An Infectious Disease Prediction Method Based on K-Nearest Neighbor Improved Algorithm

International Journal of Database Management Systems (IJDMS ) Vol.11, No.1, February 2019
DOI: 10.5121/ijdms.2019.11102 19
AN INFECTIOUS DISEASE PREDICTION METHOD
BASED ON K-NEAREST NEIGHBOR IMPROVED
ALGORITHM
Yaming Chen1
, Weiming Meng2
, Fenghua Zhang3
,Xinlu Wang4
and Qingtao Wu5
1,2&4
Computer Science and Technology, Henan University of Science and Technology,
Luo Yang, China
3
Computer Technology, Henan University of Science and Technology, Luo Yang, China
5
Professor, Henan University of Science and Technology, Luo Yang, China
ABSTRACT
With the continuous development of medical information construction, the potential value of a large
amount of medical information has not been exploited. Excavate a large number of medical records of
outpatients, and train to generate disease prediction models to assist doctors in diagnosis and improve
work efficiency.This paper proposes a disease prediction method based on k-nearest neighbor improvement
algorithm from the perspective of patient similarity analysis. The method draws on the idea of clustering,
extracts the samples near the center point generated by the clustering, applies these samples as a new
training sample set in the K-nearest neighbor algorithm; based on the maximum entropy The K-nearest
neighbor algorithm is improved to overcome the influence of the weight coefficient in the traditional
algorithm and improve the accuracy of the algorithm. The real experimental data proves that the proposed
k-nearest neighbor improvement algorithm has better accuracy and operational efficiency.
KEYWORDS
Data Mining,KNN, Clustering,Maximum Entropy
1. INTRODUCTION
Data mining, as a new technology for data information reuse, can effectively provide important
information for hospital management decisions [1]. It uses the database, artificial intelligence and
mathematical statistics as the main technical pillars for technical management and decision-
making [2]. In the process of medical information management, application data mining
technology can better organize and classify medical and health information to establish a
management model, which can effectively summarize data and provide valuable information for
the efficient operation of medical work [3].
The prediction method based on medical big data is still in the stage of contending, and can be
divided into two categories: one is based on data to train and generate predictive model [4-6],
which is based on supervised machine learning for specific purposes. Methods, such as
classification, regression, deep neural network, etc., generate training models to predict clinical
events in unknown patients; the other is to measure patient distances, establish patient similar
groups, and predict target patients by similar group characteristics In the case of this, this is called

20
patient similarity analysis [7-11]. Specifically, patient similarity analysis refers to the selection of
clinical concepts (such as diagnosis, symptoms, examination, family history, past history,
exposure environment, drugs, surgery, genes, etc.) as characteristics of patients in a specific
medical environment. Quantitative analysis of the distance between concepts in the complex
concept semantic space, thereby dynamically measuring the distance between patients, screening
out similar groups of patients similar to index patients. In the medical big data scenario, the
various features of the patient-like group can theoretically provide multiple predictions, and have
better universal characteristics than the training model for specific goals.
An improved weighted KNN algorithm uses a distance from the sample to be identified to each
test sample to weight it, that is, a path weighting.That is to say, we give a path weight variable.
When our sample to be identified is close to which test sample, we give it a larger weight, and
instead give it a small weight [12]. The semi-supervised SVM-KNN classification method makes
full use of unmarked data, and through basic experiments, when the known training information
is insufficient, the fusion training of a large number of unmarked data can improve the final
classification accuracy [13]. The KNN algorithm is parallelized on the Spark platform, which can
effectively solve the problem of low efficiency of searching historical database in the
neighborhood search process of KNN algorithm and improve the computational efficiency of the
algorithm [14]. The above methods only consider efficiency or accuracy and do not fully
integrate.
In this paper, based on clustering and maximum entropy, K-nearest neighbor algorithm based on
Clustering and Maximum Entropy (CME_KNN) is proposed. In the data preprocessing,
clustering is used to extract clusters. The generated samples near the center point are applied to
the K-nearest neighbor algorithm as a new training sample set, and the improved K-nearest
neighbor algorithm based on the maximum entropy preserves all the uncertainties, closest to the
natural state, without The weight is determined, so the formula is not affected by subjective
factors, overcoming the significant shortcomings of traditional algorithms. Based on the data
training results, design feedback mechanism, continuously optimize the model, and gradually
improve the accuracy of the prediction results.
2. RELATED WORK
2.1 DATA MINING
Data mining combines theories and techniques of high-performance computing, machine
learning, artificial intelligence, pattern recognition, statistics, data visualization, database
technology, and expert systems [15]. The era of big data is both an opportunity and a challenge
for data mining. Analyze big data, establish appropriate systems, continually optimize and
improve the accuracy of decision-making, so as to be more conducive to mastering and adapting
to the multi-terminal changes in the market [16]. In the era of big data, data mining has been
recognized as the most commonly used data analysis method in various fields. At present,
domestic and foreign scholars mainly study the application of classification, optimization,
recognition and prediction in data mining in many fields.
2.1.1 Classification.
Along with the progress of the times and the rapid development of science and technology, as a
populous country, China's public data generated in health care and aging society has grown

