A method for AI model maternal and child assistant to go from question-answering to accurate recommendation

Abstract

The present invention discloses a method for an AI model maternal and infant assistant from question-answering to precise recommendation, which belongs to the field of artificial intelligence technology, including user demand input and analysis, user portrait construction and dynamic update, product recommendation and optimization, feedback optimization and strategy adjustment, and solves the technical problems of ensuring the real-time and accuracy of user portraits, being able to flexibly respond to changes in user needs and interests, and improving the accuracy of recommendations and user satisfaction. The present invention ensures the precise understanding and processing of complex needs, ensures the real-time and accuracy of user portraits, is able to flexibly respond to changes in user needs and interests, and realizes more precise and personalized recommendations. Improve the accuracy of recommendations and user satisfaction, so that users can understand why a certain product is recommended, thereby enhancing users' trust in the recommendation system.

Description

Technical Field

[0001] The present invention belongs to the field of artificial intelligence technology, and in particular relates to a method for an AI model maternal and child assistant to go from question-answering to accurate recommendation. Background Art

[0002] With the rapid development of artificial intelligence technology, AI-based maternal and child assistant applications have been widely introduced into the market, mainly including dialogues with users through natural language processing (NLP) technology, providing product recommendations, parenting knowledge, health monitoring and other services. Existing AI maternal and child assistants mainly use big data analysis and machine learning algorithms to help parents make decisions based on the baby's needs (such as feeding, sleep, health, etc.).

[0003] Deficiencies of existing technology:

[0004] Static nature and limitations of user portraits: Existing user portrait construction mostly relies on historical user behavior and past data, lacks a dynamic adjustment mechanism, and cannot promptly reflect changes in user needs and interests. Especially when user needs change frequently, traditional portraits are slow to update, which can easily lead to inaccurate recommendations.

[0005] The singleness of the recommendation system: Most existing recommendation systems only use a single recommendation algorithm with low algorithm optimization capabilities, resulting in limited recommendation accuracy and personalization.

[0006] Lack of explainability: Although existing recommendation systems can provide product recommendations, their reasons for recommendations often lack explainability. Users cannot clearly understand why a certain product is recommended to them, which may reduce users' trust and satisfaction. Summary of the invention

[0007] The purpose of this invention is to provide a method for an AI model maternal and child assistant from question and answer to accurate recommendation, which solves the technical problems of ensuring the real-time and accuracy of user portraits, being able to flexibly respond to changes in user needs and interests, and improving the accuracy of recommendations and user satisfaction.

[0008] To achieve the above object, the present invention adopts the following technical solution:

[0009] A method for an AI model maternal and child assistant from question-answering to accurate recommendation includes the following steps:

[0010] Step 1: The user enters the requirements in the dialogue interface. The requirements are natural language text. The semantic analysis module understands the requirements through semantic analysis and identifies the core requirements using NLP technology.

[0011] If the demand is vague, the semantic analysis module uses the BERT model to analyze and further interact with the user to clarify the demand; the semantic analysis module extracts keywords from the core demand, generates demand tags, and analyzes user emotional preferences;

[0012] Step 2: The data analysis module builds a preliminary portrait by analyzing the user's historical behavior and preferences, and dynamically updates the portrait based on real-time feedback. It continuously optimizes the label system in the user portrait to ensure the dynamic update and accuracy of the user portrait, monitor changes in user interests, and adjust the portrait and recommendation strategy;

[0013] Step 3: The product recommendation module generates a preliminary candidate product pool based on demand tags through a weighted matching algorithm, and further optimizes product recommendations by combining collaborative filtering and deep learning models;

[0014] Step 4: The feedback optimization module establishes a feedback mechanism, collects feedback through user behavior, and dynamically adjusts the recommendation strategy based on the feedback. The feedback mechanism provides explainability to help users understand the reasons for the recommendation.

[0015] Preferably, when executing step 1, the following steps are specifically included:

[0016] Step 1-1: Demand input, specifically, the user enters the demand in text form in the APP dialog box, and the semantic analysis module understands the user's core demand through semantic analysis and generates the demand text;

[0017] Step 1-2: Demand fuzziness processing: When the demand is ambiguous, the semantic parsing module uses the BERT model to perform semantic analysis on the user input and extract the potential demand intention;

[0018] If the intent is still unclear, the semantic parsing module interacts with the user through conversational AI, asking supplementary questions to clarify the needs;

[0019] Step 1-3: Requirement label generation, specifically, the semantic parsing module automatically extracts keywords from the requirement text and generates requirement labels;

[0020] Introduce sentiment analysis module to analyze user preferences for products based on their sentiment input

[0021] Steps 1-4: Requirements preprocessing and mapping, specifically the semantic parsing module classifies and standardizes the requirement labels, and uses the label mapping algorithm to map the natural language requirements to the predefined categories in the knowledge base.

