Method and device for determining target users

Abstract

The present invention discloses a method and device for determining target users, and relates to the field of computer technology. A specific implementation of the method includes: determining the target group to which each item in a preset item set belongs in a predetermined item group; for each target group: obtaining multiple preliminary users of the target group, inputting the original feature data of each preliminary user for each item in the target group into a pre-trained feature generation model, and obtaining the optimized feature data of the preliminary user in the target group; inputting the optimized feature data into a pre-trained user classification model, and obtaining the classification result of the preliminary user in the target group; and using the classification result of each preliminary user in each target group to determine whether the preliminary user is the target user corresponding to the item set. This implementation can accurately identify the target user based on the predetermined item group.

Description

Technical Field

[0001] The present invention relates to the field of computer technology, and in particular to a method and device for determining a target user. Background Art

[0002] With the rapid development of Internet technology and new business models based on the Internet, service providers need to reach target users more accurately, such as high-potential users with high consumption tendencies. At present, service providers generally build target user judgment models based on fixed item category dimensions or brand dimensions. Such models are generally binary classification machine learning models, which mainly build statistical features through user behavior information on categories or brands and user attribute information, and use whether the user consumes as the target variable.

[0003] In the process of realizing the present invention, the inventors found that the prior art has at least the following problems: First, since fixed item categories or brands sever the associations between items of different categories or different brands, the prior art cannot accurately characterize the characteristics of users' behavior trajectories for a variety of items, and thus cannot describe their consumption expectations for a specific item pool through user behavior; second, the models in the prior art are associated with fixed item categories or brands, and their scope of use is therefore relatively fixed, which is inconsistent with the dynamically changing associations between items based on user behavior; in addition, the prior art requires a lot of manpower and material resources to perform ETL (extract, transform, load) on the original data, which is not conducive to achieving rapid iteration of business. Summary of the invention

[0004] In view of this, an embodiment of the present invention provides a method and device for determining target users, which can accurately identify target users based on predetermined item groups and automatically construct features using machine learning methods, thereby avoiding the adverse effects of performing ETL processing on actual business.

[0005] To achieve the above objective, according to one aspect of the present invention, a method for determining a target user is provided.

[0006] The method for determining a target user in an embodiment of the present invention comprises: determining the target group to which each item in a preset item set belongs in a predetermined item group; wherein the item group is determined based on the number of occurrences and / or the order of occurrence of multiple items in the same user behavior sequence; for each target group: obtaining multiple preliminary users of the target group, inputting the original feature data of each preliminary user for each item in the target group into a feature generation model that has been pre-trained and corresponds to the target group, and obtaining optimized feature data of the preliminary user in the target group; inputting the optimized feature data into a user classification model that has been pre-trained and corresponds to the target group, and obtaining a classification result of the preliminary user in the target group; and using the classification result of each preliminary user in each target group to determine whether the preliminary user is a target user corresponding to the item set.

[0007] Optionally, the item group is determined by the following steps: generating an item graph for characterizing the association relationship between the multiple items based on the number of occurrences and / or the order of occurrence of the multiple items in the same user behavior sequence; performing graph embedding processing on the item graph to obtain an embedding vector for each of the multiple items; and clustering the multiple items using the embedding vector to obtain multiple item groups.

[0008] Optionally, the method further includes: after generating the item map: inputting the attribute data and title data of the isolated items in the item map into a pre-trained item similarity model to obtain at least one similar item of the isolated item; and adding the association relationship between the isolated item and the similar item in the item map.

[0009] Optionally, the determining of the target group to which each item in the preset item set belongs in the predetermined item group includes: for any item in the item set, determining whether the predetermined item group contains the identifier of the item; if so, determining the item group containing the identifier of the item as the target group to which the item belongs; otherwise, inputting the attribute data and title data of the item into the item similarity model to obtain similar items of the item; determining the target group to which the item belongs using the item group to which the similar items belong; and, after determining the target group to which the item belongs, adding the item to the target group.

[0010] Optionally, the feature generation model is a deep neural network model, comprising an input layer, a first hidden layer, a second hidden layer, and an output layer connected in sequence; wherein the first hidden layer comprises a first perception layer, a first standardization layer, and a first activation layer; the second hidden layer comprises a second perception layer, a second standardization layer, and a second activation layer; the input data of the second perception layer is the data obtained by concatenating the output data of the first activation layer and the original feature data; the input data of the output layer is the data obtained by concatenating the output data of the second activation layer and the original feature data.

[0011] Optionally, the original feature data of any preliminary user for any item in the corresponding target group includes: attribute data of the preliminary user, attribute data of the item and behavior record data of the preliminary user for the item; and obtaining the optimized feature data of the preliminary user in the target group includes: inputting the original feature data of any preliminary user for each item in the corresponding target group into the feature generation model, and obtaining multiple multi-dimensional data of the preliminary user output by the second activation layer; dividing each initial dimension of the multi-dimensional data into at least one target dimension according to a numerical range, and converting the multi-dimensional data into an optimized feature data of the preliminary user in the target dimension; wherein the value of the optimized feature data in any target dimension is: the number of times the value of the multi-dimensional data corresponding to the target dimension is in the numerical range corresponding to the target dimension.

[0012] Optionally, the method of using the classification result of each preliminary user in each target group to determine whether the preliminary user is a target user corresponding to the set of items includes: using the classification result of each preliminary user in each target group and a weight value pre-set for each target group to obtain a weighted judgment result of the preliminary user; when the weighted judgment result is greater than a preset threshold, determining the preliminary user as the target user corresponding to the set of items.

[0013] Optionally, the user classification model is an XGBoost model; and the feature generation model and the user classification model corresponding to the same item group are trained by the following steps: connecting the feature generation model and the user classification model into a comprehensive model; wherein the input data of the user classification model is: optimized feature data converted from the output data of the second activation layer in the feature generation model; constructing a training sample set applied to the comprehensive model; wherein each training sample in the training sample set includes the original feature data as the feature part and the user classification result as the label part; and using the training sample set to train the feature generation model and the user classification model.

