Is there any synthetic / artificial datasets for concept drifting data? I want to visualize the performance of some clustering algorithm when data experiences concept drift and changes over time.
This is a tough question. "Concept drift" is a highly abstracted term that assumes a well developed methodological protocol of text-based constructs temporal evolution. This protocol does not exist, yet.
Standard time series solutions are very well-developed for continuously distributed information, are focused on the analysis of HAC residuals and, obviously, include the variants of Holt-Winters decompositions, exponential smoothing, Box-Jenkins, ARCH, GARCH, HARCH functional forms, spectral decompositions, and so on. Take your pick.
Nonstandard approaches to longitudinal clustering of principal components originated in the 70s with guys working in the (then) hot area of multidimensional scaling and were based on higher order tensor ranks and SVD decompositions such as CANDECOMP and PARAFAC or Pieter Kroonenberg's three-mode approaches:
Time series clustering has been getting a lot of attention in recent years. One excellent overview is Aggarwal and Reddy's book, Data Clustering. However, temporal clustering of conceptual constructs is not a well developed topic in their collection of essays.
Andreas Brandmaier's Permutation Distribution Cluster (PDC) approach to clustering aperiodic time series is well developed conceptually as well as offering routines that can be run in R.
Your challenge is different from so-called "off the shelf" methods like these. The problem is that even cross-sectional, text-based analysis is a wild, wild west of competing solutions -- without even introducing a temporal component.
Honestly though, I think some of the most interesting solutions aren't strictly focused on concept "clustering," which implies a mostly statistical answer, but are more involved with dimension reduction and the longitudinal evolution of core constructs. For instance, computational sociologists interested in the drift of concepts fundamental to American political history have done some terrific work building on social network-type analyses that include some stunning visuals. Visit this paper by Peter Bearman, at Columbia's INCITE Center which is an analysis of over 200 years of US Presidential State of the Union addresses and the evolution of core constructs underlying them, Lexical shifts, substantive changes, and continuity in State of the Union discourse:
Bearman also has a paper on Big Data and Historical Social Science which raises concerns about, e.g., "answering previously largely intractable questions about the timing and sequencing of events, and of event boundaries...based on accounts which rest on narrative sentences."
Given all of that as background, is there data which presumes concept drift? One possible source are the Large Scale Hierarchical Text Classification Challenges"
DMOZ is another possible source since it has text metadata with a temporal component, although it likely the case that you may have to define and build the constructs yourself.
Goggle's GDELT project might be another option:
I've always liked the Million Song Database which has millions of pop song features dating back to the 1960s:
Here's a paper on the evolution of genres of pop music that is based on that data:
And here's a paper on music and aesthetic analysis called *cantometrics":
Yelp just released 13T of research data that may work for you:
And, here's a laundry list of data sources in Github that may have something useful:
Apologies for the lack of canned, easy solutions. Good luck!