Cross-scene cognitive ability evaluation method and system based on fine-grained migration

Abstract

The application provides a cross-scene cognitive ability evaluation method based on fine-grained migration, which comprises the following steps: taking cognitive data of a user in a first scene as source domain data and taking cognitive data of the user in a second scene as target domain data; taking a source domain data set as a training set, training a random forest classifier, and generating a source domain model; obtaining a test accuracy rate of the individual classifier on a target domain data set and an information gain difference from source domain features to target domain features; clustering all the individual classifiers into multiple clusters according to the test accuracy rate and the information gain difference; updating the individual classifiers in each cluster by using a corresponding growth mechanism to obtain a target domain model; and evaluating the cognitive ability of the user in the second scene through the target domain model. The application also provides a cross-scene cognitive ability evaluation system based on fine-grained migration and a data processing device.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of ubiquitous computing, and particularly relates to a cognitive function evaluation method and system based on fine-grained transfer learning and oriented to cross tasks. BACKGROUND

[0002] Cognitive health is the ability to think, learn and remember clearly, and is an important part of brain health. Cognitive health is related to the normal development of individuals throughout their life cycle, from the adolescent stage of continuous development of cognitive ability, to the adult stage of optimal cognitive ability, to the elderly stage of gradual degradation of cognitive ability. Having healthy cognitive ability is the basis for independent living, learning skills and adapting to society. In recent years, many factors related to lifestyle, environment, etc. such as lack of sleep, lack of physical exercise, smoking, alcoholism, excessive intake of high-fat or high-sodium food, population aging, environmental changes have made individual cognitive health an increasingly serious family and social crisis. Statistics from the Medical School of Heidelberg University in Germany show that the incidence of Attention Deficit Hyperactivity Disorder (ADHD) in the child population is about 5% to 7%, and about 51.1 million children worldwide are affected by ADHD. In addition, statistics from the World Health Organization (WHO) in 2019 show that there are about 50 million elderly people worldwide with cognitive ability decline symptoms, and the incidence of dementia in the elderly population over 60 years old is as high as 5% to 8%. At present, there is no effective treatment for cognitive health-related diseases such as attention deficit hyperactivity disorder and senile dementia. Training and rehabilitation at the early stage of the disease, alleviating specific symptoms, and delaying the development of the disease are the only feasible response at present.

[0003] Ubiquitous computing is a new computing model that adapts to the development of computing technology, emphasizing the integration of information, physical and social space, and breaking through the time and space limitations of traditional computing models through ubiquitous computing services. Unlike traditional cognitive health monitoring in a limited environment, cognitive health monitoring in a ubiquitous non-limited environment can achieve unrestricted, long-term, and cross-scene cognitive health monitoring through wearable devices, natural human-computer interaction, and adaptive computing. However, in different ubiquitous computing scenarios such as medical consultation, community life, and home care, due to the differences in data sensing devices and sensing processes, it often brings the problem of inconsistent data distribution in different scenarios, making it difficult for existing models to adapt across scenes. Transfer learning relaxes the requirement of data distribution consistency in traditional machine learning, and can simulate the human analogical reasoning process to apply the knowledge learned in known fields to the reasoning analysis of other related fields, providing a new way to solve the cross-scene applicability of cognitive ability evaluation in ubiquitous environments. For example, Long et al. combined feature transfer and deep neural networks to achieve cross-domain transfer of computer vision models. Although transfer learning has achieved excellent experimental results in the field of computer vision, it is limited by the strong interpretability requirements of medical consultations and the small sample of data. Therefore, there are still some challenges in the application of traditional transfer learning methods in the field of medical health. Compared with the commonly used data sets in the field of computer vision, the related research on cognitive ability evaluation is often limited by the small sample problem due to the limited number of patients, the difficulty of annotation, and the high cost of data collection. The related data sets are usually smaller in scale, usually containing only more than ten subjects. Existing research is often limited by the small sample problem, making it difficult to achieve high-precision cognitive ability evaluation model construction. Traditional transfer learning methods are mostly based on the design of deep neural networks, and in such a small sample set, it is difficult to use gradient descent algorithm to optimize the deep neural network model. In addition, in the process of constructing the cognitive evaluation model related to medical health, the interpretability of the model is also crucial, and the explainability is the basis for further exploring the health-related diagnosis and treatment indicators. SUMMARY

[0004] To solve the above problems, the application combines the random forest model with strong explainability with the idea of transfer learning to establish an effective adaptive cognitive evaluation model. The cross-scene cognitive ability evaluation method comprises the following steps: taking cognitive data of a user in a first scene as source domain data and taking cognitive data of the user in a second scene as target domain data; taking a source domain data set as a training set to train a random forest classifier and generate a source domain model; taking a target domain data set as a test set to test individual classifiers of the source domain model and obtain a test accuracy of the individual classifiers; extracting source domain features of the source domain data set and target domain features of the target domain data set to obtain an information gain difference from the source domain features to the target domain features; clustering all the individual classifiers into multiple clusters according to the test accuracy and the information gain difference; updating individual classifiers in each cluster using a corresponding growth mechanism to obtain a target domain model; and evaluating cognitive ability of the user in the second scene through the target domain model.

