Skip to main content

Table 1 Performance analysis of forecasting methods

From: Crime forecasting: a machine learning and computer vision approach to crime prediction and prevention

No.

Forecasting method

Accuracy

Limitations

Observation

References

1

Decision tree (J48)

59.15%

Took more time of 0.76 s to build model compared to 0.09 s of other models.

They took J48 naïve Bayesian and ZeroR and compared them by running tests.

[58]

2

KNN (K = 5)

66.6939%

Data filling algorithms needs to be added to increase the accuracy.

In their research they try to prove that higher accuracy can be achieved if GBWKNN filling algorithm and KNN classification algorithm is combined.

[56]

3

KNN (K = 10)

87.03%

Naïve Bayesian has slightly higher accuracy.

They essentially divided data into critical and non-critical and then compared it in 5 classification algorithms and noted that naïve Bayesian, neural networks, and KNN predict better than the SVM and decision tree.

[57]

4

Naïve Bayes classifier

87.00%

Cannot be applied to the dataset having large number of features.

They implemented a novel crime detection naïve Bayes method for crime prediction and analysis.

[60]

5

Decision tree

83.9519%

As they are unstable, a small change in data can lead to a large change in the structure.

They showed that decision tree performed better than the naïve Bayesian with the same crime dataset, using WEKA.

[59]

6

Naïve Bayes

65.59%

Computational speed, robustness, scalability, and interpretability were not taken into consideration.

This paper presented comparative analysis on the accuracy of k-NN, naive Bayescand decision tree algorithms in predicting crimes and criminal actions.

[61]

7

Autoregressive integrated moving average (ARIMA)

The mean absolute error and standard deviation of the model are

(1) Test average test standard deviation is 0.0867 ± 0.0293;

(2) Training average training standard deviation is 0.0413 ± 0.0084.

They have described this model as quite complex compared to others.

This paper is about the cons of fuzzy cognitive maps with respect to time series prediction. The ARIMA uses the auto-correlation parameters.

[62]

8

Regression model

They first took 10 crimes per month the expected forecast for absolute percent error (APE) was 42%. When 20 crimes were taken the expected forecast APE was 28%, and at 25 crimes per month the expected forecast APE was 25%. After 30 crimes they 13.5% error.

This research date backs 20 years.

They aim at predicting crimes 30 days ahead.

They conduct the experiment in Pittsburgh.

[63]

9

SVM

Over 10 months of experiment its accuracy was 84.37%.

The challenge that they indicated that we may face in the future could be to locate the best point at which spatial knowledge is available.

They compare different model to analyze which has the best chance at predicting hotspots.

[64]

10

Random Forrest Regressor

Ninety-seven percent accuracy in predicting crimes

They got this high accuracy in previous recorded crimes to actually predict crimes in real life will be a challenge.

They had divided their data into 2 parts 80% of it was used to train the model and the rest 20% was used to test the model in this the model achieved the score of 90%.

[65]