following problems: First, there is a lack of precise correlation between hairstyles and wig products. Existing wig e-commerce platforms mainly rely on keyword searches or category browsing (such as categorized by "long hair," "short hair," "curly hair," etc.), requiring consumers to spend a lot of time browsing products and making it difficult to find wigs that highly match their target hairstyle. The correlation between wig products and hairstyles mainly relies on manual labeling, which is inefficient and prone to omissions.

[0003] Second, the recommendation system lacks personalization due to its limited scope. Existing recommendation systems are mostly based on "popular products" or "purchase history," failing to consider personalized factors such as users' economic strength, budget range, and brand preferences. Users with different spending power have vastly different needs for wigs, including multiple dimensions such as material (synthetic fiber, blended, human hair), price (hundreds to thousands of yuan), and brand (affordable brands to luxury brands). Existing systems cannot meet these diverse needs.

[0004] Third, it is difficult to accurately grasp users' budget preferences. A consumer's budget range is influenced by various factors such as income level, consumption habits, and purchasing motivation; it is a dynamic and implicit variable. Existing systems cannot accurately predict users' budget preferences, leading to recommended product prices that may be too high or too low, affecting users' purchase decisions and conversion rates.

[0005] Fourth, the feedback mechanism for recommendation effectiveness is inadequate. Existing systems typically train recommendation models based on static data, lacking a dynamic adjustment mechanism driven by real-time user feedback. User actions such as clicking, browsing, saving, and purchasing recommendations are not effectively used to optimize the recommendation algorithm, causing recommendation effectiveness to decline over time.

[0006] Therefore, there is an urgent need for an intelligent recommendation system and method for wig products that can accurately match hairstyles, consider users' personalized needs, automatically learn users' budget preferences, and continuously optimize the recommendation effect. Summary of the Invention

[0007] This invention aims to solve the above-mentioned technical problems and provides an intelligent recommendation system and method for wig products based on hairstyle matching and user profile. Through intelligent matching of hairstyle and product, multi-dimensional product stratification, user profile construction and dynamic recommendation algorithm, it achieves accurate personalized recommendation of wig products.

[0008] To achieve the above objectives, the present invention provides the following technical solution: Technical Solution 1: A method for intelligent recommendation of wig products based on hairstyle matching and user profile, comprising the following steps: S1: Construct a hairstyle feature database and a wig product database, including: S1.1: Collect hairstyle sample images and extract hairstyle feature vectors using a multimodal deep neural network; S1.2: Collect images and attribute information of wig products, and extract multi-dimensional features of wig products; S1.3: Establish the initial association between hairstyles and wig products.

[0009] S2: Hairstyle - Wig Product Smart Matching, including: S2.1: Receive the hairstyle ID viewed by the user and obtain the feature vector of the target hairstyle; S2.2: Perform a similarity search in the wig product database and calculate the overall similarity between the wig product and the hairstyle; S2.3: Filter wig products with a comprehensive similarity exceeding the threshold and generate a candidate product set.

[0010] S3: Building user profiles, including: S3.1: Collect user historical behavior data, including browsing, clicking, favorites, purchases, and other behaviors; S3.2: Estimate the user's budget preference range based on historical purchasing behavior; S3.3: Learn users' brand and material preferences based on historical behavior; S3.4: Build user profiles, including dimensions such as budget preference, brand preference, and material preference.

[0011] S4: Multi-dimensional product stratification, including: S4.1: Based on price range, categorize candidate products into economy, standard, high-end, and luxury models; S4.2: Based on material type, categorize candidate products into synthetic fibers, blended fibers, and human hair; S4.3: Based on brand level, candidate products are categorized into affordable brands, mainstream brands, and luxury brands; S4.4: Construct a three-dimensional hierarchical matrix of products.

[0012] S5: Personalized recommendation ranking, including: S5.1: Calculate the multi-dimensional matching score for each candidate product, including hairstyle similarity, budget matching, brand matching, and material matching; S5.2: Dynamically adjust the weights of each dimension based on user profiles; S5.3: Calculate the overall recommendation score using the following formula: Score = ω1×Sim + ω2×Bud + ω3×Bra + ω4×Mat Where: Sim represents hairstyle similarity, Bud represents budget matching, Bra represents brand matching, and Mat represents material matching; ω1, ω2, ω3, and ω4 are dynamic weights, and ω1+ω2+ω3+ω4=1; S5.4: Sort by comprehensive recommendation score in descending order and generate a recommendation list.

