On information retrieval metrics designed for evaluation with incomplete relevance assessments
 2.2k Downloads
 34 Citations
Abstract
Modern information retrieval (IR) test collections have grown in size, but the available manpower for relevance assessments has more or less remained constant. Hence, how to reliably evaluate and compare IR systems using incomplete relevance data, where many documents exist that were never examined by the relevance assessors, is receiving a lot of attention. This article compares the robustness of IR metrics to incomplete relevance assessments, using four different sets of gradedrelevance test collections with submitted runs—the TREC 2003 and 2004 robust track data and the NTCIR6 Japanese and Chinese IR data from the crosslingual task. Following previous work, we artificially reduce the original relevance data to simulate IR evaluation environments with extremely incomplete relevance data. We then investigate the effect of this reduction on discriminative power, which we define as the proportion of system pairs with a statistically significant difference for a given probability of Type I Error, and on Kendall’s rank correlation, which reflects the overall resemblance of two system rankings according to two different metrics or two different relevance data sets. According to these experiments, Q′, nDCG′ and AP′ proposed by Sakai are superior to bpref proposed by Buckley and Voorhees and to RankBiased Precision proposed by Moffat and Zobel. We also point out some weaknesses of bpref and RankBiased Precision by examining their formal definitions.
Keywords
Evaluation metrics Relevance assessments Test collections Incompleteness1 Introduction
An information retrieval (IR) test collection comprises a document collection, a set of search requests and a set of manually judged relevant documents for each request. Following TREC ^{1} parlance, hereafter the search requests will be referred to as topics, and the set of relevant documents will be referred to as qrels. The methodology of using a test collection for evaluating and comparing IR techniques was established in the 1960s, through the Cranfield 2 Test by Cleverdon (1967). Since then, laboratory experiments using test collection have always played a central role for the progress of IR techniques, as they are objective, efficient and repeatable (Voorhees 2002).
Now, in the twentyfirst century, evaluation using test collections is still a necessity for most IR researchers. However, the 1990s actually saw a major departure from the original Cranfield 2 experiments, with the advent of TREC, NTCIR ^{2} and CLEF ^{3} evaluation efforts for constructing very largescale test collections. On the surface, the main difference between the Cranfield 2 test collection and the modern test collections is the document collection size: The Cranfield 2 test collection contained only 1,400 documents; The TREC and NTCIR test collections typically contain between half a million to one million documents. However, a more important difference is that while the small scale Cranfield 2 collection had complete relevance assessments, the modern test collections do not: It is simply not feasible to examine the documents exhaustively for these large scale collections.
 1.
Participants submit their “runs” (collections of ranked lists for each topic, where each ranked list usually contains up to 1,000 documents) to the organisers;
 2.
For each topic, organisers take the top k (typically 100) documents from some of the submitted runs and obtain a list of unique documents, i.e, a document pool;
 3.
For each topic, assessors judge the relevance of all documents within the pool.
Hence, despite the fact that pooling is a very efficient way of collecting relevant documents, qrels formed through pooling are possibly incomplete. That is, there may exist relevant documents within the document collection, which none of the participating systems managed to retrieve and therefore are outside the qrels (Voorhees 2002).
 (a)
Is it possible to reliably compare two participating systems and judge which is superior? Were the test collection less incomplete, would this judgement be the same?
 (b)
Is it possible to reliably compare a participating system and a new system that never contributed to the pool? How about two new systems? In short, is the incomplete test collection reusable?
For these reasons, IR evaluation using incomplete relevance assessments is receiving more attention than ever.
One obvious approach to tackling these problems is to devise IR effectiveness metrics that are robust to relevance data incompleteness: We say that an IR metric is robust to incompleteness if system comparison results based on an incomplete set of relevance data are similar to those based on a less incomplete one. This article follows this approach, and more specifically, addresses the issues mentioned in (a) above. Below, we discuss three existing studies that are directly related to the present one. We shall discuss other related work in Sect. 2.
Buckley and Voorhees (2004) proposed an IR evaluation metric called bpref (binary preference) which is highly correlated with Average Precision (AP) when full relevance assessments are available and is yet more robust when the relevance assessments are reduced. Recent TREC tracks have used this metric along with AP. Bpref penalises a system if it ranks a judged nonrelevant document above a judged relevant one, and is indepedendent of how the unjudged documents are retrieved.
More recently, Moffat et al. (2007) introduced an IR evaluation metric called RankBiased Precision (RBP) which they claimed is suitable for evaluation with incomplete relevance data. RBP assumes that the probability that the user moves from a document at Rank r to Rank (r + 1) is a constant p, regardless of the relevance (level) of the document at Rank r. As it does not have a recall component, adding more relevant documents to the qrels always increases the RBP score.
Sakai (2007a) reported that applying Qmeasure (or simply, Q), AP and normalised Discounted Cumulative Gain (nDCG) to a condensed list, i.e., a ranked list of documents obtained by removing all unjudged documents from the original list, is a simpler and a better solution than bpref for handling relevance data incompleteness. The metrics applied to condensed lists will hereafter be referred to as Q′, AP′ and nDCG′, respectively.
This article compares the robustness of Q(′), AP(′), nDCG(′), bpref and RBP to incomplete relevance assessments, using four different sets of gradedrelevance test collections with submitted runs—the TREC 2003 and 2004 robust track data and the NTCIR6 Japanese and Chinese IR data from the crosslingual task. We believe that evaluating IR systems using graded relevance is important for the progress of IR because, if one adheres to IR evaluation based on binary relevance, it would be very difficult for him to devise an IR algorithm that can retrieve highly relevant documents on top of partially relevant ones. Following previous work, we artificially reduce the original relevance data to simulate IR evaluation environments with extremely incomplete relevance data. We then investigate the effect of this reduction on discriminative power (Sakai 2006b, 2007b), which we define as the proportion of system pairs with a statistically significant difference for a given probability of Type I Error, and on Kendall’s rank correlation (Voorhees 2001)), which reflects the overall resemblance of two system rankings according to two different metrics or two different qrels. According to these experiments, Q′, nDCG′ and AP′ are superior to bpref and RBP.
This article generalises a recent study by Sakai (2007a), in that (1) While he used the NTCIR3 and NTCIR5 Japanese/Chinese data, we use TREC robust track data and NTCIR6 Japanese/Chinese data to obtain more general and substantial conclusions; (2) We compare RBP with the other metrics, after discussing some properties of the metrics that immediately follow from their formal definitions.
The remainder of this article is organised as follows. Section 2 provides an overview of related studies. Section 3 defines and discusses the characteristics of AP(′), Q(′), nDCG(′), bpref and RBP. Section 4 describes the TREC and NTCIR data we used for comparing the robustness of these metrics to relevance data incompleteness. Section 5 compares the metrics in terms of discriminative power based on statistical significance tests. Section 6 compares the metrics in terms of Kendall’s rank correlation between the entire system rankings. Finally, Sect. 7 concludes this article.
2 Related work
This section provides an overview of previous work related to the present study.
