IO4. EXPERT SYSTEM FOR TRAINING RECOMMENDATIONS …

Project reference: 2015-1-LU01-KA202-001353

Project acronym: BIM4VET

Project title: Standardized Vocational Education and Training for BIM in EU

KA2- Cooperation for Innovation and the exchange of good practices strategic partnerships for vocation education and training

This project has been funded with support from the European Commission – Erasmus+ programme. This document reflects the views only of the authors, and the Commission as well the Anefore cannot be held responsible for any use which may be made of the information contained therein.

IO4. EXPERT SYSTEM FOR TRAINING

RECOMMENDATIONS GENERATION

Dissemination level Public Activity 4 – Fuzzy logic method for training recommendation system WP Leader CEA Contributor(s) LIST/ Cardiff University Author(s) Jean-Philippe Poli Status (F: final, D: draft) F

IO4. Expert system for training recommendations generation 1 1

Disclaimer The information in this document is provided as is and no guarantee or warranty is given that the information is fit for any particular purpose. The user thereof uses the information at its sole risk and liability. The document reflects only the authors’ view and the European Commission, and the Anefore are not responsible for any use that may be made of the information it contains.

Copyright © Copyright 2017 BIM4VET Consortium This document may not be copied, reproduced, or modified in whole or in part for any purpose without written permission from the BIM4VET Consortium. In addition to such written permission to copy, reproduce, or modify this document in whole or part, an acknowledgement of the authors of the document and all applicable portions of the copyright notice must be clearly referenced. All rights reserved.

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License


History Version Description Lead author Date 1.1 First version J.P. Poli 01/03/2018 1.2 Review A. Gueriero 15/03/2018 1.3 Final version A. Gueriero 18/04/2018 Diffusion Version Sent to Date 1.1 All consortium partners 01/03/2018 1.2 All consortium partners 15/03/2018 1.3 All consortium partners 18/04/2018 Acknowledgements BIM4VET has received funding from the Erasmus+ programme (Key Action: Cooperation for innovation and the exchange of good practices, Action Type: Strategic Partnerships for vocational education and training). The partners of the project are LIST, Cardiff University, and Commissariat à l’Energie Atomique et aux Energies Alternatives.


Acronym

BIM Building Information Modeling


Table of contents 1. INTRODUCTION 5

2. EXPRESSIFTM AS A TRAINING RECOMMENDER SYSTEM 5 2.1. WHAT IS EXPRESSIFTM ? 5 2.2. ADAPTING EXPRESSIFTM TO A RECOMMENDATION TASK 7 2.2.1. PRINCIPLE 7 2.2.2. FUZZY INCLUSION 7 2.2.3. RULEBASE 8

3. COMMUNICATION 9

4. CONCLUSION 10

5. REFERENCES 10

Table of figures Figure 1 – Overview of ExpressIFTM .................................................................................................................. 6 Figure 2 – Caption of the rule editor .................................................................................................................. 6 Figure 3- Linguistic variable "temperature" and its three terms ......................................................................... 6 Figure 4: Example of fuzzy inclusion ................................................................................................................. 8 Figure 5: Linguistic variable to quantify the match between profile/prerequesites and outcomes/goals ........... 8 Figure 6: Linguistic variable that represents the recommendation .................................................................... 9


1. Introduction This document is the final report of IO4, describing the recommender system which is involved in BIM4VET project. A recommender system [1] aims at ranking different items regarding users’ preferences. They have been intensively studied for the last decades with the proliferation of e-commerce (Amazon, etc.). There are two main approaches for recommendation. On the one hand, collaborative filtering approaches build a model from a user’s past behavior (in the Amazon example, the previously purchased items) as well as similar decisions made by the others users. This approach needs many data to recommend accurately items to users. This is not acceptable in our case since: - We do not store the previously completed training for the users ; - Users have different goals ; - We need a recommender from the very beginning : these approaches suffer from cold start problem. On the other hand, content-based filtering use the description of an item to propose items with similar description. This is more convenient in BIM4VET since we use the same referential of 25 competencies to describe a training, the users’ goal, etc. To address the training recommendation problem, we chose to use a fuzzy expert system to check if the training prerequisites match with user competencies and if the training outcomes match with the user’s goal. Fuzzy logic has been intensively applied to recommendation, by fuzzifying collaborative filtering and content-based filtering approaches [2]. The next section presents ExpressIFTM, the fuzzy expert system developed at CEA, and how the problem has been formalized. We then discuss about the access to ExpressIFTM recommendation. We finally conclude.

2. ExpressIFTM as a training recommender system 2.1. What is ExpressIFTM ?

ExpressIFTM is a software developed by CEA LIST since 2010 which aims at modeling expert knowledge with rules written as close as natural language as possible in order to reproduce automatically a human expert reasoning on data [3]. More precisely, ExpressIFTM is what is recently called an xAI (eXplanable Artificial Intelligence). This term, introduced by DARPA in 2016, applies to artificial intelligence softwares which can explain their decisions, learn from data (numerical data, images, etc.) and accessible from intuitive human-machine interfaces. These are precisely our roadmap while developing ExpressIFTM, as shown in Figure 1 and Figure 2, which present respectively an overview of ExpressIFTM and a caption of the user interface for rule authoring.


