CN117194966A

CN117194966A - Training method and related device for object classification model

Info

Publication number: CN117194966A
Application number: CN202210589182.5A
Authority: CN
Inventors: 樊鹏
Original assignee: Tencent Technology Shenzhen Co Ltd
Current assignee: Tencent Technology Shenzhen Co Ltd
Priority date: 2022-05-27
Filing date: 2022-05-27
Publication date: 2023-12-08

Abstract

The application discloses a training method and a related device of an object classification model, which can be used for scenes such as cloud technology, artificial intelligence, intelligent traffic, auxiliary driving and the like, and can be used for acquiring a sample object data set comprising a first positive sample set and an unlabeled sample set. And removing the intermediate sample from the first positive sample set to obtain a second positive sample set, and adding the intermediate sample to the unlabeled sample set to obtain an original sample set to be labeled. And training a neural network model on the second positive sample set and the original sample set to be annotated according to the object characteristics of the sample objects in the constructed sample object data set to obtain an object classification model. In the process of training the object classification model, the neural network model obtained through training is utilized to classify the sample objects in the original sample set to be marked until each sample object in the original sample set to be marked has a classification label, so that positive and negative sample are balanced, the classification accuracy of the sample objects in the original sample set to be marked is ensured, and the business effect of information recommendation is improved.

Description

Training method and related device for object classification model

Technical Field

The application relates to the technical field of computers, in particular to a training method and a related device of an object classification model.

Background

In the operation process of various products, the products need to be promoted, so that new users are continuously introduced to the products. In general, users who may be interested in the product should be selected as a target of information related to the product to improve popularization efficiency.

For this reason, it is necessary to recognize the user's intention to judge the user who may be interested in the product as a delivery object. In the related art, although users who may be interested in a product may be identified by training an object classification model, information may be recommended to the users by using the object classification model obtained by training.

However, the object classification model provided by the related technology is not very high in recognition accuracy of whether the user is a user interested in the product, so that the service effect of information recommendation is reduced.

Disclosure of Invention

In order to solve the technical problems, the application provides a training method and a related device for an object classification model, which effectively balance initial positive and negative samples, ensure better generalization in the classification process, ensure the classification accuracy of sample objects in an original sample set to be marked, and further improve the service effect of follow-up information recommendation.

The embodiment of the application discloses the following technical scheme:

in one aspect, an embodiment of the present application provides a training method for an object classification model, where the method includes:

obtaining a sample object data set aiming at a target product, wherein the sample object data set comprises a first positive sample set and an unlabeled sample set, the sample objects in the first positive sample set are provided with positive labels, the positive labels are classification labels, the positive labels are used for marking the sample objects with willingness to execute conversion operation aiming at the target product, the sample objects in the unlabeled sample set are not provided with classification labels, and the classification labels are used for marking whether the sample objects have willingness to execute conversion operation aiming at the target product or not;

constructing object features of a sample object in the sample object dataset;

selecting a plurality of sample objects from the first positive sample set as intermediate samples, removing the intermediate samples from the first positive sample set to obtain a second positive sample set, and adding the intermediate samples to the unlabeled sample set to obtain an original sample set to be labeled;

training a neural network model on the second positive sample set and the original sample set to be annotated according to the object characteristics to obtain the object classification model;

And in the process of training the neural network model to obtain the object classification model, classifying the sample objects in the original sample set to be marked by using the neural network model obtained in the training process until each sample object in the original sample set to be marked has a classification label, so as to obtain a target positive example set, a target negative example set and the object classification model.

In one aspect, an embodiment of the present application provides a training method for an object classification model, where the apparatus includes an obtaining unit, a building unit, a determining unit, and a training unit:

the obtaining unit is configured to obtain a sample object data set for a target product, where the sample object data set includes a first positive sample set and an unlabeled sample set, the sample objects in the first positive sample set have positive labels, the positive labels are classification labels, the positive labels are used for marking sample objects with willingness to perform a conversion operation for the target product, the sample objects in the unlabeled sample set do not have classification labels, and the classification labels are used for marking whether the sample objects have willingness to perform a conversion operation for the target product;

The construction unit is used for constructing object features of the sample objects in the sample object data set;

the determining unit is used for selecting a plurality of sample objects from the first positive sample set as intermediate samples, removing the intermediate samples from the first positive sample set to obtain a second positive sample set, and adding the intermediate samples to the unlabeled sample set to obtain an original sample set to be labeled;

the training unit is used for training a neural network model on the second positive sample set and the original sample set to be annotated according to the object characteristics to obtain the object classification model;

in the process of training the neural network model to obtain the object classification model, the training unit is used for classifying the sample objects in the original sample set to be marked by using the neural network model obtained in the training process until each sample object in the original sample set to be marked has a classification label, so as to obtain a target positive example set, a target negative example set and the object classification model.

In one aspect, an embodiment of the present application provides a computer device including a processor and a memory:

The memory is used for storing program codes and transmitting the program codes to the processor;

the processor is configured to perform the method of any of the preceding aspects according to instructions in the program code.

In one aspect, embodiments of the present application provide a computer readable storage medium for storing program code for performing the method of any one of the preceding aspects.

In one aspect, embodiments of the present application provide a computer program product comprising a computer program which, when executed by a processor, implements the method of any of the preceding aspects.

According to the technical scheme, when the object classification model is trained, a sample object data set aiming at a target product can be obtained, the sample object data set comprises a first positive sample set and a non-labeled sample set, sample objects in the first positive sample set are provided with positive labels, the positive labels are used for marking the sample objects with willingness to execute conversion operation aiming at the target product, the sample objects in the non-labeled sample set are not provided with classification labels, and the classification labels are used for marking whether the sample objects have willingness to execute conversion operation aiming at the target product or not. Selecting a plurality of sample objects from the first positive sample set as intermediate samples, removing the intermediate samples from the first positive sample set to obtain a second positive sample set, and adding the intermediate samples to the unlabeled sample set to obtain an original sample set to be labeled. And training a neural network model on the second positive sample set and the original sample set to be annotated according to the object characteristics of the sample objects in the constructed sample object data set to obtain an object classification model. In the process of training the neural network model to obtain an object classification model, classifying sample objects in an original sample set to be marked by using the neural network model obtained in the training process until each sample object in the original sample set to be marked has a classification label, so as to obtain a target positive example set, a target negative example set and an object classification model. Because the intermediate sample is a positive sample selected from the first positive sample set and the behavior of the unknown positive sample in the original sample set to be marked is consistent, the unknown positive sample can be reliably evaluated by using the intermediate sample, and the positive and negative samples are effectively balanced, so that the positive and negative samples have better generalization in the classification process, and the sample objects in the original sample set to be marked are better classified. And because the object classification model is continuously updated to divide the sample objects in the original sample set to be marked, the sample objects in the original sample set to be marked without the marked sample can be purified each time under the condition that the object classification model is in the best state, the classification accuracy of the sample objects in the original sample set to be marked is ensured, and the service effect of the follow-up information recommendation is further improved.

Drawings

In order to more clearly illustrate the embodiments of the application or the technical solutions of the prior art, the drawings that are required in the embodiments or the description of the prior art will be briefly described, it being obvious that the drawings in the following description are only some embodiments of the application, and that other drawings may be obtained according to these drawings without inventive faculty for a person skilled in the art.

Fig. 1 is an application scenario architecture diagram of a training method of an object classification model according to an embodiment of the present application;

FIG. 2 is a flowchart of a training method of an object classification model according to an embodiment of the present application;

FIG. 3 is a diagram illustrating an exemplary architecture of a training method for an object classification model according to an embodiment of the present application;

FIG. 4 is an exemplary diagram of an aggregation approach provided by an embodiment of the present application;

FIG. 5 is a block diagram of an ISPYST algorithm framework provided by an embodiment of the application;

FIG. 6 is a flowchart of performing parameter optimization on a neural network model based on QPSO according to an embodiment of the present application;

FIG. 7 is a flow chart of a method for quantum measurement according to an embodiment of the present application;

fig. 8 is a diagram illustrating an example network structure of a CNN according to an embodiment of the present application;

FIG. 9 is a schematic diagram of a convolution operation of a convolution layer according to an embodiment of the present disclosure;

FIG. 10 is a schematic diagram illustrating a pooling operation of a pooling layer according to an embodiment of the present application;

fig. 11 is a schematic structural diagram of a fully-connected layer according to an embodiment of the present application;

FIG. 12 is a detailed flowchart of a method for training an object classification model based on an ISPYST algorithm and BQPSO according to an embodiment of the present application;

fig. 13 is a schematic diagram showing comparison of business effects of a different method according to an embodiment of the present application;

FIG. 14 is a block diagram of a training apparatus for an object classification model according to an embodiment of the present application;

fig. 15 is a block diagram of a terminal according to an embodiment of the present application;

fig. 16 is a block diagram of a server according to an embodiment of the present application.

Detailed Description

Embodiments of the present application are described below with reference to the accompanying drawings.

The object classification model provided in the related art is determined based on a non-deep learning method, and can perform object classification. However, in the scene of "predicting whether an object has a desire to perform a conversion operation for a target product", the object features are complex, and the object features are difficult to be explicitly expressed in terms of data characterization. The method is not suitable for the scene of predicting whether an object has willingness to execute conversion operation aiming at a target product, and is easy to form a data island and poor in service effect.

In addition, because the labeling needs to consume a large amount of manpower and material resources, the sample objects with the classification labels are fewer, and even the problems of abnormal proportion of positive and negative samples and serious lower positive samples exist.

In order to solve the technical problems, the embodiment of the application provides a training method and a related device for an object classification model, wherein a positive sample selected from a first positive sample set is used as an intermediate sample to be added into an unlabeled sample set to obtain an original sample set to be labeled, and the intermediate sample is removed from the first positive sample set to obtain a second positive sample set, so that a neural network model is trained on the second positive sample set and the original sample set to be labeled to obtain the object classification model. Because the intermediate sample is a positive sample selected from the first positive sample set and the behavior of the unknown positive sample in the original sample set to be marked is consistent, the unknown positive sample can be reliably evaluated by using the intermediate sample, and the positive and negative samples are effectively balanced, so that the positive and negative samples have better generalization in the classification process, and the sample objects in the original sample set to be marked are better classified. And because the object classification model is continuously updated to divide the sample objects in the original sample set to be marked, the sample objects in the original sample set to be marked without the marked sample can be purified each time under the condition that the object classification model is in the best state, the classification accuracy of the sample objects in the original sample set to be marked is ensured, and the service effect of the follow-up information recommendation is further improved.

The embodiment of the application can be applied to various object classification scenes, wherein the object classification is mainly used for distinguishing whether the object has the intention of executing the conversion operation on the target product or not, so that in the subsequent information recommendation process, the related information of the target product is recommended to the object with the intention of executing the conversion operation on the target product, thereby improving the service effect and providing direct help for the refined operation. The conversion operation may be an operation that brings benefits or advantageous effects to the operation, and the object or sample object may be a user, for example.

In different scenarios, the intent of performing the conversion operation for the target product may be different, for example in a video conference scenario, the target product may be a video conference payment, and the intent of performing the conversion operation for the target product may be a user's intent of having a video conference payment, i.e. the conversion operation is a payment. The video conference can use video conference software, and the video conference software has the functions of 300-person online conference, full-platform one-key access, intelligent audio and video noise reduction, beauty, background blurring, conference locking, screen watermarking and the like. The video conference software also provides a real-time sharing screen, supports online document collaboration and the like; in another example, in a second-hand cart purchase scenario, the target product may be a second-hand cart, and whether the user has a desire to perform a conversion operation for the target product may be whether the user has a desire to purchase the second-hand cart, i.e., the conversion operation is a purchase; for another example, in a video viewing scenario, the target product may be video, whether there is a desire to perform a conversion operation for the target product may be whether the user has a desire to view the video, i.e., the conversion operation is viewing, and so on.

It should be noted that, the embodiment of the present application may be implemented by means of Cloud Technology (Cloud Technology), which refers to a hosting Technology that unifies serial resources such as hardware, software, networks, etc. in a wide area network or a local area network, so as to implement calculation, storage, processing and sharing of data. The cloud technology is a generic term of network technology, information technology, integration technology, management platform technology, application technology and the like based on cloud computing business model application, can form a resource pool, and is flexible and convenient as required.

For example, cloud computing (cloud computing) is a computing model that distributes computing tasks over a resource pool of a large number of computers, enabling various application systems to acquire computing power, storage space, and information services as needed. The network that provides the resources is referred to as the "cloud". Resources in the cloud are infinitely expandable in the sense of users, and can be acquired at any time, used as needed, expanded at any time and paid for use as needed.

In another example, big data (Big data) refers to a data set which cannot be captured, managed and processed by a conventional software tool within a certain time range, and is a massive, high-growth-rate and diversified information asset which needs a new processing mode to have stronger decision-making ability, insight discovery ability and flow optimization ability. With the advent of the cloud age, big data has attracted more and more attention, and special techniques are required for big data to effectively process a large amount of data within a tolerant elapsed time. Technologies applicable to big data include massively parallel processing databases, data mining, distributed file systems, distributed databases, cloud computing platforms, the internet, and scalable storage systems.

Embodiments of the present application may also relate to, for example, artificial intelligence (Artificial Intelligence, AI), which is a theory, method, technique, and application system that simulates, extends, and extends human intelligence using a digital computer or a machine controlled by a digital computer, senses the environment, obtains knowledge, and uses the knowledge to obtain optimal results. The artificial intelligence software technology mainly comprises a computer vision technology, a voice processing technology, a natural language processing technology, machine learning/deep learning and other directions.

Machine Learning (ML) is a multi-domain interdisciplinary, involving multiple disciplines such as probability theory, statistics, approximation theory, convex analysis, algorithm complexity theory, etc. It is specially studied how a computer simulates or implements learning behavior of a human to acquire new knowledge or skills, and reorganizes existing knowledge structures to continuously improve own performance. Machine learning is the core of artificial intelligence, a fundamental approach to letting computers have intelligence, which is applied throughout various areas of artificial intelligence. Machine learning and deep learning typically include techniques such as artificial neural networks, confidence networks, reinforcement learning, transfer learning, induction learning, teaching learning, and the like. For example, object classification models may be trained using machine learning and deep learning.

It should be noted that, in the embodiment of the present application, the training method of the object classification model may be performed by a computer device, where the computer device may be a terminal, a server, or a combination of a terminal and a server, which is not limited in the embodiment of the present application.

As shown in fig. 1, fig. 1 shows an application scenario architecture diagram of a training method of an object classification model. In this application scenario, taking a server as an example, the application scenario may include a server 101 and a terminal 102, where the server 101 may be an independent physical server, may be a server cluster or a distributed system formed by a plurality of physical servers, or may be a cloud server that provides cloud computing services. The terminal 102 may be a terminal corresponding to an object to be identified, and the terminal 102 may be a smart phone, a tablet computer, a notebook computer, a desktop computer, an intelligent voice interaction device, an intelligent home appliance, a vehicle-mounted terminal, an aircraft, or the like, but is not limited thereto. The terminal 102 and the server 101 may be directly or indirectly connected through wired or wireless communication, and the present application is not limited herein.

The method provided by the embodiment of the application can be applied to various scenes needing object classification, including but not limited to cloud technology, artificial intelligence, intelligent transportation, auxiliary driving and the like.

In the application scenario architecture diagram shown in fig. 1, the server 101 may be configured to train an object classification model, and after the training of the object classification model is completed, the server 101 may use the object classification model to perform object classification online, so as to perform recommendation of related information of the target product according to the classification result.

In training the object classification model, the server 101 may obtain a sample object data set for a target product, where the target product may be an operator provided product of various kinds to perform conversion operations. The sample object data set comprises a first positive sample set and an unlabeled sample set, sample objects in the first positive sample set are provided with positive labels, the positive labels are used for marking the sample objects and have willingness to execute conversion operation aiming at target products, and the positive labels can be obtained by manual pre-marking. However, since labeling requires a lot of manpower and material resources, there are many unlabeled sample objects, that is, sample objects in an unlabeled sample set do not have a classification tag for marking whether the sample object has a desire to perform a conversion operation with respect to a target product.

In order to obtain more sample objects with classification labels to train the neural network model and ensure balance of positive and negative samples as much as possible, the server 101 may select a plurality of sample objects from the first positive sample set as intermediate samples, remove the intermediate samples from the first positive sample set to obtain a second positive sample set, and add the intermediate samples to the unlabeled sample set to obtain an original sample set to be labeled. And training a neural network model on the second positive sample set and the original sample set to be annotated according to the object characteristics of the sample objects in the constructed sample object data set to obtain an object classification model.

