An intelligent recommendation method and system for government and enterprise information products based on hybrid model

Abstract

The present invention discloses an intelligent recommendation method and system for government and enterprise information products based on a hybrid model. Based on the hybrid model and customer characteristics, the product-package sales product candidate pool is recalled, covering new product recommendations, package sales product recommendations, information product recommendations and other multi-scenario businesses; at the sorting layer, the candidate products are scored according to the FM model, and the recommendation results are output according to the size of the scores; at the business layer, the model output rules and business characteristics of each stage are combined to provide corresponding recommendation explanations for different customer recommendation information. The present invention intelligently recommends government and enterprise solutions centered on government and enterprise customers, realizing the transformation from "product sales" to "customer orientation". Intelligent recommendations for government and enterprise customers, based on the user big data of government and enterprises, realize the automatic calculation of business opportunities according to the needs of different business scenarios, and complete marketing order recommendations. Business opportunity recommendations are more accurate and reliable, and the coverage rate of business opportunity recommendations is high.

Description

Technical Field

[0001] The present invention relates to the technical field of government and enterprise software development, and in particular to an intelligent recommendation method and system for government and enterprise information products based on a hybrid model. Background Art

[0002] With the saturation of the mobile user market for operators and the slowdown in revenue growth in the public market, the government and enterprise customer market has become a new revenue growth point for operators, offering significant revenue potential. Therefore, effective marketing efforts for these customers are particularly important.

[0003] Government and enterprise customers come from diverse industries, and the products and solutions offered by these industries vary significantly. The products and solutions offered to these customers are complex. Account managers primarily rely on current government and enterprise customer tags and product view information, drawing on their own business experience to make marketing recommendations. Relying on account managers' business experience to recommend business opportunities for government and enterprise customers fails to effectively utilize their information, resulting in unreliable marketing results and a poor customer experience. Summary of the Invention

[0004] The purpose of the present invention is to provide an intelligent recommendation method and system for government and enterprise information products based on a hybrid model.

[0005] The technical solution adopted in the present invention is:

[0006] An intelligent recommendation method for government and enterprise information products based on a hybrid model, comprising the following steps:

[0007] S1: At the recall layer, the customer's historical feature information and instantiated information products are used to identify the relationship between customer features and recommended products through association rule algorithms to roughly select candidate pool A from a large number of recommended products for the customer;

[0008] S2: Based on the information-based product content, a collaborative filtering algorithm is used to calculate the top n similar products for each product, which is used to solve the problem of cold start of items, and the candidate pool A is associated to obtain the candidate pool B;

[0009] S3: The product-package sales rule extracts the package sales product specification data of the information product and associates it with the customer's income and product information to match the corresponding package sales products to candidate pools A and B. At the same time, candidate products that do not match the business rules are eliminated to form the final product-package sales product candidate pool C.

[0010] S4, in the sorting layer, the recall layer data is precisely sorted using the factor decomposition model and the logistic regression model, and the corresponding factor importance ranking is output;

[0011] S5, at the business layer, aggregates model rules from different stages and outputs a recommendation explanation for each recommendation record.

[0012] Furthermore, as a preferred implementation, the specific steps of S1 are as follows:

[0013] S11: Design customer feature types, including customer income, basic business score, information type score, customer industry type, customer business nature, etc.; and discretize continuous data;

[0014] S12: Extract the package sales items accepted by the customer in the next three months and correspond them to the corresponding information products to obtain data in the form of "customer characteristics-information products";

[0015] S13: Divide the sample data into a training set and a test set in proportion; use the association rule algorithm to initialize the credibility and confidence thresholds, perform iterative training at a certain step size, select the appropriate thresholds according to the recall rate and precision rate of the training set and the test set, and only select the corresponding rule items in the strong association rules whose consequent is the product to be recommended.

[0016] S14: Complete the matching of customer candidate pool A according to the strong association rules screened out in S13.

[0017] Furthermore, as a preferred implementation method, the discretization processing of the score data in S11 is specifically as follows: each score data is divided by the total score of the type, and the clustering method is used to perform sample category and sample quantity statistics, the appropriate number of clusters and cluster centers are selected, the mean of the two adjacent categories is taken as the critical point, and a reasonable explanation is attached to each type.

