Retailing and social media platforms recommend two types of items to their users: sponsored items that generate ad revenue and nonsponsored ones that do not. The platform selects sponsored items to maximize ad revenue, often through some form of programmatic auction, and non-sponsored items to maximize user utility with a recommender system (RS). We develop a multiobjective binary integer programming model to allocate sponsored recommendations considering a dual objective of maximizing ad revenue and user utility. We propose an algorithm to solve it in a computationally eficient way. Our method can be applied as a form of post processing to an existing RS, making it widely applicable. We apply the model to data from an online grocery retailer and show that user utility for the recommended items can be improved while reducing ad revenue by a small amount. This multiobjective approach, which unifies programmatic advertising and RS, opens a new frontier for advertising and RS research and we therefore provide an extended discussion of future research topics.
• Information systems → Data mining; Collaborative filtering; Computational advertising.
A two-sided market is an intermediary platform between two distinct stakeholder groups that provide each other with network benefits. The media have long been held up as an exemplar. Media companies attract an audience of consumers with, for example, news or entertainment content (e.g., a TV program or magazine), and then sell advertisers access to the audience. The two stakeholder groups are consumers and advertisers. In recent years, a wide variety of diferent types of two-sided markets have been ∗Copyright 2019 for this paper by its authors. Use permitted under Creative Commons License Attribution 4.0 International (CC BY 4.0).
Presented at the RMSE workshop held in conjunction with the 13th ACM Conference
on Recommender Systems (RecSys), 2019, in Copenhagen, Denmark.
†Both authors contributed equally to this research.
introduced, often under the name platform rather than two-sided
market [
Related shopping examples include Ebay and Etsy. • Online travel agencies (OTAs) such as Orbitz, Expedia and Google match consumers with travel-related services (e.g., hotels, airlines, car rentals). AirBnB, VRBO and Uber are other successful platforms in the travel category. • Social media such as Facebook and LinkedIn match consumers with user-generated content and advertisers. • Online retailers often act as platforms, matching brands (advertisers) with customers.
This article studies recommendations on platforms with two stakeholder groups, users and providers (e.g., advertisers). The platform attracts users and matches them with providers through recommendations. We assume two types of provider recommendations, sponsored and non-sponsored, where sponsored providers (i.e., advertisers) pay the platform for access to the user and non-sponsored providers do not. For example, online retailing shopping platforms might show sponsored items from certain manufacturers, as well as items from manufacturers or vendors that the platform thinks the user would like.
Sponsored recommendations are often sold on programmatic
auctions where the platform can adopt various protocols to sell slots
[
The focus of this article is on deciding which sponsored items, if any, to recommend to each user. Platforms must balance between recommending sponsored items with high ad revenue and high user utility. We express the two objectives in a single equation: Q = γ (user utility) + (1 − γ )(ad revenue), (1) where γ is a parameter that determines how much weight to give to each of the two objectives. A platform that allocates sponsored recommendations with an auction has γ = 0 and is maximizing ad revenue (depending on how much the advertiser knows about the active user and how good its predictive models are, ad bids will also reflect user utility). RS that consider only the user’s utility such as those used to allocate non-sponsored recommendations have γ = 1. Rather than focusing on one or the other, this equation allows for a continuum between the two, competing objectives.
An idealized example clarifies this trade-of, which will be empirically investigated later in the paper. Suppose that advertiser A places the highest bid for a sponsored recommendation to some user. Advertiser B places a bid that is only slightly less than A’s, but would provide the user with substantially more utility than A. It would seem reasonable for the platform to give the slot to B even though it would generate less ad revenue. There could be ifnancial justifications for this decision as well, such as B may lead to higher lifetime value, where the user buys B, and ultimately is satisfied with it, and buys again from the platform into the future. If the retailer had shown A then either the user would not buy it, or would not be satisfied, and would be less likely to return in the future. Of course, an alternative scenario is also conceivable, where the ad revenue from B is substantially lower and/or B provides no advantage over A in user utility. In this second scenario the platform should display A.
This is an example of a multiobjective, multistakeholder
recommendation system (MRS) [
We formulate a post-processing binary integer programming
model to maximize the dual-objective function in equation (1) over
all users, subject to constraints that the advertiser has budget to pay
for the recommendation and the number of sponsored
recommendations to any user has not exceeded a threshold. We show how
to solve the optimization problem optimally in a computationally
eficient way. We test it using real data from an online grocery
store. With this approach, we make the following contributions.
First, we integrate the fields of computational advertising with RS.
Second, we show how to handle a situation where a platform is
interspersing non-sponsored recommendations with sponsored ones.
While sponsored recommendations guarantee the platform a
constant stream of revenue, by balancing user utility and ad revenue,
managers can generate higher customer lifetime value [
Search and display advertising is often sold using programmatic
auctions. An auction is held for an impression and diferent
advertisers place bids. Auctions are typically second-price, where the
winner pays the second-highest bid amount. Second-price auctions
have an optimal strategy for an advertiser of bidding the expected
value of the impression [
Our formulation of the MRS problem for sponsored
recommendations is related to the AdWords problem [
As we mentioned earlier, traditional RS have an exclusive focus on
one stakeholder, the user. In other words, traditional RS do not
accommodate more complicated systems with multiple objectives and
stakeholders. Multistakeholder recommender system (MRS) refers
to a designs that incorporate the interests of other parties [
Following [
We assume that a platform has item set I = {1, . . . , n}, some of which are sponsored. Our task is to identify top-k lists for each of m users in the set U = {1, . . . , m}. Let R be the m × n rating matrix, where rui is the rating of item i by user u and brui is the predicted rating from some traditional RS. Among the k items, at most smax ≤ k can be sponsored, which allows the platform to set an upper limit on the number of ads to show any single user. For now, we assume the ofline assignment problem , where the m users are known in advance. This situation arises when, for example, a system plans to send email promotions to m known users.1 Let aui be the ad revenue (or expected revenue, as discussed below) associated with displaying item i to user u. If item i is not sponsored then aui ≡ 0. Let βi be the total ad budget associated with item i, where non-sponsored items have βi ≡ 0.
We model the recommendation output via decision variable xui where xui = 1 if item i is included in the top-k list for user u and 0 if not. The m × n matrix of xui variables, denoted in matrix form as X, determines which recommendations will be delivered in the email promotion. We propose the following optimization model:
User utility γ zÕ Õ}| u ∈U i ∈I max
X subject to Õ xui = k
(∀u ∈ U) i ∈I xui ∈ {0, 1} (∀i ∈ I, u ∈ U) Õ
1(aui xui > 0) ≤ smax (∀u ∈ U) i ∈I Õ aui xui ≤ βi (∀i ∈ I)
{ brui xui +(1 − γ ) u ∈U We now discuss each part of the model, starting with the objective function in (2). The optimization problem is to select over all possible X values, i.e., the top-k lists for the m users, the values of X to maximize a two-part objective. The first part, labeled “user utility,” measures the total utility of the items recommended to all users. Product brui xui equals 0 for all items not selected, and the estimated utility for those selected; terms are summed over all users and items. The second part of the objective function, labeled “Ad revenue,” measures the total revenue from recommended sponsored items. Product aui xui equals 0 for non-recommended items and aui for recommended ones. The analyst-specified tuning parameter γ (0 < γ ≤ 1) determines how much weight to assign to the two parts. When γ = 1 the objective gives all weight to customer utility and produces the same solution as a traditional RS. When γ ≈ 0 the algorithm places nearly all the weight2 on ad revenue and mostly ignores user utility, which is how programmatic advertising auctions operate. The constraints are interpreted as follows: 1The “online” situation where we do not know which customers will require recommendations will be discussed in the future research section. This online situation arises when, for example, customers visit the system’s website, but the system does not know in advance which ones will show up on a given day. 2We do not allow γ = 0 because then the algorithm would have no way of allocating non-sponsored items when, for example, smax < k. When γ is very small, e.g., γ = 10−8, then up to smax sponsored items are selected based on the ad revenue, and the remaining non-sponsored items, which have ad revenue equal to 0 causing the ad revenue term in (2) to equal 0, are selected based on user utility. Constraining γ > 0 also addresses other degenerate cases, such as not having enough sponsored items to ifll nsmax slots. • (3): the number of recommended items is k for each user u. • (4): xui takes only values 0 or 1. • (5): the number3 of sponsored recommendations shown to user u must not exceed analyst-specified smax. Note that 1(·) is the indicator function, which equals 1 when the Boolean condition in the argument is satisfied and 0 otherwise. • (6): the total amount of ad revenue from item i does not exceed the budget for that item, βi .
