CN111241850B - Method and device for providing business model - Google Patents
Method and device for providing business model Download PDFInfo
- Publication number
- CN111241850B CN111241850B CN202010329629.6A CN202010329629A CN111241850B CN 111241850 B CN111241850 B CN 111241850B CN 202010329629 A CN202010329629 A CN 202010329629A CN 111241850 B CN111241850 B CN 111241850B
- Authority
- CN
- China
- Prior art keywords
- model
- candidate
- semantic vector
- target
- models
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
Images
Abstract
The embodiment of the specification provides a method for providing a business model, and by the method and the device provided by the embodiment of the specification, a large number of candidate models are stored in advance at a server side, candidate semantic vectors obtained by coding training data of the candidate models are stored, when a client side needs the business model, a current semantic vector determined by own business data can be uploaded, and the current semantic vector and the candidate semantic vectors are matched to select a target model according to a matching result. This approach may better protect the data privacy of the client and generate less traffic. Particularly, the server side selects a plurality of target models, and after the plurality of target models are fused, the fusion result is provided for the client side, so that the model data privacy of the server side can be effectively protected. In summary, the method for providing the business model described in this specification can improve the effectiveness of the application of the business model.
Description
Technical Field
One or more embodiments of the present specification relate to the field of computer technologies, and in particular, to a method and an apparatus for providing a business model to a client through a server based on privacy protection.
Background
With the development of computer technology, machine learning is more and more widely applied. Artificial intelligence (ArtificialIntelligence), abbreviated in english as AI. The method is a new technical science for researching and developing theories, methods, technologies and application systems for simulating, extending and expanding human intelligence. The artificial intelligence is fully based on machine learning, and great convenience is brought to human life. However, in some areas where the technology development is relatively lagged, in artificial intelligence applications, the following may be the case: (1) the data volume is rare, and the feasibility of establishing a deep model is poor; (2) lack of computational resources necessary to train machine learning models; (3) manual expertise is not sufficient, e.g. only some beginners, or practitioners with little expertise; and so on.
To address these issues, help such regions to promote artificial intelligence, they may be provided with machine learning models by the server. It may be provided with a machine learning model, for example, by: the server side receives related training data from the corresponding client side, trains out a proper machine learning model and feeds back the proper machine learning model to the client side; the client uploads the service data to be processed, and the server feeds back the processing result to the client. However, the communication data volume of the client and the server in such an operation mode is large, and it is not beneficial to protect the data privacy of both parties.
Disclosure of Invention
One or more embodiments of the present specification describe a method and apparatus for providing a business model to address one or more of the problems noted in the background.
According to a first aspect, a method for providing a business model is provided, which is applicable to a server, and the method includes: receiving a current semantic vector which is sent by a client and used for describing service data to be processed, wherein the current semantic vector is determined by processing the service data to be processed based on a first coding mode; matching the current semantic vector with each candidate semantic vector, wherein a single candidate semantic vector is determined by processing a corresponding single training data set based on a second coding mode, and the single training data set is used for training at least one candidate model; selecting a plurality of target models from the candidate models based on the matching result, and generating a first business model according to the fusion of the target models; and providing the first business model for a client so that the client performs relevant business processing on the business data to be processed according to the first business model.
In one embodiment, the current semantic vector has the same dimensionality as a single candidate semantic vector.
In one embodiment, the matching the current semantic vector with each candidate semantic vector comprises: for a single candidate semantic vector, determining the vector similarity of the current semantic vector and the single candidate semantic vector; and determining a matching result of the current semantic vector and the single candidate semantic vector according to the vector similarity.
In one embodiment, a single candidate semantic vector is determined by: aiming at each piece of training data in a single training data set corresponding to the single candidate semantic vector, determining each corresponding single data semantic vector based on the second coding mode respectively; and determining the single candidate semantic vector according to the average vector of the single data semantic vectors.
In one embodiment, each candidate model comprises a first candidate model, the matching result of a first candidate semantic vector corresponding to the first candidate model and the current semantic vector comprises a first matching degree of the first candidate semantic vector and the current semantic vector, and the first candidate model also corresponds to a first model index; the selecting a plurality of target models from the candidate models based on the matching result comprises: determining a first target score of the first candidate model according to the first matching degree and the first model index; determining whether the first candidate model is a target model according to the first target score.