21
geometrically, and the value of the data based on big data is urgently needed solve.The structure,
scale, scope and complexity of health care data are constantly expanding. Traditional calculation
methods cannot fully satisfy the analysis of medical data. Data mining technology can be based
on some characteristics of medical data: pattern polymorphism, information Loss (missing values
in the data due to personal privacy issues), timing, and redundancy classify health care data to
provide accurate decision-making for doctors or patients [17].
At the same time, China is accelerating its entry into an aging society, and the Internet is an
important medium for improving an aging society. Big data is an important technical means of
assessing an aging society. Qu Fang et al. [18] proposed the “Internet big data” model for the
realization of old-age care. The entire old-age service system is based on multi-dimensional
heterogeneous information aggregation and data fusion mining. The “Internet Big Data” pension
system integrates a variety of information and communication technologies, including
communication technology, data mining technology and artificial intelligence technology.
2.1.2 Optimization.
The traffic conditions of the roads are closely related to people's travel. With the rapid
development of the city and the improvement of living standards, the scale of motor vehicles has
gradually expanded, which has caused problems such as traffic congestion. Data mining
technology can effectively solve the optimization problem between traffic roads and logistics
networks. Pan et al. [19] proposed a data mining forecasting model, which is used to “predict
real-time” short-term traffic conditions and drive to traffic jams. The staff brought great help.
With the development of technology, online shopping has become more and more popular, and at
the same time, it has brought problems such as congestion and embarrassment of logistics and
transportation. Jingdong, one of China's largest online trading platforms, uses the UAV to detect
road condition feedback data in the era of artificial intelligence optimization, and uses data
mining technology to accurately calculate the parameters required for logistics network
transportation, which can easily and efficiently alleviate logistics. The problem of transporting
cockroaches led to the first robotic courier in China to deliver the first item to Renmin University
of China. With the increase of the length and complexity of the traffic network in the future, the
difficulty of implementing the unmanned automation strategy is also greatly increased. Only
through data mining technology can the results be quickly calculated, and the high value
generated from the complex road information can be obtained.
2.1.3 Identification.
Since the advent of digital images in the 1950s, digital images have become an indispensable
"data" in human society. In computer applications, data mining is becoming more and more
popular in image recognition applications. Representative applications are face recognition and
fingerprint recognition. Face recognition further analyzes and processes reliable and potential
data by data mining of the obtained information base, and fully prepares the data analysis work
and future development work. Wright et al. [20] described robust face recognition based on
sparse representation, and gave a detailed theoretical analysis and practical summary.
Sha Yaqing et al. [21] proposed an identity authentication scheme based on smart card and
fingerprint identification for the insecure use of username and password in the current electronic
tax filing system, and combined with fingerprint technology to construct new password
parameters, thus making The safety is significantly improved. With the continuous development

22
of data mining technology, the accuracy of big data recognition of faces and fingerprints will
become higher and higher.
2.1.4 Forecast.
Forecasting problems are the most studied issues in various fields, and their purpose is to predict
future data values or trends through historical data. Most of the historical data is time series data,
which means that they are arranged in chronological order and a series of observations are
obtained. Due to the continuous advancement of information technology, time series data has also
increased dramatically, such as weather forecasting, oil exploration, and finance. The ultimate
goal of time series data mining is to predict the trend of the future and its impact by analyzing
historical data of time series.
"Meteorology" is closely related to the ecological balance of the earth and people's normal life.
Therefore, accurate forecasting of meteorology is particularly important. Zhou Lei et al [22]
summarized the current meteorological monitoring model, based on the drought of remote
sensing data, classified the current remote sensing monitoring methods, classified the external
environmental conditions (temperature, humidity, etc.) and proposed solving complex problems.
new method.
As a non-renewable resource, oil is currently shrinking global reserves, making oil exploration
more and more important. In petroleum exploration management, the collected data has the
characteristics of large amount of data, large amount of calculation, single source of acquisition
and complex data processing flow [23], and high-performance parallel computing of large data
sets collected by data mining technology. Analysis can ensure the validity and accuracy of the
results.
In the era of big data, daily business such as banks, securities companies, and insurance
companies will generate massive amounts of data. Using the current database system, data entry,
query, and statistics can be efficiently implemented. Currently, from simple queries to It is
especially important to use data mining technology to mine knowledge and provide decision
support. The application of data mining technology in the financial industry is feasible. The
application of the theoretical basis to relevant examples includes forecasting stock indices,
discovering implicit patterns in financial time series, credit risk management and exchange rate
forecasting.
2.2 KNN ALGORITHM
The K-nearest neighbor (KNN) is also called the nearest neighbor method, which selects a certain
distance function and then calculates the distance between the test sample and the training sample
in the multidimensional space, and selects the nearest distance from it[24]. The K training sample
points are finally judged according to the category of the training sample points [25].The concept
of tree structure is introduced into the KNN algorithm to form a K-nearest neighbor on tree
(TKNN).
The KNN algorithm is different from positive classification algorithms such as neural networks
and support vector machines. It is a passive classification method. That is to say, the classification
algorithm does not need to learn the rule information according to the training sample data in
advance, and only performs the calculation classification when the classification task appears.

