CN112699402B - Wearable device activity prediction method based on federal personalized random forest - Google Patents
Wearable device activity prediction method based on federal personalized random forest Download PDFInfo
- Publication number
- CN112699402B CN112699402B CN202011577207.7A CN202011577207A CN112699402B CN 112699402 B CN112699402 B CN 112699402B CN 202011577207 A CN202011577207 A CN 202011577207A CN 112699402 B CN112699402 B CN 112699402B
- Authority
- CN
- China
- Prior art keywords
- attribute
- current
- party
- participant
- random
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
- 238000007637 random forest analysis Methods 0.000 title claims abstract description 41
- 230000000694 effects Effects 0.000 title claims abstract description 26
- 238000000034 method Methods 0.000 title claims abstract description 26
- 238000003066 decision tree Methods 0.000 claims abstract description 38
- 238000012549 training Methods 0.000 claims abstract description 32
- 230000008569 process Effects 0.000 claims abstract description 11
- 238000004891 communication Methods 0.000 claims description 4
- 230000007246 mechanism Effects 0.000 abstract description 3
- 230000004931 aggregating effect Effects 0.000 abstract description 2
- 230000006870 function Effects 0.000 description 10
- 238000013473 artificial intelligence Methods 0.000 description 4
- 238000005192 partition Methods 0.000 description 4
- 238000004422 calculation algorithm Methods 0.000 description 3
- 230000036541 health Effects 0.000 description 3
- 239000004984 smart glass Substances 0.000 description 3
- 238000013459 approach Methods 0.000 description 2
- 239000002131 composite material Substances 0.000 description 2
- 238000010586 diagram Methods 0.000 description 2
- 238000005516 engineering process Methods 0.000 description 2
- 238000010801 machine learning Methods 0.000 description 2
- 238000007781 pre-processing Methods 0.000 description 2
- 238000012545 processing Methods 0.000 description 2
- 238000000638 solvent extraction Methods 0.000 description 2
- 238000012360 testing method Methods 0.000 description 2
- 238000004458 analytical method Methods 0.000 description 1
- 238000004364 calculation method Methods 0.000 description 1
- 238000010276 construction Methods 0.000 description 1
- 238000007418 data mining Methods 0.000 description 1
- 238000013461 design Methods 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 230000003862 health status Effects 0.000 description 1
- 238000013507 mapping Methods 0.000 description 1
- 238000005259 measurement Methods 0.000 description 1
- 230000035945 sensitivity Effects 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F21/00—Security arrangements for protecting computers, components thereof, programs or data against unauthorised activity
- G06F21/60—Protecting data
- G06F21/62—Protecting access to data via a platform, e.g. using keys or access control rules
- G06F21/6218—Protecting access to data via a platform, e.g. using keys or access control rules to a system of files or objects, e.g. local or distributed file system or database
- G06F21/6245—Protecting personal data, e.g. for financial or medical purposes
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/243—Classification techniques relating to the number of classes
- G06F18/24323—Tree-organised classifiers
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N20/00—Machine learning
- G06N20/20—Ensemble learning
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- Software Systems (AREA)
- General Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Evolutionary Computation (AREA)
- General Health & Medical Sciences (AREA)
- Bioethics (AREA)
- Health & Medical Sciences (AREA)
- Artificial Intelligence (AREA)
- Medical Informatics (AREA)
- Mathematical Physics (AREA)
- Computing Systems (AREA)
- Life Sciences & Earth Sciences (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Evolutionary Biology (AREA)
- Databases & Information Systems (AREA)
- Computer Hardware Design (AREA)
- Computer Security & Cryptography (AREA)
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
Abstract
The invention discloses a wearable device activity prediction method based on federal personalized random forests, wherein each user uses a local sensitive hash function to generate a respective hash table, and each user finds out own similar user by comparing the quantity of similar sample data among all hash tables. Each user can actively initiate training, and only users similar to the user can be searched for to jointly train a random decision tree which is commonly owned by the users participating in the training. And in the process of dividing the nodes of the random decision tree, protecting the privacy of the users by using a differential privacy mechanism, disturbing the calculated information gain by each party, and aggregating the disturbed information gain of each party at one party to obtain the optimal division of the candidate attribute set. And for the generated random decision tree, the user obtains the final random forest model of the user through incremental selection, the random forest model of each user is different, and the personalized model is more suitable for the user to predict on own data.