In a further embodiment, the first model indicator comprises at least one of: accuracy, recall, F1 score; the first target score is positively correlated with both the first matching degree and the first model index.
In one embodiment, the generating a first business model from the fusion of the several target models comprises: under the condition that only the last hidden layer or output layer nodes of the target models are correspondingly consistent, adding a fusion layer for carrying out weighted average on the last hidden layer or output layer according to given weight after combining the target models to generate a first service model; and under the condition that the model structures of the target models are completely consistent, generating a first business model according to the average of each model parameter in the target models.
According to a second aspect, there is also provided an apparatus for providing a service model, which is provided at a server, the apparatus including: the system comprises an acquisition unit, a processing unit and a processing unit, wherein the acquisition unit is configured to receive a current semantic vector which is sent by a client and used for describing service data to be processed, and the current semantic vector is determined by processing the service data to be processed based on a first coding mode;
the matching unit is configured to match the current semantic vector with each candidate semantic vector, and each candidate semantic vector is determined by processing of a corresponding single training data set based on a second coding mode, wherein the single training data set is used for training at least one candidate model;
the determining unit is configured to select a plurality of target models from the candidate models based on the matching result, and generate a first business model according to the fusion of the target models;
and the providing unit is configured to provide the first service model to a client so that the client performs related service processing on the service data to be processed according to the first service model.
According to a third aspect, there is provided a computer readable storage medium having stored thereon a computer program which, when executed in a computer, causes the computer to perform the method of the first aspect.
According to a fourth aspect, there is provided a computing device comprising a memory and a processor, wherein the memory has stored therein executable code, and wherein the processor, when executing the executable code, implements the method of the first aspect.
According to the method and the device provided by the embodiment of the specification, a large number of candidate models are stored in advance at the server side, the candidate semantic vectors obtained by coding the training data of each candidate model are stored, when the client side needs the service model, the current semantic vector determined by own service data can be uploaded, the current semantic vector is matched with the candidate semantic vectors, and the target model is selected according to the matching result. This approach may better protect the data privacy of the client and generate less traffic. Particularly, the server side selects a plurality of target models, and after the plurality of target models are fused, the fusion result is provided for the client side, so that the model data privacy of the server side can be effectively protected. In summary, the method for providing the business model described in this specification can save communication and computing resources and improve the effectiveness of the application of the business model based on data privacy and by using less communication traffic.
Drawings
In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings needed to be used in the description of the embodiments are briefly introduced below, and it is obvious that the drawings in the following description are only some embodiments of the present invention, and it is obvious for those skilled in the art to obtain other drawings based on these drawings without creative efforts.
FIG. 1 illustrates an architectural diagram of an implementation of the present specification that provides a business model;
FIG. 2 illustrates a schematic diagram of a specific application scenario in which the present specification provides a business model;
FIG. 3 illustrates a flow diagram of a method of providing a business model, according to one embodiment;
4a, 4b and 4c are diagrams respectively illustrating various specific situations of generating a first business model according to each target model in the embodiment of the present specification;
FIG. 5 shows a schematic block diagram of an apparatus for providing a business model according to one embodiment.
Detailed Description
The scheme provided by the specification is described below with reference to the accompanying drawings.
First, a description will be given of an embodiment of the present invention with reference to fig. 1. As shown in fig. 1, in this implementation scenario, a computing platform provided at a server and a plurality of clients are included. Wherein, the computing platform can store a plurality of candidate models in advance. The candidate model may be trained by the computing platform in fig. 1, or may be trained by another computing platform and then stored in the computing platform in the implementation architecture of the present specification shown in fig. 1, which is not limited in the present specification. It is understood that the client shown in fig. 1 is relative to the server in the present specification, and in particular practice, the user of the client may be an application facilitator, a business platform merchant, and the like. The client can be realized by various computers, smart phones, tablet computers and the like with certain communication capacity.
The server can collect a large amount of training data in advance, and divide the training data according to various service scenes. It can be understood that a service scenario may correspond to a plurality of training data sets, for example, an order conversion rate model of a certain commodity, and a set of training data sets may be respectively selected for different regions (e.g., a first-line city, a second-line city, a third-line city, a fourth-line city, and the following cities). A batch of training data may train at least one candidate model, such as by logistic regression, decision trees, deep neural networks, each training a candidate model. In different service scenarios, the training data may be independent of each other or may be used in a cross manner, for example, a piece of commodity conversion rate data may be used in a service scenario of commodity conversion rate, and may also be used in a service scenario of user purchase behavior analysis, which is not limited in this specification.
Under the implementation framework of the specification, for a single batch of training data, a semantic vector can be determined to describe the training data. The semantic vector may be implemented by various encoding methods, such as DNN, GBDT, etc. The encoding function is to process the original data to obtain a semantic vector with a certain dimension to describe the original data. The original data is coded, so that a certain privacy data protection effect can be achieved. Alternatively, the training data may be encoded by an existing encoder. The computing platform in fig. 1 may pre-store each semantic vector corresponding to each batch of training data as a candidate semantic vector.
As shown in fig. 2, assuming that the computing platform of the server is set as a cloud, when the client needs a service model, the client may encode its own data (e.g., a small amount of training data) through a communication link and send the processed current semantic vector to the cloud, and the computing platform of the cloud selects a plurality of corresponding candidate models as target models according to a pre-stored matching relationship between each candidate semantic vector and the current semantic vector. Then, in order to protect data privacy of the server, the computing platform may fuse the selected candidate models to obtain a service model, and provide the service model for the client.
Under the technical concept of the specification, an expert model hub can be constructed by a server side to provide a large number of pre-trained candidate models, and remote client sides can obtain required business models from the server side for use, so that the model training efficiency of the client sides can be greatly improved, and the business models trained through rich training data are used in various local application programs of the client sides. Meanwhile, the privacy of the training data of the client and the server is protected, so that the problems mentioned in the background technology are solved.
It should be noted that the training data or the business data referred to in this specification may be at least one type of data such as characters, pictures, audio, video, and animation. The number of clients and computing platforms shown in fig. 1 is merely exemplary, and in practice, the number of clients may be any reasonable number.
The method of providing a business model under the technical idea of the present specification is described in detail below.
FIG. 3 illustrates a flow diagram for providing a business model according to one embodiment of the present description. The execution subject of the process may be any computer, device, server, or the like with certain computing power, such as the computing platform shown in fig. 1. As shown in fig. 3, the process of providing the business model includes: step 301, receiving a current semantic vector used for describing to-be-processed service data and sent by a client, wherein the current semantic vector is determined based on the processing of the to-be-processed service data in a first coding mode; step 302, matching the current semantic vector with each candidate semantic vector, wherein a single candidate semantic vector is determined by processing a single training data set based on a second coding mode, and the single training data set is used for training at least one candidate model; step 303, selecting a plurality of target models from the candidate models based on the matching result, and generating a first business model based on the fusion of the plurality of target models; and step 304, providing the first business model to the client, so that the client performs relevant business processing on the business data to be processed according to the first business model.
First, through step 301, a current semantic vector sent by a client for describing service data to be processed is received. The service data to be processed may be a small amount of sample data, or may be service data to be processed, which is not limited herein. The service data to be processed can be determined according to a specific service scenario. For example, in a conversion rate scenario, the user's browsing track, click preference, recent browsing history, etc. may be included, and in a user's financial default risk scenario, the user's financial default history may include age, gender, income, historical default history, etc. It is worth mentioning that in case the service data to be processed is sample data, it may not include the tag data. This is because, when the service determines the business model required by the client, the service can determine the business model only according to the proximity of the business data to be processed itself to the training data processed by the candidate model.
By encoding, it is understood that an input sequence is converted into a vector of fixed length, and the encoding process can be used to resolve the linguistic meaning in the input character sequence. For the client, the input sequence may be, for example, a feature vector extracted from the traffic data to be processed. The feature vector of the client can be extracted according to the service feature sequence given by the server. For example, a first dimension of a 10-dimensional feature vector for a certain good represents a category of the good, a second dimension represents an applicable age, and so on.
In an alternative embodiment, the encoding process may change the dimensions of the vectors, for example, processing 10-dimensional feature vectors into 20-dimensional vectors and processing 100-dimensional vectors into 30-dimensional vectors. The client can determine the dimension of the current semantic vector according to the indication of the server.
There are various encoding methods for obtaining semantic vectors, such as DNN encoder and GBDT encoder. Under the condition that a plurality of pieces of service data to be processed exist, each piece of service data to be processed can be respectively encoded to obtain corresponding semantic vectors, and then each semantic vector is fused (averaged), so that the current semantic vector is obtained. A large number of experiments show that in the process of fusing semantic vectors of a plurality of pieces of service data, the difference between coding results obtained by the existing different coding modes can be ignored. Therefore, in this specification, the service end and the client end may respectively encode the service data by using any one of these encoding methods. In order to protect data privacy, the client may select any encoding mode (referred to as a first encoding mode in this specification), and perform encoding processing on the to-be-processed service data according to parameters such as vector dimensions and hidden layer number indicated by the server, so as to obtain a current semantic vector.
Next, in step 302, the current semantic vector is matched with each candidate semantic vector. It can be understood that the server stores the candidate semantic vectors in advance. For the server, a batch of training data for training at least one candidate model may form a training data set, and one training data set may be used for training at least one candidate model and determining a candidate semantic vector.
Wherein the single candidate semantic vector may be determined based on the processing of the single training data set by the second encoding mode. Here, the second encoding scheme may be various possible encoding schemes similar to the first encoding scheme, and is not limited herein. The first encoding method and the second encoding method may be the same encoding method or different encoding methods. The candidate semantic vectors corresponding to a single training data set may be average vectors of semantic vectors obtained by encoding a plurality of training data in the first encoding mode.
Each candidate semantic vector may be of a predetermined dimension, for example 30 dimensions. Generally, the server instructs the client to perform similar semantic coding according to the number of coding network layers of the candidate semantic vector, a feature extraction mode before coding, a semantic vector dimension obtained after coding, and the like. Therefore, the current semantic vector sent by the client has the same dimension as the candidate semantic vector. Optionally, both the server and the client may perform semantic coding by using the existing coding method, and the difference between the coding results obtained by using different coding methods is negligible.
The matching of the current semantic vector and the respective candidate semantic vectors may be performed by means such as cosine similarity, euclidean distance, etc. Taking cosine similarity as an example, assuming that the current semantic vector a and the first candidate semantic vector b are both n-dimensional vectors, the cosine similarity of the vector a and the vector b may be a ratio of a sum of products of corresponding elements in the vector a and the vector b to a product of a modulus of the vector a and a modulus of the vector b, such as:
wherein the content of the first and second substances,
the element values of the ith dimension of vector a and vector b, respectively. Under the matching mode of cosine similarity, the matching result takes a value between 0 and 1. The closer the cosine similarity is to 1, the more similar the current semantic vector and the candidate semantic vector, whereas the closer the cosine similarity is to 0, the more perpendicular the current semantic vector and the candidate semantic vector are (the lower the similarity).
According to the selected matching mode, the matching result corresponding to the current semantic vector and each candidate semantic vector can be determined.
Next, in step 303, several target models are selected from the candidate models based on the matching result, and a first business model is generated based on the fusion of the several target models. It can be understood that the matching result of the current semantic vector and the candidate semantic vector can describe the similarity between the to-be-processed service data of the client and the corresponding training data.
Under the condition that the model indexes are the same, the more similar the business data to be processed is to a certain batch of training data, the better the processing effect can be obtained by processing the business data to be processed through the candidate model obtained by training with the batch of training data. The model index may be understood as an index for evaluating model performance for a specific service. For example: for candidate models considering accuracy performance, accuracy can be used as a model index, and accordingly the higher the accuracy is, the better the model effect is considered to be; for candidate models considering recall performance, recall rate can be used as a model index, and accordingly the higher the recall rate, the better the model effect is considered. In some service scenarios, such as the binary service scenario of target identification, high accuracy means reduced recall rate, and high recall rate means reduced accuracy, and at this time, the F1 score, i.e. the weighted average of accuracy and recall rate, can also be used as the model index, and accordingly the model effect is considered to be better when the F1 score is higher.
In this step 303, candidate models may be selected based on the matching results of the to-be-processed service data and the respective candidate semantic vectors. The matching result here can be described by the vector similarity in step 302, for example, the greater the similarity between the corresponding semantic vectors, the higher the matching degree.
According to one embodiment, the target model may be selected from candidate models corresponding to the candidate semantic vector having the highest degree of matching with the current semantic vector. In one embodiment, several target models may be randomly selected from the candidate models corresponding to the candidate semantic vector having the highest degree of matching with the current semantic vector. In another embodiment, a plurality of target models whose model indexes satisfy a predetermined condition (e.g., greater than a predetermined threshold), or whose model indexes are best (selected in descending order), and the like, may be selected from the candidate models corresponding to the candidate semantic vector with the highest matching degree of the current semantic vector.