[0014] To achieve the above objective, according to another aspect of the present invention, a device for determining a target user is provided.

[0015] The device for determining target users in an embodiment of the present invention may include: a group division unit, used to determine the target group to which each item in a preset item set belongs in a predetermined item group; wherein the item group is determined based on the number of occurrences and / or the order of occurrence of multiple items in the same user behavior sequence; a feature generation and classification unit, used to: for each target group: obtain multiple preliminary users of the target group, input the original feature data of each preliminary user for each item in the target group into a feature generation model that has been pre-trained and corresponds to the target group, and obtain optimized feature data of the preliminary user in the target group; input the optimized feature data into a user classification model that has been pre-trained and corresponds to the target group, and obtain the classification result of the preliminary user in the target group; a target user judgment unit, used to use the classification result of each preliminary user in each target group to judge whether the preliminary user is a target user corresponding to the item set.

[0016] Optionally, the device may further include: a group pre-determination unit, used to generate an item graph for characterizing the association relationship between the multiple items based on the number of times and / or the order of occurrence of the multiple items in the same user behavior sequence; perform graph embedding processing on the item graph to obtain an embedding vector for each of the multiple items; and use the embedding vector to cluster the multiple items to obtain multiple item groups.

[0017] To achieve the above objective, according to another aspect of the present invention, an electronic device is provided.

[0018] An electronic device of the present invention includes: one or more processors; a storage device for storing one or more programs, when the one or more programs are executed by the one or more processors, the one or more processors implement the method for determining the target user provided by the present invention.

[0019] To achieve the above objective, according to yet another aspect of the present invention, a computer-readable storage medium is provided.

[0020] A computer-readable storage medium of the present invention stores a computer program, and when the program is executed by a processor, the method for determining a target user provided by the present invention is implemented.

[0021] According to the technical solution of the present invention, the embodiments of the above invention have the following advantages or beneficial effects: in advance, an item graph is constructed offline based on the number of occurrences and / or the order of occurrence of items in the user behavior sequence, and the embedding vector of the item is determined from the item graph using the graph embedding method, and the items are clustered using the embedding vector to form multiple item groups; when it is necessary to determine the target user for the item set, the item group to which each item in the item set belongs can be determined by querying or judging similar items, and then the subsequent target user identification is performed according to the item group. Through the above settings, the association relationship between items of different categories or different brands can be accurately reflected, the inherent defects of target user identification based on item category or brand can be overcome, and it is conducive to the accurate identification of target users. At the same time, the item group can be dynamically updated according to the user behavior data, so as to be consistent with the dynamically changing association relationship between items based on user behavior. In addition, the present invention uses a deep learning method to automatically generate optimized feature data directly from the original feature data without performing ETL processing, thereby improving work efficiency and saving a lot of manpower and material resources.

[0022] The further effects of the above-mentioned non-conventional optional manner will be described below in conjunction with the specific implementation manner. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] The accompanying drawings are used to better understand the present invention and do not constitute an improper limitation of the present invention.

[0024] Figure 1 is a schematic diagram of the main steps of the method for determining a target user in an embodiment of the present invention;

[0025] Figure 2 is a schematic diagram of an item map according to an embodiment of the present invention;

[0026] Figure 3 It is a schematic diagram of the composition of the feature generation model of an embodiment of the present invention;

[0027] Figure 4 is a schematic diagram of a specific flow chart of a method for determining a target user in an embodiment of the present invention;

[0028] Figure 5 is a schematic diagram of components of an apparatus for determining a target user in an embodiment of the present invention;

[0029] Figure 6 is an exemplary system architecture diagram in which embodiments of the present invention may be applied;

[0030] Figure 7 It is a schematic diagram of the structure of an electronic device used to implement the method for determining a target user in an embodiment of the present invention. DETAILED DESCRIPTION

[0031] The following is a description of exemplary embodiments of the present invention in conjunction with the accompanying drawings, including various details of the embodiments of the present invention to facilitate understanding, which should be considered as merely exemplary. Therefore, it should be recognized by those of ordinary skill in the art that various changes and modifications may be made to the embodiments described herein without departing from the scope and spirit of the present invention. Similarly, for clarity and conciseness, the description of well-known functions and structures is omitted in the following description.

[0032] It should be pointed out that the embodiments of the present invention and the technical features therein may be combined with each other without conflict.

[0033] Figure 1 4 is a schematic diagram of main steps of a method for determining a target user according to an embodiment of the present invention.

[0034] like Figure 1 As shown, the method for determining the target user in the embodiment of the present invention can be specifically performed according to the following steps:

[0035] Step S101: determining the target group to which each item in the preset item set belongs in the predetermined item group.

[0036] In this step, the above-mentioned item set refers to a set of multiple items specified by the business personnel, and the target user that needs to be identified in the present invention is for the item set. Exemplarily, the target user may be a user who has a strong consumption tendency for multiple items in the item set. It can be understood that the specific meaning of the target user may change according to the actual scenario, and the present invention is not limited to this. Generally, the items in the item set may involve multiple categories or brands.

[0037] In an embodiment of the present invention, an item group can be predetermined offline, and the steps for determining an item group can be as follows: First, select a certain number of multiple items (e.g., all items on the service provider platform) and multiple users who have performed preset behaviors (e.g., search, browse, click, follow, add to cart, and / or submit an order, etc.) on the above multiple items in a recent period of time (e.g., within 3 months). Afterwards, generate an item graph (an item graph is a knowledge graph used to characterize the association relationship between multiple items) based on the number of occurrences and / or the order of occurrence of the above multiple items in the same user behavior sequence. It can be understood that a user behavior sequence consists of multiple items involved in a complete user behavior (e.g., submitting an order) or multiple consecutive user behaviors of the same type (e.g., multiple consecutive searches, browses, clicks, follows, or adds to carts).

[0038] For example, user behavior sequence 1 is items D, A, and B searched by user a in sequence, user behavior sequence 2 is items B and E browsed by user b in sequence, user behavior sequence 3 is items D, E, and F clicked by user b in sequence, user behavior sequence 4 is items E, C, and B followed by user c in sequence, and user behavior sequence 5 is items B and A involved in sequence when user c submits an order. An item map can be established using the number of occurrences and / or the order of occurrence of items A, B, C, D, E, and F in user behavior sequences 1-5. The above-mentioned number of occurrences refers to the number of times multiple items appear together in the same user behavior sequence. In an embodiment of the present invention, the number of occurrences can be the number of times two adjacent items appear together in the same user behavior sequence. For example, in user behavior sequence 1, the number of occurrences of D and A is 1, the number of occurrences of A and B is 1, and the number of occurrences of D and B is zero.

[0039] Exemplarily, when establishing an item graph, first determine the nodes representing various items (nodes can be represented by circles), and then add directed edges with arrows between two nodes according to the order of appearance of the items in each user behavior sequence. The direction of the arrow is consistent with the order of appearance of the items, and the number of appearances of the two nodes can be used as the weight value of the corresponding directed edge. Figure 2 FIG. 1 is a schematic diagram of an item graph established according to user behavior sequences 1-5 in an embodiment of the present invention. It can be seen that the item graph includes multiple nodes representing items and directed edges between the nodes. Figure 2 In the example, the weight value of each directed edge is 1, indicating that the number of occurrences of two adjacent items in the five user behavior sequences is 1. It can be understood that the above is a method of using both the number of occurrences and the order of occurrence to establish an item graph. If only the number of occurrences is used to establish an item graph, the edges in the item graph are all undirected edges; if only the order of occurrence is used to establish an item graph, the edges in the item graph are all directed edges without weight values.

[0040] In practical applications, the item graph established according to the above method often contains isolated items (i.e., items corresponding to isolated nodes. Isolated nodes can be defined according to actual needs, for example, defined as nodes with less than two corresponding edges), resulting in poor connectivity of the item graph. As a preferred solution, at least one edge can be added to the isolated item in the following manner to enhance the connectivity of the item graph. First, the attribute data (such as item price, category, weight, etc.) and title data (such as the words formed after word segmentation of the item title) of the isolated items in the item graph can be input into a pre-trained item similarity model to obtain at least one similar item to the isolated item; then, an edge (the edge is generally an undirected edge) between the isolated item and its similar items is added to the item graph to represent the association relationship between the isolated items, and the similarity between the isolated item and its similar items can be used as the weight value of the edge.

[0041] Preferably, the above-mentioned item similarity model can be trained using the following method. First, a training sample set including multiple training samples is constructed, wherein the feature part of each training sample is the attribute data and title data of the item, and the label part is the identifier of multiple similar items (such as inventory unit SKU) and the similarity of each similar item. Afterwards, the item similarity model is trained using the above-mentioned training sample set and an applicable machine learning algorithm. When using the trained item similarity model to obtain similar items of isolated items, the multiple similar items output by the item similarity model can be arranged in descending order of similarity, and a preset number of similar items in front can be selected to optimize the item map.

[0042] After the item graph is established, graph embedding processing (i.e., graph embedding) can be performed on the item graph to obtain embedding vectors of multiple items in the item graph. The above embedding vectors and existing clustering algorithms such as K-Means can then be used to cluster the above multiple items, thereby ultimately determining multiple item groups. The method of obtaining embedding vectors using graph embedding processing is a known method, and its specific execution details are not repeated here.

[0043] It can be understood that the item embedding vector obtained by the above method can reflect the item association relationship based on user behavior. The item group formed by clustering the embedding vector can group together multiple items with high correlation from the user's perspective, and distribute items with low correlation from the user's perspective in different item groups, thereby achieving an item classification method that is superior to categories or brands. The item classification method based on the above item groups will not break the association relationship between items of different categories or different brands, which is helpful to accurately portray the characteristics of users' behavior trajectories for multiple items and thus facilitate the accurate identification of target users. It can also flexibly combine items based on item groups to carry out precision marketing. For example, a user has purchased multiple items with high correlation and different categories and brands for many times. When the target user is identified based on categories or brands, the user is often not identified as a target user; when the target user is identified based on item groups, the multiple different categories / brands of items purchased by the user are likely to belong to the same item group, so the user is likely to be identified as a target user. It can be seen that the accuracy of the target user identification method based on item groups is higher than that of the traditional category / brand-based method. It should be noted that the item groups may also be divided according to the similarity between items or the activity of items, and the present invention is not limited to this.

[0044] The above is the process of determining the item group in an offline state. In the specific application process of step S101, after receiving the item set, the item group to which each item in the item set belongs can be first determined (hereinafter, the item group to which these items belong is referred to as the target group). Specifically, for any item in the item set, it is determined whether the item group contains the identifier (such as SKU) of the item: if so, the item group containing the identifier of the item is determined as the target group to which the item belongs; otherwise, the attribute data and title data of the item are input into the aforementioned item similarity model to obtain multiple similar items of the item, and the target group to which the item belongs is determined using the item group to which the similar items belong. Exemplarily, after obtaining multiple similar items output by the item similarity model, the multiple similar items are first arranged in descending order according to the similarity, and a preset number of similar items in front are selected, and then the item group to which each selected similar item belongs is determined, and finally the item group containing the most similar items is determined as the target group to which the item belongs. In actual scenarios, after determining the target group to which the item belongs, the item can be added to the target group for expansion.

[0045] Step S102: For each target group: obtain multiple preliminary users of the target group, input the original feature data of each preliminary user for each item in the target group into a feature generation model that has been pre-trained and corresponds to the target group, and obtain the optimized feature data of the preliminary user in the target group; input the optimized feature data into a user classification model that has been pre-trained and corresponds to the target group, and obtain the classification result of the preliminary user in the target group.

[0046] After obtaining multiple target groups through step S101, the preliminary users of each target group can be first obtained in this step. The selection rules of the preliminary users can be determined according to actual needs. For example, users who have performed preset behaviors (such as browsing, clicking, paying attention, etc.) on any item in any target group in the recent time period (for example, within 15 days) can be selected. It should be noted that the same user can be a preliminary user for multiple target groups. Thereafter, for any preliminary user of any target group, the original feature data of the preliminary user can be constructed and input into a feature generation model that has been pre-trained and corresponds to the target group to obtain preferred feature data, and then the optimized feature data can be input into a user classification model that has been pre-trained and corresponds to the target group to classify the preliminary user, and the classification result can be used to determine whether the preliminary user is a target user corresponding to the set of items.

[0047] In the embodiment of the present invention, the user classification model can be various supervised classification models, such as the XGBoost model (which is a boosted tree model formed by integrating multiple weak classifiers). The feature generation model can be a DNN (Deep Neural Networks), and its structure can be as follows: Figure 3 As shown, it includes an input layer, a first hidden layer, a second hidden layer, and an output layer connected in sequence. Among them, the first hidden layer includes a first perception layer, a first standardization layer, and a first activation layer; the second hidden layer includes a second perception layer, a second standardization layer, and a second activation layer. The first perception layer and the second perception layer are the main parts of the hidden layer, and have multiple neurons; the first standardization layer and the second standardization layer can both be BatchNo rmalization (batch data standardization layer); the activation function of the first activation layer and the second activation layer can use Leaky ReLU (Leaky Rectified Linear Unit, with leakage rectified linear unit), and the output layer can use the softmax function. In specific scenarios, the feature generation model can also use other applicable machine learning models.

[0048] In an offline state, the feature generation model and user classification model corresponding to the same item group can be trained by the following steps. First, the feature generation model and the user classification model are connected to form a comprehensive model. Generally, the second activation layer of the feature generation model can be connected to the input layer of the user classification model, and the input data of the user classification model is the optimized feature data converted from the output data of the second activation layer. After that, a training sample set applied to the above comprehensive model is constructed; wherein each training sample in the training sample set can include the original feature data as the feature part and the user classification result as the label part. The above original feature data can be the relevant data of the preliminary users of the item group, and the original feature data of any preliminary user for any item in the corresponding target group can include: the attribute data of the preliminary user (such as age, gender and / or address, etc.), the attribute data of the item, and the behavior record data of the preliminary user for the item (such as browsing records, search records, click records, attention records, shopping cart records and / or purchase records, etc., which can be the number of executions of the relevant behavior, execution time, etc.). In practical applications, the above behavior record data can be transaction-level data obtained from user behavior logs. In addition, the original feature data of any preliminary selected user is multiple pieces of data, which include the original feature data of the preliminary selected user for each item in the corresponding target group. The original feature data of the preliminary selected user for any item includes the original feature data corresponding to each preset time interval. The original feature data corresponding to any preset time interval includes the behavior record data of the preliminary selected user for the item, the attribute data of the preliminary selected user and the attribute data of the item.

[0049] Finally, the feature generation model and the user classification model are trained using the training sample set. During the data processing, the original feature data of any pre-selected user is input into the input layer of the feature generation model, and then passes through the input layer, the first perception layer, the first standardization layer, and the first activation layer to output. After that, the output data of the first activation layer is spliced ​​with the original feature data as the input data of the second perception layer, and the data passes through the second perception layer, the second standardization layer, and the second activation layer to output. After that, the output data of the second activation layer is spliced ​​with the original feature data as the input data of the output layer.

[0050] In an embodiment of the present invention, the data output by the second activation layer can be processed by the following steps to generate optimized feature data of the preliminary user in the item group. Specifically, first obtain multiple multi-dimensional data (each dimension can be called an initial dimension) of the preliminary user output by the second activation layer, then divide each initial dimension into at least one target dimension according to a numerical range, and convert the multi-dimensional data into an optimized feature data of the preliminary user in the target dimension; wherein the value of the optimized feature data in any target dimension is: the number of times the value of the multi-dimensional data in the initial dimension corresponding to the target dimension is in the numerical range corresponding to the target dimension. For example, if the initial dimensions have two feature0 and feature1, the initial user has the following 6 multi-dimensional data:

[0051] Serial number feature0 feature1 1 0.0091 0.2231 2 0.3256 0.3365 3 0.5298 0.4987 4 0.7789 0.5986 5 0.7569 0.8899 6 0.9685 0.8522

[0052] It can be seen that due to the normalization layer, the values ​​of each initial dimension of the multi-dimensional data are between 0 and 1. At this time, feature0 and feature1 are first divided into multiple target dimensions according to the preset numerical intervals. For example, feature0 is divided into three target dimensions feature0-bin0, feature0-bin1, and feature0-bin2 according to [0, 0.3], (0.3, 0.6], and (0.6, 1], and feature1 is divided into three target dimensions feature1-bin0, feature1-bin1, and feature1-bin2 according to [0, 0.4], (0.4, 0.7], and (0.7, 1]. Then, the six multi-dimensional data are converted into one optimized feature data, that is, the values ​​of the multi-dimensional data in feature0 are counted. The number of times the value is in [0,0.3], (0.3,0.6], (0.6,1] (in this case, the numbers are 1, 2, and 3 respectively), and the three numbers are used as the values ​​of the optimized feature data in the target dimensions feature0-bin0, feature0-bin1, and feature0-bin2 respectively. The number of times the value of the multi-dimensional data in feature1 is in [0,0.4], (0.4,0.7], (0.7,1] (in this case, the numbers are 2, 2, and 2 respectively), and the three numbers are used as the values ​​of the optimized feature data in the target dimensions feature1-bin0, feature1-bin1, and feature1-bin2 respectively. In this way, the following optimized feature vectors of the primary selected user can be obtained:

[0053]

[0054] Through the above settings, multiple transaction-level data corresponding to any user can be converted into one user-level data of the user, while more information of the original data can be retained, and there is no need to perform ETL processing on the original data, thereby improving work efficiency and saving a lot of manpower and material resources. In practical applications, the average value, maximum value, minimum value, etc. of the values ​​of the same initial dimension in multi-dimensional data can also be calculated to generate optimized feature data.