There are at least two approaches to tackling the relevance data incompleteness problem: One is to try to construct a better test collection more efficiently, and another is to devise or choose reliable IR metrics, given a test collection. Along the first line of research, methods for creating judgment pools efficiently were proposed by Cormack et al. (1998) and by Zobel (1998). Soboroff et al. (2001) proposed a method for ranking systems without any relevance assessments, but subsequently Aslam and Savell (2003) pointed out that the method tends to rank them by “popularity” rather than performance. More recently, Carterette et al. (2006) analyzed the distribution of AP over all possible assignments of relevance to all unjudged documents and proposed a method to construct a test collection with minimal relevance assessments; Büttcher et al. (2007) proposed to “expand” existing relevance assessments by treating them as training data for machine learning.
This article takes the latter approach, of choosing reliable IR metrics for handling relevance data incompleteness. The proposals of the aforementioned bpref (Buckley and Voorhees 2004), RBP (Moffat et al. 2007) and Q′, nDCG′ and AP′ (Sakai 2007a) fall into this category. Also along this line, Aslam et al. (2006) and Yilmaz and Aslam (2006) proposed Induced AP, Subcollection AP and Inferred AP. Induced AP is exactly what we call AP′. We do not consider Subcollection AP and Inferred AP in our present study, because (1) While the goal of Yilmaz and Aslam was to estimate the true AP values, ours is not: We prefer to explore different metrics, especially those that can handle graded relevance; (2) Both Subcollection AP and Inferred AP require knowledge of the pooled but unjudged documents, which limits their applicability; ^{4} (3) According to Bompada et al. (2007), Inferred AP is not as robust as the original nDCG for evaluation with incomplete relevance data.
Another metric proposed for handling incomplete relevance data, called RankEff (Grönqvist 2005), has been examined by Büttcher et al. (2007). However, Sakai (2008b) points out that RankEff is in fact equivalent to an existing variant of bpref called bpref_N, also known as “bpref_allnonrel” implemented in trec_eval, the standard IR evaluation software for TREC. Sakai (2007a) showed both analytically and empirically that bpref_N is not a good evaluation metric. See Sect. 3.2 for more discussions.
Büttcher et al. (2007) also used Precision at l judged documents, which relies on condensed lists just like Q′, AP′ and nDCG′. However, Precision is not a satisfactory metric for us because: (1) It ignores the ranks of retrieved relevant documents; (2) It does not average well, especially with a large document cutoff; (3) With a small document cutoff, it gives unreliable results as systems are evaluated based on a small number of observations, i.e., documents near the top of the ranked list (Sakai 2007f).
De Beer and Moens (2006) proposed rpref, a gradedrelevance version of bpref. Sakai (2007a) pointed out that it has minor bugs, and proposed rpref_relative2 by fixing them. However, he reported that it does not have any advantage over Q′, AP′ and nDCG′ despite its complexity.
Sakai (2007c) conducted a study similar to the present one, but focussed on the task of finding one highly relevant document. He showed that the application of Reciprocal Rank, Omeasure (Sakai 2006c) and P+measure (Sakai 2006a, 2007e) to condensed lists is an effective way of handling the relevance data incompleteness problem.
3 Formal definitions of the IR metrics
This section formally defines the IR metrics we consider, namely, Q(′), AP(′), nDCG(′), bpref and RBP, and also discusses their properties that immediately follow from the definitions. Among these metrics, only Q(′), nDCG(′) and RBP can handle graded relevance.
3.1 Q, AP and nDCG
Let \({\mathcal{L}}\) denote a relevance level, and let \(gain({\mathcal{L}})\) denote the gain value for retrieving an \({\mathcal{L}}\)relevant document for a particular topic. Following the NTCIR tradition (Kando 2007), this article assumes that we have Srelevant (i.e., highly relevant), Arelevant (i.e., relevant) and Brelevant (i.e., partially relevant) documents. Other documents, i.e., judged nonrelevant documents and unjudged documents, are considered nonrelevant and therefore do not carry a gain value. We let gain(S) = 3, gain(A) = 2 and gain(B) = 1 hereafter as metrics such as Q and nDCG are robust to the choice of gain values (Sakai 2007f). As for the TREC data, which only have “highly relevant” and “relevant” documents, we treat the former as Srelevant, and the latter as Brelevant. The latter were treated as Brelevant rather than Arelevant because it has been reported that there are many marginally or partially relevant documents in the TREC qrels: Sormunen (2002) reported that about one half of their TREC qrels were only marginally relevant; Sakai and Sparck Jones (2001) reported that only about 56% of the TREC qrels were highly relevant for a subcollection of the early TREC document sets.
Let \(R({\mathcal{L}})\) denote the number of \({\mathcal{L}}\)relevant documents, and let \(R=\sum_{{\mathcal{L}}}R({\mathcal{L}}).\) Let \(cg(r)=\sum_{1 \leq i \leq r} g(i)\) denote the cumulative gain at Rank r of the system output, where \(g(i)=gain({\mathcal{L}})\) if the document at Rank i is \({\mathcal{L}}\)relevant and g(i) = 0 otherwise (i.e., if the document at Rank i is either judged nonrelevant or unjudged). Let cg _{ I }(r) denote the cumulative gain of an ideal ranked output, where an ideal ranked output is one that satisfies g(r) > 0 for 1 ≤ r ≤ R and g(r) ≤ g(r − 1) for r > 1. For NTCIR, for example, listing up all Srelevant documents, followed by all Arelevant documents, followed by all Brelevant documents produces an ideal ranked output. Note that whether “nonrelevant” (i.e., either judged nonrelevant or unjudged) documents are retrieved below these relevant documents does not matter, as the nonrelevant documents do not carry gain values. Moreover, note that several “ideal” ranked outputs can exist in general, since documents can be interchanged within each relevance level.
Let isrel(r) be one if the document at Rank r is relevant and zero otherwise, and let \(count(r)=\sum_{1 \leq i \leq r}isrel(i).\) Clearly, precision at Rank r is given by \(P(r)=count(r)/r\).
Throughout this article, we let l = 1,000 as it has been reported that nDCG with a small document cutoff is unreliable (Sakai 2007f). Moreover, we let a = 2 because it has been reported that nDCG with a large logarithm base is counterintuitive and lacks discriminative power (Sakai 2007d), despite the fact that this parameter was designed to reflect persistence just like RBP’s p and Qmeasure’s β. We shall come back to this issue in Sect. 3.4.
3.2 Q′, AP′, nDCG′ and bpref
Sakai (2007a) reported that Q′, AP′ and nDCG′ are simpler and better solutions to the problem of evaluating IR systems with incomplete relevance data than bpref (Buckley and Voorhees 2004). Recall that these represent the application of Q, AP and nDCG to condensed lists, respectively.
In fact, R ≤ N in holds for all topics used in our experiments (See also Table 2), so bpref is always bpref_R in our study.
Scaling by a constant is generally not good, especially if the constant is large, because this means that the misplacement penalties with respect to the top ranked relevant documents are virtually ignored (Sakai 2007a). For example, suppose that there is a condensed list that has a judged nonrelevant document at Rank 1 and a relevant document at Rank 2. For this document at rank r′ = 2, the misplacement penalty is \(r^{\prime}count(r^{\prime})=21=1\), and \(P(r^{\prime})=1/2\). Thus, the existence of the judged nonrelevant document at Rank 1 weighs heavily in the case of \(\hbox{AP}^{\prime}\). In contrast, this has very little impact on bpref, because the misplacement penalty is divided not by r′ = 2 but by a large number, namely, R or N. In other words, bpref lacks the “top heaviness” of AP′, which is one of the main strengths of the original AP. It is clear that bpref_N (Sakai 2007a), which always uses N for scaling the misplacement penalty, suffers severely from this problem, as N is generally a very large number: See, for example, Table 2 which we shall discuss later. Sakai (2007a) showed experimentally that bpref_N indeed performs very poorly.
It should be noted that AP′ is actually implemented in trec_eval. However, it appears that it was never properly examined until Yilmaz and Aslam (2006) and Sakai (2007a) rediscovered it.
3.3 RBP
The assumption behind RBP is that the user, after examining the document at Rank r, will examine the document at Rank (r + 1) with probability p or stop scanning the ranked list with probability 1 − p. Thus the model assumes that the transition probability is independent of the relevance of the document at Rank r. Whether this assumption is realistic or not is debatable, but this does make RBP easy to interpret and to compute. Moreover, Moffat et al. (2007) argue that RBP is suitable for evaluation with incomplete relevance data as it is guaranteed to increase as more relevance judgments are added (since it does not have a recall component) and the error due to unjudged documents can be quantified.
Values of RBP for an ideal ranked output
 RBP.5  RBP.8  RBP.95 