[0018] Furthermore, as a preferred implementation method, the collaborative filtering based on the information product content in S2 calculates the similarity of the information products through keyword extraction method, keyword vectorization, and cosine similarity.

[0019] Furthermore, as a preferred embodiment, the specific steps of S2 are as follows:

[0020] S21: Data preprocessing: extract the content of information products and preprocess space symbols and paragraphs;

[0021] S22: Word segmentation: Use word segmentation algorithms to segment product content and filter out Chinese symbols, auxiliary words, quantifiers, pronouns, and function words. Professional terms are supplemented by loading custom dictionaries.

[0022] S23: Keyword extraction: Use the keyword extraction model to extract keywords from the information product content, and obtain a word dictionary and a set of keywords respectively;

[0023] S24: Similarity calculation: Vectorize the keywords corresponding to the product content based on the word dictionary, and calculate the size of the two sets of keyword vectors using the similarity formula as the similarity between the information products;

[0024] S25: According to multiple different keyword extraction methods, each information product X i According to the similarity, we can get multiple groups of n similar information products U (the top n are ranked according to the similarity) ij (j>=1) Take the intersection of three groups of similar products to get X i The final set of similar products U i =∩ j≥1 U ij , and associate candidate pool A to obtain candidate pool B.

[0025] Furthermore, as a preferred implementation method, the product-package sales rules in S3 are: extract the specification data of the package sales products corresponding to the information products and match them with the income and product information under the customer's name, match the corresponding package sales products for candidate pools A and B, and at the same time eliminate unsuitable candidate products such as lower grade and lack of product functions to form the final product-package sales candidate pool C.

[0026] Furthermore, as a preferred implementation method, S4's precise ranking uses a factor decomposition machine (FM model) to more accurately calculate whether the customer accepts a certain recommendation record and the probability of acceptance based on the content of the recall layer and customer characteristics, so as to accurately score each candidate content.

[0027] Furthermore, the present invention also discloses an intelligent recommendation system for government and enterprise information products based on a hybrid model. The system adopts the intelligent recommendation method for government and enterprise information products based on a hybrid model, and includes the following parts:

[0028] Data preprocessing: This includes feature engineering and designing customer features to make the data more suitable for the selected model algorithm. The kmeans clustering method is used to discretize continuous data to ensure that the data is suitable for the Apriori algorithm.

[0029] Recall layer: Based on the association rule algorithm, the relationship between customer characteristics and recommended products is identified to roughly select candidate pool A from the numerous recommended products for the customer. A collaborative filtering algorithm based on the information product content calculates the top n similar products for each product and associates them with candidate pool A to obtain candidate pool B. By extracting the specification data of the package sales items corresponding to the information product and correlating it with the customer's income and product information, the corresponding package sales items are matched to candidate pools A and B. At the same time, candidate products that do not match the business rules are eliminated to form a package sales candidate pool C.

[0030] Ranking layer: A factorization machine (FM model) is used to more accurately calculate the customer's acceptance and likelihood of a recommendation based on the candidate pool C and customer characteristics. Each candidate is accurately scored and the candidate set C is precisely ranked based on the score.

[0031] Business layer: Aggregate the rules of each stage and supplement them with business explanations to ensure that each recommendation record has a corresponding recommendation explanation.

[0032] The present invention adopts the above technical solution to recall candidate content based on a hybrid model according to customer characteristics and information product content, covering multiple business scenarios such as new product recommendations, package sales product recommendations, and information product recommendations; at different stages, each recommendation process has corresponding business support, that is, by aggregating model rules at different stages, a recommendation explanation is output for each recommendation record, providing business-oriented explanations on the basis of ensuring the accuracy of the recommendation, and better informing the marketing of front-line customer managers.

