Listing 1 - 10 of 31 | << page >> |
Sort by
|
Choose an application
Choose an application
Relationele leertechnieken leren patronen uit relationele gegevensbanken, die gewoonlijk uit meerdere tabellen bestaan, die met elkaar gerelateerd zijn. Deze relaties kunnen bijvoorbeeld een één-op-veel of veel-op-veel cardinaliteitsverhouding hebben. Een voorbeeld waarvoor een predictie gemaakt moet worden kan dus gerelateerd zijn aan een verzameling objecten die mogelijk relevant zijn voor de predictie. Bestaande relationele leermethoden behandelen deze verzamelingen op één van volgende manieren: door het opleggen van condities aan de elementen in de verzameling of door het gebruik van aggregaatsfuncties om ze samen te vatten. Bestaande methoden zijn niet in staat om beide benaderingen te combineren, waardoor ze bepaalde patronen niet kunnen leren. Het belangrijkste doel van dit eindwerk is het combineren van beide benaderingen, dus het aggregeren over een deelverzameling van elementen die aan een specifieke selectieconditie voldoen. Deze combinatie van aggregaten en selecties brengt verscheidene moeilijkheden met zich mee. Ten eerste wordt de zoekruimte substantieel uitgebreid en ten tweede is de algemeen-naar-specifiek ordening van de hypothesen, die verondersteld wordt door veel relationele leersystemen, geschonden. Dit impliceert dat men, gebruik makende van klassieke verfijningsoperatoren, ofwel efficiëntie ofwel volledigheid moet opgeven bij het doorzoeken van de hypotheseruimte. In dit werk ontwikkelen we een algemeen bruikbaar verfijningsraamwerk dat de volledige zoekruimte beschouwt en deze in een algemeen-naar-specifieke, en dus efficiënte, manier doorloopt. Complexe aggregaten worden ingebouwd in een bestaand relationeel leersysteem dat beslissingsbomen construeert. We argumenteren dat de algemeenheidsordening van de zoekruimte in deze context niet geschonden kan worden, en dat bijgevolg klassieke verfijningsoperatoren gebruikt kunnen worden. Om de efficiëntie te verhogen worden twee technieken voorgesteld: een toepassing van het voorgestelde verfijningsraamwerk en een opwaardering van het relationeel beslissingsboomleeralgoritme naar een systeem dat relationele gerandomiseerde bossen leert. Het gebruik van complexe aggregaten wordt ook bestudeerd in het consequent van een hypothese. Meerbepaald onderzoeken we het gebruik van complexe aggregaten in de lineaire modellen die gebouwd worden in de bladeren van relationele modelbomen. Daarvoor wordt het relationele beslissingsboomleeralgoritme uitgebreid om modelbomen te leren. De belangrijkste contributie hierbij is het ontwikkelen van een efficiënte heuristiekfunctie die geschikt is voor het leren van modelbomen. Tenslotte wordt het gebruik van complexe aggregaten geëvalueerd in twee toepassingen. In relational learning one learns patterns from relational databases, which usually contain multiple tables that are interconnected via relations. These relations may be of one-to-many or many-to-many cardinality ratios. Thus, an example for which a prediction is to be given may be related to a set of objects that are possibly relevant for that prediction. Relational classifiers differ with respect to how they handle these sets: some use properties of the set as a whole (using aggregation), some refer to properties of specific individuals of the set, however, most classifiers do not combine both. This imposes an undesirable bias on these learners. This dissertation describes a learning approach that avoids this bias, by using complex aggregates, i.e., aggregates that impose selection conditions on the set to aggregate on. This combination of aggregates and selections presents several difficulties. First, the search space is substantially increased, and second, the generality order of the hypotheses that is assumed by many relational learners is violated. This implies that one either has to give up on efficiency or on completeness when searching the hypothesis space using classical refinement operators. We develop a general refinement framework that considers the complete search space, and traverses it in a general-to-specific, hence efficient, way. Complex aggregates are included in an existing relational learner that constructs relational decision trees. We argue that in this context, the generality ordering can not be violated, and classical refinement operators can be applied. To improve efficiency, we present two techniques: an application of the developed refinement framework and an upgrade of the relational decision tree algorithm to a relational random