Method and device for evaluating feature validity, storage medium and electronic equipment

Abstract

The embodiment of the invention discloses a method and device for evaluating feature validity, a storage medium and electronic equipment, and the method comprises the steps: constructing a feature validity evaluation graph according to basic features in a data set and combined features corresponding to the basic features; and broadcasting messages from the at least one basic feature node and the at least one combined feature node to the combined feature column node in the feature validity evaluation graph. Enabling the combined feature column node to calculate an information gain of the basic feature relative to the associated feature based on a basic information entropy corresponding to the basic feature and a conditional information entropy of the basic feature relative to the associated feature, and determining a feature validity evaluation result corresponding to the basic feature according to the information gain, the basic information entropy is calculated based on the first distribution information, and the conditional information entropy is calculated based on the second distribution information.

Description

Technical Field

[0001] The present invention relates to computer technology, and in particular to a method, device, storage medium and electronic device for evaluating feature validity. Background Art

[0002] In modern data sets, the number of features may be very large. High-dimensional data can enable the model to learn more information, but it may also contain many irrelevant or redundant features, which can lead to increased model training time and decreased performance. Removing irrelevant or inefficient features can simplify the model and improve the generalization ability and prediction performance of the model. When the model contains fewer features, it is usually easier to understand and explain. In addition, data visualization in low-dimensional space is also simpler and clearer. Summary of the invention

[0003] The purpose of the embodiments of this specification is to provide a method, device, storage medium and electronic device for evaluating feature validity.

[0004] The embodiment of this specification provides a method for evaluating the effectiveness of features. The feature effectiveness evaluation method based on graph computing is a feature computing solution for big data and high dimensions. It can not only complete complex feature computing of high dimensions, but also make a reasonable evaluation of the effectiveness of features, and screen out effective features for output to strategies and models. The solution converts the relational data structure into a graph structure, performs feature computing from the perspective of graph computing, converts the relational feature data of each dimension into a graph structure, and evaluates the effectiveness of features by message broadcasting. The evaluation process has no restrictions on the input feature table structure and data magnitude, and can be reused in different business scenarios. It has high versatility and computing efficiency. The solution constructs the feature set for the required feature effectiveness calculation in the same graph, and can complete the effectiveness evaluation of all features using one message broadcast. The computing process has high parallelism and high performance. The solution calculates information entropy and information gain of features in a graph computing manner, and can complete the effectiveness evaluation of features of any dimension. It has strong versatility. The solution pre-places the feature effectiveness evaluation of the model, deletes irrelevant or inefficient features, simplifies the model, and improves the generalization ability and prediction performance of the model. The method includes:

[0005] According to the basic features in the data set and the combined features corresponding to the basic features, a feature validity evaluation graph is constructed, wherein the combined features include the basic features and the associated features corresponding to the basic features, the feature validity evaluation graph includes at least one basic feature node, at least one combined feature node, at least one associated feature node, a basic feature column node, an associated feature column node and a combined feature column node, for an associated feature node, the combined feature value of each combined feature node connected to the associated feature node includes the associated feature value of the associated feature node, the basic feature column node is used to calculate the first distribution information of the basic feature based on the at least one basic feature node, and the associated feature column node is used to calculate the second distribution information of the combined feature based on the at least one associated feature node;

[0006] In the feature validity evaluation graph, by broadcasting messages from the at least one basic feature node and the at least one combined feature node to the combined feature column node, the combined feature column node calculates the information gain of the basic feature with respect to the associated feature based on the basic information entropy corresponding to the basic feature and the conditional information entropy of the basic feature relative to the associated feature, and determines the feature validity evaluation result corresponding to the basic feature according to the information gain, wherein the basic information entropy is calculated based on the first distribution information, and the conditional information entropy is calculated based on the second distribution information.

[0007] Furthermore, the at least one basic feature node is connected to the basic feature column node via an edge, and the edge between each basic feature node and the basic feature column node corresponds to a first weight value of the basic feature value of the basic feature node.

[0008] Furthermore, the first weight value includes the number of occurrences of the basic feature value of the basic feature node in the data set.

[0009] Furthermore, the method further comprises:

[0010] In the feature validity evaluation graph, the basic feature column node receives the first weight value of the edge between the basic feature node and the basic feature column node sent by each basic feature node connected to it, and calculates the first distribution information of the basic feature based on the first weight value, wherein the first distribution information includes at least one first weight value and a corresponding first weight sum.

[0011] Furthermore, the method further comprises any of the following:

[0012] In the feature validity evaluation graph, the basic feature column node calculates the basic information entropy corresponding to the basic feature based on the first distribution information, and the combined feature column node receives the basic information entropy sent by the basic feature column node;

[0013] In the feature validity evaluation graph, the combined feature column node receives the first distribution information sent by the basic feature column node, and calculates the basic information entropy corresponding to the basic feature based on the first distribution information.

[0014] Furthermore, the feature validity evaluation graph further includes a virtual basic feature column node, the basic feature column node is connected to the virtual basic feature column node via an edge, and the virtual basic feature column node is connected to the combined feature column node via an edge;

[0015] Wherein, the method further comprises:

[0016] In the feature validity evaluation graph, the virtual basic feature column node receives the first distribution information sent by the basic feature column node, and calculates the basic information entropy corresponding to the basic feature based on the first distribution information, so that the combined feature column node receives the basic information entropy sent by the virtual basic feature column node.

[0017] Furthermore, one or more of the at least one combined feature node are connected to one of the at least one associated feature node via an edge, and for an associated feature node, the edge between each combined feature node connected to the associated feature node and the associated feature node corresponds to the second weight value of the combined feature value of the combined feature node.

[0018] Furthermore, the second weight value includes the number of occurrences corresponding to the combined feature value of the combined feature node in the data set.

[0019] Furthermore, the method further comprises:

[0020] In the feature validity evaluation graph, each associated feature node receives a second weight value of an edge between the combined feature node and the associated feature node sent by each combined feature node connected thereto, and calculates third distribution information of the combined feature relative to the associated feature based on the second weight value, wherein the third distribution information includes one or more second weight values ​​and a corresponding second weight sum;

[0021] The associated feature column node receives the third distribution information corresponding to the associated feature value of the associated feature node sent by each associated feature node, and calculates the second distribution information of the combined feature based on the third distribution information, wherein the second distribution information includes one or more second weight values ​​corresponding to each associated feature node, the sum of the second weights corresponding to the associated feature node, and the sum of the second weights corresponding to the at least one associated feature node.

[0022] Furthermore, the method further comprises:

[0023] In the feature validity evaluation graph, the associated feature column node calculates the conditional information entropy of the basic feature relative to the associated feature based on the second distribution information, so that the combined feature column node receives the conditional information entropy sent by the associated feature column node;

[0024] In the feature validity evaluation graph, the combined feature column node receives the second distribution information sent by the associated feature column node, and calculates the conditional information entropy of the basic feature relative to the associated feature based on the second distribution information.

