Feature generation, information prediction, training method and device of information prediction model

Abstract

Embodiments of the present specification provide feature generation, information prediction, and training methods and apparatuses for an information prediction model. The feature generation method includes: obtaining original features of a plurality of objects and a preset relationship table; performing missing influence analysis on each original feature to obtain a missing influence degree of each original feature; selecting a specific feature from each original feature according to the missing influence degree of each original feature; for a target object in the plurality of objects, obtaining feature information of the specific feature corresponding to each associated object according to the specific feature and the preset relationship table, wherein the associated object is an object having an association relationship with the target object; and generating an object feature of the target object according to the feature information of the specific feature corresponding to each associated object. By performing missing influence analysis on the obtained plurality of object features, a specific feature is selected, and then an object feature of a target object is generated according to an associated object and the specific feature, so that the accuracy of the generated object feature is higher.

Description

Technical Field

[0001] The embodiments in this specification relate to the field of data processing technology, and in particular to a feature generation method, an information prediction method, and a training method for an information prediction model. Background Technology

[0002] In information prediction scenarios, the first step is to collect user characteristics, then process these characteristics using neural network technology, and finally make information predictions based on the processing results. Examples include recommendation systems, credit assessments, and user mining. Clearly, the collection of user characteristics has a significant impact on the outcome of information prediction.

[0003] In practical applications, it is difficult to collect the characteristics of some users, which seriously affects the accuracy of information prediction due to the lack of user characteristics.

[0004] In cases of missing features, the median method is typically used for feature derivation, which restores the missing features. However, this method directly uses the median obtained from the features of users related to the user to fill in the missing features, which leads to inaccurate features. Therefore, a more accurate feature generation method is urgently needed. Summary of the Invention

[0005] In view of the above, embodiments of this specification provide a feature generation method. One or more embodiments of this specification also relate to an information prediction method, an information prediction model training method, a feature generation apparatus, an information prediction apparatus, an information prediction model training apparatus, a computing device, a computer-readable storage medium, and a computer program, to address the technical deficiencies existing in the prior art.

[0006] According to a first aspect of the embodiments of this specification, a feature generation method is provided, comprising:

[0007] Obtain the original characteristics of multiple objects and a preset relationship table, where the preset relationship table records the association relationships between multiple objects;

[0008] A missing feature impact analysis is performed on each original feature to obtain the missing feature impact degree of each original feature, where the missing feature impact degree characterizes the degree of impact of missing original features;

[0009] Based on the degree of impact of the absence of each original feature, specific features are selected from each original feature;

[0010] For a target object among multiple objects, based on specific features and a preset relationship table, feature information of specific features corresponding to each associated object is obtained, where associated objects are objects that have a relationship with the target object;

[0011] Based on the feature information of specific features corresponding to each associated object, generate the object features of the target object.

[0012] According to a second aspect of the embodiments of this specification, an information prediction method is provided, comprising:

[0013] Obtain the object characteristics of the target object, wherein the object characteristics are generated according to the feature generation method described above;

[0014] The object features are input into a pre-trained information prediction model to obtain information prediction results for the target object. The information prediction model is a machine learning model.

[0015] According to a third aspect of the embodiments of this specification, a method for training an information prediction model is provided, comprising:

[0016] Obtain the original features of multiple sample objects in the sample set and a preset relationship table, wherein the preset relationship table records the association relationships between multiple sample objects;

[0017] A missing feature impact analysis is performed on each original feature to obtain the missing feature impact degree of each original feature, where the missing feature impact degree characterizes the degree of impact of missing original features;

[0018] Based on the degree of impact of the absence of each original feature, specific features are selected from each original feature;

[0019] For a target sample object, feature information of specific features corresponding to each associated sample object is obtained based on specific features and a preset relationship table. The target sample object is any one of multiple sample objects, and the associated sample objects are sample objects that have a relationship with the target sample object.

[0020] Based on the feature information of specific features corresponding to each associated sample object, generate the object features of the target sample object;

[0021] The information prediction model is trained based on the object features of multiple sample objects to obtain the trained information prediction model.

[0022] According to a fourth aspect of the embodiments of this specification, a feature generation apparatus is provided, comprising:

[0023] The first acquisition module is configured to acquire the original features of multiple objects and a preset relationship table, wherein the preset relationship table records the association relationships between multiple objects;

[0024] The first analysis module is configured to perform a missing impact analysis on each original feature to obtain the missing impact degree of each original feature, wherein the missing impact degree characterizes the degree of impact of missing original features.

[0025] The first filtering module is configured to filter specific features from the original features based on the degree of missing impact of each original feature.

[0026] The first acquisition module is configured to, for a target object among multiple objects, obtain feature information of specific features corresponding to each associated object based on specific features and a preset relationship table, wherein the associated object is an object that has a relationship with the target object;

[0027] The first generation module is configured to generate object features of the target object based on the feature information of the specific features corresponding to each associated object.

[0028] According to a fifth aspect of the embodiments of this specification, an information prediction apparatus is provided, comprising:

[0029] The second acquisition module is configured to acquire the object features of the target object, wherein the object features are generated according to the aforementioned feature generation method;

[0030] The prediction module is configured to input object features into a pre-trained information prediction model to obtain information prediction results for the target object, wherein the information prediction model is a machine learning model.

[0031] According to a sixth aspect of the embodiments of this specification, a training apparatus for an information prediction model is provided, comprising:

[0032] The third acquisition module is configured to acquire the original features of multiple sample objects in the sample set and a preset relationship table, wherein the preset relationship table records the association relationship between multiple sample objects;

[0033] The second analysis module is configured to perform a missing impact analysis on each original feature to obtain the missing impact degree of each original feature, wherein the missing impact degree characterizes the degree of impact of missing original features.

[0034] The second filtering module is configured to filter specific features from the original features based on the degree of missing impact of each original feature.

[0035] The second acquisition module is configured to obtain feature information of specific features corresponding to each associated sample object based on specific features and a preset relationship table for the target sample object. The target sample object is any one of multiple sample objects, and the associated sample objects are sample objects that have an association relationship with the target sample object.

[0036] The second generation module is configured to generate object features of the target sample object based on the feature information of the specific features corresponding to each associated sample object;

[0037] The training module is configured to train the information prediction model based on the object features of multiple sample objects, and obtain the trained information prediction model.

[0038] According to a seventh aspect of the embodiments of this specification, a computing device is provided, comprising:

[0039] Memory and processor;

[0040] The memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions. When the computer-executable instructions are executed by the processor, they implement the steps of the above-described feature generation method, the above-described information prediction method, or the above-described information prediction model training method.

[0041] According to an eighth aspect of the embodiments of this specification, a computer-readable storage medium is provided that stores computer-executable instructions, which, when executed by a processor, implement the steps of the above-described feature generation method, the above-described information prediction method, or the above-described information prediction model training method.

[0042] According to a ninth aspect of the embodiments of this specification, a computer program is provided, wherein when the computer program is executed in a computer, it causes the computer to perform the steps of the above-described feature generation method, the above-described information prediction method, or the above-described information prediction model training method.

[0043] One embodiment of this specification involves obtaining the original features of multiple objects and a preset relationship table, wherein the preset relationship table records the association relationships between the multiple objects; performing a missing impact analysis on each original feature to obtain the missing impact degree of each original feature, wherein the missing impact degree characterizes the degree of influence of missing original features; selecting specific features from each original feature based on the missing impact degree of each original feature; for the target object among the multiple objects, obtaining the feature information of the specific features corresponding to each associated object based on the specific features and the preset relationship table, wherein the associated object is an object that has an association relationship with the target object; and generating the object features of the target object based on the feature information of the specific features corresponding to each associated object. By performing a missing feature analysis on the original features of multiple objects, the missing feature degree of each original feature is obtained. In other words, the missing feature degree of each original feature can be used to determine the degree of impact of the missing original feature on the final result. Thus, specific features suitable for feature generation processing are determined from each original feature. Then, based on the obtained preset relationship table, the feature information of objects that are related to the target object and their corresponding specific features are determined, and the object features of the target object are determined. This makes the generated object features based on objects that are related to the target object, so the accuracy of the generated object features of the target object is high. Attached Figure Description

[0044] Figure 1 This is a schematic diagram of the structure of a feature generation system provided in one embodiment of this specification;

[0045] Figure 2 This is a flowchart of a feature generation method provided in one embodiment of this specification;

[0046] Figure 3 This is a schematic diagram of the relationship between different objects in a feature generation method provided in one embodiment of this specification;

[0047] Figure 4a This is a flowchart of another feature generation method provided in one embodiment of this specification;

[0048] Figure 4b This is a flowchart illustrating another feature generation method provided in one embodiment of this specification;

[0049] Figure 5 This is a flowchart illustrating the processing steps of a feature generation method provided in one embodiment of this specification.