23
The rule of the KNN algorithm is to train the data sample itself. Therefore, the KNN algorithm
takes almost no time in the training phase, and its time consumption is used in the classification
phase. Moreover, as a non-parametric classification method, the KNN algorithm has the
advantages of simple and intuitive, easy to implement, high classification accuracy and
robustness for unknown and non-normally distributed data. Although the KNN method relies on
the limit theorem in principle, it is only related to a very small number of adjacent samples in
class decision making. Since the KNN method relies mainly on surrounding neighboring samples
rather than relying on the discriminant domain method to determine the class of the test sample,
the KNN method is more suitable than other methods for the sample set to be cross-over or
overlapped.
2.3 TF-IDF ALGORITHM
The TF-IDF algorithm is a statistically based method to measure the keyness of a word or phrase
in textual information. Its main principle: a word is used more frequently in the target text, but it
is used less frequently in the corpus, then it can have good text distinguishing ability [26]. The TF
value of a word in the target text refers to how often the word appears in the text. When
calculating this frequency, you need to normalize it to prevent it from being biased towards long
text with a large number of words. The formula for calculating the TF value is as follows:
,
i,
,
1
i j
j n
i j
i
f
TF
f
=
=
∑
(1)
Equation 1 is the calculation of the word frequency i
TF of the word i in the text j, where the
molecular content is the number of times the word i appears in the text, and the denominator
content is the total number of all words in the text j. Through the calculation method of the word
frequency, the bias of the word frequency i
TF on the longer text is effectively prevented. In
addition, another word frequency normalization method can be used, such as formula 2. Where
the denominator represents the maximum word frequency value that appears in the text j [27], ie
,
i,
max,
i j
j
j
f
TF
f
= (2)
The critical measure of a word can be expressed in terms of IDF values. Assuming that the
number of texts containing the word i in a text corpus collection is smaller, the i
IDF value of the
word should be larger, so that the better the distinguishing ability of the word i, the more likely it
is to be a keyword. The specific formula is as follows:
i log
1
i
N
IDF
n
=
+
(3)
Where N is the total number of texts in the text set and i
n contains the number of texts in word i.
The +1 in the denominator is to deal with the case where the denominator is 0 in the formula.

24
The calculation by the above method can be summarized into the formula 4. Suppose a word i
appears in a given text j with a higher frequency and less text in the entire text set containing the
word, then its TF-IDF value is higher, that is, the word i is easier to distinguish the document j,
can be used as a keyword.
,
- i i j i
TF IDF TF IDF
= × (4)
For the patient's text medical data information, the ik tokenizer can be used for word
segmentation. For the special treatment of the word segmentation of medical data, the proprietary
medical terminology can be included, put into the custom word segmentation dictionary, and then
the TF-IDF algorithm is used to extract the feature of the text content. The past history, current
medical history, complaints and physical examination data of the outpatient medical records were
extracted according to the above methods, and digitally represented.
3. TEST METHODS
3.1 SYSTEM ARCHITECTURE
Query
classification
Query interpretation
and normalization
Querying the
machine
Data warehouse
management system
Database management
system
Knowledge base
management system
database
Model
library
knowledge
base
Knowledge discovery
preprocessing
Pending
data
subset
Knowledge
evaluation
Generalized by
concept
Conclusion of
explanation
user
Knowledge
extraction
User query
Figure 1. Structure of a general data mining system based on data warehouse

25
A structural framework of a general data mining system based on data warehouse is used to
handle various types of data mining tasks. The model is mainly composed of user query interface,
query collaborator, data warehouse management system, knowledge base management system,
model library management system, knowledge base discovery preprocessing module, knowledge
evaluation module and conclusion expression module.
The functions of each part are briefly described as follows:
Query interface: The knowledge discovery request submitted by the user is interpreted into a
normalized query language through certain steps, and is processed by the query synergy machine.
The work of the query interface can be divided into two parts: query classification, query
interpretation and normalization.
Query coordinator: collaborative data warehouse management system, model library
management system and knowledge base management system, together to process query requests
submitted by the query interface. Querying the collaboration machine is an important part of the
system.
Data warehouse management system: directly responsible for the management of the data
warehouse, and complete the extraction of data in various heterogeneous distributed data sources,
to shield the impact of various heterogeneous data sources on the system to the maximum extent.
Knowledge Base Management System: Manage and control the knowledge base, including the
addition, deletion, update and query of knowledge. On the one hand, accepting the knowledge
base query request generated by the querying machine and performing the query, and submitting
the result to the knowledge discovery module; on the other hand, accepting the knowledge pattern
obtained by the knowledge extraction module from the final comprehensive conclusion,
depositing Knowledge base to enrich the content of the knowledge base.
Model library management system: Manage the model library. An important part of the model
library is the knowledge discovery module, which includes various data mining tools, such as:
association rule discovery sub-module; classification rule discovery sub-module; cluster analysis
discovery sub-module; feature rule discovery sub-module, model library at the same time It can
also manage various types of analysis tools. This unified management of data mining tools and
other analysis tools enables data warehouse-based data mining systems to be extended to decision
support systems.
Knowledge discovery preprocessing module: Under the cooperation of the data warehouse
management system, according to the metadata and dimension tables, the data stored in the entire
data warehouse is processed to generate the requirements that meet the requirements of the user
query and meet the requirements of the knowledge discovery tool set. Process data subsets.
Knowledge Evaluation Module: Evaluate the patterns found during the data mining phase. At this
time, if there is a redundant or irrelevant mode, it needs to be culled; if the mode does not meet
the user's requirements, then the entire discovery process needs to be returned to the discovery
stage, the data is re-selected, and a new data transformation method is used to set a new one. Data
mining parameter values, even replacing data mining algorithms.
Conclusion Expression Module: Generalize the conclusions obtained according to the semantic
hierarchy, and draw conclusions at each semantic level: then explain the conclusions obtained,

