2017 2nd IEEE International Conference on Computational Intelligence and Applications

Geospatial-Temporal Analysis andClassification of Criminal Data in Manila

Maria Jeseca C. Baculo, Charlie S. Marzan, Remedios de Dios Bulos, Conrado Ruiz
De La Salle University
Manila City, Philippines
e-mail: maria_jeseca_baculo@dlsu.edu.ph, charlie_marzan@dlsu.edu.ph, remedios.bulos@ dlsu.edu.ph, cons.ruizjr@
delasalle.ph

Abstract—The use of technology on criminal data has proven

A. Geospatial Crime Analysis
to be a valuable tool in forecasting criminal activity. Crime The mapping of crime locations and accurate detection
prediction is one of the approaches that help reduce and deter of spatial concentrations of crime aid to identify where
crimes. In this paper, we perform geospatial analysis using the crime occurrences tend to concentrate in terms of space and
kernel density estimation in ArcGIS 10 to identify the spatio- time and thus, provide important information for law
temporal hotspots in Manila, the most densely populated city enforcement crime reduction efforts.
in the Philippines. We also compared the performance One of the techniques used in geospatial analysis is the
measures of the BayesNet, Naïve Bayes, J48, Decision Stump, Kernel Density Estimation (KDE). This was one of the
and Random Forest classifiers in predicting possible crime methods investigated in the development of a web-based
activities. The results presented in this paper aim to provide geographical information system to map crimes and
insights on crime patterns as well as help law enforcement decision support [2]. Another study utilized KDE to
agencies design and implement approaches to respond to
conduct spatial and temporal analysis of street crimes in
criminal activities.
Faisalabad City of Pakistan. The study found that the
Keywords—classifiers, crime analysis, predictive methods majority of the street crimes were found in commercial and
densely populated areas [3].
KDE was also used to determine the density of sexual
I. INTRODUCTION crimes in Charlotte, Virginia. The models generated with
Crime is one of the most prevalent problem in the the use of ArcGIS 10 provided insights to better understand
Philippines. It affects the country’s economic growth as well crime patterns and improve policing and response [4]. The
as the quality of life of its citizens. Law enforcement use of GIS was also helpful in proposing a swift response
agencies (i.e. Philippine National Police) commit significant system which can identify the most probable local suspects
amount of resources to reduce and deter crime by involved in a crime case using historical data [5].
identifying crime prone areas [1]. They use web-based
database software applications to record, collate, analyze, B. Predictive Classification
and visualize crime data. Classification is a well-known technique in data mining
These agencies maintain and manage large crime which aids in distinguishing data classes, with the end goal
databases. And the advent of data mining technologies and of utilizing the model to predict the crime labels. It was
development of geographical information systems in crime applied to a crime dataset to predict the crime category for
analysis and prevention (i.e. geospatial crime analysis and different states in the US. The authors used the Back
predicting criminal activity) has led to higher crime solution Propagation (BP) and Naïve Bayes classifiers where the
efficiency, and higher crime clearance efficiency. latter outperformed BP Classifier [6].
Although crimes could occur anywhere, criminals may Another study [7] highlighted the use of classifiers in
prefer some situations and locations that are familiar areas crime prediction in Denver, USA. The algorithms compared
for these perpetrators. In this paper, we aim to identify were the BayesNet, NaiveBayes, OneR, J48, Decision Table
different variables that will most likely contribute to criminal and JRip. The Decision Tree and Naïve Bayesian classifiers
behavior. Identifying hotspots and the applications of various were used also to predict potential crime types and
classification algorithms can be used to increase the level of performed an analysis by combining findings of Denver
crime prediction and prevention. This can help law crimes’ dataset with its demographics information in order to
enforcement allocate resources efficiently by distributing capture the factors that might affect the safety of
their manpower to places where crimes will most likely neighborhoods. Results showed that the Bayesian classifier
happen. yields the best overall performance [8].
II. RELATED WORK III. PROBLEM
Data mining has paved the way for the exploration and This study focused on the geospatial-temporal analysis
application of advanced techniques in the field of crime
and implementation of classification algorithms to predict
analysis prediction. Presented in this section are some of
the studies that focused on geospatial crime analysis and criminal activities in the respective districts of Manila.
predictive classification. Specifically, it aims to achieve the following: (a) To identify
spatio-temporal hot spots using an actual dataset of crimes;