In the process of training the neural network model to obtain an object classification model, classifying sample objects in an original sample set to be marked by using the neural network model obtained in the training process until each sample object in the original sample set to be marked has a classification label, so as to obtain a target positive example set, a target negative example set and an object classification model. Because the intermediate sample is a positive sample selected from the first positive sample set and the behavior of the unknown positive sample in the original sample set to be marked is consistent, the unknown positive sample can be reliably evaluated by using the intermediate sample, and the initial positive and negative samples are effectively balanced, so that the initial positive and negative samples have better generalization in the classification process, and the sample objects in the original sample set to be marked are better classified. And because the object classification model is continuously updated to divide the sample objects in the original sample set to be marked, the sample objects in the original sample set to be marked without the marked sample can be purified each time under the condition that the object classification model is in the best state, the classification accuracy of the sample objects in the original sample set to be marked is ensured, and the service effect of the follow-up information recommendation is further improved.

After completing the training of the object classification model, the server 101 may classify the object to be identified using the object classification model, thereby determining whether to recommend the relevant information of the target product to the object to be identified according to the classification result. If the object to be identified has a desire to perform the conversion operation on the target product, the relevant information of the target product can be recommended to the object to be identified, for example, the relevant information of the target product is sent to the terminal 102 corresponding to the object to be identified for display, so that the conversion rate of the relevant information is improved; if the object to be identified does not have willingness to execute the conversion operation on the target product, the related information of the target product is not required to be recommended to the object to be identified, so that invalid recommendation is reduced, and the service effect is improved.

It should be noted that, the acquisition and processing of the data related to the user (such as the first positive sample set, the unlabeled sample set, the object feature of the sample object in the constructed sample object data set, and the related data required for constructing the object feature) according to the embodiment of the present application is allowed by the authorization of the user.

Next, a training method of the object classification model, which is provided by the embodiment of the present application, will be described in detail with reference to the accompanying drawings, taking a training method of the object classification model performed by a server as an example. Referring to fig. 2, fig. 2 shows a flowchart of a training method of an object classification model, the method comprising:

S201, a sample object data set aiming at a target product is obtained, the sample object data set comprises a first positive sample set and an unlabeled sample set, the sample objects in the first positive sample set are provided with positive labels, the positive labels are classification labels, the positive labels are used for marking the sample objects with willingness to execute conversion operation aiming at the target product, the sample objects in the unlabeled sample set are not provided with classification labels, and the classification labels are used for marking whether the sample objects have willingness to execute conversion operation aiming at the target product.

The first positive sample set may be represented by P, and the unlabeled sample set may be represented by U. The class label may be represented by a number, a symbol, or the like, taking the number as an example, for example, when the class label is 1, the class label may be a positive example label, and when the class label is 0, the class label may be a negative example label.

The technical scheme architecture of the training method for the object classification model provided by the embodiment of the application can be shown in fig. 3, and mainly comprises sample creation (for example, 301 shown in fig. 3), feature construction (for example, 302 shown in fig. 3), and a training framework (for example, an ISPYST algorithm framework shown in 303 in fig. 3) combining a hiding technology and self-learning training, wherein the sample creation is first described.

When the sample is created, the sample object which is related to the service strongly, has normal data distribution and reasonable object portrait can be found out according to the manual annotation and the service experience. Where the business strong correlation may be based on an empirical determination that the business strong correlation sample object (e.g., user), e.g., the target product is a video conference payment, it may be determined that the business strong correlation sample object (e.g., user) may be a office worker and that the work unit is often a large business on the market. The data distribution is normal and can be that various data of sample objects are compared, and the distribution is reasonable. For example, the target product is a video conference pay because pay users tend to be users of a large business on the market, and the duration of use of the game by such users tends to be relatively short. Based on the above, the positive sample objects with longer playing time length are removed according to the collected playing time length distribution of the positive sample objects. Reasonable object representation may mean that the basic characteristics of the sample object are reasonable, e.g. the target product is a video conference payment, because the paying user is often a user of a large business on the market, and then this a priori knowledge is to be met on the object representation, e.g. 24-45 years old, etc.

Normally, the ratio of positive and negative sample objects is 1:10000, and the number of positive sample objects (sample objects with positive labels) is lower than 5 ten thousand, which does not meet the minimum requirement of object classification model training. For this reason, in the training of the object classification model in the embodiment of the present application, positive sample objects and unlabeled sample objects (i.e., sample objects without classification labels) are mainly used, so that the unlabeled sample objects are classified in combination with the training framework of the hiding technology and the self-learning training to obtain sufficient and reliable positive sample objects and negative sample objects (sample objects with negative labels, which are used to indicate that the sample objects do not have willingness to perform a conversion operation for a target product). For this purpose, the embodiment of the application obtains a sample object data set aiming at a target product, wherein the sample object data set comprises a first positive sample set and an unlabeled sample set.

The target product may be various products provided by an operation to perform a conversion operation, for example, a video conference payment, a member of video software, a video, a commodity, an advertisement, and the like. The conversion operation may be an operation that brings benefit or advantageous effect to the operation, and in different scenes, whether there is a desire to perform the conversion operation for the target product may be different. For example, in a video conference scenario, the target product may be a video conference fee, and whether there is a desire to perform a conversion operation for the target product may be whether the user has a desire to pay for the video conference, i.e., the conversion operation is a fee; in another example, in a second-hand cart purchase scenario, the target product may be a second-hand cart, and whether the user has a desire to perform a conversion operation for the target product may be whether the user has a desire to purchase the second-hand cart, i.e., the conversion operation is a purchase; for another example, in a video viewing scenario, the target product may be video, whether there is a desire to perform a conversion operation for the target product may be whether the user has a desire to view the video, i.e., the conversion operation is viewing, and so on.

In one possible implementation, the method of obtaining the sample object data set for the target product may be to obtain an initial seed object for the target product, filter abnormal seed objects from the initial seed object to obtain remaining seed objects, and construct the sample object data set according to the remaining seed objects.

It will be appreciated that in the embodiment of the present application, the filtering of the abnormal seed objects may be performed in a plurality of ways, one way being that after a batch of initial seed objects are recalled coarsely based on rules, then filtering is performed based on manual screening, and finally verification is performed based on business logic.

Alternatively, a base representation of the initial seed object is obtained. The basic portrait includes some non-privacy and interaction data of the sample object for the target product, such as whether to install the mobile phone manager same as the manufacturer to which the target product belongs, whether to use the mobile phone manager disturbance interception function same as the manufacturer to which the target product belongs, and the like.

Yet another possible implementation is to calculate an abnormal object detection index, and further filter the abnormal seed object based on the abnormal object detection index. In a real business scene, false objects (users) and situations of controlling a terminal by a computer exist. In order to eliminate the influence of the non-real user on the training object classification model, abnormal object detection indexes are set based on service experience, such as flow use conditions of other products of the same manufacturer as the target product, time distribution generated by the flow and the like, so that abnormal seed objects are filtered based on the abnormal object detection indexes.

In yet another possible implementation, the outlier seed object is filtered based on a distribution outlier theorem. Outlier determination criteria are made using the "Laida criterion". The specific method comprises the following steps: assuming that a group of detection data contains only random errors, a standard deviation is obtained by calculating the detection data, a section is determined according to a certain probability, and the error exceeding the section is considered to be not random errors but coarse errors, and the data (such as an initial seed object) containing the error should be removed.

The above four ways of filtering abnormal seed objects may be used alone or in combination, which is not limited in the embodiments of the present application.

After obtaining the sample object data set, the sample object data set may be stored offline, e.g. in a distributed file system (The Hadoop Distributed File System, HDFS).

S202, object features of the sample objects in the sample object data set are constructed.

The server then builds object features for the sample objects in the sample object dataset. In one possible implementation, to construct high-dimensional feature vectors, the accuracy of the training object classification model is improved. When the server builds the object features, the basic features of the sample objects in the sample object data set can be built according to the object interaction behaviors of the sample objects in the sample object data set, the vertical type features are built according to the business features aiming at the target products, and then the feature combination is carried out according to the basic features and the vertical type features to obtain the object features.

In constructing the basic feature, the server may construct a rich object representation as the basic feature based on the object interaction behavior, and may include, for example: object base attributes, device base attributes, network connection attributes, etc. Such as: object base attribute (age, sex), device base attribute (end brand), network connection attribute (number of network connections per week, e.g. 10).

In constructing the vertical-type feature, the vertical-type feature may include, for example, a click rate, conversion rate, etc. of the sample object on the particular type of advertisement.

When the server performs feature merging according to the basic features and the vertical type features to obtain the object features, the basic features and the vertical type features of different time spans can be aggregated by combining the time dimension, then the aggregated basic features and the aggregated vertical type features are subjected to feature processing, and feature merging is performed based on the feature-processed basic features and the feature-processed vertical type features to obtain the object features (for example, as shown in 302 in fig. 3).

