Concept Drift

Concept drift refers to the mapping relationship between input data and target labels changing over time, which is different from data drift where only the input distribution changes.

Data drift or concept drift refers to the phenomenon where model performance degrades significantly when the real-world data distribution changes while the model is still trained on old data; the industry addresses this with continuous learning and active learning strategies.

模型漂移通常分为数据漂移和概念漂移两类，数据漂移指输入数据统计分布随时间变化，概念漂移指输入与输出关系本身发生变化

Data drift can usually be detected by monitoring statistical distances between the input distribution and the training distribution (such as PSI or KL divergence), whereas concept drift can only be caught through labeled online evaluation.

Data drift refers to a change in the distribution of input features P(X) while the relationship between features and labels P(Y|X) remains unchanged; concept drift is when P(Y|X) itself changes.

Rule drift refers to the phenomenon in which the underlying decision rules of an environment change non-explicitly over time; similar to concept drift, it requires the learning system to detect distribution changes and trigger model updates.

Concept drift refers to a change in the true relationship between inputs and labels.

Concept drift refers to a change in the mapping relationship between inputs and correct outputs itself, for example when an answer that was originally correct no longer holds after business rules are updated.

Model drift is a core problem that AI trading systems must continuously address, referring to the phenomenon in which a model's performance degrades over time in a production environment, and it is usually divided into two categories: data drift and concept drift.

Concept drift refers to changes over time in the composition of people and activity patterns within a facility, so that a model that performed well during training may see significantly degraded performance weeks after deployment, requiring continuous recalibration.

State-drift is different from data drift or concept drift, which focus on changes in the model's input distribution or input-output mapping and can be detected with observability tools; state-drift concerns the expiration of stored beliefs, which can fail even when the input is entirely in-distribution.

Means of addressing concept drift include online learning, sliding window retraining, and drift detection algorithms (such as ADWIN and DDM).

Concept drift refers to the statistical relationship between input data and the target variable changing over time, causing a previously trained model to gradually become ineffective; it is usually divided into three types: abrupt, gradual, and periodic.

Distribution shift includes three common types: covariate shift (the input distribution changes but the conditional distribution does not), label shift (the label marginal distribution changes) and concept drift (the conditional distribution changes over time).

Data drift and concept drift can cause model performance degradation, but may show no anomalies in service-layer health metrics (latency, throughput, error rate), requiring active detection through statistical methods such as KL divergence, PSI, and KS tests.

Data distribution drift falls into three types—covariate shift, prior probability shift, and concept drift—of which concept drift is the most insidious and dangerous.

Dataset shift primarily includes three subtypes: covariate shift (changes in the distribution of input features while the conditional label distribution remains unchanged), prior probability shift (changes in the marginal distribution of labels), and concept drift (changes in the mapping relationship between inputs and outputs)

Stock price series are typical non-stationary data, and the substantial change between the data distribution at model training time and at prediction time is referred to in statistics as 'concept drift'.

Financial fraud is an adversarial game scenario where fraud patterns continuously drift, leading to concept drift.

Financial fraud is an adversarial game scenario in which the prior probability and conditional probability distributions the model relies on continuously drift, a phenomenon known as Concept Drift