[0022] Preferably, when executing step 2, the following steps are specifically included:

[0023] Step 2-1: User historical data analysis. Specifically, the data analysis module uses collaborative filtering algorithms to analyze user historical behaviors, generate preliminary user portraits, and reveal user interest preferences. The collaborative filtering algorithm predicts user interests based on user historical ratings and generates recommendation candidates based on similar user behaviors. The specific formula is as follows:

[0024] ;

[0025] in, is the predicted rating of user u for product i, which reveals the user’s interest preference for product i; is the average rating of user u; is the neighbor set of user u; is the rating of user v on product i; is the average rating of user v; is the similarity between user u and user v;

[0026] ;

[0027] in, is the rating of product i by user u; is the rating of user v on product i; n is the number of products, and I is the set of products rated by both users u and v;

[0028] Step 2-2: Dynamic user portrait construction, specifically dynamically updating the user portrait through real-time feedback to ensure that the portrait always accurately reflects the user's current needs and interests. Real-time feedback includes clicks, browsing and purchasing behaviors;

[0029] Step 2-3: Label update and optimization. Specifically, the data analysis module continuously optimizes the label system in the user portrait based on the new user behavior and demand data to ensure the dynamic update and accuracy of the user portrait.

[0030] Step 2-4: Monitor and provide feedback on changes in user interests. Specifically, the data analysis module monitors changes in user interests in real time, analyzes user behavior, and adjusts user profiles in a timely manner.

[0031] Preferably, when executing step 3, the following steps are specifically included:

[0032] Step 3-1: The product recommendation module screens products based on user demand tags through a weighted matching algorithm to generate a preliminary candidate product pool. The weighted matching algorithm calculates the matching score between products and user needs based on user preferences for product features. The specific formula is as follows:

[0033] ;

[0034] in: The matching score of user u to business i; is the weight of feature f; is the value of product i on feature f (such as price, category, etc.); User u’s preference on feature f (such as price sensitivity, category preference, etc.);

[0035] Step 3-2: Product recommendation module Based on the method of combining collaborative filtering and deep learning model, the product recommendation module further optimizes the product candidate pool as follows:

[0036] Step 3-2-1: The collaborative filtering model adjusts the product candidate pool through user similarity and item similarity, including the user-product rating prediction formula and weighted summation formula:

[0037] User-item rating prediction formula:

[0038] Assuming there are user u and product i, the goal of collaborative filtering is to predict user u's rating of product i:

[0039] ;

[0040] in, is the predicted rating of user u for product i; is the average rating of user u; is the rating of user v on product i, is the neighbor set of user u; is the average rating of user v; is the similarity between user u and user v;

[0041] Calculation method:

[0042] ;

[0043] in, is the rating of product i by user u; is the rating of user v on product i; n is the number of products, and I is the set of products rated by both users u and v;

[0044] Weighted sum formula:

[0045] The weighted sum formula is used to accurately combine the similarities between products and users, specifically to adjust the recommendation effect of the collaborative filtering model:

[0046] ;

[0047] is the weight between user u and user v;

[0048] Step 3-2-2: The deep learning model is specifically a matrix decomposition model. Assume is the user embedding matrix, is the item embedding matrix, where m is the number of users, n is the number of items, and k is the embedding dimension;

[0049] Given user u and item i, the formula for predicting the rating is as follows:

[0050] ;

[0051] is the embedding vector of user u; is the embedding vector of product i;

[0052] The goal of the matrix factorization model is to learn the embedding vectors of users and items by minimizing the loss function:

[0053] ;

[0054] Among them, S is a rating dataset (i.e., user-item rating pairs); is the actual rating of product i by user u; Predict the rating of user u for product i; is a regularization parameter to prevent overfitting;

[0055] Step 3-2-3: Collaborative filtering and deep learning models are combined through weighted fusion method:

[0056] ;

[0057] is the recommendation score calculated by the collaborative filtering model; is the recommendation score obtained by the calculation result of the deep learning model; It is a parameter that adjusts the output weights of collaborative filtering and deep learning models;