Description
Technical Field
The invention relates to the technical field of federal learning, in particular to a wearable device activity prediction method based on federal personalized random forests.
Background
Activities of daily life are closely related to the health of people. In recent years, with the development of wearable technologies, people track their activities by using wearable devices such as smartphones, bracelets, smart glasses, and the like, so as to know their health conditions. Artificial intelligence technology will also be integrated into wearable devices, enabling a more intelligent, more self-conscious application modality, a trend that has been the first time at the CES2014 electronic consumption exhibition. For example, according to a series of data such as activity data generated by a user and a surrounding environment collected by the wearable device, the activity of the user in progress is accurately identified by using a trained machine learning model, so that the intelligent wearable device can provide a dedicated and personalized service for each user. Because artificial intelligence requires a large amount of data for machine learning to achieve better performance, applying artificial intelligence to wearable devices still faces two major challenges: firstly, under the scene that the personal wearable equipment has a large number of users and needs model individuation, algorithm designs are different; and secondly, data exists in an isolated island form, if the data of the user is collected for centralized training, personal information is speculated through analysis of the data, and therefore personal sensitive information is leaked.
Disclosure of Invention
The invention aims to solve the problem that personal sensitive information is easily leaked in the process of realizing artificial intelligence of the conventional wearable equipment, and provides a wearable equipment activity prediction method based on federal personalized random forests.
In order to solve the problems, the invention is realized by the following technical scheme:
the wearable device activity prediction method based on the federal personalized random forest comprises the following steps:
step 1, calculating hash values of all sample data of wearable equipment of each participant by using local sensitive hashes, sending a hash table formed by the hash values of the sample data to other participants by each participant, and finding similar sample data among different participants by the other participants according to a global hash table;
step 2, determining a similar participant set of each current participant, namely: the current participant respectively counts the quantity of similar sample data of other participants and the current participant, calculates the proportion of the quantity of the similar sample data to the total quantity of all the sample data of the other participants, and arranges other participants with the proportion at the front k bits to form a similar participant set of the current participant; wherein k is a set value;
step 3, randomly selecting part of similar participants from the similar participant set of each current participant, wherein the current participant serves as a current active party, and the selected similar participants serve as current passive parties;
step 4, under the coordination of the current active party, using sample data of the current active party and each current passive party to jointly train the optimal attribute and the attribute division value of each node of the random decision tree, and after the training is finished, the current active party and each current passive party respectively obtain a same random decision tree;
step 5, the current active party and each current passive party add the random decision tree obtained in the step 4 into respective federal random forest models in an increment selection mode;
step 6, repeating the steps 3-5 until the training times of each current participant reach the preset training times, and finally, enabling each participant to have different federal random forest models respectively;
and 7, inputting the measured data of the wearable equipment of the participant into the federal random forest model obtained in the step 6, so as to complete the activity prediction of the participant, wherein the prediction result is an activity label of the human activity recognition task.
