Source-linked AI summary
Posthoc Interpretability of Learning to Rank Models using Secondary Training Data
Jaspreet Singh, Avishek Anand
TL;DR
Black-box rankers make it difficult to understand how features influence rankings, motivating post-hoc interpretation. The paper trains tree-based interpreters on ranker-generated secondary data and finds that ranking agreement depends on the base ranker and training size, while standard IR metrics improve slowly.
Problem
The paper addresses how to understand feature influence in black-box rankers through post-hoc interpretation of their rankings.
Source: §1 · p. 1Method
It trains tree-based interpreters on secondary training data generated from base-ranker predictions, using an interpretable feature subset and comparing training sizes and base-ranker algorithms.
Source: §3 · p. 2, §3.2 · p. 3, §5 · p. 3Results
Ranking agreement increases with more secondary data for listwise base rankers, while pairwise base rankers can be reproduced with little data; standard IR metrics improve slowly, and the chosen content features perform poorly at scale.
Source: §5 · p. 4Takeaways & Limitations
The experiments show that faithful reproduction with interpretable features is possible in certain settings, though outcomes depend on base-ranker type and training size.
Source: §4 · p. 3, §5 · p. 4The study leaves active training-data selection, interpretable feature-space selection, and alternative secondary-labeling mechanisms for future work.
Source: §5 · p. 4
Abstract
from arXiv · showhide
Predictive models are omnipresent in automated and assisted decision making scenarios. But for the most part they are used as black boxes which output a prediction without understanding partially or even completely how different features influence the model prediction avoiding algorithmic transparency. Rankings are ordering over items encoding implicit comparisons typically learned using a family of features using learning-to-rank models. In this paper we focus on how best we can understand the decisions made by a ranker in a post-hoc model agnostic manner. We operate on the notion of interpretability based on explainability of rankings over an interpretable feature space. Furthermore we train a tree based model (inherently interpretable) using labels from the ranker, called secondary training data to provide explanations. Consequently, we attempt to study how well does a subset of features, potentially interpretable, explain the full model under different training sizes and algorithms. We do experiments on the learning to rank datasets with 30k queries and report results that serve show in certain settings we can learn a faithful interpretable ranker.
1 INTRODUCTION
The introduction frames post-hoc interpretability as learning how a trained ranker behaves on test data, rather than explaining its training mechanics. The proposed interpreter learns globally from ranker-generated secondary training data, using features chosen for interpretability.
1 INTRODUCTION
- The motivation is that opaque predictive systems can make feature influence difficult to understand, while transparency matters for trust and restricting discrimination.The introduction also points to regulation restricting some automated decisions and a proposed right to explanation.Source: §1 · p. 1
- The interpreter uses ranker predictions on test inputs to create secondary training data, with training, validation, and test sets formed from query-document pairs.The base ranker and interpreter share an input space, while the interpreter is assumed to have many test inputs and reasonable system familiarity.Source: §1 · pp. 1–2
- The paper assumes the interpreter uses a typically smaller, potentially interpretable feature set, while noting that which content-based features make better explanations is outside its scope.The introduction contrasts understandable content features, such as headline query-term presence, with harder-to-interpret embedding similarity features.Source: §1 · pp. 1–2
- The paper studies post-hoc interpretation of a trained ranker by learning a new model that resembles its rankings, without access to the original training regime.Its notion of interpretability is based on how closely the interpreter reproduces the base ranker’s output, rather than on an explanation of the model’s internal mechanics.Source: §1 · pp. 1–2
- The study asks whether more secondary data helps, whether the base ranker’s training algorithm affects interpreter effectiveness, and how closely a global interpretable model can mimic it.The work targets rankings from learning-to-rank models used in information retrieval and recommender systems.Source: §1 · pp. 1–2
2 RELATED WORK
The related work distinguishes global, model-agnostic interpretation of ranking models from work on local decision boundaries and notes limited attention to black-box rankers in information retrieval. The paper applies a black-box labeling strategy to train an interpretable ranker.
