TWI740106B - Product recommendation method and device - Google Patents

Abstract

The embodiments of this specification provide a product recommendation method and device, wherein the method is used to determine whether to recommend a product to be recommended to a target user. The method includes: obtaining multi-domain information associated with the target user, the information including: the target The user's purchase information in the product area of the product to be recommended and purchase information in other product areas; the user characteristic matrix of the target user is constructed based on the information in multiple fields; for a product to be recommended, multiple users who purchase the product to be recommended are acquired Based on the feature values in the matrix, obtain the product feature matrix of the product to be recommended; input the user feature matrix and the product feature matrix into the machine learning model to obtain the user preference vector and product preference vector; The user preference vector and the product preference vector obtain a selection evaluation value between the product to be recommended and the target user; when the selection evaluation value is greater than a predetermined recommendation threshold, it is determined that the product to be recommended is recommended to the target user.

Description

Product recommendation method and device

The present disclosure relates to the technical field of data processing, in particular to a product recommendation method and device.

In the field of product recommendation, cold boot and data sparseness are common problems. Cold boot means that product recommendations are made in the absence of a large amount of user data support; data sparseness means that projects that interact with users are only the tip of the iceberg of the overall project, resulting in extremely sparse data in the user project scoring matrix. For example, in the recommendation of financial wealth management products, due to the large transaction amount and low frequency of the financial wealth management industry, user behavior information is scarce, and there is no large amount of user data for product recommendation, resulting in a cold start The problem; Moreover, the purchase behavior of users' financial management products accounts for only a small part of the total registered users of financial product platforms, and the problem of data sparseness is also very prominent. So far, the most widely used personalized recommendation method is based on single-field collaborative filtering, that is, recommending products that are most similar to the user's favorites or products that are most similar to the products he has liked before to the target user. . However, how to give users a more satisfactory recommendation result under the conditions of cold boot and sparse data is an urgent problem to be solved.

In view of this, one or more embodiments of this specification provide a product recommendation method and device to improve the quality of product recommendation in the absence of information. Specifically, one or more embodiments of this specification are implemented through the following technical solutions: In a first aspect, a product recommendation method is provided. The method is used to determine whether to recommend a product to be recommended to a target user. The method includes: Acquiring multi-domain information associated with the target user, the multi-domain information including: purchase information of the target user in the product area of the product to be recommended and purchase information of other product areas; Constructing a user feature matrix of the target user according to the multi-domain information, the user feature matrix including: a plurality of feature values quantified according to the multi-domain information; For one product to be recommended, the user feature matrix of multiple users who purchased the product to be recommended is obtained, and based on the feature value in the user feature matrix of the multiple users, all State the product feature matrix corresponding to the product to be recommended; The user feature matrix and the product feature matrix are respectively input into a pre-trained machine learning model to obtain a user preference vector and a product preference vector. The user preference vector is used to indicate the preference of the target user in product purchase, so The product preference vector is used to represent the characteristics of users who purchase the product to be recommended; According to the user preference vector and the product preference vector, a selection evaluation value between the product to be recommended and the target user is obtained, and the selection evaluation value is used to indicate that the target user purchases the product to be recommended Probability When the selection evaluation value is greater than a predetermined recommendation threshold, it is determined to recommend the product to be recommended to the target user. In a second aspect, a product recommendation device is provided, the device is used to determine whether to recommend a product to be recommended to a target user, and the device includes: An information acquisition module for acquiring multi-field information associated with the target user, the multi-field information including: purchase data of the target user in the product field of the product to be recommended and purchase data of other product fields ； A user matrix construction module for constructing a user feature matrix of the target user based on the multi-domain information, the user feature matrix including: a plurality of feature values quantified according to the multi-domain information; The product matrix construction module is used to obtain the user characteristic matrix of multiple users who purchase the product to be recommended for one product to be recommended, and based on the user characteristic matrix of the multiple users To obtain the product feature matrix corresponding to the product to be recommended; The model processing module is used to input the user feature matrix and the product feature matrix into a pre-trained machine learning model to obtain a user preference vector and a product preference vector. The user preference vector is used to indicate that the target user is in Product purchase preferences, where the product preference vector is used to represent the characteristics of users who purchase the product to be recommended; The output processing module is used to obtain a selection evaluation value between the product to be recommended and the target user according to the user preference vector and the product preference vector, and the selection evaluation value is used to indicate the target use The probability of the person buying the product to be recommended; The recommendation determination module is configured to determine that the product to be recommended is recommended to the target user when the selection evaluation value is greater than a predetermined recommendation threshold. In a third aspect, a product recommendation device is provided. The device includes a memory, a processor, and computer instructions that are stored on the memory and can run on the processor. When the processor executes the instructions, the following steps are implemented: Acquiring multi-domain information associated with the target user, the multi-domain information including: purchase information of the target user in the product area of the product to be recommended and purchase information of other product areas; Constructing a user feature matrix of the target user according to the multi-domain information, the user feature matrix including: a plurality of feature values quantified according to the multi-domain information; For one product to be recommended, the user feature matrix of multiple users who purchased the product to be recommended is obtained, and based on the feature value in the user feature matrix of the multiple users, all State the product feature matrix corresponding to the product to be recommended; The user feature matrix and the product feature matrix are respectively input into a pre-trained machine learning model to obtain a user preference vector and a product preference vector. The user preference vector is used to indicate the preference of the target user in product purchase, so The product preference vector is used to represent the characteristics of users who purchase the product to be recommended; According to the user preference vector and the product preference vector, a selection evaluation value between the product to be recommended and the target user is obtained, and the selection evaluation value is used to indicate that the target user purchases the product to be recommended Probability When the selection evaluation value is greater than a predetermined recommendation threshold, it is determined to recommend the product to be recommended to the target user. The product recommendation method and device of one or more embodiments of this specification integrates user behavior data and basic information from multiple fields, and uses deep neural networks to intelligently perceive the user’s preference characteristics related to product purchases, and explain the use Select appropriate financial management products to effectively alleviate the sparse transaction data and cold start problems faced by the industry, effectively improve the accuracy of personalized recommendations of financial management products, and provide target users with more accurate recommendation services.