[0013] S6: Display of recommendation results and collection of feedback, including: S6.1: Display a list of recommended wig products on the hairstyle details page, arranged by product category; S6.2: Collect user feedback data on the recommendation results, including clicks, browsing time, favorites, purchases, and other behaviors; S6.3: Update user profiles and recommendation models based on feedback data.

[0014] Furthermore, the multi-dimensional feature extraction of wig products in step S1.2 includes: - Visual feature extraction: A multimodal deep neural network is used to extract the visual feature vector of the wig product; - Attribute Feature Extraction: Converting the non-visual attributes of wigs into numerical features, including: Material encoding: Material_Encode = One_Hot(material, {synthetic fiber, blend, human hair}) Length encoding: Length_Encode = Normalize(length, [min_length, max_length]) Color encoding: Color_Encode = RGB_to_LAB(color) Price encoding: Price_Encode = Log(price + 1) Brand Encode: Brand_Encode = Brand_Tier_Encode(brand) - Feature fusion: Visual features and attribute features are fused using a feature stitching method; Feature_product = Concat(Feature_visual, Feature_attribute) Furthermore, the comprehensive similarity calculation in step S2.2 includes: (1) Visual similarity calculation: Sim_visual = cosine_similarity(Feature_hairstyle_visual, Feature_product_visual) (2) Attribute similarity calculation: Sim_length = 1 - |Length_hairstyle - Length_product| / Max_length Sim_color = cosine_similarity(Color_hairstyle, Color_product) Sim_style = Jaccard_similarity(Style_hairstyle, Style_product) (3) Calculation of overall similarity: Sim_total = λ1×Sim_visual + λ2×Sim_length + λ3×Sim_color + λ4×Sim_style Where λ1, λ2, λ3, and λ4 are weighting coefficients, with default values ​​of 0.5, 0.2, 0.2, and 0.1, respectively.

[0015] Furthermore, the user budget preference estimation method in step S3.2 includes: Method 1: Based on historical purchase price distribution Calculate the 10th percentile P10 and 90th percentile P90 of the user's historical purchase price; Budget lower bound Budget_min = P10; Budget cap Budget_max = P90; Preferred price: Budget_preferred = mean(Historical_Prices) Method 2: Budget Inference Based on Browsing Behavior Collect the set of product prices viewed by the user: {Price_browse1, Price_browse2, ...}; Calculate the weighted average of browsing prices; Budget_preferred = Σ(Price_browse_i × Time_spent_i) / Σ(Time_spent_i) Where Time_spent_i represents the duration of time a user spends on the product.

[0016] Method 3: Hybrid Estimation Method When the user has a purchase history, method one is used; When a user has no purchase history but has browsing history, method two is used; When a user has neither a purchase history nor a browsing history, the default budget range is used.

[0017] Furthermore, the brand preference learning method in step S3.3 adopts the TF-IDF model: TF_brand = Count_brand / Total_purchases IDF_brand = log(Total_users / Users_purchased_brand) Brand_Preference_Score = TF_brand × IDF_brand Where: Count_brand represents the number of times a user has purchased this brand, and Total_purchases represents the total number of purchases made by the user. Total_users represents the total number of users in the system, and Users_purchased_brand represents the number of users who have purchased that brand.

[0018] Furthermore, the dynamic weight adjustment in step S5.2 includes: Weighting adjustment rules: 1) For price-sensitive users: Increase the budget matching weight ω2; 2) For users who prioritize quality: Increase the weight of brand matching (ω3) and material matching (ω4); 3) For users who prioritize hairstyle matching: increase the hairstyle similarity weight ω1.

[0019] Price sensitivity calculation: Price_Sensitivity = Variance(Historical_Prices) / Mean(Historical_Prices) If Price_Sensitivity > Threshold_price, the user is identified as price-sensitive.

[0020] Quality pursuit degree calculation: Quality_Orientation = (High_End_Purchases / Total_Purchases) ×Brand_Tier_Avg If Quality_Orientation > Threshold_quality, the user is identified as a quality-oriented user.

[0021] Furthermore, the formulas for calculating the matching scores for each dimension in step S5.3 are as follows: (1) Hairstyle similarity score: Sim = Sim_total The value range is [0, 1]. (2) Budget matching score: If Price_product ∈ [Budget_min, Budget_max]: Bud = 1 - |Price_product - Budget_preferred| / Budget_preferred If Price_product < Budget_min: Bud = 0.6 If Price_product > Budget_max: Bud = 0.3 (3) Brand matching score: Bra = Brand_Preference_Score(brand) / Max_Brand_Preference The value range is [0, 1]. (4) Material matching score: Mat = Material_Preference_Score(material) / Max_Material_Preference The value range is [0, 1]. Furthermore, the user profile update in step S6.3 includes: Budget preference update: Budget preferences are recalculated after each purchase; Budget_preferred_new = α×Budget_preferred_old + (1-α)×Price_latest Where α is the attenuation coefficient, which defaults to 0.8.