We now discuss the two parts of the objective function (2), total user utility and total ad revenue. Figure 1 shows some possible curves for the two parts as a function of γ . First, suppose that γ ≈ 0, indicating that the algorithm should focus on optimizing ad revenue. The amount of user utility and ad revenue are indicated by the black dots on the left side (γ ≈ 0) of the respective plots. Ad revenue is at its maximum value, while utility is at its smallest value, although the ads should produce positive utility. If we start to consider user utility by increasing the value of γ two things will happen: (1) user utility will either increase or stay the same, and (2) ad revenue will either decrease or stay the same. An important question is, how? The best scenario is when both the the user-utility 3Sometimes it is more convenient to express this constraint as an upper bound on the fraction of sponsored items rather than the number. In such cases, one can compute smax by multiplying the upper-bound fraction by k and using the algorithm as stated. and ad revenue curves are concave, shown by black, solid lines. In this case a small increase of γ from 0 produces a large increase in utility with very little loss in ad revenue. The worst case is when both curves are convex, shown by red, dashed lines, where a small increase of γ from 0 results in a large decrease in ad revenue with little improvement to utility.
We start at the right sides of the plots at the red triangles (γ = 1), where the algorithm should focus exclusively on user utility, as with a traditional RS. User utility is maximized, while ad revenue is at its lowest point. The same conclusions hold that the concave functions are the best scenario and convex the worst. If we decrease γ , allocating some weight to ad revenue, and the curves are concave, then utility decreases by a small amount while ad revenue increases substantially. The curvature will be investigated in our empirical examples for a real company.
The curves will be concave when the platform has flexibility. When there are many sponsored and non-sponsored items with roughly the same utility, the platform can substitute one item for another in a top-k list without sacrificing much utility. Likewise, when there are many sponsored items with roughly the same ad revenue, substitutions can be made to improve utility.
To solve the optimization problem, rewrite (2):
Q = = where optimization criterion cui = γ brui + (1 − γ )aui . This expression is maximized by selecting the largest cui values, subject to the constraints, leading to Algorithm 1 below. The most computationally expensive part of the algorithm is sorting the mn elements of cui in descending order. Sorting has super-linear computational complexity, requiring on the order of mn log(mn) operations, so this algorithm will be computationally feasible for very large problems. Sorting is indicated by an index function that returns a user-item pair, e.g., no other c value is larger than cu(1),i(1); also cu(h),i(h) gives the hth largest c value. The variable su counts the number of sponsored recommendations shown to user u, bi totals the amount of ad budget from item i spent, and ku counts the number of items recommended to user u. The algorithm walks through cui one element at a time, starting from largest, until each of the m users have received k recommendations.
We discuss some potentially challenging situations. In line 11 of the algorithm there could be situations where a sponsored item (ai > 0) has the highest utility (i.e., higher than remaining nonsponsored items), but either the ad budget (βi ) for the item has been exhausted or the maximum number of sponsored item slots (smax) has been filled. In these cases, the algorithm should display the item as non-sponsored, since it has the highest utility. In order to have transparency, there should not be a sponsored designation displayed next to the item and the platform should not charge for the recommendation. Note that for γ < 1 the sort order of cui changes if ad revenue is set to 0. An item has more value to the platform
Algorithm 1 Multistakeholder advertising recommendation Require: Analyst sets the weight for user utility (γ ), and the max number sponsored recommendations shown to a user (smax).
Advertiser sets budget (βi ). 1: cui ← γ brui + (1 − γ )aui , ∀u, i 2: Sort cui in descending order with index function mapping h ∈ {1, . . . , mn} to user-item pairs [u(h), i(h)] such that cu(1),i(1) ≥ cu(2),i(2) ≥ · · · ≥ cu(mn),i(mn). 3: su ← 0, ∀u ▷ Initialize counter of sponsored ads to user u 4: bi ← 0, ∀i ▷ Initialize spent ad i budget item i 5: xui ← 0, ∀u, i ▷ Initialize item to be not selected 6: ku ← 0, ∀u ▷ Initialize number recommended to user u 7: for h ← 1 until all users are assigned exactly k items do 8: if ku(h) < k then 9: 10: 11: xu(h),i(h) ← 1 ▷ Add item to top-k list ku(h) ← ku(h) + 1 ▷ Increment recommendation count if au(h),i(h) > 0 and bi(h) + au(h),i(h) ≤ βi(h) and su(h) < smax then
Label item sponsored 12: 13: su(h) ← su(h) + 1 14: bi(h) ← bi(h) + au(h),i(h) 15: end if 16: end if 17: end for ▷ Increment sponsored count ▷ Tally ad spend when it is sponsored than when it is not: γ brui + (1 − γ )aui > γ brui . A simple way to handle this is to include all sponsored items twice in the list, once as sponsored and once as non-sponsored (aui = 0), and then do the sort. An additional if-then will be necessary to guarantee that if an item has already been selected to be shown as sponsored, it cannot also be selected as non-sponsored. Second, when allocating the last (kth) slot there could be multiple items with the same value of criterion cui . Since the criterion values are tied it does not matter which one is selected, and the algorithm could select one at random or some with tie-breaking rule. 3.1
The analyst should be aware that the two parts of the objective function have diferent units: aui is likely measured in some currency such as dollars or euros, while the units for brui depend on how utility is operationalized. For example, it could be the log number of previous purchases plus one, the probability of purchase, or the estimated rating on a five-point scale. One way to avoid this problem is to use “unitless,” standardized versions of aui and brui in the objective function.