[0025] The embodiment of this specification also provides a device for evaluating feature validity, including:

[0026] A graph construction module, used to construct a feature validity evaluation graph according to basic features in a data set and combined features corresponding to the basic features, wherein the combined features include the basic features and associated features corresponding to the basic features, the feature validity evaluation graph includes at least one basic feature node, at least one combined feature node, at least one associated feature node, a basic feature column node, an associated feature column node and a combined feature column node, for an associated feature node, the combined feature value of each combined feature node connected to the associated feature node includes the associated feature value of the associated feature node, the basic feature column node is used to calculate the first distribution information of the basic feature based on the at least one basic feature node, and the associated feature column node is used to calculate the second distribution information of the combined feature based on the at least one associated feature node;

[0027] A feature validity evaluation module is used to broadcast messages from the at least one basic feature node and the at least one combined feature node to the combined feature column node in the feature validity evaluation graph, so that the combined feature column node calculates the information gain of the basic feature with respect to the associated feature based on the basic information entropy corresponding to the basic feature and the conditional information entropy of the basic feature relative to the associated feature, and determines the feature validity evaluation result corresponding to the basic feature according to the information gain, wherein the basic information entropy is calculated based on the first distribution information, and the conditional information entropy is calculated based on the second distribution information.

[0028] The embodiments of the present specification also provide a storage medium, wherein the storage medium stores a computer program, and the computer program is suitable for being loaded by a processor and executing the steps of the above method.

[0029] An embodiment of the present specification also provides an electronic device, comprising: a processor and a memory; wherein the memory stores a computer program, and the computer program is suitable for being loaded by the processor and executing the steps of the above method.

[0030] In an embodiment of the present specification, a method for evaluating the effectiveness of a feature is proposed. The feature effectiveness evaluation method based on graph computing is a feature computing solution for big data and high dimensions. It can not only complete complex feature calculations in high dimensions, but also make reasonable evaluations of the effectiveness of features, and screen out effective features for output to strategies and models. The scheme converts the relational data structure into a graph structure, performs feature calculations from the perspective of graph computing, and converts the relational feature data of each dimension into a graph structure, and performs feature effectiveness evaluation by message broadcasting. The evaluation process has no restrictions on the input feature table structure and data magnitude, and can be reused in different business scenarios, with high versatility and computing efficiency. The scheme constructs the feature set for the required feature effectiveness calculation in the same graph, and can complete the effectiveness evaluation of all features using one message broadcast. The computing process has high parallelism and high performance. The scheme calculates information entropy and information gain of features in a graph computing manner, and can complete the effectiveness evaluation of features of any dimension, with strong versatility. The scheme pre-places the feature effectiveness evaluation of the model, deletes irrelevant or inefficient features, simplifies the model, and improves the generalization ability and prediction performance of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 A flowchart of a method for evaluating feature effectiveness provided in an embodiment of this specification.

[0032] Figure 2 A schematic diagram of the graph structure of an example feature effectiveness evaluation graph provided in an embodiment of this specification.

[0033] Figure 3 A schematic diagram of the structure of a device for evaluating feature validity provided in an embodiment of this specification.

[0034] Figure 4 A schematic diagram of the structure of an electronic device provided in an embodiment of this specification. DETAILED DESCRIPTION

[0035] In order to make the purpose, technical solutions and advantages of this specification more clear, the technical solutions of this specification will be clearly and completely described below in combination with the specific embodiments of this specification and the corresponding drawings. Obviously, the described embodiments are only part of the embodiments of this specification, not all of them. Based on the embodiments in this specification, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this specification.

[0036] See also Figure 1 , is a flow chart of a method for evaluating feature validity provided in an embodiment of this specification. In an embodiment of this specification, the method for evaluating feature validity is applied to a device for evaluating feature validity (hereinafter referred to as a "feature validity evaluation device") or an electronic device equipped with a feature validity evaluation device. Figure 1 The process shown is described in detail, and the method for evaluating the effectiveness of a feature may specifically include the following steps:

[0037] S102, constructing a feature validity evaluation graph according to basic features in the data set and combined features corresponding to the basic features, wherein the combined features include the basic features and associated features corresponding to the basic features, and the feature validity evaluation graph includes at least one basic feature node, at least one combined feature node, at least one associated feature node, a basic feature column node, an associated feature column node, and a combined feature column node. For an associated feature node, the combined feature value of each combined feature node connected to the associated feature node includes the associated feature value of the associated feature node. The basic feature column node is used to calculate the first distribution information of the basic feature based on the at least one basic feature node, and the associated feature column node is used to calculate the second distribution information of the combined feature based on the at least one associated feature node.

[0038] In some embodiments, this exemplary embodiment does not specifically limit the number of features, feature types, and feature content in the data set. In some embodiments, a basic feature refers to a single feature, such as age, and the data set includes multiple basic features, each of which includes a basic feature type (e.g., age) and one or more basic feature values ​​in the data set that belong to the basic feature type (e.g., 17 years old, 18 years old, 19 years old, etc.).

[0039] In some embodiments, for a basic feature in a data set, the data set includes at least one combination feature corresponding to the basic feature, each combination feature includes the basic feature (e.g., age) and an associated feature corresponding to the basic feature (e.g., user type), and each associated feature includes an associated feature type (e.g., user type) and one or more associated feature data (e.g., 0, 1, 2, etc.) corresponding to one or more basic feature values ​​of the basic feature in the data set and belonging to the associated feature type. In some embodiments, each combination feature includes a combined feature type (e.g., age-user type) and one or more combinations and feature values ​​belonging to the combined feature type in the data set (e.g., 17 years old-1, 18 years old-1, 19 years old-1, 18 years old-2, 19 years old-0, etc.).

[0040] In some embodiments, a feature validity evaluation graph is constructed based on each basic feature in the data set and each combination feature corresponding to the basic feature. Each combination feature includes the basic feature and an associated feature corresponding to the basic feature, that is, the association relationship between the various features in the data set is constructed as a graph associated by points and edges. In some embodiments, at least one basic feature node corresponding to each basic feature is created in the feature validity evaluation graph based on one or more basic feature values ​​of each basic feature. Each basic feature node corresponds to one of the basic feature values ​​of the basic feature. For example, for the basic feature "age", the basic feature values ​​of the basic feature include 17 years old, 18 years old, 19 years old, 18 years old, and 19 years old. Then, three basic feature nodes are created in the feature validity evaluation graph. The three basic feature nodes correspond to one of the three basic feature values ​​17 years old, 18 years old, and 19 years old.

[0041] In some embodiments, for each basic feature, at least one combined feature node corresponding to the combined feature is created in the feature validity evaluation graph according to one or more combined feature values ​​of each combined feature corresponding to the basic feature, and each combined feature node corresponds to one of the combined feature values ​​of the combined feature. For example, for the basic feature "age", the combined feature corresponding to the basic feature includes "age-user type", and the combined feature includes the basic feature "age" and the associated feature "user type" corresponding to the basic feature. The combined feature values ​​of the combined feature include 17 years old-1, 17 years old-1, 18 years old-1, 19 years old-1, 18 years old-2, 18 years old-2, 19 years old-0, and 19 years old-0. Then, five combined feature nodes are created in the feature validity evaluation graph, and the five combined feature nodes correspond to one of the five combined feature values ​​17 years old-1, 18 years old-1, 19 years old-1, 18 years old-2, and 19 years old-0, respectively.

[0042] In some embodiments, for each basic feature, a basic feature column node corresponding to the basic feature is created in the feature validity evaluation graph, and each basic feature node corresponding to the basic feature is directly connected to the basic feature column node corresponding to the basic feature through an edge. The basic feature column node corresponding to each basic feature is used to calculate the first distribution information of the basic feature (i.e., the distribution statistical information of the aggregated basic features) based on at least one basic feature node corresponding to the basic feature that is directly connected to the basic feature column node through an edge.

[0043] In some embodiments, for each basic feature, according to one or more associated feature values ​​of the associated feature corresponding to the basic feature in one or more combined feature values ​​of each combined feature corresponding to the basic feature, at least one associated feature node corresponding to the combined feature is created in the feature validity evaluation graph, each associated feature node corresponds to one of the associated feature values ​​of the associated feature, and for each associated feature node, one or more combined feature nodes in at least one combined feature node corresponding to the combined feature are directly connected to the associated feature node through an edge, and the combined feature value corresponding to each combined feature node in the one or more combined feature nodes includes the associated feature value corresponding to the associated feature node. For example, for the basic feature "age", the combined feature corresponding to the basic feature includes "age-user type", and the combined feature values ​​of the combined feature include 17 years old-1, 18 years old-1, 19 years old-1, 18 years old-2, and 19 years old-0. The combined feature includes the basic feature "age" and the associated feature corresponding to the basic feature is "user type", and the associated feature values ​​of the associated feature "user type" in the combined feature include 0, 1, 1, 1, and 2. Then three associated feature nodes are created in the feature validity evaluation graph. The three associated feature nodes correspond to one of the three associated feature values ​​of 0, 1, and 2, respectively. The three combined feature nodes "17 years old-1", "18 years old-1", and "19 years old-1" are directly connected to the associated feature node "1" through edges, the combined feature node "18 years old-2" is directly connected to the associated feature node "2" through an edge, and the combined feature node "19 years old-0" is directly connected to the associated feature node "0" through an edge.

[0044] In some embodiments, for each combination feature corresponding to each basic feature, an associated feature column node corresponding to the combination feature is created in the feature validity evaluation graph, and each associated feature node corresponding to the combination feature is directly connected to the associated feature column node corresponding to the combination feature through an edge. The associated feature column node corresponding to each combination feature is used to calculate the second distribution information of the combination feature (i.e., the distribution statistical information of the aggregated combination feature) based on at least one associated feature node corresponding to the combination feature that is directly connected to the associated feature column node through an edge. It can also be referred to as calculating the second distribution information of the associated features in the combination feature (i.e., the distribution statistical information of the aggregated associated features).

[0045] In some embodiments, for each combination feature corresponding to each basic feature, a combination feature column node corresponding to the combination feature is created in the feature validity evaluation graph, and the basic feature column node corresponding to the basic feature and the associated feature column node corresponding to the combination feature are respectively directly connected to the combination feature column node corresponding to the combination feature through edges, and the combination feature column node corresponding to each combination feature is used to calculate the information gain of the basic feature regarding the associated feature based on the basic information entropy corresponding to the basic feature calculated according to the first distribution information of the basic feature obtained through the basic feature column node directly connected to the combination feature column node through an edge, and the conditional information entropy of the basic feature relative to the associated feature in the combination feature calculated according to the second distribution information of the combination feature obtained through the associated feature column node directly connected to the combination feature column node through an edge.

[0046] In step S104, in the feature validity evaluation graph, by broadcasting messages from the at least one basic feature node and the at least one combined feature node to the combined feature column node, the combined feature column node calculates the information gain of the basic feature with respect to the associated feature based on the basic information entropy corresponding to the basic feature and the conditional information entropy of the basic feature relative to the associated feature, and determines the feature validity evaluation result corresponding to the basic feature according to the information gain, wherein the basic information entropy is calculated based on the first distribution information, and the conditional information entropy is calculated based on the second distribution information.

[0047] In some embodiments, in the feature validity evaluation graph, for each combination feature corresponding to each basic feature, message sending starts from the starting node (at least one basic feature node corresponding to the basic feature, at least one combination feature node corresponding to the combination feature), and message convergence is performed in each node in the graph along the message propagation path (i.e., basic feature node->basic feature column node->path composed of multiple edges corresponding to combination feature column node, combination feature node->associated feature node->associated feature column node->path composed of multiple edges corresponding to combination feature column node), completing a message broadcasting process with the basic feature node and the combination feature node as the starting nodes. In some embodiments, in the feature validity evaluation graph, for each basic feature, each basic feature node in at least one basic feature node corresponding to the basic feature sends a specified message (for example, the basic feature value, or the weight of the edge between the basic feature node and the corresponding basic feature column node) to the basic feature column node corresponding to the basic feature by means of message broadcasting, so that the basic feature column node aggregates the received messages, calculates the first distribution information of the basic feature, and then makes the basic feature column node send the first distribution information to the combined feature column node corresponding to each combination feature corresponding to the basic feature by means of message broadcasting, so that the combined feature column node calculates the basic information entropy corresponding to the basic feature according to the received first distribution information, or makes the basic feature column node calculate the basic information entropy corresponding to the basic feature according to the first distribution information, and then sends the basic information entropy to the combined feature column node corresponding to each combination feature corresponding to the basic feature by means of message broadcasting.In some embodiments, for each combination feature corresponding to each basic feature, each combination feature node in at least one combination feature node corresponding to the combination feature sends a specified message (for example, a combination feature value, or the weight of the edge between the combination qualifying feature node and the corresponding associated feature node) to an associated feature node directly connected to the combination feature node by means of message broadcasting, so that each associated feature node calculates relative distribution information of the combination feature values ​​of all combination feature nodes directly connected to the associated feature node relative to the associated feature value of the associated feature node according to the received message, and sends the relative distribution information to the associated feature column node corresponding to the combination feature by means of message broadcasting, so that the associated feature column node aggregates the received messages, calculates the second distribution information of the combination feature, and then sends the second distribution information to the combination feature column node corresponding to the combination feature by means of message broadcasting, so that the combination feature node calculates the conditional information entropy of the basic feature relative to the associated feature in the combination feature based on the second distribution information, or, the associated feature column node directly calculates the conditional information entropy of the basic feature relative to the associated feature in the combination feature according to the second distribution information, and sends the conditional information entropy to the combination feature column node corresponding to the combination feature by means of message broadcasting. In some embodiments, in the feature effectiveness evaluation graph, for each combination feature corresponding to each basic feature, the combination feature column node corresponding to the combination feature is made to calculate the basic feature with respect to the combination feature based on the basic information entropy corresponding to the basic feature (the basic information entropy may be calculated by the combination feature column node, or may be calculated by a basic feature column node directly connected to the combination feature column node and sent to the combination feature column node by message broadcasting), the conditional information entropy of the basic feature relative to the associated feature in the combination feature (the conditional information entropy may be calculated by the combination feature column node, or may be calculated by an associated feature column node directly connected to the combination feature column node and sent to the combination feature column node by message broadcasting). The information gain of the associated features in the combined features, for example, can be obtained by calculating the difference between the basic information entropy and the conditional information entropy. For example, the basic information entropy corresponding to a certain basic feature is H(D), and the conditional information entropy of the associated features in a certain combined feature corresponding to the basic feature is H(D|A), then the information gain of the basic feature with respect to the associated features is G(D,A)=H(D)-H(D|A), where information entropy is a measure of the uncertainty of a random variable. The more uniform the distribution, the greater the entropy and the higher the uncertainty. Information gain refers to the degree to which the uncertainty of a random variable is reduced when an additional feature is known. It is usually used to determine a better feature, that is, to segment the data so that the sub-dataset has a lower entropy (more ordered) than the original data set.