[0022] Brand preference update: The brand preference score is updated after each purchase by the user; Brand_Preference_Score_new = Brand_Preference_Score_old + Δ Where Δ is the increment, determined based on the difference between the purchased brand and the original preference.

[0023] Recommended model parameter updates: The weight parameters are updated using an online gradient descent method. ω_new = ω_old - η×∇Loss Where η is the learning rate, and ∇Loss is the gradient of the loss function calculated based on user feedback.

[0024] Technical Solution 2: A smart recommendation system for wig products based on hairstyle matching and user profile, comprising: Hairstyle Feature Database Module: Used to store hairstyle sample images and their multimodal feature vectors; Wig Product Database Module: Used to store wig product information, multi-dimensional features, and product stratification information; Hairstyle-Product Matching Module: This module receives hairstyles browsed by users, performs similarity searches in the wig product database, and generates a set of candidate products. Product stratification module: Used to stratify candidate products in multiple dimensions such as price, material, and brand, and build a product stratification matrix; User profiling module: Used to collect historical user behavior data, build and update user profiles, including budget preferences, brand preferences, material preferences, etc.; Personalized recommendation engine: used to calculate multi-dimensional matching scores based on user profiles and product segmentation results, and generate personalized recommendation lists; Display module: Used to display a list of recommended wig products on the hairstyle details page, supporting layered display by product; Feedback optimization module: Used to collect user feedback data on recommendation results and update user profiles and recommendation model parameters.

[0025] The beneficial effects of this invention are: First, by employing multimodal feature association matching technology between hairstyles and wig products, precise association between hairstyles and wig products is achieved. Compared to manual annotation methods, the intelligent matching method of this invention improves efficiency by over 90%, achieves a matching accuracy rate of over 95%, and reduces the average time for users to find their target wig by 70%.

[0026] Secondly, through multi-dimensional adaptive product stratification technology, products are scientifically stratified based on price, material, and brand to meet the needs of users with different purchasing power. This results in a more than 40% increase in click-through rate for recommended results and a more than 35% increase in purchase conversion rate.

[0027] Third, by constructing user profiles and using a budget preference prediction mechanism, we can accurately grasp users' economic strength and consumption preferences. Based on a hybrid strategy using three budget estimation methods, the budget prediction accuracy rate reaches over 85%, and the proportion of recommended product prices within the user's acceptable range reaches over 90%.

[0028] Fourth, through a dynamic weight adjustment mechanism, the recommendation strategy is automatically adjusted based on user type (price-sensitive, quality-oriented, hairstyle-matching-priority), achieving truly personalized recommendations. The satisfaction differences between different user groups are significantly reduced, and overall user satisfaction increases by more than 30%.

[0029] Fifth, a user feedback-driven recommendation iteration and optimization system enables continuous learning and optimization of the recommendation model. The online learning mechanism ensures that the model can quickly adapt to changes in user preferences, with monthly performance improvements of 8-12%. Attached Figure Description

[0030] Figure 1 This is an overall flowchart of the intelligent recommendation method for wig products provided in an embodiment of the present invention; Figure 2 This is a flowchart illustrating the intelligent matching process between hairstyles and wigs provided in an embodiment of the present invention. Figure 3 This is a schematic diagram of multi-dimensional product layering provided in an embodiment of the present invention; Figure 4 This is a flowchart of the user profile construction process provided in an embodiment of the present invention; Figure 5 A flowchart of a personalized recommendation ranking algorithm provided in an embodiment of the present invention; Figure 6 The hairstyle details page provided in this embodiment of the invention displays recommended images of wig products; Figure 7 A flowchart illustrating the user feedback optimization mechanism provided in this embodiment of the invention; Figure 8 This is an overall architecture diagram of the intelligent recommendation system for wig products provided in an embodiment of the present invention. Detailed Implementation

[0031] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments.