上述的表1中，目標矩陣可以包括使用者對產品的購買選擇值，所述購買選擇值用於表示使用者是否購買產品。示例性的，購買選擇值可以包括“1”或者“0”，當數值是1時，表示使用者購買了該產品；當數值是0時，表示使用者未購買該產品。該目標矩陣可以作為深度神經網路模型的訓練目標，當訓練中的模型的輸出結果與該訓練目標之間的偏差越來越小，並且偏差達到預定閾值時，才結束模型的訓練，並將訓練結束的模型直接用於後續金融理財產品的推薦。 接著，對於目標矩陣中的各個使用者(例如，使用者A、使用者B)和各個產品(例如，產品1、產品2)，可以分別構建每個使用者的使用者特徵矩陣，並分別構建各個產品的產品特徵矩陣。並將構建的使用者特徵矩陣和產品特徵矩陣輸入待訓練的機器學習模型，輸出模型輸出矩陣，該模型輸出矩陣包括經過所述機器學習模型輸出的各個購買選擇值。在所述模型輸出矩陣和目標矩陣的偏差達到預定閾值時，模型訓練結束。 如下透過圖1詳細的描述模型訓練的過程，該過程中描述了如何構建上述的使用者特徵矩陣、產品特徵矩陣，以及如何將矩陣輸入模型以訓練模型的過程。 在步驟100中，獲取目標使用者關聯的多領域資訊，所述多領域資訊包括：目標使用者在待推薦產品的產品領域的購買資料和其他產品領域的購買資料。 本步驟中，目標使用者是待推薦產品的使用者，比如，想要向使用者A推薦產品，但是尚不知道向該使用者A推薦哪個產品，需要透過本實施例的推薦方法來確定要向使用者A推薦的產品，那麼該使用者A可以稱為目標使用者。 需要說明的是，在模型訓練中，這裡的目標使用者可以是目標矩陣中的使用者，這些使用者其實已經發生了實際的購買行為。而在後續的模型訓練結束後的模型使用中，目標使用者可以是待進行產品推薦的尚未對某些產品購買的使用者。 以金融理財產品為例，待推薦產品的產品領域即金融理財產品，目標使用者在該待推薦產品的產品領域的購買資料，例如可以包括：使用者購買某個金融理財產品的交易金額。而其他產品領域的購買資料可以是非金融理財產品的購買，例如，可以是購買衣服，購買電鍋等。所述的其他產品領域的購買資料可以是購買該其他領域產品的購買價格，比如，購買的衣服是200元，購買的電鍋是350元。而不論是待推薦產品的產品領域或者其他產品領域的購買資料，都是由目標使用者來進行購買而產生的資料。 此外，多領域資訊也不局限於上述的不同產品領域的購買資料，也可以包括其他的資訊。如下示例幾種，包括但不局限於下面的資訊： 例如，所述目標使用者的使用者屬性資訊。該使用者屬性資訊可以是使用者的性別、年齡、學歷等。 例如，所述目標使用者的關聯使用者在待推薦產品的產品領域的購買資料。其中，目標使用者的關聯使用者可以是與目標使用者具有好友關係、轉帳關係等。以好友關係為例，可以是目標使用者的好友發生過的購買金融理財產品的資料，比如，使用者A的好友使用者a購買過某個金融理財產品，且交易金額是2萬。 例如，目標使用者的借貸行為資料。該借貸行為資料可以目標使用者發生的借貸行為，借貸了某個品類的產品，且借貸的金額是多少。 在步驟102中，根據所述多領域資訊，構建所述目標使用者的使用者特徵矩陣，所述使用者特徵矩陣包括：根據所述多領域資訊量化的多個特徵值。 本步驟中，可以基於步驟100中採集到的資料進行量化，轉化為特徵值。 如下的表2示例一種使用者特徵矩陣的形式： 表2 使用者特徵矩陣

