• No se han encontrado resultados

Prueba de funcionamiento añadiendo otro nodo

This section outlines the configurations of all system variables with respect to the mar- ketplace defined in Chapters 3 and 5. We discuss them in the same order as they appeared in the two chapters.

The following four variables are defined in section 3.2. They are briefly described and configured as below.

Srepresents the number of constituent recommenders that are incorporated in our marketplace. Thus,S = 3 because we have three constituent recommenders.

Tb represents the duration of marketplace calling for bids. We set Tb = 5 seconds

because, in practice, we find that five seconds allows sufficient time for the three constituent recommenders to compute their recommendations in most cases.

M represents the number of recommendations that the marketplace calls for in each auction round. We set M = 5 because five recommendations do not over- burden the users and five is a practical number in terms of the trials.

6A typical demographic method analyzes the characteristics of people (such as age, gender and

occupation) and groups people with similar characteristics. Then, it analyzes the attributes of recom- mendations (such as textual descriptions or contents of books, colour of material of clothes and price of products), and, finally, matches people with certain characteristics to recommendations with suitable attributes. We do not analyze people’s characteristics by age, gender and the like, but by their research interests since what we recommend are only Web documents that are relevant to a particular set of interesting topics. Thus, we group people by characteristics of their interesting topics and match people to documents with relevant topics. We do not consider this method a content-based one though it also analyzes the textual contents of documents. Rather, we consider it a demographic method. This is because content-based method compute similarities between documents, whereas this method computes similarity between the characteristics of people and the attributes of recommendations. For example, a group of people share the interest topic of “machine learning”, thus, all of them would probably be interested in documents related to reinforcement learning.

Chapter 6 User Evaluations of the Recommender System 106

N represents the number recommendations that are rated by a user in one rec- ommending round. We set N = 5 because we require each user to rate all five recommendations in each recommending round. Thus, having rated all displayed recommendations, the marketplace gives as much information about the user: in- terests to the recommenders as possible.

The following variables are defined in Equation3.18in section3.3.4. They are configured as below.

Qhrepresents theupqof a shortlisted recommendation. We setQh ∈ {0,1,2,3,4,5} because, in section 6.2, we have argued that five positive levels are sufficient to specify the qualities of relevant recommendations and, additionally, a level of zero indicates irrelevant ones.

δ controls the amount of reward to good recommendations and α controls the signal of deviation from equilibria that are delivered to the recommenders. We keep δ = 1.5 and α = 1.5 as per section 4.1.1 because they have been proved to be effective and we want our user evaluation to also be effective based on our empirical studies in Chapter4.

P¯h represents the market equilibrium price for recommendations with upq level

of h (as we stated in section 6.1➋, we do not use P∗

h defined in section 3.2 to

evaluate market convergence). However, to kick-start the process, so that we do not need lots of recommending rounds to get good recommendations, we set some constant values rather than let the marketplace find them all out for itself. Based on our empirical study in section 4.2, we set these constant values to ¯P1 = 110,

¯

P2 = 120, ¯P3 = 130, ¯P4= 140 and ¯P5= 150 to differentiate recommendations with differentupqs. The system gives clearer incentives to recommenders by fixing these equilibrium prices to constant levels rather than by using the historical average bidding prices for different upq levels. This is because the historical averaging prices change from one auction round to another and this makes the rewards to the sameupqrecommendations in different auction rounds differently. Consequently, this makes it difficult for a recommender to learn the user’s preferences (reflected by rewards) to its recommendations. Moreover, since the constituent recommenders

Chapter 6 User Evaluations of the Recommender System 107

do not know these fixed equilibrium price values, a marketplace with a set of fixed equilibrium prices does not affect their bidding strategies.

PM+1 represents the highest bidding price that is not shortlisted in one auction

round. It is the basic unit reward and it controls the amount of actual reward to an agent together withδ. Like setting ¯Phwith constant values, we setPM+1= 100 in order to give quick and clear incentives to recommenders (because constant unit reward makes it easier for recommenders to learn the bidding price deviations from the equilibria).

The last system variable we need to discuss was defined in section5.2for agents to learn users’ interests.

Grepresents the number ofinq segments of a constituent recommender. In prac- tice, based on our computation of the inqs of the recommendations of our three constituent recommenders (detailed in sections6.3.2,6.3.3and6.3.4), we find that six segments are sufficient to effectively differentiate recommendations in terms of

inq. Thus, we set G= 6 for our three constituent recommenders.