In the step 5, the process of jointly training the optimal attribute and the attribute division value of each node of the random decision tree by using the sample data of the current active party and each current passive party is as follows:
step 5.1, the current active party randomly selects a part of attributes from the attribute set to form a candidate attribute set, and sends the candidate attribute set to each current passive party;
step 5.2, the current active party and each current passive party respectively randomly select a value from the range of the local sample data attribute value of each attribute of the candidate attribute set as an attribute random value; each current passive party sends the attribute random values of all the attributes of the candidate attribute set to the current active party;
step 5.3, the current active side randomly selects a value from the range of the attribute random value of each attribute of the candidate attribute set of the current active side and each current passive side as an attribute division value, and sends the attribute division values of all the attributes of the candidate attribute set to each current passive side;
step 5.4, the current active side and each current passive side pre-divide the local sample data thereof by utilizing the candidate attribute set and the corresponding attribute dividing value respectively, and calculate the information gain of each attribute in the pre-divided candidate attribute set;
step 5.5, adding Laplace noise into the information gain of each attribute in the candidate attribute set by each current passive party for disturbance to obtain the disturbance information gain of each attribute in the candidate attribute set, and sending the disturbance information gain of all attributes in the candidate attribute set to the current active party;
step 5.6, the active party carries out weighted summation on the information gain of the current active party and the disturbance information gain of each current passive party of each attribute of the candidate attribute set to obtain the comprehensive score of each attribute of the candidate attribute set, and selects the attribute with the maximum comprehensive score as the optimal attribute;
and 5.7, the current active party sends the optimal attribute and the attribute division value corresponding to the optimal attribute to each current passive party, wherein the optimal attribute and the attribute division value corresponding to the optimal attribute are the optimal attribute and the attribute division value of the current node of the trained random decision tree.
In the step 1, each participant sends the hash table formed by the hash values of the sample data to each participant through the distributed communication framework.
Compared with the prior art, the invention has the following characteristics:
1. and applying a locality sensitive hash function in the federal random forest, calculating similar sample data among users, and enabling the users to find similar users. And each user only selects similar users to perform combined modeling, so that the limitation of the random forest in a federal scene with larger user quantity and needing an individualized model is solved.
2. In the node division process, in order to find the optimal attribute division in the candidate division attribute set, the users participating in training perform noise disturbance on the information gain of the attribute division in the candidate attribute set by using a Laplacian mechanism. When deciding on node partitioning, the privacy of the user data is protected.
3. In the federal learning, an integrated learning increment selection idea is utilized, each user screens a random decision tree in a local model, the model precision is improved, the model complexity is simplified, and the storage and prediction expenses are reduced.
Drawings
FIG. 1 is a flow diagram of a data processing phase of a federated random forest approach to differential privacy protection.
FIG. 2 is a flow diagram of a training phase of a federated random forest approach to differential privacy protection.
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 below with reference to specific examples.
To accommodate the large number of users, the model requires a personalized scenario, treating each user as a compute node, and assuming each user is honest and curious. And each user uses a locality sensitive hash function to generate a respective hash table, and each user finds out the user similar to the user by comparing the quantity of similar sample data among all the hash tables. Each user can actively initiate training, and only users similar to the user can be searched for to jointly train a random decision tree which is commonly owned by the users participating in the training. And in the process of dividing the nodes of the random decision tree, protecting the privacy of the user by using a differential privacy mechanism, disturbing the calculated information gain by each party, and aggregating the disturbed information gain of each party at one party to obtain the optimal division of the candidate attribute set. For the generated random decision tree, a user obtains a final random forest model through increment selection, the random forest model of each user is different, the personalized model is more suitable for the user to predict on data of the user, and the user tracks the activity of the user by using wearable equipment such as a smart phone, a bracelet and smart glasses, so that the health condition of the user can be known.
The wearable device activity prediction method based on the federal personalized random forest specifically comprises the following steps:
first, a data processing stage (as shown in fig. 1).
Step 1, the methodComputing each participant (i.e., user of wearable device) p with Locality-Sensitive Hashing (LSH)oAnd e, the hash value of all sample data in P, wherein P is all users in the federal environment. And forming a global hash table by the sample data hash values of all parties through Allreduce operation of a distributed communication framework, and sending the global hash table to all the participating parties, wherein all the participating parties find similar sample data among different participating parties according to the received hash table.
