Process Melbourne Data using Methods presented in IJCAI15 Paper

1. Dataset
   • Simple Facts
   • Photo Scatter Plot
   • Issues and Solutions
2. Results
   • Precision, Recall and F1-score
   • Transition Matrix

NOTE: Please view this page via IPython Notebook Viewer Service, otherwise the within-page links may not work properly.

1. Dataset ⇈

Photos:

The photos were selected from YFCC100M dataset, while Melbourne's Geo-Coordinates is 37°48′49″S 144°57′47″E, data in a square, from (39.5S, 140.9E) to (35.5S, 148.5E) with accuracy = 16 are used, the total number of photos is 87,362.

POIs:

POIs are from OpenStreeMap, e.g. downloading data from one of these mirrors

clipping a bounding box of Melbourne: [140.9,-38.7, 148.5, -35.5], e.g. clipping using one of these tools

then filtering interested POI tags (described in the table below) from the tag list, the total number of POIs is 3360.

Python scripts for filtering POI tags from the clipped data using this python library is filter_node.py, e.g.

file Melb_tags.list is available here.

1.1 Simple Facts ⇈

Some simple facts of Melbourne data as well as data of four other cities used in ijcai15 paper are summaried in the table below.

The distribution of sequence length for each city was shown below.

1.3 Photo Scatter Plot ⇈

Photo Scatter Plot for Edinburgh | Glasgow | Osaka | Toronto | Melbourne

1.4 Issues & Solutions ⇈

Q: Picking POIs is a somewhat hard task

POIs picked manually according to photo scatter plot are much better than the results of k-means clustering/kernel density estimation, but still not good enough

A: With the help of OpenStreeMap and NationalMap/Google Maps, it would be much easier to select and visualize POIs.

Further processing the POI data:

1. filtering the list of POIs further, such as POIs too close, POIs with too few photos etc. (how to?)
2. classify/label each POI with assistance of online maps and its original tag. (manually?)

NOTES:

• Both NationalMap and GoogleMaps support headers in the first line, specify longitude and latitude in the first line when using NationalMap
• Satellite images from GoogleMaps are better while NationalMap does not restrict the number of marks (GoogleMaps restrict <= 2000 marks each layer), marks on NationalMap will float around sometimes when dragging the map

Q: How to deal with POIs that are too close, e.g. 0-10m?

A:

Q: POIs are generally associated with multiple labels

how to define these labels? how to label each POI?

A:

Q: Assign photo to a POI

if their distance is less than 200m according to paper seems not to be a good idea, as

• if POI is something not large, e.g. buildings, 200m seems OK
• if POI is something large, e.g. natural park, 1-2km seems to be a reasonable distance
• but, we don't know the type of POI when assigning as picking POI and assigning photos are being done at the same time

A: Assign a photo to the nearest POI if the distance between the two is less than, say 500m?

Q: Travel sequences independence assumption seems to be implausible

Users' travel sequences are generated by splitting travel history of users if their consecutive POI visits occur more than 8 hours, while a common travelling spans several days, which could be represented by several travel sequences with dependence (e.g. user preference patterns: beach-park-shopping, beach-beach-shopping etc.)

A:

2. Results ⇈

2.1 Precision, Recall and F1-score ⇈

Settings: Melbourne, $\eta$=0.5 with time-based user interest and POI popularity, 28/110 ≈ 25.5% solutions are suboptimal, leave-one-out

Box plot of Recall, Precision and F1-score

Settings: Melbourne, $\eta$=0.0 with POI popularity only, 29/110 ≈ 26.4% solutions are suboptimal, leave-one-out

Box plot of Recall, Precision and F1-score

2.2 Transition Matrix ⇈

Transition matrix for recommended sequences, $\eta$ = 0.5:

Transition matrix for recommended sequences, $\eta$ = 0.0:

Transition matrix for actual sequences: