TWI812883B - Method and server for recommending products on mobile payment platform - Google Patents

Abstract

The invention provides a method and a server for recommending products on a mobile payment platform. The method includes: obtaining specific health data of a specific user and specific environmental data associated with the specific user; obtaining multiple products provided by a mobile payment platform; estimating a product recommendation score of each product based on specific health data and specific environmental data; selecting multiple specific products from the products based on the product recommendation score of each product, and recommending the specific products to the specific user.

Description

Methods and servers for recommending products on mobile payment platforms

The present invention relates to a method and a server for recommending products, and in particular, to a method and a server for recommending products on a mobile payment platform.

Currently, e-commerce companies often use "user purchasing habits" or "location information" behind shopping malls to recommend ads or products to users. In recent years, people have focused on health management and the cultivation of exercise habits. Smart watches and bracelets (wristbands) account for about 80% of the overall wearable device market. The smart bracelet can record relevant body data such as heart rate, sleep depth, heart rate and blood pressure, and can upload the information to the cloud. Moreover, points from mobile payment applications can also be used to redeem goods. However, when the interface of the application is filled with various products, it is generally difficult for users to choose among the dazzling products.

Therefore, for those skilled in the art, how to design a mechanism that can recommend suitable products to users is actually an important issue.

In view of this, the present invention provides a method and server for recommending products on a mobile payment platform, which can be used to solve the above technical problems.

The present invention provides a method for recommending products on a mobile payment platform, including: when a specific application corresponding to the mobile payment platform is launched by a specific user, obtaining at least one specific health data of the specific user and associated with a specific use At least one specific environmental data of the user; obtain multiple products provided by the mobile payment platform; estimate a product recommendation score of each product based on at least one specific health data and at least one specific environmental data; based on the product recommendation score of each product in the multiple products Select multiple specific products from the products and recommend the multiple specific products to specific users.

The invention provides a server for recommending products on a mobile payment platform, which includes a storage circuit and a processor. The storage circuit stores multiple modules. The processor is coupled to the storage circuit and accesses the plurality of modules to perform the following steps: when a specific application corresponding to the mobile payment platform is launched by a specific user, obtain at least one specific health data of the specific user and At least one specific environmental data associated with a specific user; obtaining multiple products provided by a mobile payment platform; estimating a product recommendation score for each product based on at least one specific health data and at least one specific environmental data; based on the product recommendation score of each product A plurality of specific products are selected from the plurality of products, and the multiple specific products are recommended to a specific user.

100:Server

102:Storage circuit

104: Processor

400:Knowledge Graph

r: relationship vector

h,t: entity node vector

S210~S240: steps

Figure 1 illustrates recommendations on a mobile payment platform according to an embodiment of the present invention. Product server diagram.

FIG. 2 is a flow chart of a method for recommending products on a mobile payment platform according to an embodiment of the present invention.

FIG. 3 is a schematic diagram of a knowledge graph according to the first embodiment of the present invention.

FIG. 4 is a schematic diagram of a knowledge graph according to a second embodiment of the present invention.

Figure 5 is a schematic diagram of a TranE triplet according to a second embodiment of the present invention.

Please refer to FIG. 1 , which is a schematic diagram of a server for recommending products on a mobile payment platform according to an embodiment of the present invention. In different embodiments, the server 100 may be used, for example, to maintain a mobile payment platform, and the mobile payment platform may be designed with corresponding specific applications by the manufacturer. In one embodiment, users can access various services provided by the above-mentioned mobile payment platform by installing and opening the specific application on the computer device/smart device they are using, such as paying points, Use cash or other similar methods to purchase goods provided by the mobile payment platform, but it is not limited to this.

In addition, in some embodiments, the server 100 can obtain health data measured from multiple users (such as age, gender, vision, height, weight, blood sugar, blood pressure, heartbeat, pulse, sleep depth, etc.) from various channels. ). For example, the above health data can be measured by an IoT platform that manages the multiple users through a wearable device worn on the user, and provided to the server 100 by the IoT platform, but it is not limited to this.

