MATTHEW ANDREWS, ULAS OZEN, MARTIN I. REIMAN, and QIONG WANG
One persistent trend in wireless communications in recent years is that demand for bandwidth is exploding, whereas revenue per subscriber is either flat or growing at a much slower rate. Hence, there is a mismatch between the amount of revenue coming in from the end users (EUs) of the network and the amount of investment that is required for wireless capacity to keep pace with demand.
Variable pricing and bundling are two traditional ways for revenue enhancement. In the first case (e.g. [1]), the service provider charges different fees for the same service provided at different times of day or different locations, effectively using pricing as a control device to fit demand within capacity while discriminating between users with different needs. In the second case, the service provider offers a single package that contains both high revenue-per-bit services and low revenue-per-bit services, and uses the gains from the former to cover the loss of the latter. The potential of either method ultimately rests on the pockets of EUs and, hence, is limited by the latter's budget.
In this chapter, we investigate an approach whereby the service provider can tap into an alternative source of revenue, originating from sales of advertisements or products and channelled by the content provider in the form of sponsorship of viewing. The gain comes from removing inefficiency of the current arrangement under which the content provider derives profit from its content (e.g., by showing advertisements) while EUs pay for the cost of viewing it. To maximize its profit, the provider naturally wants to increase the number of views of its content but is unlikely to get help from EUs who are wary about wasting their precious bandwidth quotas on content such as embedded advertisements. The reluctance is strong not only because the use of bandwidth can be heavy (in the case when an advertisement includes rich video) but also because EUs are uncertain and thus cannot control such use. Unlike voice service, which is charged by minutes, data service is sold in units of bytes and it is harder for the EUs to interpret how many bytes will be consumed when they perform a web action. Typically, when an EU clicks on a link, the only indication theywould get that the resulting web page is large is if it takes a long time to load.
We consider a solution that allows the content provider to “sponsor” its content so that it does not get charged to the EUs’ monthly quotas. The arrangement removes EUs’ concern about paying an uncertain amount of bandwidth cost for content that is of little immediate value to them. As a consequence, more content will be accessed, not only because some of it is free but also because users are effectively given more quota. The content provider's profit increases as long as the cost of sponsoring stays below the new revenue from increased viewing of its content. Moreover, the provider's image also improves as fewer users will think of it as an irresponsible party that pushes costly and worthless materials to them. Content sponsoring also benefits the service provider by giving it the opportunity to charge content providers who, as the “over-the-top” companies, have a greater willingness to pay than EUs. Income from this new source enables the service provider to recover the value of some of the mobile services that it is enabling and use that revenue to finance capacity expansion.
The concept of sponsored content introduces the following set of research questions. In this chapter, we shall mostly focus on the first one.
Networks that allow the option of content provider pricing were studied in References 2 and 3. The first paper considers a network utility maximization (NUM) setting where both the content provider and the EU have a utility and we wish to set prices at both end points and route traffic so that the aggregate utility minus the cost of the network is maximized. This setup was considered for both an Internet Service Provider (ISP) in a competitive market where the prices are determined by the market as well as a monopolistic ISP that has complete freedom to set prices. The second paper [3] considers a model where there is one local ISP for the content provider and one local ISP for the EUs. Two mechanisms are studied in this setting. In the first, there is no collusion and each ISP tries to selfishly maximize profit. In the second, a Nashbargaining solution is used to determine the profit division between the two ISPs.
There have also been a number of studies motivated by the question of network neutrality and some of these are related to the concept of sponsored content. In Reference 4, Economides and Hermalin consider a situation where the service provider partitions bandwidth and charges content providers for access to a given bandwidth partition. Conditions are presented for which a single partition is welfare maximizing. Mitra and Wang [5] study a model where the service provider maintains two pipes that can be accessed by content providers: a best-effort pipe and a managed bandwidth pipe that provides improved quality for an additional fee. In this setting, the service provider optimizes over the amount of best-effort bandwidth and the price of managed bandwidth. A key feature of this model is that when new applications enter the market, they typically rely on the best-effort pipe and so the rate of new service generation critically depends on the amount of best-effort bandwidth that is available. Lastly, in Reference 6, Njoroge et al. consider a model where each service provider controls access to a certain subset of EUs. Moreover, each service provider can charge a content provider for access to “its” consumers. The paper describes a setting in which this nonneutral regime can drive higher investment and, hence, maximize social welfare.