In order to obtain the similarity of any two sample data in the combined data without exposing the original data to other parties, a widely used p-stable LSH function is adopted. According to LSH, if two sample data are similar, they have a higher probability of being hashed to the same value. Thus, by applying multiple LSH functions, sample data is mapped to multiple hash values. Hash function Fa,bExpressed as:
wherein v represents a d-dimensional sample data vector, a is a d-dimensional random vector, b ∈ (0, r) is a random number, r is the segment length of a straight line, and a straight line in the space is divided into a plurality of segments with the length r and equal length. Points in the space are mapped to a straight line through a mapping function, and points mapped to the same segment are endowed with the same hash value. It is understood that if two points in the space are closer to each other, the higher the probability that they are mapped to the same segment, and the functions of the hash function family are established according to the difference between a and b.
Given L randomly generated p-stable hash functions, each party first computes the hash value of its sample data. We use the distributed communication framework Allreduce operation to build L global hash tables (H)1,H2...HL) The input of AllReduce is a hash table formed by instance id of each participant and hash value thereofPropagating the aggregated hash table to each party to compute similarityAnd (4) information.
Each party gets a global hash table. For one participant poE.g. P, first at the participant PgFinding a participant P in an element PoSimilar sample data of (2). Comparison of poEach sample data with pgAnd if the number of the same hash value of each sample data is greater than the threshold value t, the two sample data are considered to be similar.
Step 2, each participant poDetermine its set of similar participants Ps。
For one participant poE.g. P. First, find out the current participant poAt the other participant pgTaking the similar sample data, merging the similar sample data, and counting other participators pgWith the current participant poThe number of similar sample data. Then, calculating the occupation of similar sample data in other participants pgThe ratio of the similar sample data ratios (i.e., the ratio of the similar sample data to the number of all sample data) for all sample data. Finally, the other participants p of the similar sample data ratio top-k are scaledg(i.e. the participator with similar sample data in the top k bits in proportion) as the current participator poOf similar participants Ps。
In practice, steps 1-2 are pre-processing procedures that can be done once and reused in multiple trainings. Only when there is an update in the training data, the preprocessing needs to be performed again.
And II, a joint training stage (shown in figure 2).
Step 3, for the participant pmE.g. P, each training from its set of similar participants PsPart of participants selected at randomUsed for jointly training a random decision tree. At this time, pmThe random forest model is regarded as the active part of the training, the random forest model is in an active training state in the training process of the current random decision tree, and the similar part P of the random forest model*Viewed as aAnd the random forest model of the passive side is in a passive training state in the training process of the current random decision tree.
Step 4, initiative party pmCoordinating passive parties P*Using the master pmAnd each passive side P*The sample data is used for jointly training the random decision tree to carry out the optimal attribute and the attribute division value of each node, the construction of the random decision tree is jointly completed, and at the moment, the initiative pmCoordinating all passive parties P*And simultaneously has an identical random decision tree.
Decision trees are among the most efficient and widely used techniques in many data mining algorithms. The decision tree is a tree composed of a plurality of judgment nodes, and samples are classified through judgment of node attributes and attribute values of each layer. The partition attributes and attribute values on each node are selected based on a goal: the 'purity' of the data of each divided node is the highest, that is, samples contained in the sub-nodes divided by the attribute and the attribute value belong to the same category as much as possible. The random decision tree is formed by adding randomness on the basis of the decision tree so as to improve the generalization capability and the training speed of the decision tree. The random forest model consisting of a plurality of random decision trees can accommodate the error of a single random tree, thereby achieving better prediction effect.