In an embodiment of the present invention, the server 100 can also obtain the current health data and environmental data (such as temperature, humidity, longitude and latitude, air pressure, rainfall, time, etc.) of each user when purchasing a certain product, and related users. Behavior (such as the cumulative number of purchases of a product, etc.), and used as a basis for recommending products to other users in the future, the relevant details will be detailed later.

As shown in FIG. 1 , the server 100 may include a storage circuit 102 and a processor 104 . The storage circuit 102 is, for example, any type of fixed or removable random access memory (Random Access Memory, RAM), read-only memory (Read-Only Memory, ROM), flash memory (Flash memory), hardware disc or other similar device, or a combination of such devices, which may be used to record multiple codes or modules.

The processor 104 is coupled to the storage circuit 102 and can be a general purpose processor, a special purpose processor, a conventional processor, a digital signal processor, a plurality of microprocessors, one or more combined digital signal processing Microprocessors, controllers, microcontrollers, Application Specific Integrated Circuits (ASICs), Field Programmable Gate Arrays (FPGAs), and any other types of integrated circuits at the core of the processor , state machines, Advanced RISC Machine (ARM)-based processors, and the like.

In an embodiment of the present invention, the processor 104 can access the modules and program codes recorded in the storage circuit 102 to implement the method of recommending products on a mobile payment platform proposed by the present invention, the details of which are described in detail below.

Please refer to FIG. 2 , which illustrates a mobile support system according to an embodiment of the present invention. Flowchart of the method to pay for recommended products on the platform. The method in this embodiment can be executed by the server 100 in Figure 1. The details of each step in Figure 2 will be described below with reference to the components shown in Figure 1.

First, in step S210, the processor 104 may obtain specific health data of the specific user and specific environmental data associated with the specific user when a specific application corresponding to the mobile payment platform is launched by the specific user. For example, if the processor 104 detects that the specific user launches the specific application on the smart device used by the processor 104, the processor 104 can obtain the information of the specific user through the Internet of Things platform. Specific health data (such as pulse, blood pressure, etc.) measured by wearable devices or other related sensors for the specific user, and specific environmental data measured for the environment in which the specific user is located (such as temperature, humidity, etc.), but may not be limited to this.

Afterwards, in step S220, the processor 104 may obtain multiple commodities provided by the mobile payment platform. In different embodiments, the products provided by the mobile payment platform may include various household appliances, sporting goods, healthy food, etc., but may not be limited thereto.

In step S230, the processor 104 may estimate the product recommendation score of each product based on the specific health data and the specific environmental data.

In an embodiment of the present invention, the processor 104 can estimate the product recommendation score of each product based on similar principles. To facilitate the explanation of the concept of the present invention, only products for one of the products (hereinafter referred to as the first product) will be described below. A score estimation mechanism is recommended, but the invention may not be limited thereto.

In a first embodiment, the processor 104 may be configured to assign each specific health Convert the data into the corresponding specific health feature vector, convert each specific environment data into the corresponding specific environment feature vector; obtain the product feature vector corresponding to the first product; convert the specific health feature vector corresponding to each specific health data, each specific environment data The corresponding specific environment feature vector and the product feature vector corresponding to the first product are input to a specific neural network, and the specific neural network can respond to the specific health feature vector corresponding to each specific health data and the specific environmental features corresponding to each specific environmental data. The vector and the product feature vector corresponding to the first product generate a product recommendation score for the first product.

In the first embodiment, the specific neural network is trained through the first pre-training mechanism, and the relevant details and the conversion methods of the various feature vectors mentioned above will be described later.

In the second embodiment, the processor 104 may be configured to: convert each of the specific health data into a corresponding specific health feature vector, and input the specific health feature vector to the first specific neural network, wherein the first specific neural network The network generates a first score in response to the specific health feature vector; converts each specific environment data into a corresponding specific environment feature vector, and inputs the specific environment feature vector into a second specific neural network, where the second specific neural network The route generates a second score in response to the specific environment feature vector; obtains the product feature vector corresponding to the first product, and inputs the product feature vector to the third specific neural network, wherein the third specific neural network generates a product feature vector in response to the product feature vector a third score; performing a weighted operation on the first score, the second score and the third score to obtain the product recommendation score of the first product.

