[SOUND] Hi, in this session we are going to give a brief overview on clustering different types of data. We encounter various kinds of types of data. For example, the early clustering algorithm most times with the design was on numerical data. And the second type of data is category data, including the binary that most people consider as also can be handled in categorical data or category. For example, gender, race, zip code, or market-basket data, you would consider the data is discrete. There's no natural order, but certain kind of data is text data, this is very popular in social media, on the Web, or social networks. Usually if we consider a word as one dimension, then we are handling very high dimensional data, but they are very sparse. Their value usually corresponding to word frequencies. The methods to handle such data, include a combination of k-means and agglomarative method, topical modelling methods, or co-clustering methods. We are going to discuss this extensive later. Multimedia data is another kind of data needs clustering. Usually images, audio, video, those are Flickr and YouTube actually are multimedia data. They are multi-modal in a sense. They have image, audio, video, many cases they have text, as captions as well. We can consider data are contextual data, they contain both behavioral and the contextual attributes. For example for images, we can consider the position of a pixel represents its context, the value represents its behavior. For video and music data, we can consider temporal ordering of the records represents its meaning. Time-series data is another popularly encountered data like in sensors, stock markets, temporal tracking, or forecasting, those task are usually handled time-series. Time-series, we usually consider data are temporally dependent. They are consecutive, usually the interval is somehow equal. Like, we can consider time as contextual attributes, data value as behavioral attribute. Usually such analysis include correlation-based online analysis, like online clustering of stocks to find stock tickers. Or we use shape-based offline analysis, for example, we can cluster ECG based on overall shapes. Sequence data is another kind of data. Usually in weblog analysis, or biological sequence analysis, or analyze the system commands. These, we can think the placement rather than time, the contextual attribute. That means where they are. Then the similarity function, include Hamming distance, edit distance, longest common subsequences. The clustering method, we call sequence clustering may use suffix tree, may use generative model, like Hidden Markov Model. Stream data is another kind of data. Stream data can be considered as a real-time data, like water flowing in and out. It may evolve a long time. It may have concept shifts because stream coming and go. So, we need a single pass algorithm, rather you can see it again and again. And the typical method for stream clustering could be micro-clustering, that means we need to create efficient intermediate representation by some kind of a micro-clustering method. Graphs and homogeneous networks is another kind of data. This then kind of data usually, so-called homogeneous networks that means we consider networks as graphs, the nodes and edges actually are of one kind, one type. Actually any kind of data can be modeled as a graph with the nodes as different attributes and entries, and edge graph in their similarity values. The typical method for handling graph clustering could be generative models, combinatorial algorithm like graph cuts, spectral clustering method, non-negative matrix factorization methods. Then a little bit beyond homogeneous network are heterogeneous networks. This kind of network consists of multiple types of nodes and edges. Like a bibliographical data, hospital, handling disease, patients, doctors, and treatments to cluster different kinds of nodes and links together. There are some interesting algorithms, like the NetClus algorithm. Uncertain data, means the data may contain noise, may have approximate values, or multiple possible values. Usually we need to incorporate some probabilistic information, like distribution, or approximate values that were improve the quality of clustering. Finally, recently, there are lots of discussion on big data. That means, we can model systems. That may store and process very big data, like weblog analysis. Usually, like Google's MapReduce framework is you have Map function to distribute the computation across many machines. Then we have Reduce function to aggregate the results obtained from the Map step. So we can see, the data is very rich that's why the clustering is a rather sophisticated methodologies, trying to handle different kinds of data. [MUSIC]