The optimal attribute and attribute partition value of each node of the random decision tree are realized in a top-down recursion mode, information gain of each attribute in the candidate attribute set is calculated for the nodes from a root node, and the attribute value with the maximum information gain are selected as the optimal partition of the nodes. And constructing child nodes according to the division of the optimal attributes and the attribute values, and recursively calling the method by the child nodes to construct a random decision tree. The process of calculating the optimal attribute and the attribute dividing value of each node of the random decision tree is as follows:
step 4.1, initiative side pmRandomly selecting partial attributes from a known attribute set FForming a candidate attribute set, and setting a candidate attribute set F*Is sent toEach passive side P*。
Step 4.2, Passive side P*First determining the property fjMaximum value in its local sample dataAnd minimum valueThen fromRandomly selects a value in the range ofAs attribute fj∈F*A random value of (a); equally active side pmFirst determining the property fjMaximum value in its local sample dataAnd minimum valueThen fromRandomly selects a value vm,j As attribute fj∈F*Is calculated. All passive parties P*Set F of candidate attributes*Each attribute f inj∈F*Corresponding random valueSent to the master pm。
Step 4.3, initiative side pmReceiving a passive party P*Transmitted random valueAnd from the random value vm,jAnd all random valuesMaximum value ofMinimum value ofRandomly selecting a value within the rangeAs attribute fj∈F*The value of (a) is divided. Initiative side pmSet F of candidate attributes*Each attribute f in (1)j∈F*Corresponding division value vj *To the passive party P*。
Step 4.4, initiative side pmAnd each passive side P*Pre-dividing local sample data by using the candidate attribute set and the corresponding attribute dividing value respectively, and calculating the information gain q of each attribute in the pre-divided candidate attribute setIG。
Initiative side pmAnd a passive side P*According to the attribute fjAnd attribute valueDividing sample data in the nodes, and calculating divided information gain:
wherein, TiRepresenting a participant piAnd (3) currently dividing sample data of the node, wherein A represents a divided child node set, and C is a sample data label type set.Representing an attribute fjThe number of sample data of the divided child node a,the number of the sample data labels in the divided child node a is c.
Step 4.5, each passive side P*Information gain for each attribute in pre-partitioned candidate attribute setAdding Laplace noise to carry out disturbance to obtain disturbance information gain of each attribute in the candidate attribute setEach passive side P*Gain disturbance information of all attributes in candidate attribute setSent to the master pm。
Passive square P*And adding Laplace noise to the information gain for disturbance. Because the information gain is calculated directly on the data, these potential holes are exactly what the differential privacy decision tree algorithm is to prevent. As the depth of the tree increases, more privacy budgets are allocated to each layer, and the effectiveness of the model is guaranteed while privacy is protected. The sensitivity of the information gain calculation function is:
S(qIG)=log(N+1)+1/ln2
wherein, N is the size of the data set, epsilon is the privacy budget allocated to each tree by the passive party, epsilon controls the degree of privacy protection, and the smaller the privacy budget, the higher the degree of privacy protection. Different budgets are allocated for different depths of the tree, when the tree depth is large, the true count is more easily overcome by noise due to further partitioning. Thus, deeper nodes are allocated a larger budget to hold the real information. k is initialized to the height of the tree, StRepresents the total share of the privacy budget:
where ε' represents the privacy budget allocation for each level of the random tree,adding noise Lap (S (q) to the information gain divided by each attribute in the candidate attribute setIG)/(ε'/|F*|)) the information gain after adding noise disturbance is expressed as:
step 4.6, for each attribute f in the candidate attribute setj∈F*Master side pmInformation gain to the master respectivelyAnd each passive side P*Gain of disturbance informationCarrying out weighted summation to obtain an attribute fjIs given a composite score
Wherein N isiIs the number of sample data of the passive side, NmIs the sample data number of the master.
Step 4.7, initiative party pmSelecting the attribute with the highest composite scoreAs the best attribute, and sends the best attribute and its corresponding attribute score value (obtained in step 4.3) to each passive party P*. The optimal attribute and the attribute partition value corresponding to the optimal attribute areAnd (4) the best attribute and the attribute division value of the current node of the trained random decision tree.
Step 5, initiative party pmAnd each passive side P*And (4) adding the random decision trees x constructed in the step (4) into respective random forest models in an incremental selection mode.
Active side pmAnd each passive side P*When the random trees are selected in an increment mode, only the random decision trees x which contribute to the accuracy of the respective existing random forest models are selected and added into the existing random forest model models. The method contributes to the precision of the random forest model and meets the following requirements: e (X (T)) is less than or equal to E (X (T)) + X (T)), X represents the existing random forest model of the user, T represents the test data of the user, namely the test precision E of the random forest is improved after a new random tree X is added.
Step 6, repeating the steps 3-5, wherein in the process of each repetition, the master party pmAnd each passive side P*The various random forest model models are updated once (whether random decision tree x is added or not is considered as a model update). In the training phase, two training states exist for the random forest model of each participant: active and passive. When the participator is used as an active party, the random forest model is in an active training state; when the participator is used as a passive party, the random forest model is in a passive training state. And after all the participants are used as the active parties and respectively finish the training and model updating for a preset number of times, showing that all the participants finish the federal learning and finish the training. At this time, each user poBelonging to P and having different federal random forest models Xo. Because each user is trained with similar users in a combined mode, and the incremental selection process is added, the user model is personalized, and the method is more suitable for local data of the user.
And thirdly, local actual measurement stage.
Step 7, in the scene of wearable personal equipment, each user continuously collects data D generated by the individual by utilizing own equipmentoUsing the trained federal random forest model XoAnd performing activity prediction on the collected new data. When making a decision, the user acquires all the individual piecesAnd (4) predicting results of the decision tree, and then according to the votes of the results for each activity category, winning the category with the most votes, namely the activity label identified by the personalized random forest model in the human activity identification task. The user uses wearable equipment such as smart mobile phone, bracelet, smart glasses to track own activity to know the health status of oneself.
It should be noted that, although the above-mentioned embodiments of the present invention are illustrative, the present invention is not limited thereto, and thus the present invention is not limited to the above-mentioned embodiments. Other embodiments, which can be made by those skilled in the art in light of the teachings of the present invention, are considered to be within the scope of the present invention without departing from its principles.
Claims (3)
1. The wearable equipment activity prediction method based on the federal personalized random forest is characterized by comprising the following steps of:
step 1, calculating hash values of all sample data of wearable equipment of each participant by using local sensitive hashes, wherein each participant sends a hash table formed by the hash values of the sample data to other participants, and the other participants find similar sample data among different participants according to a global hash table;
step 2, determining a similar participant set of each current participant, namely: the current participant respectively counts the number of similar sample data of other participants and the current participant, calculates the proportion of the number of the similar sample data to the total amount of all the sample data of the other participants, and arranges other participants with the proportion in the front k bits to form a similar participant set of the current participant; wherein k is a set value;
step 3, each current participant randomly selects part of similar participants from the similar participant set, wherein the current participant serves as a current active party, and the selected similar participants serve as current passive parties;
step 4, under the coordination of the current active party, using sample data of the current active party and each current passive party to jointly train the optimal attribute and the attribute division value of each node of the random decision tree, and after the training is finished, the current active party and each current passive party respectively obtain a same random decision tree;
step 5, the current active party and each current passive party add the random decision tree obtained in the step 4 into respective federal random forest models in an increment selection mode;
step 6, repeating the steps 3-5 until the training times of each current participant reach the preset training times, and finally enabling each participant to have different federal random forest models respectively;
and 7, inputting the measured data of the wearable equipment of the participant into the federal random forest model obtained in the step 6, so as to complete the activity prediction of the participant, wherein the prediction result is an activity label of the human activity recognition task.