In the second embodiment, the above-mentioned first, second, and third specific neural networks are trained by the second pre-training mechanism, and the relevant details and the above-mentioned various characteristics are directed to The method of converting quantities will be explained later.

After obtaining the product recommendation scores of each product based on the above principles, in step S240, the processor 104 may select multiple specific products from the multiple products based on the product recommendation scores of each product, and add the multiple specific products to Recommended for specific users. For example, the processor 104 may recommend one or several products with higher product recommendation scores as specific products to a specific user, so that the specific user can more easily see these specific products in the interface of a specific application. , thereby increasing the purchase probability of specific users for these specific products.

The following will describe the first pre-training mechanism used for a specific neural network in the first embodiment. Briefly speaking, the processor 104 can collect the current health data, environmental data (such as temperature, humidity, longitude and latitude, air pressure, rainfall, time, etc.) of multiple users when they purchase a certain product, and related user behaviors (such as for the cumulative number of purchases of a certain product, etc.), and used to create training data that can be used to train a specific neural network.

Specifically, for a reference user among the above users, the processor 104 can collect the reference user's purchasing behavior for the reference product, as well as the reference user's health data, environmental data and reference product when purchasing the reference product. product information. Afterwards, the processor 104 can obtain the user behavior feature vector corresponding to the purchase behavior, obtain the health feature vector corresponding to the above health data, obtain the environment feature vector corresponding to the above environmental data, and obtain product information corresponding to the reference product. product feature vector.

In a first embodiment, the processor 104 may The rules convert the above purchasing behavior into the corresponding user behavior feature vector, the above health data into the corresponding health feature vector, the above environmental data into the corresponding environmental feature vector, and the product information of the reference product into the corresponding Product feature vector.

For example, assuming that one of the environmental data being considered is humidity, the designer can first pre-define corresponding description values for various humidity levels, as shown in Table 1 below.

It can be seen from Table 1 that when one of the obtained environmental data is a humidity of 50%, the processor 104 will correspondingly convert the environmental data to 1, and when the obtained environmental data is a humidity of 60%, the processor 104 104 will correspondingly convert this environment data to 0.8. The relationship between the remaining humidity and the corresponding description value can be deduced in the same way, and will not be described again here.

Similarly, for other types of environmental data (such as temperature, time, etc.), health data (such as blood pressure, height, weight, heartbeat, etc.) and reference products, designers can also set corresponding conversion relationships based on similar principles to obtain The corresponding health feature vector, environment feature vector and product feature vector corresponding to the reference product.

In one embodiment, assuming that the health data of the reference user when purchasing the reference product is as shown in Table 2, the processor 104 can convert each health data into a corresponding description value based on the corresponding predefined feature rules, and then form a health data. eigenvector. For example, the designer can design a table similar to Table 1 for blood pressure, and when the processor 104 obtains the blood pressure with a value of 140/100mmHg, it can convert the blood pressure into a corresponding description value based on the table. , for example 0.8. Based on a similar principle, the processor 104 can also estimate the reference user's obesity level description value (eg, 0.9) and tachycardia level description value (eg, 0.7) based on other health data in Table 2. Thereby, the processor 104 can convert the health data in Table 2 into the corresponding health feature vector, that is, [0.8, 0.9, 0.7], but it is not limited to this.

In one embodiment, assuming that the environment data of the reference user when purchasing the reference product is as shown in Table 3, and the product information of the reference product is as shown in Table 4, the processor 104 can also classify each environment based on the corresponding predefined feature rules. The data and product information are converted into corresponding description values, and then formed into an environment feature vector (for example, [0.4, 0.7]) and a product feature vector (for example, [0.8, 0.3, 0.6]), but it is not limited to this.

In addition, for the reference user's purchasing behavior when purchasing the reference product (such as the cumulative number of purchases of the reference product, etc.), the processor 104 can also convert it into a corresponding user behavior feature vector based on similar principles, but it may not be limited to this.

In the first embodiment, the designer can establish a knowledge graph based on the concept of knowledge extension. For example, assuming that 1 is "high blood pressure → cause → excessive salt intake" and 2 is "potato chips → composition → high sodium content", then the related inference can be "high blood pressure → eat less potatoes" piece".