[0055] The above is the offline training process of the feature generation model and the user classification model. In the specific application process involved in step S102, for the target group to which each item in the aforementioned item set belongs, a plurality of preliminary users of the target group can be first obtained, and the original feature data of each preliminary user for each item in the target group is input into the feature generation model that has been pre-trained and corresponds to the target group, thereby obtaining the optimized feature data of the preliminary user in the target group. After that, the optimized feature data is input into the user classification model that has been pre-trained and corresponds to the target group, thereby obtaining the classification result of the preliminary user in the target group. Generally, the classification result is related to whether the preliminary user has a consumption tendency within a preset time period in the future (for example, within 5 days), and can be represented by two different numerical values. When the preliminary user has a consumption tendency, the classification result is 1; when the preliminary user does not have a consumption tendency, the classification result is zero.

[0056] Step S103: using the classification result of each preliminary selected user in each target group, determining whether the preliminary selected user is a target user corresponding to the item set.

[0057] After step S102, the classification result of each preliminary user in each target group can be obtained. In this step, the classification results of each target group can be combined to determine whether the preliminary user is the target user corresponding to the set of items. Specifically, first, the weighted discrimination result of the preliminary user is obtained using the classification result of each preliminary user in each target group and the weight value pre-set for each target group (that is, the weighted sum of the classification results of the preliminary user in each target group. Generally, when a user is not a preliminary user of a target group, it can be considered that the classification result of the user in the target group is zero); then determine whether the weighted discrimination result is greater than a preset threshold: if so, the preliminary user is determined as the target user corresponding to the above-mentioned set of items; otherwise, it is determined that the preliminary user is not the target user corresponding to the above-mentioned set of items.

[0058] Figure 4 FIG. 1 is a schematic diagram of a specific flow chart of a method for determining a target user in an embodiment of the present invention. Figure 4As shown, the scheme of the present invention can be divided into an offline training stage and a target user determination stage. In the offline training stage, firstly, an item graph is constructed according to the co-occurrence relationship (i.e., the number of occurrences and / or the order of occurrence) of items in the user behavior sequence, and the connectivity of isolated items in the item graph is enhanced by using an item similarity model trained by item data (including item attribute data and title data). Then, a graph embedding method is used to obtain item embedding vectors from the item graph, and the item embedding vectors are clustered to form multiple item groups. Finally, a training sample set is constructed using the original feature data of the primary selected users in the item group and the optimized feature data to train the feature generation model and the user classification model. In the target user determination stage, firstly, the identifier of each item in the item set is queried in the item group. If it is found, the target group to which the item belongs is directly determined. If it is not found (indicating that the item is a new item), the item similarity model is used to obtain similar items of the item to determine the target group to which the item belongs. Afterwards, the trained feature generation model is used to obtain the optimized feature data of each preliminary user in each target group, and the trained user classification model is used to calculate the classification result of each preliminary user. Finally, the classification results are weighted and aggregated to determine the target user corresponding to the item set.

[0059] In the technical solution of the embodiment of the present invention, an item graph is constructed in advance in an offline state according to the number of occurrences and / or the order of occurrence of items in the user behavior sequence, and the embedding vector of the item is determined from the item graph using a graph embedding method, and the items are clustered using the embedding vector to form multiple item groups; when it is necessary to determine the target user for the item set, the item group to which each item in the item set belongs can be determined by querying or judging similar items, and then the subsequent target user identification is performed according to the item group. Through the above settings, the association relationship between items of different categories or different brands can be accurately reflected, the inherent defects of target user identification based on item category or brand can be overcome, and it is conducive to the accurate identification of target users. At the same time, the item group can be dynamically updated according to the user behavior data, so as to be consistent with the dynamically changing association relationship between items based on user behavior. In addition, the present invention uses a deep learning method to automatically generate optimized feature data directly from the original feature data without performing ETL processing, thereby improving work efficiency and saving a lot of manpower and material resources.

[0060] It should be noted that, for the convenience of description, the aforementioned method embodiments are described as a series of action combinations, but those skilled in the art should be aware that the present invention is not limited by the described action sequence, and some steps can actually be performed in other sequences or simultaneously. In addition, those skilled in the art should also be aware that the embodiments described in the specification are all preferred embodiments, and the actions and modules involved are not necessarily necessary to implement the present invention.

[0061] In order to better implement the above-mentioned solution of the embodiment of the present invention, relevant devices for implementing the above-mentioned solution are also provided below.

[0062] See also Figure 5 As shown, the device 500 for determining a target user provided by an embodiment of the present invention may include a group division unit 501 , a feature generation and classification unit 502 , and a target user determination unit 503 .

[0063] Among them, the group division unit 501 can be used to determine the target group to which each item in the preset item set belongs in the predetermined item group; wherein the item group is determined based on the number of occurrences and / or the order of occurrence of multiple items in the same user behavior sequence; the feature generation and classification unit 502 can be used to: for each target group: obtain multiple preliminary users of the target group, input the original feature data of each preliminary user for each item in the target group into a feature generation model that has been pre-trained and corresponds to the target group, and obtain the optimized feature data of the preliminary user in the target group; input the optimized feature data into a user classification model that has been pre-trained and corresponds to the target group, and obtain the classification result of the preliminary user in the target group; the target user judgment unit 503 can be used to use the classification result of each preliminary user in each target group to determine whether the preliminary user is the target user corresponding to the item set.

[0064] In an embodiment of the present invention, the device 500 may further include: a group pre-determination unit, which is used to generate an item graph for characterizing the association relationship between the multiple items based on the number of times and / or the order of occurrence of the multiple items in the same user behavior sequence; perform graph embedding processing on the item graph to obtain an embedding vector for each of the multiple items; and use the embedding vector to cluster the multiple items to obtain multiple item groups.

[0065] In practical applications, the group pre-determination unit can be further used to: input the attribute data and title data of the isolated items in the item map into a pre-trained item similarity model to obtain at least one similar item of the isolated item; and add the association relationship between the isolated item and the similar item in the item map.

[0066] In a specific application, the group division unit 501 can be further used to: for any item in the item set, determine whether the predetermined item group contains the identifier of the item; if so, determine the item group containing the identifier of the item as the target group to which the item belongs; otherwise, input the attribute data and title data of the item into the item similarity model to obtain similar items of the item; determine the target group to which the item belongs using the item group to which the similar items belong; and, after determining the target group to which the item belongs, add the item to the target group.

[0067] As a preferred solution, the feature generation model is a deep neural network model, comprising an input layer, a first hidden layer, a second hidden layer, and an output layer connected in sequence; wherein the first hidden layer comprises a first perception layer, a first standardization layer, and a first activation layer; the second hidden layer comprises a second perception layer, a second standardization layer, and a second activation layer; the input data of the second perception layer is the data obtained by concatenating the output data of the first activation layer and the original feature data; the input data of the output layer is the data obtained by concatenating the output data of the second activation layer and the original feature data.

[0068] Preferably, in an embodiment of the present invention, the original feature data of any preliminary user for any item in the corresponding target group includes: the attribute data of the preliminary user, the attribute data of the item and the behavior record data of the preliminary user for the item; and the feature generation and classification unit 502 can be further used to: input the original feature data of any preliminary user for each item in the corresponding target group into the feature generation model, and obtain multiple multi-dimensional data of the preliminary user output by the second activation layer; divide each initial dimension of the multi-dimensional data into at least one target dimension according to a numerical range, and convert the multi-dimensional data into an optimized feature data of the preliminary user in the target dimension; wherein the value of the optimized feature data in any target dimension is: the number of times the value of the multi-dimensional data corresponding to the target dimension is in the numerical range corresponding to the target dimension.

[0069] In some embodiments, the target user judgment unit 503 can be further used to: obtain a weighted judgment result of the preliminary user by utilizing the classification result of each preliminary user in each target group and the weight value pre-set for each target group; when the weighted judgment result is greater than a preset threshold, determine the preliminary user as the target user corresponding to the set of items.

[0070] In addition, in an embodiment of the present invention, the user classification model is an XGBoost model; and the device 500 may further include a training unit for training a feature generation model and a user classification model corresponding to the same item group. The training unit may be used to: connect the feature generation model and the user classification model into a comprehensive model; wherein the input data of the user classification model is: optimized feature data converted from the output data of the second activation layer in the feature generation model; construct a training sample set applied to the comprehensive model; wherein each training sample in the training sample set includes the original feature data as the feature part and the user classification result as the label part; and train the feature generation model and the user classification model using the training sample set.

[0071] In the technical solution of the embodiment of the present invention, an item graph is constructed in advance in an offline state according to the number of occurrences and / or the order of occurrence of items in the user behavior sequence, and the embedding vector of the item is determined from the item graph using a graph embedding method, and the items are clustered using the embedding vector to form multiple item groups; when it is necessary to determine the target user for the item set, the item group to which each item in the item set belongs can be determined by querying or judging similar items, and then the subsequent target user identification is performed according to the item group. Through the above settings, the association relationship between items of different categories or different brands can be accurately reflected, the inherent defects of target user identification based on item category or brand can be overcome, and it is conducive to the accurate identification of target users. At the same time, the item group can be dynamically updated according to the user behavior data, so as to be consistent with the dynamically changing association relationship between items based on user behavior. In addition, the present invention uses a deep learning method to automatically generate optimized feature data directly from the original feature data without performing ETL processing, thereby improving work efficiency and saving a lot of manpower and material resources.

[0072] Figure 6 An exemplary system architecture 600 is shown to which the method for determining a target user or the apparatus for determining a target user according to an embodiment of the present invention can be applied.

[0073] like Figure 6 As shown, the system architecture 600 may include terminal devices 601, 602, 603, a network 604 and a server 605 (this architecture is only an example, and the components included in the specific architecture may be adjusted according to the specific application). The network 604 is used to provide a medium for a communication link between the terminal devices 601, 602, 603 and the server 605. The network 604 may include various connection types, such as wired, wireless communication links or optical fiber cables.

[0074] Users can use terminal devices 601, 602, 603 to interact with server 605 through network 604 to receive or send messages, etc. Various client applications can be installed on terminal devices 601, 602, 603, such as an application for determining target users (only an example).

[0075] The terminal devices 601 , 602 , and 603 may be various electronic devices having a display screen and supporting web browsing, including but not limited to smart phones, tablet computers, laptop computers, and desktop computers, etc.

[0076] The server 605 may be a server that provides various services, such as a computing server that provides support for an application for determining a target user operated by a user using the terminal devices 601, 602, and 603 (for example only). The computing server may process the received target user identification request and feed back the processing result (for example, the determined target user - for example only) to the terminal devices 601, 602, and 603.

[0077] It should be noted that the method for determining the target user provided in the embodiment of the present invention is generally executed by the server 605 , and accordingly, the device for determining the target user is generally set in the server 605 .

[0078] It should be understood that Figure 6 The number of terminal devices, networks and servers in the embodiment is only for illustration. Any number of terminal devices, networks and servers may be provided according to implementation requirements.

[0079] The present invention also provides an electronic device. The electronic device of an embodiment of the present invention includes: one or more processors; a storage device for storing one or more programs, when the one or more programs are executed by the one or more processors, the one or more processors implement the method for determining the target user provided by the present invention.

[0080] Reference below Figure 7 , which shows a schematic diagram of the structure of a computer system 700 of an electronic device suitable for implementing an embodiment of the present invention. Figure 7 The electronic device shown is only an example and should not bring any limitation to the functions and scope of use of the embodiments of the present invention.

[0081] like Figure 7As shown, the computer system 700 includes a central processing unit (CPU) 701, which can perform various appropriate actions and processes according to a program stored in a read-only memory (ROM) 702 or a program loaded from a storage part 708 into a random access memory (RAM) 703. In the RAM 703, various programs and data required for the operation of the computer system 700 are also stored. The CPU 701, the ROM 702, and the RAM 703 are connected to each other via a bus 704. An input / output (I / O) interface 705 is also connected to the bus 704.

[0082] The following components are connected to the I / O interface 705: an input section 706 including a keyboard, a mouse, etc.; an output section 707 including a cathode ray tube (CRT), a liquid crystal display (LCD), etc., and a speaker, etc.; a storage section 708 including a hard disk, etc.; and a communication section 709 including a network interface card such as a LAN card, a modem, etc. The communication section 709 performs communication processing via a network such as the Internet. A drive 710 is also connected to the I / O interface 705 as needed. A removable medium 711, such as a magnetic disk, an optical disk, a magneto-optical disk, a semiconductor memory, etc., is installed on the drive 710 as needed so that a computer program read therefrom is installed into the storage section 708 as needed.

[0083] In particular, according to the embodiments disclosed in the present invention, the process described in the main step diagram above can be implemented as a computer software program. For example, the embodiments of the present invention include a computer program product, which includes a computer program carried on a computer-readable medium, and the computer program contains program code for executing the method shown in the main step diagram. In the above embodiment, the computer program can be downloaded and installed from the network through the communication part 709, and / or installed from the removable medium 711. When the computer program is executed by the central processing unit 701, the above functions defined in the system of the present invention are executed.

[0084] It should be noted that the computer-readable medium shown in the present invention may be a computer-readable signal medium or a computer-readable storage medium or any combination of the above two. The computer-readable storage medium may be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, device or device, or any combination of the above. More specific examples of computer-readable storage media may include, but are not limited to: an electrical connection with one or more wires, a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the above. In the present invention, a computer-readable storage medium may be any tangible medium containing or storing a program that can be used by or in combination with an instruction execution system, device or device. In the present invention, a computer-readable signal medium may include a data signal propagated in a baseband or as part of a carrier wave, which carries a computer-readable program code. This propagated data signal may take a variety of forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination of the above. Computer readable signal media may also be any computer readable medium other than computer readable storage media, which may send, propagate or transmit a program for use by or in conjunction with an instruction execution system, apparatus or device. The program code contained on the computer readable medium may be transmitted using any appropriate medium, including but not limited to: wireless, wire, optical cable, RF, etc., or any suitable combination of the above.

[0085] The flow chart and block diagram in the accompanying drawings illustrate the possible architecture, function and operation of the system, method and computer program product according to various embodiments of the present invention. In this regard, each box in the flow chart or block diagram can represent a module, a program segment or a part of a code, and the above-mentioned module, program segment or a part of a code contains one or more executable instructions for realizing the specified logical function. It should also be noted that in some alternative implementations, the functions marked in the box can also occur in a different order from the order marked in the accompanying drawings. For example, two boxes represented in succession can actually be executed substantially in parallel, and they can sometimes be executed in the opposite order, depending on the functions involved. It should also be noted that each box in the block diagram or flow chart, and the combination of the boxes in the block diagram or flow chart can be implemented with a dedicated hardware-based system that performs a specified function or operation, or can be implemented with a combination of dedicated hardware and computer instructions.

[0086] The units involved in the embodiments of the present invention may be implemented by software or hardware. The units described may also be set in a processor, for example, it may be described as: a processor includes a group division unit, a feature generation and classification unit, and a target user judgment unit. The names of these units do not constitute a limitation on the units themselves under certain circumstances. For example, the group division unit may also be described as a "unit that provides a target group to the feature generation and classification unit."

[0087] As another aspect, the present invention also provides a computer-readable medium, which may be included in the device described in the above embodiment; or may exist independently and not be assembled into the device. The above computer-readable medium carries one or more programs, and when the above one or more programs are executed by the device, the steps executed by the device include: determining the target group to which each item in the preset item set belongs in the predetermined item group; wherein the item group is determined according to the number of occurrences and / or the order of occurrence of multiple items in the same user behavior sequence; for each target group: obtaining multiple preliminary users of the target group, inputting the original feature data of each preliminary user for each item in the target group into a feature generation model that has been pre-trained and corresponds to the target group, and obtaining the optimized feature data of the preliminary user in the target group; inputting the optimized feature data into a user classification model that has been pre-trained and corresponds to the target group, and obtaining the classification result of the preliminary user in the target group; using the classification result of each preliminary user in each target group to determine whether the preliminary user is the target user corresponding to the item set.

[0088] In the technical solution of the embodiment of the present invention, an item graph is constructed in advance in an offline state according to the number of occurrences and / or the order of occurrence of items in the user behavior sequence, and the embedding vector of the item is determined from the item graph using a graph embedding method, and the items are clustered using the embedding vector to form multiple item groups; when it is necessary to determine the target user for the item set, the item group to which each item in the item set belongs can be determined by querying or judging similar items, and then the subsequent target user identification is performed according to the item group. Through the above settings, the association relationship between items of different categories or different brands can be accurately reflected, the inherent defects of target user identification based on item category or brand can be overcome, and it is conducive to the accurate identification of target users. At the same time, the item group can be dynamically updated according to the user behavior data, so as to be consistent with the dynamically changing association relationship between items based on user behavior. In addition, the present invention uses a deep learning method to automatically generate optimized feature data directly from the original feature data without performing ETL processing, thereby improving work efficiency and saving a lot of manpower and material resources.

[0089] The above specific implementations do not constitute a limitation on the protection scope of the present invention. It should be understood by those skilled in the art that various modifications, combinations, sub-combinations and substitutions may occur depending on design requirements and other factors. Any modification, equivalent substitution and improvement made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A method for determining a target user, characterized in that: include: Determine the target group to which each item in the preset item set belongs in the predetermined item group; wherein the item group is determined based on the number of occurrences and / or the order of occurrence of multiple items in the same user behavior sequence; For each target group: obtaining a plurality of preliminary users of the target group, inputting the original feature data of each preliminary user for each item in the target group into a feature generation model that has been pre-trained and corresponds to the target group, and obtaining the optimized feature data of the preliminary user in the target group; inputting the optimized feature data into a user classification model that has been pre-trained and corresponds to the target group, and obtaining the classification result of the preliminary user in the target group; and Using the classification result of each preliminary selected user in each target group, determining whether the preliminary selected user is a target user corresponding to the set of items; The obtaining of the optimized feature data of the preliminary selected user in the target group includes: inputting the original feature data of any preliminary selected user for each item in the corresponding target group into the feature generation model, obtaining a plurality of multi-dimensional data of the preliminary selected user output by the second activation layer in the feature generation model; dividing each initial dimension of the multi-dimensional data into at least one target dimension according to a numerical interval, and converting the multi-dimensional data into an optimized feature data of the preliminary selected user in the target dimension; wherein the value of the optimized feature data in any target dimension is: the number of times the value of the multi-dimensional data in the initial dimension corresponding to the target dimension is in the numerical interval corresponding to the target dimension; The feature generation model is a deep neural network model, including an input layer, a first hidden layer, a second hidden layer, and an output layer connected in sequence; wherein the first hidden layer includes a first perception layer, a first standardization layer, and a first activation layer; the second hidden layer includes a second perception layer, a second standardization layer, and a second activation layer; the input data of the second perception layer is the data obtained by concatenating the output data of the first activation layer and the original feature data; the input data of the output layer is the data obtained by concatenating the output data of the second activation layer and the original feature data.

2. The method according to claim 1, characterized in that The item group is determined by the following steps: Generate an item map for characterizing the association relationship between the multiple items according to the number of occurrences and / or the order of occurrence of the multiple items in the same user behavior sequence; Performing graph embedding processing on the item graph to obtain an embedding vector for each of the multiple items; as well as The multiple items are clustered using the embedding vectors to obtain multiple item groups.

3. The method according to claim 2, characterized in that The method further comprises: after generating the item map: Inputting the attribute data and title data of the isolated item in the item atlas into a pre-trained item similarity model to obtain at least one similar item of the isolated item; Adding the association relationship between the isolated item and the similar items in the item map.

4. The method according to claim 3, characterized in that The step of determining the target group to which each item in the preset item set belongs in the predetermined item group includes: For any item in the item set, determine whether the predetermined item group contains the identifier of the item: If so, the item group containing the identifier of the item is determined as the target group to which the item belongs; Otherwise, the attribute data and title data of the item are input into the item similarity model to obtain similar items of the item; the target group to which the item belongs is determined using the item group to which the similar items belong; and after determining the target group to which the item belongs, the item is added to the target group.

5. The method according to claim 1, characterized in that The original feature data of any preliminary selection user for any item in the corresponding target group includes: the attribute data of the preliminary selection user, the attribute data of the item, and the behavior record data of the preliminary selection user for the item.

6. The method according to claim 1, characterized in that The step of using the classification result of each preliminary selected user in each target group to determine whether the preliminary selected user is a target user corresponding to the set of items includes: The weighted discrimination result of each preliminary selected user is obtained by using the classification result of each preliminary selected user in each target group and the weight value pre-set for each target group; When the weighted discrimination result is greater than a preset threshold, the preliminary selected user is determined as a target user corresponding to the item set.

7. The method according to claim 1 or 5, characterized in that: The user classification model is an XGBoost model; and the feature generation model and the user classification model corresponding to the same item group are trained by the following steps: The feature generation model and the user classification model are connected to form a comprehensive model; wherein the input data of the user classification model is: the optimized feature data formed by converting the output data of the second activation layer in the feature generation model; Constructing a training sample set applied to the comprehensive model; wherein each training sample in the training sample set includes original feature data as a feature part and a user classification result as a label part; and The feature generation model and the user classification model are trained using the training sample set.

8. A device for determining a target user, characterized in that: include: A group division unit, used to determine the target group to which each item in the preset item set belongs in the predetermined item group; wherein the item group is determined based on the number of occurrences and / or the order of occurrence of multiple items in the same user behavior sequence; The feature generation and classification unit is used to: for each target group: obtain a plurality of preliminary selected users of the target group, input the original feature data of each preliminary selected user for each item in the target group into a feature generation model that has been pre-trained and corresponds to the target group, and obtain the optimized feature data of the preliminary selected user in the target group; input the optimized feature data into a user classification model that has been pre-trained and corresponds to the target group, and obtain the classification result of the preliminary selected user in the target group; a target user determination unit, configured to determine whether each preliminary selected user is a target user corresponding to the set of items by using the classification result of each preliminary selected user in each target group; The feature generation and classification unit is further used to: input the original feature data of any preliminary selected user for each item in the corresponding target group into the feature generation model, and obtain multiple multi-dimensional data of the preliminary selected user output by the second activation layer in the feature generation model; divide each initial dimension of the multi-dimensional data into at least one target dimension according to a numerical interval, and convert the multi-dimensional data into an optimized feature data of the preliminary selected user in the target dimension; wherein the value of the optimized feature data in any target dimension is: the number of times the multi-dimensional data is in the numerical interval corresponding to the target dimension at the value of the initial dimension corresponding to the target dimension; The feature generation model is a deep neural network model, including an input layer, a first hidden layer, a second hidden layer, and an output layer connected in sequence; wherein the first hidden layer includes a first perception layer, a first standardization layer, and a first activation layer; the second hidden layer includes a second perception layer, a second standardization layer, and a second activation layer; the input data of the second perception layer is the data obtained by concatenating the output data of the first activation layer and the original feature data; the input data of the output layer is the data obtained by concatenating the output data of the second activation layer and the original feature data.

9. The device according to claim 8, characterized in that The device further comprises: The group predetermination unit is used to generate an item graph for characterizing the association relationship between the multiple items according to the number of times and / or the order of occurrence of the multiple items in the same user behavior sequence; perform graph embedding processing on the item graph to obtain an embedding vector for each of the multiple items; and cluster the multiple items using the embedding vector to obtain multiple item groups.

10. An electronic device, characterized in that: include: one or more processors; a storage device for storing one or more programs, When the one or more programs are executed by the one or more processors, the one or more processors implement the method according to any one of claims 1 to 7.

11. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the method according to any one of claims 1 to 7 is implemented.

Images