Figure 1 – Overview of ExpressIFTM

Figure 2 – Caption of the rule editor

ExpressIFTM allows to model knowledge using Fuzzy Logic. This logic, introduced in 1965 by L.A. Zadeh

[4] [5], is a multi-valued logic which truth values are comprise between [0 ; 1], instead of the classic two values “true” and “false”. Introducing continuous truth-values, fuzzy logic can handle the uncertainty of the knowledge (e.g. a « hot temperature ») and the inaccuracy of the input values. To achieve that, terms close to natural language are formalized by simple-shaped mathematical functions instead of mathematical operators (e.g. comparison operators). Figure 3 shows an example of three terms modeling respectively a cold, average and hot temperature: “temperature” is called a linguistic variable. Fuzzy logic allows using rules, which are close to the natural language, without mathematical operators, which tend to more comprehensible rule bases, particularly by non-mathematician users.

Figure 3- Linguistic variable "temperature" and its three terms

Whereas other fuzzy expert systems only use classical operators such as AND, OR and NOT, ExpressIFTM offers various operators for temporal [6], spatial [7], or even spatio-temporal reasoning [8]. These operators can be seen as the vocabulary, which can be used in the rules, representing knowledge in an even more natural syntax.


ExpressIFTM is thus extendable and we can formalize specific vocabulary for recommendation. In the next section, we formalize the recommendation problem.

2.2. Adapting ExpressIFTM to a recommendation task

2.2.1. Principle

In the BIM4VET project, the trainings prerequisites and outcomes and the user’s competencies and goals are represented the same way regarding 25 competencies identified before. Each competency receives a mark from 0 to 5. This is the only information we will provide for recommendation. Expert systems need rules to execute to make decision from the information described before. In the case of training recommendation, there are no such experts. We thus interested in common sense recommendation. We can make a comparison with recommendation performed by humans. For instance, when parents recommend to their child some optional classes, they are not experts in that field but they make their decision under the knowledge about their child and their partial knowledge about the classes. The common sense reasoning [9] is, in general, a challenge in artificial intelligence. Even if artificial intelligence can perform very complex reasoning on very large problems, the natural deduction, like humans do, is still a problem to tackle. Moreover, if we stick with the class recommendation example, human recommendation does not rely on matching exactly competencies and prerequisites. If the competencies approximately match with the prerequisites, then the class is still recommended. Our goal in this IO is to reproduce this common sense recommendation [10] [11] with ExpressIFTM. Formally, the problem can be seen as a ranking problem in which trainings are given a score: the higher the score, the higher the recommendation. Since ExpressIFTM needs rules to make decisions, we will provide rule like “IF the competencies match strongly the prerequisites AND the outcomes match the goals THEN the training is highly recommended”, and so on. As you may have noticed, we limited the criterion taken into account by the expert system to the ones which can be fuzzified. Indeed, the recommendation has to there are also the language and the distance of the training which have In the next section, we will see the formalization of the “match” keyword.

2.2.2. Fuzzy inclusion

Competencies, prerequisites, outcomes and goals can be seen as histograms, which values are from 0 to 5 over the 25 competencies. The goal is to formalize the “match” operator between two such histograms. After normalization (dividing by 5 each value of the histograms), they can be seen as membership functions and so the competencies, prerequisites, outcomes and goals are now fuzzy sets defined by: - A discrete domain with 25 values ; - A discrete membership function with the corresponding normalized histogram. The matching problem is now an inclusion problem: actually, the user’s profile matches with the prerequisites if the prerequisites are included in the user’s profile. In the crisp world, it consists in having all the bin of the histogram representing the prerequisites less or equal than the corresponding bin in the user’s profile. However, this is too strict for a common sense recommendation: for instance, if only one prerequisites’ bin is just a little bit higher, the training should have quite a high score (but not the maximum score). This trick is achieved by using the fuzzy inclusion defined in [12]:

𝑚𝑖𝑛 𝐻! 𝑣𝑎𝑙 , 𝐻!(𝑣𝑎𝑙)!"#

𝐻! 𝑣𝑎𝑙!"#


where A and B are two fuzzysets and HX(val) is the membership value of element val in fuzzy set X.

Figure 4: Example of fuzzy inclusion

The fuzzy inclusion gives values between 0 and 1 instead of a Boolean value. Hence, if we have an inclusion in the crisp sense of the word, the result will be 1. The more the level of the user is not satisfying, the less the value of recommendation. In Figure 4, we can see two fuzzy sets: the green one may represent a user’s profile, and the red one the prerequesites of a training. The inclusion is Hence, the “matches” operator which will be used in the rule is defined by:

A matches B ≡ incl(B,A).

