CN111190487A - Method for establishing data analysis model - Google Patents
Method for establishing data analysis model Download PDFInfo
- Publication number
- CN111190487A CN111190487A CN201911391335.XA CN201911391335A CN111190487A CN 111190487 A CN111190487 A CN 111190487A CN 201911391335 A CN201911391335 A CN 201911391335A CN 111190487 A CN111190487 A CN 111190487A
- Authority
- CN
- China
- Prior art keywords
- model
- user
- data
- cloud model
- server
- 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.)
- Pending
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F3/00—Input arrangements for transferring data to be processed into a form capable of being handled by the computer; Output arrangements for transferring data from processing unit to output unit, e.g. interface arrangements
- G06F3/01—Input arrangements or combined input and output arrangements for interaction between user and computer
- G06F3/011—Arrangements for interaction with the human body, e.g. for user immersion in virtual reality
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N20/00—Machine learning
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
-
- G—PHYSICS
- G16—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR SPECIFIC APPLICATION FIELDS
- G16H—HEALTHCARE INFORMATICS, i.e. INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR THE HANDLING OR PROCESSING OF MEDICAL OR HEALTHCARE DATA
- G16H80/00—ICT specially adapted for facilitating communication between medical practitioners or patients, e.g. for collaborative diagnosis, therapy or health monitoring
Abstract
The invention provides a method for establishing a data analysis model, which comprises the steps of learning and pre-training an initial cloud model by adopting a deep neural network at a server side based on current user data, and issuing the initial cloud model to different users; the user builds a user model of the user based on the received cloud model according to the data characteristics of the user, learns and trains by adopting a deep neural network, and transmits the trained user model back to the server; fusing the received user model based on a preset period by the server to obtain a new cloud model; and the user adjusts the user model of the user based on the received new cloud model, learns and trains by adopting a deep neural network, and returns the trained user model to the server. The model framework constructed based on the method can solve the problems of data islanding and individuation by combining the federal learning and homomorphic encryption technology, constructs a powerful machine learning model by summarizing data from different organizations, and can protect the privacy of users.
Description
Technical Field
The invention relates to the field of data analysis, in particular to the field of wearable health monitoring, specifically to a method for establishing data sharing models of different users, and more specifically to a method for establishing data analysis models.
Background
In recent years, with the rapid development of computing technologies, wearable technologies based on wearable devices such as smart phones, wristbands, smart glasses and the like can help people to know their health conditions more. Daily behavioral activities are closely related to human health and may be early signs of certain cognitive diseases. For example, changes in gait can lead to small vessel disease or stroke. Research shows that daily behavior activities can be identified by monitoring activities of a user by wearing a sensor. These wearable devices can easily access people's health information, including behavioral activities, sleep, exercise, and the like. The existing research shows that the wearable health monitoring technology can provide early warning for several cognitive diseases, such as Parkinson's disease, cerebrovascular diseases and the like, and can also be applied to other fields including mental health assessment, fall detection, sports monitoring and the like.
In fact, with the continuous progress of the technology, the application trend of wearable health monitoring based on wearable technology is more and more obvious. It is noted, however, that conventional healthcare applications typically build models by aggregating all user data. However, in practical applications, data is often isolated and cannot be easily shared due to privacy concerns, and the built model lacks the ability to be personalized. In wearable health monitoring field, a large amount of user data is usually used to train a machine learning model to track the health condition of a user, and conventional machine learning methods such as support vector machine, decision tree, hidden markov model, etc. are applied in many health monitoring fields.
Existing wearable health monitoring faces two major challenges.
The first challenge, as shown in fig. 1, is that in real life, data usually exists in an isolated island form, and different organizations and organizations have a large amount of data, but the data cannot be shared due to privacy and security issues, and in fig. 1, when the same user uses different products from two companies, his data is stored in the two companies and cannot be shared, which makes it difficult to train a high-performance model by using the data. In addition, in china, the united states and the european union have recently enforced protection of user data through different standardization systems, so that it is impossible to directly obtain a large amount of user data in practical applications.
Another important challenge is how to achieve personalization. In the prior art, most methods are based on a common server model applied to almost all users, a satisfactory machine learning model is trained by obtaining enough user data, and then the model is distributed to all user devices to track daily health information, and the process is not personalized. However, different users have different physical characteristics and daily behavior patterns, and thus a common model cannot achieve personalized healthcare.
Currently, there is no effective method to address the above challenges in the field of wearable health monitoring.
Disclosure of Invention
Therefore, the present invention is directed to overcome the above-mentioned drawbacks of the prior art, and to provide a new method for creating an analysis model, which is capable of creating personalized analysis models for different users and preventing data leakage.
According to a first aspect of the present invention, there is provided a method of building a data analysis model, comprising:
s1, learning and pre-training an initial cloud model at a server side by adopting a deep neural network based on current user data, and issuing the initial cloud model to different users; in the process of training the initial cloud model, an optimized cloud model loss function is taken as a training target, wherein the loss function is as follows:
S2, the user constructs a user model of the user based on the received cloud model according to the data characteristics of the user, learns and trains by adopting a deep neural network, and transmits the trained user model back to the server; in the process of training the user personalized model, a loss function of the optimized user personalized model is taken as a training target, wherein the loss function is as follows:
S3, fusing the received user models based on a preset period by the server to obtain a new cloud model; preferably, the fusing the received user models by the server adopts a model averaging method, which includes:
s31, randomly selecting a plurality of user models from the received user models;
s32, adding the parameters of the user model selected in the step S31, and taking the average value as the parameters of the new fused cloud model:
wherein f iss' denotes a new cloud model after fusion, and K denotes the number of selected user models.