2. The method for predicting the activity of wearable equipment based on the federal personalized random forest as claimed in claim 1, wherein in the step 5, the optimal attribute and the attribute dividing value of each node of the random decision tree are jointly trained by using sample data of the current active party and each current passive party by the following process:
step 5.1, the current active party randomly selects a part of attributes from the attribute set to form a candidate attribute set, and sends the candidate attribute set to each current passive party;
step 5.2, the current active party and each current passive party respectively randomly select a value from the range of the local sample data attribute value of each attribute of the candidate attribute set as an attribute random value; each current passive party sends the attribute random values of all the attributes of the candidate attribute set to the current active party;
step 5.3, the current active party randomly selects a value from the range of the attribute random value of each attribute of the candidate attribute set of the current active party and each current passive party as an attribute division value, and sends the attribute division values of all the attributes of the candidate attribute set to each current passive party;
step 5.4, the current active party and each current passive party pre-divide the local sample data thereof by using the candidate attribute set and the corresponding attribute dividing value respectively, and calculate the information gain of each attribute in the pre-divided candidate attribute set;
step 5.5, adding Laplace noise into the information gain of each attribute in the candidate attribute set by each current passive party for disturbance to obtain the disturbance information gain of each attribute in the candidate attribute set, and sending the disturbance information gain of all attributes in the candidate attribute set to the current active party;
step 5.6, the active party carries out weighted summation on the information gain of the current active party and the disturbance information gain of each current passive party of each attribute of the candidate attribute set to obtain the comprehensive score of each attribute of the candidate attribute set, and selects the attribute with the maximum comprehensive score as the optimal attribute;
and 5.7, the current active party sends the optimal attribute and the attribute division value corresponding to the optimal attribute to each current passive party, wherein the optimal attribute and the attribute division value corresponding to the optimal attribute are the optimal attribute and the attribute division value of the current node of the trained random decision tree.
3. The method of claim 1, wherein in step 1, each participant sends a hash table comprising hash values of sample data to each participant through a distributed communication framework.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202011577207.7A CN112699402B (en) | 2020-12-28 | 2020-12-28 | Wearable device activity prediction method based on federal personalized random forest |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202011577207.7A CN112699402B (en) | 2020-12-28 | 2020-12-28 | Wearable device activity prediction method based on federal personalized random forest |
Publications (2)
Publication Number | Publication Date |
---|---|
CN112699402A CN112699402A (en) | 2021-04-23 |
CN112699402B true CN112699402B (en) | 2022-06-17 |
Family
ID=75512510
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN202011577207.7A Active CN112699402B (en) | 2020-12-28 | 2020-12-28 | Wearable device activity prediction method based on federal personalized random forest |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN112699402B (en) |
Families Citing this family (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN114118601B (en) * | 2021-12-02 | 2024-02-13 | 安徽大学 | Random forest traffic prediction method based on differential privacy protection |
CN114530028B (en) * | 2022-02-14 | 2023-04-25 | 大连理工大学 | Campus student intelligent bracelet monitoring system and method based on LoRa communication and federal learning |
Citations (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN105447525A (en) * | 2015-12-15 | 2016-03-30 | 中国科学院软件研究所 | Data prediction classification method and device |
CN108108657A (en) * | 2017-11-16 | 2018-06-01 | 浙江工业大学 | A kind of amendment local sensitivity Hash vehicle retrieval method based on multitask deep learning |
CN109284626A (en) * | 2018-09-07 | 2019-01-29 | 中南大学 | Random forests algorithm towards difference secret protection |
