Quantifying the limited and gradual concept drift assumption

The impact of data difficulty factors on classification of imbalanced and concept drifting data streams

“…Severity and magnitude are equivalent criteria that have been defined as the amount of changes in the joint probability distribution caused by a drift [50], and as the degree of difference between two points of time [79], respectively. Several different distance measures can be used to characterize drifts according to their severity or magnitude [32,50,64,80], such as Kullback-Leibler Divergence, Hellinger Distance and Total Variation Distance. The Hellinger Distance and the Total Variation Distance have the advantage of being symmetric.…”

“…Another common way to categorize drifts is based on their rate of change [18], also known as speed [50]. Typically, drifts are categorized as abrupt (also called sudden) if they cause sudden changes, or gradual if their underlying joint probability distribution slowly evolves over time [18,64]. Slow evolution can refer to a period of time where two distinct distributions co-exist in the problem, with the new distribution slowly taking over the old one [26,50].…”

Classificação de fluxos de dados com mudança de conceito e latência de verificação

“…Virtualmente todos os trabalhos publicados que lidam com mudança de conceito, e usam bases reais, fazem premissas sobre a natureza das mudanças de conceito existentes a partir de argumentação informal. Até onde vai nosso conhecimento, Sarnelle et al (2015) publicou a primeira tentativa de formalizar e quantificar tais premissas. O trabalho lida com mudanças de conceito que se dão pela movimentação dos eventos de uma classe no espaço de atributos, com alguma direção e velocidade.…”

Dynamic adversarial mining - effectively applying machine learning in adversarial non-stationary environments.

“…However, these techniques make limiting assumptions on the type and nature of drifts, to make it applicable to such latent updates. Only drift in existing classes with slow changes over time, are possible to detect using these approaches [109].…”