The wifi task is one of the last projects of the data analytics curriculum.
When you reach it, you will see that the research is composed of a train set and validation set, both labelled in order to train and evaluate your model.
But to make things more interesting, you will be evaluated by a different set, called test set. Here is the description of it:
The private test dataset is blind and all the data about the user and location have been removed. It is the dataset that we provided to the 2015 EvAAL-ETRI competitors at the IPIN conference to evaluate their IPS.
This private test set is a CSV file with the same structure as the public UJIIndoorLoc samples:
- 001-520 RSSI levels
- 521-523 Real world coordinates of the sample points
- 524 BuildingID
- 525 SpaceID
- 526 Relative position with respect to SpaceID
- 527 UserID
- 528 PhoneID
- 529 Timestamp
Fields 521 to 527 have been filled with 0’s (Anonimized testing)
Field 528 contains new values:
- value 25: Nexus 5 with 5.0.1
- value 26: Orange Rono with 4.4.2
- value 27: D2303 with 4.4.4
- value 28: Wildfire S A510e with 4.2.2
- value 29: GT-I9505 with 4.4.2
You will receive this set one day after the presentation. Ask your mentor to send you it when it’s due.
In order to submit your results, create a csv with 3 columns: LATITUDE, LONGITUDE, FLOOR (in this order). Make sure to create these columns in this sequence and without quotes. Also, make sure that your csv doesn’t contain a “index column” and that your submission has 5180 rows.
Your mentor will create a folder on the google drive with your name and you should include your submission in that folder. You can have up to 5 different files inside the folder, containing different strategies that you want to be evaluated.
The name of the csv file is up to you, but you might want to create a tag of your strategy, such as “Andre_rf_pca”.
As this is a blind test, the results might take a couple of days to be evaluated.
The competition organizers evaluate results based on the positioning error which is computed in two dimensions (latitude and logitude) as the euclidean distance between the real and estimated locations. Penalties are added for floor error (4m) and building error (50m):
error = distanceˡᵃᵗ,ˡᵒⁿ(estimated, real) + 4 ∗ abs(estimated_floor – real_floor) + 50 ∗ (estimated_bld != real_bld)
Finally the metric used in the competition was the third quartile (75th percentile) of the error in all test samples.
Any questions around this point, ask your mentor or send an email to Gabriel@ubiqum.com (mentor Barcelona)