S4, the user adjusts the user model of the user based on the received new cloud model, meanwhile, deep neural network learning and training are adopted, the trained user model is transmitted back to the server, and the step S3 is repeatedly executed. The adjustment of the user model based on the received cloud model means that parameters of the first four layers of the cloud model are fixed, and feature alignment is performed between the cloud model and the output of the second full connection layer of the user model to achieve adjustment of the user model. The feature alignment aims at optimizing an adjusted user model loss function, wherein the adjusted user model loss function is as follows:
wherein lCORALIn order to be a function of the alignment loss of the feature,
representing the norm of a square Hilbert-Schmidt matrix, CS，STCovariance matrices representing the source and target domain weights, respectively, η represents a trade-off factor.
According to another aspect of the present invention, there is provided a wearable healthcare analysis system, comprising: the system comprises a server, a cloud model and a plurality of user models; wherein the content of the first and second substances,
the server is configured to learn and pre-train an initial cloud model to be issued to different users based on current user data, continuously update the cloud model based on a user model fed back by the users, and issue the cloud model to the users to adjust the user model;
each user model is configured to build its own user model based on the received cloud model according to own data characteristics while learning and training by adopting a deep neural network, and to transmit the trained user model back to the server.
Compared with the prior art, the invention has the advantages that: according to the invention, the problems of data islanding and model personalization are solved by combining federal learning, transfer learning and homomorphic encryption technologies, a powerful machine learning model is constructed by summarizing data from different organizations, after a cloud server model is constructed, the framework provided by the invention can realize personalized model learning for each user by using transfer learning, and the framework can realize incremental updating. The method has high efficiency and expandability, and compared with the traditional learning method, in the behavior recognition experiment based on the smart phone, the recognition accuracy is improved by 5.3%, and the method can be deployed in a plurality of health monitoring applications to continuously improve the performance of the health monitoring applications in real life.
Drawings
Embodiments of the invention are further described below with reference to the accompanying drawings, in which:
FIG. 1 is a schematic diagram of a data island problem of wearable health monitoring analysis in the prior art;
FIG. 2 is a block diagram of a framework for modeling data analysis according to an embodiment of the present invention;
FIG. 3 is a schematic diagram of feature alignment during a process of building a data analysis model according to an embodiment of the present invention;
fig. 4 is a schematic diagram illustrating the extensibility of a model constructed by a method for establishing a data analysis model according to an embodiment of the present invention compared with a model constructed by another method.
Detailed Description
In order to make the objects, technical solutions and advantages of the present invention more apparent, the present invention is further described in detail by embodiments with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.
For better understanding of the present invention, a background corresponding to the technology adopted by the present invention will be described first.
Federal machine learning, originally proposed by google, is a popular field of machine learning in recent years, where they train distributed machine learning models based on mobile phones distributed around the world, with the key idea of protecting user data during this process. Later, some researchers began to focus on privacy-preserving machine learning, joint multitask learning, and personalized federal learning. Federal learning can solve the data islanding problem by training privacy preserving models in the network. Federal learning is mainly divided into three categories: horizontal federal learning, vertical federal learning, federal migratory learning. The method is based on the category of federal transfer learning, and can be applied to the field of wearable health monitoring due to the fact that the federal transfer learning can efficiently achieve privacy protection and solve the problems of data islands and the like.
Migration learning attempts to migrate knowledge from a labeled data domain (i.e., a source domain) to an unlabeled data domain (i.e., a target domain). In the context of transfer learning, where data domains are often distinct but related, which makes knowledge transfer possible, the key idea of transfer learning is to reduce the distribution differences between the different data domains. For this reason, there are mainly two methods: 1. example weighting methods, reusing samples of the source domain according to some weighting algorithm; 2. and the characteristic matching method is used for performing subspace learning or distribution alignment by utilizing the subspace geometric structure so as to reduce the marginal distribution or conditional distribution difference between the source domain and the target domain. In recent years, the deep migration learning method achieves excellent performance in various migration learning tasks, because more deeply migratable feature representations can be learned by using a deep convolutional neural network compared with the conventional method.
For example, N is used to represent the number of different users, s1，s2，...，sNRepresents said N users by
To represent the data of each user, the conventional method needs to gather all the user data to get
Then training a model
However, in the prior art, in order to prevent the existence of an information island caused by data exposure of a user, the difficulty of summarizing all data is large, and a trained model cannot meet the requirements of different users. In the invention, all data are gathered in a federal learning mode to train a federal model
Wherein each user's data is not exposed to other users
Representing the model accuracy, the goal of the invention is to ensure that the accuracy of the federal model is not less than or even better than that of the traditional method:
where Δ represents a non-negative minimum.
According to an embodiment of the present invention, taking N users as an example, as shown in fig. 2, the method of the present invention is adopted to establish an analysis model a, a model B, a … model N (N models) and a cloud model for users A, B, … N (N users), and the method of the present invention adopts a federal migration learning method to establish an analysis model, including establishing a cloud model at a server side and an individualized user model that a user trains himself based on the cloud model, and feeding back the user model to the server side to update the cloud model, and sending the updated cloud model to the user, and repeating this way, so that the cloud model is continuously updated to better satisfy individualized requirements of different user data. In the process of establishing an analysis model, three aspects are mainly included: federal learning, migratory learning, and model incremental updates. For better understanding, use
Respectively representing the summarized labeled data samples at the server side and the data generated by the user, wherein n represents the total number of the data samples, i represents the serial number of the ith data sample, and xiRepresenting data samples, yiThe class label corresponding to the data sample is expressed by fsAnd fuA server-side model and a user model are shown separately, each aspect being described in detail below.
First, federal learning.
Firstly, training a cloud model of a server side based on public data, wherein the cloud model is a public model stored on a cloud server, and is a public model which is trained on the server side by using all available public data relative to each user model; and then distributing the cloud model to all users, wherein each user trains and fine-tunes their own model on their own data, and in the process of training the own user model by the users, as shown in fig. 3, fixing the parameters of the first 4 layers of the cloud model, so that the parameters are not updated during the training process, and only the parameters of the later layers are updated. The trained user models are then uploaded to the cloud to assist in training new cloud models. It is worth noting that the process does not share any user data or information, but shares the encrypted model parameters, so that the problem of data islanding can be solved under the condition of ensuring the privacy of the user. Federal learning is the main computational model of the invention, which relates to the whole process of model construction and parameter sharing, and after learning a server model, it can be directly applied to users, which is what traditional healthcare applications do. It is clear that the sample data in the server has a highly different probability distribution than the data generated by each user, and therefore the generic model does not perform well enough in terms of personalization. In addition, user models cannot be easily updated continuously due to privacy security issues, and federal learning can effectively address this issue.
The invention adopts federal learning to realize the training and sharing of the encryption model, and aims to solve the challenge of data islanding and simultaneously ensure the privacy and the safety of user data. This section mainly comprises two sections: cloud model learning and user model learning. After the trained server cloud model is obtained, the server cloud model is distributed to the user side to help the user train and fine-tune the user model. For each user, the method fine-tunes the model of the user with the help of a pre-trained server cloud model, so that the model is more suitable for some behavior characteristics of the user. In the invention, the deep neural network is adopted to learn the cloud model and the user model. The deep neural network performs end-to-end feature learning and training of the classifier by taking the original input of user data as input, and then the objective function to be optimized by the cloud model is as follows:
by optimizing the cloud model, the prediction is more accurate according to the input data. Wherein the content of the first and second substances,
a loss function representing the network model, e.g., cross entropy loss for classifying problems, and Θ represents all the parameters to be learned, including weights and bias terms.
After the pre-trained cloud model is obtained, it is distributed to all users, from the "wall" { Q in FIG. 21、…QN-1It can be seen that direct sharing of user information is prohibited. In federal learning, homomorphic encryption can be used to avoid information leakage. Similarly, the encryption schemes for the weights and bias terms of the model follow the same idea. For a real number a, the invention uses<a>To indicate its homomorphic encryption result, in which for any two real values a, b, there are<a>+<b>＝<a+b>. Therefore, it is possible to proceed without any leakage of privacy information between users with assuranceThe parameters of the line model are shared. Through federal learning, the method can ensure that all data of the user are collected under the condition of user privacy, and the optimization goal of any user u model is as follows:
by optimizing the user model, the prediction is more accurate according to the input data.
When all user models fuAfter training and fine adjustment are completed, the models are uploaded and collected to the server, and then the server performs model fusion on the user models, so that updating of the cloud model is achieved. According to an embodiment of the invention, by adopting a model fusion mode of model averaging, an average model can achieve better convergence performance in the aspect of reducing loss. In consideration of the computational complexity of model fusion, in each training period, the invention only randomly selects K user models from all the user models to perform model fusion, and the process is that the parameters of corresponding model structures of the K user models are added and averaged to obtain a fused model, namely a new cloud model, so that the new cloud model encodes the parameter information of a plurality of user models, and the updating method of the cloud model can be formalized as follows:
wherein K represents the number of randomly selected users, f'sRepresenting a new cloud model after model fusion, fukRepresenting a selected user model, after a plurality of training cycles, the updated cloud model f's can cover parameter information of almost all user models, so that better generalization performance can be achieved. Given the computational burden, the server can make scheduled updates (e.g., every night), and then get a new server model f's. It is noted that the new cloud model is based on the experience of all other users, and therefore, the new cloud model is betterGeneralization performance.
Second, transfer learning
The purpose of the transfer learning is to complete model transfer, as shown in the right part of fig. 2, each user can perform personalized training by integrating a cloud model and previous user models and data thereof to obtain a new personalized user model, and in this step, as a large distribution difference exists between the cloud model and the user model, the model transfer is completed by using the transfer learning, so that the personalized model learning is realized. It is worth noting that all parameter sharing procedures do not involve user data leakage, but are done by homomorphic encryption.
Federal learning solves the problem of data islanding under the condition of ensuring privacy and safety of user data, so that the method can use all summarized user data to construct a model. Another important challenge is the personalization of the model, which does not necessarily perform well for a specific user due to the distribution difference between the user and the cloud data sample, although the present invention can directly use the cloud model, and the general cloud model in the server only learns the rough features common to all users, but cannot learn the fine-grained features of a specific user.
In the present invention, the present invention uses transfer learning to build a personalized user model for each user. It has been the surface of research that in deep neural networks, lower-level features encode more low-level features that are highly migratory, while higher-level features learn specific features that are appropriate for certain specific tasks. Based on this, after obtaining the parameters of the cloud model, according to one embodiment of the present invention, the users are subjected to migration learning to learn a personalized model suitable for each user.
Referring to fig. 3, a process of transfer learning for a particular neural network is shown, which includes two convolutional layers, two max-pooling layers, two fully-connected layers, and a classification layer for classification. The network model is used for behavior recognition of a person, wherein the input data is an activity signal of a user, and the output is an activity category corresponding to the activity signal. In model migration, convolutional layers are intended to extract low-level features of the activity, so the invention fixes the parameters of these layers to the max pooling layer, i.e. does not update their parameters in back-propagation, as for the two fully connected layers, since they belong to the higher layer, it is the higher-level abstract features that are extracted, which are more focused on specific tasks, so their parameters are updated during training. Using the softmax function as the classification function, it can be expressed as:
wherein z iscProbability, y, of the learned class ciRepresenting the final classification result, e represents the index e in the mathematics, and the index j represents the predicted result for the jth input data sample.
The invention adapts to the input of different fields by performing feature alignment between the outputs of the second fully-connected layer of the model (such as 'feature alignment' in figure 3), normalizes the weight, performs the feature alignment operation between the outputs of the second fully-connected layer before the final classification layer given the cloud model and the user model from the server to obtain a loss function, and combines the loss function with the loss function of the classification layer to be used as the optimization target of the user model. This feature-aligned objective function is used to align the second order statistics between the inputs (the feature-aligned inputs are the output of the second fully-connected layer of the cloud model and the user model). Formally, this alignment loss function can be expressed as:
wherein the content of the first and second substances,
representing the norm of a square Hilbert-Schmidt matrix, CS，CTCovariance matrices representing source and target domain weights, respectively. Wherein, the source field refers to the data field with label, that is, the data are all the tape classη represents a trade-off factor, the final optimization objective function of the user model after feature alignment can be expressed as:
the specific learning process of the transfer learning of the invention is as follows: the user model optimization target is minimized through model training, and then the distribution difference between the cloud model feature distribution and the user feature distribution is reduced, so that the feature distribution of the universal cloud model is more consistent with the specific feature distribution of each user, and therefore feature migration, namely personalized user model learning, is realized, and thousands of people and thousands of faces are realized.
Third, the model is updated incrementally.
The purpose of the incremental updating of the model is that when a user has new data, the incremental updating of the model can be obtained by updating the user model and the cloud model simultaneously, so that the personalization degree of the model is continuously improved.
When a user has new data, the method and the system can update the user model and the cloud model simultaneously to obtain incremental update of the model. Thus, the longer a user uses a product, the more personal data the user accumulates and the higher the degree of model personalization. The invention can also adopt other machine learning methods except the deep network, such as gradient boosting decision trees, random forests and the like, and the lightweight models can be deployed to wearable equipment with calculation limitation, so that the invention is more universal for practical wearable health monitoring application.
The specific process of incremental updating comprises the following steps: firstly, a cloud model is constructed, the cloud model is issued to all users, then each user finely trains the user model fu of the user according to data information of the user to update the user model, then the user side uploads the models to the server side, and the server side carries out model fusion by using a model averaging method after taking the users
And updating the cloud model. And then, the updated user models are uploaded to the server again to update the cloud models, and the effect of incremental updating is achieved by repeating the steps.
In order to better understand the present invention, the following is a further description of the invention and the effects achieved by it through an example experiment.
In order to verify the effectiveness of the data analysis model established for wearable healthcare, the invention performs behavior recognition experiments in several public data sets.
Data set: the invention uses a public data set UCI Smartphone for human behavior recognition, which contains 6 collected activities of 30 users, including walking, going upstairs, going downstairs, sitting, standing, lying down, the age of the 30 users varying from 19 to 48 years, each user wearing a Smartphone in the waist, using a cellphone-embedded accelerometer and gyroscope to acquire user acceleration and 3-axis angular velocity at a constant rate of 50Hz, and then manually tagging these data categories. The data sets obtained were divided into two groups, 70% as training data and 30% as test data, for a total of 10299 training examples, and the statistical information of the data sets is shown in table 1 below:
TABLE 1 statistical information of data sets
| Number of users | Number of movements | Acquisition rate | Sensor with a sensor element | Number of examples |
| 30 | 6 | 50Hz | Accelerometer, gyroscope | 10299 |
In order to construct an environment suitable for verifying the invention, the invention changes the standard setting of a data set, 5 users (user IDs 26-30) are extracted and regarded as isolated users which cannot be shared due to privacy security, and the data of the other 25 users are used for training a cloud model. Therefore, it is an object of the present invention to improve the behavior recognition accuracy of 5 isolated users without breaking user data privacy and security using a cloud model and the 5 users, which is, in short, a variant of the framework shown in fig. 2, in which the number of users is N-5.
The comparison method comprises the following steps: in order to show the superiority of the method, the method is compared with the traditional learning method, namely the server cloud model is used for testing the behavior recognition accuracy of each user, and for the sake of simplicity in expression, the method is represented by NoFed.
The invention also compares the performance with K-Nearest Neighbors (KNN), Support Vector Machines (SVM) and Random Forest (RF), and uses cross validation to adjust the hyper-parameters of all comparison methods. For fairness, the invention performed 5 total experiments and recorded the average accuracy. The advantages of the present invention are more fully reflected in comparison to these methods.
Evaluation criteria: the invention selects the classification precision on the user behavior recognition data as the criterion of the method performance evaluation, and is widely applied to the evaluation of a large number of machine learning related methods. The accuracy is calculated by the following formula
Wherein y (x) and
respectively representing the true label and the predicted result.
And (4) classification results: the results of comparing the behavior recognition progress of the experiment of the present invention are shown in table 2. As can be seen from the table, the framework of the invention achieves the best classification precision for all users, and compared with NoFed, the average precision of the invention is improved by 5.3%; compared with the traditional method, the method of the invention also greatly improves the identification precision. The method of the present invention is highly advantageous compared to several other methods.