[0058] Step 3-3: The product recommendation module solves the cold start problem of new users or new products based on the content recommendation algorithm. Specifically:

[0059] Generate an initial profile based on the user’s basic information;

[0060] Combined with the user's browsing history, the user profile is further improved. The user profile includes the user's preference for a certain feature of the product;

[0061] For new products, during the cold start phase, we use a content-based recommendation algorithm that combines product features and the user's initial profile to make recommendations:

[0062] ;

[0063] in, Predict the rating of user u for product i;

[0064] Product characteristics:

[0065] ;

[0066] in, represents the jth feature of product i, the threshold of j ranges from 1 to m, and m is the total number of features;

[0067] The preference for the feature is expressed as:

[0068] ;

[0069] in, represents the preference of the jth feature of user u, where the threshold of j ranges from 1 to m;

[0070] and Represent the modulus length of user portrait and product feature vector respectively; Represents the dot product of the user portrait and the product feature vector;

[0071] Step 3-4: The product recommendation module optimizes the recommendation strategy through the reinforcement learning algorithm and adopts the ε-greedy method to avoid the recommendation being concentrated on a certain category of products.

[0072] Preferably, when executing step 4, the following steps are specifically included:

[0073] Step 4-1: The feedback optimization module establishes a feedback collection mechanism, where users provide feedback on recommended products through clicks, ratings, and purchase behaviors; the feedback optimization module collects diverse user feedback through guided feedback question templates;

[0074] Step 4-2: The feedback optimization module uses the reinforcement learning algorithm Q-learning to dynamically optimize the recommendation strategy based on user feedback;

[0075] Step 4-3: The feedback optimization module provides explainable feedback to help users understand why a certain product is recommended;

[0076] Interpretable feedback includes feature importance visualizations and decision tree-based explanations:

[0077] Feature importance visualization uses the interpretability technique LIME to display the weight of each feature in the recommendation decision;

[0078] The decision tree-based explanation specifically uses the decision tree model to analyze the relationship between user behavior and recommended products, and uses the tree model path to show why a certain product is recommended to the user;

[0079] Step 4-4: For cold start users, the feedback module collects their initial preferences through guided feedback and adjusts the recommended content.

[0080] The method of the AI ​​model maternal and infant assistant from question and answer to accurate recommendation described in the present invention solves the technical problems of ensuring the real-time and accuracy of user portraits, being able to flexibly respond to changes in user needs and interests, and improving the accuracy of recommendations and user satisfaction. The present invention introduces the BERT model and conversational AI, which can intelligently analyze and clarify the user's vague needs, further accurately extract user intentions and generate demand labels, ensure accurate understanding and processing of complex needs, and adopt a strategy of dynamically updating user portraits to continuously optimize and adjust user portraits based on real-time feedback (such as user clicks, purchases, and browsing behaviors). This ensures the real-time and accuracy of user portraits, and is able to flexibly respond to changes in user needs and interests. By combining collaborative filtering, deep learning (such as matrix decomposition), and reinforcement learning, a combination of multiple recommendation algorithms achieves more accurate and personalized recommendations. According to different scenarios (such as cold start, changes in user interests), the recommendation strategy is flexibly adjusted to improve the accuracy of recommendations and user satisfaction. This solution introduces LIME (local interpretable model-agnostic explanation) and decision tree-based interpretability technology, so that users can understand why a certain product is recommended, thereby enhancing users' trust in the recommendation system. BRIEF DESCRIPTION OF THE DRAWINGS

[0081] Figure 1 It is the main flow chart of the present invention;

[0082] Figure 2 is a flow chart of step 1 of the present invention;

[0083] Figure 3 is a flow chart of step 2 of the present invention;

[0084] Figure 4 is a flow chart of step 3 of the present invention;

[0085] Figure 5 It is a flow chart of step 4 of the present invention. DETAILED DESCRIPTION

[0086] Depend on Figure 1-Figure 5 The method of an AI model maternal and child assistant from question answering to accurate recommendation includes the following steps:

[0087] Step 1: The user enters the requirements in the dialogue interface. The requirements are natural language text. The semantic analysis module understands the requirements through semantic analysis and identifies the core requirements using NLP technology.

[0088] If the demand is vague, the semantic analysis module uses the BERT model to analyze and further interact with the user to clarify the demand; the semantic analysis module extracts keywords from the core demand, generates demand tags, and analyzes user emotional preferences;

[0089] When executing step 1, the specific steps include:

[0090] Step 1-1: Demand input, specifically, the user inputs the demand in text form in the APP dialog box, and the semantic analysis module understands the user's core demand through semantic analysis and generates the demand text; in this embodiment, semantic analysis is performed through natural language processing (NLP).

