• No results found

Best Recommendation’s Identification

5.3 Evaluation

5.3.2 Simulating and Evaluating the Strategy

5.3.2.4 Best Recommendation’s Identification

To evaluate the learning strategy’s ability to identify the best recommendation (from the viewpoint of the user, i.e. the top upq item) quickly and bid it consistently, we use the same set of experiments that were used to assess market convergence. We then trace the topupq item highlighted by a randomly selected learning agent with a good recommendation method and a corresponding one from a non-learning agent in Fig.5.5 (a) and (b) respectively. We do this by plotting this top upq items’ bidding prices with circle points in the figures. To clearly display the points of the trace and not to damage the quality of lines (representing the three displayed bids), we do not display the points when this item is raised by other agents. From Fig. 5.5(a), we can see that this item’s bidding price keeps increasing till it converges to the first bid price of the displayed items. This means that as long as the randomly selected agent chooses this particular item to bid in an auction (after the market converges), it is always displayed in the top position displayed to the user. However, in contrast, this phenomenon in a market without learning agents proceeds slowly (see Fig. 5.5(b)). This means that a learning market can satisfy the user quicker than a non-learning one. Additionally, a learning market raises the best recommendation more frequently (39 times by the selected learning agent, see Fig. 5.5(a)) than a market without learning capability (13 times by the corresponding non-learning agent, see Fig.5.5(b)).

5.4

Summary

This chapter presents the learning problem that a recommender agent faces in our mar- ketplace. Specifically, the agent needs to classify its recommendations into differentinq

categories and to quickly identify and frequently suggest those categories that highly interest a user so as to maximize its revenue, while still satisfying the user. By sim- ulating and evaluating our Q-learning strategy, we show that the strategy can always come to the optimal solution, is able to quickly identify the effectiveinqcategories and frequently suggest items from these categories, and enables the agents to make more revenue than those without a learning capability.

Chapter 5 Learning Users’ Interests 92

In sum, this chapter has developed a reinforcement learning algorithm and a Boltzmann exploration strategy for the recommender agents to learn the users’ interests. This chap- ter has also proved the effectiveness of recommender agents using the learning strategy in our marketplace. With this learning capability, our marketplace converges quicker and suggests the best items more quickly and frequently than without it.

Chapter 6

User Evaluations of the

Recommender System

With the marketplace designed, simulated and formally analyzed, we now need to eval- uate the feasibility and the efficiency with real users of our market-based approach to recommender systems. To do this, we implemented a market-based recommender system that incorporates three typically-used recommendation methods (content-based, collab- orative and demographic). We then arranged for a number of people (thirty-one in this case) to use our system so that we could record various aspects of their interactions and the system’s outputs. These records were then analyzed in order to provide a user evaluation of the efficiency of our system.

With the user evaluations of our system, we have shown that(i)multiple constituent rec- ommenders contribute to the recommendations that are placed in front of users through the marketplace, (ii) the marketplace converges with respect to most of the users,(iii) the market-based recommender’s top recommendation is the best item of those suggested by whatever constituent recommenders for most of the users most of the time, and(iv) the marketplace is able to seek out the best recommendations for a given user most of the time and place these among the top positions in the recommendation sidebar most of the time. By undertaking these user trials, this chapter contributes to the thesis by showing that the market-based approach is capable in practice (as well as in theory) of coordinating multiple recommendation methods and effectively identifying the best recommendations quickly and frequently.

Chapter 6 User Evaluations of the Recommender System 94

Specifically, section 6.1 defines the metrics that are used to evaluate our system. Sec- tion 6.2 outlines a user’s task in terms of using the recommender system. Section 6.3 details the system configurations in terms of the three recommendation methods. Con- sequently, section 6.4 discusses the evaluation results and section 6.5 summarizes our findings from this aspect of the work.

