Learning and Decision Making with Interval Valued Data

21 concepts14 papers2 proposals

When we observe or measure the real world, we often compress a range of values into a single, "crisp" number. A sensor recording wind speeds over an hour might report an average of 12 m/s; an expert evaluating a supplier might provide a single score of 8 out of 10; a web service log might report an average latency of 200ms.

This compression is a destructive operation. By collapsing a distribution or a range into a single point estimate, we discard critical structural information: the variance, the boundaries of possibility, measurement noise, and epistemic uncertainty.

To prevent this information loss, a growing body of research focuses on interval-valued data representation. Instead of mapping observations to exact scalar values, this paradigm treats data entries as continuous numerical ranges $[L, U]$. By carrying these intervals directly through mathematical, statistical, and machine learning pipelines, models can make decisions and predictions that are inherently robust to real-world fluctuations.


Preserving Structural Uncertainty

Why do intervals outperform traditional point estimates? The core intuition is that intervals preserve the upper and lower bounds of a phenomenon, allowing downstream algorithms to reason about the entire envelope of possibility rather than an arbitrary average.

In dynamic physical systems, this preservation is vital. For instance, in short-term wind speed prediction, Ronay Ak and Enrico Zio demonstrate that treating input data as intervals inherently captures real-world measurement uncertainty and environmental fluctuations. Their work shows that interval-valued inputs consistently yield more stable and reliable prediction intervals than traditional crisp methods, especially in highly volatile scenarios.

A similar advantage appears in digital infrastructure management. When users select web services based on Quality of Service (QoS) metrics, traditional exact queries fail if the user does not know the exact performance ranges available. Chen Ding et al. address this in QoS Browsing for Web Service Selection by representing performance metrics as bounded intervals. This preserves provider-guaranteed ranges and real-world variability, allowing users to visualize the performance "landscape" and iteratively zoom in on optimal services without losing uncertainty information to point-based averages.


Adapting Core Algorithms to Interval Geometry

Traditional machine learning and statistical algorithms expect points in space, not boxes or lines. To process interval data natively, the underlying mathematical operations—such as distance metrics, loss functions, and optimization constraints—must be redesigned to respect interval boundaries.

Regression and System Modeling

In foundational work on interval data analysis, Hisao Ishibuchi et al. identified a core consistency issue in traditional regression when applied to non-crisp datasets. In Interval Data Analysis by Revised Interval Regression Model, the authors integrated the optimization of upper and lower bounds into a single linear programming framework. This mathematical unification ensures that the resulting models remain logically and physically consistent across the entire input space, simultaneously representing the minimum and maximum possible values of the observed phenomenon. The authors extended this principle to complex, nonlinear systems in Interval GMDH and Its Application, utilizing a self-organizing modeling approach to estimate both the "outer bounds" (containing all data) and "inner bounds" (contained within the data) of uncertain engineering processes.

Clustering and Pattern Matching

Standard fuzzy clustering algorithms often fail when faced with noisy interval data because they force every data point to belong to a cluster. To handle the fuzzy boundaries inherent in range-based data, Jianlu Luo et al. developed an improved clustering algorithm in An IMPROVED FCM-POSSIBLE CLUSTERING ALGORITHM FOR INTERVAL DATA. By relaxing membership constraints using possibility theory, their mechanism calculates distances and memberships based directly on multi-dimensional interval boundaries. This allows the algorithm to model internal consistency and effectively ignore outliers that fall outside the true data intervals.

Neural network architectures can similarly be adapted to process ranges natively. Shengjun Xue et al. designed a specialized neural network-based interval pattern matcher in A Neural Network-Based Interval Pattern Matcher, enabling nuanced pattern recognition for environmental and forecasting problems where data naturally falls into ranges rather than discrete coordinates.


Probabilistic and Generative Semantics

Treating intervals as simple geometric bounds is highly effective, but statistical rigor requires understanding the underlying data-generation process. Researchers have approached this from both generative and combinatorial perspectives.

Instead of merely fitting bounds to observed intervals, Xin Zhang and Boris Béranger proposed a generative framework in Constructing likelihood functions for interval‐valued random variables. Their approach models the latent, real-valued data generation process that manifests as the observed ranges. By constructing likelihood functions based on this underlying process, they achieve more interpretable parameters and robust statistical inference, capturing information that simpler endpoint-focused approaches discard.