According to another embodiment, the target model may be selected according to a degree of match with the current semantic vector (e.g., positively correlated with the aforementioned similarity), and the model index resulting in a target score. The target score can be positively correlated with the matching degree and the model index respectively. For example, a target model is selected from the candidate models by using a weighted average of the matching degree and the model index, or a product of the matching degree and the model index, as a target score. At this time, several target models may be selected in order of a predetermined score threshold or target score from large to small.
In other embodiments, the target model may be selected in other reasonable manners, which are not listed here.
For the selected object models, if the number is one, the one object model may be used as the first business model, and if the number is multiple, the object models may be fused to generate the first business model. In general, in order to protect the privacy of the server model training data and the candidate models, a plurality of target models may be selected and fused to generate the first business model.
It will be appreciated that candidate models for processing similar training data will typically have similar or identical model structures. The two models have similar model structures, and can indicate that the two models have consistent structures at least in the last hidden layer or output layer. For example, in the classification model, the layer structures representing the classification probabilities (not necessarily corresponding in order) on several classification categories before a specific classification category is obtained are consistent. Typically, the layer before the specific classification category is obtained is usually the last hidden or output layer. For another example, in a scoring model, the scoring result is similar to the probability result of classifying into a certain category. The same model structure, for example, has the same number of hidden layers and the same number of model parameters on each hidden layer.
In an embodiment, each target model has the similar model structure as described above, and at this time, after each target model is combined, the last hidden layer or output layer is subjected to weighted fusion to obtain the first service model. The weighted weights of each target model may be positively correlated with the model indices or may be randomly generated. Alternatively, to control the model parameter value range of the target model, each weight may be a normalized weight (i.e., the sum of weights is 1). As shown in fig. 4a and 4b, in this case, it is equivalent to combine the object models together and add a fusion layer to the model parameters with weighting weights in the last hidden layer (or output layer). When the output layer of a single target model is the probability classified into a specific certain category or the scoring result, the probabilities or scoring scores of the output layers of the respective target models may be weighted and averaged, as shown in fig. 4a, and the fusion layer shown by the dotted line is added after the respective target models are combined. When the last hidden layer or output layer of a single object model obtains the probability of classifying into each category, or scores on multiple categories, weighted average can be performed on the probability values or scores on each category, as shown in fig. 4b, and a fusion layer shown by a dotted line is added after each object model is combined.
In another embodiment, each target model has the same structure, and at this time, the corresponding parameters of each target model may be weighted-averaged or added-averaged to obtain the first service model. In the case of a weighted average, the weighted weight may, for example, be positively correlated with the model index. In this case, the target models are merged, that is, the total amount of model parameters of the first business model is consistent with respect to the total amount of model parameters of the target models. As shown in fig. 4c, unlike fig. 4a and 4b, fig. 4c is a complete fusion of the corresponding model parameters of the respective object model, rather than adding a fusion layer after the last hidden or output layer of the respective object model as shown in fig. 4a and 4 b.
In other embodiments, there may be other reasonable fusion modes for the target model, which are not described herein again.
The first business model is then provided to the client, via step 304. According to the steps 301 to 303, a target model is selected by a candidate semantic vector which is more matched with a current semantic vector uploaded by the client based on the service data to be processed, and a first service model is generated based on the fusion of the target models, so that the first service model is determined by the server to be more suitable for the client. The server can feed back the first service model as a task processing result to the corresponding client.
The client may perform the relevant business process according to the first business model pair. For example, the service data to be processed is input into a first service model, and relevant service results, such as classification categories, scoring results, target identification results and the like, are determined according to output results of the first service model. It can be understood that, under the condition that the service data to be processed is sample data, the client can also finely tune the first service model according to the service data to be processed, so that the first service model is more suitable for the service requirement of the first service model.
In a possible design, the server can also update the candidate model stored locally to better adapt to the client requirements. For example, according to one embodiment, for a candidate model that is selected as the target model less frequently than a predetermined threshold (e.g., 1) within a predetermined time period (e.g., 2 years), the candidate model may be deleted or updated according to more training data and then re-used as the candidate model.