forest inducer. The use of complex aggregates is also studied in the consequent of a hypothesis. More precisely, we investigate the use of complex aggregates in the linear models built by a relational model tree learner. This involves upgrading the relational decision tree algorithm to a relational model tree learning system. The main contribution in this work is the development of an efficient heuristic function suitable for learning model trees. Bij automatisch leren wordt gezocht naar patronen in gegevens. Vaak zijn die patronen predictief: er wordt een bepaalde eigenschap van de gegevens voorspeld. Zo kunnen we bijvoorbeeld in een gegevensbank van moleculen zoeken naar patronen die voorspellen wanneer een molecule mutageen (schadelijk voor het DNA) is. We concentreren ons op gegevens die zich in een relationele gegevensbank bevinden. Dit betekent dat elk gegeven (ook voorbeeld genoemd) waarvoor een predictie gemaakt moet worden, verspreid zit over meerdere tabellen die aan mekaar gerelateerd zijn. Bijvoorbeeld voor de moleculen kunnen we naast een tabel "molecule" ook een tabel "atoom" hebben die voor elke molecule de bijhorende atomen beschrijft. De relatie tussen molecule en atoom heeft een één-op-veel cardinaliteitsverhouding. Een voorbeeld waarvoor een predictie gemaakt moet worden kan dus gerelateerd zijn aan een verzameling objecten, die mogelijk relevant zijn voor de predictie. De patronen die gezocht worden bestaan meestal uit een verzameling "ALS conditie DAN predictie" regels. Als er in de conditie gebruik gemaakt wordt van een verzameling gerelateerde objecten, dan gebeurde dit tot nog toe op een van volgende manieren: door het nagaan of er in de verzameling een object zit dat voldoet aan een aantal eigenschappen, of door het samenvatten van de verzameling m.b.v. aggregaatfuncties zoals maximum, minimum, cardinaliteit,... Bestaande systemen zijn niet in staat beide aanpakken te combineren, hetgeen een belangrijke beperking legt op de condities die ze kunnen construeren. Het belangrijkste doel van dit eindwerk is het combineren van beide benaderingen, dus het aggregeren over een deelverzameling van elementen die aan een specifieke selectieconditie voldoen. Een voorbeeld van een regel met een dergelijke combinatie is "als de maximale (aggregaat) lading van de koolstof (selectieconditie) atomen van de molecule groter is dan 0.6, dan is de molecule mutageen". Deze combinatie van aggregaten en selecties brengt verscheidene moeilijkheden met zich mee. Ten eerste groeit de zoekruimte van mogelijke patronen significant en ten tweede wordt het moeilijker om deze zoekruimte op een gestructureerde en efficiënte manier te doorlopen. Deze moeilijkheden worden uitgebreid besproken en een oplossing wordt voorgesteld. Daarna worden de zogenaamde complexe aggregaten ingebouwd in een relationeel leersysteem dat beslissingsbomen bouwt en wordt ook het gebruik van complexe aggregaten in het predictie gedeelte van patronen onderzocht. Tenslotte wordt het gebruik van complexe aggregaten ge"evalueerd in twee toepassingen.
Academic collection --- 681.3*I26 <043> --- Learning: analogies; concept learning; induction; knowledge acquisition; language acquisition; parameter learning (Artificial intelligence)--See also {681.3*K32}--Dissertaties --- Theses --- 681.3*I26 <043> Learning: analogies; concept learning; induction; knowledge acquisition; language acquisition; parameter learning (Artificial intelligence)--See also {681.3*K32}--Dissertaties
Choose an application
The three volume proceedings LNAI 10534 – 10536 constitutes the refereed proceedings of the European Conference on Machine Learning and Knowledge Discovery in Databases, ECML PKDD 2017, held in Skopje, Macedonia, in September 2017. The total of 101 regular papers presented in part I and part II was carefully reviewed and selected from 364 submissions; there are 47 papers in the applied data science, nectar and demo track. The contributions were organized in topical sections named as follows: Part I: anomaly detection; computer vision; ensembles and meta learning; feature selection and extraction; kernel methods; learning and optimization, matrix and tensor factorization; networks and graphs; neural networks and deep learning. Part II: pattern and sequence mining; privacy and security; probabilistic models and methods; recommendation; regression; reinforcement learning; subgroup discovery; time series and streams; transfer and multi-task learning; unsupervised and semisupervised learning. Part III: applied data science track; nectar track; and demo track.