[0005] The cross-scene cognitive ability evaluation method comprises the following steps: taking cognitive data of a user in a first scene as source domain data and taking cognitive data of the user in a second scene as target domain data; taking a source domain data set as a training set to train a random forest classifier and generate a source domain model; taking a target domain data set as a test set to test individual classifiers of the source domain model and obtain a test accuracy of the individual classifiers; extracting source domain features of the source domain data set and target domain features of the target domain data set to obtain an information gain difference from the source domain features to the target domain features; clustering all the individual classifiers into multiple clusters according to the test accuracy and the information gain difference; updating individual classifiers in each cluster using a corresponding growth mechanism to obtain a target domain model; and evaluating cognitive ability of the user in the second scene through the target domain model.

[0006] The cross-scene cognitive ability evaluation method comprises the following steps: taking cognitive data of a user in a first scene as source domain data and taking cognitive data of the user in a second scene as target domain data; taking a source domain data set as a training set to train a random forest classifier and generate a source domain model; taking a target domain data set as a test set to test individual classifiers of the source domain model and obtain a test accuracy of the individual classifiers; extracting source domain features of the source domain data set and target domain features of the target domain data set to obtain an information gain difference from the source domain features to the target domain features; clustering all the individual classifiers into multiple clusters according to the test accuracy and the information gain difference; updating individual classifiers in each cluster using a corresponding growth mechanism to obtain a target domain model; and evaluating cognitive ability of the user in the second scene through the target domain model.

[0007] The cross-scene cognitive ability evaluation method comprises the following steps: taking cognitive data of a user in a first scene as source domain data and taking cognitive data of the user in a second scene as target domain data; taking a source domain data set as a training set to train a random forest classifier and generate a source domain model; taking a target domain data set as a test set to test individual classifiers of the source domain model and obtain a test accuracy of the individual classifiers; extracting source domain features of the source domain data set and target domain features of the target domain data set to obtain an information gain difference from the source domain features to the target domain features; clustering all the individual classifiers into multiple clusters according to the test accuracy and the information gain difference; updating individual classifiers in each cluster using a corresponding growth mechanism to obtain a target domain model; and evaluating cognitive ability of the user in the second scene through the target domain model.

[0008] The application further provides a cross-scene cognitive ability evaluation system based on fine-grained migration, comprising: a data acquisition module configured to acquire source domain data and target domain data; wherein cognitive data of a user in a first scene is the source domain data, and cognitive data of the user in a second scene is the target domain data; a model generation module configured to use a source domain data set as a training set to train a plurality of individual classifiers to generate an initial decision tree model; a clustering module configured to cluster the individual classifiers of the source domain model; wherein the individual classifiers are tested using a target domain data set as a test set to obtain a test accuracy of the individual classifiers; source domain features of the source domain data set and target domain features of the target domain data set are extracted to obtain an information gain difference from the source domain features to the target domain features; all the individual classifiers are clustered into a plurality of clusters according to the test accuracy and the information gain difference; an updating module configured to update the individual classifiers in each cluster using a corresponding growth mechanism to obtain a target domain model; and an evaluation module configured to evaluate the cognitive ability of the user in the second scene through the target domain model.

[0009] The cross-scene cognitive ability evaluation system comprises a clustering module, which comprises: obtaining feature set fitness and split threshold fitness of the individual classifiers using the test accuracy and the information gain difference; and performing a clustering operation on the individual classifiers by fusing prior knowledge using the feature set fitness and the split threshold fitness as prior knowledge.

[0010] The cross-scene cognitive ability evaluation system comprises a clustering module, which comprises: obtaining feature set fitness and split threshold fitness of the individual classifiers using the test accuracy and the information gain difference; and performing a clustering operation on the individual classifiers by fusing prior knowledge using the feature set fitness and the split threshold fitness as prior knowledge.

[0011] The cross-scene cognitive ability evaluation system comprises an updating module, wherein the growth mechanism comprises: an updating threshold mechanism configured to change a threshold of a split attribute of the individual classifier; and a structure adjustment mechanism configured to split or delete the individual classifier.

[0012] The application further provides a computer readable storage medium storing computer executable instructions, wherein when the computer executable instructions are executed, the cross-scene cognitive ability evaluation method based on fine-grained migration is implemented.

[0013] The application further provides a data processing device comprising the computer readable storage medium, wherein when a processor of the data processing device invokes and executes the computer executable instructions in the computer readable storage medium, the data processing device performs the cross-scene cognitive ability evaluation based on fine-grained migration.

[0014] The cross-scene cognitive ability evaluation method based on fine-grained migration of the application realizes effective migration between a source domain and a target domain, and can adapt to a small sample quantity and difficult labeling scene. BRIEF DESCRIPTION OF DRAWINGS

[0015] Figure 1 is a flow chart of the cross-scene cognitive ability evaluation method based on fine-grained migration of the application.

[0016] Figure 2 is a framework diagram of the parameter adaptive fine-grained migration learning method of the application.

[0017] Fig. 3 is a schematic diagram of ablation experiment results of the application.

[0018] Figure 4A , 4B is a schematic diagram of parameter sensitivity analysis experiment results of the application.