26
and present the discovered patterns to the users in a visual way, or Transform the results into a
natural language form that is easy for users to understand.
3.2 BUILDING KNN BASED ON CLUSTERING AND MAXIMUM ENTROPY
3.2.1 CLUSTERING TO BUILD A STREAMLINED SAMPLE SET
Clustering technology is a very effective data analysis method. Clustering algorithm refers to the
clustering analysis of data without any data prior information. This algorithm is also called
unsupervised learning method. Clustering learning is one of the earliest methods used in pattern
recognition and data mining tasks, and is used to study large databases in various applications, so
clustering algorithms for big data are receiving more and more attention [28] . In many practical
problems, traditional clustering algorithms sometimes fail to obtain effective clustering results
because there is no prior analysis of the data itself. In some practical problems, a priori
knowledge of a small amount of data is obtained, including class labeling and data point
partitioning constraints (such as pairwise constraint information), how to use these only a small
amount of prior knowledge to a large number of no priors Clustering analysis of knowledge data
has become a very important and urgent problem to be solved.
In the data preprocessing, the clustering method is used to extract the samples near the middle
point generated by the cluster, and these samples are applied as a new training sample set in the
K-nearest neighbor algorithm.
Firstly, each type of training sample set is clustered separately, and then the sample near the
center point of the cluster is used as a subclass of the category to represent all the samples of the
category, so that the classification accuracy can be ensured. The number of original training
samples is greatly reduced, and the calculation speed of the K-near algorithm is greatly improved.
Because CURE [29] (Clustering Using Representatives) clustering algorithm solves the problem
of preference sphere and similar size, it also has excellent performance in eliminating isolated
points (noise). This paper uses CURE clustering algorithm. However, the premise of using the
CURE algorithm is that the number of clusters N needs to be provided, and the number of clusters
clustered in practical applications is often unpredictable, and a range will be given to N in the
algorithm, such as [1, 8]. The clustering calculation is performed for each N in the range, and
then the classification accuracy of the clustered model is tested one by one, and finally the N
value with the highest classification accuracy is selected.
3.2.2 CONSTRUCTING A DISTANCE METRIC FUNCTION BASED ON MAXIMUM ENTROPY
Most of the improved KNN algorithms for the distance metric function are based on the
improvement of the Euclidean distance or the cosine of the vector angle, ie the weight adjustment
factor is increased. The determination of feature weights may be determined based on the effect
of each feature on the classification, or may be set by the role of the feature in the training sample
set. Therefore, the introduction of the weight adjustment coefficient has certain subjectivity,
which naturally affects the accuracy of the classification.
Entropy is originally a concept used to describe the degree of molecular chaos in
thermodynamics. The more chaotic the molecule, the larger the entropy. With the advent of
information theory, entropy has been introduced into information science [30]. In 1948, Shannon
proposed the concept of information entropy, and believed that people's understanding of things

27
is uncertain. This degree of uncertainty is called information entropy. The information entropy of
a random event is defined as: the random variable η, η has n possible cases ( 1 2
, , , n
η η η
L ), and
the probability of each case is 1 2
, , , n
p p p
L , then it is not Determine the degree, that is, the
information entropy is:
n
1
( ) - log
i i
i
H p p
η
=
= ∑ (5)
Similarly, the continuous variable η, the density function is ( )
p η then the information entropy is
defined as:
( ) - p( )log ( )
H p d
η η η η
Ω
= ∫ (6)
The principle of Maximum entropy (ME) was proposed by Jaynes in 1957. He pointed out that
when only partial information of unknown distribution is grasped, the distribution obtained when
entropy is maximized is the most objective and closest to the actual distribution. The most
uncertain is the most unbiased, the principle of maximum entropy emphasizes that the determined
probability distribution should be consistent with the known partial information, that is, the
measured data or sample, without any subjective assumptions or estimates for the unknown
information. [31]. It can be seen that the probability distribution obtained according to the
principle of maximum entropy is the closest to the actual distribution.
In recent years, the concept of maximum entropy has been widely used in many fields, such as
information theory, pattern recognition, and transportation. It represents a measure of the
uncertainty of things. When the entropy is maximum, it means that the most uncertain things, that
is, the closest to the actual state, which is the main advantage of the maximum entropy method.
Maximum entropy is a general method for solving the inverse problem of a solution that cannot
be determined due to insufficient data. It is often used for linear and nonlinear model estimation.
It turns out that the maximum entropy essence is a kind of similarity measure, that is, the distance
measure between the observed value and the actual value. This kind of measure is more effective
when the sample data are positive. Therefore, we introduce the maximum entropy instead of the
Euclidean distance as the distance measure function of KNN. Experiments show that the method
can effectively classify the data.
The most striking feature of the maximum entropy technique is that many different features can
be integrated into a probabilistic model, and there is no independent requirement for these
features, and the maximum entropy model has the characteristics of training time period and
small classification complexity, so we use The maximum entropy is used as the distance function
yj between the training sample xi and the test sample. The specific formula is as follows:
( )
1
_ , y log
k
l
j
k
i j j k
k i
y
d ME x y
x
=
= ∑ (7)
Let 0log0=0. Use the formula instead of the Euclidean distance in the TR_KNN algorithm. Since
the maximum entropy retains all the uncertainties and is closest to the natural state, the weight is

28
not determined. Therefore, the maximum entropy metric formula is not affected by subjective
factors, which removes the shortcomings of the subjectivity of traditional algorithms.
3.2.3 IMPROVED KNN ALGORITHM
An improved KNN algorithm combining clustering algorithm and maximum entropy. The
improved algorithm steps are as follows:
Step 1:
Use the CURE clustering algorithm to streamline the training sample set.
Step 2:
Construct a training sample set and a test sample set. The training sample set is represented as Ω ,
i i
={(x ,c )|i=1,2 ,n}
Ω L
， ,where 1 2
(x ,x , ,x )
l
i i i i
x = L is a l-dimensional vector, that is, the
feature dimension is 1, xl
i representing the j-th feature component value of the i-th training
sample. i
c represents the corresponding category of the i-th sample, and i
c belongs to the label
set C,C {1,2, , t}
= L , and t is the number of categories. The test sample set is expressed as Φ ,
j
{y | j 1,2, , }
m
Φ = = L ,where yl
j is the i-th eigen component value of the j-th test sample.
Step3:
Determine the initial value of K;
Step 4:
Apply the maximum entropy formula to calculate the distance between each test sample and the
training sample, and sort the distance, and select the K samples closest to the test sample as the K
neighbors of the test sample;
Step 5:
Find the main categories: set K neighbors 1 2 k
x ,x , ,x
′ ′ ′
L , the corresponding category labels are
1 2 t
c ,c , ,c
′ ′ ′
L , these category labels belong to the label set C. The queried test samples are
classified according to the categories of K neighbors and applying the maximum probability. The
probability used refers to the proportion of each category appearing in K neighbors, calculated by
dividing the number of samples in each of the K neighbors by K. Then the category with the
highest probability is recorded as the main category.Let 1 2 3
{s ,s ,s , ,s }
t
S = L be a collection of
sample sizes for each of the K neighbors. Then
*
argmax(s / )
r C
k
τ
τ
∈
= , the sample to be tested i
y
is classified into *
τ classes.
Step6:
Evaluation. If the classification effect is not good, return to Step3 and continue the steps from
Step3 to Step6, otherwise the algorithm ends.

29
The maximum entropy is used as its distance metric function, and no introduction of weight
coefficients is involved, and the principle of maximum entropy does not introduce subjective
estimates or assumptions. Therefore, it is completely objective and the classification result is
more accurate.
4 RESULTS AND CONCLUSIONS
4.1 EXPERIMENTAL SYSTEM IMPLEMENTATION
This research developed a smart medical information system to provide a disease prediction
method for doctors based on specific conditions. According to the patient's past history, current
medical history, chief complaint, physical examination, advanced training data model, according
to the CME_KNN algorithm, the patient's diagnosis is inferred, and the doctor can provide a
reference for the doctor, and the doctor can also judge the result of the estimated patient
diagnosis. Feedback information, the existing data model is gradually modified until it is more
complete. Improve the efficiency of doctors and contribute to the automation of medical care. The
operating environment of this system is: processor Inter Core i5-4460 3.20GHz, memory 8GB,
hard disk 931.41GB, operating system is Windows 10 x64, programming language is C#.
4.2 EXPERIMENTAL RESULTS
4.2.1 RELEVANT DEFINITIONS
In order to evaluate the performance of the CME_KNN algorithm and the TR_KNN algorithm,
four standard UCI (http://archive.ics.uci.edu/ml/datasets/Adult) data sets are selected in this
paper. Table 1 gives the four data sets. specific information. The Abalone data set is incomplete.
For ease of use, only the first 100 sets of data for each class are selected from the three categories
of the standard Abalone data set. The performance of the two algorithms is analyzed from
multiple dimensions by comparing the two algorithms in various cases.
Table 1. UCI data sets used in the experiment
Data set Attribute type Number of samples Number of
categories
Number of
features
Iris Real 150 3 4
Wine Integer,Real 178 3 13
Abalone Real 300 3 8
Balance Categorical 910 5 4
The classification effect evaluation indicators commonly used are the recall rate (r), precision
ratio (p), Fi test value, macro average and micro-average, etc. The macro-average index is the
evaluation of the average situation of the class, and the micro-average index is the sample. The
evaluation of the average situation, the classification performance evaluation indicators used in
this paper have macro average recall rate ( ma -
cro r ), macro average precision rate ( ma -
cro p )
and macro F1 measurement value ( 1
ma -F
cro measure ), the calculation formula of each
evaluation index is as follows:

30
1
ma -
n
k
k
r
cro r
n
=
=
∑
(8)
1
ma -
n
k
k
p
cro p
n
=
=
∑
(9)
1k
1
1
F
ma -F
n
k
cro
n
=
=
∑
(10)
In Equation 8, k
r represents the recall rate, /
k k k
r a b
= , where k
a represents the number of
correctly tested in the Kth test sample, k
b represents the number of Kth samples; t is the number
of categories of the test sample In Equation 9, k
p is the precision, /
k k k
p a d
= , where k
d
represents the number of test samples tested as Class K samples. In Formula 10, 1K
F is the F1
measurement, 1 2 / ( )
K k k k k
F r p r p
= + , which combines the recall rate k
r and the precision k
p
into one measurement. The macro F1 measurement is the average of all individual category F1
values.
In addition, classification accuracy is often used as an evaluation index for classification
performance. Classification accuracy refers to the proportion of samples correctly classified in the
entire test sample set. The formula is
Sort the correct number of samples
Total number of all samples
ACC = (11)
4.2.2 EXPERIMENTAL RESULTS BASED ON DATA SETS
In the experiment comparing CME_KNN algorithm with TR_KNN algorithm, we set the
percentage of training samples in the whole data set to 2/3, k value is 10, the experimental results
are shown in the table.
Table 2. Comparison of classification results of CME_KNN and TR_KNN based on four data sets
Data set rTR/% rME/% pTR/% pME/% F1TR F1ME
Iris 95.83 97.92 97.92 98.04 96.86 97.98
Wine 96.66 96.80 94.71 97.22 95.67 97.00
Abalone 86.87 89.74 78.04 78.89 82.22 83.97
Balance 63.15 75.80 69.30 77.50 66.08 76.64
In the table, compared with the TR_KNN algorithm, the CME_KNN algorithm has improved in
all the evaluation indexes of all data sets. On the wine dataset, the macro average recall rate is the
smallest from TR_KNN to CME_KNN, but the value still reaches 0.14%. However, all the
indicator values of the Balance data set are significantly improved, and the classification effect is

31
obvious. On the Balance data set, ma -
cro r , ma -
cro p , and 1
ma -F
cro measure are increased
by 12%, 8%, and 0.1, respectively. The average growth rate of ma -
cro r , ma -
cro p , and
1
ma -F
cro measureon these four data sets was 4.4375%, 2.92%, and 3.69%. Therefore, the
CME_KNN algorithm has better classification performance than the TR_KNN algorithm.
4.2.3 INFLUENCE OF DIFFERENT K VALUES ON CLASSIFICATION ACCURACY
k is one of the key parameters affecting the classification accuracy of KNN. In general, if the
value of K is very small, the less classification information obtained from the training sample set,
the classification accuracy of the KNN algorithm is also relatively low. Within a certain range,
the classification accuracy of the KNN algorithm gradually increases as the K value increases.
However, when K exceeds a certain atmosphere, the classification accuracy of the KNN
algorithm begins to decrease, because too many neighbor samples are selected at the time of
classification, and a large amount of noise is generated for the classification information obtained
from the training sample set.
In order to illustrate the influence of K value on classification accuracy, four data sets were
selected to verify the improved algorithm and the traditional algorithm. The experiment gave K,
5, 10, 15, and 20 values respectively, and set the training samples throughout. The percentage of
the data set is 2/3, and the experimental results are shown in the figure.
a）Iris b) Wine
c) Abalone d) Balance
5 10 15 20
K
93.5
94
94.5
95
95.5
96
96.5
97
97.5
98
TRKNN
CMEKNN
5 10 15 20
K
91
92
93
94
95
96
97
98
99
Accuracy(%)
TRKNN
CMEKNN
5 10 15 20
K
86.5
87
87.5
88
88.5
89
89.5
90
Accuracy(%)
TRKNN
CMEKNN
5 10 15 20
K
68
70
72
74
76
78
80
Accuracy(%)
TRKNN
CMEKNN

