Ranking Metrics

rank_metrics.py

"""Information Retrieval metrics

Useful Resources:
http://www.cs.utexas.edu/~mooney/ir-course/slides/Evaluation.ppt
http://www.nii.ac.jp/TechReports/05-014E.pdf
http://www.stanford.edu/class/cs276/handouts/EvaluationNew-handout-6-per.pdf
http://hal.archives-ouvertes.fr/docs/00/72/67/60/PDF/07-busa-fekete.pdf
Learning to Rank for Information Retrieval (Tie-Yan Liu)
"""
import numpy as np


def mean_reciprocal_rank(rs):
    """Score is reciprocal of the rank of the first relevant item

    First element is 'rank 1'.  Relevance is binary (nonzero is relevant).

    Example from http://en.wikipedia.org/wiki/Mean_reciprocal_rank
    >>> rs = [[0, 0, 1], [0, 1, 0], [1, 0, 0]]
    >>> mean_reciprocal_rank(rs)
    0.61111111111111105
    >>> rs = np.array([[0, 0, 0], [0, 1, 0], [1, 0, 0]])
    >>> mean_reciprocal_rank(rs)
    0.5
    >>> rs = [[0, 0, 0, 1], [1, 0, 0], [1, 0, 0]]
    >>> mean_reciprocal_rank(rs)
    0.75

    Args:
        rs: Iterator of relevance scores (list or numpy) in rank order
            (first element is the first item)

    Returns:
        Mean reciprocal rank
    """
    rs = (np.asarray(r).nonzero()[0] for r in rs)
    return np.mean([1. / (r[0] + 1) if r.size else 0. for r in rs])


def r_precision(r):
    """Score is precision after all relevant documents have been retrieved

    Relevance is binary (nonzero is relevant).

    >>> r = [0, 0, 1]
    >>> r_precision(r)
    0.33333333333333331
    >>> r = [0, 1, 0]
    >>> r_precision(r)
    0.5
    >>> r = [1, 0, 0]
    >>> r_precision(r)
    1.0

    Args:
        r: Relevance scores (list or numpy) in rank order
            (first element is the first item)

    Returns:
        R Precision
    """
    r = np.asarray(r) != 0
    z = r.nonzero()[0]
    if not z.size:
        return 0.
    return np.mean(r[:z[-1] + 1])


def precision_at_k(r, k):
    """Score is precision @ k

    Relevance is binary (nonzero is relevant).

    >>> r = [0, 0, 1]
    >>> precision_at_k(r, 1)
    0.0
    >>> precision_at_k(r, 2)
    0.0
    >>> precision_at_k(r, 3)
    0.33333333333333331
    >>> precision_at_k(r, 4)
    Traceback (most recent call last):
        File "<stdin>", line 1, in ?
    ValueError: Relevance score length < k


    Args:
        r: Relevance scores (list or numpy) in rank order
            (first element is the first item)

    Returns:
        Precision @ k

    Raises:
        ValueError: len(r) must be >= k
    """
    assert k >= 1
    r = np.asarray(r)[:k] != 0
    if r.size != k:
        raise ValueError('Relevance score length < k')
    return np.mean(r)


def average_precision(r):
    """Score is average precision (area under PR curve)

    Relevance is binary (nonzero is relevant).

    >>> r = [1, 1, 0, 1, 0, 1, 0, 0, 0, 1]
    >>> delta_r = 1. / sum(r)
    >>> sum([sum(r[:x + 1]) / (x + 1.) * delta_r for x, y in enumerate(r) if y])
    0.7833333333333333
    >>> average_precision(r)
    0.78333333333333333

    Args:
        r: Relevance scores (list or numpy) in rank order
            (first element is the first item)

    Returns:
        Average precision
    """
    r = np.asarray(r) != 0
    out = [precision_at_k(r, k + 1) for k in range(r.size) if r[k]]
    if not out:
        return 0.
    return np.mean(out)


def mean_average_precision(rs):
    """Score is mean average precision

    Relevance is binary (nonzero is relevant).

    >>> rs = [[1, 1, 0, 1, 0, 1, 0, 0, 0, 1]]
    >>> mean_average_precision(rs)
    0.78333333333333333
    >>> rs = [[1, 1, 0, 1, 0, 1, 0, 0, 0, 1], [0]]
    >>> mean_average_precision(rs)
    0.39166666666666666

    Args:
        rs: Iterator of relevance scores (list or numpy) in rank order
            (first element is the first item)

    Returns:
        Mean average precision
    """
    return np.mean([average_precision(r) for r in rs])


def dcg_at_k(r, k, method=0):
    """Score is discounted cumulative gain (dcg)

    Relevance is positive real values.  Can use binary
    as the previous methods.

    Example from
    http://www.stanford.edu/class/cs276/handouts/EvaluationNew-handout-6-per.pdf
    >>> r = [3, 2, 3, 0, 0, 1, 2, 2, 3, 0]
    >>> dcg_at_k(r, 1)
    3.0
    >>> dcg_at_k(r, 1, method=1)
    3.0
    >>> dcg_at_k(r, 2)
    5.0
    >>> dcg_at_k(r, 2, method=1)
    4.2618595071429155
    >>> dcg_at_k(r, 10)
    9.6051177391888114
    >>> dcg_at_k(r, 11)
    9.6051177391888114

    Args:
        r: Relevance scores (list or numpy) in rank order
            (first element is the first item)
        k: Number of results to consider
        method: If 0 then weights are [1.0, 1.0, 0.6309, 0.5, 0.4307, ...]
                If 1 then weights are [1.0, 0.6309, 0.5, 0.4307, ...]

    Returns:
        Discounted cumulative gain
    """
    r = np.asfarray(r)[:k]
    if r.size:
        if method == 0:
            return r[0] + np.sum(r[1:] / np.log2(np.arange(2, r.size + 1)))
        elif method == 1:
            return np.sum(r / np.log2(np.arange(2, r.size + 2)))
        else:
            raise ValueError('method must be 0 or 1.')
    return 0.


def ndcg_at_k(r, k, method=0):
    """Score is normalized discounted cumulative gain (ndcg)

    Relevance is positive real values.  Can use binary
    as the previous methods.

    Example from
    http://www.stanford.edu/class/cs276/handouts/EvaluationNew-handout-6-per.pdf
    >>> r = [3, 2, 3, 0, 0, 1, 2, 2, 3, 0]
    >>> ndcg_at_k(r, 1)
    1.0
    >>> r = [2, 1, 2, 0]
    >>> ndcg_at_k(r, 4)
    0.9203032077642922
    >>> ndcg_at_k(r, 4, method=1)
    0.96519546960144276
    >>> ndcg_at_k([0], 1)
    0.0
    >>> ndcg_at_k([1], 2)
    1.0

    Args:
        r: Relevance scores (list or numpy) in rank order
            (first element is the first item)
        k: Number of results to consider
        method: If 0 then weights are [1.0, 1.0, 0.6309, 0.5, 0.4307, ...]
                If 1 then weights are [1.0, 0.6309, 0.5, 0.4307, ...]

    Returns:
        Normalized discounted cumulative gain
    """
    dcg_max = dcg_at_k(sorted(r, reverse=True), k, method)
    if not dcg_max:
        return 0.
    return dcg_at_k(r, k, method) / dcg_max


if __name__ == "__main__":
    import doctest
    doctest.testmod()

Hey, this is a nice implementation for those metrics! Would you mind clarifying the license by sticking a license header on top? Thanks in any case :)

great work!

nicely done!

nice work

great!! and im looking for pairwise AUC

Thanks, nice work!

But NDCG for recommender systems evaluation needs to account whole list of user ratings i.e. ratings that where not included in recommendation delivery. My implementation of recsys NDCG.

Here's my small contribution; nDCG as defined by Kaggle: https://www.kaggle.com/wiki/NormalizedDiscountedCumulativeGain
The only difference is: r → 2^r - 1

import numpy as np


def dcg_at_k(r, k):
    r = np.asfarray(r)[:k]
    if r.size:
        return np.sum(np.subtract(np.power(2, r), 1) / np.log2(np.arange(2, r.size + 2)))
    return 0.


def ndcg_at_k(r, k):
    idcg = dcg_at_k(sorted(r, reverse=True), k)
    if not idcg:
        return 0.
    return dcg_at_k(r, k) / idcg

Hey I am a beginner and I was trying to find NDCG score for my similarity matrix for 5 iterations.But the error coming is" the truth value of an array with more than one element is ambiguous. Use a.any() or a.all()" .Could somebody help me out by what this means?Thanks a ton

I agree with Kell18. That does not only hold for recommender systems but also for IR: the ideal ranking should be the ranking of all judged items in the collection for the query. Not only the retrieved ones.