Table 2 action recognition classification accuracy (%)
| User' s | KNN | SVM | RF | NoFed | The invention |
| P1 | 83.8 | 81.9 | 87.5 | 94.5 | 98.8 |
| P2 | 86.5 | 96.9 | 93.3 | 94.5 | 98.8 |
| P3 | 92.2 | 97.2 | 88.9 | 93.4 | 100.0 |
| P4 | 83.1 | 95.9 | 91.0 | 95.5 | 99.4 |
| P5 | 90.5 | 98.6 | 91.6 | 92.6 | 100.0 |
| AVG | 87.2 | 94.1 | 90.5 | 94.1 | 99.4 |
As can be seen from the above experimental results, the deep method (NoFed and the present invention) achieves better results than the conventional method, which has to rely on the manually-made feature learning, due to the strong representation capability of the deep neural network. Another advantage of deep learning is that the model can be updated gradually on-line without retraining, whereas traditional methods require further additions to the algorithm. Such attributes are of great value in federal migration learning.
And (3) evaluating expandability: to evaluate the scalability of the present invention, it compares it with two methods: 1. fine tuning, namely performing fine tuning training on the network of each user only without performing feature alignment operation, thereby reducing the distribution difference between data fields; 2. the loss of alignment of the present invention is replaced by the loss of migration using the Maximum Mean Difference (MMD). The comparative results are shown in FIG. 4.
As can be seen from fig. 4, the average accuracy of the result of the transfer learning is improved by about 4% compared with that of the non-transfer (NoFed) method, which indicates that the transfer learning of the present invention is efficient and highly scalable. Thus, the present invention is a general framework that can be extended to many applications by other transfer learning algorithms.
Fine particle size analysis: for a more detailed analysis of the present invention, the present invention compares the confusion matrix of the present invention and the method without migration (NoFed), which is referred to as an effective measure of the performance of the display method, as it provides a fine-grained classification result for each task. For simplicity, the invention shows in Table 3 the confusion matrix for user 2, where the accuracy (P), recall (R) and harmonic score (F1) are all calculated to give a more comprehensive result, with accuracy (P), C1 for walking, C2 for going upstairs, C3 for going downstairs, C4 for sitting, C5 for standing, and C6 for lying down, with the results of the other users following the same trend.
TABLE 3 non-migratory learning (NoFed) and Classification results report of the present invention
Combining the results of tables 2 and 3, it is clear that the present invention can achieve not only the best accuracy, but also the best precision, recall and harmonic mean. The confusion matrix shows that the invention can significantly reduce the error rate of behavior recognition classification, especially in the category C1 (walking), which means that the invention is an effective algorithm for recognizing the activity because walking is the most common activity in health care.
In summary, the analysis model framework for wearable health care designed by the invention is the first federal migration learning framework for the field of wearable health care, and collects data from different users or organizations without destroying privacy and security of the data, and realizes personalized model learning through knowledge migration. The framework proposed by the present invention is efficiently extensible, can serve as a standard framework for many healthcare applications, and can be easily deployed into other healthcare applications. The invention can summarize data from different users or organizations without destroying privacy and security, and solves the challenge of data islanding; the invention also uses the transfer learning technology, and the personalization of the user model can be realized through model transfer, and the more new data of the user is generated, the better the model personalization effect obtained through transfer learning is. The framework proposed by the present invention is efficient and extensible, it can be a standard framework for many healthcare applications, and it can be easily deployed into other healthcare applications.
It should be noted that, although the steps are described in a specific order, the steps are not necessarily performed in the specific order, and in fact, some of the steps may be performed concurrently or even in a changed order as long as the required functions are achieved.
The present invention may be a system, method and/or computer program product. The computer program product may include a computer-readable storage medium having computer-readable program instructions embodied therewith for causing a processor to implement various aspects of the present invention.
The computer readable storage medium may be a tangible device that retains and stores instructions for use by an instruction execution device. The computer readable storage medium may include, for example, but is not limited to, an electronic memory device, a magnetic memory device, an optical memory device, an electromagnetic memory device, a semiconductor memory device, or any suitable combination of the foregoing. More specific examples (a non-exhaustive list) of the computer readable storage medium would include the following: a portable computer diskette, a hard disk, a Random Access Memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), a Static Random Access Memory (SRAM), a portable compact disc read-only memory (CD-ROM), a Digital Versatile Disc (DVD), a memory stick, a floppy disk, a mechanical coding device, such as punch cards or in-groove projection structures having instructions stored thereon, and any suitable combination of the foregoing.
Having described embodiments of the present invention, the foregoing description is intended to be exemplary, not exhaustive, and not limited to the embodiments disclosed. Many modifications and variations will be apparent to those of ordinary skill in the art without departing from the scope and spirit of the described embodiments. The terminology used herein is chosen in order to best explain the principles of the embodiments, the practical application, or improvements made to the technology in the marketplace, or to enable others of ordinary skill in the art to understand the embodiments disclosed herein.
Claims (10)
1. A method of creating a data analysis model comprising
S1, learning and pre-training an initial cloud model at a server side by adopting a deep neural network based on current user data, and issuing the initial cloud model to different users;
s2, the user constructs a user model of the user based on the received cloud model according to the data characteristics of the user, learns and trains by adopting a deep neural network, and transmits the trained user model back to the server;
and S3, fusing the received user models based on a preset period by the server to obtain a new cloud model.
2. The method of creating a data analysis model of claim 1, further comprising:
s4, the user adjusts the user model of the user based on the received new cloud model, meanwhile, deep neural network learning and training are adopted, the trained user model is transmitted back to the server, and the step S3 is repeatedly executed.
3. The method of claim 2, wherein in the process of training the initial cloud model, an optimized cloud model loss function is taken as a training target, wherein the loss function is as follows:
4. The method of claim 1, wherein in the step S2, in the process of training the user-customized model, an optimized user-customized model loss function is taken as a training objective, wherein the loss function is:
5. The method according to claim 2, wherein the step S4 of adjusting the user model based on the received cloud model means fixing the parameters of the first four layers of the cloud model and performing feature alignment between the cloud model and the output of the second fully-connected layer of the user model to adjust the user model.
6. The method of claim 5, wherein feature alignment targets optimization of an adjusted user model loss function, wherein the adjusted user model loss function is:
wherein lCORALIn order to be a function of the alignment loss of the feature,
representing the norm of a square Hilbert-Schmidt matrix, CS，CTCovariance matrices representing the source and target domain weights, respectively, η represents a trade-off factor.
7. The method according to claim 1, wherein the step S3 of fusing the received user models by model averaging comprises:
s31, randomly selecting a plurality of user models from the received user models;
s32, adding the parameters of the user model selected in the step S31, and taking the average value as the parameters of the new fused cloud model:
wherein f iss' denotes a new cloud model after fusion, and K denotes the number of selected user models.
8. A wearable healthcare analysis system, comprising:
the system comprises a server, a cloud model and a plurality of user models; wherein the content of the first and second substances,
the server is configured to learn and pre-train an initial cloud model to be issued to different users based on current user data, continuously update the cloud model based on a user model fed back by the users, and issue the cloud model to the users to adjust the user model;
each user model is configured to build its own user model based on the received cloud model according to own data characteristics while learning and training by adopting a deep neural network, and to transmit the trained user model back to the server.
9. A computer-readable storage medium having embodied thereon a computer program, the computer program being executable by a processor to perform the steps of the method of any one of claims 1 to 7.
10. An electronic device, comprising:
one or more processors;
storage means for storing one or more programs which, when executed by the one or more processors, cause the electronic device to carry out the steps of the block link node management method according to any one of claims 1 to 7.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201911391335.XA CN111190487A (en) | 2019-12-30 | 2019-12-30 | Method for establishing data analysis model |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201911391335.XA CN111190487A (en) | 2019-12-30 | 2019-12-30 | Method for establishing data analysis model |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN111190487A true CN111190487A (en) | 2020-05-22 |
Family
ID=70709522
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201911391335.XA Pending CN111190487A (en) | 2019-12-30 | 2019-12-30 | Method for establishing data analysis model |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN111190487A (en) |
Cited By (29)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111709761A (en) * | 2020-06-08 | 2020-09-25 | 中国联合网络通信集团有限公司 | Distributed navigation menu generation method, device, equipment and storage medium |
| CN111738440A (en) * | 2020-07-31 | 2020-10-02 | 支付宝(杭州)信息技术有限公司 | Model training method and system based on field self-adaptation and federal learning |
| CN111783139A (en) * | 2020-06-29 | 2020-10-16 | 京东数字科技控股有限公司 | Federal learning classification tree construction method, model construction method and terminal equipment |
| CN111832729A (en) * | 2020-07-06 | 2020-10-27 | 东南数字经济发展研究院 | Distributed deep learning reasoning deployment method for protecting data privacy |
| CN111967309A (en) * | 2020-07-03 | 2020-11-20 | 西安电子科技大学 | Intelligent cooperative identification method and system for electromagnetic signals |
| CN111985650A (en) * | 2020-07-10 | 2020-11-24 | 华中科技大学 | Activity recognition model and system considering both universality and individuation |
| CN111985562A (en) * | 2020-08-20 | 2020-11-24 | 复旦大学 | End cloud collaborative training system for protecting end-side privacy |
| CN112052942A (en) * | 2020-09-18 | 2020-12-08 | 支付宝(杭州)信息技术有限公司 | Neural network model training method, device and system |
| CN112100295A (en) * | 2020-10-12 | 2020-12-18 | 平安科技（深圳）有限公司 | User data classification method, device, equipment and medium based on federal learning |
| CN112233118A (en) * | 2020-12-15 | 2021-01-15 | 南京可信区块链与算法经济研究院有限公司 | Eye fundus lesion image identification method and system based on incremental learning |
| CN112259238A (en) * | 2020-10-20 | 2021-01-22 | 平安科技（深圳）有限公司 | Electronic device, disease type detection method, apparatus, and medium |
| CN112580085A (en) * | 2021-02-22 | 2021-03-30 | 支付宝(杭州)信息技术有限公司 | Model training method and device |
| CN112637883A (en) * | 2020-12-09 | 2021-04-09 | 深圳智芯微电子科技有限公司 | Federal learning method with robustness to wireless environment change in power Internet of things |
| CN112668044A (en) * | 2020-12-21 | 2021-04-16 | 中国科学院信息工程研究所 | Privacy protection method and device for federal learning |
| CN112768056A (en) * | 2021-01-14 | 2021-05-07 | 新智数字科技有限公司 | Disease prediction model establishing method and device based on joint learning framework |
| CN112754473A (en) * | 2021-01-08 | 2021-05-07 | 华中科技大学鄂州工业技术研究院 | Spontaneous heating downloading and method for falling detection by fusing transfer learning algorithm |
| CN112819177A (en) * | 2021-01-26 | 2021-05-18 | 支付宝(杭州)信息技术有限公司 | Personalized privacy protection learning method, device and equipment |
| CN112836829A (en) * | 2021-01-26 | 2021-05-25 | 支付宝(杭州)信息技术有限公司 | Personalized privacy protection learning method, device and equipment |
| CN112926126A (en) * | 2021-03-31 | 2021-06-08 | 南京信息工程大学 | Federal learning method based on Markov random field |