Computer science. --- Computer security. --- Data mining. --- Artificial intelligence. --- Image processing. --- Computers. --- Computer Science. --- Data Mining and Knowledge Discovery. --- Artificial Intelligence (incl. Robotics). --- Image Processing and Computer Vision. --- Information Systems Applications (incl. Internet). --- Systems and Data Security. --- Computing Milieux. --- Automatic computers --- Automatic data processors --- Computer hardware --- Computing machines (Computers) --- Electronic brains --- Electronic calculating-machines --- Electronic computers --- Hardware, Computer --- Pictorial data processing --- Picture processing --- Processing, Image --- AI (Artificial intelligence) --- Artificial thinking --- Intellectronics --- Intelligence, Artificial --- Intelligent machines --- Machine intelligence --- Thinking, Artificial --- Algorithmic knowledge discovery --- Factual data analysis --- KDD (Information retrieval) --- Knowledge discovery in data --- Knowledge discovery in databases --- Mining, Data --- Informatics --- Computer privacy --- Computer system security --- Computer systems --- Computers --- Cyber security --- Cybersecurity --- Electronic digital computers --- Protection of computer systems --- Security of computer systems --- Protection --- Security measures --- Computer vision. --- Data protection. --- Artificial Intelligence. --- Database searching --- Science --- Data protection --- Security systems --- Hacking --- Machine vision --- Vision, Computer --- Artificial intelligence --- Image processing --- Pattern recognition systems --- Bionics --- Cognitive science --- Digital computer simulation --- Electronic data processing --- Logic machines --- Machine theory --- Self-organizing systems --- Simulation methods --- Fifth generation computers --- Neural computers --- Machine learning --- Optical data processing. --- Application software. --- Cybernetics --- Calculators --- Cyberspace --- Application computer programs --- Application computer software --- Applications software --- Apps (Computer software) --- Computer software --- Optical computing --- Visual data processing --- Integrated optics --- Photonics --- Optical equipment
Choose an application
The three volume proceedings LNAI 10534 – 10536 constitutes the refereed proceedings of the European Conference on Machine Learning and Knowledge Discovery in Databases, ECML PKDD 2017, held in Skopje, Macedonia, in September 2017. The total of 101 regular papers presented in part I and part II was carefully reviewed and selected from 364 submissions; there are 47 papers in the applied data science, nectar and demo track. The contributions were organized in topical sections named as follows: Part I: anomaly detection; computer vision; ensembles and meta learning; feature selection and extraction; kernel methods; learning and optimization, matrix and tensor factorization; networks and graphs; neural networks and deep learning. Part II: pattern and sequence mining; privacy and security; probabilistic models and methods; recommendation; regression; reinforcement learning; subgroup discovery; time series and streams; transfer and multi-task learning; unsupervised and semisupervised learning. Part III: applied data science track; nectar track; and demo track.
Computer science. --- Computer security. --- Data mining. --- Artificial intelligence. --- Image processing. --- Computers. --- Computer Science. --- Data Mining and Knowledge Discovery. --- Artificial Intelligence (incl. Robotics). --- Image Processing and Computer Vision. --- Information Systems Applications (incl. Internet). --- Systems and Data Security. --- Computing Milieux. --- Automatic computers --- Automatic data processors --- Computer hardware --- Computing machines (Computers) --- Electronic brains --- Electronic calculating-machines --- Electronic computers --- Hardware, Computer --- Pictorial data processing --- Picture processing --- Processing, Image --- AI (Artificial intelligence) --- Artificial thinking --- Intellectronics --- Intelligence, Artificial --- Intelligent machines --- Machine intelligence --- Thinking, Artificial --- Algorithmic knowledge discovery --- Factual data analysis --- KDD (Information retrieval) --- Knowledge discovery in data --- Knowledge discovery in databases --- Mining, Data --- Informatics --- Computer privacy --- Computer system security --- Computer systems --- Computers --- Cyber security --- Cybersecurity --- Electronic digital computers --- Protection of computer systems --- Security of computer systems --- Protection --- Security measures --- Computer vision. --- Data protection. --- Artificial Intelligence. --- Data protection --- Security systems --- Hacking --- Database searching --- Science --- Machine vision --- Vision, Computer --- Artificial intelligence --- Image processing --- Pattern recognition systems --- Bionics --- Cognitive science --- Digital computer simulation --- Electronic data processing --- Logic machines --- Machine theory --- Self-organizing systems --- Simulation methods --- Fifth generation computers --- Neural computers --- Machine learning --- Optical data processing. --- Application software. --- Cybernetics --- Calculators --- Cyberspace --- Application computer programs --- Application computer software --- Applications software --- Apps (Computer software) --- Computer software --- Optical computing --- Visual data processing --- Integrated optics --- Photonics --- Optical equipment
Choose an application
This book contains a selection of the best papers of the 34th Benelux Conference on Artificial Intelligence, BNAIC/ BENELEARN 2022, held in Mechelen, Belgium, in November 2022. The 11 papers presented in this volume were carefully reviewed and selected from 134 regular submissions. They address various aspects of artificial intelligence such as natural language processing, agent technology, game theory, problem solving, machine learning, human-agent interaction, AI and education, and data analysis.
Artificial intelligence. --- Computer engineering. --- Computer networks. --- Social sciences—Data processing. --- Education—Data processing. --- Artificial Intelligence. --- Computer Engineering and Networks. --- Computer Application in Social and Behavioral Sciences. --- Computers and Education. --- Communication systems, Computer --- Computer communication systems --- Data networks, Computer --- ECNs (Electronic communication networks) --- Electronic communication networks --- Networks, Computer --- Teleprocessing networks --- Data transmission systems --- Digital communications --- Electronic systems --- Information networks --- Telecommunication --- Cyberinfrastructure --- Electronic data processing --- Network computers --- Computers --- AI (Artificial intelligence) --- Artificial thinking --- Electronic brains --- Intellectronics --- Intelligence, Artificial --- Intelligent machines --- Machine intelligence --- Thinking, Artificial --- Bionics --- Cognitive science --- Digital computer simulation --- Logic machines --- Machine theory --- Self-organizing systems --- Simulation methods --- Fifth generation computers --- Neural computers --- Distributed processing --- Design and construction
Choose an application
This book contains a selection of the best papers of the 34th Benelux Conference on Artificial Intelligence, BNAIC/ BENELEARN 2022, held in Mechelen, Belgium, in November 2022. The 11 papers presented in this volume were carefully reviewed and selected from 134 regular submissions. They address various aspects of artificial intelligence such as natural language processing, agent technology, game theory, problem solving, machine learning, human-agent interaction, AI and education, and data analysis.
Choose an application
Choose an application
The machine learning research domain has been known to generate great applications in a variety of contexts. Over the years, a plethora of algorithms and training methods are developed to detect anomalies in datasets, identify objects in images, recognize speech or address any other task. In this work, these modern technologies are applied in yet another business context. The scope of this work is to automate the processing of outgoing invoices which are generated each month by a professional service provider. The invoices to be processed by this application are composed of a set of invoice lines. The invoice lines are generated based on time registrations which represent performances which are to be invoiced. Hence, the subject dataset is represented in a multiple-instance structure with the invoices being the bags, each holding a variable number of invoice lines. Active learning is one of the subdomains in machine learning which allows a learner to query labels for data points from an unlabeled dataset. By letting an algorithm select the most informative instances for labelling in an intelligent manner, the learner can predicte targets based on fewer labeled training instances compared to regular supervised learning techniques. In this work a set of active learning experiments are conducted comparing the performance of uncertainty sampling and query-by-committee query strategies. The multiple-instance dataset is transformed using instance based embedding (MILES) and principal component analysis. The results provide evidence that the active learning methods can successfully be applied in a multiple-instance context, using the mentioned transformations. Using uncertainty sampling as query strategy for the active learner led to the best results, nearly matching the performance of a regular passive learner.
Choose an application
In recent years, people are getting used to reading news and obtaining information online. Nevertheless, the information and choices on the internet are overwhelming. To tackle this, recommender systems are developed to provide useful or interesting suggestions to readers. Nowadays, their existence is omnipresent and essential. However, most news RSs focus solely on the readers' preferences while ignoring other stakeholders' utilities and goals. To solve this, we propose a multi-stakeholder news recommender system, by first selecting the stakeholders and defining their utilities in news recommendations, and then incorporating multi-stakeholder concerns with re-rank methods. The performance of the RS is tested dynamically in two-by-two experiment design. The results demonstrate that our methods are capable of increasing the utilities of different parties, with the accuracy decreasing in a moderate way. Our study provides new insights into understanding and developing the multi-stakeholder RS in the news field.
Choose an application
There is currently a rapid adoption of AI in high-risk application areas, like in healthcare, biometric identification, emergency service dispatching, law enforcement, etc. Therefore, it is time for a more human-centric approach in order to enable for trust and transparency. One of the requirements to guarantee transparency is explainability, which has led to new research on explainable artificial intelligence. The aim of explainable artificial intelligence is to enable human understanding of an AI system's output. In this thesis, the selected explainability method is `SHapley Additive exPlanations' (SHAP), since it is the only consistent feature attribution technique for explaining a single instance's prediction. This thesis extends the SHAP feature importance framework to a significance framework for time-to-event model comparison. The existing framework for Cox Proportional Hazards and eXtreme Gradient Boosting is replicated, verified, and extended to random forest methodology. This enables direct comparison between classical statistical methods and machine learning models, to overcome lack of trust in AI. A first extension of the framework considers random survival forests. Because the efficient Tree SHAP algorithm does not support survival data, the computationally expensive Exact SHAP is used. Run time constraints result in lack of confidence interval generation. Extreme hazard ratios due to wide SHAP value ranges were overcome by rescaling random survival forest SHAP values by MaxAbs division. A second extension of the framework considers random forest regression, which transforms survival outcomes first by recasting the survival setting. The efficient Tree SHAP algorithm supports the regression setting, so confidence intervals are generated. Squeezed hazard ratios due to narrow SHAP value ranges were overcome by rescaling random forest regression SHAP values by MaxAbs division. Results from experiments on four data sets generally show hazard ratios which closely agree between gradient boosting and random forest methodology. However, the random forest regression extension outperforms the random survival forest extension based on speed, confidence intervals, and consistency with gradient boosting results. These extensions can lead to new insights by identifying risk stratification factors more precisely, which can have an impact on more personalised treatment of patients.
Listing 1 - 10 of 31 | << page >> |
Sort by
|