And combining the time dimension, the mode of aggregating the basic features and the vertical type features of different time spans can select three modes of summation, median and standard deviation. For example, the vertical type features corresponding to the sample objects in a period of time are aggregated, for example, the period of time may be about half a year/about 3 months/about 1 month/about 1 week, and the mode of selecting and summing to aggregate may be shown in fig. 4, where (a) the identified graph is a vertical type feature before aggregation, (b) the identified graph is a vertical type feature after aggregation, and (a) the identified graph and (b) in the identified graph, the vertical type features corresponding to the vertical type features before and after aggregation are represented in the same filling mode. For example, four squares (1, 2, 3, 0) with white filling are polymerized to obtain a vertical type feature value of 1.5; four square grids (0, 3, 7 and 2) in a diagonal filling mode from bottom left to top right are adopted, and the value of the vertical type characteristic obtained after polymerization is 3; four squares (0, 3, 0, 4) in a diagonal filling mode from top left to bottom right are adopted, and the value of the vertical type characteristic obtained after polymerization is 1.75; four squares (5, 1, 0, 4) with curve filling gave a vertical type feature value of 2.5 after polymerization.

The feature processing method for the aggregated basic feature and the vertical type feature may include normalization processing to obtain a normalized numerical feature, discretization processing to obtain a discretized non-numerical feature, and the like. The embodiment of the application mainly introduces discretization processing. The manner in which the discretization is performed may vary for different features. The discretization process mainly comprises the following modes:

One-Hot Encoding (One-Hot Encoding): also known as one-bit valid encoding, is to encode N states using N-bit state registers, each state having its own register bit, and only one of the bits is valid at any time. I.e. only one bit is a 1 and the rest are zero values. One-Hot Encoding may be directed primarily to gender features, region features, etc. in the underlying features, e.g., for sample objects (e.g., users), one-Hot Encoding becomes: male: (1, 0), female: (0, 1);

frequency Encoding (Count Encoding), encoding according to the number and frequency of category occurrences. For example, for a WiFi point of interest (Point of Interest, POI) feature of a sample object (user), count Encoding would be used to identify the sample object and the interest level of this POI. For example, the user has consumed 3 times the POI of 'food-Chinese dish-Yue dish'.

The merging code (Consolidation Encoding), consolidation Encoding means that a plurality of values exist under the variable corresponding to certain characteristics, and the values can be summarized into the same information. The system version feature of the terminal of the× operating system, for example, includes three values of "4.2", "4.4", and "5.0", which can be generalized to "low version× system" based on experience.

And combining the processed basic features and the vertical type features to obtain object features, and storing the object features in an HDFS offline, so that the quick access of the subsequent flow is facilitated. For each sample object in the sample object data set, the object feature input to the neural network model is an N1 numerical vector, such as (1,0,31,4,0.2,9.3,8.8, …,0,0,1,2,34).

By the method, the high-dimensional feature vector can be produced, so that object features comprising richer information are obtained, and the accuracy of the training object classification model is improved.

S203, selecting a plurality of sample objects from the first positive sample set as intermediate samples, removing the intermediate samples from the first positive sample set to obtain a second positive sample set, and adding the intermediate samples to the unlabeled sample set to obtain an original sample set to be labeled.

In the embodiment of the application, it is relatively easy to collect a large number of sample objects without classification labels, but it is much more difficult to obtain a large number of sample objects with classification labels, and the self-training algorithm is an algorithm which is studied more in semi-supervised learning, and can train a classifier (also called an object classification model and can be represented by C) by using a small number of sample objects with classification labels, then predict the sample objects without classification labels by using the object classification model, and add a high confidence result to the collection of sample objects with classification labels, and iterate until all sample objects without classification labels are divided. This process is actually that the classifier uses its own prediction to boost itself.

In the embodiment of the application, in order to solve the problems of abnormal proportion of positive and negative examples and serious low number of positive examples, a training framework combining a hiding technology and self-learning training can be adopted to train the neural network model. The self-learning training may be a positive example no-label (Positive and Unlabeled, PU) learning, where PU learning refers to a process in which a training set trains a classifier with only a positive example sample set and an unlabeled sample set. The training framework combining the hiding technology and the self-learning training can be, for example, a PU learning method (PU learning method combining spy technology and semi-enhanced self-training, SPYST) combining the hiding technology with the semi-supervised self-training, or a PU learning method (An improved spy technique combined with semi-enhanced self-training for PU learning algorithm, ISPYST) combining the hiding technology with the semi-supervised self-training after improving the hiding technology. The classifier (neural network model) may be, for example, a convolutional neural network (Convolutional Neural Network, CNN), a cyclic neural network (Recurrent Neural Network, RNN), etc., and the present application will be mainly described by taking the example that the neural network model is CNN.

When training a neural network model in combination with a training framework of concealment techniques and self-learning training to obtain an object classification model, first the server selects a plurality of sample objects from a positive sample set as intermediate samples. When the intermediate samples are selected, the spatial distribution information of the first positive sample set can be mined, the clustering center of the first positive sample set is calculated after the spatial distribution information of the first positive sample set is grasped, and the intermediate samples are selected according to the distance between the clustering center and the clustering center. In general, a sample object closer to the cluster center is selected as an intermediate sample (may also be referred to as a hidden sample, and may be denoted by S) according to the distance. Specifically, a manner in which a plurality of sample objects may be selected from the positive sample set as intermediate samples may be to determine a centroid vector (i.e., a cluster center) of the first positive sample set, calculate a distance between each sample object in the first positive sample set and the centroid vector, and then determine the intermediate samples according to the distances. The distance between the middle sample and the centroid vector is generally smaller, i.e., the middle sample is a sample object closer to the centroid vector.

In one possible implementation, the method of determining the centroid vector of the first positive example set of samples may comprise the steps of:

1. Randomly selecting one sample object from the first positive sample set P as an initial centroid vector u;

2. randomly selecting t sample objects from P as intermediate samples;

3. initializing a Cluster Cluster as an empty set;

4. for j=1, …, n, the distance Dj between the sample object Pj and the centroid vector u is calculated as follows:

wherein Pj represents the j-th sample object in the first positive sample set P, u represents the centroid vector, and Dj represents the distance between Pj and the centroid vector u;

5. updating cluster=cluster & { Pj };

6. the new centroid vector u is recalculated for all sample objects in the Cluster as follows:

wherein P epsilon Cluster represents the number of sample objects in the Cluster Cluster, and Cluster serving as denominator represents the number of sample objects in the Cluster;

7. repeating the steps 2-5 until the centroid vector u is not changed any more, and outputting the centroid vector u.

In calculating the distance between each sample object in the first positive sample set and the centroid vector, the calculation can be performed by the following formula:

where d represents the distance between each sample object in the first positive sample set and the centroid vector, pj represents the j-th sample object in the first positive sample set P, u represents the finally determined centroid vector, and n represents the number of sample objects in the first positive sample set P.

It should be noted that, in the embodiment of the present application, the purpose of determining the intermediate sample according to the distance is to determine the sample object that is closer to the centroid vector as the intermediate object. In this case, the manner of determining the intermediate sample according to the distance may include a plurality of manners, and one manner may be to select a sample object having a distance smaller than a preset threshold value as the intermediate object according to the distance corresponding to each sample object.

Alternatively, the distances corresponding to each sample object may be arranged in order from small to large, and then the sample objects arranged in the first t samples are selected as the middle samples.

In another way, the distances corresponding to each sample object are arranged in order from large to small, and then the sample objects arranged in the next t samples are selected as the middle samples.

By the method, the spatial distribution information of the first positive sample set can be mined, the spatial distribution information can show the spatial structure of the first positive sample set, the centroid vector of the first positive sample set is calculated after the spatial structure of the first positive sample set is grasped, and a sample object which is close to the centroid vector is found out and used as an intermediate sample. Therefore, the intermediate samples are closer to the centroid vector in the space structure, and the real information quantity of the included positive sample is larger, so that the distribution condition of the unknown positive in the unlabeled sample set can be more effectively reflected.

After the intermediate sample is selected, the intermediate sample can be removed from the first positive sample set to obtain a second positive sample set, and the intermediate sample is added to the unlabeled sample set to obtain an original sample set to be labeled, i.e. the second positive sample set can be represented as P-S, and the original sample set to be labeled can be represented by U+S. Referring to fig. 5, fig. 5 shows a block diagram of an ISPYST algorithm framework, a second positive sample set P, a non-labeled sample set U, a second positive sample set P-S, and original to-be-labeled sample sets u+s may be respectively shown in fig. 5.

