WO2020048061A1 - 产品推荐方法、装置、计算机设备和存储介质 - Google Patents

产品推荐方法、装置、计算机设备和存储介质 Download PDF

Info

Publication number
WO2020048061A1
WO2020048061A1 PCT/CN2018/124391 CN2018124391W WO2020048061A1 WO 2020048061 A1 WO2020048061 A1 WO 2020048061A1 CN 2018124391 W CN2018124391 W CN 2018124391W WO 2020048061 A1 WO2020048061 A1 WO 2020048061A1
Authority
WO
WIPO (PCT)
Prior art keywords
product
information
customer
word
recommendation
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Ceased
Application number
PCT/CN2018/124391
Other languages
English (en)
French (fr)
Inventor
金戈
徐亮
肖京
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Ping An Technology Shenzhen Co Ltd
Original Assignee
Ping An Technology Shenzhen Co Ltd
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Ping An Technology Shenzhen Co Ltd filed Critical Ping An Technology Shenzhen Co Ltd
Publication of WO2020048061A1 publication Critical patent/WO2020048061A1/zh
Anticipated expiration legal-status Critical
Ceased legal-status Critical Current

Links

Classifications

    • GPHYSICS
    • G06COMPUTING OR CALCULATING; COUNTING
    • G06QINFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
    • G06Q40/00Finance; Insurance; Tax strategies; Processing of corporate or income taxes
    • G06Q40/06Asset management; Financial planning or analysis
    • GPHYSICS
    • G06COMPUTING OR CALCULATING; COUNTING
    • G06FELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00Pattern recognition
    • G06F18/20Analysing
    • G06F18/24Classification techniques
    • G06F18/241Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
    • G06F18/2411Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on the proximity to a decision surface, e.g. support vector machines
    • GPHYSICS
    • G06COMPUTING OR CALCULATING; COUNTING
    • G06NCOMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00Computing arrangements based on biological models
    • G06N3/02Neural networks
    • G06N3/04Architecture, e.g. interconnection topology
    • G06N3/045Combinations of networks
    • GPHYSICS
    • G06COMPUTING OR CALCULATING; COUNTING
    • G06NCOMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00Computing arrangements based on biological models
    • G06N3/02Neural networks
    • G06N3/08Learning methods
    • G06N3/084Backpropagation, e.g. using gradient descent
    • GPHYSICS
    • G06COMPUTING OR CALCULATING; COUNTING
    • G06QINFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
    • G06Q40/00Finance; Insurance; Tax strategies; Processing of corporate or income taxes
    • G06Q40/08Insurance

Definitions

  • This application relates to the field of computers, and in particular, to a method, an apparatus, a computer device, and a storage medium for product recommendation.
  • the main purpose of this application is to provide a product recommendation method, device, computer equipment, and storage medium, which are intended to effectively recommend products to customers.
  • This application proposes a product recommendation method, including:
  • Output a product corresponding to the second representation layer vector and recommend it to the customer.
  • This application also provides a product recommendation device, including:
  • An obtaining unit configured to obtain first retrieval information of a client, and search a preset first corpus dictionary for a word vector corresponding to each word in the first retrieval information to obtain a first word corresponding to the first retrieval information Vector matrix
  • a deep learning unit configured to input the first word vector matrix into a first recommendation model trained based on an LSTM model to learn, and output a first representation layer vector corresponding to product information of the first word vector matrix;
  • a similarity calculation unit configured to find a second representation layer vector with the highest similarity to the first representation layer vector in the trained product representation layer vector matrix of the corresponding product
  • a recommendation unit is configured to output a product corresponding to the second representation layer vector and recommend the product to the customer.
  • the present application further provides a computer device including a memory and a processor, where the memory stores computer-readable instructions, and when the processor executes the computer-readable instructions, implements the steps of any one of the foregoing methods.
  • the present application also provides a computer non-volatile readable storage medium having computer readable instructions stored thereon, which are executed by a processor to implement the steps of the method according to any one of the above.
  • the product recommendation method, device, computer equipment, and storage medium of the present application vectorize the customer search information in the historical data and the product information of the product they purchased, and then input it into the LSTM model for training to obtain the first recommendation model. Then use the first recommendation model to calculate the product corresponding to the customer's search term to recommend the appropriate product to the customer, which can effectively recommend the product to the customer without over recommending the popular product.
  • FIG. 1 is a schematic flowchart of a product recommendation method according to an embodiment of the present invention
  • FIG. 2 is a schematic block diagram of a product recommendation device according to an embodiment of the present invention.
  • FIG. 3 is a schematic block diagram of a product recommendation device according to an embodiment of the present invention.
  • FIG. 4 is a schematic block diagram of a product recommendation device according to an embodiment of the present invention.
  • FIG. 5 is a schematic block diagram of a product recommendation device according to an embodiment of the present invention.
  • FIG. 6 is a schematic block diagram of a product recommendation device according to an embodiment of the present invention.
  • FIG. 7 is a schematic block diagram of a structure of a recommendation unit according to an embodiment of the present invention.
  • FIG. 8 is a schematic block diagram of a structure of a recommendation module according to an embodiment of the present invention.
  • FIG. 9 is a schematic block diagram of a computer device according to an embodiment of the present invention.
  • the present application provides a product recommendation method, including steps:
  • the first word vector matrix is input to a first recommendation model trained based on the LSTM model to learn, and a first representation layer vector corresponding to product information of the first word vector matrix is output;
  • the first search information refers to the search information input by the customer, which includes keywords extracted by the customer according to his or her desire to purchase.
  • the keywords that he enters generally include Children, accidents, medical care, etc.
  • look up the word vector of each word in the first corpus dictionary and then arrange the word vector of each word in the order of the keywords to form the first Word vector matrix.
  • the above-mentioned first corpus dictionary is a dictionary in which word vectors and words are made one by one by a company that sells the above-mentioned products, and the words and word vectors contained therein are common words in their corresponding fields.
  • the process of making the first corpus dictionary includes: inputting text to be formed into a word vector into a DSSM (Deep Structured Semantic Models) model, calculating a word vector of the text through the DSSM model, and then writing the text and its corresponding The word vector is put into the first corpus dictionary.
  • the keywords can be extracted by the sales staff based on the customer's oral content, and then manually input into the computer equipment to complete the above step S1; or the customer's requirements are formed into text information.
  • the first recommendation model obtained based on the Long Short-Term Memory (LSTM) model training refers to that the first recommendation model is obtained by training the LSTM model.
  • the specific training process is: 1
  • the historical customer search information is formed into a plurality of word vector matrices through the above-mentioned first corpus dictionary, and the product information of the product purchased by the customer is formed into a corresponding plurality of word vector matrices through the first corpus dictionary to form the obtained word vector matrix.
  • the word vector matrix corresponding to the retrieval information of each customer is related to the word vector matrix corresponding to the product information of the product purchased by the customer; 2 the set of word vector matrices is divided into a training set and a test set; 3 the test The set of sample data is input into the LSTM model for training, and then the test set is used to verify the trained LSTM model. If the verification is passed, the first recommended model described above is obtained.
  • the above training process belongs to supervised training, that is, the historical customer retrieval information and customer purchased product information are information with associated tags.
  • the first recommended model trained is entered after the first word vector matrix of customer retrieval information is input. Output a first representation layer vector corresponding to product information.
  • each of the product representation layer vector matrices of the corresponding products (such as insurance products) that have been trained needs to be calculated
  • the similarity of the representation layer vector and find the second representation layer vector with the highest similarity.
  • the similarity can be calculated using Euclidian distance, Manhattan distance, and Minkowski distance. ) Or an algorithm based on cosine similarity.
  • the specific similarity calculation formula is as follows:
  • y Q is the representation layer vector of the retrieved information
  • Y D is the representation layer vector of the product information
  • T D / T Q is the length of the corresponding sentence
  • R is the similarity.
  • the product corresponding to the second representation layer vector is the product most suitable for the customer ’s purchase. Therefore, the product corresponding to the second representation layer vector is output as a product recommended to a new customer.
  • several second representation layer vectors with high similarity to the first representation layer vector may also be selected, and the products corresponding to each second representation layer vector may be output according to the similarity from high to high. Low order gives suggestions for different purchase strengths.
  • the input first retrieval information is "Children, Insurance, Accident”
  • the first representation layer vector output by the first recommendation model is closest to a second representation layer vector in a preset product representation layer vector matrix.
  • the insurance product corresponding to the second representation layer vector is children's accident insurance, and the children's medical insurance corresponding to the second surface layer vector with a slightly lower similarity to the first representation layer vector corresponding to the above "children, insurance, accident” will also be formed.
  • Recommended insurance products are launched to increase customer selectivity.
  • step S4 of outputting the product corresponding to the second representation layer vector and recommending it to the customer the method includes:
  • a product is recommended to the customer, the customer does not necessarily purchase the recommended product, so the product purchased by the customer is recorded, and the product information of the product purchased by the customer is vectorized to obtain a first product vector Matrix, and then store the first product vector matrix and the first word vector matrix (the vector matrix corresponding to the customer's first retrieval information) in a specified database for later use as training samples.
  • the above-mentioned designated data amount refers to a total amount of data storing the first product vector matrix and the first word vector matrix. When the amount of data reaches the specified threshold, it means that the number of samples meets the requirements, and then the above first recommendation model is trained.
  • the sample data in the database is used to continue training the first recommendation model to obtain a new first recommendation model, which can improve Accuracy of the above-mentioned first recommendation model.
  • the training method is the same as the above-mentioned method for training the first recommendation model, and is not repeated here.
  • the recommendation result is made more accurate by continuously optimizing the first recommendation model.
  • step S4 of outputting the product corresponding to the second representation layer vector and recommending it to the customer the method includes:
  • S412 Vectorize the new product information to obtain a second product vector matrix, and input the second product vector matrix into the backward recommendation model trained based on the LSTM model to output a third representation layer vector corresponding to the second product feature vector matrix.
  • the presentation layer vector is a vector corresponding to the information retrieved by the customer;
  • the specified contact methods generally include a mobile phone number, an email address, and other contact methods that can receive messages sent by others without adding a friend.
  • a mobile phone number can be used to directly send a text message to a customer.
  • the customer's information is associated with the first word vector in order to find the customer by the word vector.
  • the first word vector matrix and the customer's information are in a one-to-one mapping relationship.
  • the above-mentioned reverse recommendation model obtained based on the LSTM model training is obtained by training the LSTM model.
  • the specific training process is as follows: 1 Pass the historical customer retrieval information and customer purchased product information through the first A corpus dictionary is vectorized and a corresponding set of word vector matrices is obtained. 2 The set of word vector matrices is divided into a training set and a test set. 3 The sample data of the test set is input into the LSTM model for training, and then the test is performed. The set verifies the trained LSTM model. If the verification passes, the above-mentioned reverse recommendation model is obtained.
  • the above training process belongs to supervised training, that is, historical customer retrieval information and customer purchased product information are information with associated labels.
  • the trained backward recommendation model when the second product vector matrix of product information is input to the backward recommendation After the model, the reverse recommendation model will output a fifth representation layer vector corresponding to the customer retrieval information.
  • step S413 it is to calculate the similarity between the third representation layer vector and each first word vector matrix in the customer database, and then the first word whose similarity with the third representation layer vector reaches the specified requirements Vector matrix extraction process.
  • the above specified requirement refers to a first word vector matrix whose similarity with the third representation layer vector reaches a specified threshold, which may be one or more.
  • the above-mentioned specified threshold will be appropriately set to facilitate finding a large number of customer groups. If only one with the highest similarity is found, it obviously does not meet the original intention of finding customers to recommend products.
  • the extracted first word vector matrix is a word vector matrix with a high degree of similarity to the third representation layer vector output by the reverse recommendation model, so its corresponding customer should have a relatively high The desire to purchase and the corresponding purchasing power, so it is more efficient to send new products to the corresponding customers, and recommend resale.
  • the above-mentioned first retrieval information of the client is obtained, and a word vector corresponding to each word in the first retrieval information is found in a preset first corpus dictionary to obtain a first Before step S1 of the word vector matrix, it includes:
  • the designated website is generally a website that sells the above products, and the types of products sold on this website are similar.
  • the products sold are all insurance products, but the types of insurance are different, such as personal accident insurance. Products, children's medical insurance products, major illness insurance products, etc.
  • the website will introduce the products it sells, and the introduction can be used as product information.
  • step S102 because the types of the products are similar, the same words will repeatedly appear in the product information. Therefore, the same words need to be extracted, and each of the repeated words should be kept separately. To prevent multiple vectorizations of the same repeated word in the later period.
  • the remaining words are the words that do not appear repeatedly, and each repeated word only retains a set of words after it.
  • the text in the set is input into the DSSM model, and the word vector of the text is calculated by the DSSM model, and then the text and its corresponding word vector are put together to form the above-mentioned first corpus dictionary.
  • the set of refined words is the words that have appeared in all product introductions on the website, and only one set of words is retained for repeated words, it can not only meet the vectorized use of words, but also It can reduce the difficulty and workload of forming the first corpus dictionary.
  • step S4 of outputting the product corresponding to the second representation layer vector and recommending it to the customer the method includes:
  • an electronic archive is performed.
  • the customer's purchase situation can be automatically obtained from the electronic archive, that is, whether the customer has purchased the same product as the recommendation.
  • recommendations are given to customers for purchase recommendations, customers do not necessarily buy products based on recommendations.
  • the above The recommended model is not ideal, so stop using it.
  • the third count purchased by all customers within a specified period of time may also be recorded, and then the first count or the second count may be compared with the third count, and according to the result of the ratio, it may be determined whether the first count needs to be disabled.
  • the above specified time length refers to a longer period such as a quarter or a month, in order to obtain more data samples, and the obtained ratio is more available.
  • the above-mentioned step S4 of outputting a product corresponding to the second representation layer vector and recommending it to the customer includes:
  • the above introduction information is a related introduction to the product;
  • the above sales data generally includes the number of products sold, etc .;
  • the above-mentioned customer distribution refers to the age distribution and regional distribution of customers who purchased the product That is, through the distribution of this customer, you can know the situation of the customers who purchased the product, such as children's life insurance, etc.
  • the customers they purchased there were more female customers aged 20-30, and slightly fewer elderly customers aged 50-70.
  • There are fewer male customers, etc. so that customers can understand what kind of people are buying this insurance product, so that customers can choose whether to buy based on their own.
  • personal accident insurance has a lot of customers in northern winter.
  • the above-mentioned step S43 of introducing the output introduction information, sales data, and customer distribution of the output product into the recommendation information to the customer includes:
  • S432 Encapsulate the visualized drawing and introduction information of the output product into a document to form the recommendation information to the customer.
  • the sales data and customer distribution are made into a visual drawing, which is more convenient for new customers to view.
  • the geographical distribution in the customer distribution is reflected on a map, so that customers can know the regional differences at a glance. Wait.
  • the sales data can be reflected in the form of a curve, such as the sales trend curve from the beginning of the product's sales to the present, or the monthly sales volume, which greatly improves the speed at which customers can view recommended information.
  • the aforementioned products may include virtual products such as insurance products, financial products, and other physical products.
  • the product recommendation method in the embodiment of the present application vectorizes the customer retrieval information in the historical data and the product information of the product purchased by the customer, and then inputs it into the LSTM model for training to obtain a first recommendation model, and then uses the first recommendation model Calculate the products corresponding to the customer's search terms to recommend suitable products to customers, which can effectively recommend products to customers without over recommending popular products.
  • the present application further provides a product recommendation device, including steps:
  • An obtaining unit 10 is configured to obtain first search information of a client, and search a preset first corpus dictionary for a word vector corresponding to each word in the first search information to obtain a first search information corresponding to the first search information.
  • Word vector matrix
  • a deep learning unit 20 configured to input the first word vector matrix into a first recommendation model trained based on an LSTM model to learn, and output a first representation layer vector corresponding to product information of the first word vector matrix;
  • the similarity calculation unit 30 is configured to find a second representation layer vector with the highest similarity to the first representation layer vector in the trained product representation layer vector matrix of the corresponding product;
  • the recommendation unit 40 is configured to output a product corresponding to the second representation layer vector and recommend the product to the customer.
  • the first search information refers to search information input by a customer, which includes keywords extracted by the customer according to his or her desire to purchase. For example, if the customer wants to know about child insurance, the input keywords generally include Children, accidents, medical care, etc. After extracting these keywords, look up the word vector of each word in the first corpus dictionary, and then arrange the word vector of each word in the order of the keywords to form the first Word vector matrix.