如上表2，在進行特徵值的量化之前，可以首先進行產品的粗細微性處理。粗細微性處理是將資料集中較為細化的資料轉化為概括性、綜合度較高的資料。若對於一個產品品類的購買資料，在所述產品品類下購買的產品數量達到粗細微性處理條件，則將所述產品品類下的多個產品進行粗細微性處理。舉例來說，假設其他產品領域的購買資料包括了購買衣服、電鍋等多個比較細的品類，並且，在電鍋這一個品類上目標使用者就購買了博愛思DFB-B 0.8L、奧克斯AR-Y0801、洛貝LBF-091BM等小容量電鍋處理為0-1L非電腦迷你電鍋，美的MB-WHS30C96、米家壓力IH、松下SR-AE101-K等家用全自動智慧電鍋等多種電鍋。那麼如果在使用者特徵矩陣構建時，將這些產品劃分的很細，比如，表2中的其他產品購買行為中，包括產品1、產品2、產品3等很多個產品，如上述的博愛思DFB-B 0.8L、奧克斯AR-Y0801、洛貝LBF-091BM等多個產品，那麼將造成很大的計算壓力。因此，粗細微性處理可以將細微性水準較細的特徵維度匯總到一個相對粗糙的細微性水準。 例如，博愛思DFB-B 0.8L、奧克斯AR-Y0801、洛貝LBF-091BM等小容量電鍋處理為0-1L非電腦迷你電鍋，美的MB-WHS30C96、米家壓力IH、松下SR-AE101-K等家用全自動智慧電鍋處理為3L-4L智慧微電腦電鍋。而是否對一個產品品類的購買資料進行粗細微性處理，可以設置粗細微性處理條件。例如，該條件可以是在所述產品品類下購買的產品數量達到一定的數量閾值，比如，在同一個產品品類下的產品數量達到了6個以上。而對於表2中的使用者屬性資訊、社交關係、金融理財產品購買與借貸行為等特徵維度，由於其特徵維度少、資訊含量高，可以不用進行粗細微性處理。 如下分別說明如何進行各個維度的特徵值量化，其中需要說明的是，如下的量化方法僅是示例，實際實施中並不局限於此，可以按照其他量化標準執行： 1)對於目標使用者及與目標使用者建立社交關係使用者的金融理財產品購買行為： 例如，可以根據購買的金融理財產品的交易金額，將交易金額合理劃分為多個區間，比如表2中的“＜P1”、“P1-P2”、“P2-P3”等多個區間。若使用者購買該金融理財產品的金額處於該區間內，則標記為1；否則為0。 其中，表2中的社交關係欄，與目標使用者具有關聯關係的使用者購買金融理財產品的購買資料，由於關聯關係的使用者可能是多個使用者，可以先將與目標使用者建立社交關係的所有使用者的交易金額進行平均，根據平均值的金額進行標記。比如，如果平均值處於區間“P1-P2”，則可以在對應該區間的特徵值標記1。 2)對於其他產品的購買行為： 如上所述的，其他產品的購買資料進行了粗細微性處理，處在同一粗糙細微性水準的可以有多個產品，並且這些產品的價格上可以具有相對較大的差異。此時可以以價格這一指標將該品類下所有產品合理劃分到表2的各個屬性區間，並將使用者購買頻次總體0-1標準化後的值作為其屬性值，反應目標使用者購買該品類下該價格區間內產品的頻繁程度。 舉例來說：假設目標使用者在其他產品的購買行為中，在3L-4L智慧微電腦電鍋這一品類下，購買了博愛思DFB-B 0.8L、奧克斯AR-Y0801、洛貝LBF-091BM等小容量電鍋處理為0-1L非電腦迷你電鍋，美的MB-WHS30C96、米家壓力IH、松下SR-AE101-K等家用全自動智慧電鍋，即同一品類下購買了多種產品。那麼可以根據這些產品各自的購買價格，查看在“＜P1”區間內購買的產品數量，並將該數量作為對應該區間的特徵值。比如，在所述的“＜P1”區間內購買了3個產品，則特徵值是3；在所述的“P1-P2”區間內購買了1個產品，則對應該區間的特徵值可以是1。 3)對於目標使用者的借貸行為： 例如，該借貸行為的量化與金融理財產品的量化類似，同樣是將借貸金額合理劃分為多個區間，若使用者借貸該品類產品的金額處於該區間內，則標記為1；否則為0。 4)對於使用者的基本資訊： 例如，對於數值型變數，如年齡，可以按照與交易金額相同的方法進行劃分。示例性的，18歲~25歲對應一個量化值，26歲~35歲對應一個量化值。 例如，對於類別型變數，如性別、學歷，則可以將變數因數編碼後標註。示例性的，本科學歷可以對應一個量化值，研究生學歷可以對應一個量化值。 在步驟104中，對於多個產品，獲取購買所述產品的多個使用者的所述使用者特徵矩陣，並基於所述多個使用者的使用者特徵矩陣中的特徵值，得到所述產品的產品特徵矩陣。 本步驟中的產品是金融理財產品。本步驟可以構建產品特徵矩陣，一個產品特徵矩陣可以對應一個產品，該產品可以是目標矩陣中的各個產品。其中，產品特徵矩陣的構建可以基於使用者特徵矩陣。 例如，以一個產品為例，購買該金融理財產品的有多個使用者，每一個使用者都構建了表2所示的使用者特徵矩陣。那麼可以基於多個使用者分別對應的多個使用者特徵矩陣，將特徵值進行加權平均。 In order to enable those skilled in the art to better understand the technical solutions in one or more embodiments of this specification, the following will combine the drawings in one or more embodiments of this specification to compare the The technical solution is described clearly and completely. Obviously, the described embodiments are only a part of the embodiments in this specification, rather than all the embodiments. Based on one or more embodiments of this specification, all other embodiments obtained by a person of ordinary skill in the art without creative work shall fall within the protection scope of the present disclosure. One or more embodiments of this specification provide a product recommendation method when the data is sparse. The description of the method takes the recommendation of financial management products as an example, but it is understandable that this method can also be applied to other cold-starting features. Product recommendation scenarios. Among them, the recommendation method integrates user behavior data from different fields, and uses behavior information in other fields than the product field to be recommended to alleviate data sparseness and cold boot problems in the product field to be recommended. Because the user's purchase behavior in other fields can also reflect the user's identity characteristics, environmental characteristics, life taste and other information that can help reflect the user's product purchase preference, the recommendation in the recommended product field is also a good reference. In addition, the recommendation method also uses a machine learning model. For example, taking a deep neural network as an example, the output result of the deep neural network model is used to assist in product recommendation. Of course, the deep neural network model can be trained first, and the trained model can be used for product recommendation. Model training: First, train a deep neural network model for financial product recommendation. The target matrix for model training can be constructed according to the actual collected data of product purchase. The actual collected data may include the user's purchase data of the product, for example, it may be the user's actual purchase record of the financial wealth management product. For example, the user A purchased the fund J1, and the user B purchased the stock G1 and the fund J1. User C purchases fund J2, and so on. According to the above-mentioned actual collected data, a target matrix can be constructed. The following Table 1 illustrates a target matrix, but it is not limited to this: Table 1 Target matrix