A more desirable common unit would be to express brui in the
same monetary unit as aui , e.g., dollars. Assigning a monetary
interpretation, however, is a dificult task requiring substantial modeling.
In addition to quantifying the probability of a current purchase,
one would also have to consider the long-term implications. If
a recommended item produces a sale, the retailer receives some
short-term profit from selling the item, but also has a customer with
better recency-frequency-monetary (RFM) status: recency is now
zero days, frequency is one larger and monetary value is greater
[
Algorithm 1 requires the analyst to provide values of aui , the ad revenue associated with item i for user u. Notice that there are two subscripts, allowing diferent revenue amounts for diferent users and items. This subsection discusses the source of the aui values. We intend for our algorithm to be general and work for many current RS as well as systems that have not yet been devised. We assume that advertisers can set an ad budget and there are auctions for exposures. The advertiser can select items on which to place bids, and may have the ability to modify the bid amount based on additional information about the user, e.g., the advertiser could increase the bid for users in a certain demographic segment.
There are currently diferent revenue models for computational advertising. For example, advertisers can pay for a view, click, lead, conversion or engagement, such as signing up for an email list or newsletter. These models have corresponding metrics such as cost-per-view (CPV, which is often expressed in the cost for 1000 impressions, abbreviated CPM), cost-per-click (CPC), etc. In the simplest case where there is a fixed cost per exposure and the value aui is a constant. In more complicated cases, aui is an expected amount, which may be computed from a predictive model or various auction strategies. For example, there could be a model estimating the chance of a conversion and probability of winning the auction based on a certain bid. Or there could be a model where platform chooses from auction strategies like a first-price auction, secondprice auction, etc. However, we focus on the algorithm and take ad revenue as a constant. 4
We return to the example in the introduction between advertisers A and B, where A has bid more than B and won the auction. We investigate the sensitivity of the ad revenue and utility curves as a function of γ using data from an online grocery retailer. 4.1
The data is from an online grocery retailer in the United States, operating in 12 states and Washington D.C., with approximately 25,550 unique stock-keeping units (SKUs), 99,556 unique users, and 2,286,831 orders. The first step in building a RS is to define what an “item” is by aggregating the SKUs, since each size, flavor, etc. has its own SKU. For example, in the subcategory “soda,” the brand “Diet Coke” is separate from “Coke Classic.” The retailer provided a taxonomy of 136 subcategories, which, when crossed with brand names, gives 3,966 combinations. We clean the data further by combining diferent brand lines of the same product into a single brand. For example, within diapers “Pampers” is a single brand, while Huggies has many sub-brands such as “Little movers,” “Overnites,” “Pull ups,” etc., which are collapsed into one Huggies brand. We count the number of orders that contain each subcategory-brand combination. With this approach, the brand “Dean’s” in the Ice Cream subcategory is a distinct brand from Dean’s milk, Dean’s juice and Dean’s cottage cheese, but all flavors and sizes of Dean’s ice cream are grouped together. Many brandsubcategory combinations have small order counts. We minimize rare items by relabeling any subcategory-brand combination with fewer than 2500 orders as having brand “other,” which gives a total of n=1,021 combinations, defining an “item” for the algorithm.