[0091] The following are some usage scenarios of this embodiment. For example, a new mother inputs in the dialogue interface: "My baby always refuses to fall asleep at night, what should I do?" The semantic analysis module will identify the key requirement "baby refuses to fall asleep" and identify the user's anxiety through the sentiment analysis module, and infer that the user may need relevant suggestions or products to help the baby sleep.

[0092] Step 1-2: Demand fuzziness processing: When the demand is ambiguous, the semantic parsing module uses the BERT model to perform semantic analysis on the user input and extract the potential demand intention;

[0093] If the intent is still unclear, the semantic parsing module interacts with the user through conversational AI, asking supplementary questions to clarify the needs;

[0094] In this embodiment, when the user inputs "My baby has been crying a lot recently, I wonder if he is teething", the semantic parsing module uses the BERT model to analyze and identify "crying" and "teething" as potential demand intentions. The semantic parsing module may further ask: "Does your baby bite things or drool?" to help users clarify their needs and ensure accurate recommendations.

[0095] In this embodiment, the conversational AI can adopt the Rasa conversational AI framework to achieve local deployment.

[0096] Step 1-3: Requirement label generation, specifically, the semantic parsing module automatically extracts keywords from the requirement text and generates requirement labels;

[0097] The sentiment analysis module is introduced to analyze the user's preference for products based on the sentiment tendencies input by the user.

[0098] In this embodiment, an existing sentiment analysis library (such as TextBlob or VADER) is used as a custom component and inserted into the NLU pipeline of Rasa to process the text and return the result of sentiment analysis.

[0099] For example, if the sentiment analysis results of a user indicate anxiety, the semantic analysis module may focus on recommending products that can relieve the baby’s discomfort, such as teethers and pacifiers, while displaying emotional copy to help the user reduce anxiety, such as “It is common for babies to feel some discomfort during teething. You can try these products to relieve your baby’s discomfort.”

[0100] Steps 1-4: Requirements preprocessing and mapping, specifically the semantic parsing module classifies and standardizes the requirement labels, and uses the label mapping algorithm to map the natural language requirements to the predefined categories in the knowledge base.

[0101] For example, a user asks: "My baby is a little constipated, can you recommend some foods that help digestion?" The semantic analysis module maps this demand to the "digestive health" category, such as the knowledge base indicating that the category is divided into recommending suitable baby probiotics, supplementary foods that assist digestion, and other related products. In this embodiment, the knowledge base content is preset content.

[0102] Step 2: The data analysis module builds a preliminary portrait by analyzing the user's historical behavior and preferences, and dynamically updates the portrait based on real-time feedback. It continuously optimizes the label system in the user portrait to ensure the dynamic update and accuracy of the user portrait, monitor changes in user interests, and adjust the portrait and recommendation strategy;

[0103] When executing step 2, the specific steps include:

[0104] Step 2-1: User historical data analysis. Specifically, the data analysis module uses collaborative filtering algorithms to analyze user historical behaviors, generate preliminary user portraits, and reveal user interest preferences. The collaborative filtering algorithm predicts user interests based on user historical ratings and generates recommendation candidates based on similar user behaviors. The specific formula is as follows:

[0105] ;

[0106] in, is the predicted rating of user u for product i, which reveals the user’s interest preference for product i; is the average rating of user u; is the neighbor set of user u; is the rating of user v on product i; is the average rating of user v; is the similarity between user u and user v;

[0107] ;

[0108] in, is the rating of product i by user u; is the rating of user v on product i; n is the number of products, and I is the set of products rated by both users u and v;

[0109] This embodiment collects the user's purchase history, ratings (if any) or browsing behavior (e.g., purchased diaper A, high-end milk powder B). Then, a collaborative filtering algorithm is used to analyze similar users (e.g., users who have purchased diaper A and milk powder B) to infer the user's possible interest in other products.

[0110] User A's rating (or purchase behavior): Diaper A: Rated 5; Milk powder B: Rated 4;

[0111] User B's rating (or purchase behavior): Diaper A: Rated 4; Milk powder B: Rated 5;

[0112] Through collaborative filtering, the data analysis module will find that the behaviors of user A and user B are similar, and recommend other products that user A may like. For example, if user B also purchased stroller C, the data analysis module will recommend it to user A.

