A product recommendation method, apparatus, and storage medium

Abstract

This invention provides a product recommendation method, apparatus, and storage medium, belonging to the field of product recommendation technology. The method includes: importing multiple original customer consumption data and multiple customer consumption data to be recommended, and constructing an initial prediction model; training the initial prediction model with all original customer consumption data and all customer consumption data to be recommended to obtain a target prediction model; and using the target prediction model to predict the normalized customer consumption data to be recommended to obtain an original importance score. This invention improves customer identification efficiency and recommendation accuracy, can quickly identify potential customers from a massive existing customer base, eliminates reliance on manual experience, meets real-time recommendation needs, enables multi-product collaborative analysis, and can adapt to complex business scenarios.

Description

Technical Field

[0001] This invention relates to the field of product recommendation technology, specifically to a product recommendation method, apparatus, and storage medium. Background Technology

[0002] With the rapid development and application of artificial intelligence and recommendation system technologies in government and enterprise marketing, product recommendation methods have become a core technical means to improve marketing efficiency and achieve precise reach for government and enterprise clients. Currently, product recommendations for government and enterprise clients in the industry still primarily rely on traditional methods. The core of these methods is to summarize and organize the manual marketing experience and established business rules of sales personnel, solidify them into recommendation logic through programmatic and coded means, and then input basic customer data into this logic to complete product recommendations. This is an experience-driven and rule-based recommendation model. While it played a certain role in the early stages of marketing standardization, with the surge in the number of government and enterprise clients, the increasing complexity of product systems, and the ever-increasing demands for response speed in marketing scenarios, the technical shortcomings and application pain points of this type of recommendation method have become increasingly prominent. It can no longer meet the refined, efficient, and intelligent marketing needs of government and enterprises, specifically exhibiting the following three core problems, while also giving rise to the additional problem of a lack of multi-product association recognition capabilities: I. Inefficient customer identification makes it difficult to meet the needs of mining a massive number of existing customers. The government and enterprise business system contains a massive amount of existing customer data. Customer needs vary significantly across different industries, sizes, and spending power. Traditional recommendation methods rely entirely on the manual experience of sales personnel to identify potential customers. However, the experience and time available to these personnel are limited, making it impossible to quickly filter out potential customers with a willingness to purchase products from this vast and multi-dimensional customer data. Furthermore, manual screening lacks standardized and quantifiable criteria, making it susceptible to subjective influences. This results in a narrow coverage and low accuracy in identifying potential customers, leading to the loss of valuable customer resources and ultimately inefficient marketing efforts, preventing the effective allocation of marketing resources.

[0003] Second, rule-based recommendation models lack flexibility and cannot adapt to complex business scenarios. The core logic of traditional recommendation methods relies on manually defined, fixed business rules. These rules are often simple, linear judgments, failing to capture the complex, non-linear relationships between government and enterprise customer characteristics and product needs. For example, potential product demands arising from the combined effects of multiple dimensions such as customer industry attributes, registered capital, communication consumption habits, and existing product portfolios are difficult to accurately characterize through manual rules. Furthermore, the market environment, government and enterprise customer needs, and enterprise product systems are all dynamically changing. However, rule updates and iterations depend entirely on manual refactoring and rewriting, resulting in cumbersome processes, long cycles, and rule updates lagging behind the pace of business changes. This leads to a disconnect between the recommendation logic and actual business needs, causing a continuous decline in the relevance of the recommendation results.

[0004] Third, the real-time performance of the model and recommendation process is insufficient, failing to meet the rapid response requirements of marketing scenarios. Marketing scenarios are highly time-sensitive, and the needs of government and enterprise clients are often immediate. Recommendation systems must be able to quickly analyze customer data and output recommendations to support frontline marketing personnel in their immediate operations. However, traditional recommendation methods rely on generalized models that haven't been optimized for lightweight and efficient performance in government and enterprise product recommendation scenarios. Model training consumes significant time processing data, and prediction speeds for individual customers are slow, resulting in a lengthy overall process from customer data input to recommendation output. Furthermore, the fixed, programmed rules lack the rapid computation and response capabilities to handle new customer data, failing to meet the core needs of marketing scenarios for real-time recommendations and rapid decision-making, thus missing optimal opportunities for customer conversion.

[0005] In summary, existing product recommendation methods based on human experience and fixed rules have significant technical shortcomings in terms of customer identification efficiency, recommendation flexibility, real-time response capability, and multi-product correlation analysis. These shortcomings have become key issues restricting the intelligent and precise development of government and enterprise product marketing. Summary of the Invention

[0006] The technical problem to be solved by the present invention is to provide a product recommendation method, apparatus and storage medium to address the shortcomings of the prior art.

[0007] The technical solution of the present invention to solve the above-mentioned technical problems is as follows: A product recommendation method, comprising the following steps: Import multiple sets of original customer consumption data and multiple sets of customer consumption data to be recommended, and build an initial prediction model; The initial prediction model is analyzed using all the original customer consumption data and all the customer consumption data to be recommended, to obtain the target prediction model corresponding to each product and the product feature importance vector corresponding to each product. The feature sensitivity difference coefficients of each product feature importance vector are analyzed to obtain multiple normalized importance scores corresponding to each product and the target product feature set corresponding to each product. Recommendation results are obtained by performing recommendation analysis on all the normalized importance scores and all the target product feature sets using all the target prediction models.

