A support vector method for optimizing average precision

By Yisong Yue, Thomas Finley, Filip Radlinski, Thorsten Joachims. Annual International ACM SIGIR Conference on Research and Development in Information Retrieval, 2007

777 citations recorded by the source corpus.

Subfields: Artificial Intelligence, Computer Vision and Pattern Recognition, Information Systems

View Paper: https://www.semanticscholar.org/paper/094ce8379a16030a7a993b2f70af46916f7ea99b

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Artificial Intelligence

In Computer Vision and Pattern Recognition

In Information Systems

Citations in the corpus

Few-Shot Learning Through an Information Retrieval Lens (2017, methodology): “The scoring function that [20] used is borrowed from [21] and [22]:”
Dueling Posterior Sampling for Preference-Based Reinforcement Learning (2019, background): “e than one bit of information per preference elicitation; and batch supervised settings such as learning to rank (Herbrich et al., 1999; Chu and Ghahramani, 2005; Joachims, 2005; Burges et al., 2005; Yue et al., 2007; Burges et al., 2007; Liu, 2009). 3 PROBLEM STATEMENT Preliminaries. We consider ﬁx”
Good Practice in Large-Scale Learning for Image Classification (2014, background): “While the one-versus-rest strategy is computationally efficient and yields competitive results in practice, it is—at least in theory—clearly suboptimal with respect to a strategy directly optimizing a ranking loss [9], [17], [38], [75], [82], [86], [88].”
dipIQ: Blind Image Quality Assessment by Learning-to-Rank Discriminable Image Pairs (2017, methodology): “Representative algorithms include SoftRank [79], SVMmap [80], and RankGP [81].”
Ranking with ordered weighted pairwise classification (2009, methodology): “Sy.y To recover an SVM for interdependent output spaces, we .rst notice that [1 - t]= maxb.”
Early exit optimizations for additive machine learned ranking systems (2010, background): “Recent work has addressed the possibility of directly optimizing ranking measures such as MAP [32] or NDCG [28].”
"On the (Non-)existence of Convex, Calibrated Surrogate Losses for Ranking" (2012, background): “(iii) Since r can only take a ﬁnite set of values, we can prove that ℓ is r -calibrated if and only if: is the mean of R ( p, σ ) for σ ∈ arg sort( s ) .”
The scoring of poses in protein-protein docking: current capabilities and future directions (2013, methodology): “Further, listwise ranking methods could be employed where, instead of viewing each pose independently and assigning it a score, and instead of comparing pairs of poses to determine which is superior, the whole list is ranked simultaneously as an inseparable set [137-142].”
Unachievable Region in Precision-Recall Space and Its Effect on Empirical Evaluation (2012, methodology): “Additionally, some algorithms, such as SVM-MAP (Yue et al., 2007) and SAYU (Davis et al., 2005), explicitly optimize the AUCPR of the learned model.”
Listwise approach to learning to rank: theory and algorithm (2008, methodology): “Let X be the input space whose (2)In a broad sense, methods directly optimizing evaluation measures, such as SVM-MAP (Yue et al., 2007) and AdaRank (Xu & Li, 2007) can also be regarded as listwise algorithms.”
A Theoretical Analysis of NDCG Type Ranking Measures (2013, background): “Although early results of learning to rank often reduce ranking problem to classiﬁcation or regression [15, 18, 23, 26, 5], recently there is evidence that learning a ranking function by optimizing a ranking measure such as NDCG is a promising approach [33, 36].”
Seq2Slate: Re-ranking and Slate Optimization with RNNs (2018, background): “Similarly, in listwise approaches the loss depends on the full permutation of items [e.g., Cao et al., 2007, Yue et al., 2007].”
Learning to Rank for Information Retrieval and Natural Language Processing (2011, neutral): “A>B, B>C) – Implicit relevance judgment: derived from click-through data • Creation of List – List (or permutation) of documents is given – Ideal but difficult to implement Implicit Relevance Judgment Doc A Doc B Doc C B > A ranking of documents at search system users often clicked on Doc B ordered ”
Scalable Learning of Non-Decomposable Objectives (2016, methodology): “While the work of [27] demonstrates the merit of optimizing MAP instead of accuracy for reasonably sized domains, scalability is still hindered by the use of a cutting plane training algorithm that requires a costly identification of the most violated constraint.”
Beyond position bias: examining result attractiveness as a source of presentation bias in clickthrough data (2010, methodology): “Our feature representation consists primarily of indicator step functions defined using raw signals such as title bolding differences and click position (similar to Section 4.2 in [ 25 ]).”
Ranking Measures and Loss Functions in Learning to Rank (2009, background): “Note that recently people try to directly optimize ranking measures [17, 12, 14, 18].”
CLiMF: learning to maximize reciprocal rank with collaborative less-is-more filtering (2012, methodology): “However, the AUC measure does not reflect well the quality of the recommendation lists, since it is not a top-biased measure [34], i.”
End-to-End Training of Object Class Detectors for Mean Average Precision (2016, methodology): “Using the same dynamic-programming approach as [17], they apply it to the case of single-class AP with a model based on R-CNN [3], without NMS at training time.”
Yahoo! Learning to Rank Challenge Overview (2010, methodology): “The second sub-category tries to optimize the IR measure during training: examples include AdaRank (Xu and Li, 2007), several algorithms based on structured estimation (Chapelle et al., 2007; Yue et al., 2007; Chakrabarti et al., 2008) and SoftRank (Taylor et al.”