We only examine regular users who have placed 15 or more orders and have purchased within the past six months (i.e., recency ≤ 0.5 years), giving a universe of 19,360 users. We draw a simple random sample of size n=2,000 users, and build our R matrix with roughly two million entries (2,000 users × 1,021 items). We use implicit ratings by taking the log number of past purchases, which provides gradation between products that are never purchased, purchased occasionally, and purchased often by a user.
In our data items were not sponsored. In order to study sponsored items, we assume that some fraction of the branded items is sponsored. Private-label items are not allowed to be sponsored because they are owned by the retailer. Likewise, we exclude “other” subcategories, e.g., “other cottage cheese” because they would be composed of multiple small brands. We assign a $1000 budget to each branded item for sponsored recommendations and assume ad revenue (ai ) drawn randomly from the Gaussian distribution N (µ = 1, σ = 0.2). The algorithm is then executed, and the results are compared for diferent values of γ . We conduct robustness check by varying the fraction of sponsored items. 4.2
Our algorithm post-processes ratings from a traditional RS to
address the MRS problem. Hence, an RS must first predict ratings for
items each user has not rated yet. We compare item- and user-based
collaborative filtering (IBCF and UBCF). For UBCF, we use cosine
similarity and vary the number of nearest neighbors (NNs): 5, 10, 20,
30, 50, 70, 90, 100 and 110. The performance of both traditional RS
is measured by the root mean square error (RMSE). We use five-fold
cross validation to evaluate traditional RS algorithms and follow
the established model-building process outlined in section 3.2 of
[
After estimating the ratings with UBCF, we test our algorithm. The goal is to recommend the top 20 items (ku = 20) to each user, among which, at most three (smax ≤ 3) can be sponsored. To put both ratings (rˆui ) and ad revenue (aui ) values on the same scale, we use their Z -scores to calculate cui . Figure 2 displays the results of the algorithm for user utility and ad revenue as a function of γ with diferent percentages (10%, 20%, 30%, 40%, 50%) of sponsored items. In all five cases, we see that as γ (the weight given to user utility) increases from zero to one, user utility increases and ad revenue decreases monotonically. The ad revenue plot (bottom) is mostly concave as we had hoped. The user utility curve (top), however, is convex for small values of γ , but mostly concave for large values.
When γ ≈ 0, all weight is given to optimizing ad revenue. This can be considered the status quo for sponsored ads. As shown in ad revenue plot (Figure2, bottom), when the number of sponsored items increase, there are more options to choose from and ad revenue grows. In the user utility plot, having fewer options in sponsored items would lead the algorithm to choose items with higher user utility, generating higher overall user utility in the case of 10%, compared to 50% of sponsored items.
After γ reaches approximately 0.5, the ad revenue plots show a faster decrease in the cases with a larger number of sponsored items. With more sponsored items in the inventory, as γ increases, cui puts more weight on user utility and the algorithm has more ways to select sponsored items that have lower ad revenue but higher ratings. At the end, the sum of ad revenue has a faster decline in the cases with more sponsored items as to when the number of sponsored items is limited, since those limited items might have had higher ad revenue with higher ratings. In all cases, the ideal scenario with both curves concave is partially realized.
We compare our proposed algorithm for MRS with traditional RS,
i.e., UBCF. While evaluating MRS, previous studies have shown that
a slight decrease in one objective improves over the previous system
if there is a substantial increase in competing objectives [
We can also evaluate our algorithm by plotting total ad revenue
against total user utility for diferent values of γ , as shown in
Figure 3. For comparison, the axes have been scaled so that 0 indicates
the minimum value of ad revenue or utility, and 100 indicates the
maximum value. The values of the points on the curve correspond
to diferent γ values. This resembles a receiver operator curve (ROC)
[
Although our optimization model and algorithm to solve it make important contributions to the fields of computational advertising and RS, we believe this to be only the beginning of a larger area for future research. We therefore provide an extended discussion of open research questions and extensions. The algorithm should be evaluated with field tests, on data sets from diferent industries, and with more traditional RS algorithms.
Understand the relationship between customer utility and
customer lifetime value (CLV). When a platform recommends items
with lower utility than others, how does this afect CLV? If this
relationship is understood, then the trade-of can be between ad
revenue and CLV, which are both measured in the same units (e.g,
dollars). Then γ could potentially become a variable that is selected
by the optimization algorithm instead of being set by the manager.
Game theory [
Understand how to handle situations where a user has high utility
for a sponsored item. For example, if the item with the highest utility
for user u is sponsored, should the system show a “sponsored” label?
