Presentation on theme: "M INING H IGH -S PEED D ATA S TREAMS Presented by: Yumou Wang Dongyun Zhang Hao Zhou."— Presentation transcript:
M INING H IGH -S PEED D ATA S TREAMS Presented by: Yumou Wang Dongyun Zhang Hao Zhou
I NTRODUCTION The world’s information is doubling every two years. From 2006 to 2011, the amount of information grew by a factor of 9 in just five years.
I NTRODUCTION By 2020 the world will generate 50 times the amount of information and 75 times the number of "information containers" However, IT staff to manage it will grow less than 1.5 times. Current algorithms can only deal with small amount of data less than a day’s data of many applications. For example, banks, telecommunication companies.
I NTRODUCTION Problems : When new examples arrive at a higher rate than they can be mined, the amount of unused data grows without bounds as time progresses. Today, to deal with these huge amount of data in a responsible way is very important. Mining these continuous data streams brings unique opportunities, but also new challenges.
B ACKGROUND Design Criteria for mining High Speed Data Streams It must be able to build a model using at most one scan of the data. It must use only a fixed amount of main memory. It must require small constant time per record.
B ACKGROUND Usually, use KDD system to operate this examples when they arrive. Shortcomings: learning model learned are highly sensitive to example ordering compare to the batch model. Others can produce the same model as batch version but very slower.
C LASSIFICATION M ETHOD Input: Examples of the form (x,y), y is the class label, x is the vector of attributes. Output: A model y=f(x), predict the classes y of future examples x with high accuracy.
D ECISION T REE One of the most effective and widely-used classification methods. A decision tree is a decision support tool that uses a tree- like graph or model. Decision trees are commonly used in machine learning.
B UILDING A D ECISION T REE 1. Starting at the root. 2. Testing all the attributes and choose the best one according to some heuristic measure. 3. Split one node into branches and leaves. 4. Recursively replacing leaves by test nodes.
E XAMPLE OF D ECISION T REE
P ROBLEMS There are some problems existed in traditional decision tree. Some of them assume that all training data examples can be stored simultaneously in main memory. Disadvantages: Limited the number of examples can be learned from. Disk-based decision tree learners: examples in disk, repeatedly reading them. Disadvantages: expensive when learning complex trees.
H OEFFDING T REES Designed for extremely large datasets Main idea: To find the best attribute at a given node by considering only a small subset of the training examples that pass through the node. Using how many examples is sufficient
H OEFFDING B OUND Definition: The statistical result that can decide how many examples “n” using by each node is called Hoeffding bound. Assume: R—the range of variable r n independent observations mean: r’ With probability 1-δ, the true mean of r is at least r’-є
H OEFFDING B OUND This function is a decreasing function n is bigger, the є is smaller It is the difference between true value and mean value of r.
H OEFFDING T REE A LGORITHM
Inputs: S -> is a sequence of examples, X -> is a set of discrete attributes, G(.) -> is a split evaluation function, δ -> is one minus the desired probability of choosing the correct attribute at any given node. Outputs: HT -> is a decision tree.
H OEFFDING T REE A LGORITHM Goal: Ensure that, with a high probability, the attribute chosen using n examples, is the same as that would be chosen using infinite examples. Let Xa be the attribute with the highest observed G’ and Xb be with second highest attribute. After seeing n examples. Let ΔG’ = G’(Xa) – G’(Xb) ΔG’ > Thus a node needs to accumulate examples from the stream until becomes smaller than ΔG.
H OEFFDING T REE A LGORITHM The algorithm constructs the tree using the same procedure as ID3. It calculates the information gain for the attributes and determines the best attributes. At each node it checks for condition ΔG >. If the condition is satisfied, then it creates child nodes based on the test at the node. If not it streams in more training examples and carries out the calculations till it satisfies the condition.
H OEFFDING T REE A LGORITHM Memory cost d—number of attributes c—number of classes v—number of values per attribute l—number of leaves in the tree The memory cost for each leaf is O(dvc) The memory cost for whole tree is O(ldvc)
A DVANTAGES OF H OEFFDING T REE 1. Can deal with extremely large datasets. 2. Each example to be read at most once in a small constant time. Makes it possible to mine online data sources. 3. Build very complex trees with acceptable computational cost.
VFDT—V ERY F AST D ECISION T REE
Filtering out poor attributes Dropping early Reduces memory consumption Initialization Can be initialized with other existing tree Set a head start Rescans
T ESTS —C ONFIGURATION
T ESTS —S YNTHETIC DATA
Time consumption 20m examples VFDT takes 5752s to read, 625s to process 100k examples C4.5 takes 36s VFDT takes 47s
T ESTS — PARAMETERS W/ & w/o over-pruning
T ESTS — PARAMETERS
T ESTS — WEB DATA For predicting accesses 1.89m examples 61.1% with most common class examples for testing
T ESTS — WEB DATA Decision dump 64.2% accuracy 1277s to learn C4.5 with 40MB memory 74.5k examples 2975s to learn 73.3% accuracy VFDT-bootstrapped with C m examples 1450s to learn after initialization(983s to read)
T ESTS — WEB DATA
M INING T IME -C HANGING D ATA S TREAMS
W HY IS VFDT NOT E NOUGH ? VFDT, assume training data is a sample drawn from stationary distribution. Most large databases or data streams violate this assumption – Concept Drift : data is generated by a time- changing concept function, e.g. Seasonal effects Economic cycles Goal: –Mining continuously changing data streams –Scale well
W HY IS VFDT NOT E NOUGH ? Common Approach: when a new example arrives, reapply a traditional learner to a sliding window of w most recent examples –Sensitive to window size If w is small relative to the concept shift rate, assure the availability of a model reflecting the current concept Too small w may lead to insufficient examples to learn the concept –If examples arrive at a rapid rate or the concept changes quickly, the computational cost of reapplying a learner may be prohibitively high.
CVFDT CVFDT (Concept-adapting Very Fast Decision Tree learner) –Extend VFDT –Maintain VFDT’s speed and accuracy –Detect and respond to changes in the example- generating process
CVFDT ( CONTD.) With a time-changing concept, the current splitting attribute of some nodes may not be the best anymore. An out dated subtree may still be better than the best single leaf, particularly if it is near the root. – Grow an alternative subtree with the new best attribute at its root, when the old attribute seems out-of-date. Periodically use a bunch of samples to evaluate qualities of trees. – Replace the old subtree when the alternate one becomes more accurate.
H OW CVFDT W ORKS
S AMPLE E XPERIMENT R ESULT
C ONCLUSION AND F UTURE W ORK CVFDT is able to maintain a decision-tree up- to—date with a window of examples by using a small constant amount of time for each new examples that arrives. Empirical studies show that CVFDT is effectively able to keep its model up-to-date with a massive data stream even in the face of large and frequent concept shifts. Future Work : Currently CVFDT discards subtrees that are out-of-date, but some concepts change periodically and these subtrees may become useful again – identifying these situations and taking advantage of them is another area for further study.