[0032] Example 1: Intelligent Recommendation Method for Wig Products like Figure 1 As shown, this embodiment provides a method for intelligent recommendation of wig products based on hairstyle matching and user profile, including the following steps: Step S1: Build a hairstyle feature database and a wig product database 1.1 Construction of Hairstyle Feature Database We collected sample images of over 5000 common hairstyles, capturing images of each hairstyle from multiple angles and under various lighting conditions. We employed the same multimodal deep neural network as used in the hairdresser's recommended patent to extract hairstyle features, including: 1) Texture features: 256 dimensions 2) Structural features: 256 dimensions 3) Color characteristics: 128 dimensions The fused feature vector has a dimension of 640 and is stored in a vector database.

[0033] 1.2 Construction of Wig Product Database We collected information on over 10,000 wig products, including product images and attribute information.

[0034] Product image feature extraction: The Python code is as follows: def extract_product_features(product_image, product_info): """ Extracting multi-dimensional features of wig products """ # Visual Feature Extraction (Network with the Same Hairstyle) visual_features = multi_modal_extractor.extract(product_image) # Output 640-dimensional visual feature vector # Attribute Feature Extraction attribute_features = { 'material': encode_material(product_info['material']), # 3D one-hot 'length': normalize_length(product_info['length']), # 1D normalization 'color': rgb_to_lab(product_info['color']), # 3D LAB color 'price': np.log(product_info['price'] + 1), # 1D logarithmic encoding 'brand_tier': encode_brand_tier(product_info['brand']) # 3D one-hot } # Fusion Features fused_features = np.concatenate([ visual_features, attribute_features['material'], [attribute_features['length']], attribute_features['color'], [attribute_features['price']], attribute_features['brand_tier'] ]) return fused_features # 640 + 11 = 651 dimensions (End of Python code) 1.3 Establishment of Initial Relationships For each wig product, its similarity to all hairstyles is calculated, and the hairstyle with the highest similarity is selected as the initial associated hairstyle. Manual verification and adjustment of the association relationships are also supported.

[0035] Step S2: Hairstyle - Smart Matching of Wig Products like Figure 2 As shown, when a user browses a hairstyle, the system executes the following matching process: 2.1 Obtaining the characteristics of the target hairstyle Retrieve the feature vector from the hairstyle database based on the hairstyle ID: Feature_hairstyle = Hairstyle_DB.get_feature(hairstyle_id) 2.2 Similarity Search Perform a similarity search in the wig product database: The Python code is as follows: def match_hairstyle_with_products(hairstyle_id, threshold=0.7): """ Match hairstyles with wig products """ # Get hairstyle features hairstyle_feature = hairstyle_db.get_feature(hairstyle_id) # Search all wig products all_products = product_db.get_all_products() # Calculate similarity matched_products = [] For product in all_products: # Visual similarity sim_visual = cosine_similarity( hairstyle_feature['visual'], product.feature['visual'] ) # Attribute Similarity sim_length = calculate_length_similarity( hairstyle_feature['length'], product.feature['length'] ) sim_color = cosine_similarity( hairstyle_feature['color'], product.feature['color'] ) sim_style = calculate_style_similarity( hairstyle_feature['style'], product.feature['style'] ) # Overall Similarity sim_total = 0.5 * sim_visual + 0.2 * sim_length + \ 0.2 * sim_color + 0.1 * sim_style if sim_total >= threshold: matched_products.append({ 'product': product, 'similarity': sim_total }) # Sort by similarity matched_products.sort(key=lambda x: x['similarity'], reverse=True) return matched_products (End of Python code) 2.3 Generate a candidate product set Products with a similarity exceeding the threshold (0.7) are selected to generate a candidate product set. Typically, the top 20-30 candidate products are returned.

[0036] Step S3: Build User Profiles like Figure 4 As shown, the system constructs user profiles that include multiple dimensions.