In the foregoing Table 1, the target matrix may include the user's purchase selection value for the product, and the purchase selection value is used to indicate whether the user purchases the product. Exemplarily, the purchase selection value may include "1" or "0". When the value is 1, it means that the user has purchased the product; when the value is 0, it means that the user has not purchased the product. The target matrix can be used as the training target of the deep neural network model. When the deviation between the output result of the training model and the training target becomes smaller and smaller, and the deviation reaches a predetermined threshold, the training of the model is ended, and The trained model is directly used for the recommendation of subsequent financial management products. Then, for each user (for example, user A, user B) and each product (for example, product 1, product 2) in the target matrix, the user characteristic matrix of each user can be constructed separately, and Product feature matrix of each product. The constructed user feature matrix and product feature matrix are input into the machine learning model to be trained, and the model output matrix is output. The model output matrix includes each purchase selection value output by the machine learning model. When the deviation between the model output matrix and the target matrix reaches a predetermined threshold, the model training ends. The following describes the process of model training in detail through Figure 1. The process describes how to construct the above-mentioned user feature matrix, product feature matrix, and how to input the matrix into the model to train the model. In step 100, multi-domain information related to the target user is obtained, and the multi-domain information includes: purchase data of the target user in the product field of the product to be recommended and purchase data of other product fields. In this step, the target user is the user of the product to be recommended. For example, if you want to recommend a product to user A, but you don’t know which product to recommend to user A, you need to use the recommendation method of this embodiment to determine the product to be recommended. A product recommended to user A, then user A can be called a target user. It should be noted that in model training, the target users here can be users in the target matrix, and these users have actually made purchases. In the use of the model after the subsequent model training, the target user may be a user who has not yet purchased some products for product recommendation. Taking a financial wealth management product as an example, the product area of the product to be recommended is a financial wealth management product, and the purchase information of the target user in the product area of the product to be recommended may include, for example, the transaction amount of a user purchasing a financial wealth management product. The purchase information in other product areas can be the purchase of non-financial wealth management products, for example, the purchase of clothes, the purchase of electric cookers, and so on. The purchase information of the other product field may be the purchase price of the product in the other field. For example, the purchased clothes are 200 yuan, and the purchased electric cooker is 350 yuan. Regardless of whether it is the product area of the product to be recommended or the purchase data of other product areas, it is data generated by the target user to make a purchase. In addition, the multi-field information is not limited to the above-mentioned purchase information of different product areas, and may also include other information. There are several examples as follows, including but not limited to the following information: For example, the user attribute information of the target user. The user attribute information can be the user's gender, age, education, etc. For example, the purchase data of the related users of the target user in the product field of the product to be recommended. Among them, the associated user of the target user may be a friend relationship or a transfer relationship with the target user. Taking the friend relationship as an example, it can be the data of the purchase of financial management products by the friends of the target user. For example, the friend user a of user A has purchased a certain financial management product, and the transaction amount is 20,000. For example, the target user’s lending behavior data. The loan behavior data can be the loan behavior of the target user, the loan of a certain category of products, and the amount of the loan. In step 102, a user feature matrix of the target user is constructed based on the multi-domain information, the user feature matrix including: a plurality of feature values quantified according to the multi-domain information. In this step, quantification can be performed based on the data collected in step 100 and converted into characteristic values. The following Table 2 illustrates a form of user characteristic matrix: Table 2 User characteristic matrix

As shown in Table 2, before the quantification of the characteristic value, the coarseness and fineness of the product can be processed first. The coarse and subtle processing is to transform the more detailed data in the data set into the more general and comprehensive data. If for the purchase information of a product category, the number of products purchased under the product category reaches the rough and fine processing conditions, then multiple products under the product category are subjected to rough and subtle processing. For example, suppose that the purchase information in other product fields includes the purchase of clothes, electric cookers and other relatively detailed categories, and in the electric cooker category, the target user purchases Boaisi DFB-B 0.8L, Aux AR-Y0801, Robe LBF-091BM and other small-capacity electric cookers are processed into 0-1L non-computer mini electric cookers, Midea MB-WHS30C96, Mijia Pressure IH, Panasonic SR-AE101-K and other household automatic smart electric cookers, etc. Electric pot. Then if the user feature matrix is constructed, these products are divided very finely. For example, the other product purchase behaviors in Table 2 include product 1, product 2, product 3 and many other products, such as the above-mentioned Boaisi DFB -B 0.8L, Aux AR-Y0801, Robe LBF-091BM and many other products, it will cause a lot of calculation pressure. Therefore, the coarse and fine processing can summarize the feature dimensions of the finer level of fineness to a relatively coarse level of fineness. For example, small-capacity electric cookers such as Boaisi DFB-B 0.8L, Aux AR-Y0801, Robbe LBF-091BM, etc. are processed into 0-1L non-computer mini electric cookers, Midea MB-WHS30C96, Mijia Pressure IH, Panasonic SR-AE101 -K and other household automatic smart electric cookers are processed into 3L-4L smart microcomputer electric cookers. Whether to perform coarse and fine processing on the purchase information of a product category, you can set the coarse and fine processing conditions. For example, the condition may be that the number of products purchased under the product category has reached a certain number threshold, for example, the number of products under the same product category has reached 6 or more. As for the characteristic dimensions of user attribute information, social relations, financial product purchase and lending behavior in Table 2, because of its small characteristic dimensions and high information content, it is not necessary to perform detailed processing. The following describes how to quantify the eigenvalues of each dimension. It should be noted that the following quantification methods are only examples, and are not limited to this in actual implementation, and can be implemented in accordance with other quantification standards: 1) For target users and The target user establishes the social relationship user’s financial and wealth management product purchase behavior: For example, the transaction amount can be reasonably divided into multiple intervals according to the transaction amount of the purchased financial wealth management product, such as "＜P1", "P1 in Table 2 -P2", "P2-P3" and many other intervals. If the user's purchase amount of the financial wealth management product is within this range, it is marked as 1; otherwise, it is 0. Among them, the social relationship column in Table 2 refers to the purchase data of financial and wealth management products purchased by users who have an associated relationship with the target user. Since the associated users may be multiple users, you can first establish a social relationship with the target user. The transaction amount of all users of the relationship is averaged and marked according to the average amount. For example, if the average value is in the interval "P1-P2", then 1 can be marked in the characteristic value of the corresponding interval. 2) For the purchase behavior of other products: As mentioned above, the purchase information of other products has been processed in detail. There can be multiple products at the same rough and subtle level, and the prices of these products can be relatively relatively high. Big difference. At this time, all products in the category can be reasonably divided into the attribute ranges in Table 2 based on the price index, and the standardized value of the user’s purchase frequency of 0-1 as the attribute value, reflecting the target user’s purchase of the category The frequency of products in this price range. For example: Suppose the target user purchases other products in the category of 3L-4L smart microcomputer electric cooker, and purchases Boaisi DFB-B 0.8L, Aux AR-Y0801, Robe LBF-091BM, etc. The small-capacity electric cooker is processed into a 0-1L non-computer mini electric cooker, such as Midea MB-WHS30C96, Mijia Pressure IH, Panasonic SR-AE101-K and other household automatic smart electric cookers, that is, a variety of products have been purchased under the same category. Then you can check the number of products purchased in the "<P1" interval according to the respective purchase prices of these products, and use this quantity as the characteristic value of the corresponding interval. For example, if 3 products are purchased in the "<P1" interval, the characteristic value is 3; if 1 product is purchased in the "P1-P2" interval, the characteristic value corresponding to the interval can be 1. 3) For the borrowing behavior of the target user: For example, the quantification of the borrowing behavior is similar to the quantification of financial wealth management products, and the borrowing amount is also reasonably divided into multiple ranges. If the amount of the user borrowing the product is within this range , It is marked as 1; otherwise, it is 0. 4) Basic information for users: For example, for numeric variables, such as age, it can be divided in the same way as the transaction amount. Exemplarily, 18 to 25 years old corresponds to a quantified value, and 26 to 35 years old corresponds to a quantified value. For example, for categorical variables, such as gender and educational background, the variable factors can be coded and labeled. Exemplarily, a bachelor degree may correspond to a quantitative value, and a graduate degree may correspond to a quantitative value. In step 104, for multiple products, obtain the user characteristic matrix of multiple users who purchased the product, and obtain the product based on the characteristic value in the user characteristic matrix of the multiple users Product feature matrix. The products in this step are financial wealth management products. In this step, a product feature matrix can be constructed. A product feature matrix can correspond to a product, and the product can be each product in the target matrix. Among them, the construction of the product feature matrix can be based on the user feature matrix. For example, taking a product as an example, there are multiple users who purchase the financial wealth management product, and each user constructs the user characteristic matrix shown in Table 2. Then, the eigenvalues can be weighted and averaged based on multiple user feature matrices respectively corresponding to multiple users.

