stanford-futuredata / asap

ASAP: Prioritizing Attention via Time Series Smoothing

ASAP: Prioritizing Attention via Time Series Smoothing

Please consult the ASAP demo site or our VLDB 2017 paper for details.

asap's Issues

Python 3 does not have a csv.next() method

the csv reader code appears to be broken here at ASAP.py#L222 because '_csv.reader' object has no attribute 'next' in Python 3. Instead the code on L222 should read args._head = next(icsv) for Python 3.

An easy fix is to either call sys.version_info to check for Python version, or at least include a note stating that this code depends on a Python 2 version of the csv package.

FYI: ASAP in papers we love

💥 papers-we-love/papers-we-love#446

Timeseries data consistency

How to cope when shrinking a timeseries dataset?

I noticed that:
x = Array.apply(null, Array(y.length)).map(function (_, i) {return i;}) in index.html only gives us the indexes related to y`s length

I tried doing something like this but I doubt its correctness..

smooth_val = 100
var y = smooth(rsdata["values"], smooth_val)
step - Math.ceil(rsdata["timestamps"].length/smooth_val)
x = rsdata["timestamps"].filter((x,i) => i%step == 0)

does anyone have a solution?

if (largestFeasible > 0) check is useless

ASAP/ASAP-optimized.js

Line 248 in 24e95b2

var largestFeasible = -1;

ASAP/ASAP-optimized.js

Line 267 in 24e95b2

if (largestFeasible > 0) {

or there is a bug?

Python Script Giving Different Results from Demo Site

I created some fake data like this:

n = 201
x = np.linspace(0, 1, n)
y = np.sin(4*np.pi*x*x)
chirp = 0.5*np.sin((2.0*np.pi*x)/0.05)
chirp_inx = np.argwhere((x > 0.4) & (x < 0.45))
y[chirp_inx] = chirp[chirp_inx]
noise = np.random.normal(0, 0.1, n)
y_noise = y + noise

So, my data is in y_noise. The smoothed values generated from the Python script:

ASAP.smooth(y_noise)

is different from the values shown from the demo site. You can tell because the smoothed values generated from the Python script produces a min-value of 0.010680608603816815 and a max-value of 0.48671462099352242. Whereas the demo site produces a plot that has a min-value near -1.0 and a max-value of +1.0.

Moving average implementation appears incorrect

I was looking into why the numpy version of SMA didn't match the for loop version, and looking at each step in the for loop version, it appears that it is not doing the final window correctly. Is this intentional or am I misinterpreting what is going on? I haven't tested the JS version, but appears to be identical to the python one.

In [1]: x = [42,75,3,5,99,22,88]


In [2]: SMA(x, 3, 1)
window: [42, 75, 3] 	 sum(  120.0000) / count(    3.0000) =    40.0000  
window: [75, 3, 5] 	 sum(   83.0000) / count(    3.0000) =    27.6667  
window: [3, 5, 99] 	 sum(  107.0000) / count(    3.0000) =    35.6667  
window: [5, 99, 22, 88] 	 sum(  214.0000) / count(    4.0000) =    53.5000  
window: [99, 22, 88, 88] 	 sum(  297.0000) / count(    4.0000) =    53.5000  
Out[2]: [40.0, 27.666666666666668, 35.666666666666664, 53.5]

def SMA(data, _range, slide):
    ret = []
    s = 0.0
    c = 0.0
    window_start = 0
    window = []
    for i in range(len(data)):
        if i-window_start >= _range or i==len(data)-1:
            if i==len(data)-1 or c==0:
                s += data[i]
                window.append(data[i])
                c += 1
            ret.append( s/c )
            print("window: {} \t sum({:10.4f}) / count({:10.4f}) = {:10.4f}  ".format(window, s, c, ret[-1]))
            old_start = window_start
            while window_start < len(data) and window_start-old_start < slide:
                s -= data[window_start]
                window = window[1::]
                c -= 1
                window_start += 1
        s += data[i]
        window.append(data[i])
        c += 1
    print("window: {} \t sum({:10.4f}) / count({:10.4f}) = {:10.4f}  ".format(window, s, c, ret[-1]))
    return ret