Machine Learning

• security and privacy

  • anomaly and intrusion detection

  • traffic identification

• wireless network

  • channel estimation

  • cognitive radio

  • localization

• feedback loops

  • congestion contro.l

  • fault prediction and repair

• optimization problems

  • network design and resource management

  • placement of devices

• data collectoin

• data processing

• model trraining

• test data

• test data processing

• outcome

  • e.g. measuring of accuracy…

• supervised

• unsupervised

• reinforcement learning

• collect / gather labeled data

• split into train and test dataset

• determine input features of training dataset (should have enough knowledge so that model can accurately predict output)

• determine suitable algorithm such as SVM, decision tree, …

• execute algo on training dataset

• evaluate accuracy of model by providing test dataset

• regression

• classification

• binary

• multi class

• clustering

  • k-Means

  • hierarchical

• association

  • anomaly detection

  • PCA

• value k as input

• choose k datapoints randomly -> centroids of each group

• asign remainig datapoints to nearest centroid

• calculate within cluster sum of squares

• repreat step 2-4 until minimum within-cluster sum of squares are obtained

• check for dependency of one data item on another data item

• -> maps acordingly…

• => tries to find interesting relations or associations among variables of dataset

• => basically works on if else concept… (e.g. if A then B…)

• if else consequences

• e.g. if buys cereal also expected to buy milk

• -> if cereal then milk

• cereal called antecedent

• milk called consequent

• support

  • frequency of antecedent (how often occurs in dataset)

• confidence

  • how often rule has found to be true (how often antecedent and consequent occur together)

• lift

  • strenght of any rule

• support (A,C) / (support(A) * support (C))

• lift = 1 -> probability of occurence of antecedent and consequent independent of each other

• lift > 1 -> degree of which to item sets are dependent of each other

• lift < 1 -> tells that one item is substitute for other items -> one item has negative effect on another

• dataset with fixed number of items (e.g. a, b, c, d)

  
  

• generating itemsets from set of item

• -> itemsets consist of all posisble combinations of antecedents and consequents

• -> e.g. a => b, a => c, a => {b,c}, b => a, …

• for each itemset calculate the confidence value

• if confidence value above threshold -> itemset becomes a rule

• different metrics for different ML algos

• regression:

  • RMS deviation

  • the lower, the better

• classificatoin:

  • accuracy (correct class / all class)

  • precision (TP / TP + FP) -> how many of positive are actual positive

  • F1 score (2 TP / 2 TP + FP + FN)

  • -> optimally, all 1

• clustering:

  • WCSS -> if no other value is smaller… (error based)

• association

  • the hogher the threshold confidence value, the mroe solid the rule

  • optimally confidence of 1

• detect if flows or packets are malicioöus

• based on decision,, perform action

• machine learning often used in liuterature for this task

• fully connected

• convolutional

• recurrent

• graph

• networks can be abstracted as graph (routing, …)

• => apply graph based model to this graph abstraction of network…

• standard ML methods cannot directly process graphs

• -> typically process fixed sized inputs

• need method for mapping graphs to those fixed size inputs

solution:

• reduce graphs to fixed size matrix

• -> messagte propagation between nodes

• each node has value

• those values are propagated along edges of the graph

• nodse used those propagated values to update their initial value

• node i -> hidden state vector i

• hidden state depends on neighboring nodes

• message passing split into message function M, aggregation function A, update function U

• apply f to hidden state until fixpoint is reached

• -> apply g to get output