On the one hand, showing such a label can decrease the probability
of purchase [
Model components of ad revenue (aui ). Ad auctions tend to be second-price, meaning that the highest bidder A pays the secondhighest bid. This is from a game-theoretic result that keeps the auction stable, but it makes modeling revenue more complex. The second-price issue could be addressed by multiplying the highest bid by a constant that adjusts to the second-highest bid. In the proposed model, the rank is determined by user utility as well as the bid order. The ranking function is Bid × P (click|Ad) = expected revenue. Example: A is a spammer P (click|A) = 0.001; B is a real company P (click|B) = 0.1. A has to bid 100 × B’s bid to win the auction, not just B’s bid plus some ϵ, and would have to pay an amount that provides expected revenue = B for each click. The second-price criterion is very dificult to model under these circumstances.
Develop algorithms for when items have diferent margins. A
grocery retailer might make more profit from selling a private-label
product, and recommend it as a non-sponsored item over other
items [
u ∈U i ∈I u ∈U i ∈I Optimize over smax. Setting smax to a small value restricts ad inventory, which will likely increase auction bids and thus the ad revenue per slot, whereas larger values increase the supply of inventory and reduce the bids. This raises the possibility of reducing sponsored ad inventory without reducing ad revenue. One would need to know the elasticities of the supply and demand curves, which could be an important study on its own.
Develop algorithms to handle online situation, where users arrive
over time and the platform does not know which ones will visit.
A simple example will illustrate the problem. Suppose that we are
to recommend k = 1 item and that there are n = 2 items A and B,
each of which is sponsored and with enough budget for exactly
one exposure today. There are m = 2 users, 1 and 2. The criterion
values are c1A = 3, c1B = 2, c2A = 5 and c2B = 2. Suppose that user
1 visits the platform first, and so showing item A is optimal since
c1A = 3 > 2 = c1B . Next user 2 visits, but since A has exhausted its
ad budget, item B is shown giving c2B = 2 and a total criteria of
3 + 2 = 5. The optimal solution to the of-line problem, however,
is to show user 1 item B and user 2 item A, giving a total criteria
of 5 + 2 = 7 > 5. Search advertising faces the same issue [
Improve the eficiency of Algorithm 1. The example in the previous paragraph also illustrates why the sorting in Algorithm 1 must be done over all mn items to find an optimal solution. Sorting within a user, as in the example, or within items may lead to a sub-optimal solution. While sorting has super-linear computational complexity and is therefore possible for large data sets, there might be ways improve the eficiency. For example, the slowest part of the algorithm is the sort, which can be sped up by reducing the number of cui values. If the inventory is, say, 10% of mn, one could drop all values less than, say, the 85th percentile, reducing the sort cost substantially. Knowing percentiles, however, requires sorting, unless the shape of the distribution is known. The cui values for the grocery data tend to be moderately right-skewed. The mean, variance and other moments could be found with a single pass, e.g., while computing cui . This would give information about the quantiles without sorting, which could then be used to drop useritem pairs with small cui values. 6
This study exemplifies how platform managers can consider both ad revenue and user utility while recommending items to the users. This is a paradigm shift from extant computational advertising approaches that focus on maximizing ad revenue and only consider user utility through click probabilities. Our approach treats the problem of MRS in the platform settings as a post-processing optimization step to be applied after user utilities have been estimated by an existing recommendation algorithm. Thus, our approach is not dependent on any particular RS algorithm, and can be widely applied. Our approach allows platform managers to trade-of ad revenue for higher user utility. Our empirical simulation suggests that shifting some weight to user utility will increase utility substantially while decreasing ad revenue only slightly, although this should be confirmed with field tests. Thus, we make computational advertising more user-centric by explicitly trading-of ad revenue to improve user utility while recommending sponsored items.
Advertising and RS have developed independently of one another and the two have starkly diferent underlying philosophies: the goal of advertising has been to persuade while the goal of RS has been to help users find items of interest to them. The ultimate goal—selling a product or, more generally, increasing CLV—may be the same, but the approaches are completely diferent. This study attempts to unify the two by balancing ad revenue and user utility in the specific context of sponsored recommendations. A more general question is whether this hybrid approach could be applied to other areas of advertising: can media selection and targeting work better by explicitly considering user utility? We believe this is a profoundly important question, as users find new ways to avoid ads.
We thank Özge Sürer for helpful discussions and Northwestern University’s Spiegel Research Center for supporting the second and third authors.