978-1-5386-2030-4/17/$31.00 ©2017 IEEE 6

and (b) To implement different classification algorithms to attributes included in the data set were cleaned and no other
predict possible types of crimes. noise and anomalies were found. The data cleaning process
helped improve the quality of data prior to subjecting it to
IV. DATA DESCRIPTION model development and evaluation.
2) Data integration
The dataset that was used were actual crimes in Manila,
Philippines. The data were collected manually from Manila One of the preprocessing techniques employed in this
Police District (MPD) Office. From their incident report files, study is data integration which involved the merging and
the relevant attributes for the study were identified and derivation of attributes to existing columns in the raw data
encoded in CSV format. The dataset contains information set. The resulting attributes were used as features in the
about the crimes incurred from the year 2012 to 2016. These implementation of classification algorithms. The final dataset
incident data included the type of crime, date and time of included the 5 original attributes from MPD as well as
occurrence, location, location type, isHoliday, and Rain. A derived attributes from these fields.
total of 7,108 instances were gathered from the MPD. After The attribute IsHoliday is derived from the date
data cleaning, the resulting dataset included 6,369 incidents. compared with the list of holidays per year collected from
Table I shows the distribution of incident types. Manila City Hall. The attribute Rain was also derived from
the date, where each of the dates were compared to the data
TABLE I. INCIDENT TYPES AND COUNT provided by the Department of Science and Technology
(DOST).
Incident Type Count
The dataset gathered from MPD did not include the
Carnapping 1731
coordinates of the locations of the crime per district. The
Gun Shooting 506
Homicide/Murder 1155 addresses were geocoded in MS Excel using a custom
Physical Injury 1046
function that uses Google’s geocoding service.
Robbery/Theft 1144
Sexual Abuse 787
Total 6369

V. METHODOLOGY
Manila is one of the cities of Metro Manila, the capital
region of the Philippines. It is composed of sixteen districts:
Binondo, Ermita, Intramuros, Malate, Paco, Pandacan, Port
Area, Quiapo, Sampaloc, San Andres, San Miguel, San
Nicolas, Santa Ana, Santa Cruz, Santa Mesa, and Tondo. It is
the most densely populated with 71,263 persons per square
kilometer in 2015 and ranks second to Quezon City in the
highest number of crimes from 2010 to 2015. With around
60,000 establishment operating in the city, it is one of the
major centers of commerce ranking third in 2016’s Cities
and Municipalities Competitive Index of the Philippines. Fig.
1 displays the map of Manila. Figure 1. Study area

A. Data Preprocessing The resulting coordinates were then divided into two
Data preprocessing is imperative in data mining. This attributes named Longitude and Latitude. These were used to
process cleans and transforms the raw crime data gathered plot the different points of crime incidents to the map. The
from MPD into its appropriate representation for crime coordinates also played an important role in identifying the
analysis. The following data preprocessing techniques were inconsistencies in the Location attribute. The WGS 1984
performed in the data set. Web Mercator coordinate system in ArcMap 10 was used to
1) Data cleaning plot the coordinates of these crimes in Manila.
3) Data reduction
Data cleaning was performed to the instances in the data
set which have missing and inconsistent values for crime Extracting patterns from multiple attributes becomes
location. The missing values were manually filled up and challenging when it is represented in voluminous divisions.
some were ignored since they belong to cities outside Manila. Data reduction helps remove irrelevant attributes by
During geocoding the addresses, some of the locations selecting the attributes and values that best represent the data
produced noise since the coordinates returned by the to be mined.
geocoding function lies outside the study area. Some of the Originally, the incident report handed by the MPD had 9
districts in Manila were split and some were united between number of attributes which included the nature of case, date,
2012 and 2016. For their case, the old names of the streets time, location, location type, identification of suspects and
with districts were changed to the new ones. All key victims/complainant, status and facts of the case, investigator,

7
and station concerned. Among these attributes, five were three temporal perspectives: time of the day, day of the
selected. These were the type of crime, date and time of week, and quarters of the year.
occurrence, location, and location type. The number of
instances were also reduced since some of the crime types C. Predictive Classification
given are irrelevant to the area being studied. Classification models are used to predict categorical
4) Data transformation class labels. In this paper, five algorithms for classification
were implemented to predict the possible crime locations
Transforming data entails converting values to forms
along with other variables that contribute to the occurrence
appropriate for mining. In this paper, the values representing
of crime. The tool used for the implementation of these
the attributes Weather, Location and Time were transformed
algorithms was WEKA (Waikato Environment for
to minimize the distinct values the attributes represent. This
Knowledge Analysis) version 3.9. It is a software that
helped in narrowing the circumstances leading to the
contains the visualization tools and algorithms useful for
prediction of possible crime location.
predictive modeling and data analysis. The attributes
In the data set, the resulting weather attribute had 9 included in the dataset were the District, LocationType,
distinct values after data integration. Through the use of Time, IsHoliday, and Rain.
concept hierarchy, it was then reduced to 2 values (Yes/No)
and the attribute was renamed Rain. The same technique VI. RESULTS AND DISCUSSION
was also used in the location. Instead of having hundreds of
A. Kernel Density Estimation Maps
distinct values from the streets of Manila, these were
reduced to the 16 districts where these areas belong. The Fig. 2 illustrates the distribution of crimes over a day. It
diversity of time was also reduced into a four-hour interval shows that the highest crime rate happens from 8:01 PM to
resulting to 6 distinct values. Table II shows the attributes 12:00 MN followed by the incidents from midnight to 4:00
generated after data preprocessing. AM. On the other hand, the lowest crime rate occurs in the
hours between 4:01 AM to 8:00 AM.
TABLE II. DATASET ATTRIBUTES
Attribute Description Data
To identify the spatial characteristics of the incidents,
Name Type Fig. 3 maps the incidents in a four-hour interval where the
Crime_Category Type of crime Nominal blue areas represent the highest density of incidents and
District The location where the crime happened Nominal yellow green areas signify the lowest.
Location_Type The nature of location where the crime Nominal
happened
Year The year when the crime happened Nominal
DayNominal The day of the week when the crime Nominal
happened
Longitude The vertical measure of the location of the Real
crime
Latitude The horizontal measure of the location of Real
the crime
IsHoliday Specifies whether the date of occurrence Nominal
falls under holidays
Rain Specifies whether it rained during the Nominal Figure 2. Distribution of crimes over a day
execution of the crime

B. Geospatial-Temporal Analysis 12:01 AM-4:00 AM 4:01 AM-8:00 AM 8:01 AM-12:00

The crime incidents were graphically presented using

heat map and hot spots which indicate the density of these
crimes in a particular location. Geospatial-temporal analysis
involves mapping crime location and time while overlaying
other demographic information to predict criminal activities.
12:01 PM-4:00 PM 4:01 PM-8:00 PM 8:01 M-12:00
In order to perform geospatial analysis in the dataset,
ArcGIS 10, was used in point mapping the coordinates of
the location and to produce the spatio-temporal heat maps
using Kernel Density Estimation (KDE) for points features.
ArcGIS is a desktop based GIS tool which can be used for
hotspot analysis [3, 9]. KDE is one of the popular
techniques used in identifying precise hotspots by using all
the data points in the area to produce an estimate of density Figure 3. KDE of crimes over a day
[10]. KDE was used to visualize the density of crimes across

8
 Fig. 6 exhibits that majority of these crimes happen in
the months of July to September (3rd quarter) with the
highest clusters found in the districts Tondo, San Nicolas,
Port Area, Ermita, and Malate (Fig. 7). These clusters are
also prevalent in the three remaining KDE and attained the
highest density in the time of the day and day of the week
analysis.