[0113] Step 2-2: Dynamic user portrait construction, specifically dynamically updating the user portrait through real-time feedback to ensure that the portrait always accurately reflects the user's current needs and interests. Real-time feedback includes clicks, browsing and purchasing behaviors;

[0114] Step 2-3: Label update and optimization. Specifically, the data analysis module continuously optimizes the label system in the user portrait based on the new user behavior and demand data to ensure the dynamic update and accuracy of the user portrait.

[0115] In this embodiment, the user's behavior specifically includes clicking, purchasing, collecting, and rating.

[0116] Step 2-4: Monitor and provide feedback on changes in user interests. Specifically, the data analysis module monitors changes in user interests in real time, analyzes user behavior, and adjusts user profiles in a timely manner.

[0117] Step 3: The product recommendation module generates a preliminary candidate product pool based on demand tags through a weighted matching algorithm, and further optimizes product recommendations by combining collaborative filtering and deep learning models;

[0118] When executing step 3, the specific steps include:

[0119] Step 3-1: The product recommendation module screens products based on user demand tags through a weighted matching algorithm to generate a preliminary candidate product pool. The weighted matching algorithm calculates the matching score between products and user needs based on user preferences for product features. The specific formula is as follows:

[0120] ;

[0121] in: The matching score of user u to business i; is the weight of feature f; is the value of product i on feature f (such as price, category, etc.); User u’s preference on feature f (such as price sensitivity, category preference, etc.);

[0122] In this embodiment, it is assumed that user 1 inputs a demand to purchase a diaper suitable for a newborn baby and is relatively sensitive to budget.

[0123] User 1 has browsed the following products: diapers, rating: 4; milk powder, rating: 5; baby strollers, rating: 2.

[0124] The weighted matching algorithm scores each product:

[0125] Assumption: User 1's demand tags include: price (more sensitive, weight is 0.7) category (prefer diapers, weight is 0.3);

[0126] Predefined product features: Product A (diapers): price = 30, category = diapers; Product B (milk powder): price = 60, category = milk powder; Product C (baby stroller): price = 200, category = baby stroller;

[0127] Weighted matching score calculation:

[0128] For product A (diapers): matching score (u, A) = 0.7 × 30 + 0.3 × 1 = 21.3;

[0129] For product B (milk powder): matching score (u, B) = 0.7 × 60 + 0.3 × 0 = 42 + 0 = 42;

[0130] For product C (baby carriage): matching score (u, C) = 0.7 × 200 + 0.3 × 0 = 140;

[0131] Recommendation results: Based on the weighted matching score, product A (diapers) will be selected as the best recommended product because it has the highest matching score.

[0132] Step 3-2: Product recommendation module Based on the method of combining collaborative filtering and deep learning model, the product recommendation module further optimizes the product candidate pool as follows:

[0133] Step 3-2-1: The collaborative filtering model adjusts the product candidate pool through user similarity and item similarity, including the user-product rating prediction formula and weighted summation formula:

[0134] User-item rating prediction formula:

[0135] Assuming there are user u and product i, the goal of collaborative filtering is to predict user u's rating of product i:

[0136] ;

[0137] in, is the predicted rating of user u for product i; is the average rating of user u; is the rating of user v on product i, is the neighbor set of user u; is the average rating of user v; is the similarity between user u and user v;

[0138] Calculation method:

[0139] ;

[0140] in, is the rating of product i by user u; is the rating of user v on product i; n is the number of products, and I is the set of products rated by both users u and v;

[0141] Weighted sum formula:

[0142] The weighted sum formula is used to accurately combine the similarities between products and users, specifically to adjust the recommendation effect of the collaborative filtering model:

[0143] ;

[0144] is the weight between user u and user v;

[0145] Step 3-2-2: The deep learning model is specifically a matrix decomposition model. Assume is the user embedding matrix, is the item embedding matrix, where m is the number of users, n is the number of items, and k is the embedding dimension;

[0146] Given user u and item i, the formula for predicting the rating is as follows:

[0147] ;

[0148] is the embedding vector of user u; is the embedding vector of product i;

[0149] The goal of the matrix factorization model is to learn the embedding vectors of users and items by minimizing the loss function:

[0150] ;

[0151] Among them, S is a rating dataset (i.e., user-item rating pairs); is the actual rating of product i by user u; Predict the rating of user u for product i; is a regularization parameter to prevent overfitting;