[0019] Figure 5A , 5B , 5C, 5D, 5E and 5F are schematic diagrams of time complexity analysis experiment results of the application.

[0020] Figure 6 is a schematic diagram of a data processing device of the application. DETAILED DESCRIPTION

[0021] In order to make the objects, technical solutions and advantages of the application clearer, the cross-scene cognitive ability evaluation method and system based on fine-grained migration proposed by the application are further described in detail below with reference to the drawings. It should be understood that the specific implementation methods described herein are only used to explain the application and do not limit the application.

[0022] To solve the small sample and labeling difficulty challenges in cognitive function evaluation, the application proposes a cross-scene migration learning method, namely a parameter adaptive fine-grained migration learning method (PAFG-TL), which can realize adaptive migration between a source domain model and a target domain model only by relying on a small number of training samples in the target domain. PAFG-TL is a general model migration learning framework suitable for ensemble learning, which is realized based on a random forest model and consists of two parts, namely a parameter adaptive individual classifier evaluation strategy (PAME) and a domain adaptive tree growing mechanism (DATG).

[0023] Wherein, PAME is based on decision tree feature set and threshold fitness, and clusters different individual classifiers in a clustering manner, specifically, PAME measures the splitting threshold and feature set of each individual classifier to measure the fitness of the source domain model in the target domain data. PAME uses the information gain difference of the source domain-target domain construction feature and the test accuracy of the target domain data to measure the feature set fitness and the splitting threshold fitness respectively, and uses a two-dimensional vector to describe each individual classifier, and then, through the Gaussian mixture clustering method of fusing prior knowledge, the clustering analysis is performed on different individual classifiers.

[0024] DATG realizes the fine-grained evolutionary growth of the decision tree through different strategies. After the clustering analysis of different individual classifiers is completed, DATG provides different growth mechanisms for the clustered individual classifiers. For example, updating the attribute threshold: updating the threshold of each splitting attribute from top to bottom, for the internal nodes that can be reached by the target domain data, this strategy can recalculate the threshold of the node according to the information gain; structure adjustment: including structure expansion and reduction, which can update the existing model by reducing useless internal nodes and splitting impure leaf nodes.

[0025] In other words, the present application provides a model migration method for cognitive function evaluation, which is a parameter adaptive fine-grained transfer learning method based on ensemble learning. The method of the present application first quantitatively evaluates each individual classifier in ensemble learning based on the information gain difference of the source domain-target domain construction feature and the test accuracy of the target domain data; then, according to the quantitative evaluation result, the Gaussian mixture clustering method of fusing prior knowledge is used to perform clustering analysis on different individual classifiers; finally, different adaptive growth strategies are adopted for the individual classifiers in different clusters to realize the migration and update between the source domain and the target domain model.

[0026] In order to evaluate the effectiveness of PAFG-TL, the present application has carried out experimental verification on both clinical cognitive evaluation and public benchmark data sets. The experimental results show that PAFG-TL can effectively improve the cross-scene cognitive ability evaluation effect.

[0027] I. First, the following definitions are made:

[0028] Transfer learning: Transfer learning uses the knowledge learned in the source domain to solve the data prediction problem of the target domain, usually, the source domain data is represented as n S =|D S | is the data size of the source domain; the target domain data is represented as n T =|D T | is the data size of the target domain. is the feature space, For category space, To predict the number of features in the problem, To predict the number of categories in the problem. Specifically, in this invention, DS can be the user's cognitive health data in the first scenario, D T It can be the user's cognitive health data in the second scenario, where the first scenario and the second scenario are different universal computing scenarios such as medical consultation, community life, and home monitoring, and the first scenario is different from the second scenario, but the present invention is not limited thereto.

[0029] Random Forest: Random forests aim to learn forests containing... An ensemble classifier with h decision trees Source domain data For the training dataset, any sample in the training dataset In the individual classifier h j The prediction result can be expressed as It is an individual classification h j In category The prediction results above During the construction of the j-th individual classifier, the random forest uses the bootstrap data sampling method from the training dataset D. S The filter contains n S Training subset of samples When selecting splitting attributes, random forests use feature randomization methods from... Select k features from the given features as candidate features, and then select the best feature from the candidate feature set based on segmentation metrics such as information gain. The final prediction result can be calculated using different model fusion strategies such as averaging, voting, and learning. Taking the voting method as an example, the final calculation result... Among them, w j This indicates the weight of the weighted vote.

[0030] II. The Parameter Adaptive Fine-Grained Transfer Learning Method of the Invention

[0031] Figure 1 This is a flowchart of the cross-scenario cognitive ability assessment method based on fine-grained transfer of the present invention. Figure 1 As shown, the cross-scenario cognitive ability assessment method of the present invention includes:

[0032] Step S1: Obtain the user's cognitive data in the first scenario as the source domain data; obtain the user's cognitive data in the second scenario as the target domain data.

[0033] Step S2: Using the source domain data as the training dataset, train the random forest model to obtain the source domain model;

[0034] Step S3, calculating the feature set fitness of the source domain model to the target domain sample;

[0035] Step S4, calculating the segmentation threshold fitness of the source domain model to the target domain sample;

[0036] Step S5, clustering different individual classifiers through a Gaussian mixture clustering method with prior knowledge fusion;

[0037] Step S6, obtaining C1, C2, C3 and C4 different clustering clusters;

[0038] For the individual classifier in the C1 cluster, the classifier is reconstructed using the target domain data;

[0039] For the individual classifier in the C2 cluster, the classifier is updated using an updated threshold strategy;

[0040] For the individual classifier in the C3 cluster, the classifier is updated using a structure adjustment strategy;

[0041] For the individual classifier in the C4 cluster, the individual classifier is kept unchanged;

[0042] Step S7, merging the individual classifiers in each cluster R=C1∪C2∪C3∪C4;

[0043] The parameter adaptive fine-grained transfer learning (PAFG-TL) method of the application is a model migration method, which realizes the adaptation of the model from the source domain to the target domain through a supervised manner using small-scale target domain data. Figure 2 As shown in the overall framework diagram of the PAFG-TL method, the PAFG-TL method is composed of a parameter adaptation based model evaluation strategy (PAME) and a domain adaptative tree growing mechanism (DATG), and is respectively used for evaluating the individual classifier and updating the individual classifier.

[0044] The PAFG-TL is realized based on a random forest, and is integrated by multiple decision trees. According to the error-ambiguity decomposition theory of Krogh et al., the generalization error of the ensemble learning model can be expressed as: wherein, Err、 and respectively represent the model error of ensemble learning, the average generalization error of individual classifier and the "divergence" degree of individual classifier, the overall performance of ensemble learning model is determined by both the accuracy and diversity of individual classifier, and the key to improve the performance of target domain model is to improve the adaptability of individual classifier in target domain and increase the diversity of individual classifier. Therefore, the basic idea of PAFG-TL method is to evaluate each individual classifier respectively, and to formulate different growth strategies according to the matching degree of existing source domain model and target domain data, so as to ensure the individualized evolution and update of individual classifier and improve the diversity of overall ensemble learning model.

[0045] 1. Parameter adaptive individual classifier evaluation strategy

[0046] The individual decision tree is in binary or multi-branch tree structure, and the decision path is determined by the attribute category and the segmentation threshold at each node. Therefore, the determination of the segmentation threshold and the selection of the feature set are the keys to measure the adaptability of the source domain model in the target domain data, and the information gain difference of the source domain-target domain constructed feature and the test accuracy of the target domain data are used to measure the adaptability of the feature set and the adaptability of the segmentation threshold respectively.

[0047] The information gain is a commonly used index to measure the quality of node segmentation in the decision tree: IG(D,a)=Ent(D)-Ent(D|a), wherein a is the segmentation attribute, Ent(D) is the information entropy on the current data set D, p l is the probability of the lth sample in the data set D, is the conditional probability when the data set D is divided into subsets using the attribute a. The information gain of all features can form the information gain set To evaluate the information gain value of each feature, the present application sorts , and uses to represent the sorting result of the information gain of each feature. For example, if is {2.3, 6.7, 4.5, 1.2}, then is {3, 1, 2, 4}. Assuming that A m is the feature set when the mth individual decision tree is constructed, IG(D S ,A m ) is the information gain set on the source domain data D S , IG(D T ,A m ) is the information gain set on the target domain set D T , IGR(D S ,A m ) and IGR(D T ,A m) are the ranking results of information gain respectively. To evaluate the fitness of feature set of each individual decision tree, the present application defines the following indexes:

[0048] (1) The final feature set fitness index F(h m )

[0049] In information theory, information gain is generally used to measure the quality of a partition F'(h m ):

[0050]

[0051] Wherein, |IGR(D S ,a)-IGR(D T ,a)| is used to measure the difference of the importance of attribute a in the source domain and the target domain, and +1 is used to avoid the case that the difference is zero when IGR(D S ,a) = IGR(D T ,a).

[0052] The present application uses |A m | to normalize the value of F'(h m ) to obtain F(h m ), so that the value range of F(h m ) is limited between [0, 1]:

[0053]

[0054] The closer the value of F(h m ) is to 1, the higher the fitness of the feature set of the source domain model h m to the target domain data is; otherwise, the closer the value is to 0, the lower the fitness of the feature set of the source domain model h m to the target domain data is.

[0055] (2) The test accuracy of the target domain data T(h m )

[0056] The present application uses the test accuracy of the target domain data to measure the fitness of the partition threshold, which is defined as:

[0057]

[0058] Wherein, is a symbol function, and |D T | is used to normalize the recognition accuracy of the individual decision tree h m between [0, 1]. The closer the value of T(h m ) is to 1, the higher the fitness of the source domain model h mThe higher the segmentation threshold fitness of the target domain data; otherwise, the value is closer to 0, indicating that the source domain model h m The lower the segmentation threshold fitness of the target domain data.

