Enlightenment of stay home in conjunction with Shibuya Ward official virtual Shibuya Halloween project
The "Shibuya Data Consortium" led by Future Design Shibuya (hereinafter referred to as Future Design Shibuya) is a virtual Shibuya Halloween project officially recognized by Shibuya Ward so that you can enjoy Shibuya Halloween, which is also a social issue, at your stay home. Using the human flow data provided by KDDI Co., Ltd. (hereinafter referred to as KDDI) and deep learning, and using the AI model built by GAUSS Co., Ltd., we will analyze and announce the human flow forecast data around Shibuya Station on the day of Halloween. did.
<Analysis summary>
・ Overall, it is predicted that a total of 1.77 million people (*) will stay and move to Hachiko Mesh during the Halloween period.
・ Compared to 2020, the Halloween period is expected to be 137.6%, and the last two days are expected to be 140.4%.
About forecast requirements
Based on the 26-month demographic data provided by KDDI CORPORATION from July 2019 to August 2021, the 125m mesh around Shibuya Hachiko is expected to be particularly crowded during the Halloween period (10 / 25-10 / 31). ) People flow forecast.
〇 Data provided by KDDI CORPORATION (KDDI Location Data)
Statistical data processed and aggregated in units such as time zone and standard area mesh for GPS location information, etc. that is periodically accumulated from au smartphone terminals with individual consent in any target period and target area. The service to be provided.
https://k-locationdata.kddi.com/
〇 Data requirements
・ Target area Around Hachiko (125m square)
・ Target period July 2019-August 2021 (for 26 months)
・ Data type Gender Age-based hourly staying / moving population
Vital data of the target Hachiko Mae mesh code (52293596111) and other data such as weekdays / Saturdays, Sundays, and holidays, events, weather / temperature, emergency declaration, number of corona infected people, etc., which are considered to be related to human flow, are used as input data. Using deep learning, we predicted the gender of the Hachiko pre-mesh code during the Halloween period (10 / 25-10 / 31), and the hourly staying population and mobile population by age group.
About prediction method
It was done in five steps: "data collection", "data preprocessing", "model learning", "test model selection", and "execution of prediction".
① Data collection
Collect data from various sources in CSV and JSON format for input data.
② Data preprocessing
In the input data, normalize the number of people, one-hot encoding, date / time conversion. In addition, feature quantity engineering is performed and unnecessary items are deleted. Execute data organization of input data.
③ Model learning
A total of 8 patterns of various types of NN (neural network) models and ML (opportunity learning) models are created and learned. Adjust hyperparameters and layers to improve accuracy.
* Creation model
1. Stacked LSTM
2. CNN LSTM
3. Vanilla LSTM
4. CNN
5. Linear Regression
6. Ridge
7. Gradient Boosting
8. KNN regressor
④ Test model selection
The input data is divided into 95% training data and 5% test data, and the test is executed.
Select the best model using MAE (Mean Absolute Error) as an evaluation index.
As a result, the CNN LSTM model, which resulted in MAE 7.61, was selected as the optimal predictive model for an average of 788.77 people.
⑤ Execution of forecast
Optimal Predictive Model Execute forecasts with CNN LSTM and create output data.
Prediction result
■ Overall
Overall, it is estimated that a total of 1.77 million people (*) will stay and move to Hachiko Mesh during the Halloween period. Compared to 2020, the Halloween period is expected to be 137.6% and the last two days is expected to be 140.4%.
<Halloween period (10 / 25-10 / 31)>
<Last 2 days (10 / 30-10 / 31)>
■ By time zone
In the forecast by time zone, 18:00 is the forecast of the peak of human flow.
■ By age
By age group, people in their twenties are the most common, with a Halloween period of 139.7% compared to 2020 and a forecast of 144.8% for the last two days.
<Halloween period (10 / 25-10 / 31)>
<Last 2 days (10 / 30-10 / 31)>
In the future, we will compare the forecast values with the same data on the day and continue activities to improve the accuracy of the forecast toward next year, and will continue to contribute to solving social issues. We also plan to hold a SOCIAL INNOVATION WEEK 2021 (https://social-innovation-week-shibuya.jp/) verification session and solicit ideas to improve accuracy for the next fiscal year.
Person flow data provided:
KDDI CORPORATION https://www.kddi.com/
KDDI Location Data https://k-locationdata.kddi.com/
Data analysis:
GAUSS Co., Ltd. https://gauss-ai.jp/
* Regular member of Japan Deep Learning Association
[GAUSS Co., Ltd.]
An AI startup company founded in 2017. With the mission of "creating the future you imagine," we will utilize AI in various industries to provide new value to society. In 2018, we released the AI platform "GAUSS Foundation Platform" that enables easy AI development, and it is currently used in many solution platforms. Among them, "GAUDi EYE", a video solution for the construction and manufacturing industries, provides new value to many sites including companies listed on the First Section of the Tokyo Stock Exchange. In addition, "AI Horse Racing Prediction SIVA", which predicts horse racing using machine learning technology, is a typical AI that has been involved in research and development since its establishment.
GAUSS Co., Ltd. https://gauss-ai.jp/