[0050] Figure 6 This is a schematic diagram illustrating the processing steps of a feature generation method provided in one embodiment of this specification;

[0051] Figure 7 This is a flowchart of an information prediction method provided in one embodiment of this specification;

[0052] Figure 8 This is a flowchart illustrating a training method for an information prediction model provided in one embodiment of this specification;

[0053] Figure 9 This is a schematic diagram of the structure of a feature generation device provided in one embodiment of this specification;

[0054] Figure 10 This is a schematic diagram of the structure of an information prediction device provided in one embodiment of this specification;

[0055] Figure 11 This is a schematic diagram of the structure of a training device for an information prediction model provided in one embodiment of this specification;

[0056] Figure 12 This is a structural block diagram of a computing device provided in one embodiment of this specification. Detailed Implementation

[0057] Many specific details are set forth in the following description to provide a full understanding of this specification. However, this specification can be implemented in many other ways than those described herein, and those skilled in the art can make similar extensions without departing from the spirit of this specification. Therefore, this specification is not limited to the specific implementations disclosed below.

[0058] The terminology used in one or more embodiments of this specification is for the purpose of describing particular embodiments only and is not intended to be limiting of the one or more embodiments of this specification. The singular forms “a,” “described,” and “the” as used in one or more embodiments of this specification and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used in one or more embodiments of this specification refers to and includes any or all possible combinations of one or more associated listed items.

[0059] It should be understood that although the terms first, second, etc., may be used to describe various information in one or more embodiments of this specification, such information should not be limited to these terms. These terms are only used to distinguish information of the same type from one another. For example, first may also be referred to as second without departing from the scope of one or more embodiments of this specification, and similarly, second may also be referred to as first. Depending on the context, the word "if" as used herein may be interpreted as "when," "when," or "in response to a determination."

[0060] First, the terms and concepts used in one or more embodiments of this specification will be explained.

[0061] A heterogeneous graph is a structure that describes a set of objects, where certain pairs of objects are "related" in some sense. These objects correspond to mathematical abstractions called vertices (also known as nodes or points), and each related pair of vertices is called an edge (also known as a link or line). A heterogeneous graph can contain more than one type of node and edge.

[0062] Derived features: Derived features are new features obtained through feature learning using the original data. Derived features are generally caused by two reasons: changes in the data itself, resulting in many features that were not originally present; or, during feature learning, the algorithm generates derived features based on certain relationships between features, and sometimes derived features better reflect the relationships between data features.

[0063] Small and medium-sized enterprise (SME) graph: This is a graph-based data structure where nodes represent micro and small enterprises, and links represent relationships between SMEs, such as holding relationships, joint ventures, upstream and downstream relationships, and money transfer relationships.

[0064] Embedding layer: A common layer in deep learning network models, mainly used to handle vector representations of sparse features. It not only solves the problem of one-hot vector length, but also represents the similarity between features.

[0065] Coverage: refers to the proportion of valid feature values ​​among all users.

[0066] Importance: refers to the usefulness of the input features during the processing.

[0067] Information gain (Kullback–Leibler divergence): In probability and information theory, information gain is asymmetric and is used to measure the difference between two probability distributions P and Q. Information gain describes the difference between encoding using Q and encoding using P. Typically, P represents the distribution of samples or observations, or it may be a precisely calculated theoretical distribution. Q represents a theory, model, description, or approximation of P.

[0068] In information prediction scenarios, predictions need to be based on user data and user characteristics. However, without high-quality data, there are no high-quality data mining results. Missing data values ​​are a common problem in data analysis. When the missing percentage is small, missing records can be discarded directly or processed manually. However, in real-world data, missing data often constitutes a significant proportion. In this case, manual processing is very inefficient, and discarding missing records will result in the loss of a large amount of information, creating a systematic discrepancy between incomplete and complete observations. Analyzing such data may lead to incorrect conclusions. There are various reasons for missing data values: in marketing scenarios, due to new customers, inactive users, and data access restrictions, missing user characteristics are unavoidable. When missing data values ​​occur, we will fill in the missing features.

[0069] Existing methods primarily use global numerical statistical features or user profiles for imputation. For example, statistical methods use the mean, weighted mean, and median for numerical data, and the most common mode for categorical data. Model-based methods often use other existing fields as the target variable to predict the most likely imputed value. If the column with missing values ​​is a numerical variable, a regression model is used; if it's a categorical variable, a classification model is used. Expert imputation is also crucial for a small number of significant data records. Other methods exist, such as randomization, special value methods, and multiple imputation, but these methods are coarse-grained, and the imputed data is not strongly correlated with the original data, potentially increasing noise. Therefore, a highly accurate feature generation method is urgently needed.

[0070] To address the aforementioned issues, this embodiment of the specification obtains the original features of multiple objects and a preset relationship table, wherein the preset relationship table records the association relationships between multiple objects; performs a missing feature impact analysis on each original feature to obtain the missing feature impact degree of each original feature, wherein the missing feature impact degree characterizes the degree of impact of missing original features; based on the missing feature impact degree of each original feature, specific features are selected from each original feature; for the target object among multiple objects, based on the specific features and the preset relationship table, the feature information of the specific features corresponding to each associated object is obtained, wherein the associated object is an object that has an association relationship with the target object; based on the feature information of the specific features corresponding to each associated object, the object features of the target object are generated. By performing a missing feature analysis on the original features of multiple objects, the missing feature degree of each original feature is obtained. In other words, the missing feature degree of each original feature can be used to determine the degree of impact of the missing original feature on the final result. Thus, specific features suitable for feature generation processing are determined from each original feature. Then, based on the obtained preset relationship table, the feature information of objects that are related to the target object and their corresponding specific features are determined, and the object features of the target object are determined. This makes the generated object features based on objects that are related to the target object, so the accuracy of the generated object features of the target object is high.

[0071] This specification provides a feature generation method. One or more embodiments of this specification also relate to an information prediction method, an information prediction model training method, a feature generation apparatus, an information prediction apparatus, an information prediction model training apparatus, a computing device, and a computer-readable storage medium, which will be described in detail in the following embodiments.

[0072] See Figure 1 , Figure 1A schematic diagram of a feature generation system according to an embodiment of this specification is shown, such as... Figure 1 As shown, the system includes server 0101 and client 0102.

[0073] The server 0101 obtains the original features and the preset relationship table from the interaction information generated by the interaction with the client 0102. Then, it determines the specific features based on the original features. After that, it obtains the feature information of the specific features corresponding to each associated object based on the specific features and the preset relationship table, and further generates the object features of the target object.

[0074] Specifically, the process involves acquiring the original features of multiple objects and a pre-defined relationship table, which records the relationships between the objects. A missing feature impact analysis is performed on each original feature to obtain its missing feature degree, which characterizes the extent of the impact of missing features. Based on the missing feature degree, specific features are selected from the original features. For the target object among the multiple objects, based on the specific features and the pre-defined relationship table, feature information of the specific features corresponding to each associated object is obtained, where associated objects are those related to the target object. Based on the feature information of the specific features corresponding to each associated object, object features of the target object are generated. By performing a missing feature impact analysis on the acquired original features of multiple objects, the missing feature degree of each original feature is obtained. This allows the determination of the impact of missing original features on the final result, thereby identifying specific features suitable for feature generation. Then, based on the acquired pre-defined relationship table, the feature information of objects related to the target object and their corresponding specific features is determined, thus identifying the object features of the target object. This ensures that the generated object features are based on objects related to the target object, resulting in high accuracy of the generated object features.

[0075] See Figure 2 , Figure 2 A flowchart of a feature generation method according to an embodiment of this specification is shown, which specifically includes the following steps.

[0076] Step 202: Obtain the original features of multiple objects and a preset relationship table, wherein the preset relationship table records the association relationships between the multiple objects.

[0077] Specifically, original features refer to the original features obtained directly based on the user's activity information on the Internet. For example, original features could be the user's browsing characteristics of an app, the information characteristics and evaluation characteristics of a product, etc. A preset relationship table refers to a pre-set relationship table that records the associations between multiple objects. For example, a preset relationship table could be a pre-set list of company employees, a pre-set relationship table for sharing shopping links, etc. An association refers to a relationship between multiple objects. For example, the association between multiple objects could be a transfer relationship, a social relationship, a supply chain relationship, or a friend relationship, etc. The specific choice depends on the actual situation, and this specification does not impose any limitations on this embodiment.

[0078] It should be noted that, since the final requirement is to generate the object features of the target object, it is first necessary to obtain the original features of multiple objects and a preset relationship table. This allows specific features to be determined from the original features of multiple objects, further determining the object features of the target object, and ultimately supplementing the features of the target object. For example, in the information recommendation scenario, information can be recommended to the target object based on the determined object features.

[0079] It should also be noted that there are many ways to obtain the original characteristics and preset relationship table of multiple objects. One possible implementation is to directly obtain the original characteristics and preset relationship table of multiple objects from the client; another possible implementation is that the server automatically generates the original characteristics and preset relationship table of multiple objects based on the information generated during the communication with the client.

[0080] By applying the scheme of the embodiments of this specification, by obtaining the original features of multiple objects and a preset relationship table, wherein the preset relationship table records the association relationships between multiple objects, it is possible to subsequently determine specific features based on the original features, determine the associated objects that have associations and relationships with the target object based on the preset relationship table, and finally determine the object features of the target object based on the specific features and the associated objects, thereby speeding up the determination of the object features of the target object.