32
Figure 2. The accuracy of CME_KNN and TR_KNN in different data sets varies with k value
In the figure, on the Iris dataset, the accuracy of TR_KNN and CME_KNN reaches the maximum
at k=5 and 15. On the wine dataset, the accuracy of the improvement before and after k=15
reaches the maximum. For the Abalone dataset, When k=20, the classification accuracy is the
highest. On the Balance dataset, the classification accuracy reaches a maximum when k=10. In
these four images, the trend of the curves varies, and the k value obtained when the classification
effect is the best is different. Since the value of the experiment k is too small, and k is worth the
jitter, the graph a , b, d can partially show the trend of classification accuracy with k worth, but
the c map can not reflect the influence of k value on classification accuracy. Therefore, it is
concluded that the value of k has no definite method, and it can only be adjusted continuously
until the classification accuracy is optimal, and the optimal k value for the different data sets is
not necessarily the same, although There is no definite method for determining the k value, but
the classification accuracy of the same k-value CME_KNN algorithm is significantly better than
the TR_KNN algorithm.
4.2.4 EFFECT OF THE PERCENTAGE OF TRAINING SAMPLES IN THE SAMPLE SET ON
CLASSIFICATION ACCURACY
The proportion of training samples in the entire sample set during the classification process also
affects the classification results. In order to verify the influence of the proportion of training
samples on the classification results, four sample sets were selected for the verification
experiments of CME_KNN and TR_KNN algorithms. The specific proportions of the selected
training samples were 1/3, 1/2, 2/3, respectively. 5 Experimental results are shown in the table.
Table 3. Effect of the proportion of training sample sets on classification accuracy (application CME_KNN
and TR_KNN)
Classification
accuracy
Ptrs Iris Wine Abalone Balance
ACCTR/% 1/3 92.93 89.83 85.33 68.26
1/2 93.33 88.64 86.36 71.81
2/3 95.83 91.38 87.88 71.52
ACCME/% 1/3 94.95 91.53 88.89 74.38
1/2 95.67 94.32 88.89 80.18.
2/3 97.92 94.83 89.90 78.48
As the proportion of training samples in the whole data set increases, the classification accuracy
of the CME_KNN algorithm and the TR_KNN algorithm gradually increases. The bold data in
the table is an outlier, that is, it does not conform to the increasing trend, but overall, for the same
data set. In the case of the same algorithm, when the proportion of training samples is 1/3, 1/2,
and 2/3, the classification accuracy of the knn algorithm increases sequentially. The classification
accuracy of the CME_KNN algorithm is higher than that of the TR_KNN algorithm.
5. CONCLUSIONS
The calculation amount of KNN algorithm is exponentially increased with the increase of sample
set data. From the perspective of reducing the amount of data, it is proposed to simplify the
training sample set by CURE clustering algorithm. Most improved algorithms based on KNN use
the method of adding attribute weights to calculate the distance between two samples. This

33
method has certain subjectivity and affects the accuracy of classification.An improved KNN
algorithm based on maximum entropy is proposed.
The experimental results show that the improved algorithm proposed in this paper is more
effective than the traditional KNN algorithm in terms of recall rate and precision. However, the
distance metric function has a logarithmic term and therefore only applies if the feature is
positive. The improved KNN algorithm proposed in this paper still has some shortcomings, which
requires further research and improvement.
ACKNOWLEDGEMENTS
Participating in the information construction project of Yongcheng Central Hospital made me
grow up a lot and master some necessary skills. Thank the laboratory for providing such a good
environment. Thank the teachers and students for helping me. Thank you for helping me to grow
up.
REFERENCES
[1] Ian H. Witten, Eibe Frank, & Mark A. Hall.(2005).Data Mining: Practical Machine Learning Tools
and Techniques (Third Edition).
[2] Burges, C. J. C. . (1998). A tutorial on support vector machines for pattern recognition. Data Mining
& Knowledge Discovery, 2(2), 121-167.
[3] Rjeily, C. B., Badr, G., Hassani, A. H. E., & Andres, E. (2019). Medical Data Mining for Heart
Diseases and the Future of Sequential Mining in Medical Field. Machine Learning Paradigms.
[4] Stiglic, G. , Brzan, P. P. , Fijacko, N. , Wang, F. , Delibasic, B. , & Kalousis, A. , et al. (2015).
Comprehensible predictive modeling using regularized logistic regression and comorbidity based
features. PLOS ONE, 10(12), e0144439.
[5] Nguyen, P. , Tran, T. , Wickramasinghe, N. , & Venkatesh, S. . (2016). Deepr: a convolutional net for
medical records.
[6] Choi, E. , Bahadori, M. T. , Kulas, J. A. , Schuetz, A. , Stewart, W. F. , & Sun, J. . (2016). Retain: an
interpretable predictive model for healthcare using reverse time attention mechanism.
[7] Hoogendoorn, M. , El Hassouni, A. , Mok, K. , Ghassemi, M. , & Szolovits, P. . (2016). Prediction
using patient comparison vs. modeling: a case study for mortality prediction. Conf Proc IEEE Eng
Med Biol Soc, 2016, 2464-2467.
[8] Sharafoddini, A. , Dubin, J. A. , & Lee, J. . (2017). Patient similarity in prediction models based on
health data: a scoping review. Jmir Med Inform, 5(1), e7.
[9] Zhang, P. , Wang, F. , Hu, J. , & Sorrentino, R. . (2014). Towards personalized medicine: leveraging
patient similarity and drug similarity analytics. Amia Jt Summits Transl Sci Proc, 2014, 132-136.
[10] Ng, K. , Sun, J. , Hu, J. , & Wang, F. . (2015). Personalized predictive modeling and risk factor
identification using patient similarity. Amia Jt Summits Transl Sci Proc, 2015, 132-136.
[11] Sherry-Ann, B. . (2016). Patient similarity: emerging concepts in systems and precision medicine.
Frontiers in Physiology, 7.
[12] Jiang, L. , Cai, Z. , Wang, D. , & Zhang, H. . (2014). Bayesian citation-knn with distance weighting.
International Journal of Machine Learning & Cybernetics, 5(2), 193-199.