Please refer to FIG. 3 , which is a schematic diagram of a knowledge graph according to the first embodiment of the present invention. In this embodiment, Figure 3 is, for example, a knowledge graph established by the designer based on the concept of knowledge extension. After obtaining the above health feature vector, environment feature vector, product feature vector and user behavior feature vector, the processor 104 These feature vectors can be filled in corresponding nodes in Figure 3 accordingly, but are not limited thereto.

In the second embodiment, the designer can pre-establish a knowledge graph, and this knowledge graph can include multiple connected nodes, where each node can correspond to one of health factors, environmental factors, product information factors, and user behavior factors. One, but not limited to this.

Please refer to FIG. 4 , which is a schematic diagram of a knowledge graph according to a second embodiment of the present invention. In Figure 4, the designer can establish each node and related connection relationships in advance to form a knowledge graph 400. Afterwards, the processor 104 can execute the TransE algorithm, the PTransE algorithm or the TransR algorithm based on the knowledge graph 400 to generate the feature vector corresponding to each node.

In order to facilitate understanding of the concept of the present invention, the TransE algorithm will be used as an example for description below, but the present invention is not limited thereto. Please refer to FIG. 5 , which is a schematic diagram of a TranE triplet according to a second embodiment of the present invention. In Figure 5, for two entity node vectors h and t connected by a relationship vector r, the objective function of TransE is to minimize d(h+r,t) of the correct sample and maximize d(h '+r,t'), where d(x,y) is the vector distance, and the L2 square difference (Euclidean distance) is used to calculate the distance of the relationship. For example, the high blood pressure vector will be closer to the blood pressure, stroke, and potato chip vectors, but farther from the gender, temperature, and vegetable pie vectors.

，其中k為向量維度，每一筆訓練資料為(h,r,t)組合；(2)從訓練資料取樣出n筆資料做批次，針對每一筆資料，產生出對應錯誤關聯 資料(固定h,r替換t，或固定r,t替換h)；(3)針對這批正確及關聯錯誤資料T，使用最小化邊界法則迭代更新L=Σ _T [γ+d(h+r,t)-d(h'+r,t')]₊，其中d是L1或L2距離，+表示使用距離相減為正值的部分，γ是可調整的邊界參數；(4)重複步驟(2)、(3)直到收斂。 In one embodiment, the specific steps for the processor 104 to execute TransE may include: (1) Initializing all entity nodes and associated vectors as random numbers distributed among

, where k is the vector dimension, and each piece of training data is a combination of ( h, r, t ); (2) Sample n pieces of data from the training data to make a batch, and for each piece of data, the corresponding error correlation data (fixed h ,r replaces t , or fixed r,t replaces h ); (3) For this batch of correct and associated error data T , use the minimization boundary rule to iteratively update L =Σ _T [ γ + d ( h + r,t )- d ( h' + r,t' )] ₊ , where d is the L 1 or L 2 distance, + represents the part using the distance subtracted to a positive value, γ is an adjustable boundary parameter; (4) Repeat step (2) ), (3) until convergence.

到

之間。一般而言，知識圖譜400中有關聯節點的距離會較小，反之無關聯的距離會較大，如下表五所例示。 In Figure 4, there are a total of 12 entity nodes and 6 types of associations (12 in total). Assuming that the vector dimension is 3 (i.e., k is 3) and the batch size is 1, then using L1 distance, the initial vector values will be distributed in -

arrive

between. Generally speaking, the distance between related nodes in the knowledge graph 400 will be smaller, and conversely, the distance between unrelated nodes will be larger, as shown in Table 5 below.

In the second embodiment, after the processor 104 executes the TransE algorithm based on the knowledge graph 400, the obtained feature vectors of each node can be exemplified in Table 6 below.

After that, the processor 104 can find the user behavior feature vector corresponding to the purchase behavior, the health feature vector corresponding to the health data, the environment feature vector corresponding to the environment data, and the product corresponding to the reference product from the feature vectors of each node. Product feature vector of information.

Based on the first embodiment or the second embodiment, the user behavior feature vector corresponding to the purchasing behavior, the health feature vector corresponding to the health data, the environment feature vector corresponding to the environment data, and the product information corresponding to the reference product are obtained. After the feature vector, the processor 104 can generate reference training data based on the user behavior feature vector, health feature vector, environment feature vector and product feature vector, and train a specific neural network based on the reference training data.

In the first embodiment, the processor 104 may, for example, combine the user behavior feature vector, the health feature vector, the environment feature vector and the product feature vector into a super vector, and use this super vector as a piece of training data to train the specific Neural network, but is not limited to this.

In one embodiment, the training objective function for training a specific neural network can be characterized as: p' ( y | x ) = (1- α - β ). p ( y | x )+ α ． h ( x )+ β . h ( y ) , where p ( y | x ) is the probability of user weight. In addition, the cross entropy of the loss function for training a specific neural network can be characterized as: H ( p , q )=-Σ _t p ( t )log q ( t ), where p is the training corpus distribution (i.e., p' ( y | _ _

It should be understood that when the processor 104 executes step S230, each specific health data can be converted into a corresponding specific health feature vector based on the concepts described in the first/second embodiment, and each specific environment data can be converted into The corresponding specific environment feature vector, and obtain the product feature vector corresponding to the first product. And, in In the first embodiment, the processor 104 may combine the specific health feature vector corresponding to each specific health data, the specific environment feature vector corresponding to each specific environment data, and the product feature vector corresponding to the first product into a super vector, and then The supervector is input to the trained specific neural network, so that the specific neural network generates a product recommendation score for the first product. From another point of view, it can be known from the training objective function of a specific neural network that the product recommendation score of the first product can be understood as the probability that a specific user purchases or clicks on the first product and that the first product helps improve specific usage. A comprehensive score of the patient's health, but may not be limited to this.

The following second embodiment will describe the second pre-training mechanism. In the second embodiment, the above-mentioned second pre-training mechanism may be used to jointly train the first, second, third and fourth specific neural networks, for example. Similar to the first embodiment, the processor 104 in the second embodiment can also collect current health data and environmental data (such as temperature, humidity, longitude and latitude, air pressure, rainfall, time) of multiple users when purchasing a certain product. etc.), product information and related user behavior (such as the cumulative number of purchases of a product, etc.) as training data for training the first, second, third and fourth specific neural networks.

Specifically, when executing the second pre-training mechanism, the processor 104 may collect the reference user's purchasing behavior for the reference product, as well as the reference user's health data, environmental data and product information of the reference product when purchasing the reference product. Furthermore, the processor 104 can obtain the user behavior feature vector corresponding to the purchase behavior, obtain the health feature vector corresponding to the health data, obtain the environment feature vector corresponding to the environment data, and obtain the product feature vector corresponding to the product information of the reference product. . above For the steps, please refer to the description of the first and second embodiments in the first embodiment, and will not be described again here.

In the second embodiment, after obtaining the user behavior feature vector corresponding to the purchasing behavior, the health feature vector corresponding to the health data, the environment feature vector corresponding to the environment data, and the product feature vector corresponding to the product information of the reference product, The processor 104 may jointly train the first, second, third and fourth specific neural networks based on the health feature vector, the environment feature vector, the product feature vector and the user behavior feature vector respectively.

In short, the first, second, third and fourth specific neural networks can respectively be based on corresponding training data and objective functions (such as p' ( y | x )) and training objectives (such as minimizing H ( p , q )) for training, but may not be limited to this.

It should be understood that when the processor 104 executes step S230, each specific health data can be converted into a corresponding specific health feature vector based on the concepts described in the first/second embodiment, and each specific environment data can be converted into The corresponding specific environment feature vector, and obtain the product feature vector corresponding to the first product. Furthermore, in the second embodiment, the processor 104 may use a first specific neural network to convert the specific health feature vector corresponding to each specific health data into a first score, and use a second specific neural network to convert the specific health feature vector corresponding to each specific environmental data into a first score. The specific environment feature vector is converted into a second score, a third specific neural network is used to convert the product feature vector corresponding to the first product into a third score, and then the first, second, and third scores are multiplied by the corresponding weights and added The product recommendation score is always the first product.