CN111598186A (en) * | 2020-06-05 | 2020-08-28 | 腾讯科技(深圳)有限公司 | Decision model training method, prediction method and device based on longitudinal federal learning |
CN111742313A (en) * | 2018-02-14 | 2020-10-02 | 万思伴股份有限公司 | System, apparatus and method for privacy preserving context authentication |
Family Cites Families (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US10095552B2 (en) * | 2016-02-05 | 2018-10-09 | Sas Institute Inc. | Automated transfer of objects among federated areas |
US10848519B2 (en) * | 2017-10-12 | 2020-11-24 | Charles River Analytics, Inc. | Cyber vaccine and predictive-malware-defense methods and systems |
-
2020
- 2020-12-28 CN CN202011577207.7A patent/CN112699402B/en active Active
Patent Citations (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN105447525A (en) * | 2015-12-15 | 2016-03-30 | 中国科学院软件研究所 | Data prediction classification method and device |
CN108108657A (en) * | 2017-11-16 | 2018-06-01 | 浙江工业大学 | A kind of amendment local sensitivity Hash vehicle retrieval method based on multitask deep learning |
CN111742313A (en) * | 2018-02-14 | 2020-10-02 | 万思伴股份有限公司 | System, apparatus and method for privacy preserving context authentication |
CN109284626A (en) * | 2018-09-07 | 2019-01-29 | 中南大学 | Random forests algorithm towards difference secret protection |
CN111598186A (en) * | 2020-06-05 | 2020-08-28 | 腾讯科技(深圳)有限公司 | Decision model training method, prediction method and device based on longitudinal federal learning |
Non-Patent Citations (1)
Title |
---|
不完全数据集的差分隐私保护决策树研究;沈思倩等;《计算机科学》;20170630;第44卷(第6期);第139-143页 * |
Also Published As
Publication number | Publication date |
---|---|
CN112699402A (en) | 2021-04-23 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN112084422B (en) | Account data intelligent processing method and device | |
CN111932386B (en) | User account determining method and device, information pushing method and device, and electronic equipment | |
WO2021203854A1 (en) | User classification method and apparatus, computer device and storage medium | |
CN112699402B (en) | Wearable device activity prediction method based on federal personalized random forest | |
CN109766454A (en) | A kind of investor's classification method, device, equipment and medium | |
CN109344326A (en) | A kind of method for digging and device of social circle | |
US11397757B2 (en) | Relevance estimation and actions based thereon | |
Avrachenkov et al. | Network partitioning algorithms as cooperative games | |
CN109960755B (en) | User privacy protection method based on dynamic iteration fast gradient | |
CN112418525A (en) | Method and device for predicting social topic group behaviors and computer storage medium | |
Gokulkumari et al. | Analyze the political preference of a common man by using data mining and machine learning | |
Flexer et al. | Mutual proximity graphs for improved reachability in music recommendation | |
Maroulis et al. | Context-aware point of interest recommendation using tensor factorization | |
Liao et al. | GRBMC: An effective crowdsourcing recommendation for workers groups | |
Sun et al. | Overlapping community detection based on information dynamics | |
Liu et al. | Federated personalized random forest for human activity recognition | |
Silva et al. | Social media as a source of sensing to study city dynamics and urban social behavior: Approaches, models, and opportunities | |
CN110196951A (en) | User matching method and equipment | |
Taha et al. | A system for analyzing criminal social networks | |
CN110929171B (en) | Interest point recommendation method based on scene and social network location | |
Nia et al. | Leveraging social interactions to suggest friends | |
Uddin et al. | Proposing logical table constructs for enhanced machine learning process | |
CN115147154A (en) | Social network group influence maximization implementation method | |
Mukhina et al. | Forecasting of the Urban Area State Using Convolutional Neural Networks | |
Gao | Design and Implementation of 3D Animation Data Processing Development Platform Based on Artificial Intelligence |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
GR01 | Patent grant | ||
GR01 | Patent grant | ||
TR01 | Transfer of patent right | ||
TR01 | Transfer of patent right |
Effective date of registration: 20240126 Address after: 230000 floor 1, building 2, phase I, e-commerce Park, Jinggang Road, Shushan Economic Development Zone, Hefei City, Anhui Province Patentee after: Dragon totem Technology (Hefei) Co.,Ltd. Country or region after: China Address before: 541004 No. 15 Yucai Road, Qixing District, Guilin, the Guangxi Zhuang Autonomous Region Patentee before: Guangxi Normal University Country or region before: China |