[0081] Step 204: Perform a missing impact analysis on each original feature to obtain the missing impact degree of each original feature, wherein the missing impact degree characterizes the degree of influence of missing original features.

[0082] Specifically, missing feature analysis refers to analyzing the impact of missing features. For example, a pre-defined neural network model can be used to input original features to obtain the missing feature impact of each feature. The processing performed by the neural network model is the missing feature analysis. The missing feature impact refers to the degree to which the missing original feature affects the final result. For example, the missing feature impact can refer to feature coverage, feature importance, feature information gain, etc., and the specific choice depends on the actual situation. This specification does not impose any limitations on this aspect in the embodiments.

[0083] It should be noted that the original features are the original features of multiple objects. To perform missing impact analysis on the original features and obtain the missing impact degree of each original feature, it is necessary to identify the original features with the greater missing impact degree from the multiple original features. Based on the original features with the greater missing impact degree, subsequent feature derivation processing is performed to obtain the object features of the target object.

[0084] For example, a missing feature analysis was performed on time features, login features, and information features, and the missing feature impact was determined as follows: the coverage of time features was a1, the importance was a2, and the information gain was a3; the coverage of login features was b1, the importance was b2, and the information gain was b3; and the coverage of information features was c1, the importance was c2, and the information gain was c3.

[0085] By applying the scheme of the embodiments of this specification, the missing impact of each original feature is analyzed to obtain the missing impact degree of each original feature. The missing impact degree represents the degree of influence of the missing feature, so that specific features can be determined based on the missing impact degree of each original feature, thereby accelerating the determination of the object features of the target object.

[0086] In one optional implementation, the step of performing a missing feature analysis on each original feature to obtain the missing feature degree of each original feature includes the following specific steps:

[0087] Statistically analyze the category attribute parameters of each of the original features;

[0088] Calculate the influence index of each original feature based on the category attribute parameters of each original feature;

[0089] Based on the influence index of each original feature, the degree of missing influence of each original feature is determined.

[0090] Specifically, category attribute parameters refer to the attribute parameters of the category to which the feature belongs. For example, the category attribute parameters of the original feature include the number of feature categories, the probability that the feature belongs to a specific category, etc. Influence indicators refer to indicators that judge the degree of influence. For example, the influence indicators of the original feature can be the information gain of the feature, the coverage of the feature, the importance of the feature, etc. The specific selection depends on the actual situation, and the embodiments in this specification do not impose any limitations on this.

[0091] It should be noted that the influence index of each original feature needs to be calculated, so the category attribute parameters of each original feature need to be statistically analyzed in advance. The missing influence of the original feature cannot be obtained directly based on the category attribute parameters of the original feature. Therefore, after obtaining the category attribute parameters of the original feature, the influence index needs to be calculated first, and then the missing influence of the original feature needs to be determined based on the influence index.

[0092] The calculation process for the influence index of each original feature, based on the statistically derived category attribute parameters, is as follows:

[0093] First, we count the number of categories |y| and p for each original feature. k This represents the probability that the label category of a certain original feature is k.

[0094] Information entropy is one of the most commonly used metrics for measuring the purity of a sample set. It describes the uncertainty of the occurrence of various possible events in an information source. Its calculation formula is as follows (1):

[0095]

[0096] Ent(D) refers to information entropy. The smaller the information entropy, the higher the information content.

[0097] Assuming a discrete feature *a* has V possible values, if *a* is used to partition the sample set D, V branch nodes will be generated, where the v-th branch node contains the value of feature *a* in D. v The sample is denoted as D. v We can calculate D v The information entropy is calculated by normalizing it according to the quantity of D, and the formula for calculating information gain is as follows (2):

[0098]

[0099] Gain(D,a) refers to information gain. The greater the information gain, the higher the discriminative power of feature a.

[0100] Cover(a v ) is a v The feature coverage is given by the following formula (3):

[0101]

[0102] GINI(D) represents the feature importance of feature D. Feature importance is calculated using the Gini coefficient, p k The probability that a sample belongs to the kth category is given by the following formula (4).

[0103]

[0104] It should be further noted that there are many ways to determine the missing impact based on the influence indicators of each original feature. One possible approach is to directly use the influence indicators of each original feature as the missing impact of each original feature. For example, if the coverage of the time feature is calculated to be a1, the importance to be a2, and the information gain to be a3; the coverage of the login feature is calculated to be b1, the importance to be b2, and the information gain to be b3; and the coverage of the information feature is calculated to be c1, the importance to be c2, and the information gain to be c3, then the values ​​of the influence indicators (coverage, importance, and information gain) of the time feature, login feature, and information feature are directly used as the missing impact. Another possible implementation is to weight the various influence indicators of each original feature according to different weights, and use the weighted result as the missing influence of each original feature. For example, if the coverage of the time feature is a1, the importance is a2, and the information gain is a3; the coverage of the login feature is b1, the importance is b2, and the information gain is b3; and the coverage of the information feature is c1, the importance is c2, and the information gain is c3, and the weight of the coverage is x, the weight of the importance is y, and the weight of the information gain is z, then the missing influence of the time feature is a1*x + a2*y + a3*z; the missing influence of the login feature is b1*x + b2*y + b3*z; and the missing influence of the information feature is c1*x + c2*y + c3*z.

[0105] The scheme of the embodiments of this specification is applied by statistically analyzing the category attribute parameters of each original feature and calculating the influence index of each original feature based on the statistically analyzed category attribute parameters of each original feature. This further determines the missing influence degree of each original feature, so that specific features can be determined based on the missing influence degree of each original feature. In this way, determining specific features from each original feature based on the missing influence degree improves the accuracy of determining the object features of the target object.

[0106] Step 206: Based on the degree of missing impact of each original feature, select specific features from the original features.

[0107] Specifically, a specific feature refers to a feature that can be derived from that feature.

[0108] It should be noted that, based on the degree of missing impact of each original feature, specific features are selected from each original feature. Since the selected specific features need to be fused to generate the object features of the target object, it is necessary to filter each original feature according to the degree of missing impact of each original feature, and select the original features whose degree of missing impact meets the requirements as the specific features.

[0109] It should be further noted that there are many ways to filter features. One possible implementation is to pre-set a missing value threshold. If the missing value of a certain original feature is less than the threshold, then that original feature is selected as the specific feature. Another possible implementation is to set a range for the missing value. If the missing value of a certain original feature falls within this range, then that original feature is selected as the specific feature.

[0110] The scheme implemented in this specification improves the rigor and accuracy of selecting specific features by filtering out qualified original features as specific features based on the degree of impact of the absence of each original feature, and by selecting qualified features from multiple original features.

[0111] In one optional implementation, the missing impact includes coverage and information gain; the step of selecting specific features from the original features based on the missing impact of each original feature includes the following specific steps:

[0112] From the original features, specific features with coverage less than a preset coverage threshold and information gain greater than a preset information gain threshold are selected.

[0113] Specifically, coverage refers to the proportion of any one of the original features to the total number of objects. Information gain refers to the degree to which information uncertainty is reduced under a given condition, that is, the degree to which the uncertainty of the prediction result is reduced if a certain original feature is missing. The preset coverage threshold is a pre-set limit value for coverage. For example, if the preset coverage threshold is 0.7, then if the coverage of a certain original feature is 0.8, then that original feature is discarded. The preset information gain is a pre-set limit value for information gain. For example, if the information gain threshold is 0.5, then if the information gain of a certain original feature is 0.4, then that original feature is discarded. A specific feature refers to a feature selected from the original features whose coverage is less than the coverage threshold and whose information gain is greater than the preset information gain. For example, if the preset coverage threshold is 0.7 and the information gain threshold is 0.5, then an original feature with a coverage of 0.5 and an information gain of 0.7 is determined as a specific feature. The specific selection is based on the actual situation, and the embodiments in this specification do not impose any limitations on this.

[0114] It's important to note that from the original features, specific features with coverage less than a preset coverage threshold and information gain greater than a preset information gain threshold are selected. When selecting specific features, it's crucial to choose those with low coverage and high information gain. This prevents the oversmoothing of object features when determining the target object using the identified specific features. In real-world scenarios, features with high information gain are generally user behavior features. Due to varying user activity levels, these features typically have low coverage, and the required specific features are precisely these types of features.

[0115] It should be further explained that, based on the missing impact of each original feature, one possible implementation for selecting a specific feature from the original features is to determine the original feature whose missing impact, which meets the missing impact threshold, is determined after weighting each original feature by influence indicators to obtain its missing impact. Another possible implementation is to determine the original feature whose missing impact, which meets the corresponding missing impact threshold, is determined after any influence indicator has been determined as its missing impact. Yet another possible implementation is to determine the original feature whose missing impact, which meets the missing impact threshold for some or all types, is determined after determining all calculated influence indicators as their missing impact.

[0116] The solution implemented in this specification determines the object features of the target object by identifying original features from each original feature whose coverage is less than a preset coverage threshold and whose information gain is greater than a preset information gain threshold. This avoids the problem of overly smooth object features and makes the determined object features more accurate.