In the above process, the method for providing a service model provided in the embodiments of the present specification stores a large number of candidate models in advance at the server, and stores candidate semantic vectors obtained by encoding training data of each candidate model, and when the client needs a service model, can upload current semantic vectors determined by own sample data or service data to be processed, and match the current semantic vectors with the candidate semantic vectors, so as to select a target model according to a matching result. This approach may better protect the data privacy of the client and generate less traffic. Particularly, the server side selects a plurality of target models, and after the plurality of target models are fused, the fusion result is provided for the client side, so that the model data privacy of the server side can be effectively protected. In summary, the method for providing the business model described in this specification can save communication and computing resources and improve the effectiveness of the application of the business model based on data privacy and by using less communication traffic.
According to an embodiment of another aspect, an apparatus for providing a business model is also provided. The service data may be various types of data such as text, image, voice, video, animation, etc. The device can be arranged at a server side. FIG. 5 shows a schematic block diagram of an apparatus for providing a business model according to one embodiment. As shown in fig. 5, the apparatus 500 includes: the acquiring unit 51 is configured to receive a current semantic vector which is sent by the client and used for describing the to-be-processed service data, and the current semantic vector is determined based on the processing of the to-be-processed service data in the first coding mode; a matching unit 52 configured to match the current semantic vector with each candidate semantic vector, where a single candidate semantic vector is determined by processing a corresponding single training data set based on the second encoding mode, and the single training data set is used for training at least one candidate model; a determining unit 53 configured to select a plurality of target models from the candidate models based on the matching result, and generate a first business model according to fusion of the plurality of target models; the providing unit 54 is configured to provide the first service model to the client, so that the client performs relevant service processing on the to-be-processed service data according to the first service model.
It is to be understood that, in some embodiments, the apparatus 500 may further include a storage unit, configured to store each candidate semantic vector and each candidate model, which is not described herein again.
According to an embodiment, the matching unit 52 may be further configured to:
determining the vector similarity of the current semantic vector and a single candidate semantic vector aiming at the single candidate semantic vector;
and determining the matching result of the current semantic vector and the single candidate semantic vector according to the vector similarity.
In an optional implementation manner, each candidate model includes a first candidate model, a matching result of a first candidate semantic vector corresponding to the first candidate model and a current semantic vector includes a first matching degree of the first candidate semantic vector and the current semantic vector, and the first candidate model further corresponds to a first model index; the determining unit 53 may be further configured to:
determining a first target score of the first candidate model according to the first matching degree and the first model index;
determining whether the first candidate model is the target model according to the first target score.
Wherein the first model indicator may comprise at least one of: accuracy, recall, F1 score; the first target score is positively correlated with the first matching degree and the first model index.
According to one possible design, the determining unit 53 is further configured to:
under the condition that only the last hidden layer or output layer nodes of a plurality of target models are correspondingly consistent, adding a fusion layer for carrying out weighted average on the last hidden layer or output layer according to given weight after combining a plurality of target models to generate a first service model;
and under the condition that the model structures of the target models are completely consistent, generating a first business model according to the average of each model parameter in the target models.
It should be noted that the apparatus 500 shown in fig. 5 is an apparatus embodiment corresponding to the method embodiment shown in fig. 3, and the corresponding description in the method embodiment shown in fig. 3 is also applicable to the apparatus 500, and is not repeated herein.
According to an embodiment of another aspect, there is also provided a computer-readable storage medium having stored thereon a computer program which, when executed in a computer, causes the computer to perform the method described in connection with fig. 3.
According to an embodiment of yet another aspect, there is also provided a computing device comprising a memory and a processor, the memory having stored therein executable code, the processor, when executing the executable code, implementing the method in conjunction with fig. 3.
Those skilled in the art will recognize that, in one or more of the examples described above, the functions described in the embodiments of this specification may be implemented in hardware, software, firmware, or any combination thereof. When implemented in software, the functions may be stored on or transmitted over as one or more instructions or code on a computer-readable medium.
The above embodiments are only intended to be specific embodiments of the technical concept of the present disclosure, and should not be used to limit the scope of the technical concept of the present disclosure, and any modification, equivalent replacement, improvement, etc. made on the basis of the technical concept of the embodiments of the present disclosure should be included in the scope of the technical concept of the present disclosure.
Claims (14)
1. A method for providing a business model, which is applicable to a server side, wherein the business model is a machine learning model, and the method comprises the following steps:
receiving a current semantic vector which is sent by a client and used for describing service data to be processed, wherein the current semantic vector is determined by processing the service data to be processed based on a first coding mode;
matching the current semantic vector with each candidate semantic vector, wherein a single candidate semantic vector is determined by processing a corresponding single training data set based on a second coding mode, and the single training data set is used for training at least one candidate model;
selecting a plurality of target models from the candidate models based on the matching result, and generating a first business model according to the fusion of the target models;
and providing the first business model for a client so that the client performs relevant business processing on the business data to be processed according to the first business model.
2. The method of claim 1, wherein the current semantic vector has the same dimensionality as a single candidate semantic vector.
3. The method of claim 1, wherein the matching the current semantic vector to respective candidate semantic vectors comprises:
for a single candidate semantic vector, determining the vector similarity of the current semantic vector and the single candidate semantic vector;
and determining a matching result of the current semantic vector and the single candidate semantic vector according to the vector similarity.
4. The method of claim 1, wherein a single candidate semantic vector is determined by:
aiming at each piece of training data in a single training data set corresponding to the single candidate semantic vector, determining each corresponding single data semantic vector based on the second coding mode respectively;
and determining the single candidate semantic vector according to the average vector of the single data semantic vectors.
5. The method of claim 1, wherein each candidate model comprises a first candidate model, the matching result of a first candidate semantic vector corresponding to the first candidate model and the current semantic vector comprises a first matching degree of the first candidate semantic vector and the current semantic vector, and the first candidate model further corresponds to a first model index; the selecting a plurality of target models from the candidate models based on the matching result comprises:
determining a first target score of the first candidate model according to the first matching degree and the first model index;
determining whether the first candidate model is a target model according to the first target score.
6. The method of claim 5, wherein the first model metric comprises at least one of: accuracy, recall, F1 score; the first target score is positively correlated with both the first matching degree and the first model index.
7. The method of claim 1, wherein the generating a first business model from the fusion of the number of target models comprises:
under the condition that only the last hidden layer or output layer nodes of the target models are correspondingly consistent, adding a fusion layer for carrying out weighted average on the last hidden layer or output layer according to given weight after combining the target models to generate a first service model;
and under the condition that the model structures of the target models are completely consistent, generating a first business model according to the average of each model parameter in the target models.
8. An apparatus for providing a business model, which is provided at a server side, wherein the business model is a machine learning model, the apparatus comprising:
the system comprises an acquisition unit, a processing unit and a processing unit, wherein the acquisition unit is configured to receive a current semantic vector which is sent by a client and used for describing service data to be processed, and the current semantic vector is determined by processing the service data to be processed based on a first coding mode;
the matching unit is configured to match the current semantic vector with each candidate semantic vector, and each candidate semantic vector is determined by processing of a corresponding single training data set based on a second coding mode, wherein the single training data set is used for training at least one candidate model;
the determining unit is configured to select a plurality of target models from the candidate models based on the matching result, and generate a first business model according to the fusion of the target models;
and the providing unit is configured to provide the first service model to a client so that the client performs related service processing on the service data to be processed according to the first service model.
9. The apparatus of claim 8, wherein the matching unit is further configured to:
for a single candidate semantic vector, determining the vector similarity of the current semantic vector and the single candidate semantic vector;
and determining a matching result of the current semantic vector and the single candidate semantic vector according to the vector similarity.
10. The apparatus of claim 8, wherein each candidate model comprises a first candidate model, the matching result of a first candidate semantic vector corresponding to the first candidate model and the current semantic vector comprises a first matching degree of the first candidate semantic vector and the current semantic vector, and the first candidate model further corresponds to a first model index; the determination unit is further configured to:
determining a first target score of the first candidate model according to the first matching degree and the first model index;
determining whether the first candidate model is a target model according to the first target score.
11. The apparatus of claim 10, wherein the first model metric comprises at least one of: accuracy, recall, F1 score; the first target score is positively correlated with both the first matching degree and the first model index.
12. The apparatus of claim 8, wherein the determining unit is further configured to:
under the condition that only the last hidden layer or output layer nodes of the target models are correspondingly consistent, adding a fusion layer for carrying out weighted average on the last hidden layer or output layer according to given weight after combining the target models to generate a first service model;
and under the condition that the model structures of the target models are completely consistent, generating a first business model according to the average of each model parameter in the target models.