[0037] 3.1 Collect user historical behavior data Collect users' historical behavior on the platform: 1) Browsing behavior: Products viewed by the user and the duration of their browsing. 2) Click behavior: Users click on a product to view details. 3) Favoriting behavior: Products that users favorite. 4) Purchasing behavior: The goods and prices actually purchased by the user. 3.2 Budget Preference Estimation The Python code is as follows: def estimate_budget_preference(user_id): """ Estimate users' budget preferences """ purchase_history = get_purchase_history(user_id) browsing_history = get_browsing_history(user_id) Method 1: Based on purchase history if purchase_history: prices = [p['price'] for p in purchase_history] budget = { 'min': np.percentile(prices, 10), 'max': np.percentile(prices, 90), 'preferred': np.mean(prices) } return budget # Method 2: Based on browsing history (weighted average) elif browsing_history: total_time = 0 weighted_sum = 0 for browse in browsing_history: price = browse['product']['price'] time_spent = browse['time_spent'] weighted_sum += price * time_spent total_time += time_spent preferred_price = weighted_sum / total_time if total_time > 0else 500 budget = { 'min': preferred_price * 0.5, 'max': preferred_price * 2, 'preferred': preferred_price } return budget # Method 3: Default Budget else: return { 'min': 100, 'max': 1000, 'preferred': 500 } (End of Python code) 3.3 Brand Preference Learning Using the TF-IDF model to learn brand preferences: The Python code is as follows: def learn_brand_preference(user_id): """ Learning user brand preferences """ purchases = get_purchases(user_id) if not purchases: return {} # Count the number of times users purchase each brand brand_counts = {} for p in purchases: brand = p['product']['brand'] brand_counts[brand] = brand_counts.get(brand, 0) + 1 total_purchases = len(purchases) # Calculate TF-IDF brand_preferences = {} for brand, count in brand_counts.items(): #TF tf = count / total_purchases # IDF total_users = get_total_users() users_purchased_brand = count_users_purchased(brand) idf = math.log(total_users / users_purchased_brand) # TF-IDF brand_preferences[brand] = tf * idf # Normalization max_score = max(brand_preferences.values()) if brand_preferenceselse 1 for brand in brand_preferences: brand_preferences[brand] / = max_score return brand_preferences (End of Python code) 3.4 Material Preference Learning The Python code is as follows: def learn_material_preference(user_id): """ Learning user material preferences """ purchases = get_purchases(user_id) if not purchases: return {'Synthetic Fiber': 0.33, 'Blended': 0.33, 'Human Hair': 0.34} material_counts = {'Synthetic fiber': 0, 'Mixed': 0, 'Human hair': 0} for p in purchases: material = p['product']['material'] if material in material_counts: material_counts[material] += 1 total = sum(material_counts.values()) for material in material_counts: material_counts[material] / = total return material_counts (End of Python code) Step S4: Multi-dimensional Product Layering like Figure 3 As shown, candidate products are stratified in multiple dimensions.

[0038] 4.1 Price Stratification The Python code is as follows: def classify_by_price(products): """ Tiered by price """ price_tiers = { 'economy': [], # Economy 'standard': [], # Standard model 'premium': [], # Premium model 'luxury': [] # Luxury model } for product in products: price = product['price'] if price < 200: price_tiers['economy'].append(product) elif price < 800: price_tiers['standard'].append(product) elif price < 2000: price_tiers['premium'].append(product) else: price_tiers['luxury'].append(product) return price_tiers (End of Python code) 4.2 Material layering The Python code is as follows: def classify_by_material(products): """ Layered by material """ material_tiers = { 'synthetic': [], # chemical fiber 'mixed': [], # Mixed 'human': [] # Real person } for product in products: material = product['material'] if material in material_tiers: material_tiers[material].append(product) return material_tiers (End of Python code) 4.3 Brand Segmentation The Python code is as follows: def classify_by_brand(products): """ stratified by brand level """ brand_tiers = { 'budget': [], # Affordable Brands 'mainstream': [], # Mainstream brands 'luxury': [] # Luxury Brands } brand_tier_map = load_brand_tier_mapping() for product in products: brand = product['brand'] tier = brand_tier_map.get(brand, 'mainstream') if tier in brand_tiers: brand_tiers[tier].append(product) return brand_tiers (End of Python code) Step S5: Personalized Recommendation Ranking like Figure 5 As shown, calculate the overall recommendation score for each product.