[0117] Step 208: For the target object among the multiple objects, based on the specific feature and the preset relationship table, obtain the feature information of the specific feature corresponding to each associated object, wherein the associated object is an object that has an association relationship with the target object.

[0118] Specifically, a target object refers to the object from which feature generation is needed. For example, in practical applications, it could refer to an object lacking certain features; when generating features for it, this object can be called the target object. Associated objects are objects that have a relationship with the target object. For example, if object A is the target object, and object A has a transfer relationship with object B, or a cooperative relationship with object C, then object B and object C are both associated objects of object A. Specific features are the categories of original features that satisfy the filtering criteria, and the corresponding original features within those categories. Feature information refers to the attributes, categories, and other information of a feature. For example, if the feature of an object is logging in at night, then its feature information could be a time feature: the time is night, and the behavior is logging in.

[0119] It should be noted that, for a target object among multiple objects, based on specific features and a preset relationship table, the feature information of the specific features corresponding to each associated object is obtained. To generate features for the target object, it is first necessary to filter out specific features from each original feature, and then determine which objects among the objects related to the target object have the specific feature. The determined objects are used as associated objects of the target object, and then the feature information of the specific features of the associated objects can be obtained.

[0120] It should be further explained that there are many ways to obtain the feature information of the specific features corresponding to each associated object for the target object, based on specific characteristics and a preset relationship table. One possible implementation is to first filter the associated objects that are related to the target object from the preset relationship table, and then obtain the feature information of the specific features corresponding to these associated objects. Furthermore, it is also possible to filter the objects containing the specific feature from the filtered associated objects as the final associated objects, and then obtain the feature information of the specific features corresponding to the final associated objects. Another possible implementation is to first construct an object relationship graph based on the specific features and the preset relationship table, and then obtain the feature information of the specific features corresponding to the associated objects based on the object relationship graph.

[0121] By applying the solution of the embodiments of this specification, when a target object is determined, specific information of the specific features corresponding to the associated objects of the target object can be determined according to specific features and a preset relationship table. The preset relationship table is used to find the associated objects that have a relationship with the target object, and the feature information of the specific features of the associated objects is further determined. Then, the object features of the target object are generated based on the feature information, which increases the accuracy of determining the object features of the target object.

[0122] In one optional implementation, the step of obtaining feature information of the specific feature corresponding to each associated object based on the specific feature and the preset relationship table for a target object among the plurality of objects includes the following specific steps:

[0123] For the target object among the plurality of objects, an object relationship graph is constructed based on the specific features and the preset relationship table, wherein each node in the object relationship graph records the feature information of the specific features, and the edges in the object relationship graph represent the association relationship;

[0124] Based on the object relationship graph, determine each neighbor node of the node corresponding to the target object;

[0125] Read the feature information of the specific feature corresponding to each associated object from each neighbor node.

[0126] Specifically, an object relationship graph is a heterogeneous graph constructed from multiple objects and a pre-defined relationship table. For example, if object A has a transfer relationship with object B and a cooperative relationship with object C, then object A is connected to objects B and C with edges. These edges represent the relationships between object A and objects B and C, namely, the transfer relationship and the cooperative relationship. Furthermore, each node in the object relationship graph records feature information for specific characteristics. For example, if the specific characteristic is determined to be time, then the node corresponding to object A will only record the time feature information. In practical applications, this can be used to create a graph for small and medium-sized enterprises. A neighbor node is a node corresponding to an associated object of the target object. For example, if object A has a transfer relationship with object B and a cooperative relationship with object C, then object A is connected to objects B and C with edges. In the object relationship graph constructed from objects A, B, and C, the nodes corresponding to objects B and C are called the neighbor nodes of the node corresponding to object A.

[0127] It should be noted that, for a target object among multiple objects, an object relationship diagram is first constructed based on specific characteristics and a preset relationship table. This can be done by first determining the target object, then finding related objects that are associated with the target object from the preset relationship table, and then constructing the object relationship diagram based on the target object, related objects, the relationship between the target object and related objects, and specific characteristics.

[0128] It should be further explained that, based on the object relationship graph, the neighboring nodes of the node corresponding to the target object are determined, and then the feature information of the specific features corresponding to each associated object is read from each neighboring node. This is done by first finding the neighboring nodes adjacent to the node corresponding to the target object in the object relationship graph. For example, its neighboring nodes can be first-order neighboring nodes, second-order neighboring nodes, ... K-order neighboring nodes. Then, the feature information of the specific features recorded by each node is read from each neighboring node.

[0129] In practical applications, there are many ways to determine the neighboring nodes of a target object based on the object relationship graph. The specific method to be selected depends on the actual situation, and the embodiments in this specification do not impose any limitations on this.

[0130] See Figure 3 , Figure 3 This diagram illustrates different object relationship graphs according to an embodiment of this specification, representing a feature generation method. In one possible implementation of this specification, neighbor node selection is performed for object relationship graphs composed of objects with simple relationships and those composed of objects with sparse relationships, such as... Figure 3 (a) and (b) in the text.

[0131] Figure 3 In (a) and (b), there are nodes A, B, C and D, where "i" represents a node.

[0132] against Figure 3 In the two object relationship graphs shown in (a) and (b), the method for determining the neighbor nodes corresponding to the target object is as follows:

[0133] The missing features of the target object have varying degrees of similarity to the K-order directly connected neighbor nodes in the object relationship graph. For object relationship graphs composed of objects with simple or sparse relationships, the number of first-order neighbor nodes is small, and some may even have missing features. In this case, the information of the first-order neighbor nodes is insufficient to generate the object features of the target object. Therefore, it is necessary to search for the information of K-order neighbor nodes and read the feature information of the specific features corresponding to the K-order neighbor nodes. K-order neighbors actually refer to the neighbor nodes in the K-order neighborhood based on the node corresponding to the target object. K is generally 3. The advantage of this implementation method is that it is relatively simple and easy to implement. The disadvantage is that it does not select neighbors and directly determines all nodes in the corresponding order neighborhood, which will lead to excessive data noise.

[0134] In another possible implementation of this specification, neighboring nodes are selected based on the spatial distance and feature distance between nodes in the object relationship graph, such as... Figure 3 (c) in the middle.

[0135] Figure 3 In (c), there are distances from node A to nodes B, C, and D. Nodes B and C, which are shorter than node A, are determined to be the choice of their corresponding feature information as derived features of node A, where "TRUE" indicates certainty. Node D, which is farther from node A, is not determined as a derived feature of node A, where "FALSE" indicates uncertainty.

[0136] against Figure 3 In case (c), the method for determining the neighbor nodes corresponding to the target object is as follows:

[0137] To explore spatially and feature-proximity neighbors, an unsupervised algorithm with an encoder structure can be used to obtain the node embedding. The spatial and feature distances between the node corresponding to the target object and other neighboring nodes are calculated and represented by the inner product of the node embeddings. The inner product of the embeddings of multiple nodes and the node corresponding to the target object is calculated. The larger the inner product, the more similar the node is to the node corresponding to the target object. The K neighboring nodes with the largest inner products can be selected according to the requirements.

[0138] In another possible implementation of this specification, neighboring nodes are selected based on the different weights between nodes in the object relationship graph, such as... Figure 3 (d) in the middle.

[0139] Figure 3 In (d), there are nodes A, B, C and D, where "i" represents a node. The weights of nodes A and B are 0.3, B and C are 0.25, and B and D are 0.09.

[0140] against Figure 3 In case (d), the method for determining the neighbor nodes corresponding to the target object is as follows:

[0141] The neighboring nodes are determined based on the weights between different neighboring nodes and the target node corresponding to the target object in the object relationship graph. Specifically, the nodes in the object relationship graph are connected by edges, and there are weights on these edges. We can use these weights and set a weight threshold to select neighboring nodes. For example, if node A communicates with node B 6 times and node A communicates with node C 4 times, then if node A is selected as the target node corresponding to the target object, we construct an object relationship graph for node A. In this graph, the edge weight between node A and node B is 0.6, the edge weight between node A and node C is 0.4, and the weight threshold is 0.5. Therefore, node A can be selected as a neighboring node.

[0142] It is worth noting that there are many ways to obtain the feature information of each associated object based on specific features and a preset relationship table. One possible approach is to obtain the feature information of the associated object corresponding to the specific feature by constructing an object relationship graph. Another possible approach is to directly find the associated objects that are related to the target object and have specific features through the preset relationship table and specific features, and then obtain the feature information of these associated objects corresponding to the specific feature.

[0143] The scheme of the embodiments of this specification is used to construct an object relationship graph for a target object and specific features and a preset relationship table among multiple objects. Then, based on the object relationship graph, each neighbor node of the node corresponding to the target object is determined. Furthermore, the feature information of the specific features of the associated objects corresponding to each neighbor node is determined. Subsequently, the feature information of the specific features of the objects corresponding to each neighbor node determined in the object relationship graph can be fused to determine the object features of the target object, thereby improving the accuracy of determining the object features.

[0144] Step 210: Generate the object features of the target object based on the feature information of the specific features corresponding to each associated object.