[0008] Another technical solution of the present invention to solve the above-mentioned technical problems is as follows: A product recommendation device, comprising: The import module is used to import multiple original customer consumption data sets and multiple customer consumption data sets to be recommended. The model analysis module is used to construct an initial prediction model. It performs model analysis on the initial prediction model using all the original customer consumption data and all the customer consumption data to be recommended, and obtains the target prediction model corresponding to each product and the product feature importance vector corresponding to each product. The difference coefficient analysis module is used to analyze the feature sensitivity difference coefficient of each product feature importance vector to obtain multiple normalized importance scores corresponding to each product and the target product feature set corresponding to each product. The recommendation result acquisition module is used to perform recommendation analysis on all the normalized importance scores and all the target product feature sets through all the target prediction models to obtain product recommendation results.

[0009] Based on the product recommendation method described above, the present invention also provides a product recommendation system.

[0010] Another technical solution of the present invention to solve the above-mentioned technical problems is as follows: a product recommendation system, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, the product recommendation method described above is implemented.

[0011] Based on the above-described product recommendation method, the present invention also provides a computer-readable storage medium.

[0012] Another technical solution of the present invention to solve the above-mentioned technical problems is as follows: a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the product recommendation method as described above.

[0013] The beneficial effects of this invention are as follows: by analyzing the initial prediction model using original customer consumption data and customer consumption data to be recommended, a target prediction model and a product feature importance vector are obtained. By analyzing the feature sensitivity difference coefficient of the product feature importance vector, a normalized importance score and a target product feature set are obtained. By analyzing the recommendation results of the target prediction model using the normalized importance score and the target product feature set, product recommendation results are obtained. This improves the efficiency of customer identification and the accuracy of recommendation, enables the rapid identification of potential customers from a large number of existing customers, eliminates the reliance on human experience, meets real-time recommendation needs, enables multi-product collaborative analysis, and can adapt to complex business scenarios. Attached Figure Description

[0014] Figure 1 A flowchart illustrating the product recommendation method provided in an embodiment of the present invention; Figure 2A schematic diagram of the original customer consumption data and multiple customer consumption data to be recommended in the product recommendation method provided in the embodiments of the present invention; Figure 3 This is a schematic diagram illustrating the feature sensitivity difference coefficient division of the product recommendation method provided in this embodiment of the invention. Figure 4 This is a text conversion diagram of the product recommendation method provided in an embodiment of the present invention; Figure 5 A schematic diagram of the preset hierarchical calling rules for the product recommendation method provided in this embodiment of the invention; Figure 6 A block diagram of a product recommendation device provided in an embodiment of the present invention. Detailed Implementation

[0015] The principles and features of the present invention are described below with reference to the accompanying drawings. The examples given are only for explaining the present invention and are not intended to limit the scope of the present invention.

[0016] Figure 1 This is a flowchart illustrating a product recommendation method provided in an embodiment of the present invention.

[0017] like Figure 1 and 2 As shown, a product recommendation method includes the following steps: S1: Import multiple original customer consumption data and multiple customer consumption data to be recommended, and build an initial prediction model; S2: Train the initial prediction model using all the original customer consumption data and all the customer consumption data to be recommended to obtain a target prediction model corresponding to each product, and use the target prediction model to predict the normalized customer consumption data to be recommended to obtain a product feature importance vector corresponding to each product. The product feature importance vector includes the original importance score corresponding to each product feature. S3: Normalize each of the original importance scores to obtain the normalized importance scores corresponding to each of the product features; S4: Analyze the feature sensitivity difference coefficient for each of the normalized importance scores, and combine the analysis results to obtain the target product feature set corresponding to each of the products; S5: The feature sets of each target product are predicted by each target prediction model to obtain multiple predicted probabilities and multiple recommended star ratings corresponding to each customer. S6: Import multiple product business data for each of the aforementioned customers, and perform recommendation analysis on all the aforementioned customers based on all the predicted probabilities, all the aforementioned recommendation star ratings, all the aforementioned normalized importance scores, and all the aforementioned product business data to obtain product recommendation results.

[0018] It should be understood that the initial prediction model can be the LightGBM model, and the original customer consumption data and the customer consumption data to be recommended can be data such as customer industry, registered capital, communication consumption, ordered products, and city code.

[0019] Specifically, using the feature_importance attribute of LightGBM (i.e., the target prediction model), the feature importance vector (i.e., the product feature importance vector) of each product model is extracted, as shown in the following formula: F_p = [f_p1, f_p2, ..., f_pn], Where F_p represents the feature importance vector of product p (i.e., the product feature importance vector), and f_pi represents the importance score of the i-th feature to product p (i.e., the original importance score).

[0020] It should be understood that for each type of product (such as cloud networking, internet leased lines, etc.), a differentiated feature set (i.e., original customer consumption data and customer consumption data to be recommended) is designed according to the product business characteristics, and a LightGBM prediction model (i.e., the initial prediction model) is independently constructed.

[0021] In the above embodiments, a target prediction model is obtained by training an initial prediction model using original customer consumption data and consumption data of customers to be recommended. The original importance score is obtained by predicting the normalized consumption data of customers to be recommended using the target prediction model. The normalization of the original importance score is then used to obtain a normalized importance score. The feature sensitivity difference coefficient of the normalized importance score is analyzed, and the analysis results are combined to obtain a target product feature set. The prediction probability and recommendation star rating are obtained by predicting the target product feature set using the target prediction model. Based on the prediction probability, recommendation star rating, normalized importance score, and product business data, the product recommendation result is obtained through customer recommendation analysis. This improves the efficiency of customer identification and the accuracy of recommendation. It can quickly identify potential customers from a large number of existing customers, eliminate the reliance on human experience, meet real-time recommendation needs, realize multi-product collaborative analysis, and adapt to complex business scenarios.

[0022] Optionally, as an embodiment of the present invention, the process of training the initial prediction model using all the original customer consumption data and all the customer consumption data to be recommended to obtain a target prediction model corresponding to each product, and using the target prediction model to predict the normalized customer consumption data to be recommended to obtain a product feature importance vector corresponding to each product includes: According to the preset product business characteristic classification rules, the original customer consumption data and the customer consumption data to be recommended are classified and processed to obtain the original customer consumption data and the customer consumption data to be recommended after multiple classifications corresponding to each product. Data cleaning processing is performed on the original customer consumption data after each category and the customer consumption data to be recommended after multiple categories corresponding to each product, to obtain multiple cleaned original customer consumption data and multiple cleaned customer consumption data to be recommended corresponding to each product. The initial prediction model is trained by using multiple original customer consumption data after cleaning corresponding to each of the products to obtain the target prediction model corresponding to each of the products. Each of the target prediction models is used to predict the consumption data of multiple cleaned customers to be recommended for each product, thereby obtaining the product feature importance vector corresponding to each product.

[0023] In the above embodiments, the initial prediction model is trained using the original customer consumption data and the consumption data of the customers to be recommended to obtain the target prediction model. The target prediction model is then used to predict the normalized consumption data of the customers to be recommended to obtain the product feature importance vector. This improves the efficiency of customer identification and the accuracy of recommendation, and enables the rapid identification of potential customers from a large number of existing customers.

[0024] Optionally, as an embodiment of the present invention, such as Figure 3 As shown, the process of analyzing the feature sensitivity difference coefficients of each of the normalized importance scores and aggregating the analysis results to obtain the target product feature set corresponding to each of the products includes: The first formula is used to calculate the normalized importance score for each of the aforementioned products and the normalized importance score for any remaining product, respectively, to obtain the feature sensitivity difference coefficient corresponding to each of the product features. , in, For the first The feature sensitivity difference coefficient corresponding to each product feature For the first The first of the products Normalized importance scores corresponding to each product feature For the first The first of the products Normalized importance scores corresponding to each product feature It is a smoothing factor; If the feature sensitivity difference coefficient is greater than the preset first feature sensitivity threshold, and the normalized importance score corresponding to the feature sensitivity difference coefficient is not 0, then the product feature corresponding to the feature sensitivity difference coefficient is taken as the first feature to be processed, thereby obtaining multiple first features to be processed. If the feature sensitivity difference coefficient is less than or equal to the preset first feature sensitivity threshold and greater than the preset second feature sensitivity threshold, then the product feature corresponding to the feature sensitivity difference coefficient is taken as the second feature to be processed, thereby obtaining multiple second features to be processed. If the feature sensitivity difference coefficient is less than or equal to the second feature sensitivity threshold, then the product feature corresponding to the feature sensitivity difference coefficient is taken as the third feature to be processed, thereby obtaining multiple third features to be processed. All the third features to be processed are sorted in ascending order of the normalized importance scores corresponding to the third features to be processed, and the first K sorted third features to be processed are taken as the fourth features to be processed, thus obtaining multiple fourth features to be processed. All the first features to be processed, all the second features to be processed, and all the fourth features to be processed are classified according to product categories to obtain multiple fifth features to be processed corresponding to each product. The multiple fifth features to be processed corresponding to each product are then combined to obtain the target product feature set corresponding to each product.

[0025] It should be understood that the normalized importance score for any remaining product refers to the normalized importance score for any other product besides the current product.

[0026] Specifically, the "Feature Sensitivity Divergence" (FSD) coefficient is proposed to quantify the degree of difference in features among different products: FSD_i = 1 - (2×min(f_Ai, f_Bi)) / (f_Ai + f_Bi +ε), in: - f_Ai and f_Bi represent the normalized importance of feature i in product A and product B, respectively; -ε is a smoothing factor (value 0.001) to prevent division by zero errors; - FSD_i ∈ [0, 1]. The larger the value, the greater the difference in sensitivity of the feature among different products.

[0027] It should be understood that, taking eparchy_code (city code) as an example: - Importance of cloud networking: 0 (0 after normalization); - Importance of Internet dedicated line: 2066 (0.54 after normalization); - FSD = 1 - (2×min(0, 0.54)) / (0 + 0.54 + 0.001) = 1.0 (maximum difference).

[0028] It should be understood that the input is: target product type P, original feature set X, importance matrix F of each product feature, and the output is: adaptive feature set X'.

[0029] Specifically, calculate the FSD value (i.e., feature sensitivity difference coefficient) of all features, classify the features into three levels according to the FSD value. For the target product P, as follows: - Eliminate product-sensitive features (i.e., the first to-be-processed features) with FSD > 0.7 and importance 0 in P; - Retain general features with FSD ≤ 0.3 (i.e., the second to-be-processed features); - For weakly sensitive features with 0.3 < FSD ≤ 0.7 (i.e., the fourth to-be-processed features), retain Top-K according to the importance in P, and return the adaptive feature set X' (i.e., the target product feature set).

[0030] In the above embodiments, the feature sensitivity difference coefficient is analyzed for each normalized importance score respectively, and the target product feature set is obtained by integrating the analysis results, getting rid of the dependence on manual experience, meeting the real-time recommendation requirements, enabling multi-product collaborative analysis, and being able to adapt to complex business scenarios.

[0031] Optionally, as an embodiment of the present invention, the process of performing recommendation analysis on all the customers according to all the prediction probabilities, all the recommendation stars, all the normalized importance scores, and all the product business data to obtain the recommendation results of the products includes: Respectively perform feature contribution degree analysis on the multiple product business data of each customer and the multiple normalized importance scores corresponding to each product, and obtain a customer feature contribution degree ranking table corresponding to each customer; Respectively perform analysis of recommendation reasons on each customer feature contribution degree ranking table and the multiple recommendation stars corresponding to each customer, and obtain the recommendation reasons corresponding to each customer; Import the multiple product service priorities corresponding to each customer, and calculate the multiple predicted probabilities, multiple recommendation star ratings, and multiple product service priorities corresponding to each customer using the second formula to obtain multiple comprehensive scores corresponding to each customer. The second formula is: , in, For the first The overall score corresponding to each product. , as well as All are weighting coefficients. For the first The predicted probability for each product. For the first Recommended star rating for each product. For the first Product business priority corresponding to each product; The maximum values ​​of multiple comprehensive scores corresponding to each customer are filtered out, and the maximum comprehensive score corresponding to each customer is obtained after filtering. Recommendation information for each customer is constructed by using the products corresponding to each maximum comprehensive score, the predicted probabilities corresponding to each maximum comprehensive score, the recommendation star rating corresponding to each maximum comprehensive score, the customer feature contribution ranking table corresponding to each customer, and the recommendation reasons corresponding to each customer. All of the recommendation information is then used as the product recommendation result.

[0032] It should be understood that an adaptive feature set (i.e., the target product feature set) is used as input to call the corresponding product's LightGBM model (i.e., the target prediction model) for prediction: - Output 1: Predicted probability (a continuous value between 0 and 1); - Output 2: Recommended star rating (divided into 1-5 stars based on probability thresholds).

[0033] It should be understood that for each customer, the following outputs are provided: recommended product name (i.e., product), recommended star rating (1-5 stars), predicted probability value (i.e., predicted probability), recommendation reason (generated by a large model or template), and Top-3 key recommendation factors (i.e., customer feature contribution ranking table). Among them, the recommended product name is obtained from the product definition table through the corresponding product name, product description and other information.

[0034] Specifically, multiple product recommendations can be output simultaneously for the same customer, sorted by comprehensive score, as follows: Score =α ×Prob +β ×Star +γ ×Priority, in: Prob: Predicted probability; Star: Recommended rating; Priority: Product and business priority; α, β, γ: Weighting coefficients (configurable).

[0035] Specifically, the output for each customer is: Recommendation levels: Divided into 5 stars based on threshold strategy, from high to low; Reason for recommendation: Generate interpretable recommendation criteria based on feature importance; Marketing advice: Match industry solutions with marketing messaging.

[0036] In the above embodiments, product recommendation results are obtained by performing recommendation analysis on all customers based on all predicted probabilities, all recommendation star ratings, all normalized importance scores, and all product business data. This improves the efficiency of customer identification and the accuracy of recommendation, enables the rapid identification of potential customers from a massive existing customer base, eliminates reliance on human experience, meets real-time recommendation needs, enables collaborative analysis of multiple products, and can adapt to complex business scenarios.

[0037] Optionally, as an embodiment of the present invention, the process of performing feature contribution analysis on multiple product business data of each customer and multiple normalized importance scores corresponding to each product to obtain a customer feature contribution ranking table corresponding to each customer includes: Calculate the mean of each product feature corresponding to all the customers to obtain the mean of the product feature corresponding to each product feature; Calculate the standard deviation of each product feature corresponding to all the customers, and obtain the standard deviation of the product feature corresponding to each product feature; The contribution of multiple features corresponding to each customer is calculated using the third formula, which involves analyzing the multiple product business data of each customer, the mean of the product features corresponding to each product feature, the standard deviation of the product features corresponding to each product feature, and the multiple normalized importance scores corresponding to each product feature. The third formula is as follows: , in, For the first The feature contribution of each product feature For the first The first customer Product business data corresponding to each product feature For the first The first of the products Normalized importance scores corresponding to each product feature For the first The average value of product features corresponding to each product feature. For the first The standard deviation of product characteristics corresponding to each product characteristic; The multiple feature contributions corresponding to each customer are sorted in descending order of feature contribution, and the top N sorted feature contributions are plotted as customer feature contribution ranking tables, thereby obtaining customer feature contribution ranking tables corresponding to each customer.

[0038] Specifically, for each customer to be recommended, their personalized feature contribution ranking is calculated as follows: c_i = f_pi × (v_ui - μ_i) / σ_i, in: - f_pi represents the global importance of feature i in product p (i.e., the normalized importance score). - v_ui represents the value of customer u on feature i (i.e., product business data). - μ_i and σ_i are the mean (i.e., the product characteristic mean) and standard deviation (i.e., the product characteristic standard deviation) of feature i.

[0039] It should be understood that the output format is: C_u = [(feature1, contribution1), (feature2, contribution2), ..., (featurek, contributionk)], sorted in descending order of contribution, and the top-3 are taken as the core basis for the recommendation reason (i.e., customer feature contribution ranking table).

[0040] In the above embodiments, feature contribution analysis is performed on multiple product business data and normalized importance scores of each customer to obtain a customer feature contribution ranking table, which improves the customer identification efficiency and recommendation accuracy, and can quickly identify potential customers from a massive number of existing customers.

[0041] Optionally, as an embodiment of the present invention, such as Figure 4 and 5 As shown, the process of analyzing the recommendation reasons for each customer by examining the contribution ranking table of each customer's characteristics and the multiple recommendation star ratings corresponding to each customer, and obtaining the recommendation reasons for each customer, includes: Each of the customer feature contribution ranking tables is converted into text to obtain customer semantic description text corresponding to each customer. According to the preset hierarchical calling rules, the model is called for multiple recommendation star ratings corresponding to each customer to obtain the reason generation model corresponding to each customer. Each of the aforementioned reason generation models generates a reason for each customer's semantic description text, thereby obtaining a recommendation reason for each customer.

[0042] It should be understood that, as Figure 4 As shown, the numerical feature contributions output by the small model (i.e., the customer feature contribution ranking table) are transformed into natural language descriptions (i.e. customer semantic description text) that the large model can understand. Based on the feature contribution ranking, prompt words for the large model are dynamically constructed.

[0043] It should be understood that the reasoning generation model can be a large model or a preset template.

[0044] Specifically, such as Figure 5 As shown, the constructed prompt words (i.e., customer semantic description text) are sent to the Large Language Model (LLM) (i.e., reason generation model) to generate personalized recommendation reasons. The decision to call the large model is based on the recommendation star rating, balancing effectiveness and cost.

[0045] In the above embodiments, the reasons for recommendation are analyzed by sorting the contribution of each customer feature and the recommendation star rating, which improves the efficiency of customer identification and the accuracy of recommendation, and can quickly identify potential customers from a large number of existing customers.

[0046] Optionally, as an embodiment of the present invention, it further includes: The fourth equation is used to calculate the importance vectors of each product feature and the remaining product feature importance vectors, respectively, to obtain multiple importance similarities corresponding to each product. The fourth equation is: , in, For the first The product and the first The similarity in importance of individual products For the first Product feature importance vectors corresponding to each product For the first Product feature importance vectors corresponding to each product; If the importance similarity is greater than a preset similarity threshold, then cross-recommendation processing is performed on the customers corresponding to the importance similarity.

[0047] Preferably, the preset similarity threshold can be 0.6.

[0048] It should be understood that any remaining product feature importance vector refers to any other product feature importance vector besides the current product feature importance vector.

[0049] Specifically, based on the feature importance similarity (i.e., importance similarity) between products, related products are identified and cross-recommended, as follows: Similarity calculation: Sim(A,B) = F_A·F_B / (||F_A||×||F_B||), when Sim>0.6, cross-recommendation is triggered.

[0050] It should be understood that cross-recommendation can be interpreted as: when a customer is suitable to be recommended product A, they are also suitable to be recommended product B; or it can be interpreted as recommending products A and B simultaneously.

[0051] In the above embodiments, the importance similarity is calculated by calculating the importance vector of each product feature and the importance vector of any remaining product feature, which improves the customer identification efficiency and recommendation accuracy, and can quickly identify potential customers from a large number of existing customers.

[0052] Optionally, as another embodiment of the present invention, the present invention adopts a six-stage process: "first, independent modeling of multiple products; second, feature sensitivity analysis; third, adaptive feature selection; fourth, small model prediction; fifth, feature contribution bridging; and sixth, large model reason generation," to achieve accurate product recommendation based on feature sensitivity adaptation and collaboration between large and small models, as detailed below: STEP 1: Design differentiated feature sets for different products and model them independently. Extract the feature importance vectors for each product and discover that the same feature has significantly different importance in different products.

[0053] STEP 2: Propose the FSD formula to quantify the sensitivity differences of features among different products. The larger the FSD value, the stronger the product specificity of the feature.

[0054] STEP 3: Based on the FSD value, the features are divided into three categories: product sensitive, weakly sensitive, and general. Differentiated strategies such as elimination, weight adjustment, and sharing are adopted for different categories.

[0055] STEP 4: Use the small model to predict probabilities and star ratings, calculate the personalized feature contribution ranking for each customer, and output the Top-3 key recommendation factors.

[0056] STEP 5: Transform feature contribution into semantic descriptions and dynamically construct large model prompts containing customer profiles and key factors.

[0057] STEP 6: Use the large model to generate recommendation reasons, and adopt a star-rating hierarchical calling strategy. High-value customers must be called, and low-value customers are downgraded to templates.

[0058] STEP 7: Output recommendation results, supporting multi-product collaborative recommendation and cross-recommendation based on feature similarity.

[0059] Alternatively, as another embodiment of the present invention, the key points of the present invention are as follows: 1. FSD formula: Quantifying the product specificity of features; 2. Three-level feature classification: a refined feature management strategy; 3. Feature contribution bridging: Information transfer mechanism between large and small models; 4. Tiered application strategy: a dynamic balance between effectiveness and cost; 5. Collaborative architecture of large and small models: complete technology chain.

[0060] Alternatively, as another embodiment of the present invention, the protection points of the present invention are as follows: (1) Feature Sensitivity Difference Coefficient (FSD) and Adaptive Feature Selection: FSD is a quantitative indicator proposed for the first time in this invention, and no similar method has been found in the prior art; the three-level feature classification strategy based on FSD is original.

[0061] (2) Feature contribution bridging mechanism: Design a feature contribution calculation formula to transform the prediction results of the small model into a personalized feature contribution ranking, and then generate prompt words that the large model can understand through semantic transformation. This bridging mechanism realizes the effective collaboration between the large and small models. In the existing technology, the large and small models are usually used independently and lack a structured collaboration method.

[0062] (3) The large model hierarchical calling strategy based on the recommendation star rating: Based on the recommendation star rating output by the small model, it is decided whether to call the large model. High-value customers must call the large model, while low-value customers use templates to downgrade. The confidence of the recommendation result is associated with the large model calling decision, so as to achieve a dynamic balance between cost and effect, which is technically innovative.

[0063] Optionally, as another embodiment of the present invention, this invention discloses a product recommendation method based on feature sensitivity adaptation and large-scale model collaboration, belonging to the field of artificial intelligence and recommendation systems. Addressing the problems of low accuracy in multi-product recommendations and insufficient interpretability of recommendation reasons resulting from the use of a unified feature set in existing recommendation methods, this invention proposes the following technical solution: First, a LightGBM model is independently trained for each product category, and feature importance vectors are extracted; second, a "Feature Sensitivity Difference Coefficient" (FSD) is proposed to quantify the importance differences of features among different products, classifying features into three categories: product-sensitive features, weakly sensitive features, and general features, achieving adaptive feature selection; then, the personalized feature contribution of each customer is calculated, and the structured output of the small model is transformed into semantic prompts understandable by the large model through a "feature contribution bridging" mechanism; finally, a large language model is called to generate interpretable recommendation reasons, and a hierarchical calling strategy based on recommendation star rating is adopted to control costs. This invention achieves collaboration between efficient prediction by the small model and intelligent interpretation by the large model, significantly improving the interpretability of recommendation reasons.

[0064] Figure 6 This is a module block diagram of a product recommendation device provided in an embodiment of the present invention.

[0065] Alternatively, as another embodiment of the present invention, such as Figure 6 As shown, a product recommendation device includes: The import module is used to import multiple original customer consumption data sets and multiple customer consumption data sets to be recommended. The training module is used to build an initial prediction model. The initial prediction model is trained using all the original customer consumption data and all the customer consumption data to be recommended, so as to obtain the target prediction model corresponding to each product. The first prediction module is used to predict the normalized consumption data of the customers to be recommended through the target prediction model, and obtain the product feature importance vector corresponding to each product. The product feature importance vector includes the original importance score corresponding to each product feature. The normalization module is used to normalize each of the original importance scores to obtain the normalized importance scores corresponding to each of the product features. The difference coefficient analysis module is used to analyze the feature sensitivity difference coefficient of each of the normalized importance scores, and to combine the analysis results to obtain the target product feature set corresponding to each of the products. The second prediction module is used to predict each of the target product feature sets using each of the target prediction models, and to obtain multiple predicted probabilities and multiple recommendation star ratings corresponding to each customer. The import module is also used to import multiple product business data of each of the customers; The recommendation result acquisition module is used to perform recommendation analysis on all customers based on all the predicted probabilities, all the recommendation star ratings, all the normalized importance scores, and all the product business data to obtain product recommendation results.

[0066] Optionally, another embodiment of the present invention provides a product recommendation system, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the product recommendation method described above. This system can be a computer or similar system.

[0067] Optionally, another embodiment of the present invention provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the product recommendation method as described above.

[0068] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0069] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working process of the above-described apparatus and unit can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0070] In the several embodiments provided in this application, it should be understood that the disclosed apparatus and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative. For instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed.

[0071] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of the embodiments of the present invention, depending on actual needs.

[0072] Furthermore, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0073] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0074] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A product recommendation method characterized by, Includes the following steps: Import multiple sets of original customer consumption data and multiple sets of customer consumption data to be recommended, and build an initial prediction model; The initial prediction model is trained using all the original customer consumption data and all the customer consumption data to be recommended to obtain a target prediction model corresponding to each product. The target prediction model is then used to predict the normalized customer consumption data to be recommended to obtain a product feature importance vector corresponding to each product. The product feature importance vector includes the original importance score corresponding to each product feature. Each of the original importance scores is normalized to obtain a normalized importance score corresponding to each of the product features. The feature sensitivity difference coefficients of each of the normalized importance scores are analyzed, and the analysis results are combined to obtain the target product feature set corresponding to each of the products. Each target prediction model is used to predict the feature set of each target product to obtain multiple predicted probabilities and multiple recommended star ratings corresponding to each customer. Import multiple product business data for each of the aforementioned customers, and perform recommendation analysis on all the aforementioned customers based on all the aforementioned predicted probabilities, all the aforementioned recommendation star ratings, all the aforementioned normalized importance scores, and all the aforementioned product business data to obtain product recommendation results.

2. The product recommendation method according to claim 1, characterized in that, The process of training the initial prediction model using all the original customer consumption data and all the customer consumption data to be recommended, to obtain a target prediction model corresponding to each product, and then using the target prediction model to predict the normalized customer consumption data to be recommended, to obtain the product feature importance vector corresponding to each product, includes: According to the preset product business characteristic classification rules, the original customer consumption data and the customer consumption data to be recommended are classified and processed to obtain the original customer consumption data and the customer consumption data to be recommended after multiple classifications corresponding to each product. Data cleaning processing is performed on the original customer consumption data after each category and the customer consumption data to be recommended after multiple categories corresponding to each product, to obtain multiple cleaned original customer consumption data and multiple cleaned customer consumption data to be recommended corresponding to each product. The initial prediction model is trained by using multiple original customer consumption data after cleaning corresponding to each of the products to obtain the target prediction model corresponding to each of the products. Each of the target prediction models is used to predict the consumption data of multiple cleaned customers to be recommended for each product, thereby obtaining the product feature importance vector corresponding to each product.

3. The product recommendation method according to claim 2, characterized by, The process of analyzing the feature sensitivity difference coefficients of each of the normalized importance scores and aggregating the analysis results to obtain the target product feature set corresponding to each of the products includes: The first formula is used to calculate the normalized importance score for each of the aforementioned products and the normalized importance score for any remaining product, respectively, to obtain the feature sensitivity difference coefficient corresponding to each of the product features. ， wherein, is the feature sensitivity difference coefficient corresponding to the jth product feature, is the normalized importance score corresponding to the jth product feature in the ith product, is the normalized importance score corresponding to the jth product feature in the ith product, is the normalized importance score corresponding to the jth product feature in the ith product, is a smoothing factor;​​​​ If the feature sensitivity difference coefficient is greater than the preset first feature sensitivity threshold, and the normalized importance score corresponding to the feature sensitivity difference coefficient is not 0, then the product feature corresponding to the feature sensitivity difference coefficient is taken as the first feature to be processed, thereby obtaining multiple first features to be processed. If the feature sensitivity difference coefficient is less than or equal to the preset first feature sensitivity threshold and greater than the preset second feature sensitivity threshold, then the product feature corresponding to the feature sensitivity difference coefficient is taken as the second feature to be processed, thereby obtaining multiple second features to be processed. If the feature sensitivity difference coefficient is less than or equal to the second feature sensitivity threshold, then the product feature corresponding to the feature sensitivity difference coefficient is taken as the third feature to be processed, thereby obtaining multiple third features to be processed. All the third features to be processed are sorted in ascending order of the normalized importance scores corresponding to the third features to be processed, and the first K sorted third features to be processed are used as fourth features to be processed, thereby obtaining multiple fourth features to be processed. All the first features to be processed, all the second features to be processed, and all the fourth features to be processed are classified according to product categories to obtain multiple fifth features to be processed corresponding to each product. The multiple fifth features to be processed corresponding to each product are then combined to obtain the target product feature set corresponding to each product.

4. The product recommendation method according to claim 3, characterized in that, The process of performing recommendation analysis on all customers based on all predicted probabilities, all recommended star ratings, all normalized importance scores, and all product business data to obtain product recommendation results includes: Feature contribution analysis was performed on multiple product business data of each customer and multiple normalized importance scores corresponding to each product to obtain a customer feature contribution ranking table corresponding to each customer. The recommendation reasons for each customer are analyzed by ranking the contribution of each customer's characteristics and the multiple recommendation stars corresponding to each customer. Import the multiple product service priorities corresponding to each customer, and calculate the multiple predicted probabilities, multiple recommendation star ratings, and multiple product service priorities corresponding to each customer using the second formula to obtain multiple comprehensive scores corresponding to each customer. The second formula is: ， in, For the first The overall score corresponding to each product. , as well as All are weighting coefficients. For the first The predicted probability for each product. For the first Recommended star rating for each product. For the first Product business priority corresponding to each product; The maximum values ​​of multiple comprehensive scores corresponding to each customer are filtered out, and the maximum comprehensive score corresponding to each customer is obtained after filtering. Recommendation information for each customer is constructed by using the products corresponding to each maximum comprehensive score, the predicted probabilities corresponding to each maximum comprehensive score, the recommendation star rating corresponding to each maximum comprehensive score, the customer feature contribution ranking table corresponding to each customer, and the recommendation reasons corresponding to each customer. All of the recommendation information is then used as the product recommendation result.

5. The product recommendation method according to claim 4, characterized in that, The process of performing feature contribution analysis on multiple product business data of each customer and multiple normalized importance scores corresponding to each product to obtain a customer feature contribution ranking table corresponding to each customer includes: Calculate the mean of each product feature corresponding to all the customers to obtain the mean of the product feature corresponding to each product feature; Calculate the standard deviation of each product feature corresponding to all the customers, and obtain the standard deviation of the product feature corresponding to each product feature; The contribution of multiple features corresponding to each customer is calculated using the third formula, which involves analyzing the multiple product business data of each customer, the mean of the product features corresponding to each product feature, the standard deviation of the product features corresponding to each product feature, and the multiple normalized importance scores corresponding to each product feature. The third formula is as follows: ， in, For the first The feature contribution of each product feature For the first The first customer Product business data corresponding to each product feature For the first The first of the products Normalized importance scores corresponding to each product feature For the first The average value of product features corresponding to each product feature. For the first The standard deviation of product characteristics corresponding to each product characteristic; The multiple feature contributions corresponding to each customer are sorted in descending order of feature contribution, and the top N sorted feature contributions are plotted as customer feature contribution ranking tables, thereby obtaining customer feature contribution ranking tables corresponding to each customer.

6. The product recommendation method according to claim 4, characterized in that, The process of analyzing the contribution ranking table of each customer feature and the multiple recommendation stars corresponding to each customer to obtain the recommendation reasons for each customer includes: Each of the customer feature contribution ranking tables is converted into text to obtain customer semantic description text corresponding to each customer. According to the preset hierarchical calling rules, the model is called for multiple recommendation star ratings corresponding to each customer to obtain the reason generation model corresponding to each customer. Each of the aforementioned reason generation models generates a reason for each customer's semantic description text, thereby obtaining a recommendation reason for each customer.

7. The product recommendation method according to claim 4, characterized in that, Also includes: The fourth equation is used to calculate the importance vectors of each product feature and the remaining product feature importance vectors, respectively, to obtain multiple importance similarities corresponding to each product. The fourth equation is: ， in, For the first The product and the first The similarity in importance of individual products For the first Product feature importance vectors corresponding to each product For the first Product feature importance vectors corresponding to each product; If the importance similarity is greater than a preset similarity threshold, then cross-recommendation processing is performed on the customers corresponding to the importance similarity.

8. A product recommendation device, characterized in that, include: The import module is used to import multiple original customer consumption data sets and multiple customer consumption data sets to be recommended. The training module is used to build an initial prediction model. The initial prediction model is trained using all the original customer consumption data and all the customer consumption data to be recommended, so as to obtain the target prediction model corresponding to each product. The first prediction module is used to predict the normalized consumption data of the customers to be recommended through the target prediction model, and obtain the product feature importance vector corresponding to each product. The product feature importance vector includes the original importance score corresponding to each product feature. The normalization module is used to normalize each of the original importance scores to obtain the normalized importance scores corresponding to each of the product features. The difference coefficient analysis module is used to analyze the feature sensitivity difference coefficient of each of the normalized importance scores, and to combine the analysis results to obtain the target product feature set corresponding to each of the products. The second prediction module is used to predict each of the target product feature sets using each of the target prediction models, and to obtain multiple predicted probabilities and multiple recommendation star ratings corresponding to each customer. The import module is also used to import multiple product business data of each of the customers; The recommendation result acquisition module is used to perform recommendation analysis on all customers based on all the predicted probabilities, all the recommendation star ratings, all the normalized importance scores, and all the product business data to obtain product recommendation results.

9. A product recommendation device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the product recommendation method as described in any one of claims 1 to 7.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the product recommendation method as described in any one of claims 1 to 7.

Images