Figure 4. Distribution of crimes over a week B. Predictive Classification

Five algorithms for classifications were applied to the
SUNDAY MONDAY TUESDAY WEDNESDAY
crime dataset obtained from MPD. These algorithms were
BayesNet, Naïve Bayes, J48, Decision Stump, and Random
Forest. The configuration for J48 and Random Forest
classifiers as to the number of seeds was modified and
changed to 10 while the remaining algorithms used its
default configuration in WEKA.
THURSDAY FRIDAY SATURDAY
Bayesian networks (BNs) are graphical models for
reasoning under uncertainty, where the nodes represent
variables (discrete or continuous) and arcs represent direct
connections between them [11]. Meanwhile, Naïve Bayes
classifier is a probabilistic which when given an input gives
a probability distribution of set of all classes rather than
Figure 5. KDE of crimes over a day providing a single output. It makes use of all the attributes
contained in the data, and analyses them individually as
though they are equally important and independent of each
other [12].
On the other side, J-48 is a predictive machine-learning
model that decides the target value (dependent variable) of a
new sample based on various attribute values of the
available data. A decision stump is also a machine
learning model consisting of a one-level decision tree. A
Figure 6.Distribution of crimes over a year
decision stump makes a prediction based on the value of just
1ST QUARTER 2ND QUARTER 3RD QUARTER 4TH QUARTER a single input feature. Further, random forests are an
ensemble learning method for classification and regression
that construct a number of decision trees at training time and
outputting the class that is the mode of the classes output by
individual trees [13].
In order to extract patterns, the dataset was divided
Figure 7. KDE of crimes over a year based on the type of crime with selected district with the
most occurrences. The attributes included in the dataset
The maps show that crimes are most likely to happen at were the District, LocationType, Time, IsHoliday, and Rain.
around 8:01 PM to 12:00 PM in the districts: Tondo, Port The selected class was the District attribute and the
Area, Ermita, and Malate. There is also a high density of experiment aimed to predict the time and location that
crime in San Nicolas and Port Area from 8:00 PM to 4:00 crimes may occur.
AM. It can also be gleaned that there is a high density of The performance measures selected to compare the
criminal activities along Ermita and Malate with nearby results were the percentage of correctly classified instances,
universities, parks, and malls from 12:00 PM to 12:00 MN. kappa statistic, precision, and Receiver Operating
In the day of the week analysis, it is evident in Fig. 4 that Characteristic (ROC) Area. The percentage of correctly
the greatest number of incidents happen during the classified instances measures the accurate predictions made
weekends while the lowest crime rate happens during by the model. The Kappa statistic is the measure of how
Wednesdays. The KDE maps illustrated in Fig. 5 manifests well the classifiers performed in comparison with its
the areas with high density of incidents per day. Areas with performance by chance. Its value ranges from 0 to 1 where 0
high clusters include Malate during Monday and Friday, is the expected value if the results were generated by chance.
Tondo’s highest clusters happen during the weekends. Port Moreover, precision pertains to how often the classifier
Area also has high concentration of crimes during generated the correct prediction. The area under the ROC
Wednesdays.

9
curve has values ranging from 0.5 to 1. A perfect test has a Correctly
ROC
ROC Area of 1 and a useless test has 0.5. Classifier Classified Kappa Precision
Area
Instances
The results illustrate the performance of the selected Carnapping
classifiers in the carnapping, gun shooting, physical injury, BayesNet 65.37% .5348 .660 .833
and robbery/theft datasets. These datasets were selected Naïve Bayes 65.47% .5358 .661 .833
because of their high prediction accuracy where patterns can J48 73.03% .6342 .730 .891
be extracted. For murder and sexual assault, the predictions Random Forest 73.03% .6344 .730 .896
Decision Stump 35.85% .0983 .248 .583
that the models produced were at random because of their Gun Shooting
low kappa statistics. BayesNet 77.41% .5077 .772 .775
Naïve Bayes 77.78% .5146 .775 .775
1) Prediction: K-Fold validation J48 73.84% .4316 .734 .755
The k-fold cross validation method was used to run the Random Forest 76.34% .4944 .761 .784
Decision Stump 77.06% .4897 .771 .684
test 10 times and uses the final fold for testing. Out of the Robbery/Theft
six experimental datasets, the crimes concerning carnapping, BayesNet 70.18% .5994 .715 .888
robbery, and physical injury resulted to similar results. Naïve Bayes 70.41% .6026 .718 .888
It can be seen in Table III that in these datasets, J48 and J48 77.51% .6936 .776 .911
Random Forest classifiers gained the highest percentage of Random Forest 77.51% .6958 .775 .919
Decision Stump 51.24% .2595 .357 .654
correct classified instances and performance measures. With
Physical Injury
the computed ROC Area higher than 0.5, both classifiers BayesNet 73.03% .4942 .729 .819
have relatively good performance over all possible Naïve Bayes 73.03% .4942 .729 .821
thresholds. On the other hand, the Naïve Bayes classifier J48 86.31% .7617 .863 .958
provided the highest accurate prediction, precision, and Random Forest 85.68% .7509 .856 .956
Decision Stump 69.09% .385 .544 .650
kappa statistic in the Gun Shooting dataset. However, in
terms of the test performance, the model generated using the TABLE IV. ACCURACY AND PERFORMANCE MEASURES ON
Random Forest classifier got the highest ROC Area. PERCENTAGE SPLIT

2) Prediction: percentage split Train Correctly

ROC
Classifier Data Classified Kappa Precision
Percentage split was also used in dividing the dataset Area
(%) Instances
into training and testing set, a portion of the instances was Carnapping
used to build the model then the learning was tested using BayesNet 80% 66.35% .5512 .695 .855
the remaining instances. During the implementation of the 70% 66.25% .5474 .695 .840
Naïve Bayes 80% 66.35% .5512 .695 .853
algorithms, the percentages used for the training data were 70% 65.93% .5432 .692 .839
70% and 80%. Table IV demonstrates that when the test J48 80% 74.88% .6612 .754 .897
data is less, the number of correct prediction increases with 70% 73.19% .6369 .733 .881
exclusion to the two Bayesian classifiers under Random 80% 75.83% .6721 .761 .905
Robbery/Theft and Physical Injury datasets. In addition, Forest 70% 71.29% .6107 .714 .893
Decision 80% 38.86% .1257 .202 .619
when the number of trained data is less, the accuracy of
Stump 70% 37.54% .1047 .193 .604
prediction is reduced. It is also apparent that the values Gun Shooting
computed using percentage split is higher than that of the k- BayesNet 80% 80.36% .571 .806 .853
fold cross validation. This is because in the k-fold cross 70% 75% .4793 .756 .795
validation method, the test is run several times before using Naïve Bayes 80% 80.36% .571 .806 .862
the final fold for testing. 70% 75% .4793 .756 .797
J48 80% 80.36% .571 .806 .785
In contrast with the result using the k fold validation 70% 75% .4793 .756 .799
method, the Random Forest classifier has the highest Random 80% 82.14% .6196 .820 .823
performance among the other algorithms. However, the Forest 70% 76.19% .5056 .766 .767
performance measures in the robbery/theft and physical Decision 80% 80.36 .571 .806 .774
injury datasets are identical. In both k-fold validation and Stump 70% 75% .4793 .756 .733
Robbery/Theft
percentage split, J48 and Random Forest gained the highest
BayesNet 80% 75.15% .6705 .770 .913
correct predictions with minimal differences in its kappa 70% 73.12% .6451 .753 .905
statistic, precision, and ROC area. However, the Random Naïve Bayes 80% 75.15% .6705 .770 .912
Forest Classifier got the highest ROC area value in all four 70% 72.73% .64 .750 .904
datasets. J48 80% 81.07% .7457 .828 .941
70% 79.45% .7247 .801 .935
TABLE III. ACCURACY AND PERFORMANCE MEASURES ON 10-FOLD Random 80% 79.88% .7311 .801 .929
CROSS VALIDATION Forest 70% 80.23% .7349 .811 .926
Decision 80% 51.48% .2878 .351 .710
Stump 70% 50.19% .2774 .344 .710