In one embodiment, it is assumed that the goods provided by the mobile payment platform include Dust mite vacuum cleaner, probiotics, dehumidifier, oatmeal, tummy toner and fat burning tea. In this case, if the specific health data of the specific user shows that the specific user has high body fat or high blood sugar, and the specific environmental data corresponding to the environment of the specific user shows that the environment is relatively humid, the processor 104 can based on the above Specific health data and specific environmental data estimate individual product recommendation scores for the above products. Assuming that the product recommendation scores of the dust mite-removing vacuum cleaner and the fat-burning tea are relatively high (that is, the processor 104 determines that a specific user is more likely to purchase these two products and/or these two products are more beneficial to the specific user's health), the process The device 104 can preferentially recommend these two products to specific users, but is not limited to this.

To sum up, the present invention at least has the following characteristics: (1) The present invention can use wearable devices to collect a lot of personalized health information and current environmental factors and upload them to the cloud; (2) The present invention can collect a lot of personalized health information and environmental factors. Factors use knowledge graphs to establish feature vectors; (3) The present invention will combine neural network-like deep learning methods to use feature vectors such as user behavior, health factors, environmental factors, products, etc., and train one or more specific neural networks in advance to perform Subsequent product recommendation scores are sorted; (4) The present invention actively helps users recommend product information that is suitable for the user. In this way, the present invention can combine environmental factors and personal health factors to achieve personalized and accurate product recommendation marketing.

Although the present invention has been disclosed above through embodiments, they are not intended to limit the present invention. Anyone with ordinary knowledge in the technical field may make some modifications and modifications without departing from the spirit and scope of the present invention. Therefore, The protection scope of the present invention shall be determined by the appended patent application scope.

S210~S240: steps

Claims (10)

A method for recommending products on a mobile payment platform, including: a processor, when a specific application corresponding to the mobile payment platform is launched by a specific user, obtains at least one specific health data of the specific user and is associated with the At least one specific environmental data of the specific user, wherein the at least one specific environmental data includes at least one of temperature, humidity, longitude and latitude, air pressure and rainfall; the processor obtains multiple products provided by the mobile payment platform; the processing The processor estimates a product recommendation score of each product through one or more specific neural networks based on the at least one specific health data and the at least one specific environmental data; the processor estimates a product recommendation score of each product based on the product recommendation score of each product in the Select multiple specific products from these products and recommend these specific products to the specific user. The method of claim 1, wherein the products include a first product, and the step of the processor estimating the product recommendation score of each product based on the at least one specific health data and the at least one specific environmental data includes: The processor converts each specific health data into a corresponding specific health feature vector; the processor converts each specific environment data into a corresponding specific environment feature vector; the processor obtains a product corresponding to the first product Feature vector; the processor combines the specific health feature vector corresponding to each specific health data, the specific environment feature vector corresponding to each specific environment data, and the first product corresponding The product feature vector is input to a specific neural network, and the specific neural network responds to the specific health feature vector corresponding to each specific health data, the specific environment feature vector corresponding to each specific environment data, and the first The product feature vector corresponding to the product generates the product recommendation score of the first product. The method of claim 2, wherein the specific neural network is trained through a first pre-training mechanism, wherein the first pre-training mechanism includes: the processor collects a reference user's information on a reference product. Purchase behavior, as well as at least one health data, at least one environmental data and product information of the reference product when the reference user purchases the reference product; the processor obtains a user behavior feature vector corresponding to the purchase behavior, obtains the corresponding From the at least one health feature vector of the at least one health data, obtain at least one environment feature vector corresponding to the at least one environmental data and obtain a product feature vector corresponding to the product information of the reference product; the processor is based on the use The behavioral feature vector, the at least one health feature vector, the at least one environmental feature vector and the product feature vector generate a reference training data; the processor trains the specific neural network based on the reference training data. The method as described in claim 3, wherein the first training mechanism trains a training objective function of the specific neural network represented by: p' ( y | x ) = (1- α - β ). p ( y | x )+ α ． h ( x )+ β . h ( y ) , where p ( y | x ) is the probability of user weights, where the cross entropy of the loss function for training this particular neural network is represented by:

, where p is the training corpus distribution, q is the predicted distribution of the specific neural network, t is the training data of the specific neural network, and the training goal of the specific neural network includes minimizing the cross entropy of the loss function. The method of claim 1, wherein the products include a first product, and the step of the processor estimating the product recommendation score of each product based on the at least one specific health data and the at least one specific environmental data includes: The processor converts each specific health data into a corresponding specific health feature vector, and inputs the specific health feature vector to a first specific neural network, wherein the first specific neural network responds to the specific health feature The vector generates a first score; the processor converts each specific environment data into a corresponding specific environment feature vector, and inputs the specific environment feature vector to a second specific neural network, wherein the second specific neural network The route generates a second score in response to the specific environment feature vector; the processor obtains a product feature vector corresponding to the first product, and inputs the product feature vector to a third specific neural network, wherein the third specific The neural network generates a third score corresponding to the product feature vector; The processor performs a weighted operation on the first score, the second score and the third score to obtain the product recommendation score of the first product. The method of claim 5, wherein the first specific neural network, the second specific neural network, the third specific neural network and a fourth specific neural network are trained by a second pre-training mechanism Thus, the second pre-training mechanism includes: the processor collects a purchase behavior of a reference user for a reference product, and at least one health data, at least one environmental data and at least one environmental data of the reference user when purchasing the reference product. Product information of the reference product; the processor obtains a user behavior feature vector corresponding to the purchase behavior, obtains at least one health feature vector corresponding to the at least one health data, and obtains at least one health feature vector corresponding to the at least one environmental data. Environmental feature vectors and obtaining a product feature vector corresponding to the product information of the reference product; the processor is based on the at least one health feature vector, the at least one environmental feature vector, the product feature vector and the user behavior feature vector respectively The first specific neural network, the second specific neural network, the third specific neural network and the fourth specific neural network are jointly trained. The method as described in claim 3 or 6, including: the processor converts the purchase behavior into the corresponding user behavior feature vector according to a plurality of predefined feature rules, and converts the at least one health data into the corresponding at least one A health feature vector, converting the at least one environmental data into the corresponding at least one environmental feature vector, and converting the product information of the reference product into the corresponding product feature vector. The method as described in claim 3 or 6, including: the processor establishing a knowledge graph, wherein the knowledge graph includes a plurality of connected nodes, each node corresponding to a health factor, an environmental factor, a product information factor and One of the user behavior factors; the processor executes a TransE algorithm, a PTransE algorithm or a TransR algorithm based on the knowledge graph to generate a feature vector corresponding to each node; the processor performs Find the user behavior feature vector corresponding to the purchase behavior, the at least one health feature vector corresponding to the at least one health data, and the at least one environment feature vector corresponding to the at least one environmental data from the feature vectors of the node. and the product feature vector corresponding to the product information of the reference product. The method of claim 6, wherein the second training mechanism jointly trains the first specific neural network, the second specific neural network, the third specific neural network and the fourth specific neural network. A training objective function is represented by: p' ( y | x ) = (1- α - β ). p ( y | x )+ α ． h ( x )+ β . h ( y ) , where p ( y | x ) is the probability of user Weight, wherein the cross entropy of the loss function for jointly training the first specific neural network, the second specific neural network, the third specific neural network and the fourth specific neural network is represented by:

, where p is the training corpus distribution, q is the predicted distribution of the specific neural network, t is the training data of these specific neural networks, and the first specific neural network, the second specific neural network, the The training objectives of the third specific neural network and the fourth specific neural network include minimizing the cross entropy of the loss function. A server for recommending products on a mobile payment platform includes: a storage circuit that stores a plurality of modules; and a processor that is coupled to the storage circuit and accesses the modules to perform the following steps: corresponding to the When a specific application of the mobile payment platform is launched by a specific user, at least one specific health data of the specific user and at least one specific environmental data associated with the specific user are obtained, wherein the at least one specific environmental data includes temperature , humidity, latitude and longitude, air pressure and rainfall; obtain multiple products provided by the mobile payment platform; estimate a product recommendation score for each product based on the at least one specific health data and the at least one specific environmental data; based on The product recommendation score of each product selects multiple specific products among the products and recommends the specific products to the specific user.

Images