And S204, training a neural network model on the second positive sample set and the original sample set to be annotated according to the object characteristics to obtain an object classification model.

S205, in the process of training the neural network model to obtain an object classification model, classifying sample objects in an original sample set to be marked by using the neural network model obtained in the training process until each sample object in the original sample set to be marked has a classification label, and obtaining a target positive example set, a target negative example set and an object classification model.

After the second positive sample set and the original sample set to be marked are obtained, the server can train the neural network model on the second positive sample set and the original sample set to be marked according to the object characteristics to obtain an object classification model.

In the process of training the neural network model to obtain an object classification model, classifying sample objects in an original sample set to be marked by using the neural network model obtained in the training process until each sample object in the original sample set to be marked has a classification label, so as to obtain a target positive example set, a target negative example set and an object classification model. Wherein, the target positive example set can be represented by RP, and the target negative example set can be represented by RN.

Because the intermediate sample is a positive sample selected from the first positive sample set and the behavior of the unknown positive sample in the original sample set to be marked is consistent, the unknown positive sample can be reliably evaluated by using the intermediate sample, and the positive and negative samples are effectively balanced, so that the positive and negative samples have better generalization in the classification process, and the sample objects in the original sample set to be marked are better classified. And the object classification model is continuously updated to divide the sample objects in the original sample set to be marked, so that the sample objects in the original sample set to be marked without the marked sample can be purified each time under the condition that the object classification model is in the best state, and the classification accuracy of the sample objects in the original sample set to be marked is ensured.

In one possible implementation manner, in the process of training the neural network model to obtain the object classification model, classifying the sample objects in the original sample set to be marked by using the neural network model obtained by training until each sample object in the original sample set to be marked has a classification label, and obtaining the target positive example set, the target negative example set and the object classification model may be that the sample objects in the original sample set to be marked are classified by using the neural network model obtained by training to obtain an initial positive example set and an initial negative example set, and the initial positive example set is used as a new sample set to be marked; retraining the neural network model on the second positive example set and the initial negative example set; classifying the new sample set to be marked by utilizing the neural network model obtained by training, updating the second positive sample set, the new sample set to be marked and the initial negative sample set based on the classification result, and re-executing the step of re-training the neural network model on the second positive sample set and the initial negative sample set until each sample object in the new sample set to be marked has a classification label, thereby obtaining a target positive sample set, a target negative sample set and an object classification model.

Wherein, the initial positive example set can be represented by RN, and the initial negative example set can be represented by RP. And (3) re-giving positive example labels to the second positive example sample set, giving negative example labels to sample objects in the RN, hiding the RP, and regarding the RP as a new sample set to be marked, wherein the new sample set to be marked can be represented by U'. Then, classifying the U 'by using the trained neural network model, updating the second positive example sample set P, the new sample set to be marked U' and the initial negative example set RN based on the classification result, and re-executing the step of re-training the neural network model on the second positive example sample set and the initial negative example set until each sample object in the new sample set to be marked has a classification label, thereby obtaining a target positive example set, a target negative example set and an object classification model (see the diagram shown in fig. 5).

In one possible implementation manner, the training is utilized to obtain a neural network model to classify sample objects in an original sample set to be marked, so that an initial positive example set and an initial negative example set can be obtained; calculating a second posterior probability of the intermediate sample belonging to the positive example sample, wherein the second posterior probability can be expressed by Prs; then determining a probability threshold based on the second posterior probability Prs, the probability threshold being represented by theta; then, an initial negative example set is determined according to the sample object with the first posterior probability smaller than the probability threshold, and an initial positive example set is determined according to the original sample set to be annotated and the initial negative example set. For the original sample set to be marked u= { U1, U2, U3 … … }, ui represents the ith sample object in the original sample set to be marked, if the first posterior probability corresponding to ui is smaller than the probability threshold, adding the first posterior probability to the initial negative example set, wherein the initial positive example set is a set obtained by removing sample objects belonging to the initial negative example set from the original sample set to be marked, namely rp=u-RN.

In one possible implementation manner, the training is utilized to obtain a neural network model to classify the new sample set to be marked, and the second positive sample set, the new sample set to be marked and the initial negative sample set are updated based on the classification result, wherein the training is utilized to obtain the neural network model to classify the new sample set to be marked, so that the posterior probability maximum value of the category of the sample object in the new sample set to be marked is obtained, and the posterior probability maximum value is used as the classification result; arranging the maximum posterior probability of the category of the sample object in the new sample set to be marked according to the sequence from the big to the small; selecting a sample object with the maximum posterior probability ordered at the front f bits; and updating the second positive example sample set, the new sample set to be marked and the initial negative example set according to the sample object of the front f bits and the corresponding classification label.

The maximum posterior probability value can be expressed as Pr, the maximum posterior probability values of the categories to which the sample objects in the new sample set to be marked belong are arranged in order from large to small, the larger the maximum posterior probability value Pr is, the higher the confidence level is, so that the sample objects with the maximum posterior probability value ordered in the front f bits are selected, and the sample objects with the front f bits and the corresponding classification labels are updated into a second positive sample set, a new sample set to be marked and an initial negative sample set. In general, f selected sample objects need to be subtracted from the new sample set to be annotated U' to update the new sample set to be annotated.

By selecting the sample object of the first f bits, the sample object of the first f bits is a high confidence sample and carries more effective distribution information, thereby being beneficial to retraining the neural network model. And iterating the process circularly until the sample objects in the new sample set to be marked are completely divided. The aim of refining and purifying the rough classification result of the hiding technology is achieved, and therefore the PU learning frame with higher classification efficiency is obtained.

The specific implementation method of the SPYST algorithm framework is introduced, and the core function of the SPYST algorithm framework is to perform self-training purification on an unlabeled sample set, extract more reliable high-quality positive samples, reduce the inclination ratio of the positive samples and the negative samples, and improve the absolute magnitude of the positive samples.

After the training of the object classification model is completed, the server can classify the object to be identified by using the object classification model, so as to determine whether to recommend relevant information of a target product to the object to be identified according to the classification result. Specifically, the server may obtain the object to be identified, construct a target object feature of the object to be identified, and further determine a classification label of the object to be identified according to the target object feature through the object classification model, where the classification label may be used as a classification result, and the classification label is used to indicate whether the object to be identified has a desire to perform a conversion operation with respect to the target product.

If the classification label indicates that the object to be identified has a willingness to execute the conversion operation on the target product, the relevant information of the target product can be recommended to the terminal corresponding to the object to be identified, for example, the relevant information of the target product is sent to the terminal corresponding to the object to be identified for display, so that the conversion rate of the relevant information is improved; if the classification label indicates that the object to be identified does not have willingness to execute the conversion operation on the target product, the related information of the target product is not required to be recommended to the object to be identified, so that invalid recommendation is reduced, and the service effect is improved.

In the embodiment of the application, the training frame combining the hiding technology and the self-learning training is adopted to train the neural network model so as to improve the accuracy and the recognition efficiency of the object classification model obtained by training, and the parameter optimizing process of the neural network model can be improved in the training process, so that the training efficiency of the object classification model is improved, and the accuracy and the recognition efficiency of the object classification model obtained by training are further improved.

In this case, according to the object characteristics, the method of training the neural network model on the second positive sample set and the original sample set to be annotated may be to update the parameters of the neural network model on the second positive sample set and the original sample set to be annotated according to the object characteristics, so as to obtain the optimal network parameters; and determining an object classification model according to the optimal network parameters.

In one possible implementation, a particle swarm optimization (Particle Swarm Optimization, PSO) algorithm may be employed for parameter optimization. On the basis, a PSO-based improved algorithm can be adopted to perform parameter optimization, and the improved Particle swarm optimization algorithm can comprise a Quantum-weighted Particle swarm optimization algorithm (Quantum-Bahaved Particle Swarm Optimizationwith Binary Encoding, BQPSO) and the like. The embodiment of the application mainly describes parameter optimization based on an improved particle swarm optimization algorithm, and the parameter optimization based on the improved particle swarm optimization algorithm is shown as 304 in fig. 3. If the neural network model comprises a plurality of network layers, according to the object characteristics, updating parameters of the neural network model on a second positive sample set and an original sample set to be marked, wherein the optimal network parameters can be obtained by respectively encoding the parameters of the plurality of network layers to obtain multidimensional particle vectors; utilizing an improved quantum particle swarm optimization algorithm to determine the global optimal position of a population corresponding to the multidimensional particle vector; in the process of determining the global optimal position of the population corresponding to the multidimensional particle vector by utilizing the improved quantum particle swarm optimization algorithm, if the termination condition of the improved quantum particle swarm optimization algorithm is reached, decoding and outputting the global optimal position of the population when the termination condition is reached, and obtaining the optimal network parameters.