[0039] 5.1 Determine User Type The Python code is as follows: def classify_user_type(user_profile): """ Determine user type """ user_type = { 'price_sensitive': False, 'quality_oriented': False, 'style_priority': False } purchases = get_purchases(user_profile['user_id']) if purchases: prices = [p['price'] for p in purchases] # Price Sensitivity price_variance = np.var(prices) / np.mean(prices) user_type['price_sensitive'] = (price_variance > 0.5) # Quality pursuit high_end_ratio = sum(1 for p in purchases if p['price'] >1000) / len(purchases) avg_brand_tier = np.mean([p['product']['brand_tier'] for p inpurchases]) quality_score = high_end_ratio * avg_brand_tier user_type['quality_oriented'] = (quality_score > 0.3) # Default is hairstyle priority type if not (user_type['price_sensitive'] or user_type['quality_oriented']): user_type['style_priority'] = True return user_type (End of Python code) 5.2 Dynamic Weight Adjustment The Python code is as follows: def adjust_weights(user_type): """ Adjust weights based on user type """ weights = { 'similarity': 0.4, # Hairstyle similarity 'budget': 0.3, # Budget match 'brand': 0.2, # Brand fit 'material': 0.1 # Material matching degree } if user_type['price_sensitive']: weights['budget'] = 0.5 weights['similarity'] = 0.3 weights['brand'] = 0.1 weights['material'] = 0.1 elif user_type['quality_oriented']: weights['brand'] = 0.35 weights['material'] = 0.25 weights['similarity'] = 0.25 weights['budget'] = 0.15 elif user_type['style_priority']: weights['similarity'] = 0.6 weights['budget'] = 0.2 weights['brand'] = 0.1 weights['material'] = 0.1 return weights (End of Python code) 5.3 Calculate the matching score The Python code is as follows: def calculate_matching_scores(products, user_profile, weights): """ Calculate the multi-dimensional matching score of the product """ scored_products = [] for product in products: # Hairstyle similarity sim = product['similarity'] # Budget Matching price = product['price'] budget = user_profile['budget_preference'] if budget['min'] <= price <= budget['max']: bud = 1 - abs(price - budget['preferred']) / budget['preferred'] elif price < budget['min']: bud = 0.6 else: bud = 0.3 # Brand Match brand = product['brand'] brand_prefs = user_profile.get('brand_preference', {}) bra = brand_prefs.get(brand, 0.5) # Material matching material = product['material'] material_prefs = user_profile.get('material_preference', {}) mat = material_prefs.get(material, 0.33) # Overall Score total_score = ( weights['similarity'] * sim + weights['budget'] * bud + weights['brand'] * bra + weights['material'] * mat ) scored_products.append({ 'product': product, 'scores': { 'similarity': sim, 'budget': bud, 'brand': bra, 'material': mat, 'total': total_score } }) # Sort by total score scored_products.sort(key=lambda x: x['scores']['total'], reverse=True) return scored_products (End of Python code) Step S6: Displaying Recommendation Results and Collecting Feedback 6.1 Display the recommendation list The hairstyle details page displays recommended wig products, arranged by product category, such as... Figure 6 As shown.

[0040] 6.2 Collecting User Feedback Collect user feedback on the recommendation results: 1) Click behavior: Users click to view product details 2) Browsing time: The amount of time a user spends on a product page. 3) Add to favorites behavior: Users add items to their favorites list. 4) Purchase behavior: Users purchase goods 6.3 Update user profiles and models like Figure 7 As shown, the system will be updated based on user feedback.

[0041] The Python code is as follows: def update_user_profile(user_id, feedback): """ Update user profile """ profile = get_user_profile(user_id) # Update budget preferences if feedback['type'] == 'purchase': price = feedback['product']['price'] old_preferred = profile['budget_preference']['preferred'] new_preferred = 0.8 * old_preferred + 0.2 * price profile['budget_preference']['preferred'] = new_preferred # Update brand preferences if feedback['type'] in ['click', 'purchase']: brand = feedback['product']['brand'] increment = 0.05 if feedback['type'] == 'purchase' else 0.01 if brand in profile['brand_preference']: profile['brand_preference'][brand] += increment else: profile['brand_preference'][brand] = increment # Normalization max_score = max(profile['brand_preference'].values()) for b in profile['brand_preference']: profile['brand_preference'][b] / = max_score # Update material preferences (similar to brand preferences) save_user_profile(user_id, profile) def update_model_weights(feedback_list): """ Online model weight update """ # Calculate loss based on feedback loss = compute_loss(feedback_list) # Gradient Descent Update learning_rate = 0.01 For weight in weights: weights[weight] -= learning_rate * gradients[weight] save_weights(weights) (End of Python code) Example 2: Intelligent Recommendation System for Wig Products like Figure 8 As shown, this embodiment provides an intelligent recommendation system for wig products based on hairstyle matching and user profiles, including the following modules: 1. Hairstyle Feature Database Module The hairstyle feature database is implemented using the Milvus vector database, storing feature vectors for over 5000 hairstyles.

[0042] 2. Wig Product Database Module The wig product database is implemented using MongoDB, storing product information, feature vectors, and hierarchical information.

[0043] 3. Hairstyle-Product Matching Module The hairstyle-product matching module is implemented using Python and FastAPI, providing similarity search services.

[0044] 4. Product Layering Module The product stratification module enables multi-dimensional classification, including price stratification, material stratification, and brand stratification.

[0045] 5. User Profile Module The user profiling module collects and processes user behavior data to build multi-dimensional user profiles.