R = 1  .5  .2  .05 
R = 10  .9990  .8926  .4013 
R = 100  1  1  .9941 
R = 1,000  1  1  1 
Moreover, Table 1 shows the extreme cases of when R = 1: it can be observed that the RBP of an ideal ranked output (i.e., one that has the only one relevant document at Rank 1) can range from 0.05 (p = 0.95) and 0.5 (p = 0.5), since RBP in this case equals 1 − p. Thus the user’s persistence, i.e., the probability of moving from a document from Rank r to that at Rank (r + 1), influences the effectiveness value of the same ranked output quite drastically, even though only the document at Rank 1 is being examined. Whether this is a desirable feature for an IR metric is also debatable. In contrast, Q, nDCG and AP are by definition guaranteed to be one whenever the system output is an ideal ranked output regardless of the value of R, since they are based on comparing the system output with the ideal one.^{5}
We further argue that depending on recall is not necessarily bad. The real user may have some idea of the number of relevant documents, due to his background knowledge, or if not, by looking at the total number of hits shown in the IR interface. Moreover, even if this is not the case, a good IR performance metric is not necessarily one that closely mimics “user satisfaction.” For example, a user may be very satisfied with the ranked output, having found a decent document, but he may have missed ten other documents that are in fact more relevant than the one he has found. That is, the user may be happy, just because he is ignorant. From a conscientious system developer’s point of view, if he knows that there are ten relevant documents that should be retrieved, then he would design a system that can retrieve as many of them as possible rather than a system that makes the user “happy” by showing just one relevant document and hiding the other relevant ones completely. Hence Q and AP depend directly on R, the number of judged relevant documents, and even nDCG depends on it indirectly, as it relies on an ideal ranked output which lists up all relevant documents.
3.4 Topheaviness of RBP, AP, Q and nDCG
In Fig. 1, the topheaviness curve of AP is almost completely hidden by that of Qmeasure, because in a binary relevance environment, Qmeasure = AP holds if there is no relevant document below Rank R, while Qmeasure > AP holds if there is at least one relevant document below Rank R (Sakai 2006c). Thus the AP curve actually begins to deviate from the Qmeasure one at Rank 11 in the graph at the top (where R = 10).
It can also be observed that the topheaviness curves of nDCG have a minor problem: nDCG with a logarithm base of 2 cannot distinguish between a system that has a relevant document at Rank 1 and one that has a relevant document at Rank 2. This is because, according to the original definition of nDCG (which we stick to), gain discounting cannot be applied to ranks above a(=2). This is precisely why using a large a with nDCG is no good (Sakai 2007d): it makes the topheaviness curve even flatter.^{6} It should also be noted that the topheaviness curve for nCG (Järvelin and Kekäläinen 2002; Kekäläinen 2005), the undiscounted version of nDCG, is a completely flat line. That is, to nCG, it does not matter at all at which rank the relevant document is found. This explains why nCG performs very poorly: and the same goes for Precision at l (Sakai 2007f).
To sum up, the IR metrics we consider in this study all have a mechanism, each in its own way, of penalising relevant documents found near the bottom of the ranked list. But the graphs suggest that using p = 0.5 for RBP may not be good for reliable evaluation. This we will verify in our experiments described below.
4 Full and reduced data
TREC and NTCIR data used in our experiments
 TREC03  TREC04  NTCIR6J  NTCIR6C 

#Topics  50  49  50  50 
#Documents  Approx. 529,000  858,400  901,446  
Pool depth  125  100  100  100 
Average N  925.5  654.6  1157.9  999.4 
Range N  [292, 2050]  [132, 1371]  [480, 2732]  [414, 1907] 
Average R  33.2  41.2  95.3  88.1 
Range R  [4, 115]  [3, 161]  [4, 311]  [15, 400] 
Srelevant  8.1  12.5  2.5  21.6 
Arelevant  –  –  61.1  30.4 
Brelevant  25.0  28.8  31.7  36.1 
#Teams  16  14  12  11 
#All runs  78  110  73  45 
#Runs used for rank correlation  30  30  30  30 
To examine the effect of relevance data incompleteness on the IR metrics, we created reduced relevance data from the full relevance data, following the original methodology by Buckley and Voorhees (2004): First, for each topic, we created a randomised list of judged relevant documents of size R, and a separate randomised list of judged nonrelevant documents of size N. Then, for each reduction rate j ∈ {90, 70, 50, 30, 10}, we created a reduced set of relevance data by taking the first R _{ j } and N _{ j } documents from the two lists, respectively, where \(R_{j}=\hbox{max}(1, truncate(R{*}j/100))\) and \(N_{j}=\hbox{max}(10,truncate(N{*}j/100))\). The contents 1 and 10 have been copied from Buckley and Voorhees (2004), representing the minimum number of judged relevant and nonrelevant documents required for a topic, respectively. In practice, the constant 10 was seldom used since N was generally very large. This stratified sampling is essentially equivalent to random sampling from the entire set of judged documents (Yilmaz and Aslam 2006).
The above method of random sampling from the original qrels may be criticised: Possibly, a better method of studying the effect of incompleteness would be to use the actual pools for each topic, and vary the pool depth. However, we prefer to be faithful to the methodology by Buckley and Voorhees as one of the main goals of this study is to contrast their claims regarding bpref (Buckley and Voorhees 2004) with our new findings. We will report on our “shallow pool” experiments elsewhere (Sakai 2008a).
5 Discriminative power
This section compares the robustness of IR metrics to incomplete relevance assessments in terms of discriminative power using Sakai’s Bootstrap Sensitivity Method (Sakai 2006b, 2007b). The input to this method are a test collection, a set of runs, an IR metric, and the significance level α for bootstrap hypothesis tests (Efron and Tibshirani 1993). Using resampled topic sets, the method conducts a paired bootstrap hypothesis test for every system pair, and computes the discriminative power, i.e., for how many system pairs the IR metric was able to detect a significant difference, and the estimated overall performance difference required to achieve that significance. For this purpose, one thousand bootstrap samples of topics were created for each data set, by sampling with replacement from the original topic set. Details of Sakai’s Bootstrap Sensitivity Method can be found elsewhere (Sakai 2006b; 2007b).
It should be stressed that we are comparing the discriminative power of different IR metrics while holding α constant in a common experimental environment, where α is the probability of Type I Error, representing the chance of concluding that two systems are different even though the truth is they are equivalent. In other words, IR metrics with high discriminative power are those that can reliably detect meaningful differences between two systems. The Boostrap Sensitivity Method is closely related to the swap method proposed by Voorhees and Buckley (2002) which compares two systems using two different topics sets and examines whether the two outcomes are the same. Unlike the Boostrap Sensitivity Method, the swap method lacks a theoretical foundation and is not directly related to statistical significance tests. However, Sakai (2006b, 2007b) reported that the two methods yield very similar results when used for comparing different IR metrics.
Discriminative power at α = 0.05 with 100% qrels
 Disc. power (%)  Diff. required 
 Disc. power (%)  Diff. required 

(a) TREC03  (c) NTCIR6J  
Q  80/120 = 66.7  0.07  nDCG  48/66 = 72.7  0.09 
Q′  77/120 = 64.2  0.07  nDCG′  47/66 = 71.2  0.10 
AP  77/120 = 64.2  0.07  Q  47/66 = 71.2  0.08 
AP′  77/120 = 64.2  0.09  Q′  47/66 = 71.2  0.09 
nDCG  71/120 = 59.2  0.08  AP  46/66 = 69.7  0.10 
nDCG′  71/120 = 59.2  0.08  AP′  46/66 = 69.7  0.09 
bpref_R  69/120 = 57.5  0.08  bpref_R  42/66 = 63.6  0.12 
RBP.8  57/120 = 47.5  0.08  RBP.95  42/66 = 63.6  0.07 
RBP.95  55/120 = 45.8  0.04  RBP.8  40/66 = 60.6  0.08 
RBP.5  45/120 = 37.5  0.12  RBP.5  36/66 = 54.5  0.10 
(b) TREC04  (d) NTCIR6C  
Q  63/91 = 69.2  0.08  nDCG′  43/55 = 78.2  0.10 
Q′  62/91 = 68.1  0.08  Q  42/55 = 76.4  0.07 
AP  61/91 = 67.0  0.07  nDCG  42/55 = 76.4  0.09 
AP′  61/91 = 67.0  0.07  RBP.95  42/55 = 76.4  0.06 
nDCG  58/91 = 63.7  0.08  AP′  42/55 = 76.4  0.07 
nDCG′  58/91 = 63.7  0.09  bpref_R  42/55 = 76.4  0.08 
bpref_R  57/91 = 62.6  0.09  AP  41/55 = 74.5  0.08 
RBP.95  45/91 = 49.5  0.05  Q′  40/55 = 72.7  0.08 
RBP.8  36/91 = 39.6  0.09  RBP.8  35/55 = 63.6  0.09 
RBP.5  30/91 = 33.0  0.12  RBP.5  27/55 = 49.1  0.13 