[0048] In some embodiments, for each basic feature, the information gain of the basic feature with respect to the associated features in various combined features is calculated based on the combined feature column nodes corresponding to the various combined features corresponding to the basic feature, and then the feature validity evaluation result corresponding to the basic feature is determined based on the information gain of the basic feature with respect to the various associated features. The feature validity evaluation result refers to the evaluation result obtained by evaluating the basic feature values ​​in which one or more combined features corresponding to the basic feature are taken as the effective features of the basic feature. For example, if the information gain of the basic feature with respect to the associated features in a certain combined feature is greater than or equal to a preset threshold, it can be determined that the basic feature value in the combined feature is the effective feature of the basic feature. For another example, a preset number of combined features with the largest information gain of the basic feature with respect to the corresponding associated features are selected from all the combined features corresponding to the basic feature, and the basic features in the preset number of combined features are taken as the effective features of the basic feature.

[0049] The embodiment of this specification proposes a method for evaluating the effectiveness of a feature. The feature effectiveness evaluation method based on graph computing is a feature computing solution for big data and high dimensions. It can not only complete complex feature calculations in high dimensions, but also make reasonable evaluations of the effectiveness of features, and screen out effective features for output to strategies and models. The solution converts the relational data structure into a graph structure, performs feature calculations from the perspective of graph computing, converts the relational feature data of each dimension into a graph structure, and performs feature effectiveness evaluation by message broadcasting. The evaluation process has no restrictions on the input feature table structure and data magnitude, and can be reused in different business scenarios. It has high versatility and computing efficiency. The solution constructs the feature set for the required feature effectiveness calculation in the same graph, and can complete the effectiveness evaluation of all features using one message broadcast. The computing process has high parallelism and high performance. The solution calculates information entropy and information gain of features in a graph computing manner, and can complete the effectiveness evaluation of features of any dimension. It has strong versatility. The solution pre-places the feature effectiveness evaluation of the model, deletes irrelevant or inefficient features, simplifies the model, and improves the generalization ability and prediction performance of the model.

[0050] In some embodiments, the at least one basic feature node is connected to the basic feature column node by an edge, and the edge between each basic feature node and the basic feature column node corresponds to the first weight value of the basic feature value of the basic feature node. In some embodiments, in the feature validity evaluation graph, for each basic feature, at least one basic feature point node corresponding to the basic feature is connected to the basic feature column node corresponding to the basic feature by an edge, and the direction of the edge is from the basic feature node to the basic feature column node, and the edge between each basic feature node in the at least one basic feature node and the basic feature column node corresponds to the first weight value of the basic feature value of the basic feature node, wherein the first weight value may be set by the training personnel corresponding to the data set, or may also be obtained after data analysis of the basic feature value, for example, the basic feature value is input into a trained weight model to obtain the first weight value corresponding to the basic feature value output by the weight model. It should be noted that the above method of determining the first weight value is only for example, not for limitation. Those skilled in the art should understand that any method of determining the first weight value can be included in the protection scope of this specification.

[0051] In some embodiments, the first weight value includes the number of occurrences of the basic feature value of the basic feature node in the data set. In some embodiments, the first weight value corresponding to the edge between a basic feature node and a basic feature column node directly connected to it is the number of occurrences of the basic feature value of the basic feature node in the data set. For example, if the basic feature value appears 3 times in the data set, the corresponding first weight value is 3.

[0052] In some embodiments, the method further includes: in the feature validity evaluation graph, the basic feature column node receives the first weight value of the edge between the basic feature node and the basic feature column node sent by each basic feature node connected to it, and calculates the first distribution information of the basic feature based on the first weight value, wherein the first distribution information includes at least one first weight value and a corresponding first weight sum. In some embodiments, in the feature validity evaluation graph, for each basic feature, each basic feature node corresponding to the basic feature sends the first weight value of the edge between it and the basic feature column node directly connected to the basic feature node corresponding to the basic feature node to the basic feature column node through message broadcasting along the outgoing direction, so that the basic feature node calculates the first distribution information of the basic feature according to the at least one first weight value it receives, and the first distribution information includes the at least one first weight value and the sum of weights corresponding to the at least one first weight value. This example embodiment does not specifically limit the data structure of the first distribution information. For example, the first distribution information can be a Map (a key-value pair set) structure, such as Map[[0, sum(weight)], [weight1, weight2...]], wherein the key (key) of the Map is an array (such as Tuple2, a binary array), the value (value) of the Map is a list (such as List), and sum(weight) refers to the sum of weights corresponding to multiple first weight values ​​including weight 1, weight 2, etc. For example, the first weight values ​​are 2, 2, and 1, respectively, and the corresponding first weight sum is 5, then the corresponding first distribution information is Map[[0, 5], [2, 2, 1]].

[0053] In some embodiments, the method further includes any of the following: in the feature validity evaluation graph, the basic feature column node calculates the basic information entropy corresponding to the basic feature based on the first distribution information, so that the combined feature column node receives the basic information entropy sent by the basic feature column node; in the feature validity evaluation graph, the combined feature column node receives the first distribution information sent by the basic feature column node, and calculates the basic information entropy corresponding to the basic feature based on the first distribution information. In some embodiments, in the feature validity evaluation graph, for each basic feature, the basic feature column node corresponding to the basic feature will calculate the basic information entropy corresponding to the basic feature based on the first distribution information of the basic feature, and then send the basic information entropy to the combined feature column node corresponding to each combined feature corresponding to the basic feature through message broadcasting, wherein the specific calculation formula of the basic information entropy is as follows:

[0054]

[0055] Among them, H(D) refers to the basic information entropy corresponding to the basic feature, k = 1, 2, ... K, K refers to the number of basic feature nodes corresponding to the basic feature, C k Refers to the first weight values ​​corresponding to each basic feature node corresponding to the basic feature, and D refers to the sum of the first weights corresponding to all basic feature nodes corresponding to the basic feature. For example, the first distribution information is Map[[0, 5], [2, 2, 1]], the first weight values ​​are 2, 2, 1 respectively, and the sum of the first weights is 5, then the corresponding basic information entropy = -2 / 5*log2 / 5-2 / 5*log2 / 5-1 / 5*log1 / 5. In some embodiments, the first distribution information of the basic feature can also be sent by the basic feature column node corresponding to the basic feature to the combination feature column node corresponding to each combination feature corresponding to the basic feature by means of message broadcasting, and the combination feature column node calculates the basic information entropy corresponding to the basic feature based on the first distribution information it receives.