[0046] 6. Personalized Recommendation Engine The personalized recommendation engine calculates product matching scores and generates a recommendation list.

[0047] 7. Display Module The display module provides a front-end interface to show the recommendation results.

[0048] 8. Feedback Optimization Module The feedback optimization module collects user feedback and updates user profiles and recommendation models.

[0049] System performance indicators: 1) Hairstyle-product matching accuracy: ≥95% 2) Budget forecast accuracy: ≥85% 3) Recommended click-through rate: ≥25% (40% improvement over the benchmark) 4) Purchase conversion rate: ≥8% (35% improvement over the benchmark) 5) System response time: ≤300ms The above description is only a specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the protection scope of the present invention.

Claims

1. A method for intelligent recommendation of wig products based on hairstyle matching and user profile, characterized in that, Includes the following steps: S1: Construct a hairstyle feature database and a wig product database, including: S1.1: Collect hairstyle sample images and extract hairstyle feature vectors using a multimodal deep neural network; S1.2: Collect images and attribute information of wig products, and extract multi-dimensional features of wig products, including visual features and attribute features; S1.3: Establish the initial association between hairstyles and wig products. S2: Hairstyle - Wig Product Smart Matching, including: S2.1: Receive the hairstyle ID viewed by the user and obtain the feature vector of the target hairstyle; S2.2: Perform a similarity search in the wig product database and calculate the overall similarity between the wig product and the hairstyle; S2.3: Filter wig products with a comprehensive similarity exceeding the threshold and generate a candidate product set. S3: Building user profiles, including: S3.1: Collect user historical behavior data, including browsing, clicking, favorites, and purchasing behavior; S3.2: Estimate the user's budget preference range based on historical purchase or browsing behavior; S3.3: Learn users' brand and material preferences based on historical behavior; S3.4: Build user profiles, including dimensions such as budget preference, brand preference, and material preference. S4: Multi-dimensional product stratification, including: S4.1: Based on price range, categorize candidate products into economy, standard, high-end, and luxury models; S4.2: Based on material type, categorize candidate products into synthetic fibers, blended fibers, and human hair; S4.3: Based on brand level, candidate products are categorized into affordable brands, mainstream brands, and luxury brands; S4.4: Construct a three-dimensional hierarchical matrix of products. S5: Personalized recommendation ranking, including: S5.1: Calculate the multi-dimensional matching score for each candidate product, including hairstyle similarity, budget matching, brand matching, and material matching; S5.2: Determine user type based on user profile and dynamically adjust the weight of each dimension; S5.3: Calculate the overall recommendation score using the following formula: Score = ω1×Sim + ω2×Bud + ω3×Bra + ω4×Mat Where: Sim represents hairstyle similarity, Bud represents budget matching, Bra represents brand matching, and Mat represents material matching; ω1, ω2, ω3, and ω4 are dynamic weights, and ω1+ω2+ω3+ω4=1; S5.4: Sort by comprehensive recommendation score in descending order and generate a recommendation list. S6: Display of recommendation results and collection of feedback, including: S6.1: Display a list of recommended wig products on the hairstyle details page, arranged by product category; S6.2: Collect user feedback data on recommendation results, including clicks, browsing time, favorites, and purchase behavior; S6.3: Update user profiles and recommendation model parameters based on feedback data.

2. The method according to claim 1, characterized in that, The multi-dimensional feature extraction of wig products in step S1.2 includes: Visual feature extraction: A multimodal deep neural network is used to extract the visual feature vectors of the wig products; Attribute feature extraction: Convert the non-visual attributes of wig products into numerical features, including material code, length code, color code, price code, and brand code; Feature fusion: Visual features and attribute features are fused using a feature splicing method to generate a comprehensive feature vector for wig products.

3. The method according to claim 1, characterized in that, The comprehensive similarity calculation in step S2.2 includes: Visual similarity calculation: Cosine similarity is used to calculate the visual feature similarity between hairstyles and wig products; Attribute similarity calculation: Calculate length similarity, color similarity, and style similarity respectively; Overall similarity calculation: Sim_total = λ1×Sim_visual + λ2×Sim_length + λ3×Sim_color + λ4×Sim_style Where λ1, λ2, λ3, and λ4 are weighting coefficients, with default values ​​of 0.5, 0.2, 0.2, and 0.1, respectively.