[0033] This invention intelligently recommends solutions for government and enterprise customers, shifting from a "product sales" approach to a "customer-centric" approach. Based on big data from government and enterprise users, this intelligent recommendation automatically calculates business opportunities and recommends marketing orders based on the needs of different business scenarios. This makes business opportunity recommendations more accurate and reliable, with a high coverage rate. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments;

[0035] Figure 1 A flowchart of a method for recommending an information product-package sales item according to an embodiment of the present invention;

[0036] Figure 2 A flowchart of feature engineering in one embodiment of the present invention;

[0037] Figure 3 This is a flowchart of collaborative filtering based on information product content in one embodiment of the present invention. DETAILED DESCRIPTION

[0038] In order to make the purpose, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application.

[0039] like Figures 1 to 3 As shown in one, the present invention discloses an intelligent recommendation method for government and enterprise information products based on a hybrid model, which includes the following steps:

[0040] S1: At the recall layer, the customer's historical feature information and instantiated information products are used to identify the relationship between customer features and recommended products through association rule algorithms to roughly select candidate pool A from a large number of recommended products for the customer;

[0041] Specifically, the core idea of ​​association rules is to generate candidate items and their support through concatenation, and then prune them to generate frequent itemsets. By designing parameters such as customer features, confidence thresholds, and support thresholds, the frequency and confidence of the occurrences between customer features and target products are identified, and corresponding rules are selected as candidate recall rules to form a candidate set.

[0042] The recall layer uses the Apriori algorithm to train the model data. Initial support and confidence thresholds are set to 0.03 and 0.5, respectively. Iterative training is performed with step sizes of 0.005 and 0.05 until data convergence. The rule with the highest predicted F1 score and the consequent being the recommended information product is selected.

[0043] S2: Based on the information-based product content, a collaborative filtering algorithm is used to calculate the top n similar products for each product, which is used to solve the problem of cold start of items, and the candidate pool A is associated to obtain the candidate pool B;

[0044] S3: The product-package sales rule extracts the package sales product specification data of the information product and associates it with the customer's income and product information to match the corresponding package sales products to candidate pools A and B. At the same time, candidate products that do not match the business rules are eliminated to form the final product-package sales product candidate pool C.

[0045] S4, in the sorting layer, the recall layer data is precisely sorted using the factor decomposition model and the logistic regression model, and the corresponding factor importance ranking is output;

[0046] S5, at the business layer, aggregates model rules from different stages and outputs a recommendation explanation for each recommendation record.

[0047] Further, if Figure 2 As shown in the figure, as a preferred implementation, S1 designs relevant customer features for feature engineering to make the data more suitable for the selected model algorithm. In this example, the kmeans clustering method is used to discretize the continuous data. The specific steps are as follows:

[0048] S11: Design customer feature types, including customer income, basic business score, information type score, customer industry type, customer business nature, etc.; and discretize continuous data;

[0049] S12: Extract the package sales items accepted by the customer in the next three months and correspond them to the corresponding information products to obtain data in the form of "customer characteristics-information products";

[0050] S13: Divide the sample data into a training set and a test set in proportion; use the association rule algorithm to initialize the credibility and confidence thresholds, perform iterative training at a certain step size, select the appropriate thresholds according to the recall rate and precision rate of the training set and the test set, and only select the corresponding rule items in the strong association rules whose consequent is the product to be recommended.

[0051] S14: Complete the matching of customer candidate pool A according to the strong association rules screened out in S13.

[0052] Furthermore, as a preferred implementation method, the discretization processing of the score data in S11 is specifically as follows: each score data is divided by the total score of the type, and the clustering method is used to perform sample category and sample quantity statistics, the appropriate number of clusters and cluster centers are selected, the mean of the two adjacent categories is taken as the critical point, and a reasonable explanation is attached to each type.

[0053] Further, if Figure 3 As shown, as a preferred implementation method, the collaborative filtering based on the information product content in S2 calculates the similarity of information products through keyword extraction method, keyword vectorization, and cosine similarity.

[0054] Furthermore, as a preferred embodiment, the specific steps of S2 are as follows:

[0055] S21: Data preprocessing: extract the content of information products and preprocess space symbols and paragraphs;

[0056] S22: Word segmentation: Use word segmentation algorithms to segment product content and filter out Chinese symbols, auxiliary words, quantifiers, pronouns, and function words. Professional terms are supplemented by loading custom dictionaries.

[0057] S23: Keyword extraction: Use the keyword extraction model to extract keywords from the information product content, and obtain a word dictionary and a set of keywords respectively;

[0058] Specifically, TF-IDF, LSI, and LDA models are used to extract keywords for each product content;

[0059] When using TF-IDF to calculate product content keywords, a dictionary of IDF values ​​is generated based on the training dataset. When calculating the TF-IDF of each word, it is directly read from the dictionary. For words that do not appear in the dictionary, it is assumed that they only appear in one document. At the same time, the denominator is added with 1 for smoothing when calculating the IDF value. Calculate the top ten keyword set K1 for each product content;

[0060] When using the LSI and LDA models to calculate product content keywords: Build a dictionary; Use the BOW model for vectorization; Use TF-IDF to weight each word to obtain a weighted vector representation; Directly call the model trained by the gensim module; Set the number of topics to 4 to generate the word-topic and topic-document distributions for the dataset; Based on the word-topic and topic-document distributions, calculate the top ten keyword sets K2 and K3 for each product content.

[0061] S24: Similarity calculation: Vectorize the keywords corresponding to the product content based on the word dictionary, and calculate the size of the two sets of keyword vectors using the similarity formula as the similarity between the information products;

[0062] Specifically, similarity calculation: for product X i and X j Use BOW model vectorization and TF-IDF to weight the corresponding keyword set K ik (k=1,2,3) and K jk (k=1,2,3) to get the weighted keyword vector V ik =(V ik_1 ,V ik_2 ,V ik_3 ,...,V ik_m )(k=1,2,3) and V jk =(V jk_1 ,V jk_2 ,V jk_3 ,...,V jk_m )(k=1,2,3), the size between two sets of keyword vectors is calculated by cosine similarity.

[0063] The similarity calculation formula is defined as follows:

[0064]

[0065] S25: According to multiple different keyword extraction methods, each information product X i According to the similarity, we can get multiple groups of n similar information products U (the top n are ranked according to the similarity) ij (j>=1) Take the intersection of three groups of similar products to get X i The final set of similar products U i =∩ j≥1 U ij , and associate candidate pool A to obtain candidate pool B.

[0066] Specifically, for product X i The top ten similarity ranking sets U calculated based on different keyword extraction methodsik (k=1,2,3), take the intersection of the three as the product X i Similar set U i =U i1 ∩U i2 ∩U i3 , candidate pool A is associated with similar set U i Get candidate pool B;

[0067] Furthermore, as a preferred implementation method, the product-package sales rules in S3 are: extract the specification data of the package sales products corresponding to the information products and match them with the income and product information under the customer's name, match the corresponding package sales products for candidate pools A and B, and at the same time eliminate unsuitable candidate products such as lower grade and lack of product functions to form the final product-package sales candidate pool C.

[0068] Furthermore, as a preferred implementation method, S4's precise ranking uses a factor decomposition machine (FM model) to more accurately calculate whether the customer accepts a certain recommendation record and the probability of acceptance based on the content of the recall layer and customer characteristics, so as to accurately score each candidate content.

[0069] Furthermore, the present invention also discloses an intelligent recommendation system for government and enterprise information products based on a hybrid model. The system adopts the intelligent recommendation method for government and enterprise information products based on a hybrid model, and includes the following parts:

[0070] Data preprocessing: This includes feature engineering and designing customer features to make the data more suitable for the selected model algorithm. The kmeans clustering method is used to discretize continuous data to ensure that the data is suitable for the Apriori algorithm.

[0071] Recall layer: Based on the association rule algorithm, the relationship between customer characteristics and recommended products is identified to roughly select candidate pool A from the numerous recommended products for the customer. A collaborative filtering algorithm based on the information product content calculates the top n similar products for each product and associates them with candidate pool A to obtain candidate pool B. By extracting the specification data of the package sales items corresponding to the information product and correlating it with the customer's income and product information, the corresponding package sales items are matched to candidate pools A and B. At the same time, candidate products that do not match the business rules are eliminated to form a package sales candidate pool C.