6.1

Evaluation Metrics

In seeking to evaluate our system with real users, the first step is to identify the properties that we would like our market-based recommender system to exhibit. This then gives us the requirements against which we perform our evaluation. In particular, we are interested in the following metrics (the first, second and fourth metrics are the most important system properties selected from those discussed in sections 4.2 and 4.4; the third metric is defined according to the essential purpose of the market-based approach to recommender systems discussed in section 1.1):

Balanced Output Contribution

There are three constituent recommenders incorporated in our marketplace (each of which exploits a different recommendation method, see section6.3for details). Here we term the recommendations suggested by one constituent recommender and eventually displayed (shortlisted) to users as that recommender’soutput con- tributions. For a given user, it might be the case that one recommender makes the significant majority of output contributions and the others make very few output contributions. In this case, we say that the recommender that contributes the majority of outputsdominates the marketplace. Such domination with respect to a specific user is not necessarily a bad thing (because it means the dominating recommender has learnt this user’s interests more efficiently and therefore con- tributes more good recommendations than the remaining recommenders). How- ever, it would be a problem if the same method dominates the user population across all their various interests. Indeed, if the users’ interests literally follow a uniform distribution among a number of potential interesting browsing topics (meaning that different users have different interests and no one interesting topic

Chapter 6 User Evaluations of the Recommender System 95

dominates the majority population of the users), if one constituent recommender dominates the marketplace for most users most of the time, the marketplace es- sentially degenerates to the single dominant method. To capture the fact that multiple methods actively work simultaneously, generally speaking, we expect the different constituent recommenders to make balanced (broadly similar) output contributions with respect to a number of users with various interests. This met- ric is important because, on the one hand, compared to the equal opportunity of bidding that the marketplace gives to different constituent recommenders (dis- cussed in simulations in section 4.2.4), this metric further evaluates the fairness of the marketplace in terms of output contribution in a real environment (with real users and real recommendations) and this cannot be done in the design and simulation stages. On the other hand, the balanced output contribution metric eventually verifies the fact that the marketplace works as a means of coordinating multiple different recommendation methods and ensures the marketplace does not degenerate to a single method.

Market Convergence

As highlighted in sections4.2and4.4, market convergence is a key desirable char- acteristic of our system. Such convergence is important because it ensures that the system makes an effective shortlist of recommendations, gives the appropriate incentives to the constituent recommenders, gives equal opportunity of bidding to different constituent recommenders, makes the marketplace stable, and seeks out the best recommendations frequently. Now section4.2.1showed that convergence happened with our simulated users, but here we want to ensure that it does also happen with real ones.

In more detail, to demonstrate the market convergence for each user, we evaluate whether the bidding prices for all upq levels1 have an overall tendency to con- verge to their corresponding equilibrium prices (meaning that the bidding prices for items with a upqvalue of 1 tend to converge to the equilibrium price forupq

of 1, prices for items with a upq value of 2 tend to converge to the equilibrium price forupqof 2, and so on). In our previous simulations (section4.2), we evalu- ated the convergence by validating whether the bidding prices for each shortlisted

1The upq of a recommendation is identical to the user’s rating throughout this chapter. This is

Chapter 6 User Evaluations of the Recommender System 96

advertisement slot converged to a small oscillation around a constant level after a number of auction rounds. Thus this approach directly observes the economic market equilibrium price (the point where the demand meets the supply) for each advertisement slot. However, in practice, we need quick market convergence so as to quickly suggest high quality recommendations to a real user without spending a vast amount of time (i.e. hundreds of auction rounds as per section 4.2) be- fore good recommendations come out. Therefore, instead of evaluating how prices deviate from corresponding equilibria of different advertisement slots, we seek to evaluate the tendency towards market convergence by evaluating how the bidding prices for each upq level deviate from their corresponding equilibria prices. In- deed, we have already demonstrated that the convergence of prices for different

upqs is consistent with the convergence of prices for different advertisement slots in section 3.5. This is because as the bidding prices for different advertisement slots converge, prices of recommendations of a specificupq also converge (other- wise, with recommendations of at least one specificupqlevel not converging, prices with respect to advertisement slots do not converge). More formally, with respect to a specificupqlevel ¯Q, in a specific auction round ¯a, there areNQ¯ recommenda-

tions (from whatever constituent recommenders) with bidding pricesP1,P2, · · ·, PNQ¯ being rated at ¯Qlevel by a user and the corresponding equilibrium price is ¯P