[0145] It should be noted that the object features of the target object are generated based on the feature information of the specific features corresponding to each associated object. Depending on the different feature information contained in the target object, there are two possible ways to generate the object features of the target object. One possible implementation is to directly fuse the feature information of the specific features corresponding to each associated object, and use the fused feature directly as the object features of the target object. Another possible implementation is to fuse the feature information of the specific features corresponding to each associated object, and use the fused feature along with the feature information of the specific features originally contained in the target object as the object features of the target object. The specific method should be chosen according to the actual situation, and this specification does not impose any limitations on the embodiments.

[0146] The scheme of the embodiments of this specification generates the object features of the target object by using the feature information of the specific features corresponding to each associated object, so that the object features of the generated target object are obtained by filtering from each associated object, and the accuracy of the object features of the generated target object is higher.

[0147] In one optional implementation, the step of generating the object features of the target object based on the feature information of the specific features corresponding to each associated object includes the following specific steps:

[0148] The feature information of the specific features corresponding to each associated object is fused to obtain fused feature information;

[0149] Based on the fused feature information, the object features of the target object are generated.

[0150] Specifically, fused feature information refers to feature information formed by the fusion of multiple feature information. It represents the fused feature information of specific features of the target object and its associated objects. For example, if object A is the target object, object B is an associated object of object A, and object C is an associated object of object A, then the fused feature information can be obtained by fusing the feature information of the specific features corresponding to object B and object C.

[0151] It should be noted that the feature information of the specific features corresponding to each associated object is fused to obtain fused feature information. Further, the object features of the target object are generated based on the fused feature information. There are many ways to perform fusion processing. For example, it can be to calculate the average value of the feature information of the specific features corresponding to each associated object, or to sum the feature information of the specific features corresponding to each associated object, or to calculate the standard deviation of the feature information of the specific features corresponding to each associated object.

[0152] It is worth noting that the fused feature information obtained by fusing the feature information of specific features corresponding to each associated object can not only supplement the features of the target object, but also describe the average level of the community in which the node corresponding to the target object is located.

[0153] In practical applications, the three fusion processing methods can be used simultaneously to obtain three fused feature information.

[0154] By applying the scheme of the embodiments of this specification, the feature information corresponding to specific features of each associated object is fused to obtain fused feature information, and the object features of the target object are further generated. This results in high accuracy based on the fused feature information, and further improves the accuracy of the generated object features of the target object.

[0155] In one possible implementation, the step of generating the object features of the target object based on the fused feature information includes the following specific steps:

[0156] When the target object has the original feature information of the specific feature, the object feature of the target object is determined based on the fused feature information and the original feature information.

[0157] Specifically, original feature information refers to the feature information of a specific characteristic that is originally present. For example, original feature information may be the attribute information, identification information, etc. of a specific characteristic of a target object.

[0158] It should be noted that when the target object itself possesses original feature information with specific characteristics, the fused feature information and the original feature information are used as new specific features of the target object, i.e., object features. If the target object itself does not possess original feature information with specific characteristics, the fused feature information is directly used as the object features of the target object.

[0159] The solution implemented in this specification incorporates the fused feature information as part of the object features of the target object, regardless of whether the target object originally possesses or does not have specific feature information. The fused feature information is obtained by processing the feature information of each associated object corresponding to specific features, making the generated object features of the target object more complete and accurate.

[0160] In one possible implementation, the fusion process includes at least one of the following three methods: averaging, summing, and calculating the standard deviation.

[0161] Specifically, averaging refers to summing the feature information of specific characteristics of each associated object, and then dividing the sum by the number of associated objects to obtain the fused feature information, as shown in the following formula (5):

[0162]

[0163] in, X is the average value, n represents the number of feature information, and X is the average value. i This is the i-th feature information.

[0164] Specifically, summation refers to summing the feature information of specific characteristics of each associated object. The summation value is the fused feature information, as shown in the following formula (6):

[0165]

[0166] Where ∑ is the summation value, n represents the number of feature information, and X i This is the i-th feature information.

[0167] Specifically, calculating the standard deviation refers to calculating the standard deviation of specific features of each associated object. The value obtained by calculating the standard deviation is the fused feature information, as shown in the following formula (7):

[0168]

[0169] Where σ is the standard deviation, n represents the number of features, and X i For the i-th feature information, The average value of the feature information ( The feature information in the data includes not only the feature information corresponding to specific features of each associated object, but also the feature information corresponding to specific features of the target object.

[0170] It should be noted that the average value obtained by averaging serves as a supplement to the specific features of the target object and can be directly added to the features of the target object as the object features; summation provides the neighbor information of the nodes corresponding to the target object with missing features. The more source neighbor information is added, the higher the confidence of the supplemented features. Furthermore, the magnitude of the feature after summation reflects the number of neighbors of a node; standard deviation provides the expectation and degree of deviation between the supplemented features and the original features.

[0171] By applying the scheme of the embodiments in this specification, the feature information of specific features corresponding to each associated object is fused by the calculation methods of summation, averaging, and standard deviation. This makes the determined fused feature information more rigorous and more diverse, and further makes the object features of the determined target object more numerous and accurate.

[0172] See Figure 4a and Figure 4b , Figure 4a A flowchart of another feature generation method provided according to an embodiment of this specification is shown. Figure 4bThe flowchart shows another feature generation method provided according to an embodiment of the present specification, which specifically includes the following steps:

[0173] Step 402: Feature selection.

[0174] It should be noted that the coverage, information gain, importance, and IV value of each original feature are calculated, and then based on the feature index values of each original feature, Top-K specific features are selected from multiple original features. Here, the IV (Information Value) is the information value, which represents the contribution degree of the feature to the target prediction and the prediction ability of the feature. Generally speaking, the higher the IV value, the stronger the prediction ability of the feature and the higher the information contribution degree.

[0175] Step 404: Construct an object relationship graph.

[0176] Based on the specific features and the preset relationship table, an object relationship graph for the target object is constructed. The features included in the nodes of the object relationship graph are only specific features and the feature information of specific features, where "i" represents a node.

[0177] Step 406: Determine neighbor nodes.

[0178] Based on the object relationship graph for the target object, the neighbor nodes of the node corresponding to the target object are determined. Here, the "dashed box" represents the neighborhood information, and the neighbor nodes are determined based on the neighborhood information.

[0179] Step 408: Derived features.

[0180] Specifically, the derived feature refers to the object feature of the target object generated based on the feature information of the specific features corresponding to each associated object, as shown in Figure 4b "Derivation" in it.

[0181] The scheme implemented in this specification involves obtaining the original features of multiple objects and a preset relationship table, wherein the preset relationship table records the association relationships between multiple objects; performing a missing impact analysis on each original feature to obtain the missing impact degree of each original feature, wherein the missing impact degree characterizes the degree of influence of missing original features; selecting specific features from each original feature based on the missing impact degree of each original feature; for the target object among multiple objects, obtaining the feature information of the specific features corresponding to each associated object based on the specific features and the preset relationship table, wherein the associated object is an object that has an association relationship with the target object; and generating the object features of the target object based on the feature information of the specific features corresponding to each associated object. By performing a missing feature analysis on the original features of multiple objects, the missing feature degree of each original feature is obtained. In other words, the missing feature degree of each original feature can be used to determine the degree of impact of the missing original feature on the final result. Thus, specific features suitable for feature generation processing are determined from each original feature. Then, based on the obtained preset relationship table, the feature information of objects that are related to the target object and their corresponding specific features are determined, and the object features of the target object are determined. This makes the generated object features based on objects that are related to the target object, so the accuracy of the generated object features of the target object is high.

[0182] The following is in conjunction with the appendix Figure 5 Appendix Figure 6 Taking the application of the feature generation method provided in this specification in an information recommendation scenario as an example, the feature generation method will be further explained. Among other things, Figure 5 A flowchart illustrating the processing steps of a feature generation method according to an embodiment of this specification is shown. Figure 6 The diagram illustrates a process flow of a feature generation method according to an embodiment of this specification, specifically including the following steps.

[0183] Step 502: Obtain object A (time feature 11, spatial feature 10), object B (time feature null, spatial feature 15), object C (time feature 12, spatial feature 14), and object D (time feature 30, spatial feature 20), and ensure that objects A, B, C, and D all have a relationship of sending shopping links to each other.

[0184] The values ​​of the temporal and spatial characteristics of objects A, B, C, and D mentioned above are shown in [reference needed]. Figure 6 (a) in the middle.

[0185] Step 504: Statistically analyze the category attribute parameters of each time feature and spatial feature to obtain the influence index of time features and spatial features. The coverage of time features is 75%, the information gain is 0.8, and the importance is 0.8. The coverage of spatial features is 100%, the information gain is 0.6, and the importance is 0.7.

[0186] The values ​​of the characteristic indices for the temporal and spatial characteristics of objects A, B, C, and D mentioned above are shown in [reference needed]. Figure 6 (b) in the middle.

[0187] Step 506: Based on the influence indicators of time and space characteristics, determine the degree of influence of missing time and space characteristics, which is the influence indicator.

[0188] Step 508: Select specific features from the temporal and spatial features that have a coverage of less than 80% of the preset coverage threshold and an information gain greater than 0.7 of the preset information gain threshold as temporal features.

