200 likes | 282 Views
A Learning System for Decision Support in Telecommunications. Filip Železný, Olga Štěpánková (Czech Technical University in Prague) Jiří Zídek (Atlantis Telecom). Where we come from. Czech Technical University in Prague Faculty of Electrical Engineering Dept. of Cybernetics
E N D
A Learning System for Decision Support in Telecommunications Filip Železný, Olga Štěpánková (Czech Technical University in Prague) Jiří Zídek (Atlantis Telecom)
Where we come from • Czech Technical University in Prague • Faculty of Electrical Engineering • Dept. of Cybernetics • The Gerstner Laboratory for Intelligent Decision Making and Control • Machine Learning Group http://cyber.felk.cvut.cz/gerstner/machine-learning Železný, Štěpánková, Zídek: A Learning System for Decision Support in Telecommunications
The ML Group • Research on ML principles • Instance-based learning, Relational learning, Randomization in search, etc. • Application projects • Mainly data mining in areas such as • Medical (heart surgery mortality predictions, subgroup discovery in Spa-patients), Industrial (Intelligent pump diagnosis), etc. • Telecommunications Železný, Štěpánková, Zídek: A Learning System for Decision Support in Telecommunications
The telecommunication project • “Intelligent Telephone Assistant” • Equipe a private branch exchange (PBX, switchboard) with intelligent behaviour • Helps callers automatically • E.g. Find the correct person to connect to the caller upon recognition of the caller’s number • Learns from previous events • I.e. from calls assisted by humans • Stored in a logging file of the exchange Železný, Štěpánková, Zídek: A Learning System for Decision Support in Telecommunications
What we should learn from… • Logging (history) file of the PBX operation Železný, Štěpánková, Zídek: A Learning System for Decision Support in Telecommunications
… is not so obvious • Related records spread One event (transferred call) Another simultaneous event (transferred call) Železný, Štěpánková, Zídek: A Learning System for Decision Support in Telecommunications
One time axis One event (transferred call) Another event (transferred call) Železný, Štěpánková, Zídek: A Learning System for Decision Support in Telecommunications
Therefore, the plan is to: • Reconstruct stored events • Associate related records • Those related to one event (incoming call) • We know how (expert knowledge) • Recognize the sequence of actions in events • E.g. transfers or attempts to transfer the caller btw. internal lines • We do not know how (materials do not say how actions map to sequences of records) • Learn decision support rules from the event descriptions Železný, Štěpánková, Zídek: A Learning System for Decision Support in Telecommunications
Or, in boxes: Telecomm.Traffic LoggingData Telephone Exchange Rules EventReconstruction EventDescriptions Prediction Železný, Štěpánková, Zídek: A Learning System for Decision Support in Telecommunications
Learning action patterns • Generate event examples • (manual generation) • Nature of examples • Consist of variable number of records • Contain structured data types • Use multiple relations t(time(19,43,48),[1,2],time(19,43,48),e,li,empty,d,empty,empty,ex, [0,6,0,2,3,3,0,5,3,3],empty,anstr([0,0,5,0,0,0]),fe,fe,id(4)). t(time(19,43,48),[1,2],time(19,43,50),e,lb,e(relcause),d,dr,06,ex [0,6,0,0,0,0,0,0,0,0],empty,anstr([0,0,5,0,0,0]),fe,fe,id(5)). ex_ans([0,6,0,2,3,3,0,5,3,3],[1,2]). hangsup([0,6,0,2,3,3,0,5,3,3]). This was stored This “happened”(our description) Železný, Štěpánková, Zídek: A Learning System for Decision Support in Telecommunications
Descriptive ILP setting • ILP = Inductive Logic Programming • Find first-order clauses true in all given interpretations • Our examples ~ interpretations • May also use a background theory to prove clauses • Clauses must comply to a given grammar • E.g. heads (conclusions) consist of names of actions Železný, Štěpánková, Zídek: A Learning System for Decision Support in Telecommunications
Rules • One of the rules • “same_num/2” defined in the background theory Action “external answered call” occurred if… ex_ans(EX1,DN1):- t(D1,IT1,DN1,ET1,e,li,empty,d,EF1,FI1,ex,EX1,empty,ANTR1,CO1,DE1,ID1), t(D2,IT2,DN2,ET2,e,lb,RC2,d,EF2,FI2,ex,EX2,empty,ANTR2,CO2,DE2,ID2) IT2=ET1, ANTR2=ANTR1, same_num(EX1,EX2). ... these records were stored, connected in time, with the same answering port With the “same” caller’s id. (May have different suffices) Železný, Štěpánková, Zídek: A Learning System for Decision Support in Telecommunications
Using the rules… ?-recognize([id(60216),id(60218),id(60224),id(60228),id(60232),id(60239)]) EVENT STARTS. 648256849 rings on 32 - call accepted, 32 attempts to transfer 0648256849 to 16 with notification, but 16 refused, 32 notifies 12 and transfers 0648256849 to 12, 12 attempts to transfer 0648256849 to 28 with notification, but 28 does not respond, 12 notifies 26 and transfers 0648256849 to 26, call terminated. EVENT STOPS. Železný, Štěpánková, Zídek: A Learning System for Decision Support in Telecommunications
Event recognition performance • Proportion of recognized events Železný, Štěpánková, Zídek: A Learning System for Decision Support in Telecommunications
Recognition allowed for visualisation • Frequency of internal transfers of external incoming calls • Some interestingobservations! Železný, Štěpánková, Zídek: A Learning System for Decision Support in Telecommunications
Predicting within events (1) • Ongoing work • Training data • Structural representation of events incoming( date(8,28),time(15,29,19),[0,3,1,5,4,5,9,6,7,5],[3,2], transfer([[2,9],[2,8]], transfer([[2,6]], unavailable))). Železný, Štěpánková, Zídek: A Learning System for Decision Support in Telecommunications
Predicting within events (2) • Background knowledge relations • Predicates that • Map dates to weekdays • (client habits may depend of particular weekdays) • Extract prefix of incoming numbers • Binds callers from the same company, region etc. • Etc. • We try to collect more relevant background knowledge • E.g. regular absence of employees, etc. Železný, Štěpánková, Zídek: A Learning System for Decision Support in Telecommunications
Predictive ILP setting • Input: • Positive example set P (Prolog facts) • Negative example set N (Prolog facts) • Background knowledge B (Prolog theory) • Output • Hypothesis H (Prolog theory) • Such that H & B logically entails • all pP • no nN Železný, Štěpánková, Zídek: A Learning System for Decision Support in Telecommunications
Predictive rules • Example of a rule found: • “if a number starting with 0250- calls the receptionist on Monday, it is always transferred to line 10.” • Such rules allow for • Decision support • Automation • Problem: • Small “coverage” of found rules • Need more relevant background knowledge Železný, Štěpánková, Zídek: A Learning System for Decision Support in Telecommunications
Conclusions • Inductive Logic Programming serves very well for induction from structural and multirelational telecommunication data • Successful reconstruction of events from switchboard logging file • Some signs of predictive induction, but we must collect more relevant background knowledge Železný, Štěpánková, Zídek: A Learning System for Decision Support in Telecommunications