| WO2021120676A1 (en) * | 2020-06-30 | 2021-06-24 | 平安科技（深圳）有限公司 | Model training method for federated learning network, and related device |
| CN113052333A (en) * | 2021-04-02 | 2021-06-29 | 中国科学院计算技术研究所 | Method and system for data analysis based on federal learning |
| CN113077056A (en) * | 2021-03-29 | 2021-07-06 | 上海嗨普智能信息科技股份有限公司 | Data processing system based on horizontal federal learning |
| CN113128701A (en) * | 2021-04-07 | 2021-07-16 | 中国科学院计算技术研究所 | Sample sparsity-oriented federal learning method and system |
| CN113179244A (en) * | 2021-03-10 | 2021-07-27 | 上海大学 | Federal deep network behavior feature modeling method for industrial internet boundary safety |
| CN113807157A (en) * | 2020-11-27 | 2021-12-17 | 京东科技控股股份有限公司 | Method, device and system for training neural network model based on federal learning |
| WO2021259090A1 (en) * | 2020-06-23 | 2021-12-30 | 华为技术有限公司 | Method and apparatus for federated learning, and chip |
| CN113902473A (en) * | 2021-09-29 | 2022-01-07 | 支付宝(杭州)信息技术有限公司 | Training method and device of business prediction system |
| WO2022236831A1 (en) * | 2021-05-14 | 2022-11-17 | 北京小米移动软件有限公司 | Model learning method, model learning apparatus, and storage medium |
| CN117373066A (en) * | 2023-12-07 | 2024-01-09 | 华侨大学 | Pedestrian re-identification method and system based on Yun Bian searching federal deep learning method |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107025205A (en) * | 2016-01-30 | 2017-08-08 | 华为技术有限公司 | A kind of method and apparatus of training pattern in distributed system |
| CN109886417A (en) * | 2019-03-01 | 2019-06-14 | 深圳前海微众银行股份有限公司 | Model parameter training method, device, equipment and medium based on federation's study |
| CN110399742A (en) * | 2019-07-29 | 2019-11-01 | 深圳前海微众银行股份有限公司 | A kind of training, prediction technique and the device of federation's transfer learning model |
| CN110443063A (en) * | 2019-06-26 | 2019-11-12 | 电子科技大学 | The method of the federal deep learning of self adaptive protection privacy |
| CN110555467A (en) * | 2019-08-13 | 2019-12-10 | 深圳创新奇智科技有限公司 | industrial data classification method based on model migration |
-
2019
- 2019-12-30 CN CN201911391335.XA patent/CN111190487A/en active Pending
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107025205A (en) * | 2016-01-30 | 2017-08-08 | 华为技术有限公司 | A kind of method and apparatus of training pattern in distributed system |
| CN109886417A (en) * | 2019-03-01 | 2019-06-14 | 深圳前海微众银行股份有限公司 | Model parameter training method, device, equipment and medium based on federation's study |
| CN110443063A (en) * | 2019-06-26 | 2019-11-12 | 电子科技大学 | The method of the federal deep learning of self adaptive protection privacy |
| CN110399742A (en) * | 2019-07-29 | 2019-11-01 | 深圳前海微众银行股份有限公司 | A kind of training, prediction technique and the device of federation's transfer learning model |
| CN110555467A (en) * | 2019-08-13 | 2019-12-10 | 深圳创新奇智科技有限公司 | industrial data classification method based on model migration |
Cited By (36)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111709761A (en) * | 2020-06-08 | 2020-09-25 | 中国联合网络通信集团有限公司 | Distributed navigation menu generation method, device, equipment and storage medium |
| WO2021259090A1 (en) * | 2020-06-23 | 2021-12-30 | 华为技术有限公司 | Method and apparatus for federated learning, and chip |
| CN111783139A (en) * | 2020-06-29 | 2020-10-16 | 京东数字科技控股有限公司 | Federal learning classification tree construction method, model construction method and terminal equipment |
| WO2021120676A1 (en) * | 2020-06-30 | 2021-06-24 | 平安科技（深圳）有限公司 | Model training method for federated learning network, and related device |
| CN111967309B (en) * | 2020-07-03 | 2024-02-06 | 西安电子科技大学 | Intelligent cooperative identification method and system for electromagnetic signals |
| CN111967309A (en) * | 2020-07-03 | 2020-11-20 | 西安电子科技大学 | Intelligent cooperative identification method and system for electromagnetic signals |
| CN111832729A (en) * | 2020-07-06 | 2020-10-27 | 东南数字经济发展研究院 | Distributed deep learning reasoning deployment method for protecting data privacy |
| CN111985650A (en) * | 2020-07-10 | 2020-11-24 | 华中科技大学 | Activity recognition model and system considering both universality and individuation |
| CN111985650B (en) * | 2020-07-10 | 2022-06-28 | 华中科技大学 | Activity recognition model and system considering both universality and individuation |
| CN111738440A (en) * | 2020-07-31 | 2020-10-02 | 支付宝(杭州)信息技术有限公司 | Model training method and system based on field self-adaptation and federal learning |
| CN111985562A (en) * | 2020-08-20 | 2020-11-24 | 复旦大学 | End cloud collaborative training system for protecting end-side privacy |
| CN112052942A (en) * | 2020-09-18 | 2020-12-08 | 支付宝(杭州)信息技术有限公司 | Neural network model training method, device and system |
| CN112052942B (en) * | 2020-09-18 | 2022-04-12 | 支付宝(杭州)信息技术有限公司 | Neural network model training method, device and system |
| CN112100295A (en) * | 2020-10-12 | 2020-12-18 | 平安科技（深圳）有限公司 | User data classification method, device, equipment and medium based on federal learning |
| CN112259238A (en) * | 2020-10-20 | 2021-01-22 | 平安科技（深圳）有限公司 | Electronic device, disease type detection method, apparatus, and medium |
| CN113807157A (en) * | 2020-11-27 | 2021-12-17 | 京东科技控股股份有限公司 | Method, device and system for training neural network model based on federal learning |
| CN112637883B (en) * | 2020-12-09 | 2023-04-28 | 深圳智芯微电子科技有限公司 | Federal learning method with robustness to wireless environment change in electric power Internet of things |