[0072] Specifically, the recall layer primarily screens candidate products. The model data is trained using the Apriori algorithm, with initial support and confidence thresholds set to 0.03 and 0.5, respectively. Iterative training is performed with step sizes of 0.005 and 0.05 until data convergence. The rule with the highest predicted F1 score and the consequent being the recommended information technology product is selected.

[0073] Ranking layer: A factorization machine (FM model) is used to more accurately calculate the customer's acceptance and likelihood of a recommendation based on the candidate pool C and customer characteristics. Each candidate is accurately scored and the candidate set C is precisely ranked based on the score.

[0074] Business layer: Aggregate the rules of each stage and supplement them with business explanations to ensure that each recommendation record has a corresponding recommendation explanation.

[0075] The present invention adopts the above technical solution, and recalls the candidate pool of package sales products based on the hybrid model and customer characteristics, covering multiple business scenarios such as new product recommendations, package sales product recommendations, and information product recommendations; scores the candidate products according to the FM model at the sorting layer, and outputs the recommendation results based on the size of the scores; at the business layer, it provides corresponding recommendation explanations for different customer recommendation information in combination with the model output rules and business characteristics of each stage. The present invention recalls candidate content based on the hybrid model according to customer characteristics and information product content, covering multiple business scenarios such as new product recommendations, package sales product recommendations, and information product recommendations; at different stages, each recommendation process has corresponding business support, that is, by aggregating the model rules of different stages, a recommendation explanation is output for each recommendation record, providing business explanations on the basis of ensuring the accuracy of the recommendation, and better informing the marketing of front-line customer managers.

[0076] This invention intelligently recommends solutions for government and enterprise customers, shifting from a "product sales" approach to a "customer-centric" approach. Based on big data from government and enterprise users, this intelligent recommendation automatically calculates business opportunities and recommends marketing orders based on the needs of different business scenarios. This makes business opportunity recommendations more accurate and reliable, with a high coverage rate.

[0077] Obviously, the described embodiments are part of the embodiments of the present application, rather than all of the embodiments. In the absence of conflict, the embodiments in the present application and the features in the embodiments can be combined with each other. The components of the embodiments of the present application generally described and shown in the drawings here can be arranged and designed in various different configurations. Therefore, the detailed description of the embodiments of the present application is not intended to limit the scope of the application for protection, but merely represents the selected embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of this application.

Claims

1. An intelligent recommendation method for government and enterprise information products based on a hybrid model, characterized by: It includes the following steps: S1, at the recall layer, uses customer historical feature information and instantiated information products to identify the relationship between customer features and recommended products through an association rule algorithm, and roughly selects a candidate pool A from a large number of recommended products for the customer. The specific steps of S1 are as follows: S11: Design customer feature types, including customer income, basic business score, information type score, customer industry type, and customer business nature; and discretize continuous data; S12: Extract the package sales items accepted by the customer in the next three months and map them to the corresponding information products to obtain data in the form of "customer characteristics-information products"; S13: Divide the sample data into a training set and a test set in proportion; use the association rule algorithm to initialize the credibility and confidence thresholds, perform iterative training at a certain step size, and select appropriate thresholds based on the recall and precision of the training and test sets. Only select the rule items whose consequents in the strong association rules are the corresponding products to be recommended; S14: Complete the matching of customer candidate pool A according to the strong association rules screened in S13; S2: A collaborative filtering algorithm based on the information-based product content is used to calculate the top n similar products for each product, which is used to solve the problem of cold start of items, and associates candidate pool A to obtain candidate pool B; S3: The product-package sales rule extracts the package sales product specification data of the information product and associates it with the customer's income and product information to match the corresponding package sales products to candidate pools A and B. At the same time, candidate products that do not match the business rules are eliminated to form the final product-package sales product candidate pool C. S4, in the sorting layer, the recall layer data is precisely sorted using the factor decomposition model and the logistic regression model, and the corresponding factor importance ranking is output; S5, at the business layer, aggregates model rules from different stages and outputs a recommendation explanation for each recommendation record.

2. The method for intelligently recommending government and enterprise information products based on a hybrid model according to claim 1, characterized in that: The data discretization processing in S11 is specifically as follows: each score data is divided by the total score of the type, and the clustering method is used to count the sample categories and sample quantities, the appropriate number of clusters and cluster centers are selected, the mean of the two adjacent categories is taken as the critical point, and a reasonable explanation is attached to each type.

3. The method for intelligently recommending government and enterprise information products based on a hybrid model according to claim 1, characterized in that: In S2, collaborative filtering based on information product content calculates the similarity of information products through keyword extraction method, keyword vectorization, and cosine similarity.

4. The method for intelligently recommending government and enterprise information products based on a hybrid model according to claim 1 or 3, characterized in that: The specific steps of S2 are as follows: S21: Data preprocessing: extract the content of information products and preprocess space symbols and paragraphs; S22: Word segmentation: Use a word segmentation algorithm to segment product content and filter out Chinese symbols, auxiliary words, quantifiers, pronouns, and function words. Professional terms are supplemented by loading a custom dictionary. S23: Keyword extraction: Use the keyword extraction model to extract keywords from the information product content, and obtain a word dictionary and a set of keywords respectively; S24: Similarity calculation: Vectorize the keywords corresponding to the product content based on the word dictionary, and calculate the size of the two sets of keyword vectors using the similarity formula as the similarity between the information products; S25: According to multiple different keyword extraction methods, each information product X i Take the top n rankings according to the similarity to get multiple groups of n similar information products. U ij , take the intersection of three groups of similar products and get X i The final set of similar products U i = and j >=1, and associate candidate pool A to obtain candidate pool B.

5. The method for intelligently recommending government and enterprise information products based on a hybrid model according to claim 1 is characterized by: The product-package sales rules in S3 are: extract the specification data of the corresponding package sales products of the information products and match them with the income and product information under the customer's name, match the corresponding package sales products for candidate pools A and B, and at the same time eliminate unsuitable candidate products with lower grades and lack of product functions to form the final product-package sales candidate pool C.

6. The method for intelligently recommending government and enterprise information products based on a hybrid model according to claim 1, characterized in that: S4's precise ranking uses a factor decomposition machine to more accurately calculate whether a customer will accept a certain recommendation record and the likelihood of acceptance based on the content of the recall layer and customer characteristics, thereby accurately scoring each candidate content.

7. An intelligent recommendation system for government and enterprise information products based on a hybrid model, using the intelligent recommendation method for government and enterprise information products based on a hybrid model according to any one of claims 1 to 6, characterized in that: The system consists of the following parts: Data preprocessing: This includes feature engineering and designing customer features to make the data more suitable for the selected model algorithm. The kmeans clustering method is used to discretize continuous data to ensure that the data is suitable for the Apriori algorithm. Recall layer: Using an association rule algorithm to identify the relationship between customer characteristics and recommended products, a candidate pool A is roughly selected from a large number of recommended products. Specifically, the package sales products that the customer will accept in the next three months are mapped to the corresponding information products to obtain data in the form of "customer characteristics-information products". The sample data is then divided into training and test sets according to a certain ratio. The association rule algorithm is used to initialize the credibility and confidence thresholds, and iterative training is performed at a certain step size. Appropriate thresholds are selected based on the recall and precision of the training and test sets. Only rule items whose consequents are recommended products in strong association rules are selected. Matching is completed to customer candidate pool A based on the selected strong association rules. A collaborative filtering algorithm based on the information product content calculates the top n similar products for each product and associates them with candidate pool A to obtain candidate pool B. By extracting the specification data of the package sales products corresponding to the information products and correlating them with the customer's income and product information, corresponding package sales products are matched to candidate pools A and B. At the same time, candidate products that do not match the business rules are eliminated to form a package sales candidate pool C. Ranking layer: A factorization machine is used to more accurately calculate the customer's acceptance and likelihood of a recommendation based on the candidate pool C and customer characteristics. Each candidate is accurately scored and the candidate set C is precisely ranked based on the score. Business layer: Aggregate the rules of each stage and supplement them with business explanations to ensure that each recommendation record has a corresponding recommendation explanation.

Images