itec - European Schoolnet

More documents

Recommendations

Info

iTEC Project Title: ITEC-D10_2_V1-1 041102012.Docx Number of criteria Rating according to the i-th criterion Utility function for the i-th criterion Taking as a foundation existing theories and research in this field, (Adomavicius, Manouselis, & Kwon, 2011), (Lakiotaki, Matsatsinis, & Tsoukià ands, 2011), (Liu, Mehandjiev, & Xu, 2011) propose to treat the recommendation problem as a multicriteria decision problem according to the methodology in Roy (1996) to model this class of problems. According to this methodology, we can identify four analysis steps or levels when analysing a decision-making problem: • Level 1. Object of Decision. In this first stage we define the purpose of the decision system to be developed among those below: o Choice: to select the most appropriate elements for a given user. o Sorting: to classify all the elements available into categories, according to their appropriateness for a given user. o Ranking: to sort the elements according to their appropriateness for a given user, to provide a sorted list, from best to worst alternative. o Description: to describe the appropriateness of an element according to the rating criteria. • Level 2. Family of Criteria. The second stage is devoted to identifying the criteria to be used. A criterion is any artefact or factor that enables the comparison of different alternatives according to a given point of view (Roy, 1996). More precisely, we can formally define a criterion as a real-valued, non-decreasing function, defined for the set of alternatives , to compare the apropriateness of these alternatives: where: (3) i-th rating criterion scale Rating of according to the i-th criterion Rating vector of according to the criteria Single-criterion approximations have to take into account a single factor to capture all possible relevance aspects. On the other side, multicriteria approximations consider multiple factors to represent different points of view. The inherent complexity of multicriteria systems requires the identification of a consistent family of criteria (Roy, 1996): Let be the function expressing the preferences of a given user to decide between two alternatives and according to criterion . A family of criteria is consistent if it satisfies the properties below: o Monotonic: if and only if, for each pair of alternatives , for criterion , and for all criteria , then alternative is preferred to Page 39/96
iTEC Project Title: ITEC-D10_2_V1-1 041102012.Docx o Exhaustive: if and only if, for each alternative pair , for criterion then alternative is equivalent to alternative o Non-redundant: if and only if the removal of a criterion implies the violation of at least one of the properties below. Table 2 illustrates the most used criterion types (Jacquet-Lagrèze & Siskos, 2001). Table 2. Classification of criteria used in multicriteria recommendation systems. Measurable Ordinal Supports a quantitative measurement according to a rating scale. Defines an ordered set of acceptable values to use a qualitative or descriptive scale. Probabilistic Represents ratings’ uncertainty according to probability distributions. Fuzzy Ratings are represented according to the position in a given interval of a rating scale. • Level 3. Global Preference Model. This stage is devoted to developing a model to compute global preferences, i.e. a strategy to aggregate the different criteria to express the preferences among several alternatives in the element set. According to (Vincke, 1986) and (Jacquet-Lagrèze & Siskos, 2001) we can identify four modelling categories: o Multi-Objective Optimization models. Criteria are expressed as multiple constraints of a multiobjective optimization problem. We may cite (Malekmohammadi, Zahraie, & Kerachian, 2011) among the solutions using this model. o Multi-attribute utility/value theory (Value-focused). A value system is built to aggregate preferences for each criterion. After that, these marginal preferences are aggregated into a single utility function . These solutions are typically named MAUT (Multi-Attribute Utility Theory) or MAVT (Multi-Attribute Value Theory). Contributions based on this technique: (Abbas, 2010), (Lakiotaki, Tsafarakis, & Matsatsinis, 2008). o Outranking approaches. The main idea behind this approach is that a complete classification of solutions is not always needed to assist decision. It is based on a set of decisions for pairs of elements to obtain binary relations in a decision set. In this way, relations such as “a” is incompatible with “b”, “a” is preferred to “b” or “a” and “b” are equivalents insofar as preferences are concerned are possible. Contributions based on this technique: (Doumpos, Marinakis, Marinaki, & Zopounidis, 2009). o Preference disaggregation models. The preference model is inferred from global preferences by analysing past decisions. In many cases, these models are considered as sub-models from one of the models discussed above, as they try to infer the model in a concrete way, typically value function or outranking relations. Contributions based on this technique: (Fernandez, Navarro, & Bernal, 2009). • Level 4. Decision Support Process. Once the previous stages are completed, we need to design and implement the processes, methods or systems needed to select the appropriate set of alternatives. According to (Adomavicius, Manouselis, & Kwon, 2011), we can identify three MCDM recommender types: Page 40/96
Page 1 and 2: iTEC Project Title: ITEC-D10_2_V1-1
Page 37: iTEC Project Title: ITEC-D10_2_V1-1
Page 89 and 90:
iTEC Project Title: ITEC-D10_2_V1-1
Page 91 and 92:
Page 93 and 94:
Page 95:
show all

itec - European Schoolnet

Create successful ePaper yourself

Delete template?

Save as template?