4. The method according to claim 1, characterized in that, The user budget preference estimation in step S3.2 includes: Method 1: Based on historical purchase price distribution Calculate the 10th and 90th percentiles of the user's historical purchase prices as the budget range; Budget lower limit = P10, budget upper limit = P90, preferred price = mean(Historical_Prices) Method 2: Budget Inference Based on Browsing Behavior Budget_preferred = Σ(Price_browse_i × Time_spent_i) / Σ(Time_spent_i) Where Price_browse_i is the price of the browsed item, and Time_spent_i is the duration of the browsing session. Method 3: Hybrid Estimation Method When the user has a purchase history, method one is used; When a user has no purchase history but has browsing history, method two is used; When a user has neither a purchase history nor a browsing history, the default budget range is used.

5. The method according to claim 1, characterized in that, The brand preference learning in step S3.3 adopts the TF-IDF model: TF_brand = Count_brand divided by Total_purchases IDF_brand = log(Total_users divided by Users_purchased_brand) Brand_Preference_Score = TF_brand * IDF_brand Where: Count_brand represents the number of times a user has purchased this brand, and Total_purchases represents the total number of purchases made by the user. Total_users represents the total number of users in the system, and Users_purchased_brand represents the number of users who have purchased that brand.

6. The method according to claim 1, characterized in that, The user type determination in step S5.2 includes: Price-sensitive users: Calculate price sensitivity: Price_Sensitivity = Variance(Historical_Prices) / Mean(Historical_Prices). If Price_Sensitivity > Threshold_price, the user is identified as a price-sensitive user. For quality-oriented users: Calculate the quality orientation: Quality_Orientation = (High_End_Purchases / Total_Purchases) × Brand_Tier_Avg. If Quality_Orientation > Threshold_quality, the user is identified as a quality-oriented user. Hairstyle matching priority users: If the conditions of price sensitivity and quality pursuit are not met, they are judged as hairstyle matching priority users.

7. The method according to claim 1, characterized in that, The dynamic weight adjustment rule in step S5.2 is as follows: For price-sensitive users: ω1=0.3, ω2=0.5, ω3=0.1, ω4=0.1; For users who prioritize quality: ω1=0.25, ω2=0.15, ω3=0.35, ω4=0.25; For users who prioritize hairstyle matching: ω1=0.6, ω2=0.2, ω3=0.1, ω4=0.

1.

8. The method according to claim 1, characterized in that, The formulas for calculating the matching scores for each dimension in step S5.3 are as follows: Hairstyle similarity score: Sim = Sim_total, with a value range of [0,1]; Budget match score: If Price_product ∈ [Budget_min, Budget_max]: Bud = 1 - |Price_product - Budget_preferred| / Budget_preferred If Price_product < Budget_min: Bud = 0.6 If Price_product > Budget_max: Bud = 0.3 Brand fit score: Bra = Brand_Preference_Score(brand) / Max_Brand_Preference, with a value range of [0,1]. Material matching score: Mat = Material_Preference_Score(material) / Max_Material_Preference, with a value range of [0,1].

9. The method according to claim 1, characterized in that, The user profile update in step S6.3 includes: Budget preference update: Budget_preferred_new = α×Budget_preferred_old + (1-α)×Price_latest Where α is the attenuation coefficient, with a default value of 0.8, and Price_latest is the latest purchase price; Brand preference update: Brand_Preference_Score_new = Brand_Preference_Score_old + Δ Where Δ is the increment, determined based on the difference between the purchased brand and the original preference; Recommended model parameter updates: The weight parameters are updated using an online gradient descent method, ω_new = ω_old - η×∇Loss. Where η is the learning rate, and ∇Loss is the gradient of the loss function calculated based on user feedback.

10. A smart recommendation system for wig products based on hairstyle matching and user profile, characterized in that, include: Hairstyle Feature Database Module: Used to store hairstyle sample images and their multimodal feature vectors; Wig Product Database Module: Used to store wig product information, multi-dimensional features, and product stratification information; Hairstyle-Product Matching Module: This module receives hairstyles browsed by users, performs similarity searches in the wig product database, and generates a set of candidate products. Product stratification module: Used to stratify candidate products in multiple dimensions such as price, material, and brand, and build a product stratification matrix; User profiling module: used to collect historical user behavior data, build and update user profiles, including budget preferences, brand preferences, and material preferences; Personalized recommendation engine: used to calculate multi-dimensional matching scores based on user profiles and product segmentation results, and generate personalized recommendation lists; Display module: Used to display a list of recommended wig products on the hairstyle details page, supporting layered display by product; Feedback optimization module: Used to collect user feedback data on recommendation results and update user profiles and recommendation model parameters.

Images