Can you give me an idea of how to use your function if I have a vector of binary (ground truth) labels and then an output from an ALS model, for example: [ 1.09253478e-01 1.97033856e-01 5.51080274e-01 ..., 1.77992064e-03 1.90066773e-12 1.74711004e-04]

When evaluation my model using AUC, I can just feed in the binary ground truth vector and the output from my ALS model as the predicted scores as is, but I am wondering how this would work with your model if I am considering, for example, k=10 recommendations and would like to use NDCG to evaluate the output.

The reference URL in the comments for ndcg_at_k does not work. I believe this is the current URL for the referenced document: http://web.stanford.edu/class/cs276/handouts/EvaluationNew-handout-6-per.pdf

There is a typographical error on the formula referenced in the original definition of this function:
http://www.stanford.edu/class/cs276/handouts/EvaluationNew-handout-6-per.pdf
log2(i) should be log2(i+1)

The formulas here are derived from
https://en.wikipedia.org/wiki/Discounted_cumulative_gain#Discounted_Cumulative_Gain

@gumption find this issue, see more:
https://gist.github.com/gumption/b54278ec9bab2c0e0472816d1d7663be

+1 to kell18 and suzanv. quote from wiki on ideal DCG:

sorting all relevant documents in the corpus by their relative relevance, producing the maximum possible DCG through position p, also called Ideal DCG (IDCG) through that position.

how about recall@K ?

@bwhite
I think you have an error in the average_precision metric.

So for example:

from rank_metrics import average_precision

relevance_list = [[1, 1, 1], [1, 1, 0], [1, 0, 0]]

for r in relevance_list:
    print(average_precision(r))

Will print:

1.0
1.0
1.0

Instead of:

1.
0.6666666666666666
0.3333333333333333

So in the metric's return you should replace np.mean(out) with np.sum(out) / len(r)

there is bug in dcg_at_k

print(ndcg_at_k([14, 2, 0, 0], 5)) # output would be 1
print(ndcg_at_k([2, 14, 0, 0], 5)) #output would be 1

correct dcg_at_k would be

def dcg_at_k(r, k, method=0):
    r = np.asfarray(r)[:k]
    if r.size:
        if method == 0:
            return r[0] + np.sum(r[1:] / np.log2(np.arange(3, r.size + 2))) ### fix here
        elif method == 1:
            return np.sum(r / np.log2(np.arange(2, r.size + 2)))
        else:
            raise ValueError('method must be 0 or 1.')
    return 0.

@bwhite
I think you have an error in the average_precision metric.

So for example:

from rank_metrics import average_precision

relevance_list = [[1, 1, 1], [1, 1, 0], [1, 0, 0]]

for r in relevance_list:
    print(average_precision(r))

Will print:

1.0
1.0
1.0

Instead of:

1.
0.6666666666666666
0.3333333333333333

So in the metric's return you should replace np.mean(out) with np.sum(out) / len(r)

+1

@bwhite
I think you have an error in the average_precision metric.

So for example:

from rank_metrics import average_precision

relevance_list = [[1, 1, 1], [1, 1, 0], [1, 0, 0]]

for r in relevance_list:
    print(average_precision(r))

Will print:

1.0
1.0
1.0

Instead of:

1.
0.6666666666666666
0.3333333333333333

So in the metric's return you should replace np.mean(out) with np.sum(out) / len(r)

@lucidyan , @cuteapi

The code is correct if you assume that the ranking list contains all the relevant documents that need to be retrieved. In your example, the query with ranking list r=[1,0,0] retrieves 3 documents, but only one is relevant, which is in the top position, so your Average Precision is 1.0. Note that Mean Average Precision assumes that each query is independent of each other, and in your example, there is no reason to believe that every query has to retrieve always 3 relevant documents. For your example r=[1,1,0] and r=[1,0,0] the relevant documents are 2 and 1 respectively, because the code assumes that the total number of 1's in your list is the total number of relevant documents (there are supposed to be no misses in the ranked list).

This is a strong assumption in this code, but it does not make the implementation incorrect. You need to be aware that the ranking list that you pass has to contain all the positions where relevant documents appear. If that is not the case, you need to use another implementation that takes into account recall, which is the missing piece in this code.

is IDCG calculated across all the queries or IDCG for each query for calculating NDCG?

Has anyone made this into a pypi package? If not @bwhite, would you mind if I went ahead and made it into one? I'd of course accredit you. I'm just tired of copying and pasting this code because it is super useful haha

Made this into a cute little pypi package if anyone is interested: https://github.com/ncoop57/cute_ranking

i'm trying to learning about ndcg, but there's one thing i don't understand about the code. What is "method" variable ? is that something that produce same two maximum weights ?

method variable distinguishes between two ways of giving weights to relevances while calculating DCG