[0059] To realize Gaussian mixture clustering, a Gaussian mixture distribution containing mixed components is defined as: Where μ k and∑ k represent the parameters of the kth Gaussian distribution, α k represents the mixing coefficient and satisfies If the training set D = {x1, x2, …, x n} is generated by a Gaussian mixture distribution, let the random variable represent the Gaussian mixture component generating the sample x i , and the prior probability P(z i = k) of z i is α k . According to Bayes' theorem, the posterior distribution γ i of z ik is:

[0060]

[0061] In addition, when the Gaussian distribution is known, the Gaussian mixture clustering divides the sample set D into cluster clusters The cluster label λ i of the sample x i can be represented as:

[0062]

[0063] The Prior Knowledge based Gaussian Mixture Model (PGMM) process is similar to the standard Gaussian mixture clustering process, except that the PGMM method needs to pre-specify the distribution range of each cluster center, and continuously limit the cluster center distribution within the limited range in the subsequent process. The number of cluster centers is set to and the individual classifiers are divided into four cluster clusters The Gaussian mean distribution of each cluster cluster is defined in the range of μ k ∈R k , where In the PGMM clustering process, the present application will be The estimated prior of four clusters is incorporated into the clustering process. In the initial stage, the mean of the Gaussian distribution is initialized to be within the range of and further adjusted during the clustering process according to .

[0064] 2. Domain-adaptive decision tree growing mechanism

[0065] (1) The update threshold process uses the divergence index to fine-tune the feature threshold of the existing random forest model, which is defined as follows:

[0066]

[0067] 2JSD(P,Q)=KL(P||M)+KL(Q||M) (7)

[0068] where D v is the dataset of the reachable node v, a(v) is the split attribute of node v, and τ(v) is the existing split threshold of node v. Q L and Q R are the label distributions of dataset D v when the threshold is τ(v), Q' L and Q' R are the label distributions of dataset D v when the new threshold is used. M=(P+Q) / 2 is the mean of distributions P and Q, and the Jensen-Shanno divergence (JSD) is a symmetric extension of the Kullback-Leibler divergence (KL) that can measure the distance between two distributions and calculate the distance metric value. In particular, JSD(Q' L , Q L ) and JSD(Q' R , Q R ) represent the JSD distances of the left and right child nodes of node v when it uses the existing threshold τ(v) and the candidate threshold, respectively. Therefore, the divergence value can optimize the feature threshold of node v in consideration of the target training data.

[0069] (2) Structural adjustment includes both expansion and reduction of the structure, which can update the existing model by reducing useless internal nodes and splitting impure leaf nodes. Structural adjustment defines two loss metric indicators for internal node v, including subtree loss and leaf loss, which respectively evaluate the loss value of the subtree rooted at v and the loss value when node v is pruned as a leaf node. Nodes with a subtree loss greater than a leaf loss or data-unreachable nodes will be pruned.

[0070] 3. Parameter adaptive fine-grained transfer random forest

[0071] In the implementation process of the parameter adaptive fine-grained transfer random forest algorithm, the algorithm input is an existing ensemble learning model of a source domain The information gain set of the source domain And target domain data D T The output is an updated ensemble learning model. PAFG-TL first calculates the information gain set of the target domain And sorts And respectively, to obtain And Then, according to formula (2) and formula (3), calculate the two-dimensional set Used to describe each individual classifier from the feature set adaptability and the threshold adaptability; finally, through the Gaussian mixture clustering method of fusing prior knowledge, the individual classifiers are clustered and analyzed, and different updating strategies are used to incrementally update the individual classifiers in different sets.

[0072] III. Experimental evaluation

[0073] (1) Data set and preprocessing

[0074] Clinical cognitive assessment data set: The clinical cognitive assessment data set uses a touch screen-based cognitive assessment test to measure the cognitive state of the elderly, which consists of 4 different single tasks and 12 different double tasks. During the experiment, data collection was carried out in two scenarios: 1) Scenario one (labeled as TL): 61 subjects participated in data collection, including 20 patients with mild cognitive impairment (age: 68.25±6.15, 8 males and 12 females) and 41 healthy elderly people (age: 67.36±4.76, 21 males and 20 females). The test system runs on a Huawei M5 tablet computer (screen size 10.1 inches, resolution 1920x1200); during data collection, all subjects performed 4 assessment tasks, namely single task I, single task II, single task III and single task IV. 2) Scenario two (labeled as VS): 37 subjects participated in data collection, including 25 patients with mild cognitive impairment (age: 65.08±9.68, 16 males and 9 females) and 12 healthy people (age: 39.44±2.31, 7 males and 5 females). The test system runs on a NanoPi M4 single-chip computer (screen size 21.5 inches, resolution 1920x1200); during data collection, all subjects performed 4 assessment tasks, including single task II, double task AII, double task BII and double task CII. For the clinical cognitive assessment data set, the invention extracts five types of features, including number-based features Time-based features Speed-based features Angle-based features and drop distribution-based features