For example, taking the age in the basic information as an example, each user who purchases the product has a characteristic value corresponding to the age, and the characteristic values of multiple users can be weighted and averaged to obtain a comprehensive characteristic value corresponding to an age.

For another example, taking category 1 in other product purchase behaviors in Table 2 as an example, each user of multiple users has a feature value corresponding to category 1, and the feature values of multiple users can be calculated Weighted average to obtain a comprehensive feature value corresponding to category 1.

It can also be seen that each feature value in Table 2 corresponds to a different feature value position. For example, the feature value position corresponding to x1 in Table 2 is [column corresponds to "P1-P2" interval, row corresponds to "category 1"] , And the feature value position corresponding to feature value x2 is [column corresponds to "P2-P3" interval, row corresponds to "category 1"]. When constructing the product feature matrix, the feature values corresponding to the same feature value location in the user feature matrix of multiple users can be weighted and averaged to obtain the feature value corresponding to the feature value location in the product feature matrix.

That is, in each row in Table 2, the feature values of multiple users can be weighted and averaged, and finally a product feature matrix that can reflect the overall features of the users who purchased the product can be obtained.

Among them, the setting of the weight in the weighted average of the characteristic value can be determined according to the actual business situation. For example, if you think that the characteristic value of a user is more important in reflecting the overall characteristics of the user, set its weight higher.

In step 106, the user feature matrix and product feature moment Array for attribute interactive operations.

In this step, the interactive operation of the attributes of the user feature matrix and the product feature matrix can be performed. The attribute interaction operation is to establish the correlation between the attributes that are not directly related in the matrix. First, the constructed feature matrix is randomly sorted by the attribute behavior unit to generate multiple new feature matrices, and then multiple new feature matrices are spliced to produce the attribute interaction. The feature matrix. It should be noted that the attribute interaction operation can be an optional operation. After the attribute interaction operation is performed, the potential associations between different features can be found more effectively, so that the subsequent use of machine learning models to perceive user preferences will also more precise.

The principle of the attribute interaction operation of the feature matrix can be seen in Fig. 2: As shown in Fig. 2, each feature such as feature 1, feature 2, feature 3, etc. corresponds to different feature rows. Taking the user feature matrix as an example, feature 1 can be "product 1 in the purchase behavior of financial wealth management products" in Table 2, and feature 15 can be "borrowing category 1 in lending behavior" in Table 2, that is, different features Corresponds to different rows. As shown in Figure 2, it is equivalent to random movement between different rows in Table 2, random sorting by row unit, and then splicing.

In step 108, the user feature matrix and the product feature matrix after the interaction are respectively input into the machine learning model to obtain the user preference vector and the product preference vector.

In this step, the deep neural network consists of two parallel neural networks, one of which is the smart perceptron for user behavior preference, and the other is the smart perceptron for general characteristics of users who purchase the product, as shown in Figure 3. Show. The feature matrix after attribute interaction and splicing is used as the input of the parallel neural network. For example, the user feature matrix after the attribute interaction is input into a neural network, and the product feature matrix after the attribute interaction is input into another neural network.