[0056] In some embodiments, the feature validity evaluation graph further includes a virtual basic feature column node, the basic feature column node is connected to the virtual basic feature column node through an edge, and the virtual basic feature column node is connected to the combined feature column node through an edge; wherein, the method further includes: in the feature validity evaluation graph, the virtual basic feature column node receives the first distribution information sent by the basic feature column node, calculates the basic information entropy corresponding to the basic feature based on the first distribution information, and enables the combined feature column node to receive the basic information entropy sent by the virtual basic feature column node. In some embodiments, for each basic feature, the feature validity evaluation graph further includes a virtual basic feature column node corresponding to the basic feature, the basic feature column node corresponding to the basic feature is directly linked to the virtual basic feature column node through an edge, the direction of the edge is from the basic feature column node to the virtual basic feature column node, the virtual basic feature column node is directly connected to the combined feature column node corresponding to each combined feature corresponding to the basic feature through an edge, and the direction of the edge is from the virtual basic feature column node to the combined feature column node. In some embodiments, for each basic feature, the basic feature column node corresponding to the basic feature will send the first distribution information of the basic feature to the virtual basic feature column node corresponding to the basic feature (i.e., the virtual basic feature column node directly connected to the basic feature column node) by means of message broadcasting. The virtual basic feature column node will calculate the basic information entropy corresponding to the basic feature based on the first distribution information. The calculation method of the basic information entropy has been described in detail above and will not be repeated here. In some embodiments, the virtual basic feature column node will send the calculated basic information entropy to the combined feature column node corresponding to each combined feature corresponding to the basic feature (i.e., the combined feature column node directly connected to the virtual basic feature column node) by means of message broadcasting.

[0057] In some embodiments, one or more of the at least one combined feature node is connected to one of the at least one associated feature node through an edge, and for an associated feature node, the edge between each combined feature node connected to the associated feature node and the associated feature node corresponds to the second weight value of the combined feature value of the combined feature node. In some embodiments, for each combined feature corresponding to each basic feature, at least one combined feature node corresponding to the combined feature is created in the feature validity evaluation graph, each combined feature node corresponds to one of the combined feature values ​​of the combined feature, and according to one or more associated feature values ​​of the associated feature corresponding to the basic feature in the one or more combined feature values ​​of each combined feature corresponding to the basic feature, at least one associated feature node corresponding to the combined feature is created in the feature validity evaluation graph, each associated feature node corresponds to one of the associated feature values ​​of the associated feature, and for each associated feature node, one or more of the at least one combined feature node corresponding to the combined feature is directly connected to the associated feature node through an edge, and the combined feature value corresponding to each of the one or more combined feature nodes includes the associated feature value corresponding to the associated feature node, and the direction of the edge is from the combined feature node to the associated feature node. In some embodiments, for an associated feature node, each combined feature node directly connected to the associated feature node through an edge and the edge between the associated feature node corresponds to a second weight value of the combined feature value of the combined feature node, wherein the second weight value may be set by a trainer corresponding to the data set, or may be obtained by performing data analysis on the combined feature value, for example, inputting the combined feature value into a trained weight model to obtain the second weight value corresponding to the combined feature value output by the weight model. It should be noted that the above-mentioned method of determining the second weight value is only an example and not a limitation. Those skilled in the art should understand that any method of determining the second weight value may be included in the scope of protection of this specification.

[0058] In some embodiments, the second weight value includes the number of occurrences of the combined feature value of the combined feature node in the data set. In some embodiments, the second weight value corresponding to the edge between a combined feature node and its directly connected associated feature node is the number of occurrences of the combined feature value of the combined feature node in the data set. For example, if the combined feature value appears 2 times in the data set, the corresponding second weight value is 2.

[0059] In some embodiments, the method further includes: in the feature validity evaluation graph, each associated feature node receives the second weight value of the edge between the combined feature node and the associated feature node sent by each combined feature node connected to it, and calculates the third distribution information of the combined feature relative to the associated feature based on the second weight value, wherein the third distribution information includes one or more second weight values ​​and the corresponding second weight sum; and the associated feature column node receives the third distribution information corresponding to the associated feature value of the associated feature node sent by each associated feature node, and calculates the second distribution information of the combined feature based on the third distribution information, wherein the second distribution information includes one or more second weight values ​​corresponding to each associated feature node, the second weight sum corresponding to the associated feature node, and the second weight sum corresponding to the at least one associated feature node. In some embodiments, in the feature validity evaluation graph, for each associated feature node in at least one associated feature node corresponding to the associated feature in the combined feature corresponding to each basic feature, each combined feature node directly connected to the associated feature node through an edge will send the second weight value of the edge between the combined feature node and the associated feature node to the associated feature node by message broadcasting along the outgoing direction, and the associated feature node will calculate the third distribution information of one or more combined feature values ​​of the one or more combined feature nodes relative to the associated feature value of the associated feature node based on the one or more second weight values ​​received by the one or more combined feature nodes directly connected to it, the third distribution information including the one or more second weight values ​​and the sum of weights corresponding to the one or more second weight values. This example embodiment does not specifically limit the data structure of the third distribution information. For example, the third distribution information can be a Map (a key-value pair set) structure, such as Map[[0, sum(weight)], [weight1, weight2...]], wherein sum(weight) refers to the sum of weights corresponding to multiple second weight values ​​including weight1, weight2, etc. For example, the second weight values ​​are 1, 1, 1, respectively, and the corresponding first weight sum is 3, then the corresponding first distribution information is Map[[0, 3], [1, 1, 1]].In some embodiments, in the feature validity evaluation graph, for the combined feature corresponding to each basic feature, the associated feature node in at least one associated feature node corresponding to the associated feature in the combined feature will send the third distribution information calculated by it to the associated feature column node directly connected to the associated feature node (that is, the associated feature column node corresponding to the combined feature), so that the associated feature column node aggregates and counts the one or more third distribution information it receives, and calculates the second distribution information of the combined feature, the second distribution information including one or more second weight values ​​corresponding to each associated feature node in the at least one associated feature node (that is, one or more second weight values ​​corresponding to the edges between one or more combined feature nodes directly connected to each associated feature node and the associated feature node), the second weight sum of the one or more second weight values ​​corresponding to each associated feature node, and the second weight sum of all second weight values ​​corresponding to the at least one associated feature node (that is, the second weight sum of the one or more second weight values ​​corresponding to each associated feature node is summed). This exemplary embodiment does not specifically limit the data structure of the second distribution information. For example, the second distribution information may be It is a Map (a key-value pair collection) structure. For example, the third distribution information corresponding to the associated feature node P1 is Map[[0, 3], [1, 1, 1]] (corresponding to three second weight values, 1, 1, 1, and the corresponding second weight sum is 3), the third distribution information corresponding to the associated feature node P2 is Map[[0, 1], [1]] (corresponding to a second weight value 1, and the corresponding second weight sum is 1), and the third distribution information corresponding to the associated feature node P3 is Map[[0, 1], [1]] (corresponding to a second weight value 1, and the corresponding second weight sum is 1 ), the corresponding second distribution information is Map[[[3,5],[1,1,1]], [[1,5],[1]], [[1,5],[1]]], including the second weight value [1,1,1] corresponding to the associated feature node P1, the second weight value [1] corresponding to the associated feature node P2, the second weight value [1] corresponding to the associated feature node P3, the second weight sum 3 corresponding to the associated feature node P1, the second weight sum 1 corresponding to the associated feature node P2, the second weight sum 1 corresponding to the associated feature node P3, and the second weight sum 5 corresponding to these three associated feature nodes.