[0075] Disclosed benchmark datasets: Currently, there is a lack of benchmark transfer learning datasets for cognitive ability assessment. However, cognitive impairment is closely related to behavioral ability, and the studies by Roebers et al. of the University of Bern in Switzerland and Verlinden et al. of the Erasmus University Medical Center in Rotterdam, the Netherlands, reveal this intrinsic connection; various clinical commonly used cognitive ability assessment scales, such as the Edinburgh Motor Assessment (EMAS), the Scale for the Assessment and Rating of Cerebellar Ataxia (SARA), and the Unified Parkinson's Disease Rating Scale (UPDRS), also take behavioral ability as an important indicator for measuring cognitive impairment. Therefore, the present application selects four commonly used behavioral ability assessment datasets as disclosed benchmark datasets, and many transfer learning related studies also use related datasets. A brief introduction to the four disclosed benchmark datasets is shown in Table 1, including: 1) the DSADS dataset (labeled D), which collects 19 behaviors of 8 subjects through sensors worn at 5 positions; 2) the UCI-HAR dataset (labeled H), which collects 6 behaviors of 30 subjects through a sensor worn at the waist; 3) the USC-HAD dataset (labeled U), which collects 12 behaviors of 14 subjects through a sensor worn at the right front hip; and 4) the PAMAP dataset (labeled P), which collects 18 behaviors of 9 subjects through sensors worn at 3 positions. Since different datasets are collected from different sensors, worn at different positions, and contain different behaviors, in order to achieve transfer between different datasets, the present application performs uniformization processing on the initial dataset, and only uses data of the same sensor, the same behavior, and the same position for transfer. Specifically, the present application uses the acceleration and gyroscope data after three-axis synthesis as training data, and extracts 19 kinds of time domain and frequency domain features for each sensor, a total of 38 kinds; uses data of the right arm, waist, right front hip, and right wrist, and filters data of four behaviors of lying down, walking, climbing stairs, and descending stairs as the training dataset.

[0076]

[0077] Table 1

[0078] (2) Comparison methods and parameter details

[0079] To verify the experimental effect of the PAFG-TL method, the present application selects 10 existing methods as comparison methods, including:

[0080] a stratified transfer learning (STL) method;

[0081] a transfer component analysis (TCA) method;

[0082] a geodesic flow kernel (GFK) method;

[0083] a recognition model constructed only using target domain data (TarOnly, Tar);

[0084] a recognition model constructed only using source domain data (SrcOnly, Src);

[0085] a recognition model constructed using both source domain and target domain data (ComOnly, Com);

[0086] a structure expansion / reduction (SER) method;

[0087] a structure transfer (Struct) method;

[0088] a mix of SER and Struct (Mix) method;

[0089] a fine-grained adaptation random forest (FAT) method.

[0090] Wherein, Src, Tar and Com are batch learning methods; STL, TCA, GFK, SER, Struct, Mix and FAT are transfer learning methods. SER, Struct, Mix and FAT are model transfer methods based on random forest; TCA, STL and GFK are representative feature transfer methods, which can map source domain and target domain data to a unified space and reduce data distribution difference. The experimental environment is Dell Precision 5530 (Intel Core i9-8950HK / 32GB DDR3), and the integrated development environment is Matlab R2020a. Except for STL, Tar and GFK, all the comparative methods use random forest as the basic classifier. In the experimental process, the number of individual decision trees in the random forest is set to The candidate features on each split node are ( The minimum number of samples for stopping splitting of the node is 2, and the maximum depth of a single individual decision tree is maxD=10. In addition, STL, TCA and GFK, which are three attribute transfer-based learning methods, need to reduce the dimension, and the present application sets the reduced dimension to 30; the three parameters of FAT, δ1, δ2 and δ3, are set to 0.6, 0.7 and 0.8 respectively.

[0091] (3) Comparative experimental results

[0092]

[0093]

[0094] Table 2 Comparative experimental results

[0095] The comparative experiment results are shown in Table 2. Each migration experiment is repeated 10 times, and therefore, for the public benchmark dataset, the values in Table 2 are the average recognition accuracy of 10 experiments; in addition, for the clinical cognitive assessment dataset, since different scenarios (i.e., TL and VS) include four different test tasks, respectively, the migration processes TL→VS, VS→TL, VS→VS and TL→TL between two scenarios each include 16, 16, 12 and 12 different migration tasks, respectively, and the experiment results of the clinical cognitive assessment dataset are the average of the recognition accuracy of different migration tasks after averaging the results of 10 experiments of the same migration task. According to the experiment results shown in Table 2, the average accuracy of the PAFG-TL method proposed in the application is 79.4% on the clinical cognitive assessment dataset, which is better than the 10 comparative methods, and the accuracy is improved by 1.6% compared with the suboptimal method; for the migration results between different scenarios, PAFG-TL is better than the comparative methods in the four migration processes. On the public benchmark dataset, the average accuracy of the PAFG-TL method proposed in the application is 69.8%, which is better than the first nine comparative methods, but the accuracy is reduced by 0.7% compared with the FAT method; for the 12 different migration tasks, the PAFG-TL method achieves the best recognition accuracy in 4 migration tasks, and the FAT method achieves the best recognition accuracy in the remaining 8 migration tasks. Overall, on the public benchmark dataset, PAFG-TL is better than the first nine comparative methods, and worse than the FAT method proposed in the foregoing work, and a possible explanation is that PAFG-TL is a FAT parameter-independent improvement method, and it is difficult to achieve accuracy optimization through global parameter optimization.