After the neural network's convolutional layer, pooling layer and full connection operation, the neural network can obtain the user preference vector and the product preference vector respectively. Wherein, the user preference vector can be used to indicate the user's preference for product purchase, which is equivalent to indicating what kind of product a user likes to buy. The product preference vector can be used to indicate the characteristics of users who purchase the products corresponding to the product feature matrix, which is equivalent to indicating what characteristics of a product users are more inclined to purchase the product. In step 110, a model output matrix is obtained according to the user preference vector and the product preference vector output by the model, and the model output matrix includes each purchase selection value output by the machine learning model. For example, input a user feature matrix corresponding to a user into a neural network model to obtain a user preference vector; input a product feature matrix corresponding to a product into a neural network model to obtain a product preference vector. A purchase selection value can be obtained according to the user preference vector and the product preference vector. For example, the inner product of the aforementioned user preference vector and the product preference vector can be calculated to obtain a purchase selection value, which represents the probability of the aforementioned user purchasing the product. For each user in the target matrix, a user feature matrix can be constructed, and for each product, a corresponding product feature matrix can be constructed separately. According to the above method, one of the users can obtain the purchase selection value of a product. These purchase selection values can form a model output matrix, that is, each purchase selection value included in the model output matrix is a value output by the neural network model. The user’s product purchase selection value included in the target matrix is obtained based on actual collected data and is the actual purchase behavior of the user. The target matrix is the actual mutual selection matrix between the user and the product. The target matrix can be used as the training target of the neural network model. With the continuous optimization of the model, the output result of the neural network model will be closer to the actual occurrence value. In step 112, when the deviation between the model output matrix and the target matrix reaches a predetermined threshold, the model training ends. For example, it can be set to end the training of the model when the deviation between the model output matrix and the target matrix reaches a predetermined threshold. The deviation reaching the predetermined threshold value may be that the deviation value is less than or equal to the predetermined threshold value, that is, the deviation between the two is sufficiently small. Among them, the deviation of the model output matrix and the target matrix can be measured in a variety of methods. For example, the deviation can be measured using Root Mean Square Error (RMSE) or MAE (Mean Absolute Deviation). After the model training is over, when predicting the mutual selection probability between the user and the product according to the trained neural network model, the prediction will be close to the actual situation, and there is a great probability of predicting success. Use of the model after training: Assuming that the training of the two parallel neural networks has been completed, the following is an example to illustrate how to use the trained model to determine which product to recommend to the user will have a higher success rate. For example, suppose that financial management products are currently recommended to user Y, and the products to be recommended include: product C1, product C2, product C3, etc., then which financial management product to recommend to user Y will have a higher success rate. You can follow the recommended method in this example. The user feature matrix of user Y can be constructed first, and the product feature matrixes of multiple products such as product C1, product C2, and product C3 can be constructed respectively. Then, input the user feature matrix of user Y and the product feature matrix of product C1 into the parallel neural network respectively to obtain the user preference vector and the product preference vector. And based on these two vectors, the user Y's selection evaluation value of the product C1 is obtained, and the selection evaluation value is used to indicate the probability of the target user buying the evaluation product. The selection evaluation value is calculated in the same way as the purchase selection value mentioned above, except that the two names are used to distinguish it. The purchase selection value is the value calculated during model training, and the selection evaluation value is calculated when the model is used after training. The value of is used as a basis for recommending products to users. The above-mentioned products C1, product C2, product C3 and other products to be recommended can be called evaluation products, that is, it is evaluated whether these products are to be recommended to user Y. A selection evaluation value can be obtained between the product feature matrix of each product and the user feature matrix of user Y. A recommendation threshold may be set, and when the selection evaluation value is greater than a predetermined recommendation threshold, it is determined to recommend the evaluation product to the target user. For example, suppose the choice evaluation value of product C1 and user Y is 0.6, the choice evaluation value of product C2 and user Y is 0.8, the choice evaluation value of product C3 and user Y is 0.2, and the recommendation threshold is 0.55 , Then it can be determined to recommend product C1 and product C2 to user Y, but product C3 is not recommended. The personalized recommendation method of financial products in this example integrates user behavior data and basic information from multiple fields, and uses deep neural networks to intelligently perceive user preferences related to product purchases, so that users can choose the right ones. Financial wealth management products effectively alleviate the sparse transaction data and cold start problems faced by the industry, effectively improve the accuracy of personalized recommendation of financial wealth management products, provide more accurate recommendation services for target users, and become a sales platform and use promotion A powerful measure for a benign interaction between people. In order to implement the above method, at least one embodiment of this specification also provides a product recommendation device. As shown in FIG. 4, the device can be used to determine whether to recommend the product to be recommended to the target user. The device can include: an information acquisition module 41, a user matrix construction module 42, a product matrix construction module 43, a model The processing module 44, the output processing module 45, and the recommendation determination module 46. The information acquisition module 41 is used to acquire multi-domain information associated with the target user, the multi-domain information including: the target user’s purchase data in the product field of the product to be recommended and purchases in other product fields material; The user matrix construction module 42 is configured to construct a user feature matrix of the target user according to the multi-domain information, the user feature matrix including: a plurality of feature values quantified according to the multi-domain information; The product matrix construction module 43 is configured to obtain the user feature matrix of multiple users who purchase the product to be recommended for one product to be recommended, and based on the user feature matrix of the multiple users The feature value in, obtain the product feature matrix corresponding to the product to be recommended; The model processing module 44 is used to input the user feature matrix and the product feature matrix into a pre-trained machine learning model to obtain a user preference vector and a product preference vector, and the user preference vector is used to represent the target user Preference in product purchase, where the product preference vector is used to represent the