[0152] Step 3-2-3: Collaborative filtering and deep learning models are combined through weighted fusion method:

[0153] ;

[0154] is the recommendation score calculated by the collaborative filtering model; is the recommendation score obtained by the calculation result of the deep learning model; It is a parameter that adjusts the output weights of collaborative filtering and deep learning models;

[0155] In this embodiment, it is assumed that user 1 has rated product A (diapers) and product B (milk powder). The user's interest embedding vector is calculated based on historical behavior, and the following predicted rating is generated:

[0156] Predict user 1’s rating for product A (diapers): ;

[0157] Predict user 1’s rating for product B (milk powder): ;

[0158] Based on this result, product A (diapers) will be recommended to user 1 because its predicted score is higher.

[0159] Step 3-3: The product recommendation module solves the cold start problem of new users or new products based on the content recommendation algorithm. Specifically:

[0160] Generate an initial profile based on the user’s basic information;

[0161] Combined with the user's browsing history, the user profile is further improved. The user profile includes the user's preference for a certain feature of the product;

[0162] For new products, during the cold start phase, we use a content-based recommendation algorithm that combines product features and the user's initial profile to make recommendations:

[0163] ;

[0164] in, Predict the rating of user u for product i;

[0165] Product characteristics:

[0166] ;

[0167] in, represents the jth feature of product i, the threshold of j ranges from 1 to m, and m is the total number of features;

[0168] The preference for the feature is expressed as:

[0169] ;

[0170] in, represents the preference of the jth feature of user u, where the threshold of j ranges from 1 to m;

[0171] and Represent the modulus length of user portrait and product feature vector respectively; Represents the dot product of the user portrait and the product feature vector;

[0172] In this embodiment, it is assumed that user 1 is a new user with no historical rating data, but it is known that he is a parent of a newborn. The product recommendation module calculates the interest vector based on product features (such as the category and price of diapers) and the user's basic information (parent of a newborn).

[0173] The calculated score of product A (diapers) is 4.7, and the score of product B (milk powder) is 3.9, so product A (diapers) is recommended to user 1.

[0174] Step 3-4: The product recommendation module optimizes the recommendation strategy through the reinforcement learning algorithm and adopts the ε-greedy method to avoid the recommendation being concentrated on a certain type of product. The ε-greedy (epsilon-greedy strategy) method is an existing technology and will not be described in detail.

[0175] Step 4: The feedback optimization module establishes a feedback mechanism, collects feedback through user behavior, and dynamically adjusts the recommendation strategy based on the feedback. The feedback mechanism provides explainability to help users understand the reasons for the recommendation.

[0176] When executing step 4, the specific steps include:

[0177] Step 4-1: The feedback optimization module establishes a feedback collection mechanism, where users provide feedback on recommended products through clicks, ratings, and purchase behaviors; the feedback optimization module collects diverse user feedback through guided feedback question templates;

[0178] Step 4-2: The feedback optimization module uses a reinforcement learning algorithm to dynamically optimize the recommendation strategy based on user feedback;

[0179] The reinforcement learning algorithm formula is as follows:

[0180] ;

[0181] in, is in state Take action the value of From the state Take Action Immediate rewards after receiving a reward (such as user clicks, purchases, favorites, etc.); is a discount factor that indicates the impact of future rewards (usually set to a value close to 1); is the learning rate, which represents the weight between the new Q value and the old Q value; Represents the next state All possible actions under

[0182] In this example, assume that user 1 sees product A (diapers) and clicks it. The feedback optimization module gives an immediate reward (+1) based on the click behavior and updates the recommendation strategy through the reinforcement learning algorithm. If the user is observed to like diapers and click product A multiple times, the frequency of recommending diapers will increase.

[0183] Step 4-3: The feedback optimization module provides explainable feedback to help users understand why a certain product is recommended;

[0184] Interpretable feedback includes feature importance visualizations and decision tree-based explanations:

[0185] Feature importance visualization uses the interpretability technique LIME to display the weight of each feature in the recommendation decision;

[0186] The decision tree-based explanation specifically uses the decision tree model to analyze the relationship between user behavior and recommended products, and uses the tree model path to show why a certain product is recommended to the user;

[0187] Step 4-4: For cold start users, the feedback module collects their initial preferences through guided feedback and adjusts the recommended content.