[0096] (4) Ablation experiment

[0097] During the ablation experiment, the Gaussian mixture clustering method of the PAFG-TL method for fusing prior knowledge is replaced by the K-means clustering (K-means) and the traditional Gaussian mixture clustering (Gaussian Mixture Model, GMM) method one by one, and the recognition accuracy is calculated respectively. The ablation experiment results are shown in Figure 3A 、 3B Similarly, each different migration task is repeated 10 times, Figure 3A each column chart and error line in Table 2 is the average value and standard deviation of the experiment results of 56 different migration tasks; Figure 3B each column chart and error line in Table 2 is the average value and standard deviation of the experiment results of 12 different migration tasks. The experiment results on the clinical cognitive assessment dataset are shown in Figure 3AAs shown, the recognition accuracies of the three clustering methods, K-means, GMM, and PGMM, are 80.6%, 81.9%, and 82.9%, respectively, with standard deviations of 7.8%, 6.5%, and 6.8%. The PGMM method proposed in this invention outperforms the other two comparison methods, improving accuracy by 2.3% and 1.0%, respectively. Experimental results on publicly available benchmark datasets are shown below. Figure 3B As shown, the recognition accuracies of the three clustering methods, K-means, GMM, and PGMM, are 68.6%, 67.8%, and 70.2%, respectively, with standard deviations of 6.2%, 4.8%, and 6.0%. The PGMM method proposed in this invention is superior to the other two comparison methods, with accuracy improvements of 1.6% and 2.4%, respectively.

[0098] (5) Parameter sensitivity analysis

[0099] Since PAFG-TL is a parameter-independent transfer learning method, the number of decision trees in the random forest model... This is the only significantly influential variable in this method; therefore, this invention only analyzes the parameter. Impact on recognition accuracy. During the parameter sensitivity analysis experiment, the following parameters were set... Calculate in sequence The recognition accuracy under different values. Similarly, each different transfer task was repeated 10 times, and the experimental results are as follows. Figure 4A , 4B As shown. Figure 4A Each broken line and error band represents the mean and standard deviation of the experimental results for 56 different transfer tasks, respectively. Figure 4B Each broken line and error band represents the mean and standard deviation of the experimental results for 12 different transfer tasks, respectively. Experimental results on the clinical cognitive assessment dataset are shown below. Figure 4A As shown, from Figure 4A It can be seen that the recognition accuracy of PAFG-TL compared with the seven comparison methods is consistent with the parameter. Related, with With the improvement of parameters, the recognition accuracy gradually increases; compared with seven other methods, PAFG-TL shows that when parameters are... At the same time, better recognition accuracy can be achieved, possibly because... With improvements, PAFG-TL exhibits greater model diversity, thus achieving better recognition accuracy. Experimental results on public benchmark datasets are as follows: Figure 4B As shown, similarly, with parameters With the improvement of parameters, the recognition accuracy gradually increases; compared with the seven comparison methods, when the parameters are optimized... At the same time, PAFG-TL can achieve better recognition accuracy.

[0100] (6) Time Complexity Analysis

[0101] The present application analyzes the model training, updating and testing time of Tar, Src, Com, Ser, Struct, Mix, FAT and PAFG-TL methods on the clinical cognitive assessment dataset and the public benchmark dataset when the parameter takes different values, and the experimental results are shown in Table 1, Table 2, Table 3, Table 4, Table 5, Table 6, Table 7 and Table 8, wherein, Figure 5A-5F the training time of the clinical cognitive assessment dataset is shown in Table 1, Figure 5A the updating time of the clinical cognitive assessment dataset is shown in Table 2, Figure 5B the testing time of the clinical cognitive assessment dataset is shown in Table 3, Figure 5C the training time of the public benchmark dataset is shown in Table 4, Figure 5D the updating time of the public benchmark dataset is shown in Table 5, Figure 5E the testing time of the public benchmark dataset is shown in Table 6. It can be seen from Table 1 to Table 6 that the parameter Figure 5F is significantly related to the model training, updating and testing time, and the model training, updating and testing time is significantly improved with the increase of the parameter Figure 5A-5F . In terms of model training time, PAFG-TL is close to FAT method, the training time is longer than Src and Com methods, and shorter than Struct, Mix, Tar and Ser methods; in terms of model updating time, PAFG-TL is significantly higher than the other seven comparison methods, and one possible reason is that the clustering process of individual classifier model in PAFG-TL method is time-consuming; the model testing time result is close to the model training time result, and the testing time is longer than Src and Com methods, and shorter than Struct, Mix, Tar and Ser methods.

[0102] Figure 6 is a schematic diagram of the data processing device of the present application. As shown in Figure 6 ​​As shown, the embodiment of the present application further provides a computer readable storage medium and a data processing apparatus. The computer readable storage medium of the present application stores computer executable instructions, and the computer executable instructions are executed by a processor of the data processing apparatus to implement the above-mentioned cross-scene cognitive ability evaluation method based on fine-grained migration. Those skilled in the art can understand that all or part of the steps in the above-mentioned method can be completed by programs instructing related hardware (such as a processor, an FPGA, an ASIC, etc.), and the programs can be stored in a readable storage medium, such as a read-only memory, a magnetic disk or an optical disk, etc. All or part of the steps of the above-mentioned embodiments can also be implemented by using one or more integrated circuits. Accordingly, each module in the above-mentioned embodiments can be implemented in the form of hardware, such as by an integrated circuit to implement its corresponding function, or can be implemented in the form of a software function module, such as by a processor executing a program / instruction stored in a memory to implement its corresponding function. The embodiments of the present application are not limited to any specific form of combination of hardware and software.

[0103] The model migration learning method and system based on ensemble learning proposed by the present application is a migration learning method and system for cognitive function evaluation. The system mainly includes two modules of a parameter adaptive individual classifier evaluation strategy and a domain adaptive decision tree growth mechanism, wherein the parameter adaptive individual classifier evaluation strategy is used to evaluate the adaptation degree of each classifier of the source domain model to the target domain data. The strategy respectively evaluates each individual classifier from two aspects of feature set adaptability and segmentation attribute adaptability, establishes a two-dimensional measurement vector, and uses a Gaussian mixture clustering method with fused prior knowledge to perform clustering analysis on all individual classifiers. The domain adaptive decision tree growth mechanism defines different individual classifier model migration methods, including updating threshold, structure adjustment, etc., and different growth mechanisms are adapted to different individual classifier clusters, which can be used to realize adaptation of different granularities of source domain and target domain data.

[0104] The above embodiments are only used to illustrate the present application, and not to limit the present application. Those skilled in the art can make various changes and modifications without departing from the spirit and scope of the present application, and therefore all equivalent technical solutions also belong to the scope of the present application. The patent protection scope of the present application should be defined by the claims.

Claims

1. A method for assessing cross-scenario cognitive ability based on fine-grained transfer, characterized in that, include: The cognitive data of human users in the first scenario is used as the source domain data, and the cognitive data of the same human users in the second scenario is used as the target domain data. The first scenario and the second scenario are universal computing scenarios, including medical consultation, community life and home monitoring scenarios, and the first scenario is different from the second scenario. Using the source domain dataset as the training set, a random forest classifier is trained to generate the source domain model; Using the target domain dataset as the test set, the individual classifiers of the source domain model are tested to obtain the test accuracy of the individual classifiers. The source domain features of the source domain dataset and the target domain features of the target domain dataset are extracted to obtain the information gain difference from the source domain features to the target domain features. Based on the test accuracy and the information gain difference, the feature set fitness and the segmentation threshold fitness of the individual classifiers are obtained. Using the feature set fitness and the segmentation threshold fitness as prior knowledge, a Gaussian mixture clustering operation that integrates prior knowledge is performed on the individual classifiers to cluster all the individual classifiers into multiple clusters. The individual classifiers in each cluster are updated using the corresponding growth mechanism to obtain the target domain model; The target domain model is used to assess the cognitive abilities of the human user in this second scenario.

2. The cross-scenario cognitive ability assessment method as described in claim 1, characterized in that, This growth mechanism includes: The parameter adjustment mechanism changes the threshold of the segmentation attribute of the classifier for this individual. The structural adjustment mechanism splits or deletes the individual classifier.

3. A cross-scenario cognitive ability assessment system based on fine-grained transfer, characterized in that, include: The data acquisition module is used to acquire source domain data and target domain data; wherein, the cognitive data of the human user in the first scenario is used as the source domain data, and the cognitive data of the human user in the second scenario is used as the target domain data; the first scenario and the second scenario are ubiquitous computing scenarios, including medical consultation, community life and home monitoring scenarios, and the first scenario is different from the second scenario; The model generation module is used to train a random forest classifier using the source domain dataset as the training set to generate source domain models. The clustering module is used to cluster individual classifiers of the source domain model. Specifically, the individual classifiers are tested using the target domain dataset as the test set to obtain their test accuracy. Source domain features from the source domain dataset and target domain features from the target domain dataset are extracted to obtain the information gain difference from the source domain features to the target domain features. Based on the test accuracy and the information gain difference, the feature set fitness and segmentation threshold fitness of the individual classifiers are obtained. Using the feature set fitness and segmentation threshold fitness as prior knowledge, a Gaussian mixture clustering operation is performed on the individual classifiers, fusing the prior knowledge, to cluster all individual classifiers into multiple clusters. The update module is used to update the individual classifiers in each cluster using the corresponding growth mechanism to obtain the target domain model; The evaluation module is used to assess the human user's cognitive abilities in the second scenario using the target domain model.

4. The cross-scenario cognitive ability assessment system as described in claim 3, characterized in that, The growth mechanism in this update module includes: Update threshold mechanism: Change the threshold of the classifier's segmentation attribute for this individual; Structural adjustment mechanism: split or delete the individual classifier.

5. A computer-readable storage medium storing computer-executable instructions, characterized in that, When the computer-executable instructions are executed, the cross-scenario cognitive ability assessment method based on fine-grained transfer as described in claim 1 or 2 is implemented.

6. A data processing apparatus comprising a computer-readable storage medium as claimed in claim 5, wherein when a processor of the data processing apparatus retrieves and executes computer-executable instructions in the computer-readable storage medium, the data processing apparatus performs a cross-scenario cognitive ability assessment based on fine-grained transfer.

Images