13. A computer-readable storage medium, on which a computer program is stored which, when executed in a computer, causes the computer to carry out the method of any one of claims 1-7.
14. A computing device comprising a memory and a processor, wherein the memory has stored therein executable code that, when executed by the processor, implements the method of any of claims 1-7.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202010329629.6A CN111241850B (en) | 2020-04-24 | 2020-04-24 | Method and device for providing business model |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202010329629.6A CN111241850B (en) | 2020-04-24 | 2020-04-24 | Method and device for providing business model |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN111241850A CN111241850A (en) | 2020-06-05 |
| CN111241850B true CN111241850B (en) | 2020-07-17 |
Family
ID=70866270
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202010329629.6A Active CN111241850B (en) | 2020-04-24 | 2020-04-24 | Method and device for providing business model |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN111241850B (en) |
Families Citing this family (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN112700008B (en) * | 2021-01-06 | 2022-06-28 | 青岛弯弓信息技术有限公司 | Model matching processing method and system for cloud configuration platform |
| CN113190587A (en) * | 2021-04-20 | 2021-07-30 | 北京异乡旅行网络科技有限公司 | Data processing method and device for realizing service data processing |
| CN116862012A (en) * | 2022-03-28 | 2023-10-10 | 华为技术有限公司 | Machine learning model training method, business data processing method, device and system |
Family Cites Families (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US10664662B2 (en) * | 2018-04-18 | 2020-05-26 | Microsoft Technology Licensing, Llc | Multi-scale model for semantic matching |
| CN109598285A (en) * | 2018-10-24 | 2019-04-09 | 阿里巴巴集团控股有限公司 | A kind of processing method of model, device and equipment |
| CN110287332B (en) * | 2019-06-06 | 2021-10-08 | 中国人民解放军国防科技大学 | Method and device for selecting simulation model in cloud environment |
| CN110287421A (en) * | 2019-06-28 | 2019-09-27 | 北京金山安全软件有限公司 | Information content recommendation method and device and electronic equipment |
-
2020
- 2020-04-24 CN CN202010329629.6A patent/CN111241850B/en active Active
Also Published As
| Publication number | Publication date |
|---|---|
| CN111241850A (en) | 2020-06-05 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN111241850B (en) | Method and device for providing business model | |
| CN109344884B (en) | Media information classification method, method and device for training picture classification model | |
| KR102234850B1 (en) | Method and apparatus for complementing knowledge based on relation network | |
| CN105453122A (en) | Contextual mobile application advertisements | |
| EP3885966B1 (en) | Method and device for generating natural language description information | |
| CN109471978B (en) | Electronic resource recommendation method and device | |
| CN115062732B (en) | Resource sharing cooperation recommendation method and system based on big data user tag information | |
| CN113392236A (en) | Data classification method, computer equipment and readable storage medium | |
| CN113011884B (en) | Account feature extraction method, device, equipment and readable storage medium | |
| CN111324773A (en) | Background music construction method and device, electronic equipment and storage medium | |
| CN115587535A (en) | Model construction optimization method, device, storage medium, and program product | |
| CN112258250A (en) | Target user identification method and device based on network hotspot and computer equipment | |
| CN112215629B (en) | Multi-target advertisement generating system and method based on construction countermeasure sample | |
| CN109710842B (en) | Business information pushing method and device and readable storage medium | |
| CN114119123A (en) | Information pushing method and device | |
| CN111460113A (en) | Data interaction method and related equipment | |
| CN111382232A (en) | Question and answer information processing method and device and computer equipment | |
| CN116401522A (en) | Financial service dynamic recommendation method and device | |
| US20230237093A1 (en) | Video recommender system by knowledge based multi-modal graph neural networks | |
| CN115186085A (en) | Reply content processing method and interaction method of media content interaction content | |
| CN110866195B (en) | Text description generation method and device, electronic equipment and storage medium | |
| CN114996435A (en) | Information recommendation method, device, equipment and storage medium based on artificial intelligence | |
| CN117688247B (en) | Recommendation method, terminal device and storage medium | |
| CN116913278B (en) | Voice processing method, device, equipment and storage medium | |
| US11734573B2 (en) | Image element matching via graph processing |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant | ||
| REG | Reference to a national code |
Ref country code: HK Ref legal event code: DE Ref document number: 40030603 Country of ref document: HK |