2 RELATED WORK
- The paper targets the global decision surface of a learned ranking model, contrasting its scope with approaches that approximate local decision boundaries.It also distinguishes its goal from adversarial learning: success is not measured against an adversary’s cost model, and it does not isolate susceptible boundary regions.Source: §2 · p. 2
- Prior interpretability work frames human transparency and post-hoc explanations as distinct, potentially competing notions.The cited discussion also questions blanket claims that linear models are interpretable and deep neural networks are not.Source: §2 · p. 2
- The authors identify little prior information-retrieval work on interpreting black-box learning-to-rank models and propose using the black box to generate labels for an interpretable model.They situate this model-agnostic strategy alongside approaches that exploit black boxes to train new models.Source: §2 · p. 2
3 EXPERIMENTAL SETUP
The experiments use Microsoft’s learning-to-rank data to train a base ranker and generate secondary labels for interpreters. They compare ranking agreement and standard information-retrieval metrics across base-ranker types, feature sets, and training sizes.
3 EXPERIMENTAL SETUP
- A base ranker is trained on 5k queries; interpreters use nested splits from 100 to 15k queries labeled by the base ranker, with separate validation and human-labeled test sets.The dataset contains 30k queries, each with 100–300 judged documents and 132 features; the base ranker’s ranking positions are mapped to relevance levels.Source: §3 · p. 2
3.1 Evaluation measures
- Evaluation uses Kendall’s τ and τ@10 to measure preserved pairwise orderings, alongside NDCG@k and Precision@k as standard ranking metrics.The study reports τ@10 because users may be more interested in top results, and compares τ measures against rankings from the base ranker.Source: §3.1 · p. 3
3.2 System Variants
- The system variants train tree-based interpreters from secondary data against listwise and pairwise base rankers, using either all base features or a subset of interpretable features.The base rankers are LambdaMART and RankNet, respectively; the interpreters use a tree-based learning-to-rank approach.Source: §3.2 · p. 3
- The interpretable feature subset F′ contains 24 features selected as easiest for the assumed user to interpret, from query-term presence indicators to TF-IDF scores.The authors identify feature selection by user groups with varying expertise as future work.Source: §3.2 · p. 3
4 RESULTS
Interpreter performance depends on the base ranker and evaluation measure: more secondary data helps ranking agreement for listwise models, while pairwise models can be mimicked with small samples. Standard IR metrics change little as training size grows.
4 RESULTS
- With all features, interpreters do not significantly improve Precision or NDCG as training size grows, and base-ranker-generated labels appear to bound those metrics.The listwise base ranker outperforms the pairwise variant on Precision and NDCG in these experiments.Source: §4 · p. 3
- For a pairwise base ranker, ranking agreement is high after 100 queries (τ = 0.86 and τ@10= 0.74) but decreases with larger splits; for a listwise base ranker, it rises with more data.After 15k queries for the listwise case, τ and τ@10 are 0.49 and 0.74, respectively; the authors call the pairwise pattern preliminary evidence for more careful example selection.Source: §4 · p. 3
- With interpretable features, Precision and NDCG vary little as training size grows, while listwise-base ranking agreement rises and pairwise-base τ@10 is sensitive to split size.For the pairwise case, the 400-query split has the highest τ@10 value, 0.33, and nearly the highest Precision@10, 0.5535.Source: §4 · p. 3
- Under some settings, including a 400-query split with a pairwise base ranker, interpretable-feature models achieve reasonable reproduction, while overall conclusions differ by base-ranker type and metric.The authors report that increasing secondary data helps for interpretable and all-feature models with listwise base rankers; under certain pairwise settings, a subset of interpretable features provides reasonable performance.Source: §4 · p. 3
5 CONCLUSIONS & FUTURE WORK
Using ranker-generated secondary labels, the study explores interpretable rankers built from content features. Results vary with the learning approach, while future work includes data and feature selection.
5 CONCLUSIONS & FUTURE WORK
- Experiments use pairwise- and listwise-trained base rankers, secondary labels, and an interpretable space of 24 content-based features on 30k queries.Source: §5 · p. 3
- Listwise interpreters correlate more with the base ranker as secondary data grows, while pairwise-base rankings can be faithfully reproduced with little data.Source: §5 · p. 4
- Standard IR metrics improve slowly with more training data, and content-based features perform poorly even with large secondary datasets.Source: §5 · p. 4
- Potential explanation scenarios include comparing a pair, explaining an item's position against the rest, and contrasting top-ranked items with the remainder.Source: §5 · p. 4
- Future work includes active training-data selection, interpretable-feature-space selection, and alternative mechanisms for labeling secondary data.Source: §5 · p. 4