10
 Train Correctly
ROC providing us strength to finish this worthwhile undertaking.
Classifier Data Classified Kappa Precision Appreciation is also due to our mentor, Ms. Julie Ann
Area
(%) Instances
Physical Injury Salido, for her helpful insights and ideas vital to the conduct
BayesNet 80% 76.04% .5777 .765 .877 of this research. Ms. Baculo and Mr. Marzan acknowledge
70% 77.24% .5897 .784 .878 the Commission on Higher Education, in collaboration with
Naïve Bayes 80% 76.04% .5777 .765 .896 the De La Salle University (DLSU), for funding support
70% 77.24% .5897 .784 .874 through the Commission on Higher Education K-12
J48 80% 89.58% .8321 .909 .977
70% 88.28% .81 .893 .963
Transition (CHED K-12) Program.
Random 80% 89.58% .8313 .907 .974
Forest 70% 88.28% .8089 .890 .962
REFERENCES
Decision 80% 71.88% .4767 .574 .749 [1] Inquirer, P. (2016). PNP’s crime analysis goes hi-tech. [online]
Stump 70% 69.66% .4262 .550 .714 Newsinfo.inquirer.net. Available at:
http://newsinfo.inquirer.net/760990/pnps-crime-analysis-goes-hi-
VII. CONCLUSION tech [Accessed 27 Jun. 2017].
[2] G. Zhou, J. Lin, and W. Zheng, “A web-based geographical
The heat maps generated by ArcGIS 10 using the kernel information system for crime mapping and decision support”,
density estimation showed areas in Manila with the highest International Conference on Computational Problem-Solving, 2012,
number of criminal activities in the time of day, day of the pp. 147-150.
week and quarter of the year analysis. The time when [3] S. Khalid, J. Wang, M. Shakeel, and X. Nan, “Spatio-temporal
criminal activities are at peak is around 8:00 PM to 4:00 analysis of the street crime hotspots in Faisalabad city of Pakistan”,
AM. Also, most of the incidents happen during the 23rd International Conference on Geoinformatics, 2015, pp. 1-4.
[4] E. Clougherty, J. Clougherty, X. Liu, and D. Brown, “Spatial and
weekends and throughout the year, the third quarter (July-
temporal analysis of sex crimes in Charlottesville, Virginia”,
September) has the most active criminal activities based on Systems and Information Engineering Design Symposium, 2015, pp.
historical data. The hotspots identified lies in the districts of 69-74.
Tondo, Port Area, San Nicolas, Ermita, and Malate. The [5] M. Sarayanan, R. Thayyil, and S. Narayanan, “Enabling real time
identification of these hotspots may be used as input to the crime intelligence using mobile GIS and prediction methods”,
PNP in allocating their resources at a given time and European Intelligence and Security Informatics Conference, 2013,
pp. 125-128.
location.
[6] A. Babakura, M. N. Sulaiman, and M. A. Yusuf, “Improved method
We then applied the BayesNet, Naïve Bayes, J48, of classification algorithms for crime prediction”, 2014
Decision Stump, and Random Forest classifiers to six International Symposium on Biometrics and Security Technologies,
datasets categorized by the type of crime. Results showed 2014, pp. 250-255.
that generally, the Random Forest classifier outperformed [7] A. Gupta, A. Syed, A. Mohammad, and M. Halgamuge, “A
the other algorithms in both 10-fold cross validation and comparative study of classification algorithms using data mining:
percentage split methods. However, the computed accuracy Crime and accidents in Denver city USA”, International Journal of
Advanced Computer Science and Applications, 2016, pp. 374-381.
and performance values computed were close to the results [8] T. Almanie, R. Mirza, and E. Lor, “Crime prediction based on crime
generated using J48 classifier. Both algorithms got the types and using spatial and temporal criminal hotspots”,
highest percentage of correctly classified instances of 89.58% International Journal of Data Mining & Knowledge Management
but J48 got better kappa statistic, precision and ROC area Process, 2015, pp. 1-19.
values. Using the models generated by these classification [9] Z. Wang, J. Wu and B. Yu, "Analyzing spatio-temporal distribution
algorithms, law enforcement agencies may be able to predict of crime hot-spots and their related factors in Shanghai, China," 19th
International Conference on Geoinformatics, 2011, pp. 1 – 6.
the location (class label) and possible factors (attributes) [10] E. Johansson, C. Gåhlin and A. Borg, "Crime hotspots: An
that can affect the occurrence of crimes. evaluation of the KDE spatial mapping technique," European
Intelligence and Security Informatics Conference, 2015, pp. 69 – 74.
FUTURE WORK [11] I. Ben-Gal, F. Ruggeri, F. Faltin, and R. Kenett, Bayesian networks
As future extension of this research, we plan to apply encyclopedia of statistics in quality & reliability, Wiley & Sons,
more classification and clustering methods to criminal data 2007.
[12] S. G. Sathyadevan, “Crime analysis and prediction using data
sets and expand the study area to the other cities in Metro
mining”, First International Conference on Networks & Soft
Manila. We also intend to add other attributes necessary for Computing, 2014, pp. 406-412.
profiling the identification of possible suspects based on [13] P. Sharma, “Comparative Analysis of Various Decision Tree
historical data. Classification Algorithms using WEKA”, International Journal on
Recent and Innovation Trends in Computing and Communication,
ACKNOWLEDGMENT 2014, pp. 684-690.
Our deepest, profound, and sincere gratitude to the
Heavenly Father for lending us His spirit of wisdom and for

11