From Recommendation Systems to Facility Location Games

Facility location games [1992, 2010, 2014], which are extensively studied in economics, operations research and computer science, portray recommendation systems with strategic content providers incredibly well, due to the nature of recommendation systems described above. This is true since in many (and perhaps even most) recommendation systems, the computational task of matching users with content is carried out by modeling both into a joint metric space. In the vast majority of facility location games literature, it is assumed that users are attracted to their nearest facility¹¹ 1 Some exceptions appear in Subsection 1.4.. In the example above, this amounts to each user selecting the closest blog to her preference.

Given a strategy profile, i.e., the selection of each player for a particular mix of news and articles in her blog, the best (in terms of social cost) recommendation system matches each user with the nearest blog to her preference. However, the way the recommendation system operates affects the strategic behavior of players, and more precisely redefines player payoffs. As a result, it may induce a game with a different set of PNE, potentially with improved social welfare (i.e., the social cost with negative sign). This calls for the design of a mediator that takes into account strategic behavior of players, aiming at achieving low social cost in equilibrium.

Our conceptual contribution is the introduction of a mediator to facility location models, in the pursuit of better outcomes for recommendation systems with strategic content providers. We aim at designing mediators that a) induce a game with high social welfare in equilibrium, and b) intervene as little as possible.

To better describe the latter, let NIME be the No Intervention Mediator, namely the default mediator that displays each user with its nearest content. A question in this regard is whether a mediator that intervenes with the process can do better than NIME, in terms of social welfare. Clearly, one can design a mediator that dictates each player which content to select, thereby forcing the players to any desired strategy profile. Consider the DICT mediator that operates as follows: it commands each player to play a particular strategy, and if she disobeys DICT directs no user to her content. Indeed, by adopting DICT the mediator designer is guaranteed to have any specific strategy profile as the unique PNE of the induced recommendation game. However, this is a huge and potential harmful intervention, and doing so may lead to poor performance in case that specific profile is not materialized. Crucially, this can happen even when the players are rational, but some of the assumptions the mediator holds regarding the players are violated, e.g., players have constraints on their strategy space, slightly different payoff functions, they are able to form coalitions and so on.

A major aspect of our approach is the quantification of intervention cost. Given the above intuition, we define the intervention cost as the maximal increase in user social cost, which might be caused if the players use arbitrary strategies (i.e., select arbitrary contents). Formally, let $\mathcal{M}$ denote a mediator, $S$ be the set of all strategy profiles, $\boldsymbol{s}\in S$ be a strategy profile, and let $\textnormal{{SC}}(\mathcal{M},\boldsymbol{s})$ denote the social cost under $\mathcal{M}$ and $\boldsymbol{s}$. The intervention cost of a mediator $\mathcal{M}$ is defined as $\sup_{\boldsymbol{s}\in S}\{\textnormal{{SC}}(\mathcal{M},\boldsymbol{s})-\textnormal{{SC}}(\textnormal{{NIME}},\boldsymbol{s})\}$. Evidently, NIME has a zero intervention cost and high social cost, and serves as a benchmark for low intervention cost. On the other hand, DICT has high intervention cost but low social cost (when dictating the socially optimal strategy profile), and it serves as a benchmark for low social cost. The main challenge in our agenda is the design of a mediator with low intervention cost and low social cost in equilibrium.

This paper presents a case study in high welfare low intervention mediator design. The mathematical model we adopt in service of the above is based on pure location Hotelling games [1929], with the $[0,1]$ segment. Section 2 presents a formal mathematical model for the setting, stating former widely-known results for pure Hotelling games (or equivalently, recommendation games with NIME as a mediator) on the segment with uniform user distribution. In addition, in Subsection 2.3 we define the intervention cost, and bound the intervention cost of DICT.

The main technical results of the paper appear in Section 3. We introduce the Limited Intervention Mediator, LIME. We provide the intuition behind it, as well as a thorough example to illustrate how it operates. We then prove that the game induced under LIME possesses a unique PNE, which attains the minimal social cost. Then, we establish upper and lower bounds on its intervention cost, which are almost tight. We show that not only is its intervention cost lower than that of DICT for every $n$, it is also $O(\frac{1}{n})$. Since DICT and LIME have the same social cost under (the unique) PNE profile, the results given in this section provide a highly encouraging answer to the challenge given above.