For TREC03 and TREC04, Q(′), AP(′), nDCG(′) and bpref_R are more discriminative than RBP.

For NTCIR6J, Q(′), AP(′) and nDCG(′) are more discriminative than bpref_R and RBP.

For NTCIR6C, Q(′), AP(′), nDCG(′) bpref_R and RBP.95 are more discriminative than RBP.8 and RBP.5.

To sum up, the overall winners given 100% relevance data are Q(′), AP(′) and nDCG(′).
It is clear from Table 3 that small values of p for RBP hurt discriminative power. This is probably because a small p makes RBP too topheavy: as we have seen in Fig. 1, using p = 0.5 implies that IR systems are more or less evaluated based on the top 10 documents only, which makes evaluation very unreliable (Sakai 2007f). For this reason, we drop RBP.05 from our experiments henceforth.

For TREC03 and TREC04, \(\hbox{Q}^{\prime}\), \(\hbox{AP}^{\prime}\) and \(\hbox{nDCG}^{\prime}\) are more robust than other metrics to incomplete relevance assessments. The original nDCG does well for TREC04 but not for TREC03.

Similarly, for NTCIR6J and NTCIR6C, \(\hbox{Q}^{\prime}\), \(\hbox{AP}^{\prime}\), \(\hbox{nDCG}^{\prime}\) and nDCG are the most robust. (Bpref_R appears to do well for NTCIR6C, but it has a problem, as we shall discuss later using Table 5.)

RBP.95, AP and RBP.8 are at the bottom of the list, exactly in this order for all four data sets.

To sum up, the overall winners in terms of robustness to incomplete relevance assessments are \(\hbox{Q}^{\prime}\), \(\hbox{AP}^{\prime}\) and \(\hbox{nDCG}^{\prime}\). AP and RBP clearly lack the robustness. nDCG, Q and bpref_R lie in the middle.
Discriminative power at α = 0.05 with 10% qrels
 Disc. power (%)  Diff. required 
 Disc. power (%)  Diff. required 

(a) TREC03  (c) NTCIR6J  
AP′  63/120 = 52.5  0.14  Q′  46/66 = 69.7  0.10 
Q′  61/120 = 50.8  0.13  nDCG  45/66 = 68.2  0.06 
nDCG′  60/120 = 50.0  0.14  AveP′  44/66 = 66.7  0.11 
bpref_R  47/120 = 39.2  0.16  nDCG′  44/66 = 66.7  0.10 
Q  32/120 = 26.7  0.09  Q  43/66 = 65.2  0.05 
nDCG  29/120 = 24.2  0.09  bpref_R  39/66 = 59.1  0.11 
RBP.95  26/120 = 21.7  0.01  RBP.95  36/66 = 54.5  0.01 
AP  13/120 = 10.8  0.08  AP  34/66 = 51.5  0.04 
RBP.8  6/120 = 5.0  0.03  RBP.8  21/66 = 31.8  0.03 
(b) TREC04  (d) NTCIR6C  
Q′  50/91 = 54.9  0.11  Q′  39/55 = 70.9  0.11 
AP′  46/91 = 50.5  0.12  AP′  39/55 = 70.9  0.11 
nDCG′  43/91 = 47.3  0.12  bpref_R  39/55 = 70.9  0.12 
nDCG  42/91 = 46.2  0.09  nDCG′  38/55 = 69.1  0.12 
bpref_R  37/91 = 40.7  0.15  nDCG  37/55 = 67.3  0.06 
Q  29/91 = 31.9  0.11  Q  33/55 = 60.0  0.04 
RBP.95  24/91 = 26.4  0.01  RBP.95  31/55 = 56.4  0.02 
AP  15/91 = 16.5  0.09  AP  28/55 = 50.9  0.04 
RBP.8  10/91 = 11.0  0.04  RBP.8  12/55 = 21.8  0.03 
Number of significant differences detected with 10% qrels but not with 100% qrels: (i) #significant; (ii) #inconsistent; (iii) percentage
 #Significant  #Inconsistent  % 
 #Significant  #Inconsistent  % 