Next, the embodiment of the present application will mainly take the improved quantum particle swarm optimization algorithm, which is QPSO and BQPSO, as an example, and describe the determination manner of the above-mentioned optimal network parameters in detail. First, determining optimal network parameters based on QPSO is introduced.

The flowchart of the parameter optimization of the neural network model based on QPSO can be seen in fig. 6:

s601, initializing quantum particle position information in QPSO, and selecting a global optimal position.

S602, each quantum particle is decoded into a neural network model structure according to a segmented compression coding strategy.

S603, determining whether the neural network model structure is effective, if yes, executing S604, and if no, executing S607.

S604, calculating a fitness value.

S605, the individual best position Pi and the global best position Pg will be updated.

S606, determining whether a maximum evolution period is reached, if so, ending the flow and outputting Pg; if not, return to S602.

S607, quantum measurement is carried out.

When particle decoding produces an ineffective neural network structure, the quantum particles can be repeatedly measured to obtain more new positional information, which in turn decodes to produce more new CNN structures. The method of quantum measurement may be seen in fig. 7, which includes:

S701, calculating the characteristic length of the current particle and the attractor.

The calculation formulas of the characteristic length and the attractor can be as follows:

wherein mbest is the average optimal position of the population, M is the population scale, pid is the ith particle in the D dimension of the population, D takes a value between 1 and D, D is the dimension of the particle vector,and u is [0,1]]Random decimal in interval, beta is a normal number called coefficient of contraction expansion, a _id Representing local attractors, pid and Pgd represent the position information of the local and global optimal solutions of the ith particle in the d-th dimension, respectively.

S702, randomly generating a decimal.

Typically, a fraction can be randomly generated in the interval [0,1 ].

S703, generating a new individual.

In an embodiment of the present application, a new individual may be generated according to the following formula:

wherein x is _id (t+1) represents the new individual generated for the t+1st iteration, x _id (t) represents the current particle, i.e., the individual of the t-th iteration, u represents the randomly generated fraction, mbest _d Representing the population average optimal position in the d-th dimension.

S704, decoding the position information of the bits into a neural network structure according to the segmented compression coding strategy.

S705, determining whether the neural network model structure is effective, if so, executing S706, and if not, returning to S702.

S706, taking the individual as a next generation individual.

If the maximum number of iterations is reached, the process ends.

Next, determining optimal network parameters based on BQPSO is described, comprising the steps of:

a) Initializing each particle xi in the population in the form of a binary bit string such that pi=xi;

where xi= { xi1, xi2, … …, xiD }, xid represents the positional information of the ith particle in the d-th dimension.

b) The value of the group average optimal position mbest is calculated, and the calculation flow is as follows:

will P ₁ 、P ₂ 、……、P _M As input, mbest is calculated as output by the following code:

setting the initial value of sum to 0;

sum＝sum+P _i [j]

avg＝sum/M；

when avg >0.5, mbest [ j ] =1; when avg <0.5, mbest [ j ] =0; when avg=0.5, mbest [ j ] =0 or 1;

where j is a value between 1 and l, l is the length of each particle, pi is the ith particle in the population, i is a value between 1 and M, and M is the population size.

c) And calculating the fitness value of each particle in the population according to the fitness function, comparing the fitness value with the optimal value of the previous particle, and if f (xi) > f (Pi), if pi=xi, otherwise, not updating. The fitness function may include a plurality of fitness values, and in the embodiment of the present application, two calculation modes of fitness values of the Schaffer function and the RA-rastigin function may be selected.

d) Updating global optimal particles pg in the population;

e) Calculating the local attractor a according to an algorithm _id Is a value of (2);

a _id is calculated from Pid and Pgd, pid and Pgd represent the position information of the local and global optimal solutions of the ith particle in the d-th dimension, respectively.

f) Calculating p according to the formula _rd The value of p _rd The probability in the d-th dimension is expressed as follows:

wherein x is _id Representing position information of the ith particle in the d-th dimension, mbest _d Represents the average optimal position of the population in the d-th dimension, u represents a randomly generated decimal between 0 and 1, beta is a normal number called the coefficient of contraction expansion, and ld represents x _id Length d of (d) _H () Representing a computational function.

g) Calculating x according to an algorithm _id And concatenating the values of (2) to generate xi;

x _id is according to a _id And p _rd Calculated.

h) Repeating the steps until the termination condition of the improved quantum particle swarm optimization algorithm is met.

In the above-mentioned process of parameter optimization based on the improved particle swarm optimization algorithm, parameters of the neural network model need to be encoded first, taking the neural network model as CNN as an example, the network structure of the CNN may be shown in fig. 8, and the network structure of the CNN includes a plurality of network layers, which are respectively an Input layer (Input), two Convolution layers (Convolution), two pooling layers (Pool), and two full connection layers (including a hidden layer and an output layer). The configuration of each network layer varies: step size, kernel size and feature map are important parameters of the convolutional layer configuration; step size, core size, and pool type are the primary attributes of the pooling layer; the number of neurons is the only attribute of the fully connected layer. The convolution layer, pooling layer and full-connection layer are briefly described below.

CNN consists of one Input layer (Input), two Convolution layers (Convolition), two pooling layers (Pool), and two fully connected layers. The number of neurons of the output layer is equal to the number of classes in the classification problem.

A schematic diagram of the convolution operation of the convolution layer is shown in fig. 9: the convolution operation was performed starting from the upper left-hand shadow position of the input image matrix using a convolution kernel of size 3*3, and the result of this operation was 3. Since the step length is set to 1, the shadow position can also be moved leftwards or rightwards, and a 3*3 Feature Map (Feature Map) composed of 9 results is finally obtained

The schematic diagram of the pooling operation of the pooling layer is shown in fig. 10: to some extent, the more convolutional layers, the more comprehensive the extracted features. However, too many convolution kernel operations may result in too large an extracted feature dimension, and thus require too large computational resources, and thus require pooling operations to reduce the feature dimension. The pooling operation has two parameters of pooling core and step size, and common pooling operation has mean pooling and maximum pooling. The former takes the mean value of the feature matrix corresponding to the area of the kernel size, and the latter takes the maximum value of the feature matrix corresponding to the area.

The full link layer is shown in fig. 11: the first layer is responsible for inputting data (e.g., x1, x2, x 3), the last layer is responsible for outputting results (e.g., y), and the other layers are called hidden layers. Each neuron of the hidden layer is connected to a neuron of an upper layer, and is thus called a fully connected layer.

Taking the improved particle swarm optimization algorithm as a BQPSO as an example, the method for coding the parameters of the CNN is as follows:

binary encoding strategy:

the encoding process is implemented in the D-dimensional discrete search space of the BQPSO algorithm, where the D decision variables in each particle are fixed length binary sequences, as shown in table 1:

TABLE 1

ID

Parameter 1

Parameter 2

……

Each binary sequence carries the configuration of one network layer, respectively, including its ID and parameters.

Parameters of each network layer type of the CNN are binary coded as follows:

the convolutional layer coding strategy is shown in table 2:

TABLE 2

	ID	Nuclear size	Feature map	Step size	Totals to
						Value range	0 or 7	[1,8]	[1,128]	[1,4]
Binary digits	3	3	7	2	15
						Examples of the invention	0(000)	2(001)	16(0001111)	2(01)	000001000111101

The pooling layer coding strategy is shown in table 3:

TABLE 3 Table 3

	ID	Nuclear size	Step size	Type(s)	Totals to
						Value range	1 or 6	[1,4]	[1,4]	[1,2]
Binary digits	3	2	2	1	8
						Examples of the invention	0(001)	2(01)	2(01)	2(1)	00101011XXXXXXX

The full-link layer coding strategy is shown in table 4:

TABLE 4 Table 4

	ID	Neuron number	Totals to
				Value range	2 or 5	[1,2048]
Binary digits	3	11	14
				Examples of the invention	2(010)	1024(01111111111)	010011111111111X

The blank layer coding strategy is shown in table 5:

TABLE 5

	ID	Totals to
			Value range	3 or 4
Binary digits	3	3
			Examples of the invention	3(011)	011XXXXXXXXXXXX

Based on the coding strategies of the different network layers and the network structure of the CNN, the parameter coding of the CNN can be completed.

And carrying out parameter optimization on the neural network model based on a binary quantum particle swarm optimization algorithm so as to obtain a better model effect. And the prediction accuracy can be improved by predicting all samples through the neural network model with optimized parameters.