[0189] Step 510: For objects A, B, C, and D, construct an object relationship diagram based on the time characteristics and the relationships between the objects.

[0190] Step 512: Based on the object relationship graph, determine the neighbor nodes of the nodes corresponding to objects A, B, C, and D respectively; read the feature information of the time features corresponding to each associated object from the neighbor nodes, where object A is 11, object B is null, object C is 12, and object D is 30, and transfer these feature information to the corresponding target object.

[0191] The above process involves transferring time features. The changes in the values ​​of the time features of each object before and after the transfer are shown in the following figure. Figure 6 (c) in the middle.

[0192] After transferring the feature information to the corresponding target object, the transferred data can be processed to remove outliers. There are many ways to remove outliers, such as the MAD algorithm, the 3σ algorithm, the percentile method, etc.

[0193] Step 514: Perform summation, averaging, and standard deviation calculations on the time feature information corresponding to each associated object to obtain fused feature information.

[0194] Before the above-described fusion process, the sum, mean, and standard deviation of the temporal characteristic values ​​of each object are calculated, as shown in Figure 6(d); the specific steps for the above-described fusion process are detailed in [link to figure]. Figure 6In (e), "Fe1" is "numerical value", "avg" is "average value", "sd" is "standard deviation", and "null" is "non-existent". That is, if the numerical value of the time feature is "null", then the numerical value of the time feature becomes "average value of the time feature"; if the standard deviation of the numerical value of the time feature is greater than 5, then the numerical value of the time feature becomes "null".

[0195] Step 516: Based on the fused feature information and the original feature information, determine the object features of target object A, target object B, target object C, and object D.

[0196] The object characteristics of target objects A, B, and C obtained above are as follows: Figure 6 (f) in the middle.

[0197] The scheme implemented in this specification involves obtaining the original features of multiple objects and a preset relationship table, wherein the preset relationship table records the association relationships between multiple objects; performing a missing impact analysis on each original feature to obtain the missing impact degree of each original feature, wherein the missing impact degree characterizes the degree of influence of missing original features; selecting specific features from each original feature based on the missing impact degree of each original feature; for the target object among multiple objects, obtaining the feature information of the specific features corresponding to each associated object based on the specific features and the preset relationship table, wherein the associated object is an object that has an association relationship with the target object; and generating the object features of the target object based on the feature information of the specific features corresponding to each associated object. By performing a missing feature analysis on the original features of multiple objects, the missing feature degree of each original feature is obtained. In other words, the missing feature degree of each original feature can be used to determine the degree of impact of the missing original feature on the final result. Thus, specific features suitable for feature generation processing are determined from each original feature. Then, based on the obtained preset relationship table, the feature information of objects that are related to the target object and their corresponding specific features are determined, and the object features of the target object are determined. This makes the generated object features based on objects that are related to the target object, so the accuracy of the generated object features of the target object is high.

[0198] It should be noted that the technical solution of this feature generation method is similar to that described above. Figure 2 The technical solutions for feature generation methods shown belong to the same concept. Details not described in detail in these technical solutions can be found in [reference needed]. Figure 2 The technical solution of the feature generation method shown is described.

[0199] See Figure 7 , Figure 7 A flowchart of an information prediction method according to an embodiment of this specification is shown, which specifically includes the following steps.

[0200] Step 702: Obtain the object features of the target object, wherein the object features are based on the above... Figure 2 The feature generation method described herein generates the feature.

[0201] Specifically, the object characteristics of the target object refer to those formed by... Figure 2 The feature generation method in [the document] generates object features of the target object.

[0202] The solution implemented in this specification, based on the need to predict information about a target object, obtains the object features of the target object generated by the feature generation method. Subsequently, information prediction of the target object can be directly performed based on the obtained object features. With the high accuracy of the obtained object features, the accuracy of information prediction is further improved.

[0203] Step 704: Input the object features into a pre-trained information prediction model to obtain information prediction results for the target object, wherein the information prediction model is a machine learning model.

[0204] Specifically, an information prediction model refers to a machine learning model that can predict information about an object based on its characteristics.

[0205] By applying the scheme of the embodiments of this specification, the object features of the target object obtained by the feature generation method are input into a pre-trained information prediction model to obtain the information prediction result for the target object. By utilizing the efficiency of the model and the accuracy of the generated object features, the accuracy of information prediction is further improved.

[0206] See Figure 8 , Figure 8 A flowchart is shown of a training method for an information prediction model according to an embodiment of this specification, which specifically includes the following steps.

[0207] Step 802: Obtain the original features of multiple sample objects in the sample set and a preset relationship table, wherein the preset relationship table records the association relationships between the multiple sample objects.

[0208] Step 804: Perform a missing impact analysis on each original feature to obtain the missing impact degree of each original feature, wherein the missing impact degree characterizes the degree of influence of missing original features.

[0209] Step 806: Based on the degree of missing impact of each original feature, select specific features from the original features.

[0210] Step 808: For the target sample object, based on the specific feature and the preset relationship table, obtain the feature information of the specific feature corresponding to each associated sample object, wherein the target sample object is any one of the plurality of sample objects, and the associated sample object is a sample object that has an association relationship with the target sample object.

[0211] Step 810: Generate the object features of the target sample object based on the feature information of the specific features corresponding to each associated sample object.

[0212] Steps 802 to 810 and Figure 2 Steps 202 to 210 in the illustrated embodiment are the same or similar, as can be seen in detail. Figure 2 The embodiments shown are not described in detail here.

[0213] Step 812: Based on the object features of the multiple sample objects, train the information prediction model to obtain the trained information prediction model.

[0214] Specifically, an information prediction model refers to a model that can obtain information prediction results for a sample object based on the input object features.

[0215] It should be noted that the trained information prediction model can be obtained by training the information prediction model using the object features of multiple sample objects.

[0216] By applying the scheme of the embodiments of this specification, the information prediction model is trained by acquiring the object features of multiple sample objects, and the trained information prediction model is obtained. Through training, the accuracy of the information prediction model is improved.

[0217] The scheme implemented in this specification involves obtaining the original features of multiple sample objects in a sample set and a preset relationship table, wherein the preset relationship table records the association relationships between the multiple sample objects; performing a missing feature impact analysis on each original feature to obtain the missing feature impact degree of each original feature, wherein the missing feature impact degree characterizes the degree of influence of missing original features; selecting specific features from each original feature based on the missing feature impact degree of each original feature; for a target sample object, obtaining feature information of the specific features corresponding to each associated sample object based on the specific features and the preset relationship table, wherein the target sample object is any one of the multiple sample objects, and the associated sample objects are sample objects that are associated with the target sample object; generating object features of the target sample object based on the feature information of the specific features corresponding to each associated sample object; and training an information prediction model based on the object features of multiple sample objects to obtain a trained information prediction model. This paper implements a missing feature analysis on multiple acquired object features, filters out specific features, and then generates object features of sample objects based on associated objects and specific features, resulting in high accuracy of the generated object features. Furthermore, by training the information prediction model with the object features of multiple acquired sample objects, the accuracy of the information prediction model is improved through training.

[0218] With the above Figure 2 Corresponding to the feature generation method embodiments shown, this specification also provides feature generation apparatus embodiments. Figure 9 A schematic diagram of a feature generation apparatus according to one embodiment of this specification is shown. Figure 9 As shown, the device includes:

[0219] The first acquisition module 902 is configured to acquire the original features of multiple objects and a preset relationship table, wherein the preset relationship table records the association relationships between multiple objects;

[0220] The first analysis module 904 is configured to perform a missing impact analysis on each original feature to obtain the missing impact degree of each original feature, wherein the missing impact degree characterizes the degree of impact of missing original features.

[0221] The first screening module 906 is configured to select specific features from the original features based on the degree of missing impact of each original feature;

[0222] The first acquisition module 908 is configured to, for a target object among multiple objects, obtain feature information of specific features corresponding to each associated object based on specific features and a preset relationship table, wherein the associated object is an object that has an association relationship with the target object;

[0223] The first generation module 910 is configured to generate object features of the target object based on the feature information of the specific features corresponding to each associated object.

[0224] Optionally, the first analysis module 904 is further configured to: statistically analyze the category attribute parameters of each original feature; calculate the influence index of each original feature based on the category attribute parameters of each original feature; and determine the missing influence degree of each original feature based on the influence index of each original feature.

[0225] Optionally, the missing impact includes coverage and information gain; the first filtering module 906 is further configured to filter out specific features from each original feature whose coverage is less than a preset coverage threshold and whose information gain is greater than a preset information gain threshold.

[0226] Optionally, the first obtaining module 908 is further configured to construct an object relationship graph for a target object among multiple objects, based on specific features and a preset relationship table, wherein each node in the object relationship graph records feature information of specific features, and the edges in the object relationship graph represent association relationships; based on the object relationship graph, determine each neighbor node of the node corresponding to the target object; and read feature information of specific features corresponding to each associated object from each neighbor node.

[0227] Optionally, the first generation module 910 is further configured to perform fusion processing on the feature information of specific features corresponding to each associated object to obtain fused feature information; and generate object features of the target object based on the fused feature information.