  • the above-mentioned first corpus dictionary is a dictionary in which word vectors and words are made one by one by a company that sells the above-mentioned products, and the words and word vectors contained therein are common words in their corresponding fields.
  • the process of making the first corpus dictionary includes: inputting text to be formed into a word vector into a DSSM (Deep Structured Semantic Models) model, calculating a word vector of the text through the DSSM model, and then writing the text and its corresponding The word vector is put into the first corpus dictionary.
  • DSSM Deep Structured Semantic Models
  • the keywords can be extracted by the sales staff based on the customer's oral content, and then manually entered into the computer equipment to complete the vectorization of the first retrieval information; or the customer's It is required to form text information and input the text information into a keyword extraction model to extract keywords, and then vectorize the extracted keywords; it is also possible to use speech recognition technology to convert the customer's voice into text information, and then convert the text The information is input to a keyword extraction model to extract keywords and the like to complete the vectorization of the first retrieval information.
  • the first recommendation model obtained based on the LSTM (Long Short-Term Memory, long-term short-term memory network) training means that the first recommendation model is obtained by training the LSTM model.
  • the specific training process is : 1
  • the historical customer search information is formed into a plurality of word vector matrices through the above-mentioned first corpus dictionary, and the product information of the product purchased by the customer is formed into a corresponding plurality of word vector matrices through the first corpus dictionary to form the obtained word vector matrix.
  • the word vector matrix corresponding to the retrieval information of each customer is related to the word vector matrix corresponding to the product information of the product purchased by the customer; 2
  • the set of word vector matrices is divided into a training set and a test set; 3
  • the sample data of the test set is input into the LSTM model for training, and then the test set is used to verify the trained LSTM model. If the verification is passed, the above-mentioned first recommended model is obtained.
  • the above training process belongs to supervised training, that is, the historical customer retrieval information and customer purchased product information are information with associated tags.
  • the first recommended model trained is entered after the first word vector matrix of customer retrieval information is input. Output a first representation layer vector corresponding to product information.
  • the product representation layer vector matrix of the corresponding product for example, insurance product
  • the product representation layer vector matrix of the corresponding product for example, insurance product
  • the calculation of similarity can use Euclidian distance, Manhattan distance, and Minkowski distance ( Minkowski distance) or cosine similarity.
  • Minkowski distance Minkowski distance
  • cosine similarity the specific similarity calculation formula is as follows:
  • y Q is the representation layer vector of the retrieved information
  • Y D is the representation layer vector of the product information
  • T D / T Q is the length of the corresponding sentence
  • R is the similarity.
  • the product corresponding to the second representation layer vector is the product most suitable for the customer to purchase. Therefore, the product corresponding to the second representation layer vector is output as a product recommended to a new customer.
  • several second representation layer vectors with high similarity to the first representation layer vector may also be selected, and the products corresponding to each second representation layer vector may be output according to the similarity from high to high. Low order gives suggestions for different purchase strengths. For example, if the input first retrieval information is "Children, Insurance, Accident", then the first representation layer vector output by the first recommendation model is closest to a second representation layer vector in a preset product representation layer vector matrix.
  • the insurance product corresponding to the second representation layer vector is children's accident insurance, and the children's medical insurance corresponding to the second surface layer vector with a slightly lower similarity to the first representation layer vector corresponding to the above "children, insurance, accident” will also form Recommended insurance products are launched to increase customer selectivity.
  • the above product recommendation device further includes:
  • a recording unit 401 configured to record a product purchased by the customer, and the first search information
  • a vectorization unit 402 configured to vectorize first product information of a product purchased by the customer to obtain a first product vector matrix
  • An association storage unit 403, configured to store the first word vector matrix corresponding to the first search information and the first product vector matrix corresponding to the first product information of the product purchased by the user in a specified database;
  • a training unit 404 is configured to continue training the first recommendation model by using all the first product vector matrices and the first word vector matrix in the database when the amount of data in the database reaches a preset threshold to obtain a new The first recommendation model.
  • the product recommendation device recommends a product to a customer, the customer does not necessarily purchase the recommended product, so the product purchased by the customer is recorded ,
  • the above-mentioned designated data amount refers to a total amount of data storing the first product vector matrix and the first word vector matrix. When the amount of data reaches the specified threshold, it means that the number of samples meets the requirements, and then the above first recommendation model is trained.
  • the sample data in the database is used to continue training the first recommendation model to obtain a new first recommendation model, which can improve Accuracy of the above-mentioned first recommendation model.
  • the training method is the same as the above-mentioned method for training the first recommendation model, and is not repeated here.
  • the recommendation result is made more accurate by continuously optimizing the first recommendation model.
  • the foregoing product recommendation device further includes:
  • An association unit 411 configured to store the information of the customer in a customer database in association with the first word vector matrix, wherein the information of the customer includes a specified contact method
  • a calculation unit 412 configured to vectorize new product information to obtain a second product vector matrix, and input the second product vector matrix to the LSTM model training to output a third representation layer vector corresponding to the second product feature vector matrix.
  • the third representation layer vector is a vector corresponding to the customer retrieval information;
  • a searching unit 413 configured to search the customer database for a first word vector matrix that has a similarity with the third representation layer vector that meets a specified requirement
  • An extraction unit 414 configured to extract the information of the customer corresponding to the first word vector matrix that meets the requirements
  • the anti-recommendation unit 415 is configured to recommend the new product to the customer according to the extracted contact information of the user.
  • the above-mentioned specified contact methods generally include a contact method that can receive messages sent by others without adding friends, such as a mobile phone number and an email address. For example, a short message can be directly sent to a customer through a mobile phone number, and the same is true for an email address.
  • the customer's information is associated with the first word vector in order to find the customer by the word vector.
  • the first word vector matrix and the customer's information are in a one-to-one mapping relationship.
  • the above-mentioned reverse recommendation model obtained based on the LSTM model training refers to the training through the LSTM model.
  • the specific training process is as follows: 1 The historical customer retrieval information and the product information purchased by the customer pass the first A corpus dictionary is vectorized and a corresponding set of word vector matrices is obtained. 2 The set of word vector matrices is divided into a training set and a test set. 3 The sample data of the test set is input into the LSTM model for training, and then the test is performed. The set verifies the trained LSTM model. If the verification passes, the above-mentioned reverse recommendation model is obtained.
  • the above training process belongs to supervised training, that is, historical customer retrieval information and customer purchased product information are information with associated labels.
  • the trained backward recommendation model when the second product vector matrix of product information is input to the backward recommendation After the model, the reverse recommendation model will output a fifth representation layer vector corresponding to the customer retrieval information.
  • the similarity calculation of the third representation layer vector and each first word vector matrix in the customer database is performed, and then the first word whose similarity with the third representation layer vector reaches a specified requirement Vector matrix extraction process.
  • the above specified requirement refers to a first word vector matrix whose similarity with the third representation layer vector reaches a specified threshold, which may be one or more.
  • the above-mentioned specified threshold will be appropriately set to facilitate finding a large number of customer groups. If only one with the highest similarity is found, it obviously does not meet the original intention of finding customers to recommend products.
  • the first word vector matrix extracted above is a word vector matrix with a high degree of similarity to the third representation layer vector output by the reverse recommendation model, so its corresponding customers should have relative Higher purchase desire and corresponding purchasing ability, so it is more efficient to send new products to corresponding customers and recommend resale.
  • the above product recommendation device further includes:
  • the crawling unit 101 is configured to crawl relevant information of all products on a specified website to form product information corresponding to each product;
  • the retaining unit 102 is configured to extract repeated words in each product information, and only one of each repeated word is retained;
  • a vectorization unit 103 is configured to input the remaining words into the DSSM model for calculation, and obtain a word vector matrix of each word to obtain the first corpus dictionary.
  • the remaining words are no duplicates in the product information.
  • the set of words that appear, and each repeated word retains only one subsequent word.
  • the designated website is generally a website that sells the above products, and the types of products sold on this website are similar.
  • the products sold are insurance products, but the types of insurance are different, such as personal accidents Insurance products, children's medical insurance products, major illness insurance products, etc.
  • the website will introduce the products it sells, and the introduction can be used as product information.
  • the remaining words are words that do not appear repeatedly, and each repeated word retains only a set of words after the word.
  • the text in the set is input into the DSSM model, and the word vector of the text is calculated by the DSSM model, and then the text and its corresponding word vector are put together to form the above-mentioned first corpus dictionary.
  • the set of refined words is the words that have appeared in all product introductions on the website, and only one set of words is retained for repeated words, it can not only meet the vectorized use of words, but also It can reduce the difficulty and workload of forming the first corpus dictionary.
  • the foregoing product recommendation device further includes:
  • a judging unit 421, configured to obtain a product purchased by a customer, and judge whether the product is the same as a recommended product
  • the counting unit 422 is configured to add one to the count of different products if the product purchased by the customer is different from the recommended product, and obtain a first count; if the product purchased by the customer is the same as the recommended product, then purchase Add one to the count of the same product to get the second count
  • the comparison unit 423 is configured to use a first count ratio over a second count at a time node of a specified time length, and if the ratio is greater than a preset threshold, disable the first recommendation model.
  • an electronic archive is performed, and the customer's purchase situation can be automatically obtained from the electronic archive, that is, whether the customer has purchased the same product as the recommendation.
  • recommendations are given to customers for purchase recommendations, customers do not necessarily buy products based on recommendations. Record each customer's purchases and count them. When buying a different product from the recommendation, count it once. When buying the same product as the recommendation, count it once to get the first and second counts. Then calculate the ratio of the first count to the second count. If the ratio is large, it means that the recommendation system's recommendation effect is poor. For example, if the ratio is greater than 1, the number of purchases by the uneasy recommendation system is greater than the number of purchases according to the recommendation.
  • the above The recommended model is not ideal, so stop using it.
  • the third count purchased by all customers within a specified period of time may also be recorded, and then the first count or the second count may be compared with the third count, and according to the result of the ratio, it may be determined whether the first count needs to be disabled.
  • the above specified time length refers to a longer period such as a quarter or a month, in order to obtain more data samples, and the obtained ratio is more available.
  • the above-mentioned recommendation unit 40 includes:
  • An output module 41 configured to output a product corresponding to the second representation layer vector
  • a search module 42 configured to search the output introduction information, product sales data, and customer distribution for purchasing the output product according to the output product
  • the recommendation module 43 is configured to form recommendation information for the output product introduction information, sales data, and customer distribution to the customer.
  • the above introduction information is the relevant introduction of the product; the above sales data generally includes the quantity of the product sold; the above customer distribution refers to the age distribution and regional distribution of the customers who purchased the product, that is, through the customer distribution, we can know Customers who purchase this product, such as children's life insurance, etc. Among the customers they purchase, there are more female customers aged 20-30, while there are fewer elderly customers aged 50-70, and fewer male customers. Customers understand what kind of people are buying this insurance product, so that customers can choose whether to buy based on their own. For another example, personal accident insurance has a lot of customers in northern winter. It is related to the cold and easy to fall in the northern cold. Getting the geographical distribution and time distribution is also helpful for new customers to choose whether to buy or not.
  • the introduction information, sales data and customer distribution of the above insurance products are presented to new customers, which facilitates new customers to make objective judgments and reduces the workload of product sales staff.
  • the foregoing recommendation module 43 includes:
  • the encapsulation sub-module 432 is configured to encapsulate the visualization drawing and the introduction information of the output product into a document to form the recommendation information to the customer.
  • the sales data and customer distribution are made into a visual drawing, which is more convenient for new customers to view.
  • the geographical distribution in the customer distribution is reflected on a map, so that customers can see at a glance. Know geographical differences and so on.
  • the sales data can be reflected in the form of a curve, such as the sales trend curve from the beginning of the product's sales to the present, or the monthly sales volume, which greatly improves the speed at which customers can view recommended information.
  • the aforementioned products may include virtual products such as insurance products, financial products, and other physical products.
  • the product recommendation device in the embodiment of the present application vectorizes the customer retrieval information in the historical data and the product information of the product purchased by the customer, and then inputs the LSTM model for training to obtain a first recommendation model, and then uses the first recommendation model. Calculate the products corresponding to the customer's search terms to recommend suitable products to customers, which can effectively recommend products to customers without over recommending popular products.
  • an embodiment of the present application further provides a computer device.
  • the computer device may be a server, and its internal structure may be as shown in FIG. 9.
  • the computer device includes a processor, a memory, a network interface, and a database connected through a system bus.
  • the computer design processor is used to provide computing and control capabilities.
  • the memory of the computer device includes a non-volatile storage medium and an internal memory.
  • the non-volatile storage medium stores an operating system, computer-readable instructions, and a database.
  • the memory provides an environment for operating systems and computer-readable instructions in a non-volatile storage medium.
  • the database of the computer equipment is used to store customer information, a first recommendation model, a corpus dictionary, and the like.
  • the network interface of the computer device is used to communicate with an external terminal through a network connection.
  • the computer-readable instructions are executed by a processor to implement the processes of the embodiments of the methods described above.
  • FIG. 9 is only a block diagram of a part of the structure related to the solution of the present application, and does not constitute a limitation on the computer equipment to which the solution of the present application is applied.
  • the computer device in the embodiment of the present application vectorizes the customer search information in the historical data and the product information of the product they purchased, and then inputs it into the LSTM model for training to obtain a first recommendation model, and then uses the first recommendation model to calculate The product corresponding to the customer's search term to recommend the appropriate product to the customer can effectively recommend the product to the customer without over recommending the popular product.
  • An embodiment of the present application further provides a computer non-volatile readable storage medium, which stores computer-readable instructions.
  • a computer non-volatile readable storage medium which stores computer-readable instructions.