2.2.3. Rulebase

The rule base can be summarized as one rule: the more the user’s profile matches the prerequisites and the outcomes match with the goal, the more the training is recommended. Unfortunately, at the moment, we cannot express “the more… the more” rules in ExpressIFTM. So we need to quantify the match with fuzzy terms like: matches strongly, mediumly, weakly, not at all. The terms are represented in Figure 5.

Figure 5: Linguistic variable to quantify the match between profile/prerequesites and outcomes/goals


The linguistic output variable is defined as a rank between 0 and 100 and is presented in Figure 6.

Figure 6: Linguistic variable that represents the recommendation

The rules are given in the association table below:

Profile matches with prerequesites

Outcomes match with

goals

Not at all Weakly Mediumly Strongly

Not at all Null Null Null Null

Weakly Null Low Low Average

Mediumly Null Average High High

Strongly Null Average High High

3. Communication In the system BIM4VET architecture, we decided to call the recommender as a webservice. The tangible table, which queries the database to select the catalogue of trainings, preselects the training regarding several criteria: - The distance range chosen by the user ; - The languages in which the trainings are given.


Actually, those two criteria are not fuzzifiable: ExpressIFTM can manage them but there is no point in let it manage them. The tangible table then calls the webservice with the user id and the concerned trainings. The webservice is a REST/JSON API which is simply used by sending this JSON snippet: [{"UserId": 1, "TrainingIds": [3, 4, 7, 8, 37]}] which tells the recommender that the current user’s id is 1, and the preselected trainings’ id are 3, 4, 7, 8, 37. The recommender answers with a JSON snippet like: [{“TrainingId”: 3, “Score”: 97}, {“TrainingId”: 4, “Score”: 67} …] which associates to each training id the computed rank. The tangible table displays then a sorted list of trainings regarding their rank.

4. Conclusion In this work package, we have defined a new paradigm for recommendation. Extending our fuzzy logic inference engine ExpressIFTM, we are able to rank the matching between a user and different trainings. The recommender uses only the information available in the database and performs common sense recommendation matching on one side the user’s competencies and the prerequisites and on the other side the outcomes and the user’s goals. The recommender is integrated in the demonstration as a webservice in order to be queried easily.

5. Références

[1] Y. F. B. O. F.O. Isinkaye, «Recommendation systems: Principles, methods and evaluation,»

Egyptian Informatics Journal, vol. 16, n° %13, pp. 261-273, 2015. [2] R. R. Yager, «Fuzzy logic methods in recommender systems,» Fuzzy Sets and Systems, vol. 136,

n° %12, pp. 133-149, 2003. [3] J.-P. Poli et L. Boudet, «A Modular Fuzzy Expert System Architecture For Data and Event

Streams Processing,» IPMU, 2016. [4] L. Zadeh, «Fuzzy sets,» Information and Control, vol. 8, n° %13, pp. 338-353, 1965. [5] L. Zadeh, «The concept of a linguistic variable and its application to approximate reasoning,»

Information Sciences, vol. 9, n° %13, pp. 43-80, 1975. [6] J.-P. Poli, L. Boudet et D. Mercier, «Online Temporal Reasoning For Event And Data Streams

Processing,» FuzzIEEE, 2016. [7] I. Bloch, «Fuzzy spatial relationships for image processing and interpretation: a review,» Image

and Vision Computing, vol. 23, n° %12, 2005. [8] J.-M. Le Yaouanc et J.-P. Poli, «A Fuzzy Spatio-Temporal-based Approach for activity

recognition,» Proceedings of 2012 ER Conference, 2012. [9] E. Davis et G. Marcus, «Commonsense reasoning and commonsense knowledge in artificial

intelligence,» Communications of the ACM, vol. 58, n° %19, pp. 92-103, 2015. [10] A. G. B. L. d. M. S. Sierra, «CBR with Commonsense Reasoning and Structure Mapping: An

Application to Mediation,» chez Case-Based Reasoning Research and Development, 2011. [11] I. M. N. N. V. S. B De Carolis, «Using common sense in providing personalized recommendations

in the tourism domain,» chez Workshop on Context-Aware Recommender Systems, 2009. [12] M. Wygralak, «Fuzzy inclusion and fuzzy equality of two fuzzy subsets, fuzzy operations for fuzzy

subsets,» Fuzzy sets and systems, vol. 10, n° %11-3, pp. 157-168, 1983.


BIM4VET Partners

LIST / Luxembourg Institute of Science and Technology

Cardiff University

LIST / CEA tech / Commissariat à l’Energie atomique et aux énergies alternatives

BIM4VET contact & website

Contact: [email protected] [email protected] Website: http://www.bim4vet.eu/

Documents

IO4. EXPERT SYSTEM FOR TRAINING RECOMMENDATIONS …