[0228] Optionally, the first generation module 910 is further configured to determine the object features of the target object based on the fused feature information and the original feature information, provided that the target object has original feature information with specific characteristics.

[0229] Optionally, the fusion process includes at least one of the following three: averaging, summing, and standard deviation.

[0230] The scheme implemented in this specification involves obtaining the original features of multiple objects and a preset relationship table, wherein the preset relationship table records the association relationships between multiple objects; performing a missing impact analysis on each original feature to obtain the missing impact degree of each original feature, wherein the missing impact degree characterizes the degree of influence of missing original features; selecting specific features from each original feature based on the missing impact degree of each original feature; for the target object among multiple objects, obtaining the feature information of the specific features corresponding to each associated object based on the specific features and the preset relationship table, wherein the associated object is an object that has an association relationship with the target object; and generating the object features of the target object based on the feature information of the specific features corresponding to each associated object. By performing a missing feature analysis on the original features of multiple objects, the missing feature degree of each original feature is obtained. In other words, the missing feature degree of each original feature can be used to determine the degree of impact of the missing original feature on the final result. Thus, specific features suitable for feature generation processing are determined from each original feature. Then, based on the obtained preset relationship table, the feature information of objects that are related to the target object and their corresponding specific features are determined, and the object features of the target object are determined. This makes the generated object features based on objects that are related to the target object, so the accuracy of the generated object features of the target object is high.

[0231] The above is an illustrative scheme of a feature generation device according to this embodiment. It should be noted that the technical solution of this feature generation device is similar to that described above. Figure 2 The technical solutions for the feature generation methods shown belong to the same concept. For details not described in the technical solutions for the feature generation devices, please refer to the above. Figure 2 The technical solution of the feature generation method shown is described.

[0232] With the above Figure 7 Corresponding to the information prediction method embodiments shown, this specification also provides information prediction device embodiments. Figure 10 A schematic diagram of an information prediction device according to one embodiment of this specification is shown. Figure 10 As shown, the device includes:

[0233] The second acquisition module 1002 is configured to acquire object features of the target object, wherein the object features are generated according to any one of the feature generation methods of claims 1 to 7;

[0234] The prediction module 1004 is configured to input object features into a pre-trained information prediction model to obtain information prediction results for the target object, wherein the information prediction model is a machine learning model.

[0235] By applying the scheme of the embodiments of this specification, the object features of the target object obtained by the feature generation method are input into a pre-trained information prediction model to obtain the information prediction result for the target object. By utilizing the efficiency of the model and the accuracy of the generated object features, the accuracy of information prediction is further improved.

[0236] The above is an illustrative scheme of an information prediction device according to this embodiment. It should be noted that the technical solution of this information prediction device is similar to that described above. Figure 7 The technical solutions for the information prediction methods shown belong to the same concept. For details not described in the technical solutions for the information prediction devices, please refer to the above. Figure 7 The technical solution of the information prediction method shown is described.

[0237] With the above Figure 8 Corresponding to the training method embodiment of the information prediction model shown, this specification also provides an embodiment of the training device for the information prediction model. Figure 11 A schematic diagram of a training apparatus for an information prediction model according to one embodiment of this specification is shown. Figure 11 As shown, the device includes:

[0238] The third acquisition module 1102 is configured to acquire the original features of multiple sample objects in the sample set and a preset relationship table, wherein the preset relationship table records the association relationship between multiple sample objects;

[0239] The second analysis module 1104 is configured to perform a missing impact analysis on each original feature to obtain the missing impact degree of each original feature, wherein the missing impact degree characterizes the degree of impact of missing original features.

[0240] The second filtering module 1106 is configured to filter out specific features from the original features based on the degree of missing impact of each original feature;

[0241] The second obtaining module 1108 is configured to obtain feature information of specific features corresponding to each associated sample object based on specific features and a preset relationship table for the target sample object. The target sample object is any one of multiple sample objects, and the associated sample object is a sample object that has an association relationship with the target sample object.

[0242] The second generation module 1110 is configured to generate object features of the target sample object based on the feature information of the specific features corresponding to each associated sample object;

[0243] Training module 1112 is configured to train the information prediction model based on the object features of multiple sample objects, and obtain the trained information prediction model.

[0244] The scheme implemented in this specification involves obtaining the original features of multiple sample objects in a sample set and a preset relationship table, wherein the preset relationship table records the association relationships between the multiple sample objects; performing a missing feature impact analysis on each original feature to obtain the missing feature impact degree of each original feature, wherein the missing feature impact degree characterizes the degree of influence of missing original features; selecting specific features from each original feature based on the missing feature impact degree of each original feature; for a target sample object, obtaining feature information of the specific features corresponding to each associated sample object based on the specific features and the preset relationship table, wherein the target sample object is any one of the multiple sample objects, and the associated sample objects are sample objects that are associated with the target sample object; generating object features of the target sample object based on the feature information of the specific features corresponding to each associated sample object; and training an information prediction model based on the object features of multiple sample objects to obtain a trained information prediction model. This paper implements a missing feature analysis on multiple acquired object features, filters out specific features, and then generates object features of sample objects based on associated objects and specific features, resulting in high accuracy of the generated object features. Furthermore, by training the information prediction model with the object features of multiple acquired sample objects, the accuracy of the information prediction model is improved through training.

[0245] The above is an illustrative scheme of a training device for an information prediction model according to this embodiment. It should be noted that the technical solution of this training device for the information prediction model is similar to that described above. Figure 8 The technical solutions for the training methods of the information prediction models shown belong to the same concept. Details not described in the technical solutions for the training devices of the information prediction models can be found above. Figure 8 The technical solution for training the information prediction model shown is described.

[0246] Figure 12 A structural block diagram of a computing device according to one embodiment of this specification is shown. The components of the computing device 1200 include, but are not limited to, a memory 1210 and a processor 1220. The processor 1220 is connected to the memory 1210 via a bus 1230, and a database 1250 is used to store data.

[0247] The computing device 1200 also includes an access device 1240, which enables the computing device 1200 to communicate via one or more networks 1260. Examples of such networks include Public Switched Telephone Network (PSTN), Local Area Network (LAN), Wide Area Network (WAN), Personal Area Network (PAN), or combinations of communication networks such as the Internet. The access device 1240 may include one or more of any type of wired or wireless network interface (e.g., Network Interface Card (NIC)), such as an IEEE 802.11 Wireless Local Area Networks (WLAN) interface, a Wi-MAX (World Interoperability for Microwave Access) interface, an Ethernet interface, a Universal Serial Bus (USB) interface, a cellular network interface, a Bluetooth interface, a Near Field Communication (NFC) interface, and so on.

[0248] In one embodiment of this specification, the aforementioned components of the computing device 1200 and Figure 12 Other components, not shown, can also be connected to each other, for example, via a bus. It should be understood that... Figure 12 The block diagram of the computing device shown is for illustrative purposes only and is not intended to limit the scope of this specification. Those skilled in the art can add or replace other components as needed.

[0249] The computing device 1200 can be any type of stationary or mobile computing device, including mobile computers or mobile computing devices (e.g., tablet computers, personal digital assistants, laptop computers, notebook computers, netbooks, etc.), mobile phones (e.g., smartphones), wearable computing devices (e.g., smartwatches, smart glasses, etc.) or other types of mobile devices, or stationary computing devices such as desktop computers or PCs. The computing device 1200 can also be a mobile or stationary server.

[0250] The processor 1220 is configured to execute the following computer-executable instructions, which, when executed by the processor, implement the above-mentioned... Figure 2 The feature generation method shown or Figure 7 The information prediction method shown or Figure 8The steps of training the information prediction model shown are as follows.

[0251] The above is an illustrative scheme of a computing device according to this embodiment. It should be noted that the technical solution of this computing device is similar to that described above. Figure 2 The feature generation method shown or Figure 7 The information prediction method shown or Figure 8 The technical solutions for training the information prediction model shown belong to the same concept. Details not described in the technical solution for the computing device can be found in the above-mentioned... Figure 2 The feature generation method shown or Figure 7 The information prediction method shown or Figure 8 The technical solution for training the information prediction model shown is described.

[0252] This specification also provides a computer-readable storage medium storing computer-executable instructions that, when executed by a processor, implement the above-described embodiments. Figure 2 The feature generation method shown or Figure 7 The information prediction method shown or Figure 8 The steps of training the information prediction model shown are as follows.

[0253] The above is an illustrative scheme of a computer-readable storage medium according to this embodiment. It should be noted that the technical solution of this storage medium is different from that described above. Figure 2 The feature generation method shown or Figure 7 The information prediction method shown or Figure 8 The technical solutions for training the information prediction model shown belong to the same concept. For details regarding the storage medium technology solution not described in detail above, please refer to the above. Figure 2 The feature generation method shown or Figure 7 The information prediction method shown or Figure 8 The technical solution for training the information prediction model shown is described.

[0254] An embodiment of this specification also provides a computer program, wherein when the computer program is executed in a computer, it causes the computer to perform the above-described... Figure 2 The feature generation method shown or Figure 7 The information prediction method shown or Figure 8 The steps of training the information prediction model shown are as follows.