(a) TREC03  (c) NTCIR6J  
AP  13  2  15  AP  34  2  6 
Q  32  2  6  Q  43  4  9 
nDCG  29  0  0  nDCG  45  1  2 
RBP.8  6  0  0  RBP.8  21  0  0 
RBP.95  26  1  4  RBP.95  36  2  6 
bpref_R  47  7  15  bpref_R  39  1  3 
AP′  63  3  5  AP′  44  2  5 
Q′  61  5  8  Q′  46  1  2 
nDCG′  60  5  8  nDCG′  44  0  0 
(b) TREC04  (d) NTCIR6C  
AP  15  1  7  AP  28  1  4 
Q  29  0  0  Q  33  0  0 
nDCG  42  0  0  nDCG  37  0  0 
RBP.8  10  0  0  RBP.8  12  0  0 
RBP.95  24  0  0  RBP.95  31  1  3 
bpref_R  37  4  11  bpref_R  39  10  26 
AP′  46  5  11  AP′  39  1  3 
Q′  50  3  6  Q′  39  1  3 
nDCG′  43  1  2  nDCG′  38  0  0 
The above analysis was based on the number of statistically significant differences detected given incompleteness relevance data. The basic assumption here is that the set of significantly different pairs at j% reduction rate is basically a subset of one with the full relevance data. However, it might be the case that most of these conclusions at j% reduction rate are in fact inconsistent with the original conclusions with the 100% relevance data. We thus provide an additional analysis in Table 5, which is similar in spirit to the “accuracy” of Bompada et al. (2007). The table compares, for each metric, the set of significantly different pairs at 10% reduction rate with that with the full relevance data. For example, Table 5(a) shows that, for TREC03, AP detected a statistical significance for 13 cases with the 10% relevance data, but two of them (15%) are not among the set of cases detected by AP with the 100% relevance data. Assuming that the conclusions with the 100% relevance data are the ground truth, the numbers presented in the table represent “errors”. As can be seen, the number of errors are generally small, supporting the aforementioned assumption. Bpref_R, however, appears to be quite unreliable from this viewpoint as well: For example, Table 5(d) shows that as many as 10 cases out of the 39 significant differences detected by bpref_R at 10% reduction rate (See also Table 4(d)) are inconsistent with the original bpref_R results. This, again, is not good news for bpref.
6 Rank correlation
The previous section compared the discriminative power of IR metrics, by focussing on the differences between individual pairs of runs. We now compare the entire system rankings according to two different IR metrics given the original relevance data, as well as two different qrels, namely, the original one and a reduced one.
Kendall’s rank correlation between different metrics, given 100% qrels
Q  nDCG  RBP.8  RBP.95  bpref_R  AP′  Q′  nDCG′  

(a) TREC03  
AP  .931  .857  .706  .848  .922  .982  .931  .867 
Q  –  .844  .655  .807  .871  .949  .991  .853 
nDCG  –  –  .775  .853  .844  .857  .844  .991 
RBP.8  –  –  –  .821  .747  .697  .655  .775 
RBP.95  –  –  –  –  .899  .839  .798  .853 
bpref_R  –  –  –  –  –  .913  .862  .844 
AP′  –  –  –  –  –  –  .949  .867 
Q′  –  –  –  –  –  –  –  .853 
(b) TREC04  
AP  .968  .940  .747  .890  .968  .977  .945  .945 
Q  –  .936  .733  .876  .954  .972  .977  .940 
nDCG  –  –  .770  .903  .936  .936  .922  .977 
RBP.8  –  –  –  .821  .770  .733  .710  .756 
RBP.95  –  –  –  –  .913  .876  .853  .890 
bpref_R  –  –  –  –  –  .945  .931  .931 
AP′  –  –  –  –  –  –  .959  .940 
Q′  –  –  –  –  –  –  –  .945 
(c) NTCIR–6J  
AP  .968  .834  .770  .857  .908  .936  .903  .811 
Q  –  .857  .793  .880  .913  .940  .936  .834 
nDCG  –  –  .862  .885  .834  .853  .894  .968 
RBP.8  –  –  –  .830  .770  .789  .802  .839 
RBP.95  –  –  –  –  .867  .885  .890  .853 
bpref_R  –  –  –  –  –  .972  .922  .830 
AP′  –  –  –  –  –  –  .940  .848 
Q′  –  –  –  –  –  –  –  .880 
(d) NTCIR–6C  
AP  .963  .903  .885  .949  .972  .986  .936  .908 
Q  –  .922  .857  .922  .936  .949  .972  .926 
nDCG  –  –  .880  .945  .913  .899  .949  .995 
RBP.8  –  –  –  .926  .903  .880  .839  .876 
RBP.95  –  –  –  –  .968  .945  .913  .940 
bpref_R  –  –  –  –  –  .977  .917  .908 
AP′  –  –  –  –  –  –  .922  .903 
Q′  –  –  –  –  –  –  –  .954 

Q′, AP′ and nDCG′ are consistently among the most robust metrics in terms of system ranking stability. Bpref_R does well for TREC04.

As Figs. 7 and 8 show, the system rankings by AP and RBP.8 collapse as relevance data is reduced. RBP.95 is also not very good: at 30% reduction rate, its Kendall’s rank correlation with the original ranking is as low as that of AP for TREC04 and for NTCIR6J; it performs as poorly as RBP.8 for NTCIR6C.

To sum up, \(\hbox{Q}^{\prime}\), AP′ and nDCG′ are again the overall winners, and the advantage of introducing a new metric like bpref is not clear in terms of system ranking stability either. RBP is not as good as Q′, AP′ and nDCG′ in terms of system ranking stability, even with p = 0.95. Again, nDCG, Q and bpref_R lie in the middle.
Kendall’s rank correlation: 100% vs 10% qrels for each metric
(a) TREC03  (c) NTCIR6J  
AP′  .807  AP′  .899 
RBP.95  .802  Q′  .894 
Q′  .793  nDCG′  .867 
Q  .738  RBP.95  .839 
nDCG′  .724  nDCG  .821 
bpref_R  .724  bpref_R  .802 
nDCG  .715  Q  .743 
AP  .664  RBP.8  .655 
RBP.8  .503  AP  .563 
(b) TREC04  (d) NTCIR6C  
nDCG′  .890  Q′  .949 
Q′  .880  \(\hbox{nDCG}^{\prime}\)  .936 
bpref_R  .871  nDCG  .917 
AP′  .839  Q  .885 
nDCG  .798  AP′  .880 
RBP.95  .752  bpref_R  .853 
Q  .706  AP  .789 
AP  .559  RBP.8  .775 
RBP.8  .559  RBP.95  .756 
7 Conclusions
This article compared the robustness of IR metrics to incomplete relevance data, using four different sets of gradedrelevance test collections with submitted runs—the TREC 2003 and 2004 robust track data and the NTCIR6 Japanese and Chinese IR data from the crosslingual task. Our discriminative power experiments and rank correlation experiments agreed that Q′, AP′ and \(\hbox{nDCG}^{\prime}\), the application of Q, AP and nDCG to condensed lists, are more robust than other metrics to relevance data incompleteness; that AP and RBP lack the robustness; and that nDCG, Q and bpref_R lie in the middle. As these results hold across two different evaluation efforts, namely TREC and NTCIR, we believe that these findings are general. It is also interesting that \(\hbox{Q}^{\prime}\), \(\hbox{nDCG}^{\prime}\) and AP′ are comparable to one another in terms of robustness to incomplete relevance data, even though Q and nDCG are clearly superior to AP. In other words, the advantage of using graded relevance seems to disappear when condensed lists are used with very incomplete relevance data.
Our TREC03, TREC04 and NTCIR6 results, together with the NTCIR3 and NTCIR5 results reported by Sakai (2007a), provide ample evidence that Q′, AP′ and nDCG′ are not only simpler than but also superior to bpref, at least in terms of discriminative power and system ranking stability. Although we have no intention of claiming that Q′, AP′ and nDCG′ are the perfect solution to the problem of relevance data incompleteness, we believe that they are more elegant than introducing metrics like bpref and bpref_N (i.e., RankEff) that lack the “topheaviness” property of AP by definition.
Even though Moffat et al. (2007) claimed that RBP is suitable for evaluation with incomplete relevance data as its error due to unjudged documents can be quantified, we demonstrated that it has weaknesses. While RBP is interesting in that it is independent of recall, because of this very feature, it often does not equal one even for an ideal ranked output. For example, as we have discussed using Table 1, an ideal output for a topic with 10 (regular) relevant documents may receive an RBP of .4013, while an ideal output for a topic with 100 (regular) relevant documents may receive an RBP of .9941. Whether it is good to average such a measurement across topics is debatable. Moreover, our experimental results showed that small values of p make RBP unreliable, and that RBP is not as robust to incomplete relevance data as Q′, AP′ and \(\hbox{nDCG}^{\prime}\) in terms of discriminative power and system ranking stability, even with p = 0.95.
The fact that Q′, AP′ and \(\hbox{nDCG}^{\prime}\) perform clearly and consistently better than the original Q, AP and nDCG in an incomplete relevance environment implies the following: The assumption that all unjudged documents are nonrelevant is not good; It is much better to treat all unjudged documents as if they never existed, in order to let judged relevant and judged nonrelevant documents move up the ranks and hence serve as stronger pieces of evidence for computing system effectiveness.
It should be recalled, however, that we used random sampling from the original qrels in order to artificially create very incomplete test collections. We shall discuss the effect of using shallow pools and that of using fewer participating teams for forming relevance assessments elsewhere Sakai (2008a, 2008b). Moreover, although we examined the IR metrics in terms of discriminative power and Kendall’s rank correlation, there may be other criteria for choosing “good” metrics. “Simplicity” and “intuitiveness” are but a few examples, although they are difficult to quantify. Establishing a standard set of criteria for metric selection is an important goal of our future research.
As we mentioned in Sect. 2, our present study takes the approach of choosing IR metrics given a test collection with incomplete relevance data. However, the approach of constructing reliable test collections efficiently (e.g., work by Carterette et al. 2006) is equally important, and combining these two approaches is probably even more so. That is, IR metrics should perhaps be designed by taking the process of test collection construction into account. This is another research topic that needs to be explored.
Footnotes
 1.
 2.
 3.
 4.
Subcollection AP requires even more knowledge, namely, how small the subcollection with relevance assessments is compared to the entire document collection.
 5.
AP, on the other hand, has a different weakness, in that it can be one for a suboptimal ranked output in a graded relevance environment. To be more specific, AP is one as long as all top R documents are at least somewhat relevant: It does not matter if partially relevant documents are retrieved above highly relevant ones.
 6.
 7.
The statistical significance of Kendall’s rank correlation depends directly on the number of runs (Sakai 2006b).
Notes
Acknowledgments
We thank William Webber, Justin Zobel and Alistair Moffat for their criticisms and constructive comments. We also thank Ellen Voorhees for letting us use the TREC robust track data, and Alistair Moffat and Justin Zobel for providing their unpublished manuscript (Moffat and Zobel 2008).
Open Access
This article is distributed under the terms of the Creative Commons Attribution Noncommercial License which permits any noncommercial use, distribution, and reproduction in any medium, provided the original author(s) and source are credited.
References
 Aslam, J. A., & Savell, R. (2003). On the effectiveness of evaluating retrieval systems in the absence of relevance judgments. In ACM SIGIR 2003 Proceedings (pp. 361–362).Google Scholar
 Aslam, J. A., Pavlu, V., & Yilmaz, E. (2006). A statistical method for system evaluation using incomplete judgments. In ACM SIGIR 2006 Proceedings (pp. 541–548).Google Scholar
 Bompada, T., Chang, C.C., Chen, J., Kumar, R., & Shenoy, R. (2007). On the robustness of relevance measures with incomplete judgments. In ACM SIGIR 2007 Proceedings (pp. 359–366).Google Scholar
 Buckley, C., & Voorhees, E. M. (2004). Retrieval evaluation with incomplete information. In ACM SIGIR 2004 Proceedings (pp. 25–32).Google Scholar
 Burges, C., Shaked, T., Renshaw, E., Lazier, A., Deeds, M., Hamilton, N., & Hullender, G. (2005). Learning to rank using gradient descent. In ACM ICML 2005 Proceedings (pp. 89–96).Google Scholar
 Büttcher, S., Clarke, C. L. A., Yeung, P. C. K., & Soboroff, I. (2007). Reliable information retrieval evaluation with incomplete and biased judgements. In ACM SIGIR 2007 Proceedings (pp. 63–70).Google Scholar
 Carterette, B., Allan, J., & Sitaraman, R. (2006). Minimal test collections for retrieval evaluation. In ACM SIGIR 2006 Proceedings (pp. 268–275).Google Scholar
 Cleverdon, C. W. (1967). The cranfield tests on index language devices. In Aslib Proceedings (Vol. 19, pp. 173–192).Google Scholar
 Cormack, G. V., Palmer, C. R., & Clarke, C. L. A. (1998). Efficient construction of large test collections. In ACM SIGIR ’98 Proceedings (pp. 282–289).Google Scholar
 De Beer, J., & Moens, M.F. (2006). Rpref—a generalization of bpref towards graded relevance judgments. In ACM SIGIR 2006 Proceedings (pp. 637–638).Google Scholar
 Efron, B., & Tibshirani, R. (1993). Introduction to the bootstrap. Chapman & Hall/CRC.Google Scholar
 Grönqvist, L. (2005). Evaluating latent semantic vector models with synonym tests and document retrieval. In ELECTRA Workshop—Methodologies and Evaluating of Lexical Cohesion Techniques in RealWorld Applications (pp. 86–88).Google Scholar
 Järvelin, K., & Kekäläinen, J. (2002). Cumulated gainbased evaluation of IR techniques. ACM Transactions on Information Systems, 20(4), 422–446.CrossRefGoogle Scholar
 Kando, N. (2007). Overview of the Sixth NTCIR Workshop. In NTCIR6 Proceedings (pp. i–ix).Google Scholar
 Kekäläinen, J. (2005). Binary and graded relevance in IR evaluations—comparison of the effects on ranking of IR systems. Information Processing and Management, 41, 1019–1033.CrossRefGoogle Scholar
 Moffat, A., Webber, W., & Zobel, J. (2007). Strategic system comparisons via targeted relevance judgments. In ACM SIGIR 2007 Proceedings (pp. 375–382).Google Scholar
 Moffat, A., & Zobel, J. (2008). Rankbiased precision for measurement of retrieval effectiveness (under review).Google Scholar
 Sakai, T. (2006a). Bootstrapbased comparisons of IR metrics for finding one relevant document. In AIRS 2006: Lecture Notes in Computer Science 4182 (pp. 374–389). SpringerVerlag.Google Scholar
 Sakai, T. (2006b). Evaluating evaluation metrics based on the bootstrap. In ACM SIGIR 2006 Proceedings (pp. 525–532).Google Scholar
 Sakai, T. (2006c). On the task of finding one highly relevant document with high precision. Information Processing Society of Japan Transactions on Databases, 47(SIG4(TOD29)), 13–27Google Scholar
 Sakai, T. (2007a). Alternatives to Bpref. In ACM SIGIR 2007 Proceedings (pp. 71–78).Google Scholar
 Sakai, T. (2007b). Evaluating information retrieval metrics based on bootstrap hypothesis tests. Information Processing Society of Japan Transactions on Databases 48(SIG9(TOD35)), 11–28Google Scholar
 Sakai, T. (2007c). evaluating the task of finding one relevant document using incomplete relevance data. In Forum on Information Technology 2007 Information Technology Letters (pp. 91–94).Google Scholar
 Sakai, T. (2007d). On penalising late arrival of relevant documents in information retrieval evaluation with graded relevance. In Proceedings of the First International Workshop on Evaluating Information Acess (EVIA 2007) (pp. 32–43).Google Scholar
 Sakai, T. (2007e). On the properties of evaluation metrics for finding one highly relevant document. Information Processing Society of Japan Transactions on Databases, 48(SIG9(TOD35)), 29–46.Google Scholar
 Sakai, T. (2007f). On the reliability of information retrieval metrics based on graded relevance. Information Processing and Management, 43(2), 531–548.CrossRefGoogle Scholar
 Sakai, T. (2008a). Comparing metrics across TREC and NTCIR: The robustness to pool depth bias (under review).Google Scholar
 Sakai, T. (2008b). Comparing metrics across TREC and NTCIR: The robustness to system bias (under review).Google Scholar
 Sakai, T., & Sparck Jones, K. (2001). Generic summaries for indexing in information retrieval. In ACM SIGIR 2001 Proceedings (pp. 190–198).Google Scholar
 Soboroff, I., Nicholas, C., & Cahan, P. (2001). Ranking retrieval systems without relevance judgments. In ACM SIGIR 2001 Proceedings (pp. 66–73).Google Scholar
 Sormunen, E. (2002). Liberal relevance criteria of TREC—counting on negligible documents? In ACM SIGIR 2002 Proceedings (pp. 324–330).Google Scholar
 Voorhees, E. M. (2001). Evaluation by highly relevant documents. In ACM SIGIR 2001 Proceedings (pp. 74–82).Google Scholar
 Voorhees, E. M. (2002). The effect of topic set size on retrieval experiment error. In ACM SIGIR 2002 Proceedings (pp. 316–323).Google Scholar
 Voorhees, E. M. (2004). Overview of the TREC 2003 robust retrieval Track. In TREC 2003 Proceedings.Google Scholar
 Voorhees, E. M. (2005). Overview of the TREC 2004 robust retrieval track. In TREC 2004 proceedings.Google Scholar
 Voorhees, E. M., & Buckley, C. (2002). The philosophy of information retrieval evaluation. In Proceedings of CLEF 2001, Lecture Notes in Computer Science 2406 (pp. 355–370).Google Scholar
 Yilmaz, E., & Aslam, J. A. (2006). Estimating average precision with incomplete and imperfect judgments. In ACM CIKM 2006 Proceedings (pp. 102–111).Google Scholar
 Zobel, J. (1998). How Reliable are the results of largescale information retrieval experiments? In ACM SIGIR ’98 Proceedings (pp. 307–314).Google Scholar