Smart Phone Location Predictions using WiFi Signal Data

Build an algorithm to predict an individual's location using their Smart Phone and WiFi signals? Do so with a relatively high level of accuracy, precision, and recall per Client Minimum Specifications?

Check and check.

• See "Location Predictions Analysis" above for Python Pipeline Analysis.

Background

My client is developing a system to help people navigate a complex, unfamiliar interior space on a college campus. They want me to investigate the feasibility of using WiFi fingerprinting to determine a person's indoor location. If a model meets or exceeds minimum specifications, it will be incorporated into a Smart Phone app for indoor locationing on a University campus.

Business Objectives

1. Build multiple different models to predict a person’s location in indoor college campus spaces using Wifi signal data
2. Answer the business question, “Can a model be built that predicts indoor location on a college campus that meets the Client’s minimum specifications?”
3. If a model can be built meeting minimum specifications, deploy algorithm into smart phone indoor locationing app.

Skills

• Python/Jupyter Notebook IDE
• Multiple machine learning algorithms: Support Vector Machines, Decision Trees, Random Forest, K Nearest Neighbors
• Understanding the business problem and answering the business questions (classification vs. regression)
• Feature Engineering
• Sampling techniques and feature selection strategies for large dataset
• Dimensionality reduction/ zero variance variable removal
• Cross validation, parameter tuning, and post-resample comparison of model metrics for highly multi-class classification problem
• Using metrics and recall investigation for determining optimal algorithm(s) for model deployment
• Business objectives solved

Client Minimum Specifications

• Indoor location must be as precise as predicting within 10-15 feet of the indoor room, also defined as ‘SpaceID’ within source data. Relative position, or whether individual is outside or inside of room, is unnecessary.
• Performance metrics ideal for deployment:
  • Accuracy scores on test data reaches 80% or higher
  • Precision (accuracy of minority class) on test data reaches 80% or higher
  • Recall (coverage of minority class) on test data is 80% or higher
  • F1 Score for multi-class problem achieves 80% or higher

Data Description

Data was collected by 20 individuals using mobile phone devices on a college campus in Velencia, Spain. The data source is the UJIIndoorLoc WLAN database

• 19937 observations of 529 variables
• 520 of the variables (98% of dataset) are homogenous wifi-access points providing numeric values representing signal strength
• 9 remaining variables are Longitude, Latitude, Floor, BuildingID, SpaceID, Relative Position, UserID, PhoneID, and Time
• Contains no missing values

Summary and Final Model Recommendation

• Random Forest algorithms overall met or exceeded client minimum specification metrics for predicting WiFi locations on a college campus.
• Recall for location predictions on individual building datasets was higher than OOB dataset.
• Recommend deploying Random Forest algorithms by individual buildings for Indoor Locationing smart phone app

Post-resample Performance Comparison of All Models

• Random Forest outperformed Support Vector Machines metrics in OOB and by individual building datasets.

Absolute Recall Performance Comparison of Random Forest Models

• Higher recall is better
• Total locations in 75-100% recall ranges:
• Oob2: 456
• Buildings 0, 1, 2: 483
• 27 more locations in high recall range for individual buildings versus full dataset