We subsequently study three natural extensions of the basic setting. First, we discuss neutral mediators. Informally, a mediator is neutral if when two players swap their strategies, the mediator’s direction also swaps. We show an impossibility result for neutral mediators that aim at optimizing the social cost in the two-player case. Second, we design a mediator with a configurable intervention cost. This mediator is important where e.g. one seeks to design a mediator that minimizes social cost but is penalized for intervention (similarly to regularization in machine learning models). We show that for some cases of $n$, it induces Pareto optimal solutions for the setting. Third, we consider non-uniform user distributions. We propose the General Limited Intervention Mediator, which depends on the distribution quantiles solely. We then prove that it always induces a game with a unique PNE. This becomes even more striking when one recalls that PNE may not exist in facility location games (with no mediator) [1993], and are generally hard to characterize (see e.g. [1981, 2015]). We bound its intervention cost (for the uniform distribution), and show that its intervention cost is lower than that of DICT.

From an algorithmic perspective, we deal with one-dimensional problems. This may sound disappointing, but recall that we intentionally focus on relatively simple, structured problems, and that this domain is extremely well-studied in the facility location literature. Due to space constraints, the proofs are deferred to the appendix.

The notion of a mediator in a game-theoretic setting was first proposed by Aumann in his seminal work on Correlated Equilibrium [1974]. In the setting Aumann considers, a mediator may send signals to the players, where the signal is designed to drive the players to achieve better payoffs. The work of ? (?) grants stronger capabilities to the mediator, by setting constraints on participants’ behavior. A different type of mediator, considered by ? (?), is allowed to change player payoffs. As in our work, the mediator is designed to “force” players into playing some desired subset of their strategy sets with minimal interference with the payoff functions. Unlike the above work, where the mediator intervenes with the outcome of the game to improve the social welfare of the (strategic) players, in the model we study the mediator intervenes in order to improve the social welfare of the (non-strategic) users, as we describe next.

The work of ? (?) introduces a formal model of adversarial information retrieval. Each author (a player) has several strategies, where each strategy corresponds to a document she can publish. Each pair of document-query (where a document is a player’s strategy and a query represents a user) has a score, termed “quality”, which measures the extent to which the document is relevant to the query. The mediator (i.e., search engine) provides each user the document with the highest quality relative to her query, among those selected under the corresponding strategy profile. In their model, every author seeks to maximize a function that takes into account the number of queries for which her document is the most relevant (representing users directed to that document by the mediator) along with the document-query quality.

As ? show, displaying each user with the document with the highest quality (i.e., no intervention) may lead to deteriorated content and low user utility. They argue against no intervention, and claim that introducing randomization into the mediator leads to an overall user utility that transcends that of no intervention at all. Nevertheless, they neither a) provide a systematic approach for doing so under PNE profiles; nor b) show that a PNE always exists. Recently, ? (?) claim in favor of no intervention in a similar model, and show that no intervention policy leads to convergence of any better-response dynamics; thus, a PNE is guaranteed to exist. Importantly, ? (?) as well as ? (?) do not consider the task of mediator design for better outcomes.

The work of ? (?) is the first to consider mediator design in recommendation systems with strategic content providers, in a mathematical model that extends that of ? (?) and ? (?). They highlight several fairness-related properties that a mediator should arguably satisfy, along with the requirement of PNE existence. They show that the no-intervention mediator satisfies the fairness-related properties, but may lead to a game without PNEs. Then, they design a mediator that is based on the Shapley value [1952], prove it satisfies the fairness properties and the game it induces always possesses a PNE. On the other hand, ? (?) take an axiomatic approach, and do not address user utility as one of the axioms. Moreover, they show that the user utility under their proposed mediator can be arbitrarily low. In contrast, this paper stems from a user utility optimization point of view, and so are the solution concepts it proposes.

Several extensions of pure location Hotelling games have been suggested recently, assuming non-deterministic user behavior [2016, 2017, 2017]. More precisely, users do not select their nearest facility, but rather have a reaction function, mapping every strategy profile to a distribution over player indices, possibly skipping all options. The work of ? [2017] associates users with a reaction function motivated by decision theory literature, and shows that the induced facility location game always possesses a PNE, regardless of the metric space in which users are embedded. ? (?) show a result of similar flavor for a different reaction function. Interestingly, every reaction function can be implemented as a mediator. Unlike this line of research, in this paper (as in the information retrieval setting) the mediator is required to direct users to facilities w.p. 1, i.e., skipping all facilities is not an option. We elaborate on the latter in Section 5.

Finally, a different line of research in the algorithmic game theory literature [2007] is the study of facility location, in the context of approximate mechanism design without money [2009]. That literature deals with the case where only one entity dictates the place of a facility (or several facilities), while user preferences are their private information and are strategically reported (see,e.g., [2007]). In contrast, the setting we consider is a full information setting.

We remark that any mediator always directs a user to one of the facilities selected by the players; thus, the player payoffs sum to one under every strategy profile $\boldsymbol{s}$, i.e., $\sum_{i=1}^{n}\pi_{i}(\boldsymbol{s})=1$. Consequently, we view the social cost of the users as the public welfare. In the blogging example given above, we consider the distance between a user’s preferences and the actual attribute of the blog as the extent to which she is satisfied with that blog. Clearly, this perspective is identical to the one considered by [1929], albeit the motivation of [1929] is the physical world. Motivated by transportation cost, Hotelling [1929] defines the cost a customer (i.e., a user) suffers as the distance he has to travel to reach his nearest facility. This notion is naturally extended to the mediated setting.

We denote by NIME the No-Intervention Mediator. Namely, NIME sends every user to his nearest facility, breaking ties uniformly. Formally,

By implementing NIME we recover the non-mediated version of the setting, where under every strategy profile each user is attracted to his nearest facility, as is in pure location Hotelling games. Given $n$, we denote by $\boldsymbol{o}^{n}=(o^{n}_{1},\dots,o^{n}_{n})$, where $o^{n}_{i}\stackrel{{\scriptstyle\text{def}}}{{=}}\frac{2i-1}{2n}$, the sequence of $n$-socially optimal locations. These are optimal in the following sense:

It turns out that this is the unique socially optimal profile, up to renaming the players. Moreover, for any fixed strategy profile, it is apparent that employing NIME results in the lowest possible social cost w.r.t. that particular profile. Namely,

However, as we show later, the set of PNE depends on the mediator; hence, a PNE under NIME may not be in equilibrium under another mediator, and vice versa. Since we care about the social cost in equilibrium, this turns out to be crucial. Pure location Hotelling games (or equivalently $G(n,\textnormal{{NIME}})$ for any $n$) have been studied extensively in the past decade, and its equilibrium structure is widely known (see, e.g. [1975, 2014, 2018b]). For completeness, we state the following well-known results for $G(n,\textnormal{{NIME}})$:
$\bullet$ The only PNE for $n=2$ is $(\nicefrac{{1}}{{2}},\nicefrac{{1}}{{2}})$;
$\bullet$ There is no PNE for $n=3$, although a mixed NE exists [1982];
$\bullet$ For $n\in\{4,5\}$ a unique PNE exists (up to renaming the players). If $n\geq 6$, there are infinite PNEs.
For an elaborated discussion of these results, the reader is referred to [1975]. Table 1 summarizes the social cost for NIME. Notice the social cost under the worst PNE, which is a factor of two of the optimal feasible social cost.

Indeed, a mediator that implements the $n$-socially optimal locations as a PNE can greatly decrease the social cost. By intervening with the market and using a basic punishing mechanism, one can drive the players to play any arbitrary profile, by intervention that makes that profile the only PNE.

We denote by DICT the mediator³³ 3 One can think of other dictatorship mediators as well, by varying the punishment in case of disobedience. that dictates the strategy profile $\boldsymbol{o}^{n}$, the $n$-socially optimal locations. Namely, DICT operates as follows: given a strategy profile $\boldsymbol{s}$, let $A=\{i:s_{i}=o_{i}^{n}\}$. If $A$ is not empty, $\textnormal{{DICT}}(\boldsymbol{s})$ directs every user $t$ to its nearest facility from those of the players in $A$, i.e., $\textnormal{{DICT}}(\boldsymbol{s},t)_{i}=\mathds{1}_{i\in A}\cdot\textnormal{{NIME}}(\boldsymbol{s}\cap A,t)_{i}$. Otherwise, if $A$ is empty, it directs every user to a player selected at random. It is apparent that

However, sometimes the strategy profile being materialized is not the equilibrium profile. As elaborated above, a mediator intervening with the system to decrease social cost may hence have implicit negative effects. It is therefore highly desired to design a mediator that not only drives the players to a “good” equilibrium (in terms of social cost), but also does not intervene “that much”. To quantify the extent to which a mediator intervenes with the natural market, we introduce the following measure.

Indeed, the intervention cost captures the measure of intervention by the mediator. By definition of NIME, it has an intervention cost of zero, i.e., $\mathcal{I}_{n}(\textnormal{{NIME}})=0$ for every $n$. In addition, the intervention cost as defined above also demonstrates the great amount of intervention employed by DICT.

To illustrate it, consider the consequences of a strategy profile other than $\boldsymbol{o}^{n}$ being materialized under DICT. Take for example the two-player game, $G(2,\textnormal{{DICT}}{})$, and the strategy profile $\boldsymbol{s}=(s_{1},s_{2})=(0,1)$. This profile represents the case where the offered contents are varied to the extreme. Both players defy their commands; thus, DICT sends every user to a facility selected uniformly at random, and every user travels a distance of $\nicefrac{{1}}{{2}}$ in expectation. Summing over all users we get, $\textnormal{{SC}}(\textnormal{{DICT}},\boldsymbol{s})=\nicefrac{{1}}{{2}}$. The social cost under no mediation is $\textnormal{{SC}}(\textnormal{{NIME}},\boldsymbol{s})=\nicefrac{{1}}{{4}}$; hence, the intervention cost of DICT is lower bounded by $\textnormal{{SC}}(\textnormal{{DICT}}s)-\textnormal{{SC}}(\textnormal{{NIME}},\boldsymbol{s})=\nicefrac{{1}}{{4}}$. As we show in the appendix, for $n=2$ this is not only a lower bound but rather the actual intervention cost of DICT.

In fact, this profile can be extended to any $n$-player game, showing that $\mathcal{I}_{n}(\textnormal{{DICT}})\geq\nicefrac{{1}}{{4}}$ for any $n$. However, using a different construction we show a much tighter bound that depends on $n$. More precisely,

Due to its high intervention cost (among other properties) using DICT as a mediator may not be the best solution. In the next section we devise a mediator that implements $\boldsymbol{o}^{n}$ as the unique equilibrium, but unlike DICT has a substantially low intervention cost.

We now show that LIME is carefully constructed to mitigate players’ incentives. In particular, we show that $\boldsymbol{o}^{n}$ is the unique PNE of $G(n,\textnormal{{LIME}})$. First, we show that under any PNE, players are not encouraged to locate their facilities in PIIs.

Next, we leverage Lemma 2 to show that there is a unique equilibrium under LIME, which is composed of the $n$-socially optimal locations, i.e., $\boldsymbol{o}^{n}$. The case of $n=2$ is an interesting exception and is therefore discussed separately in Subsection 4.1.

See Figure 1.c for illustration. Importantly, Theorem 1 holds for any $\epsilon<\nicefrac{{1}}{{2}}$, and implies that $\boldsymbol{o}^{n}$ is an exact PNE of $G(n,\textnormal{{LIME}}{})$ and not an approximated PNE. In addition, under the profile $\boldsymbol{o}^{n}$ all the facilities are outside the PIIs; hence, every user is directed to his nearest facility. Consequently,

Therefore, the social cost of LIME under the PNE profile matches the optimal (see Table 1).

Having demonstrated that LIME obtains the optimal social cost in the (unique) equilibrium, we now claim that its intervention cost is low. By definition of the LIME mediator, it suffices to consider users inside PIIs only, as users outside PIIs are directed to their nearest facility; hence, they do not contribute to the intervention cost. First, we lower bound its intervention cost.

More precisely, the bound is $(1-\frac{\epsilon}{2})\left(\frac{2n-4}{n^{2}}\right)$, which increases as $\epsilon$ decreases. The expression given in Lemma 3 is obtained by taking $\epsilon$ to zero. Next, we move on to upper bounding its intervention cost.

The proof of Theorem 2 assumes for simplicity that $\epsilon$ (in Line 1 of Algorithm 1) is arbitrarily small. Notice that the upper bound almost matches the lower bound. We summarize the intervention cost of LIME in Table 1. Observe that not only does LIME intervene less than DICT, but also its intervention cost diminishes in a $\nicefrac{{1}}{{n}}$ scale, similarly to the optimal social cost (which is $\nicefrac{{1}}{{4n}}$).

Dilemmas other than intervention cost may arise when implementing a mediator. To illustrate such dilemmas and intertwine our recommendation games with social choice theory, we analyze the two-player game. As Theorem 1 shows, $\boldsymbol{o}^{n}$ is the unique PNE for $n\geq 3$. In $G(2,\textnormal{{LIME}})$, however, LIME does not induce a game with a unique PNE.

Notice that the selection of the socially optimal locations remains a PNE, similarly to greater values of $n$, but this PNE is not unique. To clarify the intuition behind the latter phenomena, consider the profile $(\nicefrac{{1}}{{4}},\nicefrac{{1}}{{4}})$. According to Line 1 in Algorithm 1, LIME directs all users in the segment $(\nicefrac{{1}}{{4}},\nicefrac{{3}}{{4}})$ to a random player w.p. $\epsilon$. Under this profile, random direction collides with directing these users to their nearest facility; hence, the randomness LIME employs for breaking symmetry is not enough and other tools should be applied. In fact, this problem can be easily solved by adding a lexicographic tie breaker in case both players select the same location. When augmented to LIME, this tie-breaker induces a game with the socially optimal locations as the unique PNE.

Indeed, our notion of intervention cost is not affected by such tie-breaking intervention. However, the mediator should treat players fairly, in some sense of the word. This fact is meaningful with respect to the generality of our agenda, since it introduces other considerations that involve the players rather than the users, borrowing ideas from social choice theory. The most immediate characterization of fairness in our setting is through the notion of neutrality. Formally, a mediator $\mathcal{M}$ is neutral if for every $i,j\in[n]$ and $\boldsymbol{s}=(\boldsymbol{s}_{-\{i,j\}},s_{i},s_{j})$ it holds that $\pi_{i}(\boldsymbol{s})=\pi_{j}(\boldsymbol{s}_{-\{i,j\}},s_{j},s_{i})$. Noticeably, the DICT mediator is not neutral, while the LIME mediator and all other mediators discussed in the paper and the appendix are neutral. The inability of LIME to handle the two-player game then becomes apparent: it turns out that for every neutral mediator that induces a game with a PNE, there is a PNE when both players select the same location (see the proof in the appendix). So, no neutral mediator implementing the socially optimal locations as a unique PNE exists for $n=2$, while for $n\geq 3$ we show LIME does precisely that under low intervention cost.

In some situations it might be beneficial to control the amount of intervention. A mediator may offer a tradeoff between social cost in equilibrium and intervention cost. Namely, we devise a parameterized mediator, where its parameter determines a desired intervention cost. The mediator attains the desired intervention cost but suffers some increase in its social cost in equilibrium (compared to the optimal one). For some values of $n$ we show decisive positive results for the applicability of such a mediator in creating desired tradeoffs. This extension is deferred to the appendix.

The results obtained for LIME do not hold for general user distribution, since players may have beneficial deviations. However, the case of non-uniform user distribution is appealing, especially given the fact that PNEs are hard to characterize (see, e.g., [1981, 2015]) or may not even exist [1993]. In this subsection we devise a mediator for general user distribution, showing initial results for this setting as well. Let $g$ be an arbitrary continuous function supported in the $[0,1]$ segment, and denote by $G(n,\mathcal{M},g)$ the game induced by the mediator $\mathcal{M}$, $n$ players and $g$ as the user distribution function. Let $q_{a}$ be the $a$-th quantile of $g$. Consider the General Limited Intervention Mediator described in Algorithm 2, and referred to as GLIME hereinafter for abbreviation. GLIME operates similarly to LIME, but carries out two things differently. First, it incorporates the quantiles of $g$ instead of those of the uniform distribution ($\boldsymbol{o}^{n}$ defined earlier). The other difference is given in Line 2. Instead of directing users in PIIs to their nearest facility outside that PII, it directs such users w.p. $\nicefrac{{1}}{{2}}$ to the nearest facility right of the PII (if such exists), and with the remaining probability to the nearest facility left of the PII. Importantly, GLIME induces games with a unique PNE, which is highly desired.

Next, we bound the intervention cost of GLIME. Since the analysis for general distributions is highly challenging, we leave it for future work and only focus on its intervention cost under the uniform distribution.

Indeed, while the intervention cost of GLIME is greater than that of LIME, it is better than that of DICT.