As already discussed, much of the focus of this chapter relates to how contracts should be designed between a content provider and a service provider. Similar issues of contract design have been discussed in a different context in the marketing discipline under the banner of “channel coordination” [7] and have been widely addressed in the supply chain literature (e.g., see Reference 8 for review). More recent work in References 9 and 3 has started the consideration of coordinating contracts in the network setting. Our paper fits closely with this stream of thoughts. In particular, our contracting arrangement is equivalent to the stylized buyback contract discussed in both References 7 and 8.
As already discussed, the focus of our presentation is on designing contracts between the service provider and content provider so that there is a win–win–win for all parties in the system (including the EUs). By this, we mean that all parties are better off than if the content was not sponsored. Our model differs from the prior models discussed earlier in that the underlying demand from the end users is uncertain. It is not simply a function of price and user/content provider utility. (In addition, the NUM model of Reference 2 assumes strictly concave utility functions as a consequence of demand being elastic with respect to price and, hence, does not capture a natural situation where the content provider is paid a fixed price per view by advertisers.) Lastly, our model extends beyond simple per-byte pricing and attempts to capture the notion of EU quotadynamics.
The fact that we treat underlying demand as a random variable has two effects on our analysis. First, the problem faced by the content provider is similar to the “Newsvendor” problem that is common in supply chain analysis. (In a Newsvendor problem, a retailer must purchase inventory to cover uncertain demand over a fixed time period. When the time period is over, the inventory only has a salvage value that is well below the purchasing price.) In our setting, the Newsvendor problem arises because the content provider must decide how much content to sponsor in a time period without knowing what the EU demand is. Second, the uncertain demand coupled with a reservation fee paid in advance will allow the service provider to control how much content is sponsored by the content provider, even when the latter's revenue grows in proportion to the number of views of its content.
We model the interaction between the service and the content providers as a Stackelberg game in which the service provider offers a contract parameterized by two fixed fees: a reservation fee proportional to the maximum number of views to be sponsored and a usage fee for each sponsored view that actually takes place. By accepting this contract, the content provider determines the maximum number of sponsored views and pays the corresponding reservation fee in advance, and assumes the payment of the usage fee for each view of its content by EUs, up to the aforementioned maximum.
We divide our discussions into two parts that reflect different ways of modeling EU payments.
We focus on the relationship between the service provider and a single content provider. We show that the aforementioned two-fee contract is incentive compatible: by charging a proper reservation fee, the content provider will be induced to choose the maximum number of sponsored views to optimize the total expected profit of both parties. It is also in the service provider's best interest to charge such a fee to bring about this outcome, because it can then use the per-use fee to transfer the profit to itself. In Section 10.2.5, we present a numerical example to demonstrate how the optimization might work in practice.
In this paper, we consider a contractual relationship between a single service provider (SP) and a single content provider (CP) in offering sponsored views of content (e.g., a webpage, an advertisement, or an online video). The situation is formally modeled as a Stackelberg game in which the SP is the leader who sets price parameters of the contract and the CP is the follower who responds by determining the maximum number of views it is willing to sponsor within a fixed period, for example, monthly. The purpose of sponsoring is to raise revenue by increasing the number of views of the said content. To model this effect, we assume EU will always access the said content if it is sponsored and with a smaller probability if it is not.
This problem can be naturally extended to the case of multiple CPs competing for the attentions of the EUs by sponsoring their own content. That situation leads to interesting competitive dynamics between the content providers and we leave its analysis for future work.
We define two basic models. In the first, the EUs pay for bandwidth on a per-byte “pay-as-you-go” basis. In the second model, we aim to capture the more common situation in which EUs pay for bandwidth via monthly quotas.
The EUs generate N (a random variable) potential views in a period for content items that for ease of exposition are all assumed to have the same size θ. (It would not be difficult to extend to a situation where θ is the mean size of heterogeneous content.) Let F be the cumulative distribution function of N and let . If the content is sponsored, it is viewed with probability 1. If content is not sponsored, it is viewed with probability . (The parameter q here captures the strategic behavior of the EUs.) Let denote a binomial random variable with trials and success probability q. Then .
The SP charges only the CP for sponsored content and the EUs for nonsponsored content, all on a per-byte basis. As we assume constant size for content, we denoteEUs’ payment per view for nonsponsored content by r. (The fact that an EU does not have to pay this amount for sponsored content is the main reason why the probability (=1) of viewing sponsored content is more than the probability (=q) of viewing nonsponsored content. The CP's decision is denoted by B, defined as the maximum number of views the CP is willing to sponsor. The actual number of sponsored views is, therefore, and the total number of views is . As mentioned earlier, the payment from the CP to the SP is structured as an ex ante reservation fee and an ex post usage fee. We define c to be the reservation fee per view and b to be the usage fee per view. Hence, the total revenue that the SP collects from the CP in a given period is .
It remains to define the advertising revenue earned by CP and the bandwidth cost incurred by SP. We assume that CP earns revenue a for each view and, hence, its total revenue is . (We assume that all parameters are known by the SP and leave the interesting case where a and q are private information of the CP for future work.) The cost for the SP is dependent on the total congestion on its network, which is given by a nondecreasing function . We let be the total load on the network excluding the EU's views of the CP's content. The total load without sponsoring is
and the total load with a sponsoring level of B is
So the expected congestion cost paid by the SP is
Our goal is to determine the maximum number of views that the CP should sponsor to maximize the total profit of both CP and SP. We also study the fees charged by SP that can induce this outcome.
We now describe a more refined model in which EUs do not pay for bandwidth on a per-byte basis. Each EU instead pays periodically for a base data quota and has the ability to buy additional quota in case the base quota is exhausted in a period. More formally, we assume there is a homogeneous population of K EUs, all of whom are served by a single SP who periodically charges a fixed subscription fee. At the beginning of each billing cycle, every EU gets a bandwidth quota that she can use anytime within the period. The starting point of the first cycle of EUs is uniformly distributed over a period length. When an EU has exhausted her quota before the end of a period, she can wait until she gets new quota at the beginning of the next period or refill her quota immediately by paying an additional amount d. The choice between waiting and refilling is assumed to be independent of the number of times that the EU has refilled before.
An EU's opportunity to access content within a unit time period is a Poisson distributed random variable with mean . The likelihood of an EU taking the opportunity to view the content depends on the amount of the quota she has for the remainder of the period (this is a strong assumption because it does not take into account that an EU may use her quota more aggressively when it is about to expire). We model EU's decision by a discrete-state Markov chain. States are indexed by , where EUs in the states of smaller index have more available quota left. EUs in state S have exhausted their quotas and are waiting for the next period to arrive. An EU in state i views unsponsored content with probability (), where . We remark that by using the Markov model, the periods will not have an equal length. However, we focus on this model as an approximation to a regular billing cycle because of its tractability.
We conclude this section with a brief discussion of what the SP needs to track in order to implement a sponsored content offering. As already mentioned, the interaction between the SP and CP would happen on a periodic basis, for example, monthly. In order to perform the correct optimization, the SP needs to know a number of parameters, for example, q, a, and the distribution of N. The SP could estimate q and the distribution of N by monitoring EU behavior. However, the correct value of a (the valueof a view to the CP) has to come from the CP itself. (The case in which the CP can try to “cheat” by giving an incorrect value of a generates a whole new set of interesting research questions that we will address in future work.)
Once SP has decided on the prices and CP has decided on the sponsoring level B, the system is then operated by the SP. It has to identify which traffic is associated with the CP and charge the CP or EU appropriately. This depends on whether or not the particular view is sponsored which in turn depends on whether or not we have reached the sponsoring level B.
Of all these implementation issues, probably the most challenging is identifying the traffic associated with the CP because that involves monitoring the traffic at line rate and determining the content provider from which the content is being requested.