[0060] In some embodiments, the method further includes: in the feature validity evaluation graph, the associated feature column node calculates the conditional information entropy of the basic feature relative to the associated feature based on the second distribution information, so that the combined feature column node receives the conditional information entropy sent by the associated feature column node; in the feature validity evaluation graph, the combined feature column node receives the second distribution information sent by the associated feature column node, and calculates the conditional information entropy of the basic feature relative to the associated feature based on the second distribution information. In some embodiments, in the feature validity evaluation graph, for each combined feature corresponding to each basic feature, the associated feature column node corresponding to the combined feature will calculate the conditional information entropy of the basic feature relative to the associated feature in the combined feature with the second distribution information of the combined feature, and then send the conditional information entropy to the combined feature column node corresponding to the combined feature by message broadcasting, wherein the specific calculation formula of the conditional information entropy is as follows:

[0061]

[0062] Among them, H(D|A) refers to the conditional information entropy of the basic feature relative to the associated feature in a certain combination feature corresponding to the basic feature, i = 1, 2, ... n, n refers to the number of nodes of the associated feature node corresponding to the combination feature, D i It refers to the sum of the second weights of each associated feature node corresponding to the combined feature (i.e., the sum of the second weights of all combined feature nodes directly connected to each associated feature node), D refers to the sum of the second weights of all associated feature nodes corresponding to the combined feature (i.e., the sum of the second weights of all combined feature nodes directly connected to all associated feature nodes), H(D i ) refers to the basic information entropy corresponding to each associated feature node, k = 1, 2, ... K, K refers to the number of combined feature nodes directly connected to each associated feature node corresponding to the combined feature, D ikIt refers to the second weight values ​​corresponding to each associated feature node and each combined feature node directly connected to the associated feature node. For example, the second distribution information is Map[[[3,5],[1,1,1]], [[1,5],[1]], [[1,5],[1]]], that is, the second weight values ​​corresponding to each associated feature node are [1, 1, 1], [1], [1] respectively, and the sum of the second weights corresponding to each associated feature node is 3, 1, and 1 respectively. The sum of the second weights corresponding to these three associated feature nodes is 5, then the corresponding conditional information entropy = 3 / 5*(-1 / 3*log1 / 3-1 / 3*log1 / 3-1 / 3*log1 / 3)+1 / 5*(-1*log1)+1 / 5*(-1*log1)=-3 / 5*log1 / 3-1 / 5*log1-1 / 5*log1. In some embodiments, the associated feature column node may also send the second distribution information of the combined feature to the combined feature column node corresponding to the combined feature by means of message broadcasting, and the combined feature column node may calculate the conditional information entropy of the basic feature relative to the associated feature in the combined feature based on the second distribution information it receives.

[0063] Figure 2 A schematic diagram of the graph structure of an example feature effectiveness evaluation graph provided in an embodiment of this specification.

[0064] like Figure 2As shown, the feature validity evaluation graph includes three basic feature nodes corresponding to the basic feature (age), and the basic feature values ​​corresponding to these three basic feature nodes are 17, 18, and 19 respectively. The number of occurrences of these three basic feature values ​​in the data set is 1 time, 2 times, and 2 times respectively. The feature validity evaluation graph also includes a basic feature column node age corresponding to the basic feature. The three basic feature nodes are directly connected to the basic feature column node through edges, and the direction of the edge is from the basic feature node to the basic feature column node. The weights of the edges between the three basic feature nodes and the basic feature column node are 1, 2, and 2 respectively. The three basic feature nodes send the weights of the edges between them and the basic feature column node to the basic feature column node through message broadcasting along the outgoing direction. The basic feature column node aggregates the distribution statistics of the basic features to obtain the feature distribution of the basic features [0,5][1,2,2], the feature validity evaluation graph also includes a virtual basic feature column node age-virtual corresponding to the basic feature, the basic feature column node and the virtual basic feature column node are directly connected by an edge, the direction of the edge is from the basic feature column node to the virtual basic feature column node, the basic feature column node sends the feature distribution corresponding to the basic feature to the virtual basic feature column node along the outgoing direction by message broadcasting, the virtual basic feature column node calculates the information entropy of the basic feature according to the feature distribution, and the feature validity evaluation graph also includes a combined feature column node (i.e., a mixed feature column node) age-user_type, the virtual basic feature column node and the The nodes of the combined feature column are directly connected by edges, and the direction of the edge is from the virtual basic feature column node to the combined feature column node. The virtual basic feature column node sends the information entropy of the basic feature to the combined feature column node by message broadcasting along the outgoing direction. The feature validity evaluation graph also includes five combined feature nodes corresponding to the combined feature (age-user_type) corresponding to the basic feature. The combined feature includes the basic feature (age) and the associated feature (user_type) corresponding to the basic feature. The combined feature values ​​corresponding to these five combined feature nodes are 17-1, 18-1, 19-1, 18-2, 19-0, the number of occurrences of these five combined feature values ​​in the data set is 1. The feature validity evaluation graph also includes three associated feature nodes corresponding to the combined feature. The associated feature values ​​corresponding to these three associated feature nodes are 1, 2, and 0, respectively. The three combined feature nodes corresponding to the combined feature values ​​17-1, 18-1, and 19-1 are directly connected to the associated feature node corresponding to the associated feature value 1 through an edge. The combined feature value corresponding to the combined feature value 18-2 is directly connected to the associated feature node corresponding to the associated feature value 2 through an edge. The combined feature value corresponding to the combined feature value 19-0 is directly connected to the associated feature node corresponding to the associated feature value 0 through an edge. The direction of the edge is from the combined feature node to the associated feature node. The weights of the edges between these five combined feature nodes and the corresponding associated feature nodes are all 1. These five combined feature nodes send the weights of the edges between themselves and the corresponding associated feature nodes to the corresponding associated feature nodes by means of message broadcasting along the outgoing direction. Each associated feature node calculates the feature distribution of the combined feature values ​​of each combined feature node directly connected to the associated feature node relative to the associated feature value of the associated feature node based on the received weight. The feature distributions calculated by these three associated feature nodes are [0,3][1,1,1], [0,1][1], [0,1][1], the feature validity evaluation graph also includes the associated feature column node user_type corresponding to the combined feature. The three associated feature nodes are directly connected to the associated feature column node through edges. The direction of the edge is from the associated feature node to the associated feature column node. The associated feature column node aggregates the distribution statistics of the associated features to obtain the corresponding associated feature distribution (also known as the feature distribution of the combined feature) [[3,5][1,1,1], [1,5][1], [1,5][1]], and the associated feature column node is directly connected to the combined feature column node through edges. The direction of the edge is from the associated feature node to the associated feature column node. The associated feature column node points to the combined feature column node, which calculates the conditional information entropy of the basic feature relative to the associated feature in the combined feature according to the distribution of the associated feature, and sends the conditional information entropy to the combined feature column node by message broadcasting along the outbound direction. The combined feature column node aggregates the information entropy H(D) of the basic feature and the conditional information entropy H(D|A) of the basic feature relative to the associated feature, and calculates the difference (H(D)-H(D|A)) between the information entropy H(D) of the basic feature and the conditional information entropy H(D|A) to complete the calculation of the information gain of the basic feature with respect to the associated feature. ,