It should be noted that, a detailed flowchart of a method for training an object classification model based on an ISPYST algorithm and a BQPSO according to an embodiment of the present application may be shown in fig. 12, which includes:

s1201, acquiring an initial seed object aiming at a target product.

S1202, filtering abnormal seed objects from the initial seed objects to obtain residual seed objects.

S1203, constructing a sample object data set according to the remaining seed objects.

S1204, constructing basic characteristics.

S1205, constructing vertical type features.

And S1206, carrying out feature combination according to the basic features and the vertical type features to obtain object features.

S1207, a plurality of sample objects are selected from the first positive sample set as intermediate samples.

S1208, removing the intermediate sample from the first positive sample set to obtain a second positive sample set, and adding the intermediate sample to the unlabeled sample set to obtain an original sample set to be labeled.

S1209, classifying sample objects in the original sample set to be marked by utilizing the neural network model obtained through training to obtain an initial positive example set and an initial negative example set.

S1210, taking the initial positive example set as a new sample set to be marked.

S1211, retraining the neural network model on the second positive example sample set and the initial negative example set.

S1212, classifying the new sample set to be marked by using the neural network model obtained through training, and updating the second positive example sample set, the new sample set to be marked and the initial negative example set based on the classification result.

S1213 initializing each particle xi in the population in the form of a binary bit string such that pi=xi.

S1214, calculating the value of the group average optimal position mbest.

S1215, calculating an fitness value of each particle in the population according to the fitness function, and comparing with the previous particle optimum value, if f (xi) > f (Pi), pi=xi.

S1216, updating global optimal particles pg in the population.

S1217, calculating the local attractor a according to the algorithm _id Is a value of (2).

S1218, calculate p _rd Is a value of (2).

S1219, calculating x according to the algorithm _id And concatenates the values of xi.

S1220, repeating the steps until the termination condition of the improved quantum particle swarm optimization algorithm is met.

The following is a comparison analysis of business effects of different methods, where the different methods include related art 1, related art 2 and a method provided by the embodiments of the present application, the related art 1 may be a method based on manual experience to determine a data rule, that is, product operation is based on business experience, and a rule of manual identification is set, for example, an object with a age of 35-40 years and a high living standard is considered to have a higher probability of being in a "high willingness object for video conference payment", the related art 2 may be a method based on non-deep learning, and a comparison schematic diagram of business effects of the different methods is shown in fig. 13:

As can be seen in FIG. 13, compared with other technologies, the method provided by the embodiment of the application improves the average of 182.7% from the aspect of advertisement click rate; compared with other technologies, the method provided by the embodiment of the application improves 178.41% on average, namely the business effect is obviously improved.

It should be noted that, based on the implementation manner provided in the above aspects, further combinations may be further performed to provide further implementation manners.

Based on the training method of the object classification model provided in the corresponding embodiment of fig. 2, the embodiment of the application further provides a training device 1400 of the object classification model. Referring to fig. 14, the training apparatus 1400 of the object classification model includes an acquisition unit 1401, a construction unit 1402, a determination unit 1403, and a training unit 1404:

the obtaining unit 1401 is configured to obtain a sample object data set for a target product, where the sample object data set includes a first positive sample set and an unlabeled sample set, the sample objects in the first positive sample set have positive labels, the positive labels are classification labels, the positive labels are used to mark sample objects with willingness to perform a conversion operation for the target product, the sample objects in the unlabeled sample set do not have classification labels, and the classification labels are used to mark whether the sample objects have willingness to perform a conversion operation for the target product;

The constructing unit 1402 is configured to construct object features of a sample object in the sample object dataset;

the determining unit 1403 is configured to select a plurality of sample objects from the first positive sample set as intermediate samples, remove the intermediate samples from the first positive sample set to obtain a second positive sample set, and add the intermediate samples to the unlabeled sample set to obtain an original sample set to be labeled;

the training unit 1404 is configured to train a neural network model on the second positive sample set and the original sample set to be annotated according to the object feature, so as to obtain the object classification model;

in the process of training the neural network model to obtain the object classification model, the training unit 1404 is configured to classify the sample objects in the original sample set to be labeled by using the neural network model obtained in the training process until each sample object in the original sample set to be labeled has a classification label, so as to obtain a target positive example set, a target negative example set and the object classification model.

In a possible implementation, the determining unit 1403 is specifically configured to;

determining a centroid vector of the first positive example sample set;

Calculating the distance between each sample object in the first positive sample set and the centroid vector;

and determining the intermediate sample according to the distance.

In a possible implementation manner, the determining unit 1403 is specifically configured to:

arranging the distances corresponding to each sample object in order from small to large;

sample objects arranged in the first t are selected as the intermediate samples.

In one possible implementation, the training unit 1404 is specifically configured to:

classifying sample objects in the original sample set to be marked by using the neural network model obtained through training to obtain an initial positive example set and an initial negative example set, and taking the initial positive example set as a new sample set to be marked;

retraining the neural network model on the second positive example set and the initial negative example set;

classifying the new sample set to be marked by using the trained neural network model, updating the second positive sample set, the new sample set to be marked and the initial negative sample set based on classification results, and re-executing the step of re-training the neural network model on the second positive sample set and the initial negative sample set until each sample object in the new sample set to be marked has a classification label, thereby obtaining the target positive sample set, the target negative sample set and the object classification model.

classifying the sample objects in the original sample set to be marked by using the neural network model obtained through training to obtain a first posterior probability that the sample objects in the original sample set to be marked belong to a positive sample;

calculating a second posterior probability that the intermediate sample belongs to the positive sample;

determining a probability threshold according to the second posterior probability;

and determining the initial negative example set according to the sample object of which the first posterior probability is smaller than the probability threshold, and determining the initial positive example set according to the original sample set to be annotated and the initial negative example sample set.

classifying the new sample set to be marked by using the neural network model obtained through training to obtain a posterior probability maximum value of the category of the sample object in the new sample set to be marked, wherein the posterior probability maximum value is used as the classification result;

arranging the maximum posterior probability of the category of the sample object in the new sample set to be marked according to the sequence from big to small;

Selecting a sample object with the maximum posterior probability ordered at the front f bits;

and updating the second positive example sample set, the new sample set to be marked and the initial negative example set according to the sample object of the front f bits and the corresponding classification label.

In one possible implementation, the training unit 1404 is specifically configured to;

according to the object characteristics, updating parameters of the neural network model on the second positive sample set and the original sample set to be annotated to obtain optimal network parameters;

and determining the object classification model according to the optimal network parameters.

In one possible implementation, the neural network model includes a plurality of network layers, and the training unit 1404 is specifically configured to:

encoding parameters of the network layers respectively to obtain multidimensional particle vectors;

utilizing an improved particle swarm optimization algorithm to determine the global optimal position of the population corresponding to the multidimensional particle vector;

in the process of determining the global optimal position of the population corresponding to the multidimensional particle vector by utilizing the improved particle swarm optimization algorithm, if the termination condition of the improved quantum particle swarm optimization algorithm is reached, decoding and outputting the global optimal position of the population when the termination condition is reached, so as to obtain the optimal network parameter.

In one possible implementation manner, the acquiring unit 1401 is specifically configured to:

acquiring an initial seed object aiming at the target product;

filtering abnormal seed objects from the initial seed objects to obtain residual seed objects;

and constructing the sample object data set according to the residual seed objects.

In a possible implementation manner, the construction unit 1402 is specifically configured to:

constructing basic characteristics of sample objects in the sample object data set according to the object interaction behaviors of the sample objects in the sample object data set;

constructing vertical type characteristics according to the service characteristics aiming at the target product;

and carrying out feature combination according to the basic feature and the vertical type feature to obtain the object feature.

In a possible implementation manner, the apparatus further includes a classification unit:

the acquiring unit 1401 is further configured to acquire an object to be identified;

the construction unit 1402 is further configured to construct a target object feature of the object to be identified;

the classification unit is used for determining a classification label of the object to be identified through the object classification model according to the target object characteristics, and the classification label is used for indicating whether the object to be identified has willingness to execute conversion operation aiming at a target product.

In a possible implementation manner, the apparatus further includes a recommending unit:

and if the classification label indicates that the object to be identified has a willingness to execute a conversion operation on a target product, the recommending unit is used for recommending the related information of the target product to the terminal corresponding to the object to be identified.