We start with the simplest situation in which the reservation fee . Recall that B denotes the maximum number of sponsored views, denotes the (advertising) revenue to the CP of each view, and denotes the usage fee per sponsored view paid by the CP to the SP. The revenue received by the CP is and the cost paid by the CP to the SP is . The net revenue to the CP is
where , and we have used . Let denote the maximizing value of B for given b. Then if and if . If , the CP's net revenue function becomes a constant, and hence, the CP is indifferent between any choice of . In other words, if the CP does not need to pay a reservation fee in advance for sponsoring but price b is paid for each view of sponsored content, then the CP's optimal choice of is either zero or infinity with a transition point where the CP is indifferent between sponsoring any content or not.
Now, consider the case with a per-unit reservation fee . Then CP's revenue function becomes
This is a standard Newsvendor model. If , then , that is, the CP will not sponsor any content viewing if the combined reservation and usage fees exceed the additional revenue from advertisement. If N has a continuous distribution function (F has no jumps), then
In our setting, N is a discrete random variable, so F has jumps, and there may not be a such that Eq. (10.3) holds exactly. On the other hand, N is likely to be an extremely large integer (in our numerical Example, is on the order of ), so Eq. (10.3) will hold almost exactly. In particular, because is defined by
the error we make in assuming that Eq. (10.3) holds is miniscule and will henceforth be ignored.
First, consider the case with no contract cost (i.e., ) and no revenue from EUs (i.e., ). The SP's revenue from the CP is . The SP pays a congestion cost, given by a function C. Recall that is the “baseline” congestion without the EU, and the congestion cost is given by Eq. (10.1). We remark that congestion cost may not be a convex function of B even if C is linear because is concave.
The SP wants to choose b to maximize
Recall that with , , and with , . Thus the SP wants to choose either b as large as possible subject to , in which case the SP's profit is
or , in which case the SP's profit is .The SP will choose the alternative yielding the higher profit.
Now consider the case with fixed contract cost . The SP's revenue from the CP is . So the SP wants to choose c and b to maximize
Given the relationship from Eq. (10.3), the SP's problem is equivalent to choosing and to maximize
We now consider the optimal b for a given B. Looking at the first-order derivative of the profit function
we conclude that SP wants to set b as high as possible such that . Thus with B fixed, the SP's profit is
so that the optimal profit is attained using given by
Of course, in practice the limit cannot be attained: the SP needs to keep to induce the CP to choose the SP's desired . Thus there is some small such that and , that is, while a positive reservation fee is necessary to induce optimal B, the SP is better off to keep it as low as possible and derive all its profit by setting b as high as possible.
We also remark that because of Eq. (10.3), when finding we must optimize over the support of N. In reality, we would typically wish to restrict the optimization further to between (say) and because otherwise the system would be overly sensitive to the exact values of b and c.
Suppose that SP earns revenue from the EUs, that is, where r is the revenue rate. So the SP wants to choose c and b to maximize
If , it is not beneficial for the SP to offer sponsored content option to CP because the additional revenue from CP is not high enough to compensate for the loss in revenues from the EUs . This is the case if content revenues are low (i.e., for low a values) and/or the content is popular (i.e., for high q values).
Now, consider the case with . Similarly as above, the SP's problem is equivalent to choosing B and b to maximize
Checking the derivative with respect to b, we conclude that SP wants to set b as high as possible such that as above. We can again optimize over B, the only difference being the additional term .
The system performance, that is, the aggregate profit achieved by the SP and the CP, is given by
and the SP takes the following share
Let denote the system optimal number of sponsored views, that is, . The increase of the total expected profit from sponsoring is
and from the system's perspective, sponsoring only makes sense if
From the earlier discussion, we know that the inequality is not satisfied if . However, we remark that even if , sponsored content might not generate sufficient advertising revenue to offset thecombined effect of losing EU revenue and increasing congestion cost. Such situations will be identified by the optimization of when the optimal solution . This will happen if the congestion function increases steeply, for example, if,
where,
is an increasing and convex function of B, in which case for all .
Assume now that . As the system profit is maximized at and the SP's share is increasing with b for any given B (and, hence, CP's share is decreasing with b), a contract is Pareto efficient if and only if . Therefore, Pareto efficient contracts can be characterized by the single parameter where . Under the set of Pareto efficient contracts, any allocation of additional system profit can be possible. The SP's share of profit will have a range of , whereas CP's profit will have a range of . We remark that for CP to achieve profit of , CP needs to negotiate from SP the most favorable per-unit price b such that
When r is small, b that satisfies this equality might be negative.
The findings from the above analysis can be summarized as follows.
Optimizing the above determines whether sponsoring should take place.
We now present a numerical example to illustrate the above concepts. Consider the case of a large CP for which N has a truncated normal distribution with mean views per month. (The distribution is truncated to two standard deviations on each side.) The size of the content θ is 7.416 Mbit and the CP receives $0.0125 profit for each view (before paying any sponsoring charge to the SP). We assume EUs pay at a rate $10/GB for nonsponsored content and so this translates to a cost per view of . For the SP congestion cost, we set the baseline congestion and use a piecewise linear function given by
(This stylized cost function reflects, in a simple manner, the additional costs, such as lost customer good will, of exceeding the nominal system capacity.) We set , that is, an EU is five times as likely to view the content when it is free to them than they are when they have to pay for the bandwidth. In Figures 10.1 and 10.2, we show system profit as a function of B when the standard deviation of the underlying normal distribution is and , respectively. We can see that as the uncertainty in N increases (i.e., the standard deviation increases), the optimal amount of content to sponsor decreases because there is more likelihood that the realization of N will correspond to the steep part of the SP congestion cost curve. We can also see that although the system profit is not concave, it is simple to identify the optimal value of B.
In Figure 10.3, we fix B to its optimal value (in this case, views) for the case that the standard deviation is . We then plot both SP profit and CP profit as a function of b. (Recall that c is then determined from b and B via Eq. (10.3). In particular, as b increases from 0 to , c decreases from to 0.) As b increases, more of the excess system profit generated from the sponsored content is transferred from CP to SP. As a comparison, we also plot the baseline profit for SP and CP that would occur in the case that no content is sponsored (i.e., ).
Recall that we model quota usage via a discrete-state Markov Chain with state space . An EU in state i views the content with probability . EUs in state S have exhausted their quota and are waiting for the next period to arrive. (Hence, .) When a user exhausts its quota it can pay to immediately renew it via an additional charge. An EU's opportunity to access content within a unit time period is a Poisson distributed random variable with mean . We let K denote the number of EUs and so the total number of potential views is given by the random variable , which has a Poisson distribution with mean .
Let be the transition rate from state i to state (); and be the rates from state to state S and state 0, respectively; and be the rate from state S to state 0. Transition rates between all other states are zero. Rates () reflect how fast EUs run down their quota. Define
as the fraction of EUs who choose to wait for the next period after exhausting their quotas and denote as the rate of EUs in state using up their quotas. Then
The rate is inversely related to the residual time until the next billing cycle.
The transition rates imply the following steady-state probability for an EU to be in state i:
Although, as in the previous model, there are again two possible outcomes for each unsponsored potential view—viewed or not viewed, the probability of viewing is a bit more complicated. In particular, a potential view is associated with a user in state i with probability , and such a user views content with probability . Thus the number of actual views when there are m potential views is a binomial random variable , where . Let , and note that .
Let be the base revenue that SP receives from the EUs for their regular monthly quotas and let be the rate at which EUs refill their quota “early.” From the Markov Chain transition probabilities, we have
Hence, the SP revenue from the EUs is .
We begin with the case in which the Markov Chain transition probabilities do not change when CP's content is sponsored (i.e., the users simply switch their viewing from another content provider). We can obtain similar conclusions as before with Q playing the role of q. In particular, the CP's expected profit is given by
As in Section 10.2.1, this is again a standard Newsvendor model and so the CP decision leads to a relationship of the form
The SP choice of b and c is, therefore, equivalent to a choice of b and B so long as . (Once again, therefore, we need in order for the SP to be able to control the system.)
Similar to the per-byte cases, the system profit for a fixed value of B is
Hence, as in the previous model, we can optimize system profit via a univariate optimization over B. Once the optimal value of the B has been obtained, the split between the SP and CP can be controlled by an appropriate choice of b.