| CN112637883A (en) * | 2020-12-09 | 2021-04-09 | 深圳智芯微电子科技有限公司 | Federal learning method with robustness to wireless environment change in power Internet of things |
| CN112233118A (en) * | 2020-12-15 | 2021-01-15 | 南京可信区块链与算法经济研究院有限公司 | Eye fundus lesion image identification method and system based on incremental learning |
| CN112668044A (en) * | 2020-12-21 | 2021-04-16 | 中国科学院信息工程研究所 | Privacy protection method and device for federal learning |
| CN112754473A (en) * | 2021-01-08 | 2021-05-07 | 华中科技大学鄂州工业技术研究院 | Spontaneous heating downloading and method for falling detection by fusing transfer learning algorithm |
| CN112768056A (en) * | 2021-01-14 | 2021-05-07 | 新智数字科技有限公司 | Disease prediction model establishing method and device based on joint learning framework |
| CN112836829B (en) * | 2021-01-26 | 2022-08-02 | 支付宝(杭州)信息技术有限公司 | Personalized privacy protection learning method, device and equipment |
| CN112836829A (en) * | 2021-01-26 | 2021-05-25 | 支付宝(杭州)信息技术有限公司 | Personalized privacy protection learning method, device and equipment |
| CN112819177A (en) * | 2021-01-26 | 2021-05-18 | 支付宝(杭州)信息技术有限公司 | Personalized privacy protection learning method, device and equipment |
| CN112580085A (en) * | 2021-02-22 | 2021-03-30 | 支付宝(杭州)信息技术有限公司 | Model training method and device |
| CN113179244A (en) * | 2021-03-10 | 2021-07-27 | 上海大学 | Federal deep network behavior feature modeling method for industrial internet boundary safety |
| CN113077056A (en) * | 2021-03-29 | 2021-07-06 | 上海嗨普智能信息科技股份有限公司 | Data processing system based on horizontal federal learning |
| CN112926126A (en) * | 2021-03-31 | 2021-06-08 | 南京信息工程大学 | Federal learning method based on Markov random field |
| CN112926126B (en) * | 2021-03-31 | 2023-04-25 | 南京信息工程大学 | Federal learning method based on Markov random field |
| CN113052333A (en) * | 2021-04-02 | 2021-06-29 | 中国科学院计算技术研究所 | Method and system for data analysis based on federal learning |
| CN113128701A (en) * | 2021-04-07 | 2021-07-16 | 中国科学院计算技术研究所 | Sample sparsity-oriented federal learning method and system |
| WO2022236831A1 (en) * | 2021-05-14 | 2022-11-17 | 北京小米移动软件有限公司 | Model learning method, model learning apparatus, and storage medium |
| CN113902473A (en) * | 2021-09-29 | 2022-01-07 | 支付宝(杭州)信息技术有限公司 | Training method and device of business prediction system |
| CN117373066A (en) * | 2023-12-07 | 2024-01-09 | 华侨大学 | Pedestrian re-identification method and system based on Yun Bian searching federal deep learning method |
| CN117373066B (en) * | 2023-12-07 | 2024-03-12 | 华侨大学 | Pedestrian re-identification method and system based on Yun Bian searching federal deep learning method |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN111190487A (en) | Method for establishing data analysis model | |
| US11468337B2 (en) | System and methods for facilitating pattern recognition | |
| CN115345278A (en) | Structural learning of convolutional neural networks | |
| Usman et al. | A survey on big multimedia data processing and management in smart cities | |
| Hsieh et al. | A PSO-based rule extractor for medical diagnosis | |
| Cao et al. | A bat-inspired approach to define transition rules for a cellular automaton model used to simulate urban expansion | |
| US10949653B2 (en) | Intelligent persona generation | |
| Tyagi et al. | Recurrent neural networks: Concepts and applications | |
| CN112580902B (en) | Object data processing method and device, computer equipment and storage medium | |
| CN107392164A (en) | A kind of Expression analysis method based on the estimation of Facial action unit intensity | |
| Ramkumar et al. | Deep maxout network for lung cancer detection using optimization algorithm in smart Internet of Things | |
| Adeniyi et al. | Comparison of the performance of machine learning techniques in the prediction of employee | |
| Sawant et al. | Study of Data Mining Techniques used for Financial Data Analysis | |
| Wang et al. | Cornerstone network with feature extractor: a metric-based few-shot model for chinese natural sign language | |
| Rossi et al. | A multimodal deep learning network for group activity recognition | |
| Hoang et al. | A novel time series prediction approach based on a hybridization of least squares support vector regression and swarm intelligence | |
| Luo et al. | Forest fire detection using spiking neural networks | |
| CN113762648B (en) | Method, device, equipment and medium for predicting male Wei Heitian goose event | |
| Goyal et al. | Internet of things information analysis using fusion based learning with deep Neural Network | |
| Turner et al. | Optimizing artificial neural networks using LevyChaotic mapping on Wolf Pack optimization algorithm for detect driving sleepiness | |
| Adeniyi et al. | Comparative analysis of machine learning techniques for the prediction of employee performance | |
| Alam | Ai-fairness towards activity recognition of older adults | |
| Fan et al. | Aedmts: an attention-based encoder-decoder framework for multi-sensory time series analytic | |
| Zhu et al. | Mechanism of situation element acquisition based on deep auto-encoder network in wireless sensor networks | |
| Jafarigol | Uncovering the Potential of Federated Learning: Addressing Algorithmic and Data-driven Challenges under Privacy Restrictions |
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 | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20200522 |
|
| RJ01 | Rejection of invention patent application after publication |