34
[13] Islam, M. M., Iqbal, H., Haque, M. R., & Hasan, M. K. (2018). Prediction of breast cancer using
support vector machine and K-Nearest neighbors. IEEE Region 10 Humanitarian Technology
Conference.
[14] Maillo, J. , Ramírez, Sergio, Triguero, I. , & Herrera, F. . (2016). Knn-is: an iterative spark-based
design of the k-nearest neighbors classifier for big data. Knowledge-Based Systems,
S0950705116301757.
[15] Han, J. . (2005). Data Mining: Concepts and Techniques. Morgan Kaufmann Publishers Inc.
[16] Zhang, X. , Huang, X. , & Wang, F. . (2017). The construction of undergraduate data mining course
in the big data age. International Conference on Computer Science & Education. IEEE.
[17] Holzinger, A., & Jurisica, I. (2014). Knowledge discovery and data mining in biomedical informatics:
the future is in integrative, interactive machine learning solutions.
[18] Qu Fang,& Guo Hua.(2017). "Internet + big data" pension path to achieve. Science & Technology
Review, , 35(16): 84-90.
[19] Pan, T. L. , Sumalee, A. , Zhong, R. X. , & Indra-Payoong, N. . (2013). Short-term traffic state
prediction based on temporal–spatial correlation. IEEE Transactions on Intelligent Transportation
Systems, 14(3), 1242-1254.
[20] Wright, J. , Yang, A. Y. , Ganesh, A. , Sastry, S. S. , & Ma, Y. . (2009). Robust face recognition via
sparse representation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 31(2), 210-
227.
[21] Sha Yaqing, Sun Hongwei, & Gu Ming. (2006). Electronic tax certification system based on smart
card and fingerprint identification.Computer Engineering, 32.(14):133-135.
[22] Zhou Lei, Wu Jianjun, & Zhang Jie. (2015). Research progress of remote sensing based drought
monitoring methods. Geography Science, 35(5): 630-636.
[23] Xie Wei, Liu Bin,&Liu Xin. . (2017). Petroleum seismic exploration system and software platform in
big data era. Science & Technology Review, 35(29): 172-174.
[24] Noora Abdulrahman, & Wala Abedalkhader. (2017). KNN Classifier and Naive Bayse Classifier for
Crime Prediction in San Francisco Contextdoi. International Journal of Database Management
Systems,9(4) .
[25] Juan, L. . (2012). TKNN: An Improved KNN Algorithm Based on Tree Structure. Seventh
International Conference on Computational Intelligence & Security. IEEE.
[26] Patil, S., & Kulkarni, S. (2018). Mining social media data for understanding students’ learning
experiences using memetic algorithm. Materials Today Proceedings, 5(1), 693-699.
[27] Yan, Z., Yun, Q., & Li, C. (2018). Improved KNN text classification algorithm with MapReduce
implementation. International Conference on Systems & Informatics.
[28] He, Q. , Li, N. , Luo, W. J. , & Shi, Z. Z. . (2014). A survey of machine learning algorithms for big
data. Moshi Shibie yu Rengong Zhineng/Pattern Recognition and Artificial Intelligence, 27(4), 327-
336.
[29] Metz, T. (2018). Medicine without cure?: a cluster analysis of the nature of medicine. The Journal of
Medicine and Philosophy: A Forum for Bioethics and Philosophy of Medicine, 43(3), 306-312.
[30] Stonier, T. (2018). Towards a general theory of information ii: information and entropy. Electronic
Notes in Discrete Mathematics, 21(9), 181-184.
[31] Chen, L., Gao, S., Zhang, H., Sun, Y., Ma, Z., & Vedal, S., et al. (2018). Spatiotemporal modeling of
pm2.5 concentrations at the national scale combining land use regression and bayesian maximum
entropy in china. Environment International, 116, 300-307

35
AUTHORS
Yanming Chen is a student at Henan University of Science and Technology,
master's degree.
Weiming Mengis a student at Henan University of Science and Technology,
master's degree.
Fenghua Zhang is a student at Henan University of Science and Technology,
master's degree.

An Infectious Disease Prediction Method Based on K-Nearest Neighbor Improved Algorithm

More Related Content

Similar to An Infectious Disease Prediction Method Based on K-Nearest Neighbor Improved Algorithm (20)

More from ijdmsjournal (20)

Recently uploaded (20)

An Infectious Disease Prediction Method Based on K-Nearest Neighbor Improved Algorithm