The embodiment of the application also provides computer equipment which can execute the graph data processing method. The computer device may be, for example, a terminal, taking a smart phone as an example:

fig. 15 is a block diagram showing a part of a structure of a smart phone according to an embodiment of the present application. Referring to fig. 15, the smart phone includes: radio Frequency (r.f. Frequency) circuitry 1510, memory 1520, input unit 1530, display unit 1540, sensor 1550, audio circuitry 1560, wireless fidelity (r.f. WiFi) module 1570, processor 1580, and power supply 1590. The input unit 1530 may include a touch panel 1531 and other input devices 1532, the display unit 1540 may include a display panel 1541, and the audio circuit 1560 may include a speaker 1561 and a microphone 1562. It will be appreciated that the smartphone structure shown in fig. 15 is not limiting of the smartphone, and may include more or fewer components than shown, or may combine certain components, or a different arrangement of components.

The memory 1520 may be used to store software programs and modules, and the processor 1580 performs various functional applications and data processing of the smart phone by executing the software programs and modules stored in the memory 1520. The memory 1520 may mainly include a storage program area and a storage data area, wherein the storage program area may store an operating system, an application program required for at least one function (such as a sound playing function, an image playing function, etc.), and the like; the storage data area may store data (such as audio data, phonebooks, etc.) created according to the use of the smart phone, etc. In addition, memory 1520 may include high speed random access memory, and may also include non-volatile memory, such as at least one magnetic disk storage device, flash memory device, or other volatile solid state storage device.

Processor 1580 is a control center of the smartphone, connects various parts of the entire smartphone with various interfaces and lines, performs various functions of the smartphone and processes data by running or executing software programs and/or modules stored in memory 1520, and invoking data stored in memory 1520. In the alternative, processor 1580 may include one or more processing units; preferably, the processor 1580 can integrate an application processor and a modem processor, wherein the application processor primarily processes operating systems, user interfaces, application programs, and the like, and the modem processor primarily processes wireless communications. It is to be appreciated that the modem processor described above may not be integrated into the processor 1580.

In this embodiment, the steps performed by the processor 1580 in the smart phone may be implemented based on the structure shown in fig. 15.

The computer device according to the embodiment of the present application may also be a server, as shown in fig. 16, fig. 16 is a block diagram of a server 1600 provided by the embodiment of the present application, where the server 1600 may have a relatively large difference due to different configurations or performances, and may include one or more central processing units (Central Processing Units, abbreviated as CPUs) 1622 (e.g. one or more processors) and a memory 1632, and one or more storage media 1630 (e.g. one or more mass storage devices) storing application programs 1642 or data 1644. Wherein memory 1632 and storage medium 1630 may be transitory or persistent. The program stored on the storage medium 1630 may include one or more modules (not shown), each of which may include a series of instruction operations on a server. Further, the central processor 1622 may be configured to communicate with a storage medium 1630 to execute a series of instruction operations on the storage medium 1630 on the server 1600.

The server 1600 may also include one or more power supplies 1626, one or more wired or wireless network interfaces 1650, one or more input output interfaces 1658, and/or one or more operating systems 1641, such as Windows Server, mac OS XTM, unixTM, linuxTM, freeBSDTM, and the like.

In this embodiment, the central processor 1622 in the server 1600 may perform the following steps:

constructing object features of a sample object in the sample object dataset;

According to an aspect of the present application, there is provided a computer-readable storage medium for storing program code for performing the training method of the object classification model according to the foregoing embodiments.

According to one aspect of the present application, there is provided a computer program product or computer program comprising computer instructions stored in a computer readable storage medium. The processor of the computer device reads the computer instructions from the computer-readable storage medium, and the processor executes the computer instructions, so that the computer device performs the methods provided in the various alternative implementations of the above embodiments.

The descriptions of the processes or structures corresponding to the drawings have emphasis, and the descriptions of other processes or structures may be referred to for the parts of a certain process or structure that are not described in detail.

The terms "first," "second," "third," "fourth," and the like in the description of the application and in the above figures, if any, are used for distinguishing between similar objects and not necessarily for describing a particular sequential or chronological order. It is to be understood that the data so used may be interchanged where appropriate such that the embodiments of the application described herein may be implemented, for example, in sequences other than those illustrated or otherwise described herein. Furthermore, the terms "comprises," "comprising," and "having," and any variations thereof, are intended to cover a non-exclusive inclusion, such that a process, method, system, article, or apparatus that comprises a list of steps or elements is not necessarily limited to those steps or elements expressly listed but may include other steps or elements not expressly listed or inherent to such process, method, article, or apparatus.

In the several embodiments provided in the present application, it should be understood that the disclosed systems, devices, and methods may be implemented in other manners. For example, the apparatus embodiments described above are merely illustrative, e.g., the division of the units is merely a logical function division, and there may be additional divisions when actually implemented, e.g., multiple units or components may be combined or integrated into another system, or some features may be omitted or not performed. Alternatively, the coupling or direct coupling or communication connection shown or discussed with each other may be an indirect coupling or communication connection via some interfaces, devices or units, which may be in electrical, mechanical or other form.

The units described as separate units may or may not be physically separate, and units shown as units may or may not be physical units, may be located in one place, or may be distributed on a plurality of network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

In addition, each functional unit in the embodiments of the present application may be integrated in one processing unit, or each unit may exist alone physically, or two or more units may be integrated in one unit. The integrated units may be implemented in hardware or in software functional units.

The integrated units, if implemented in the form of software functional units and sold or used as stand-alone products, may be stored in a computer readable storage medium. Based on such understanding, the technical solution of the present application may be embodied in essence or a part contributing to the prior art or all or part of the technical solution in the form of a software product stored in a storage medium, including several instructions for causing a computer device (which may be a computer, a server, or a network device, etc.) to perform all or part of the steps of the method according to the embodiments of the present application. And the aforementioned storage medium includes: a U-disk, a removable hard disk, a Read-Only Memory (ROM), a random access Memory (Random Access Memory, RAM), a magnetic disk, or an optical disk, or other various media capable of storing program codes.

The above embodiments are only for illustrating the technical solution of the present application, and not for limiting the same; although the application has been described in detail with reference to the foregoing embodiments, it should be understood by those of ordinary skill in the art that: the technical scheme described in the foregoing embodiments can be modified or some technical features thereof can be replaced by equivalents; such modifications and substitutions do not depart from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A method of training an object classification model, the method comprising:

Constructing object features of a sample object in the sample object dataset;

2. The method of claim 1, wherein the selecting a plurality of sample objects from the positive sample set as intermediate samples comprises;

determining a centroid vector of the first positive example sample set;

And determining the intermediate sample according to the distance.

3. The method of claim 2, wherein said determining said intermediate samples from said distances comprises:

4. The method according to claim 1, wherein in training the neural network model to obtain the object classification model, classifying the sample objects in the original sample set to be labeled by using the trained neural network model until each sample object in the original sample set to be labeled has a classification label, to obtain a target positive example set, a target negative example set and an object classification model, including:

5. The method of claim 4, wherein classifying the sample objects in the original set of samples to be labeled using the trained neural network model to obtain an initial positive example set and an initial negative example set, comprises:

6. The method of claim 4, wherein classifying the new sample set to be labeled using the trained neural network model and updating the second positive example sample set, the new sample set to be labeled, and the initial negative example set based on classification results comprises:

7. The method of claim 1, wherein the training a neural network model on the second positive sample set and the original sample set to be annotated according to the object features comprises;

8. The method according to claim 1, wherein the neural network model includes a plurality of network layers, and the updating parameters of the neural network model on the second positive sample set and the original sample set to be annotated according to the object features to obtain optimal network parameters includes:

9. The method of any one of claims 1-8, wherein the acquiring a sample object dataset for a target product comprises:

acquiring an initial seed object aiming at the target product;

10. The method of any of claims 1-8, wherein said constructing object features of a sample object in said sample object dataset comprises:

11. The method according to any one of claims 1-8, further comprising:

acquiring an object to be identified;

constructing target object characteristics of the object to be identified;

and determining a classification label of the object to be identified through the object classification model according to the target object characteristics, wherein the classification label is used for indicating whether the object to be identified has willingness to execute conversion operation aiming at a target product.

12. The method of claim 11, wherein the method further comprises:

if the classification label indicates that the object to be identified has a willingness to execute conversion operation aiming at a target product, recommending the relevant information of the target product to a terminal corresponding to the object to be identified.

13. The training method of the object classification model is characterized in that the device comprises an acquisition unit, a construction unit, a determination unit and a training unit:

14. A computer device, the computer device comprising a processor and a memory:

the processor is configured to perform the method of any of claims 1-12 according to instructions in the program code.

15. A computer readable storage medium for storing program code which, when executed by a processor, causes the processor to perform the method of any of claims 1-12.

16. A computer program product comprising a computer program, characterized in that the computer program, when executed by a processor, implements the method of any of claims 1-12.