characteristics of users who purchase the product to be recommended; The output processing module 45 is configured to obtain a selection evaluation value between the product to be recommended and the target user according to the user preference vector and the product preference vector, and the selection evaluation value is used to represent the target The probability that the user purchases the product to be recommended; The recommendation determination module 46 is configured to determine to recommend the product to be recommended to the target user when the selection evaluation value is greater than a predetermined recommendation threshold. In an example, the user matrix construction module 42 is also used to: if the number of products purchased under the product category reaches the rough and subtle processing conditions for the purchase information of a product category, then the product category is Many of the products are subjected to coarse and fine processing. In one example, the product matrix construction module 43 is specifically configured to perform a weighted average on the feature values corresponding to the same feature value position in the user feature matrix of the multiple users to obtain the corresponding feature values in the product feature matrix. The eigenvalue at the position of the eigenvalue. In one example, the model processing module 44 is also used to separately attribute the user feature matrix and product feature matrix before inputting the user feature matrix and product feature matrix into the pre-trained machine learning model. Interactive operation; input the interactive user feature matrix and product feature matrix into the machine learning model. The devices or modules described in the above embodiments can be implemented by computer chips or entities, or implemented by products with certain functions. A typical implementation device is a computer. The specific form of the computer can be a personal computer, a laptop computer, a cellular phone, a camera phone, a smart phone, a personal digital assistant, a media player, a navigation device, an email receiving and sending device, and a game. A console, a tablet, a wearable device, or a combination of any of these devices. For the convenience of description, when describing the above device, the functions are divided into various modules and described separately. Of course, when implementing one or more embodiments of this specification, the functions of each module can be implemented in the same or multiple software and/or hardware. The execution order of each step in the flow shown in the above figure is not limited to the order in the flow chart. In addition, the description of each step can be implemented in the form of software, hardware, or a combination thereof. For example, those skilled in the art can implement it in the form of software program code, which can be a computer capable of implementing the logic function corresponding to the step. Executable instructions. When it is implemented in software, the executable instructions can be stored in the memory and executed by the processor in the device. For example, corresponding to the above method, one or more embodiments of this specification provide a product recommendation device at the same time. The device may include a processor, a memory, and computer instructions stored on the memory and running on the processor. The processor executes the instructions to implement the following steps: Acquiring multi-domain information associated with the target user, the multi-domain information including: purchase information of the target user in the product area of the product to be recommended and purchase information of other product areas; Constructing a user feature matrix of the target user according to the multi-domain information, the user feature matrix including: a plurality of feature values quantified according to the multi-domain information; For one product to be recommended, the user feature matrix of multiple users who purchased the product to be recommended is obtained, and based on the feature value in the user feature matrix of the multiple users, all State the product feature matrix corresponding to the product to be recommended; The user feature matrix and the product feature matrix are respectively input into a pre-trained machine learning model to obtain a user preference vector and a product preference vector. The user preference vector is used to indicate the preference of the target user in product purchase, so The product preference vector is used to represent the characteristics of users who purchase the product to be recommended; According to the user preference vector and the product preference vector, a selection evaluation value between the product to be recommended and the target user is obtained, and the selection evaluation value is used to indicate that the target user purchases the product to be recommended Probability When the selection evaluation value is greater than a predetermined recommendation threshold, it is determined to recommend the product to be recommended to the target user. It should also be noted that the terms "include", "include" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, commodity or equipment including a series of elements includes not only those elements, but also Other elements that are not explicitly listed, or also include elements inherent to such processes, methods, commodities, or equipment. If there are no more restrictions, the element defined by the sentence "including a..." does not exclude the existence of other identical elements in the process, method, commodity, or equipment that includes the element. Those skilled in the art should understand that one or more embodiments of this specification can be provided as a method, a system, or a computer program product. Therefore, one or more embodiments of this specification may adopt the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware. Moreover, one or more embodiments of this specification can be implemented on one or more computer-usable storage media (including but not limited to disk memory, CD-ROM, optical memory, etc.) containing computer-usable program codes. In the form of a computer program product. One or more embodiments of this specification may be described in the general context of computer-executable instructions executed by a computer, such as a program module. Generally, program modules include routines, programs, objects, components, data structures, etc. that perform specific tasks or realize specific abstract data types. One or more embodiments of this specification can also be practiced in a distributed computing environment. In these distributed computing environments, remote processing devices connected through a communication network perform tasks. In a distributed computing environment, program modules can be located in local and remote computer storage media including storage devices. The various embodiments in this specification are described in a progressive manner, and the same or similar parts between the various embodiments can be referred to each other, and each embodiment focuses on the difference from other embodiments. In particular, as for the data processing device embodiment, since it is basically similar to the method embodiment, the description is relatively simple, and for related parts, please refer to the part of the description of the method embodiment. The foregoing describes specific embodiments of this specification. Other embodiments are within the scope of the patent application. In some cases, the actions or steps described in the scope of the patent application may be performed in a different order than in the embodiments and still achieve desired results. In addition, the processes depicted in the drawings do not necessarily require the specific order or sequential order shown in order to achieve the desired result. In some embodiments, multiplexing and parallel processing are also possible or may be advantageous. The above descriptions are only preferred embodiments of one or more embodiments of this specification, and are not intended to limit one or more embodiments of this specification. All within the spirit and principle of one or more embodiments of this specification, Any modification, equivalent replacement, improvement, etc. made should be included in the protection scope of one or more embodiments of this specification.

41‧‧‧Information Acquisition Module 42‧‧‧User Matrix Building Module 43‧‧‧Product matrix building module 44‧‧‧Model Processing Module 45‧‧‧Output Processing Module 46‧‧‧Recommended Confirmation Module

In order to more clearly describe the technical solutions in one or more embodiments of this specification or the prior art, the following will briefly introduce the drawings that need to be used in the description of the embodiments or prior art. Obviously, the figures in the following description The formulas are only some of the embodiments recorded in one or more embodiments of this specification. For those of ordinary skill in the art, without creative labor, other schemas can be obtained based on these schemas. Figure 1 is a process of model training provided by one or more embodiments of this specification; FIG. 2 is the principle of the attribute interaction operation of the feature matrix provided by one or more embodiments of this specification; 3 is a schematic diagram of neural network processing provided by one or more embodiments of this specification; Fig. 4 is a schematic structural diagram of a product recommendation device provided by one or more embodiments of this specification.

Claims (9)

A method of obtaining a selection evaluation value in a distributed computing environment. The selection evaluation value is used to indicate the probability that the target user will purchase the product to be recommended. The method is used to determine whether the local and remote processing equipment connected through a communication network executes Recommending the product to be recommended to a target user, the method includes: obtaining multi-field information associated with the target user, the multi-field information including: the target user’s purchase data in the product field of the product to be recommended and other product fields According to the multi-field information, construct the user feature matrix of the target user; for a product to be recommended, obtain the user feature matrix of multiple users who purchased the product to be recommended, and based on the multiple The feature value in the user feature matrix of each user is used to obtain the product feature matrix corresponding to the product to be recommended; perform attribute interaction operations on the user feature matrix and the product feature matrix respectively; The user feature matrix and the product feature matrix are input to the two parallel smart perceptrons in the pre-trained machine learning model to obtain the user preference vector and the product preference vector respectively. The user preference vector is used to indicate the target user’s product purchase The product preference vector is used to indicate the characteristics of the user who purchases the product to be recommended; according to the user preference vector and the product preference vector, the evaluation value of the choice between the product to be recommended and the target user is obtained; Compare the selection evaluation value with a predetermined recommendation threshold; and when it is determined that the selection evaluation value is greater than the predetermined recommendation threshold, it is determined to recommend the product to be recommended to the target user, wherein the user based on the multiple users The feature values in the feature matrix are used to obtain the product feature matrix corresponding to the product to be recommended, including: performing a weighted average on the feature values corresponding to the same feature value position in the user feature matrix of the multiple users to obtain the product feature matrix The eigenvalue corresponding to the position of the eigenvalue in. According to the method described in item 1 of the scope of the patent application, the multi-field information further includes at least one of the following: purchase data of the target user's affiliated user in the product field of the product to be recommended; the target user's user Property information. According to the method described in item 1 of the scope of patent application, the construction of the user characteristic matrix of the target user includes: if the purchase information of a product category, the number of products purchased under the product category reaches the rough and subtle processing conditions, Then, multiple products under the product category are processed with coarse and fine details, where the coarse and fine processing is to transform the more detailed purchase data in the purchase data collection into more general and comprehensive purchase data, in which, The rough and subtle processing conditions include at least one of the following: the quantity of the multiple products purchased under the product category reaches the expected A predetermined threshold; and the feature dimension in the user feature matrix is higher than a predetermined value. According to the method described in item 1 of the scope of patent application, before the user characteristic matrix and the product characteristic matrix are respectively input to the pre-trained machine learning model, the method further includes: training the machine learning model, and the training process includes The processing is as follows: construct a target matrix for model training based on the actual collection data of product purchases, the actual collection data includes the user's product purchase data, and the target matrix includes: the user's product purchase selection value determined according to the purchase data , The purchase selection value is used to indicate whether the user purchases a product; for each user in the target matrix, construct the user characteristic matrix for each user; construct each product for each product in the target matrix The product feature matrix of the product; input the constructed user feature matrix and product feature matrix into the machine learning model to be trained, and obtain the model output matrix according to the user preference vector and product preference vector output by the model, the model output matrix includes After each purchase selection value output by the machine learning model; when the deviation between the model output matrix and the target matrix reaches a predetermined threshold, the model training ends. According to the method described in item 1 of the scope of patent application, the attribute interaction operation includes: establishing a correlation relationship between the user feature matrix and the attributes that are not directly related to the product feature matrix, and first use the attributes in the matrix The behavior units are randomly sorted to generate multiple new feature matrices, and then the multiple new feature matrices are spliced to generate the user feature matrix after the attribute interaction and the product feature matrix. A device that obtains a selection evaluation value in a distributed computing environment. The selection evaluation value is used to indicate the probability that the target user purchases the product to be recommended. The device is used to determine whether or not the local and remote processing equipment connected through a communication network Recommend the product to be recommended to a target user. The device includes: an information acquisition module for acquiring multi-domain information associated with the target user. The multi-domain information includes: the target user is in the product field of the product to be recommended Purchase data and purchase data in other product areas; a user matrix construction module for constructing a user characteristic matrix of the target user based on the multi-field information; a product matrix construction module for Product, acquiring the user feature matrix of multiple users who purchased the product to be recommended, and obtaining the product feature matrix corresponding to the product to be recommended based on the feature value in the user feature matrix of the multiple users; model; The processing module is used to perform attribute interaction operations on the user feature matrix and the product feature matrix respectively, and respectively input the user feature matrix and the product feature matrix after the attribute interaction into the pre-trained machine learning The two neural networks paralleled in the model obtain the user preference vector and the product preference vector respectively. The user preference vector is used to express the target user’s preference for product purchase, and the product preference vector is used to express the purchase of the to-be-recommended The user characteristics of the product; an output processing module for obtaining a selection evaluation value between the product to be recommended and the target user based on the user preference vector and the product preference vector; and a recommendation determination module for Compare the selection evaluation value with a predetermined recommendation threshold, and when it is determined that the selection evaluation value is greater than the predetermined recommendation threshold, it is determined to recommend the product to be recommended to the target user. The product matrix construction module is specifically used The feature values corresponding to the same feature value position in the user feature matrix of the multiple users are weighted and averaged to obtain the feature value corresponding to the feature value location in the product feature matrix. According to the device described in item 6 of the scope of patent application, the user matrix construction module is also used to: if the purchase data of a product category, the number of products purchased under the product category reaches the rough and subtle processing conditions, then Perform coarse and fine processing on multiple products under the product category, where the coarse and fine processing is to convert the more detailed purchase data in the purchase data collection into more general and comprehensive purchase data, where the The coarse and subtle processing conditions include at least one of the following: the number of the multiple products purchased under the product category reaches a predetermined threshold; and The feature dimension in the user feature matrix is higher than a predetermined value. The device according to item 6 of the scope of patent application, wherein the attribute interaction operation includes: establishing a correlation between the user feature matrix and the attributes that are not directly related to the product feature matrix, and first use the attributes in the matrix The behavior units are randomly sorted to generate multiple new feature matrices, and then the multiple new feature matrices are spliced to generate the user feature matrix after the attribute interaction and the product feature matrix. A product recommendation device that obtains a selection evaluation value in a distributed computing environment. The selection evaluation value is used to indicate the probability of a target user buying the product to be recommended. The device includes local and remote processing devices connected through a communication network. Processing equipment such as memory, a processor, and computer instructions stored on the memory and running on the processor implement the following steps when the processor executes the instructions: obtain multi-domain information related to the target user, and the multi-domain information The information includes: the target user’s purchase data in the product field of the product to be recommended and other product fields; according to the multi-field information, the user characteristic matrix of the target user is constructed; for a product to be recommended, obtain The user feature matrix of the multiple users who purchased the product to be recommended, and based on the feature value in the user feature matrix of the multiple users, the product corresponding to the product to be recommended is obtained Product feature matrix; respectively perform attribute interaction operations on the user feature matrix and the product feature matrix. Input the user feature matrix and the product feature matrix after the attribute interaction into two parallel neural networks in the pre-trained machine learning model. The user preference vector and the product preference vector are obtained respectively. The user preference vector is used to indicate the preference of the target user in product purchase, and the product preference vector is used to indicate the characteristics of the user who purchases the product to be recommended; The user preference vector and the product preference vector are used to obtain a choice evaluation value between the product to be recommended and the target user, and the choice evaluation value is used to indicate the probability that the target user purchases the product to be recommended; compare the choice evaluation And a predetermined recommendation threshold; and when it is determined that the selection evaluation value is greater than the predetermined recommendation threshold, it is determined to recommend the product to be recommended to the target user, wherein the product is based on the features in the user feature matrix of the multiple users Value to obtain the product feature matrix corresponding to the product to be recommended, including: performing a weighted average on the feature values corresponding to the same feature value position in the user feature matrix of the multiple users to obtain the corresponding feature value in the product feature matrix The characteristic value of the location.

Images