Improving land use inference by factorizing mobile phone call activity matrix

Land use is shaped by human activities. Traditional methods of measuring land uses (e.g. surveys and remote sensing techniques) often have difficulties in capturing human activities. The near-ubiquitous coverage of mobile phones opens up a new way to investigate land use through human activities. We propose to analyze land use by characterizing human activity patterns based on the aggregated call volume, and apply non-negative matrix factorization to identify fundamental behavioral classes. Using tower-based call data from Dakar, Senegal, we discover two fundamental land use patterns: commercial/business/industrial (C/B/I) and residential. Then, the land use of the reception area of each cell tower can be inferred based on the weights obtained for each basis vector. To evaluate the proposed approach, the results are compared with two points-of-interest (POI) data sets obtained from OpenStreetMap and Facebook's Graph API. We have found that a majority of POIs like embassies, offices, and hotels are located in the predicted C/B/I areas; specifically, there is a strong positive correlation between estimated land use weights and the number of related POIs. Furthermore, we have shown analyzing 24-h call pattern matrix can track daily land use changes.