From a data-mining perspective, Ji Cheng et al. tackled the challenge of mining complex structures, such as Order-Preserving Submatrices (OPSMs), from uncertain data. In Mining Order-Preserving Submatrices Under Data Uncertainty: A Possible-World Approach, the authors leverage possible world semantics. Instead of treating an interval as a static block, their algorithm evaluates pattern significance across all valid scalar instantiations within the continuous numerical ranges. This allows downstream mining algorithms to account for measurement noise and fluctuations, resulting in higher-quality pattern discovery in noisy real-world datasets.


Reconciling Imprecise Preferences and Expert Judgments

Interval representations are uniquely suited for human-centric data, such as expert opinions, subjective preferences, and behavioral metrics, where forcing a single numerical value introduces artificial precision.

Multi-Attribute Decision-Making (MADM)

When experts evaluate complex strategies, their judgments are inherently vague. In Modeling Enablers of Environmentally Conscious Manufacturing Strategy: An Integrated Method, Wenyan Song and Yanru Zhou transform subjective expert assessments into rough interval numbers, capturing natural cognitive imprecision. To prevent premature information loss during data aggregation, Felix T. S. Chan and Wen Jiang developed an evidential fusion method in Evidential Supplier Selection Based on Interval Data Fusion. By retaining interval data throughout the fusion process within the Dempster-Shafer theory framework, their approach provides a more accurate reflection of original uncertainty and expert opinions.

A persistent challenge in decision-making is ranking alternatives when the relative importance (weights) of different criteria is unknown. Bizhan Lian et al. addressed this in Relative Entropy Sorting Method Based on the Preference Information of Alternative by representing both objective attribute values and subjective expert preferences as closed interval ranges. They introduced a mathematical framework based on relative entropy (Kullback-Leibler distance) to reconcile what the data says with what decision-makers want across these ranges. Similarly, Rui Xiu developed a decision model in Decision Model of Aero-Engine Design Integration Platform Implementation Solution Selection Based on Number Interval Number with Priority Order that utilizes interval numbers alongside a Lagrangian optimization model to determine attribute weights. This ensures that complex engineering selections are grounded in realistic data bounds rather than potentially inaccurate single-point approximations.

Comparing Aggregated Intervals

Once interval data is aggregated, comparing and matching these ranges requires specialized metrics. Justin Kane Gunn et al. designed a novel similarity measure in Similarity measure for aggregated fuzzy numbers from interval-valued data. Using the Interval Agreement Approach (IAA), their method combines geometric features (such as area and perimeter) with statistical measures (such as quartiles and agreement ratios) weighted via Principal Component Analysis (PCA). This allows for highly nuanced comparisons of human opinions and assessments.

Point-Based Representation (Destructive)
Real-World Engagement ──► [ Average: 4.5 Minutes ] ──► Discards Variance & Confidence
(e.g., Dwell Time)

Interval-Valued Representation (Preservative)
Real-World Engagement ──► [ Bounded Range: 3.0 to 6.0 Min ] ──► Preserves Uncertainty

This capacity to model vague human behavior also extends to recommendation systems. In collaborative filtering, models often suffer from severe data sparsity when users fail to provide explicit ratings. To circumvent this, Weike Pan et al. proposed a transfer learning framework in Transfer Learning in Collaborative Filtering with Uncertain Ratings that extracts patterns from implicit auxiliary data, such as video dwell time or engagement metrics. Because these metrics provide a range of possible user preferences rather than a crisp score, the authors mathematically incorporate these uncertain intervals directly into the predictive model, allowing the system to learn from rich, range-bound behavioral data without forcing artificial precision.

Key Foundational Work

QoS Browsing for Web Service Selection (2009) — contributed 2 concept elements to this direction.
Interval GMDH and Its Application (1989) — contributed 1 concept elements to this direction.
Interval Data Analysis by Revised Interval Regression Model (1989) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_interval_valued_data_representation_54bea5dc,
  title = {interval-valued data representation},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/interval-valued-data-representation/rdr_54bea5dc58414e478bc3b5f38a92da9a},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_interval_valued_data_representation_54bea5dc_v1,
  title = {interval-valued data representation},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/interval-valued-data-representation/rdr_54bea5dc58414e478bc3b5f38a92da9a/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake