strange regression coefficients

[vohtaski]

I am trying to fit the polynomial line into the data-set below as polynomial with order or 4 and 10,
and all coefficients are 0. Is it normal? Polynomial of high degree should normally over-fit the data,
but here it is just a line.

regression version: 2.0.0

// [[x,y]]
var data = [[0,90],[45,93],[90,92],[135,90],[180,89],[225,90],[270,88],[315,89],[360,85],[405,85],[450,90],[495,94],[540,87],[585,89],[630,91],[675,94],[720,88],[765,92],[810,76],[855,81],[900,83],[1215,94],[1260,91],[1305,85],[1350,83],[1395,85],[1440,82],[1485,81],[1530,83],[1575,82],[1620,84],[1665,84],[1710,87],[1755,87],[1800,89],[1845,86],[1890,83],[1935,88],[1980,78],[2025,78],[2070,66],[2115,59],[2250,87],[2295,94],[2430,87],[2475,88],[2520,88],[2565,87],[2700,90],[2745,93],[2790,91],[2835,86],[2880,89],[2925,88],[2970,90],[3015,89],[3060,90],[3105,88],[3150,85],[3195,91],[3240,91],[3285,85],[3330,87],[3375,87],[3420,89],[3465,93],[3510,87],[3555,90],[3600,92],[3645,91],[3690,90],[3735,88],[3780,78],[3825,90],[3870,93],[3915,91],[3960,91],[4005,90],[4050,89],[4095,94],[4140,88],[4185,89],[4230,82],[4275,77],[4320,79],[4365,103],[4635,111],[4680,88],[4725,74],[4770,81],[4815,87],[4860,80],[4905,87],[4950,80],[4995,85],[5040,87],[5085,89],[5130,91],[5175,87],[5220,90],[5265,90],[5535,82],[5580,73],[5625,87],[5670,94],[5715,95],[5760,100],[5805,100],[5850,98],[5895,101],[5940,107],[6030,95],[6075,92],[6120,94],[6165,88],[6210,86],[6255,85],[6300,89],[6345,96],[6390,94],[6435,87],[6480,89],[6525,89],[6570,93],[6615,86],[6660,90],[6705,85],[6750,88],[6795,85],[6840,86],[6885,89],[6930,85],[6975,91],[7020,84],[7065,91],[7290,89],[7560,94]];

regression.polynomial(data, {order: 4})
// "y = 0x^4 + 0x^3 + 0x^2 + -0.01x + 92.01"
regression.polynomial(data, {order: 10})
// "y = 0x^10 + 0x^9 + 0x^8 + 0x^7 + 0x^6 + 0x^5 + 0x^4 + 0x^3 + 0x^2 + -0.01x + 91.64"

Data looks like this:

[vohtaski]

I just checked the version 1.4.0 and it's working on the data above:
regression('polynomial' data, 4)

regression('polynomial' data, 10)

[gtjamesa]

I have been issues with polynomial on 2.0.0 - thanks for posting that version 1.4.0 works :)

[Tom-Alexander]

Can I please have more information about the discrepancies between 2.0.0 and 1.4.0? The coefficients are the same between versions for me. This could be caused by rounding as 2.0.0 rounds by default. It can be configured with the precision option.

[billy0059]

I had a similar issue in the "linear" method, but the problem was caused from the parameters passing.
The parameter array I passed is considered as a string array, so the addition operation sum[0] += data[n][0]; sum[1] += data[n][1]; would lead to a false result.
Simply modified to sum[0] += parseInt(data[n][0]); sum[1] += parseInt(data[n][1]); solved my problem.

[lucamattiazzi]

i'm having the same problem, I too solved it with 1.4.0, but the explanation from @billy0059 is not clear to me, why was the value casted as string?
(btw, a regression in the new version was inevitable 😄)

[Eugeny]

Here's another example of 2.0.0 providing weird results:

Input: [[1000,100],[101000,1000],[201000,2000],[301000,2500],[401604,2700.3333],[502208,2800],[602208,2900],[702208,2950],[802208,3000],[902208,3500]]

Output: y = 0x^4 + 0x^3 + 0x^2 + 0.01x + 52.61

[cboden]

Tom-Alexander was right. To get 2.0 matching with 1.4 I just had to update my precision:

regression.polynomial(data, {order: 4, precision: 20});