(see section6.3.1for the definition of equilibrium price for one specificupqlevel). The deviation from equilibrium for ¯Q in auction round ¯a is then calculated as:

DQ,¯¯a= 1 NQ¯ NQ¯ X i=1 |Pi−P¯|. (6.1)

With this definition, the ideal market convergence would be that DQ,¯¯a converges

to zero with ¯a increasing for all different ¯Qs. However, rewards to high upq

recommendations give more confident incentives to recommenders than those to low upqones. Thus, it takes more time to converge for lowupq levels than high ones. In some cases, recommenders fail to learn users’ interests with respect to very low upq recommendations and the market cannot converge on these upq

levels. Therefore, in practice, we expect DQ,¯¯a to converge for most ¯Qs, especially

the high ones.

Chapter 6 User Evaluations of the Recommender System 97

As discussed in section 1.1, we ideally want the market-based recommender to always perform as well as the best of the constituent recommenders (whatever that is for the given user in the given context). Thus, we view our market-based system as a meta-recommender whose recommendations are those shortlisted items that are displayed to the users. Specifically, we expect to see that the first displayed recommendation (in the first slot of the recommendation sidebar, see Figure 1.1) suggested by our market-based recommender at any auction round is as good, from the user’s viewpoint, as the best of the first bid items suggested by all constituent recommenders. This is important because a good recommender system is one that makes the best recommendations. To do this, we define a metric, called peak performance. A constituent recommender’s peak performance at a given auction round is defined as the upq of its first bid item, whereas the market- based recommender’s is defined as the upq of its first displayed item. Note that in the case of a constituent recommender that has no items shortlisted at an auction round, its local peak performance is zero. Therefore, we expect the market- based recommender’s peak performance to be as high as that of the best of the constituent recommenders’ for most auction rounds for most users. To do this, we define the effective peak performance point as an auction round in which the market-based recommender’s peak performance is as high as that of the best of the constituent recommenders’. With this, we statistically evaluate how many times the marketplace performs as well as the best of all constituent recommenders for all users over all auction rounds.

Best Recommendation Identification

The previous evaluation metric evaluates the qualities of recommendations from the perspective of comparison between the market-based recommender and its constituent recommenders. However, whether a recommender is able to satisfy a user is eventually decided by the user. Thus, we also need to evaluate the qualities of recommendations from the users’ point of view. This is the most important property of any kind of recommender systems. Indeed, whether a user likes a recommender system eventually depends on its ability to identify the best recommendations to him. Here, we define thebest recommendationsas the items of the highest twoupqlevels (i.e. “4” and “5”, see section6.2for the configuration of

Chapter 6 User Evaluations of the Recommender System 98

the rating value range of recommendations). As discussed in section4.4, we want our system to be able to identify and frequently suggest the best recommendations to users. To evaluate this, we define two measurements: qualified recommending round and satisfied recommending round. Specifically, with respect to a particular user, a qualified recommending round means an auction round with at least one best recommendation displayed in any advertisement slot of the recommendation sidebar, whereas a satisfied recommending round means an auction round with at least one best recommendation displayed in any of the first two advertisement slots. Thus, a satisfied recommending round must be a qualified recommending round, but a qualified recommending round may not be a satisfied recommending round. With these two measurements, we evaluate how the numbers of qualified and satisfied recommending rounds, with respect to a given user, are compared to the total number of recommending rounds.

With these metrics in place, we now outline the user trial process.