[0065] Figure 3 A schematic diagram of a structure of a device for evaluating feature effectiveness provided in an embodiment of the present specification, the device for evaluating feature effectiveness (hereinafter referred to as "feature effectiveness evaluation device 1") can be implemented as all or part of an electronic device through software, hardware, or a combination of both. According to some embodiments, the feature effectiveness evaluation device 1 includes a graph construction module 11 and a feature effectiveness evaluation module 12.

[0066] A graph construction module 11 is used to construct a feature validity evaluation graph according to basic features in a data set and combined features corresponding to the basic features, wherein the combined features include the basic features and the associated features corresponding to the basic features, and the feature validity evaluation graph includes at least one basic feature node, at least one combined feature node, at least one associated feature node, a basic feature column node, an associated feature column node, and a combined feature column node. For an associated feature node, the combined feature value of each combined feature node connected to the associated feature node includes the associated feature value of the associated feature node. The basic feature column node is used to calculate the first distribution information of the basic feature based on the at least one basic feature node, and the associated feature column node is used to calculate the second distribution information of the combined feature based on the at least one associated feature node.

[0067] The feature validity evaluation module 12 is used to broadcast messages from the at least one basic feature node and the at least one combined feature node to the combined feature column node in the feature validity evaluation graph, so that the combined feature column node calculates the information gain of the basic feature with respect to the associated feature based on the basic information entropy corresponding to the basic feature and the conditional information entropy of the basic feature relative to the associated feature, and determines the feature validity evaluation result corresponding to the basic feature according to the information gain, wherein the basic information entropy is calculated based on the first distribution information, and the conditional information entropy is calculated based on the second distribution information.

[0068] In some embodiments, the at least one basic feature node is connected to the basic feature column node via an edge, and the edge between each basic feature node and the basic feature column node corresponds to a first weight value of the basic feature value of the basic feature node.

[0069] In some embodiments, the first weight value includes the number of occurrences corresponding to the basic feature value of the basic feature node in the data set.

[0070] In some embodiments, the feature validity evaluation device 1 is further used for:

[0071] In the feature validity evaluation graph, the basic feature column node receives the first weight value of the edge between the basic feature node and the basic feature column node sent by each basic feature node connected to it, and calculates the first distribution information of the basic feature based on the first weight value, wherein the first distribution information includes at least one first weight value and a corresponding first weight sum.

[0072] In some embodiments, the feature validity evaluation device 1 is also used for any of the following:

[0073] In the feature validity evaluation graph, the basic feature column node calculates the basic information entropy corresponding to the basic feature based on the first distribution information, and the combined feature column node receives the basic information entropy sent by the basic feature column node;

[0074] In the feature validity evaluation graph, the combined feature column node receives the first distribution information sent by the basic feature column node, and calculates the basic information entropy corresponding to the basic feature based on the first distribution information.

[0075] In some embodiments, the feature validity evaluation graph further includes a virtual basic feature column node, the basic feature column node is connected to the virtual basic feature column node via an edge, and the virtual basic feature column node is connected to the combined feature column node via an edge;

[0076] The feature validity evaluation device 1 is also used for:

[0077] In the feature validity evaluation graph, the virtual basic feature column node receives the first distribution information sent by the basic feature column node, and calculates the basic information entropy corresponding to the basic feature based on the first distribution information, so that the combined feature column node receives the basic information entropy sent by the virtual basic feature column node.

[0078] In some embodiments, one or more of the at least one combined feature node are connected to one of the at least one associated feature node via an edge, and for an associated feature node, the edge between each combined feature node connected to the associated feature node and the associated feature node corresponds to the second weight value of the combined feature value of the combined feature node.

[0079] In some embodiments, the second weight value includes the number of occurrences corresponding to the combined feature value of the combined feature node in the data set.

[0080] In some embodiments, the feature validity evaluation device 1 is further used for:

[0081] In the feature validity evaluation graph, each associated feature node receives a second weight value of an edge between the combined feature node and the associated feature node sent by each combined feature node connected thereto, and calculates third distribution information of the combined feature relative to the associated feature based on the second weight value, wherein the third distribution information includes one or more second weight values ​​and a corresponding second weight sum;

[0082] The associated feature column node receives the third distribution information corresponding to the associated feature value of the associated feature node sent by each associated feature node, and calculates the second distribution information of the combined feature based on the third distribution information, wherein the second distribution information includes one or more second weight values ​​corresponding to each associated feature node, the sum of the second weights corresponding to the associated feature node, and the sum of the second weights corresponding to the at least one associated feature node.

[0083] In some embodiments, the feature validity evaluation device 1 is further used for:

[0084] In the feature validity evaluation graph, the associated feature column node calculates the conditional information entropy of the basic feature relative to the associated feature based on the second distribution information, so that the combined feature column node receives the conditional information entropy sent by the associated feature column node;

[0085] In the feature validity evaluation graph, the combined feature column node receives the second distribution information sent by the associated feature column node, and calculates the conditional information entropy of the basic feature relative to the associated feature based on the second distribution information.

[0086] The above device embodiments correspond to the method embodiments. For specific descriptions, please refer to the description of the method embodiments, which will not be repeated here. The device embodiments are obtained based on the corresponding method embodiments and have the same technical effects as the corresponding method embodiments. For specific descriptions, please refer to the corresponding method embodiments.

[0087] The embodiment of the present specification also provides a computer storage medium, which can store multiple instructions, and the instructions are suitable for being loaded by a processor to execute the method of the embodiment of the present specification.

[0088] The embodiments of the present specification also provide a computer program product, which stores at least one instruction, and the at least one instruction is loaded by the processor to execute the method of the embodiments of the present specification.

[0089] The embodiments of this specification also provide Figure 4 The structural diagram of the electronic device shown in FIG. Figure 4 At the hardware level, the electronic device includes a processor, an internal bus, a network interface, a memory, and a non-volatile memory, and may also include hardware required for other services. The processor reads the corresponding computer program from the non-volatile memory into the memory and then runs it to implement the method of the embodiment of this specification.

[0090] The systems, devices, modules or units described in the above embodiments may be implemented by computer chips or entities, or by products with certain functions. A typical implementation device is a computer. Specifically, the computer may be, for example, 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 device, a game console, a tablet computer, a wearable device, or a combination of any of these devices.

[0091] Those skilled in the art will appreciate that the embodiments of this specification may be provided as methods, systems, or computer program products. Therefore, this specification may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Moreover, this specification may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) that contain computer-usable program code.