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Business, Economics & Management (AREA)
  • General Physics & Mathematics (AREA)
  • Accounting & Taxation (AREA)
  • Finance (AREA)
  • Data Mining & Analysis (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Artificial Intelligence (AREA)
  • General Engineering & Computer Science (AREA)
  • Evolutionary Computation (AREA)
  • Development Economics (AREA)
  • Software Systems (AREA)
  • Mathematical Physics (AREA)
  • Technology Law (AREA)
  • General Business, Economics & Management (AREA)
  • Health & Medical Sciences (AREA)
  • Economics (AREA)
  • Computing Systems (AREA)
  • Biomedical Technology (AREA)
  • Biophysics (AREA)
  • Computational Linguistics (AREA)
  • Marketing (AREA)
  • Strategic Management (AREA)
  • General Health & Medical Sciences (AREA)
  • Molecular Biology (AREA)
  • Entrepreneurship & Innovation (AREA)
  • Operations Research (AREA)
  • Game Theory and Decision Science (AREA)
  • Human Resources & Organizations (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Evolutionary Biology (AREA)
  • Management, Administration, Business Operations System, And Electronic Commerce (AREA)
  • Information Retrieval, Db Structures And Fs Structures Therefor (AREA)

Abstract

本申请揭示了一种产品推荐方法、装置、计算机设备和存储介质,本申请利用第一推荐模型计算客户的检索词对应的产品,以推荐给客户合适的产品,既可以有效给客户推荐产品,又不会将热门产品过度推荐。

Description

产品推荐方法、装置、计算机设备和存储介质
本申请要求于2018年9月5日提交中国专利局、申请号为 201811032909X,申请名称为“产品推荐方法、装置、计算机设备和存储介质”的中国专利申请的优先权,其全部内容通过引用结合在本申请中。
技术领域
本申请涉及到计算机领域,特别是涉及到产品推荐方法、装置、计算机设备和存储介质。
背景技术
保险、投资理财的时候,会有相关的系统进行统计与计算,以生成推荐信息给客户买哪些保险或理财产品等。现有的推荐系统或是基于内容推荐、或是基于用户推荐,但是当新的保险或理财等产品推出时,不能有效进行推荐,而有些热门产品会被过度推荐。
技术问题
本申请的主要目的为提供一种产品推荐方法、装置、计算机设备和存储介质,旨在有效地将产品推荐给客户。
技术解决方案
本申请提出一种产品推荐方法,包括:
获取客户的第一检索信息,在预设的第一语料字典中查找所述第一检索信息中每一个字对应的词向量,得到对应所述第一检索信息的第一词向量矩阵;
将所述第一词向量矩阵输入到基于LSTM模型训练得到的第一推荐模型中学习,输出对应所述第一词向量矩阵的产品信息的第一表示层向量;
到训练好的对应产品的产品表示层向量矩阵中查找与所述第一表示层向量相似度最高的第二表示层向量;
将所述第二表示层向量对应的产品输出,并推荐给所述客户。
本申请还提供一种产品推荐装置,包括:
获取单元,用于获取客户的第一检索信息,在预设的第一语料字典中查找所述第一检索信息中每一个字对应的词向量,得到对应所述第一检索信息的第一词向量矩阵;
深度学习单元,用于将所述第一词向量矩阵输入到基于LSTM模型训练得到的第一推荐模型中学习,输出对应所述第一词向量矩阵的产品信息的第一表示层向量;
相似度计算单元,用于到训练好的对应产品的产品表示层向量矩阵中查找与所述第一表示层向量相似度最高的第二表示层向量;
推荐单元,用于将所述第二表示层向量对应的产品输出,并推荐给所述客户。
本申请还提供一种计算机设备,包括存储器和处理器,所述存储器存储有计算机可读指令,所述处理器执行所述计算机可读指令时实现上述任一项所述方法的步骤。
本申请还提供一种计算机非易失性可读存储介质,其上存储有计算机可读指令,所述计算机可读指令被处理器执行时实现上述中任一项所述的方法的步骤。
有益效果
本申请的产品推荐方法、装置、计算机设备和存储介质,将历史数据中的客户检索信息和其购买的产品的产品信息向量化,然后输入到LSTM模型中进行训练,以得到第一推荐模型,然后利用第一推荐模型计算客户的检索词对应的产品,以推荐给客户合适的产品,既可以有效给客户推荐产品,又不会将热门产品过度推荐。
附图说明
图1 为本发明一实施例的产品推荐方法的流程示意图;
图2 为本发明一实施例的产品推荐装置的结构示意框图;
图3 为本发明一实施例的产品推荐装置的结构示意框图;
图4 为本发明一实施例的产品推荐装置的结构示意框图;
图5 为本发明一实施例的产品推荐装置的结构示意框图;
图6 为本发明一实施例的产品推荐装置的结构示意框图;
图7 为本发明一实施例的推荐单元的结构示意框图;
图8 为本发明一实施例的推荐模块的结构示意框图;
图9 为本发明一实施例的计算机设备的结构示意框图。
本申请目的的实现、功能特点及优点将结合实施例,参照附图做进一步说明。
本发明的最佳实施方式
为了使本申请的目的、技术方案及优点更加清楚明白,以下结合附图及实施例,对本申请进行进一步详细说明。应当理解,此处描述的具体实施例仅仅用以解释本申请,并不用于限定本申请。
参照图1,本申请提供一种产品推荐方法,包括步骤:
S1、获取客户的第一检索信息,在预设的第一语料字典中查找所述第一检索信息中每一个字对应的词向量,得到对应所述第一检索信息的第一词向量矩阵;
S2、将所述第一词向量矩阵输入到基于LSTM模型训练得到的第一推荐模型中学习,输出对应所述第一词向量矩阵的产品信息的第一表示层向量;
S3、到训练好的对应产品的产品表示层向量矩阵中查找与所述第一表示层向量相似度最高的第二表示层向量;
S4、将所述第二表示层向量对应的产品输出,并推荐给所述客户。
如上述步骤S1所述,上述第一检索信息是指客户输入的检索信息,其包括客户根据自身购买欲望而提炼出的关键词,比如客户想要了解儿童保险,那么其输入的关键词一般包括儿童、意外、医疗等,将这些关键词提取出来之后,在上述第一语料字典中查找每一个字的词向量,然后将每一个字的词向量按照关键词的顺序排列,形成上述的第一词向量矩阵。上述的第一语料字典是售卖上述产品的公司等制作的词向量与文字一一映射的字典,其中包含的文字和词向量是其对应领域中常见的字。第一语料词典的制作过程包括:将待形成词向量的文字输入到DSSM(Deep Structured Semantic Models,深层结构化语义模型)模型中,通过DSSM模型计算出文字的词向量,然后将文字及其对应的词向量放入到第一语料字典中。本步骤中,关键词的提取可以是售卖产品的工作人员根据客户的口述内容,人为提取出来的,然后手动输入计算机设备中,以完成上述步骤S1;也可以是将客户的要求形成文本信息,并将文本信息输入到关键词提取模型中提取关键词,然后将提取出的关键词完成上述步骤S1;还可以是利用语音识别技术将客户的语音转化成文本信息,然后将文本信息输入到关键词提取模型中提取关键词等。
如上述步骤S2所述,上述基于LSTM(Long Short-Term Memory,长短期记忆网络)模型训练得到的第一推荐模型,是指第一推荐模型通过LSTM模型训练而得,具体的训练过程为:①将历史的客户检索信息通过上述第一语料字典形成多个词向量矩阵,以及将客户购买的产品的产品信息通过第一语料字典形成对应的多个词向量矩阵,形成得到的词向量矩阵的集合,其中,每一位客户的检索信息对应的词向量矩阵与该客户购买的产品的产品信息对应的词向量矩阵相互关联;②将词向量矩阵的集合分成训练集和测试集;③将测试集的样本数据输入到LSTM模型中进行训练,然后将测试集对训练后的LSTM模型进行验证,如果验证通过,则得到上述的第一推荐模型。上述训练的过程属于监督训练,即历史的客户检索信息和客户购买的产品信息是带有关联标记的信息,训练出的第一推荐模型,当输入客户检索信息的第一词向量矩阵后,会输出一个关于产品信息对应的第一表示层向量。
如上述步骤S3所述,因为上述第一推荐模型中输出的第一表示层向量是推算出来的,所以需要计算与训练好的对应产品(例如保险产品)的产品表示层向量矩阵中的每一个表示层向量的相似度,寻找相似度最高的第二表示层向量,相似度的计算可以使用欧几里得距离(Eucledian Distance)、曼哈顿距离(Manhattan Distance)、明可夫斯基距离(Minkowski distance)或者余弦相似度中一种算法进行计算。本实施例中,具体的相似度计算公式如下:
Figure dest_path_image002
其中, y Q为检索信息的表示层向量,Y D为产品信息的表示层向量,T D/T Q为对应句子的长度,R 为相似度。在其它的实施例中,无论有没有与第一表示层向量相同的第二表示层向量,都会进行上述的相似度计算,并将计算结果进行由高到低的排名。
如上述步骤S4所述,因为上述第二表示层向量与第一推荐模型输出的第一表示层向量的相似度最高,所以,第二表示层向量对应的产品是最适合上述客户购买的产品,因此将该第二表示层向量对应的产品输出,作为推荐给新客户的产品。在其它实施例中,还可以将与第一表示层向量相似度较高的几个第二表示层向量选取出来,并将各第二表示层向量对应的产品输出,按照其相似度由高到低的顺序,给出不同购买强度的建议。比如,输入的第一检索信息时“儿童、保险、意外”,那么上述第一推荐模型输出的第一表示层向量与预设的产品表示层向量矩阵中的一个第二表示层向量最接近,该第二表示层向量对应的保险产品为儿童意外保险,而与上述“儿童、保险、意外”对应的第一表示层向量相似度略低一点的第二表层向量对应的儿童医疗保险同样会形成推荐的保险产品推出,以提高客户的选择性。
在一个实施例中,上述将所述第二表示层向量对应的产品输出,并推荐给所述客户的步骤S4之后,包括:
S401、记录所述客户购买的产品,以及所述第一检索信息;
S402、将所述客户购买的产品的第一产品信息进行向量化得第一产品向量矩阵;
S403、将所述第一检索信息对应的第一词向量矩阵和用户购买的产品的第一产品信息对应的第一产品向量矩阵关联地保存到指定的数据库中;
S404、当所述数据库中的数据量达到预设的阈值,利用数据库中的全部第一产品向量矩阵和第一词向量矩阵对所述第一推荐模型进行继续训练,得到新的第一推荐模型。
如上述步骤S401至S404所述,虽然给客户推荐了产品,但是客户并不一定会购买推荐的产品,所以记录该客户购买的产品,将客户购买的产品的产品信息向量化得到第一产品向量矩阵,然后将第一产品向量矩阵与第一词向量矩阵(客户的第一检索信息对应的向量矩阵)关联存储到指定的数据库中,以便后期作为训练样本使用。上述指定数据量是指存储上述第一产品向量矩阵与第一词向量矩阵的总的数据量。当数据量达到指定阈值,则说明样本数量达到要求,然后对上述的第一推荐模型进行训练,使用数据库中的样本数据继续对第一推荐模型进行训练,得到新的第一推荐模型,可以提高上述第一推荐模型的准确性。其训练方法与上述训练第一推荐模型的方法相同,在此不在赘述。本实施例中,通过不断地优化第一推荐模型,使得推荐结果更加准确。
在另一个实施例中,上述将所述第二表示层向量对应的产品输出,并推荐给所述客户的步骤S4之后,包括:
S411、将所述客户的信息与所述第一词向量矩阵关联地存储到客户数据库中,其中,客户的信息包括指定的联系方式;
S412、将新的产品信息向量化得到第二产品向量矩阵,并输入到基于LSTM模型训练得到的反向推荐模型中,以输出对应第二产品特征向量矩阵的第三表示层向量,该第三表示层向量是对应客户检索信息的向量;
S413、到所述客户数据库中查找与所述第三表示层向量相似度达到指定要求的第一词向量矩阵;
S414、将查找到的达到要求的第一词向量矩阵对应的客户的信息提取出来;
S415、根据提取出的用户的联系方式将所述新的产品推荐给客户。
如上述步骤S411所述,上述指定的联系方式一般包括手机号码、邮箱等无需添加好友,既可以接收他人发送的消息的联系方式,比如,通过手机号码可以直接发送短信给客户,邮箱同样如此。将客户的信息与第一词向量关联,是为了可以通过词向量反向查找到客户,第一词向量矩阵与客户的信息是一对一的映射关系。
如上述步骤S412所述,上述基于LSTM模型训练得到的反向推荐模型是指通过LSTM模型训练而得,具体的训练过程为:①将历史的客户检索信息和客户购买的产品信息通过上述的第一语料字典进行向量化,并得到对应的词向量矩阵的集合;②将该词向量矩阵的集合分成训练集和测试集;③将测试集的样本数据输入到LSTM模型中进行训练,然后将测试集对训练后的LSTM模型进行验证,如果验证通过,则得到上述的反向推荐模型。上述训练的过程属于监督训练,即历史的客户检索信息和客户购买的产品信息是带有关联标记的信息,训练出的反向推荐模型,当输入产品信息的第二产品向量矩阵到反向推荐模型后,反向推荐模型会输出一个与客户检索信息对应的第五表示层向量。
如上述步骤S413所述,即为将第三表示层向量分别与客户数据库中的各第一词向量矩阵进行像似度计算,然后将与第三表示层向量相似度达到指定要求的第一词向量矩阵提取出来的过程。上述指定要求是指与第三表示层向量相似度达到指定阈值的第一词向量矩阵,其可能是一个,也可能是多个。本实施例中,上述指定阈值会适当设定,以便于查找到大量的客户群体,如只查找一个相似度最高的,显然不符合查找客户以推荐产品的初衷。
如上述步骤S414和S415所述,上述提取出的第一词向量矩阵是与反向推荐模型输出的第三表示层向量相似度较高的词向量矩阵,所以其对应的客户应该具有相对较高的购买欲望,以及相应的购买能力,所以将新的产品发送给对应的客户,推荐转销售的效率更高。
在一个实时中,上述获取客户的第一检索信息,在预设的第一语料字典中查找所述第一检索信息中每一个字对应的词向量,得到对应所述第一检索信息的第一词向量矩阵的步骤S1之前,包括:
S101、抓取指定网站上的全部产品的相关信息,形成各产品对应的产品信息;
S102、将各产品信息中重复出现的字提取出来,每一个重复的字只保留一个;
S103、将剩余的字输入到DSSM模型中计算,得出每一个字的词向量矩阵,以得到所述第一语料字典,所述剩余的字为所述产品信息中没有重复出现的字,以及每一个重复的字只保留一个后的字的集合。
如上述步骤S101所述,上述指定网站一般为售卖上述产品的网站,而该网站上售卖的产品的类型近似,比如,售卖产品均为保险产品,但是保险的种类各不相同,如人身意外险产品、儿童医疗保险产品、重大疾病保险产品等。网站上会对其售卖的产品进行介绍,该介绍既可以作为产品信息。
如上述步骤S102所述,因为产品的类型近似,所以其产品信息中会重复出现相同的字,因此需要将相同的字提取出来,并将各重复出现的字,分别保留一个即可,将其它的删除掉,防止后期对同一个重复的字进行多次的向量化工作。
如上述步骤S103所述,上述剩余的字即为没有重复出现的字,以及每一个重复的字只保留一个后的字的集合。将该集合内的文字输入到DSSM模型中,通过DSSM模型计算出文字的词向量,然后将文字及其对应的词向量放入到一起,即形成上述第一语料字典。本实施例中,因为提炼出的字的集合是网站内所有的产品介绍中出现过的字,并将重复的字只保留一个的字的集合,所以,既可以满足字的向量化使用,也可以降低第一语料字典的形成难度和工作量。
在一个实施例中,上述将所述第二表示层向量对应的产品输出,并推荐给所述客户的步骤S4之后,包括:
S421、获取客户购买的产品,并判断该产品与推荐的产品是否相同;
S422、若不相同,则在购买不同产品的计数的基础上加一,得到第一计数;若相同,则在购买相同产品的计数的基础上加一,得到的第二计数
S423、在指定时间长度的时间节点处,使用第一计数比上第二计数,若比值大于预设阈值,则停用所述第一推荐模型。
如上述步骤S421至S423所述,客户购买产品后会进行电子存档,可以从电子存档中自动获取客户的购买情况,即了解客户是否购买了与推荐相同的产品。也就是说,虽然推荐给客户购买建议,但是客户并不一定会根据推荐来购买产品。记录每一次客户的购买情况,并进行计数,当购买与推荐不同的产品时,计数一次,购买与推荐相同的产品时计数一次,即得到上述的第一计数和第二计数。然后计算第一计数比上第二计数的比值,如果比值较大,说明推荐系统的推荐效果较差,比如,比值大于1,即不安推荐系统购买的次数大于按照推荐购买的次数,显然,上述推荐模型并不理想,所以停止使用。在其它实施例中,还可以记录在指定时间长度内的全部客户购买的第三计数,然后将第一计数或第二计数比上第三计数,根据比值结果,判断是否需要停用上述第一推荐模型。上述的指定时间长度是指一个季度或者一个月等较长时间,以便于得到较多的数据样本,得出的比值的可用性更高。
在一个实施例中,上述将所述第二表示层向量对应的产品输出,并推荐给所述客户的步骤S4,包括:
S41、将所述第二表示层向量对应的产品输出;
S42、根据输出的产品,查找所述输出的介绍信息、产品的售卖数据以及购买所述输出的产品的客户分布;
S43、将所述输出的产品的介绍信息、售卖数据和客户分布形成推荐信息给所述客户。
如上述步骤S41至S43所述,上述的介绍信息即为产品的相关介绍;上述的售卖数据一般包括该产品售卖的数量等;上述客户分布是指购买该产品的客户的年龄分布、区域分布等,即通过该客户分布,可以知道购买该产品的客户的情况,如,儿童人身保险等,其购买的客户中,20-30岁的女性客户较多,而50-70岁的老人客户略少,男性客户更少等,可以使客户了解购买该保险产品的人群是什么样的人群,以便于客户结合自身选择是否购买。又比如,人身意外险,其购买的客户在北方冬天的时候较多,其与北方天寒地冻容易滑倒相关,得到此地域分布和时间分布,也有利于帮助新客户选择是否购买等。将上述的保险产品的介绍信息、售卖数据和客户分布一起呈现给新客户,方便新客户的进行客观的判断,也减少产品的售卖人员的工作量。
在一个实施例中,上述将所述输出的产品的介绍信息、售卖数据和客户分布形成推荐信息给所述客户的步骤S43,包括:
S431、将所述售卖数据和所述客户分布制作成可视化附图;
S432、将所述可视化附图以及所述输出的产品的介绍信息封装到一篇文档中形成所述推荐信息给所述客户。
如上述步骤S431和S432所述,将售卖数据和客户分布制成可视化附图,更方便新客户查看,比如客户分布中的地域分布是体现在一张地图上,使客户可以一目了然地知道地域差异等。而售卖数据可以以曲线的形式体现,如从产品开始售卖开始至今的售卖走势曲线,或者每个月的售卖量等,大大地提高客户查看推荐信息的速度。
本申请实施例中,上述产品可以包括保险产品、金融产品等虚拟的产品,也可以是其它实物产品等。
本申请实施例的产品推荐方法,将历史数据中的客户检索信息和其购买的产品的产品信息向量化,然后输入到LSTM模型中进行训练,以得到第一推荐模型,然后利用第一推荐模型计算客户的检索词对应的产品,以推荐给客户合适的产品,既可以有效给客户推荐产品,又不会将热门产品过度推荐。
参照图2,本申请还提供一种产品推荐装置,包括步骤:
获取单元10,用于获取客户的第一检索信息,在预设的第一语料字典中查找所述第一检索信息中每一个字对应的词向量,得到对应所述第一检索信息的第一词向量矩阵;
深度学习单元20,用于将所述第一词向量矩阵输入到基于LSTM模型训练得到的第一推荐模型中学习,输出对应所述第一词向量矩阵的产品信息的第一表示层向量;
相似度计算单元30,用于到训练好的对应产品的产品表示层向量矩阵中查找与所述第一表示层向量相似度最高的第二表示层向量;
推荐单元40,用于将所述第二表示层向量对应的产品输出,并推荐给所述客户。
在上述获取单元10中,上述第一检索信息是指客户输入的检索信息,其包括客户根据自身购买欲望而提炼出的关键词,比如客户想要了解儿童保险,那么其输入的关键词一般包括儿童、意外、医疗等,将这些关键词提取出来之后,在上述第一语料字典中查找每一个字的词向量,然后将每一个字的词向量按照关键词的顺序排列,形成上述的第一词向量矩阵。上述的第一语料字典是售卖上述产品的公司等制作的词向量与文字一一映射的字典,其中包含的文字和词向量是其对应领域中常见的字。第一语料词典的制作过程包括:将待形成词向量的文字输入到DSSM(Deep Structured Semantic Models,深层结构化语义模型)模型中,通过DSSM模型计算出文字的词向量,然后将文字及其对应的词向量放入到第一语料字典中。本实施例中,关键词的提取可以是售卖产品的工作人员根据客户的口述内容,人为提取出来的,然后手动输入计算机设备中,以完成第一检索信息的向量化;也可以是将客户的要求形成文本信息,并将文本信息输入到关键词提取模型中提取关键词,然后将提取出的关键词完成向量化;还可以是利用语音识别技术将客户的语音转化成文本信息,然后将文本信息输入到关键词提取模型中提取关键词等,以完成第一检索信息的向量化。
在上述深度学习单元20中,上述基于LSTM(Long Short-Term Memory,长短期记忆网络)模型训练得到的第一推荐模型,是指第一推荐模型通过LSTM模型训练而得,具体的训练过程为:①将历史的客户检索信息通过上述第一语料字典形成多个词向量矩阵,以及将客户购买的产品的产品信息通过第一语料字典形成对应的多个词向量矩阵,形成得到的词向量矩阵的集合,其中,每一位客户的检索信息对应的词向量矩阵与该客户购买的产品的产品信息对应的词向量矩阵相互关联;②将词向量矩阵的集合分成训练集和测试集;③将测试集的样本数据输入到LSTM模型中进行训练,然后将测试集对训练后的LSTM模型进行验证,如果验证通过,则得到上述的第一推荐模型。上述训练的过程属于监督训练,即历史的客户检索信息和客户购买的产品信息是带有关联标记的信息,训练出的第一推荐模型,当输入客户检索信息的第一词向量矩阵后,会输出一个关于产品信息对应的第一表示层向量。
在上述相似度计算单元30中,因为上述第一推荐模型中输出的第一表示层向量是推算出来的,所以需要计算与训练好的对应产品(例如保险产品)的产品表示层向量矩阵中的每一个表示层向量的相似度,寻找相似度最高的第二表示层向量,相似度的计算可以使用欧几里得距离(Eucledian Distance)、曼哈顿距离(Manhattan Distance)、明可夫斯基距离(Minkowski distance)或者余弦相似度中一种算法进行计算。本实施例中,具体的相似度计算公式如下:
Figure 850750dest_path_image002
其中, y Q为检索信息的表示层向量,Y D为产品信息的表示层向量,T D/T Q为对应句子的长度,R 为相似度。在其它的实施例中,无论有没有与第一表示层向量相同的第二表示层向量,都会进行上述的相似度计算,并将计算结果进行由高到低的排名。
在上述推荐单元40中,因为上述第二表示层向量与第一推荐模型输出的第一表示层向量的相似度最高,所以,第二表示层向量对应的产品是最适合上述客户购买的产品,因此将该第二表示层向量对应的产品输出,作为推荐给新客户的产品。在其它实施例中,还可以将与第一表示层向量相似度较高的几个第二表示层向量选取出来,并将各第二表示层向量对应的产品输出,按照其相似度由高到低的顺序,给出不同购买强度的建议。比如,输入的第一检索信息时“儿童、保险、意外”,那么上述第一推荐模型输出的第一表示层向量与预设的产品表示层向量矩阵中的一个第二表示层向量最接近,该第二表示层向量对应的保险产品为儿童意外保险,而与上述“儿童、保险、意外”对应的第一表示层向量相似度略低一点的第二表层向量对应的儿童医疗保险同样会形成推荐的保险产品推出,以提高客户的选择性。
参照图3,在一个实施例中,上述产品推荐装置还包括:
记录单元401,用于记录所述客户购买的产品,以及所述第一检索信息;
向量化单元402,用于将所述客户购买的产品的第一产品信息进行向量化得第一产品向量矩阵;
关联存储单元403,用于将所述第一检索信息对应的第一词向量矩阵和用户购买的产品的第一产品信息对应的第一产品向量矩阵关联地保存到指定的数据库中;
训练单元404,用于当所述数据库中的数据量达到预设的阈值,利用数据库中的全部第一产品向量矩阵和第一词向量矩阵对所述第一推荐模型进行继续训练,得到新的第一推荐模型。
在上述记录单元401、向量化单元402、关联存储单元403和训练单元404中,上述产品推荐装置虽然给客户推荐了产品,但是客户并不一定会购买推荐的产品,所以记录该客户购买的产品,将客户购买的产品的产品信息向量化得到第一产品向量矩阵,然后将第一产品向量矩阵与第一词向量矩阵(客户的第一检索信息对应的向量矩阵)关联存储到指定的数据库中,以便后期作为训练样本使用。上述指定数据量是指存储上述第一产品向量矩阵与第一词向量矩阵的总的数据量。当数据量达到指定阈值,则说明样本数量达到要求,然后对上述的第一推荐模型进行训练,使用数据库中的样本数据继续对第一推荐模型进行训练,得到新的第一推荐模型,可以提高上述第一推荐模型的准确性。其训练方法与上述训练第一推荐模型的方法相同,在此不在赘述。本实施例中,通过不断地优化第一推荐模型,使得推荐结果更加准确。
参照图4,在另一个实施例中,上述产品推荐装置还包括:
关联单元411,用于将所述客户的信息与所述第一词向量矩阵关联地存储到客户数据库中,其中,客户的信息包括指定的联系方式;
计算单元412,用于将新的产品信息向量化得到第二产品向量矩阵,并输入到基于LSTM模型训练得到的反向推荐模型中,以输出对应第二产品特征向量矩阵的第三表示层向量,该第三表示层向量是对应客户检索信息的向量;
查找单元413,用于到所述客户数据库中查找与所述第三表示层向量相似度达到指定要求的第一词向量矩阵;
提取单元414,用于将查找到的达到要求的第一词向量矩阵对应的客户的信息提取出来;
反推荐单元415,用于根据提取出的用户的联系方式将所述新的产品推荐给客户。
在上述关联单元411中,上述指定的联系方式一般包括手机号码、邮箱等无需添加好友,既可以接收他人发送的消息的联系方式,比如,通过手机号码可以直接发送短信给客户,邮箱同样如此。将客户的信息与第一词向量关联,是为了可以通过词向量反向查找到客户,第一词向量矩阵与客户的信息是一对一的映射关系。
在上述计算单元412中,上述基于LSTM模型训练得到的反向推荐模型是指通过LSTM模型训练而得,具体的训练过程为:①将历史的客户检索信息和客户购买的产品信息通过上述的第一语料字典进行向量化,并得到对应的词向量矩阵的集合;②将该词向量矩阵的集合分成训练集和测试集;③将测试集的样本数据输入到LSTM模型中进行训练,然后将测试集对训练后的LSTM模型进行验证,如果验证通过,则得到上述的反向推荐模型。上述训练的过程属于监督训练,即历史的客户检索信息和客户购买的产品信息是带有关联标记的信息,训练出的反向推荐模型,当输入产品信息的第二产品向量矩阵到反向推荐模型后,反向推荐模型会输出一个与客户检索信息对应的第五表示层向量。
在上述查找单元413中,即为将第三表示层向量分别与客户数据库中的各第一词向量矩阵进行像似度计算,然后将与第三表示层向量相似度达到指定要求的第一词向量矩阵提取出来的过程。上述指定要求是指与第三表示层向量相似度达到指定阈值的第一词向量矩阵,其可能是一个,也可能是多个。本实施例中,上述指定阈值会适当设定,以便于查找到大量的客户群体,如只查找一个相似度最高的,显然不符合查找客户以推荐产品的初衷。
在上述提取单元414和反推荐单元中,上述提取出的第一词向量矩阵是与反向推荐模型输出的第三表示层向量相似度较高的词向量矩阵,所以其对应的客户应该具有相对较高的购买欲望,以及相应的购买能力,所以将新的产品发送给对应的客户,推荐转销售的效率更高。
参照图5,在一个实时中,上述产品推荐装置还包括:
抓取单元101,用于抓取指定网站上的全部产品的相关信息,形成各产品对应的产品信息;
保留单元102,用于将各产品信息中重复出现的字提取出来,每一个重复的字只保留一个;
向量化单元103,用于将剩余的字输入到DSSM模型中计算,得出每一个字的词向量矩阵,以得到所述第一语料字典,所述剩余的字为所述产品信息中没有重复出现的字,以及每一个重复的字只保留一个后的字的集合。
在上述抓取单元101中,上述指定网站一般为售卖上述产品的网站,而该网站上售卖的产品的类型近似,比如,售卖产品均为保险产品,但是保险的种类各不相同,如人身意外险产品、儿童医疗保险产品、重大疾病保险产品等。网站上会对其售卖的产品进行介绍,该介绍既可以作为产品信息。
在上述保留单元102中,因为产品的类型近似,所以其产品信息中会重复出现相同的字,因此需要将相同的字提取出来,并将各重复出现的字,分别保留一个即可,将其它的删除掉,防止后期对同一个重复的字进行多次的向量化工作。
在上述向量化单元103中,上述剩余的字即为没有重复出现的字,以及每一个重复的字只保留一个后的字的集合。将该集合内的文字输入到DSSM模型中,通过DSSM模型计算出文字的词向量,然后将文字及其对应的词向量放入到一起,即形成上述第一语料字典。本实施例中,因为提炼出的字的集合是网站内所有的产品介绍中出现过的字,并将重复的字只保留一个的字的集合,所以,既可以满足字的向量化使用,也可以降低第一语料字典的形成难度和工作量。
参照图6,在一个实施例中,上述产品推荐装置还包括:
判断单元421,用于获取客户购买的产品,并判断该产品与推荐的产品是否相同;
计数单元422,用于若客户购买的产品与推荐的产品不相同,则在购买不同产品的计数的基础上加一,得到第一计数;若客户购买的产品与推荐的产品相同,则在购买相同产品的计数的基础上加一,得到的第二计数
比较单元423,用于在指定时间长度的时间节点处,使用第一计数比上第二计数,若比值大于预设阈值,则停用所述第一推荐模型。
本实施例中,客户购买产品后会进行电子存档,可以从电子存档中自动获取客户的购买情况,即了解客户是否购买了与推荐相同的产品。也就是说,虽然推荐给客户购买建议,但是客户并不一定会根据推荐来购买产品。记录每一次客户的购买情况,并进行计数,当购买与推荐不同的产品时,计数一次,购买与推荐相同的产品时计数一次,即得到上述的第一计数和第二计数。然后计算第一计数比上第二计数的比值,如果比值较大,说明推荐系统的推荐效果较差,比如,比值大于1,即不安推荐系统购买的次数大于按照推荐购买的次数,显然,上述推荐模型并不理想,所以停止使用。在其它实施例中,还可以记录在指定时间长度内的全部客户购买的第三计数,然后将第一计数或第二计数比上第三计数,根据比值结果,判断是否需要停用上述第一推荐模型。上述的指定时间长度是指一个季度或者一个月等较长时间,以便于得到较多的数据样本,得出的比值的可用性更高。
参照图7,在一个实施例中,上述推荐单元40,包括:
输出模块41,用于将所述第二表示层向量对应的产品输出;
查找模块42,用于根据输出的产品,查找所述输出的介绍信息、产品的售卖数据以及购买所述输出的产品的客户分布;
推荐模块43,用于将所述输出的产品的介绍信息、售卖数据和客户分布形成推荐信息给所述客户。
上述的介绍信息即为产品的相关介绍;上述的售卖数据一般包括该产品售卖的数量等;上述客户分布是指购买该产品的客户的年龄分布、区域分布等,即通过该客户分布,可以知道购买该产品的客户的情况,如,儿童人身保险等,其购买的客户中,20-30岁的女性客户较多,而50-70岁的老人客户略少,男性客户更少等,可以使客户了解购买该保险产品的人群是什么样的人群,以便于客户结合自身选择是否购买。又比如,人身意外险,其购买的客户在北方冬天的时候较多,其与北方天寒地冻容易滑倒相关,得到此地域分布和时间分布,也有利于帮助新客户选择是否购买等。将上述的保险产品的介绍信息、售卖数据和客户分布一起呈现给新客户,方便新客户的进行客观的判断,也减少产品的售卖人员的工作量。
参照图8,在一个实施例中,上述推荐模块43,包括:
可视化子模块431,用于将所述售卖数据和所述客户分布制作成可视化附图;
封装子模块432,用于将所述可视化附图以及所述输出的产品的介绍信息封装到一篇文档中形成所述推荐信息给所述客户。
在上述可视化子模块431和封装子模块432中,将售卖数据和客户分布制成可视化附图,更方便新客户查看,比如客户分布中的地域分布是体现在一张地图上,使客户可以一目了然地知道地域差异等。而售卖数据可以以曲线的形式体现,如从产品开始售卖开始至今的售卖走势曲线,或者每个月的售卖量等,大大地提高客户查看推荐信息的速度。
本申请实施例中,上述产品可以包括保险产品、金融产品等虚拟的产品,也可以是其它实物产品等。
本申请实施例的产品推荐装置,将历史数据中的客户检索信息和其购买的产品的产品信息向量化,然后输入到LSTM模型中进行训练,以得到第一推荐模型,然后利用第一推荐模型计算客户的检索词对应的产品,以推荐给客户合适的产品,既可以有效给客户推荐产品,又不会将热门产品过度推荐。
参照图9,本申请实施例中还提供一种计算机设备,该计算机设备可以是服务器,其内部结构可以如图9所示。该计算机设备包括通过系统总线连接的处理器、存储器、网络接口和数据库。其中,该计算机设计的处理器用于提供计算和控制能力。该计算机设备的存储器包括非易失性存储介质、内存储器。该非易失性存储介质存储有操作系统、计算机可读指令和数据库。该内存器为非易失性存储介质中的操作系统和计算机可读指令的运行提供环境。该计算机设备的数据库用于存储客户信息、第一推荐模型、语料字典等。该计算机设备的网络接口用于与外部的终端通过网络连接通信。该计算机可读指令被处理器执行时以实现如上述各方法的实施例的流程。
本领域技术人员可以理解,图9中示出的结构,仅仅是与本申请方案相关的部分结构的框图,并不构成对本申请方案所应用于其上的计算机设备的限定。
本申请实施例的计算机设备,将历史数据中的客户检索信息和其购买的产品的产品信息向量化,然后输入到LSTM模型中进行训练,以得到第一推荐模型,然后利用第一推荐模型计算客户的检索词对应的产品,以推荐给客户合适的产品,既可以有效给客户推荐产品,又不会将热门产品过度推荐。
本申请一实施例还提供一种计算机非易失性可读存储介质,其上存储有计算机可读指令,计算机可读指令被处理器执行时实现如上述各方法的实施例的流程。
以上所述仅为本申请的优选实施例,并非因此限制本申请的专利范围,凡是利用本申请说明书及附图内容所作的等效结构或等效流程变换,或直接或间接运用在其他相关的技术领域,均同理包括在本申请的专利保护范围内。