[0188] The method of the AI ​​model maternal and infant assistant from question and answer to accurate recommendation described in the present invention solves the technical problems of ensuring the real-time and accuracy of user portraits, being able to flexibly respond to changes in user needs and interests, and improving the accuracy of recommendations and user satisfaction. The present invention introduces the BERT model and conversational AI, which can intelligently analyze and clarify the user's vague needs, further accurately extract user intentions and generate demand labels, ensure accurate understanding and processing of complex needs, and adopt a strategy of dynamically updating user portraits to continuously optimize and adjust user portraits based on real-time feedback (such as user clicks, purchases, and browsing behaviors). This ensures the real-time and accuracy of user portraits, and is able to flexibly respond to changes in user needs and interests. By combining collaborative filtering, deep learning (such as matrix decomposition), and reinforcement learning, a combination of multiple recommendation algorithms achieves more accurate and personalized recommendations. According to different scenarios (such as cold start, changes in user interests), the recommendation strategy is flexibly adjusted to improve the accuracy of recommendations and user satisfaction. This solution introduces LIME (local interpretable model-agnostic explanation) and decision tree-based interpretability technology, so that users can understand why a certain product is recommended, thereby enhancing users' trust in the recommendation system.

Claims

1. A method for an AI model maternal and infant assistant to go from question-answering to accurate recommendation, characterized by: The steps include: Step 1: The user enters the requirements in the dialogue interface. The requirements are natural language text. The semantic analysis module understands the requirements through semantic analysis and identifies the core requirements using NLP technology. If the requirements are vague, the semantic parsing module uses the BERT model to analyze and further interact with the user to clarify the requirements; The semantic analysis module extracts keywords from core requirements, generates requirement tags, and analyzes user sentiment preferences; Step 2: The data analysis module builds a preliminary portrait by analyzing the user's historical behavior and preferences, and dynamically updates the portrait based on real-time feedback. It continuously optimizes the label system in the user portrait to ensure the dynamic update and accuracy of the user portrait, monitor changes in user interests, and adjust the portrait and recommendation strategy; Step 3: The product recommendation module generates a preliminary candidate product pool based on demand tags through a weighted matching algorithm, and further optimizes product recommendations by combining collaborative filtering and deep learning models; Step 4: The feedback optimization module establishes a feedback mechanism to collect feedback from users and dynamically adjust the recommendation strategy based on the feedback. The feedback mechanism provides explainability to help users understand the reasons for the recommendation. When executing step 4, the specific steps include: Step 4-1: The feedback optimization module establishes a feedback collection mechanism, where users provide feedback on recommended products through clicks, ratings, and purchase behaviors; the feedback optimization module collects diverse user feedback through guided feedback question templates; Step 4-2: The feedback optimization module uses the reinforcement learning algorithm Q-learning to dynamically optimize the recommendation strategy based on user feedback; Step 4-3: The feedback optimization module provides explainable feedback to help users understand why a certain product is recommended; Interpretable feedback includes feature importance visualizations and decision tree-based explanations: Feature importance visualization uses the interpretability technique LIME to display the weight of each feature in the recommendation decision; The decision tree-based explanation specifically uses the decision tree model to analyze the relationship between user behavior and recommended products, and uses the tree model path to show why a certain product is recommended to the user; Step 4-4: For cold start users, the feedback module collects their initial preferences through guided feedback and adjusts the recommended content.

2. The method of the AI ​​model maternal and infant assistant from question and answer to accurate recommendation as claimed in claim 1, characterized in that: When executing step 1, the specific steps include: Step 1-1: Demand input, specifically, the user enters the demand in text form in the APP dialog box, and the semantic analysis module understands the user's core demand through semantic analysis and generates the demand text; Step 1-2: Demand fuzziness processing: When the demand is ambiguous, the semantic parsing module uses the BERT model to perform semantic analysis on the user input and extract the potential demand intention; If the intent is still unclear, the semantic parsing module interacts with the user through conversational AI, asking supplementary questions to clarify the needs; Step 1-3: Requirement label generation, specifically, the semantic parsing module automatically extracts keywords from the requirement text and generates requirement labels; Introduce sentiment analysis module to analyze user preferences for products based on their sentiment input Steps 1-4: Requirements preprocessing and mapping, specifically the semantic parsing module classifies and standardizes the requirement labels, and uses the label mapping algorithm to map the natural language requirements to the predefined categories in the knowledge base.