[0255] The above is an illustrative scheme of a computer program according to this embodiment. It should be noted that the technical solution of this computer program is similar to that described above. Figure 2 The feature generation method shown or Figure 7 The information prediction method shown or Figure 8The technical solutions for training the information prediction models shown belong to the same concept. Details not described in the technical solutions of the computer programs can be found above. Figure 2 The feature generation method shown or Figure 7 The information prediction method shown or Figure 8 The technical solution for training the information prediction model shown is described.

[0256] The foregoing has described specific embodiments of this specification. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recited in the claims may be performed in a different order than that shown in the embodiments and may still achieve the desired result. Furthermore, the processes depicted in the drawings do not necessarily require the specific or sequential order shown to achieve the desired result. In some embodiments, multitasking and parallel processing are possible or may be advantageous.

[0257] The computer instructions include computer program code, which may be in the form of source code, object code, executable file, or some intermediate form. The computer-readable medium may include: any entity or device capable of carrying the computer program code, recording media, USB flash drive, portable hard drive, magnetic disk, optical disk, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc. It should be noted that the content included in the computer-readable medium may be appropriately added to or subtracted according to the requirements of legislation and patent practice in the jurisdiction. For example, in some jurisdictions, according to legislation and patent practice, computer-readable media may not include electrical carrier signals and telecommunication signals.

[0258] It should be noted that, for the sake of simplicity, the foregoing method embodiments are all described as a series of actions. However, those skilled in the art should understand that the embodiments in this specification are not limited to the described order of actions, because according to the embodiments in this specification, some steps can be performed in other orders or simultaneously. Furthermore, those skilled in the art should also understand that the embodiments described in this specification are all preferred embodiments, and the actions and modules involved are not necessarily essential to the embodiments in this specification.

[0259] In the above embodiments, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0260] The preferred embodiments disclosed above are merely illustrative of this specification. The optional embodiments do not exhaustively describe all details, nor do they limit the invention to the specific implementations described. Clearly, many modifications and variations can be made based on the embodiments described herein. These embodiments are selected and specifically described in this specification to better explain the principles and practical applications of the embodiments, thereby enabling those skilled in the art to better understand and utilize this specification. This specification is limited only by the claims and their full scope and equivalents.

Claims

1. A feature generation method, comprising: Obtain the original features of multiple objects and a preset relationship table, wherein the preset relationship table records the association relationships between the multiple objects, and the original features refer to the original features obtained based on the user's activity information on the Internet, including time features, login features, and information features; A missing feature impact analysis is performed on each original feature to obtain the missing feature impact degree of each original feature, wherein the missing feature impact degree characterizes the degree of impact of missing original features, and the missing feature impact degree includes coverage and information gain; Based on the degree of impact of the absence of each original feature, specific features are selected from the original features. The process of selecting specific features from the original features based on the degree of impact of the absence of each original feature includes: selecting specific features from the original features whose coverage is less than a preset coverage threshold and whose information gain is greater than a preset information gain threshold. For a target object among the plurality of objects, feature information of the specific feature corresponding to each associated object is obtained according to the specific feature and the preset relationship table, wherein the associated object is an object that has an association relationship with the target object; Based on the feature information of the specific features corresponding to each associated object, the object features of the target object are generated.

2. The method according to claim 1, wherein performing a missing feature analysis on each original feature to obtain the missing feature degree of each original feature includes: Statistically analyze the category attribute parameters of each of the original features; Calculate the influence index of each original feature based on the category attribute parameters of each original feature; Based on the influence index of each original feature, the degree of missing influence of each original feature is determined.

3. The method according to claim 1, wherein obtaining feature information of the specific feature corresponding to each associated object based on the specific feature and the preset relationship table for the target object among the plurality of objects includes: For the target object among the plurality of objects, an object relationship graph is constructed based on the specific features and the preset relationship table, wherein each node in the object relationship graph records the feature information of the specific features, and the edges in the object relationship graph represent the association relationship; Based on the object relationship graph, determine each neighbor node of the node corresponding to the target object; Read the feature information of the specific feature corresponding to each associated object from each neighbor node.

4. The method according to claim 1, wherein generating the object feature of the target object based on the feature information of the specific feature corresponding to each associated object includes: The feature information of the specific features corresponding to each associated object is fused to obtain fused feature information; Based on the fused feature information, the object features of the target object are generated.

5. The method according to claim 4, wherein generating the object features of the target object based on the fused feature information comprises: When the target object has the original feature information of the specific feature, the object feature of the target object is determined based on the fused feature information and the original feature information.

6. The method according to claim 4 or 5, wherein the fusion process includes at least one of the following three: averaging, summing, and standard deviation.

7. An information prediction method, comprising: Obtain the object features of the target object, wherein the object features are generated by the feature generation method according to any one of claims 1 to 6; The object features are input into a pre-trained information prediction model to obtain information prediction results for the target object, wherein the information prediction model is a machine learning model.

8. A training method for an information prediction model, comprising: Obtain the original features of multiple sample objects in the sample set and a preset relationship table, wherein the preset relationship table records the association relationship between the multiple sample objects, and the original features refer to the original features obtained based on the user's activity information on the Internet, including time features, login features, and information features; A missing feature impact analysis is performed on each original feature to obtain the missing feature impact degree of each original feature, wherein the missing feature impact degree characterizes the degree of impact of missing original features, and the missing feature impact degree includes coverage and information gain; Based on the degree of impact of the absence of each original feature, specific features are selected from the original features. The process of selecting specific features from the original features based on the degree of impact of the absence of each original feature includes: selecting specific features from the original features whose coverage is less than a preset coverage threshold and whose information gain is greater than a preset information gain threshold. For a target sample object, feature information of the specific feature corresponding to each associated sample object is obtained according to the specific feature and the preset relationship table, wherein the target sample object is any one of the plurality of sample objects, and the associated sample object is a sample object that has an association relationship with the target sample object; Based on the feature information of the specific features corresponding to each associated sample object, the object features of the target sample object are generated; Based on the object features of the multiple sample objects, the information prediction model is trained to obtain the trained information prediction model.

9. A feature generation apparatus, comprising: The first acquisition module is configured to acquire the original features of multiple objects and a preset relationship table. The preset relationship table records the association relationships between the multiple objects. The original features refer to the original features acquired based on the user's activity information on the Internet, including time features, login features, and information features. The first analysis module is configured to perform a missing impact analysis on each original feature to obtain the missing impact degree of each original feature, wherein the missing impact degree characterizes the degree of impact of missing original features, and the missing impact degree includes coverage and information gain; The first filtering module is configured to filter specific features from the original features based on the missing impact of each original feature. The filtering of specific features from the original features based on the missing impact of each original feature includes: filtering specific features from the original features with a coverage less than a preset coverage threshold and an information gain greater than a preset information gain threshold. The first obtaining module is configured to, for a target object among the plurality of objects, obtain feature information of the specific feature corresponding to each associated object based on the specific feature and the preset relationship table, wherein the associated object is an object that has an association relationship with the target object; The first generation module is configured to generate object features of the target object based on the feature information of the specific features corresponding to each associated object.

10. An information prediction device, comprising: The second acquisition module is configured to acquire object features of the target object, wherein the object features are generated by the feature generation method according to any one of claims 1 to 6; The prediction module is configured to input the object features into a pre-trained information prediction model to obtain an information prediction result for the target object, wherein the information prediction model is a machine learning model.

11. A training device for an information prediction model, comprising: The third acquisition module is configured to acquire the original features of multiple sample objects in the sample set and a preset relationship table. The preset relationship table records the association between the multiple sample objects. The original features refer to the original features obtained based on the user's activity information on the Internet, including time features, login features, and information features. The second analysis module is configured to perform a missing impact analysis on each original feature to obtain the missing impact degree of each original feature, wherein the missing impact degree characterizes the degree of impact of missing original features, and the missing impact degree includes coverage and information gain; The second filtering module is configured to filter out specific features from the original features based on the degree of absence of each original feature. The step of filtering out specific features from the original features based on the degree of absence of each original feature includes: filtering out specific features from the original features whose coverage is less than a preset coverage threshold and whose information gain is greater than a preset information gain threshold. The second obtaining module is configured to obtain feature information of the specific feature corresponding to each associated sample object based on the specific feature and the preset relationship table for the target sample object, wherein the target sample object is any one of the plurality of sample objects, and the associated sample object is a sample object that has an association relationship with the target sample object; The second generation module is configured to generate object features of the target sample object based on the feature information of the specific features corresponding to each associated sample object; The training module is configured to train the information prediction model based on the object features of the multiple sample objects, and obtain the trained information prediction model.

12. A computing device, comprising: Memory and processor; The memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions. When the computer-executable instructions are executed by the processor, they implement the steps of the feature generation method of any one of claims 1 to 6, the information prediction method of claim 7, or the training method of the information prediction model of claim 8.

13. A computer-readable storage medium storing computer-executable instructions that, when executed by a processor, implement the steps of the feature generation method of any one of claims 1 to 6, the information prediction method of claim 7, or the training method of the information prediction model of claim 8.

Images