Claims (20)

  1. 一种产品推荐方法,其特征在于,包括:
    获取客户的第一检索信息,在预设的第一语料字典中查找所述第一检索信息中每一个字对应的词向量,得到对应所述第一检索信息的第一词向量矩阵;
    将所述第一词向量矩阵输入到基于LSTM模型训练得到的第一推荐模型中学习,输出对应所述第一词向量矩阵的产品信息的第一表示层向量;
    到训练好的对应产品的产品表示层向量矩阵中查找与所述第一表示层向量相似度最高的第二表示层向量;
    将所述第二表示层向量对应的产品输出,并推荐给所述客户。
  2. 根据权利要求1所述的产品推荐方法,其特征在于,所述将所述第二表示层向量对应的产品输出,并推荐给所述客户的步骤之后,包括:
    记录所述客户购买的产品,以及所述第一检索信息;
    将所述客户购买的产品的第一产品信息进行向量化得第一产品向量矩阵;
    将所述第一检索信息对应的第一词向量矩阵和用户购买的产品的第一产品信息对应的第一产品向量矩阵关联地保存到指定的数据库中;
    当所述数据库中的数据量达到预设的阈值,利用数据库中的全部第一产品向量矩阵和第一词向量矩阵对所述第一推荐模型进行继续训练,得到新的第一推荐模型。
  3. 根据权利要求1所述的产品推荐方法,其特征在于,所述将所述第二表示层向量对应的产品输出,并推荐给所述客户的步骤之后,包括:
    将所述客户的信息与所述第一词向量矩阵关联地存储到客户数据库中,其中,客户的信息包括指定的联系方式;
    将新的产品信息向量化得到第二产品向量矩阵,并输入到基于LSTM模型训练得到的反向推荐模型中,以输出对应第二产品特征向量矩阵的第三表示层向量,该第三表示层向量是对应客户检索信息的向量;
    到所述客户数据库中查找与所述第三表示层向量相似度达到指定要求的第一词向量矩阵;
    将查找到的达到要求的第一词向量矩阵对应的客户的信息提取出来;
    根据提取出的用户的联系方式将所述新的产品推荐给客户。
  4. 根据权利要求1所述的产品推荐方法,其特征在于,所述获取客户的第一检索信息,在预设的第一语料字典中查找所述第一检索信息中每一个字对应的词向量,得到对应所述第一检索信息的第一词向量矩阵的步骤之前,包括:
    抓取指定网站上的全部产品的相关信息,形成各产品对应的产品信息;
    将各产品信息中重复出现的字提取出来,每一个重复的字只保留一个;
    将剩余的字输入到DSSM模型中计算,得出每一个字的词向量矩阵,以得到所述第一语料字典,所述剩余的字为所述产品信息中没有重复出现的字,以及每一个重复的字只保留一个后的字的集合。
  5. 根据权利要求1所述的产品推荐方法,其特征在于,所述将所述第二表示层向量对应的产品输出,并推荐给所述客户的步骤之后,包括:
    获取客户购买的产品,并判断该产品与推荐的产品是否相同;
    若不相同,则在购买不同产品的计数的基础上加一,得到第一计数;若相同,则在购买相同产品的计数的基础上加一,得到的第二计数;
    在指定时间长度的时间节点处,使用第一计数比上第二计数,若比值大于预设阈值,则停用所述第一推荐模型。
  6. 根据权利要求1所述的产品推荐方法,其特征在于,所述将所述第二表示层向量对应的产品输出,并推荐给所述客户的步骤,包括:
    将所述第二表示层向量对应的产品输出;
    根据输出的产品,查找所述输出的介绍信息、产品的售卖数据以及购买所述输出的产品的客户分布;
    将所述输出的产品的介绍信息、售卖数据和客户分布形成推荐信息给所述客户。
  7. 根据权利要求6所述的产品推荐方法,其特征在于,所述将所述输出的产品的介绍信息、售卖数据和客户分布形成推荐信息给所述客户的步骤,包括:
    将所述售卖数据和所述客户分布制作成可视化附图;
    将所述可视化附图以及所述输出的产品的介绍信息封装到一篇文档中形成所述推荐信息给所述客户。
  8. 一种产品推荐装置,其特征在于,包括:
    获取单元,用于获取客户的第一检索信息,在预设的第一语料字典中查找所述第一检索信息中每一个字对应的词向量,得到对应所述第一检索信息的第一词向量矩阵;
    深度学习单元,用于将第一词向量矩阵输入到基于LSTM模型训练得到的第一推荐模型中学习,输出对应所述第一词向量矩阵的产品信息的第一表示层向量;
    相似度计算单元,用于到训练好的对应产品的产品表示层向量矩阵中查找与所述第一表示层向量相似度最高的第二表示层向量;
    推荐单元,用于将所述第二表示层向量对应的产品输出,并推荐给所述客户。
  9. 根据权利要求8所述的产品推荐装置,其特征在于,还包括:
    记录单元,用于记录所述客户购买的产品,以及所述第一检索信息;
    向量化单元,用于将所述客户购买的产品的第一产品信息进行向量化得第一产品向量矩阵;
    关联存储单元,用于将所述第一检索信息对应的第一词向量矩阵和用户购买的产品的第一产品信息对应的第一产品向量矩阵关联地保存到指定的数据库中;
    训练单元,用于当所述数据库中的数据量达到预设的阈值,利用数据库中的全部第一产品向量矩阵和第一词向量矩阵对所述第一推荐模型进行继续训练,得到新的第一推荐模型。
  10. 根据权利要求8所述的产品推荐装置,其特征在于,还包括:
    关联单元,用于将所述客户的信息与所述第一词向量矩阵关联地存储到客户数据库中,其中,客户的信息包括指定的联系方式;
    计算单元,用于将新的产品信息向量化得到第二产品向量矩阵,并输入到基于LSTM模型训练得到的反向推荐模型中,以输出对应第二产品特征向量矩阵的第三表示层向量,该第三表示层向量是对应客户检索信息的向量;
    查找单元,用于到所述客户数据库中查找与所述第三表示层向量相似度达到指定要求的第一词向量矩阵;
    提取单元,用于将查找到的达到要求的第一词向量矩阵对应的客户的信息提取出来;
    反推荐单元,用于根据提取出的用户的联系方式将所述新的产品推荐给客户。
  11. 根据权利要求8所述的产品推荐装置,其特征在于,还包括:
    抓取单元,用于抓取指定网站上的全部产品的相关信息,形成各产品对应的产品信息;
    保留单元,用于将各产品信息中重复出现的字提取出来,每一个重复的字只保留一个;
    向量化单元,用于将剩余的字输入到DSSM模型中计算,得出每一个字的词向量矩阵,以得到所述第一语料字典,所述剩余的字为所述产品信息中没有重复出现的字,以及每一个重复的字只保留一个后的字的集合。
  12. 根据权利要求8所述的产品推荐装置,其特征在于,还包括:
    判断单元,用于获取客户购买的产品,并判断该产品与推荐的产品是否相同;
    计数单元,用于若客户购买的产品与推荐的产品不相同,则在购买不同产品的计数的基础上加一,得到第一计数;若客户购买的产品与推荐的产品相同,则在购买相同产品的计数的基础上加一,得到的第二计数
    比较单元,用于在指定时间长度的时间节点处,使用第一计数比上第二计数,若比值大于预设阈值,则停用所述第一推荐模型。
  13. 根据权利要求8所述的产品推荐装置,其特征在于,所述推荐单元,包括:
    输出模块,用于将所述第二表示层向量对应的产品输出;
    查找模块,用于根据输出的产品,查找所述输出的介绍信息、产品的售卖数据以及购买所述输出的产品的客户分布;
    推荐模块,用于将所述输出的产品的介绍信息、售卖数据和客户分布形成推荐信息给所述客户。
  14. 根据权利要求13所述的产品推荐装置,其特征在于,所述推荐模块,包括:
    可视化子模块,用于将所述售卖数据和所述客户分布制作成可视化附图;
    封装子模块,用于将所述可视化附图以及所述输出的产品的介绍信息封装到一篇文档中形成所述推荐信息给所述客户。
  15. 一种计算机设备,包括存储器和处理器,所述存储器存储有计算机可读指令,其特征在于,所述处理器执行所述计算机可读指令时实现产品推荐方法,该产品推荐方法包括:
    获取客户的第一检索信息,在预设的第一语料字典中查找所述第一检索信息中每一个字对应的词向量,得到对应所述第一检索信息的第一词向量矩阵;
    将所述第一词向量矩阵输入到基于LSTM模型训练得到的第一推荐模型中学习,输出对应所述第一词向量矩阵的产品信息的第一表示层向量;
    到训练好的对应产品的产品表示层向量矩阵中查找与所述第一表示层向量相似度最高的第二表示层向量;
    将所述第二表示层向量对应的产品输出,并推荐给所述客户。
  16. 根据权利要求15所述的计算机设备,其特征在于,所述将所述第二表示层向量对应的产品输出,并推荐给所述客户的步骤之后,包括:
    记录所述客户购买的产品,以及所述第一检索信息;
    将所述客户购买的产品的第一产品信息进行向量化得第一产品向量矩阵;
    将所述第一检索信息对应的第一词向量矩阵和用户购买的产品的第一产品信息对应的第一产品向量矩阵关联地保存到指定的数据库中;
    当所述数据库中的数据量达到预设的阈值,利用数据库中的全部第一产品向量矩阵和第一词向量矩阵对所述第一推荐模型进行继续训练,得到新的第一推荐模型。
  17. 根据权利要求15所述的计算机设备,其特征在于,所述将所述第二表示层向量对应的产品输出,并推荐给所述客户的步骤之后,包括:
    将所述客户的信息与所述第一词向量矩阵关联地存储到客户数据库中,其中,客户的信息包括指定的联系方式;
    将新的产品信息向量化得到第二产品向量矩阵,并输入到基于LSTM模型训练得到的反向推荐模型中,以输出对应第二产品特征向量矩阵的第三表示层向量,该第三表示层向量是对应客户检索信息的向量;
    到所述客户数据库中查找与所述第三表示层向量相似度达到指定要求的第一词向量矩阵;
    将查找到的达到要求的第一词向量矩阵对应的客户的信息提取出来;
    根据提取出的用户的联系方式将所述新的产品推荐给客户。
  18. 根据权利要求15所述的计算机设备,其特征在于,所述获取客户的第一检索信息,在预设的第一语料字典中查找所述第一检索信息中每一个字对应的词向量,得到对应所述第一检索信息的第一词向量矩阵的步骤之前,包括:
    抓取指定网站上的全部产品的相关信息,形成各产品对应的产品信息;
    将各产品信息中重复出现的字提取出来,每一个重复的字只保留一个;
    将剩余的字输入到DSSM模型中计算,得出每一个字的词向量矩阵,以得到所述第一语料字典,所述剩余的字为所述产品信息中没有重复出现的字,以及每一个重复的字只保留一个后的字的集合。
  19. 一种计算机非易失性可读存储介质,其上存储有计算机可读指令,其特征在于,所述计算机可读指令被处理器执行时实现产品推荐方法,该产品推荐方法包括:
    获取客户的第一检索信息,在预设的第一语料字典中查找所述第一检索信息中每一个字对应的词向量,得到对应所述第一检索信息的第一词向量矩阵;
    将所述第一词向量矩阵输入到基于LSTM模型训练得到的第一推荐模型中学习,输出对应所述第一词向量矩阵的产品信息的第一表示层向量;
    到训练好的对应产品的产品表示层向量矩阵中查找与所述第一表示层向量相似度最高的第二表示层向量;
    将所述第二表示层向量对应的产品输出,并推荐给所述客户。
  20. 根据权利要求19所述的计算机非易失性可读存储介质,其特征在于,所述将所述第二表示层向量对应的产品输出,并推荐给所述客户的步骤之后,包括:
    记录所述客户购买的产品,以及所述第一检索信息;
    将所述客户购买的产品的第一产品信息进行向量化得第一产品向量矩阵;
    将所述第一检索信息对应的第一词向量矩阵和用户购买的产品的第一产品信息对应的第一产品向量矩阵关联地保存到指定的数据库中;
    当所述数据库中的数据量达到预设的阈值,利用数据库中的全部第一产品向量矩阵和第一词向量矩阵对所述第一推荐模型进行继续训练,得到新的第一推荐模型。
PCT/CN2018/124391 2018-09-05 2018-12-27 产品推荐方法、装置、计算机设备和存储介质 Ceased WO2020048061A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN201811032909.XA CN109447813A (zh) 2018-09-05 2018-09-05 产品推荐方法、装置、计算机设备和存储介质
CN201811032909.X 2018-09-05