3. The method of the AI ​​model maternal and infant assistant from question and answer to accurate recommendation as claimed in claim 1, characterized in that: When executing step 2, the specific steps include: Step 2-1: User historical data analysis. Specifically, the data analysis module uses collaborative filtering algorithms to analyze user historical behaviors, generate preliminary user portraits, and reveal user interest preferences. The collaborative filtering algorithm predicts user interests based on user historical ratings and generates recommendation candidates based on similar user behaviors. The specific formula is as follows: ； in, is the predicted rating of user u for product i, which reveals the user’s interest preference for product i; is the average rating of user u; is the neighbor set of user u; is the rating of user v on product i; is the average rating of user v; is the similarity between user u and user v; ； in, is the rating of product i by user u; is the rating of user v on product i; n is the number of products, and I is the set of products rated by both users u and v; Step 2-2: Dynamic user portrait construction, specifically dynamically updating the user portrait through real-time feedback to ensure that the portrait always accurately reflects the user's current needs and interests. Real-time feedback includes clicks, browsing and purchasing behaviors; Step 2-3: Label update and optimization. Specifically, the data analysis module continuously optimizes the label system in the user portrait based on the new user behavior and demand data to ensure the dynamic update and accuracy of the user portrait. Step 2-4: Monitor and provide feedback on changes in user interests. Specifically, the data analysis module monitors changes in user interests in real time, analyzes user behavior, and adjusts user profiles in a timely manner.

4. The method of the AI ​​model maternal and infant assistant from question and answer to accurate recommendation as claimed in claim 3, characterized in that: When executing step 3, the specific steps include: Step 3-1: The product recommendation module screens products based on user demand tags through a weighted matching algorithm to generate a preliminary candidate product pool. The weighted matching algorithm calculates the matching score between products and user needs based on user preferences for product features. The specific formula is as follows: ； in: The matching score of user u for product i; is the weight of feature f; is the value of product i on feature f; User u’s preference on feature f; Step 3-2: Product recommendation module Based on the method of combining collaborative filtering and deep learning model, the product recommendation module further optimizes the product candidate pool as follows: Step 3-2-1: The collaborative filtering model adjusts the product candidate pool through user similarity and item similarity, including the user-product rating prediction formula and weighted summation formula: User-item rating prediction formula: Assuming there are user u and product i, the goal of collaborative filtering is to predict user u's rating of product i: ； in, is the predicted rating of user u for product i; is the average rating of user u; is the rating of user v on product i, is the neighbor set of user u; is the average rating of user v; is the similarity between user u and user v; Calculation method: ； in, is the rating of product i by user u; is the rating of user v on product i; n is the number of products, and I is the set of products rated by both users u and v; Weighted sum formula: The weighted sum formula is used to accurately combine the similarities between products and users, specifically to adjust the recommendation effect of the collaborative filtering model: ; is the weight between user u and user v; Step 3-2-2: The deep learning model is specifically a matrix decomposition model. Assume is the user embedding matrix, is the item embedding matrix, where m is the number of users, n is the number of items, and k is the embedding dimension; Given user u and item i, the formula for predicting the rating is as follows: ； is the embedding vector of user u; is the embedding vector of product i; The goal of the matrix factorization model is to learn the embedding vectors of users and items by minimizing the loss function: ； Among them, S is the scoring dataset; is the actual rating of product i by user u; Predict the rating of user u for product i; is a regularization parameter to prevent overfitting; Step 3-2-3: Collaborative filtering and deep learning models are combined through weighted fusion method: ; is the recommendation score calculated by the collaborative filtering model; is the recommendation score obtained by the deep learning model calculation; It is a parameter that adjusts the output weights of collaborative filtering and deep learning models; Step 3-3: The product recommendation module solves the cold start problem of new users or new products based on the content recommendation algorithm. Specifically: Generate an initial profile based on the user’s basic information; Combined with the user's browsing history, the user profile is further improved. The user profile includes the user's preference for a certain feature of the product; For new products, during the cold start phase, we use a content-based recommendation algorithm that combines product features and the user's initial profile to make recommendations: ； in, Predict the rating of user u for product i; Product characteristics: ; in, represents the jth feature of product i, the threshold of j ranges from 1 to m, and m is the total number of features; The preference for the feature is expressed as: ; in, represents the preference of the jth feature of user u, where the threshold of j ranges from 1 to m; and Represent the modulus length of user portrait and product feature vector respectively; Represents the dot product of the user portrait and the product feature vector; Step 3-4: The product recommendation module optimizes the recommendation strategy through the reinforcement learning algorithm and adopts the ε-greedy method to avoid the recommendation being concentrated on a certain category of products.

Images