The above analysis assumed that sponsoring content does not have a material affect on how fast the EUs consume their quota. It just causes the EUs to consume additional bandwidth corresponding to the sponsored content. In reality, of course, the knowledge that the CP's content is sponsored may affect the dynamics of quota usage. In particular, sponsoring may slow the rate at which quota is consumed, thereby lowering the probability that a user has to refill. In this case, the sponsoring of the content is cannibalizing the revenue that the SP obtains from the EUs.
A general analysis of this case is beyond the scope of this chapter. Here we take an initial step in the following by using a simple case to highlight issues involved. In particular, we consider a two-state model as in Figure 10.4. EUs are in state 0 if they have available quota to use and in state 1 if they do not. We define as EUs’ transition rate out of state 0 and assume it is a decreasing function of B, that is, EUs exhaust their quotas more slowly if they get more sponsored bandwidth. Let be the (constant) fraction of EUs who do not refill their quotas, so is EUs’ transition rate into state 1. We also define as the transition rate at which EUs move from state 1 back to state 0 as a result of monthly replenishment of quotas. We assume to be an increasing function of B. As more content is sponsored, those who run out of their quotas will do so later in their monthly cycles, hence, get replenishment sooner and move back to state 0 faster.
Following the above definitions, the steady-state probability for EUs to be in states 0 and 1 are
Obviously, and () are transition rates and steady-state probabilities, respectively, for the case without sponsorship. Observe that is the rate at which an EU moves from state 1 back to state 0. As more sponsored content results in fewer EUs in state 1, it is natural to assume that and should be such that decreases in B.
As before, we define as the subscription revenue that SP receives from the EUs to pay for their monthly quotas. Define
to be the rate at which EUs refill their quota “early.” For the convenience of discussion, we denote as the probability that EUs access nonsponsored content. Let . Given B, the expected profit of the system is
Compare the above with the case without sponsorship, the difference in profit is
Examining each component in the above,
is the increase of advertising revenue because of sponsoring, which increases in B, that is, more sponsoring leads to higher advertisement revenue. The incremental cost from sponsoring contents
always increases in B. The change of the refill revenue
is always negative because from Eq. (10.6),
and the right-hand side decreases in B.
In comparison with Case 1, the additional advertisement revenue in Eq. (10.7) is higher here because increases in B instead of being fixed. This extra reward is accrued by the CP. On the other hand, the last component Eq. (10.8) shows an additional negative profit impact of content sponsoring, the cannibalization of the SP's refill revenue because of slower use of EUs’ quotas. Like the bandwidth cost, this loss of revenue is assumed by the SP. To recoup its loss and share the extra gain, it is in the SP's best interest to require a more demanding transfer payment than that in Case 1 from the CP.
In this chapter, we have introduced some of the natural research questions that arise if content providers in a wireless network are allowed to sponsor their content and thereby make it free to EUs. We considered the case of a single content provider, a single service provider, and a pool of EUs and showed that it is possible to design contracts that are win–win–win for all participants in system. This was first done in a setting where EUs pay for nonsponsored content on a per-byte basis. We then extended the model to incorporate EU data quotas. A key feature of all the models that we considered is that the content has uncertain demand, that is, the number of potential views is a random variable. This naturally led to a two-component price structure with a reservation fee and a usage fee.
We believe that this work can be extended in a number of natural ways. First, many interesting questions arise when we consider multiple content providers interacting with a single SP. For example, how should the SP decide which CPs should be allowed to sponsor content? Moreover, if different content has a different value of q (i.e., the probability that an EU views nonsponsored content is dependent on the identity of that content), how should that affect the cost of sponsoring?
Other potential variations include making the cost of sponsoring dependent on the user location, the time of day, or the current congestion in the network. We can also envisage a situation where sponsored content is provided its own quality-of-service guarantees by the network. Lastly, our entire framework is predicated on the knowledge of a number of different parameters, for example, a, q, and the distribution of N. One interesting network measurement taskwould be to monitor current network traffic and then estimate each of these parameters.