Publications (1)

Publication Number Publication Date
WO2020048061A1 true WO2020048061A1 (zh) 2020-03-12

Family

ID=65530931

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2018/124391 Ceased WO2020048061A1 (zh) 2018-09-05 2018-12-27 产品推荐方法、装置、计算机设备和存储介质

Country Status (2)

Country Link
CN (1) CN109447813A (zh)
WO (1) WO2020048061A1 (zh)

Cited By (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111461827A (zh) * 2020-03-31 2020-07-28 中国银行股份有限公司 产品评价信息的推送方法和装置
CN112163963A (zh) * 2020-09-27 2021-01-01 中国平安财产保险股份有限公司 业务推荐方法、装置、计算机设备和存储介质
CN119904286A (zh) * 2024-12-10 2025-04-29 广东欣悦隆科技有限公司 一种电商产品精准推荐方法、装置、设备、介质及产品

Families Citing this family (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112749296B (zh) * 2019-10-31 2024-01-26 北京达佳互联信息技术有限公司 一种视频推荐方法、装置、服务器及存储介质
CN111932296B (zh) * 2020-07-20 2024-05-28 中国建设银行股份有限公司 一种产品推荐方法及装置、服务器、存储介质
CN113032589B (zh) * 2021-03-29 2024-01-23 北京奇艺世纪科技有限公司 多媒体文件推荐方法、装置、电子设备及可读存储介质
CN115034906A (zh) * 2022-04-20 2022-09-09 苏州圈点互动科技有限公司 一种金融产品配置方案记忆与推荐方法、系统
CN115936801A (zh) * 2022-11-10 2023-04-07 中国平安财产保险股份有限公司 基于神经网络的产品推荐方法、装置、设备和存储介质
CN116522003B (zh) 2023-07-03 2023-09-12 之江实验室 基于嵌入表压缩的信息推荐方法、装置、设备和介质

Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN104834747A (zh) * 2015-05-25 2015-08-12 中国科学院自动化研究所 基于卷积神经网络的短文本分类方法
CN107301246A (zh) * 2017-07-14 2017-10-27 河北工业大学 基于超深卷积神经网络结构模型的中文文本分类方法
CN107329949A (zh) * 2017-05-24 2017-11-07 北京捷通华声科技股份有限公司 一种语义匹配方法和系统
CN108334889A (zh) * 2017-11-30 2018-07-27 腾讯科技(深圳)有限公司 摘要描述生成方法和装置、摘要描述模型训练方法和装置
CN108334638A (zh) * 2018-03-20 2018-07-27 桂林电子科技大学 基于长短期记忆神经网络与兴趣迁移的项目评分预测方法

Family Cites Families (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN105930413A (zh) * 2016-04-18 2016-09-07 北京百度网讯科技有限公司 相似度模型参数的训练方法、搜索处理方法及对应装置
CN106874522A (zh) * 2017-03-29 2017-06-20 珠海习悦信息技术有限公司 信息推荐方法、装置、存储介质及处理器
CN108182621A (zh) * 2017-12-07 2018-06-19 合肥美的智能科技有限公司 商品推荐方法及商品推荐装置、设备和存储介质

Patent Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN104834747A (zh) * 2015-05-25 2015-08-12 中国科学院自动化研究所 基于卷积神经网络的短文本分类方法
CN107329949A (zh) * 2017-05-24 2017-11-07 北京捷通华声科技股份有限公司 一种语义匹配方法和系统
CN107301246A (zh) * 2017-07-14 2017-10-27 河北工业大学 基于超深卷积神经网络结构模型的中文文本分类方法
CN108334889A (zh) * 2017-11-30 2018-07-27 腾讯科技(深圳)有限公司 摘要描述生成方法和装置、摘要描述模型训练方法和装置
CN108334638A (zh) * 2018-03-20 2018-07-27 桂林电子科技大学 基于长短期记忆神经网络与兴趣迁移的项目评分预测方法

Cited By (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111461827A (zh) * 2020-03-31 2020-07-28 中国银行股份有限公司 产品评价信息的推送方法和装置
CN111461827B (zh) * 2020-03-31 2024-04-09 中国银行股份有限公司 产品评价信息的推送方法和装置
CN112163963A (zh) * 2020-09-27 2021-01-01 中国平安财产保险股份有限公司 业务推荐方法、装置、计算机设备和存储介质
CN112163963B (zh) * 2020-09-27 2024-04-02 中国平安财产保险股份有限公司 业务推荐方法、装置、计算机设备和存储介质
CN119904286A (zh) * 2024-12-10 2025-04-29 广东欣悦隆科技有限公司 一种电商产品精准推荐方法、装置、设备、介质及产品

Also Published As

Publication number Publication date
CN109447813A (zh) 2019-03-08

Similar Documents

Publication Publication Date Title
WO2020048061A1 (zh) 产品推荐方法、装置、计算机设备和存储介质
Athota et al. Chatbot for healthcare system using artificial intelligence
KR102564144B1 (ko) 텍스트 관련도를 확정하기 위한 방법, 장치, 기기 및 매체
CN114330354B (zh) 一种基于词汇增强的事件抽取方法、装置及存储介质
CN110807091B (zh) 一种酒店智能问答推荐与决策支持分析方法及系统
WO2020048062A1 (zh) 产品销售的智能推荐方法、装置、计算机设备和存储介质
CN111858940B (zh) 一种基于多头注意力的法律案例相似度计算方法及系统
WO2021139247A1 (zh) 医学领域知识图谱的构建方法、装置、设备及存储介质
WO2021120588A1 (zh) 语料生成方法、装置、计算机设备及存储介质
CN111680154B (zh) 一种基于深度学习的评论文本属性级情感分析方法
WO2025025953A1 (zh) 问题答案推荐方法、存储介质和电子设备
WO2022088671A1 (zh) 自动问答方法、装置、设备及存储介质
CN112036923A (zh) 一种服务评价方法、系统、装置及存储介质
CN108804677A (zh) 结合多层级注意力机制的深度学习问题分类方法及系统
CN113409907A (zh) 一种基于互联网医院的智能预问诊方法及系统
CN110738059B (zh) 一种文本相似度计算方法及系统
CN105426514A (zh) 个性化的移动应用app推荐方法
CN117931858B (zh) 数据查询方法、装置、计算机设备和存储介质
CN110929498A (zh) 一种短文本相似度的计算方法及装置、可读存储介质
JP2025114640A (ja) ヌル値又は同等の値を有するリレーショナル・テーブルのためのベクトル埋込モデル
CN113392312A (zh) 信息处理方法、系统及电子设备
US20250225008A1 (en) Efficient generation of application programming interface calls using language models, data types, and enriched schema
CN111368081A (zh) 一种精选文本内容确定方法和系统
CN111444338A (zh) 文本处理、装置、存储介质及设备
CN114385852A (zh) 音频文本检索模型训练方法、系统、电子设备和存储介质

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 18932477

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 18932477

Country of ref document: EP

Kind code of ref document: A1