Presentation on theme: "Mark Goadrich Computer Science and Mathematics"— Presentation transcript:
1 Using Relational Structure for Learning and Modeling in Biomedical and Social Domains Mark GoadrichComputer Science and MathematicsCentenary College of LouisianaNatural Science ColloquiumNovember 6th, 2007
2 Overview First-Order Logic and Machine Learning The world is full of ObjectsModel these Objects to understand the worldInductive Logic ProgrammingObjects and Relations/PropertiesAgent-Based ModelingObjects and Interactions/BehaviorsILP takes snapshots of the worldABM simulates a dynamic world
3 Bongard Problems 6 positive examples of a concept on left Bongard problems. From Mikhail Bongard, Soviet Computer Scientist6 positive examples of a concept on left6 negative examples on rightHow to learn this concept using a computer?
4 First-Order Logic using PROLOG Objectse3, t1, t2, c1Typesexample(e3)triangle(t1)triangle(t2)circle(c1)Relationshas_shape(e3, t1)has_shape(e3, t2)has_shape(e3, c1)inside(t2, c1)left(t2, t1)size(c1, 2.5)above(t2, t1)…We make lots of facts for each example. This is what is true about the world.Positive Example 3Repeat this process for each example in dataset
5 Inductive Logic Programming (ILP) Search the space of possible rules “positive(E) :- …”Judge rule quality by positive - negative coveragepositive(E):- has_shape(E, A)positive(E)positive(E):- has_shape(E, A), triangle(A)Animate sample rule at level 1 and 2positive(E) :- has_shape(E, A), has_shape(E, B),triangle(A), circle(B), inside(A, B).
6 Research Issues in ILP Enormous space to search for rules Enormous number of examplesIncorporation of continuous featuresLearning of probabilistic rulesEvaluation of rule qualitySurvey of ILP domains and future interests
7 Mutagenesis Designing effective and selective drugs Represent chemicals as atoms and bonds between thematm(127, 127_1, c, 22, )bond(127, 127_1, 127_6, 7 )Learned mutagenic rule:mutagenic(A) :- atm(A, B, c, 27, C), bond(A, D, E, 1), bond(A, B, E, 7).Bond(compound, atom1, atom2, bondtype)Atm(compound, atom, element, atomtype, charge)(A) 3,4,4'-tri-nitro-biphenyl(B) 2-nitro-1,3,7,8-tetrachlorodibenzo-1,4-dioxin(C) 1,6,-dinitro-9,10,11,12-tetrahydrobenzo[e]pyrene(D) nitrofurantoin
8 Breast Cancer Detection Large dataset of abnormalities found in mammogramsNot enough radiologistsRelational featuresMore than one abnormality per mammogramMore than one mammogram per person over timemalignant(A) :- not asymmetric(A), in_same_mammorgram(A, A2), spiculated_margin(A2), not distorted(A2)
9 Robot ScientistRepresent Metabolic Pathways as a Regulatory Network GraphKnock out genes, and then systematically deduce the unknown functionTry to learn the network from time-series microarray data
10 Social Networks People are connected by friendships into networks Each person has likes/dislikes, possibly influenced by their networkCan we learn your interests based on who you know and what they like? Targeted advertisements?
11 Netflix Prize What movies should Netflix recommend you watch next? Large relational datasetMoviesTitlesRatingsFriendsFriend’s ratingsGenre$1 million if you achieve 10% improvement over their algorithm Cinematch
12 Zendo Board game about inductive logic Master creates a rule which some 3-D pyramid structures fit and others do notPlayers build structures and try to guess the Master ruleEasier to design computer Master to decide if a structure fits the ruleHarder to design computer Player that must efficiently guess the rule
13 Crab ClawsWhat physical characteristics distinguish between two species?Within the same species, what changes due to growth, diet and their relation to predation?Find the “shock graph” of each imageUse ILP to learn differences based on these graphs
14 Agent-Based Modeling Objects have interactions with each other Flocks of Birds, Schools of FishSeparationAlignmentCohesionObjects interact with their environmentAnt Foraging, Pheromones, Traffic LawsAgent-Based Modeling (ABM)Create discrete-time computational simulationAlign models with known behaviorVary parameters to test new hypotheses
16 ConclusionsFirst-Order Logic combines with ILP and ABM to create a powerful representation of the worldResearch OpportunitiesSocial NetworksZendo PlayerClaws and Shock GraphsCellular SimulationSocial Simulation[Insert your favorite dataset here]