[0092] This specification is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of this specification. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0093] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0094] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0095] It should also be noted that the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, commodity or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, commodity or device. In the absence of more restrictions, the elements defined by the sentence "comprises a ..." do not exclude the existence of other identical elements in the process, method, commodity or device including the elements.

[0096] This specification may be described in the general context of computer-executable instructions executed by a computer, such as program modules. Generally, program modules include routines, programs, objects, components, data structures, etc. that perform specific tasks or implement specific abstract data types. This specification may also be practiced in distributed computing environments where tasks are performed by remote processing devices connected through a communication network. In a distributed computing environment, program modules may be located in local and remote computer storage media, including storage devices.

[0097] Each embodiment in this specification is described in a progressive manner, and the same or similar parts between the embodiments can be referred to each other, and each embodiment focuses on the differences from other embodiments. In particular, for the system embodiment, since it is basically similar to the method embodiment, the description is relatively simple, and the relevant parts can be referred to the partial description of the method embodiment.

[0098] The above description is only an embodiment of the present specification and is not intended to limit the present specification. For those skilled in the art, the present specification may have various changes and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present specification shall be included in the scope of the claims of the present specification.

Claims

1. A method for evaluating feature effectiveness, comprising: According to the basic features in the data set and the combined features corresponding to the basic features, a feature validity evaluation graph is constructed, wherein the combined features include the basic features and the associated features corresponding to the basic features, the feature validity evaluation graph includes at least one basic feature node, at least one combined feature node, at least one associated feature node, a basic feature column node, an associated feature column node and a combined feature column node, for an associated feature node, the combined feature value of each combined feature node connected to the associated feature node includes the associated feature value of the associated feature node, the basic feature column node is used to calculate the first distribution information of the basic feature based on the at least one basic feature node, and the associated feature column node is used to calculate the second distribution information of the combined feature based on the at least one associated feature node; In the feature validity evaluation graph, by broadcasting messages from the at least one basic feature node and the at least one combined feature node to the combined feature column node, the combined feature column node calculates the information gain of the basic feature with respect to the associated feature based on the basic information entropy corresponding to the basic feature and the conditional information entropy of the basic feature relative to the associated feature, and determines the feature validity evaluation result corresponding to the basic feature according to the information gain, wherein the basic information entropy is calculated based on the first distribution information, and the conditional information entropy is calculated based on the second distribution information.

2. According to the method of claim 1, the at least one basic feature node is connected to the basic feature column node through an edge, and the edge between each basic feature node and the basic feature column node corresponds to the first weight value of the basic feature value of the basic feature node.

3. According to the method described in claim 2, the first weight value includes the number of occurrences corresponding to the basic feature value of the basic feature node in the data set.

4. The method according to claim 2 or 3, further comprising: In the feature validity evaluation graph, the basic feature column node receives the first weight value of the edge between the basic feature node and the basic feature column node sent by each basic feature node connected to it, and calculates the first distribution information of the basic feature based on the first weight value, wherein the first distribution information includes at least one first weight value and a corresponding first weight sum.

5. The method according to claim 4, further comprising any of the following: In the feature validity evaluation graph, the basic feature column node calculates the basic information entropy corresponding to the basic feature based on the first distribution information, and the combined feature column node receives the basic information entropy sent by the basic feature column node; In the feature validity evaluation graph, the combined feature column node receives the first distribution information sent by the basic feature column node, and calculates the basic information entropy corresponding to the basic feature based on the first distribution information.

6. The method according to claim 4, wherein the feature validity evaluation graph further comprises a virtual basic feature column node, the basic feature column node is connected to the virtual basic feature column node via an edge, and the virtual basic feature column node is connected to the combined feature column node via an edge; in, The method further comprises: In the feature validity evaluation graph, the virtual basic feature column node receives the first distribution information sent by the basic feature column node, and calculates the basic information entropy corresponding to the basic feature based on the first distribution information, so that the combined feature column node receives the basic information entropy sent by the virtual basic feature column node.

7. According to the method according to claim 1, one or more of the at least one combined feature nodes are connected to one of the at least one associated feature nodes through an edge, and for an associated feature node, the edge between each combined feature node connected to the associated feature node and the associated feature node corresponds to the second weight value of the combined feature value of the combined feature node.

8. According to the method of claim 7, the second weight value includes the number of occurrences corresponding to the combined feature value of the combined feature node in the data set.

9. The method according to claim 7 or 8, further comprising: In the feature validity evaluation graph, each associated feature node receives a second weight value of an edge between the combined feature node and the associated feature node sent by each combined feature node connected thereto, and calculates third distribution information of the combined feature relative to the associated feature based on the second weight value, wherein the third distribution information includes one or more second weight values ​​and a corresponding second weight sum; The associated feature column node receives the third distribution information corresponding to the associated feature value of the associated feature node sent by each associated feature node, and calculates the second distribution information of the combined feature based on the third distribution information, wherein the second distribution information includes one or more second weight values ​​corresponding to each associated feature node, the sum of the second weights corresponding to the associated feature node, and the sum of the second weights corresponding to the at least one associated feature node.

10. The method according to claim 9, further comprising: In the feature validity evaluation graph, the associated feature column node calculates the conditional information entropy of the basic feature relative to the associated feature based on the second distribution information, so that the combined feature column node receives the conditional information entropy sent by the associated feature column node; In the feature validity evaluation graph, the combined feature column node receives the second distribution information sent by the associated feature column node, and calculates the conditional information entropy of the basic feature relative to the associated feature based on the second distribution information.

11. An apparatus for evaluating feature validity, comprising: A graph construction module, used to construct a feature validity evaluation graph according to basic features in a data set and combined features corresponding to the basic features, wherein the combined features include the basic features and associated features corresponding to the basic features, the feature validity evaluation graph includes at least one basic feature node, at least one combined feature node, at least one associated feature node, a basic feature column node, an associated feature column node and a combined feature column node, for an associated feature node, the combined feature value of each combined feature node connected to the associated feature node includes the associated feature value of the associated feature node, the basic feature column node is used to calculate the first distribution information of the basic feature based on the at least one basic feature node, and the associated feature column node is used to calculate the second distribution information of the combined feature based on the at least one associated feature node; A feature validity evaluation module is used to broadcast messages from the at least one basic feature node and the at least one combined feature node to the combined feature column node in the feature validity evaluation graph, so that the combined feature column node calculates the information gain of the basic feature with respect to the associated feature based on the basic information entropy corresponding to the basic feature and the conditional information entropy of the basic feature relative to the associated feature, and determines the feature validity evaluation result corresponding to the basic feature according to the information gain, wherein the basic information entropy is calculated based on the first distribution information, and the conditional information entropy is calculated based on the second distribution information.

12. A storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 10 are implemented.

13. An electronic device, characterized in that: include: A processor and a memory; wherein the memory stores a computer program, and the computer program is suitable for being loaded by the processor and executing the steps of the method as claimed in any one of claims 1 to 10.

14. A computer program product having at least one instruction stored thereon, characterized in that: When the at least one instruction is executed by the processor, the steps of the method described in any one of claims 1 to 10 are implemented.

Images