This post was originally published by Riley Howsden at Towards Data Science
Assessing the value that personalization brings to players.
Part two of this blog has been finished and is located here.
The concept behind a rotating store is relatively simple; only a fraction of a catalog’s items are purchasable at any given time. The store effectively hides the remaining products, and transactions on that subset are impossible until a “rotation” occurs. In most cases, a rotation takes place at some predetermined cadence known by the consumer. In this sense, a rotating store shares similarities with a “flash sale.” Under this scenario, items that were likely purchasable beforehand have now been discounted for a short time, usually 1–2 days. However, the mechanism is slightly different as the “sale” in a rotating store is the opportunity to purchase the item. While possible, it does not have to include an actual discount.
Similarities to Clothing Retailers
In theory, many physical clothing retailers act as a rotating store by shifting their displayed inventory to match the seasons. However, instead of hiding items away, the fraction of the catalog that has expired is usually sold off at a discount or sold in bulk to a reseller; this action is often due to capacity constraints. Therefore, the mechanism is quite different for physical retailers and is more focused on seasonality than the rotation described above. Online clothing retailers, such as Rue La La and Gilt, have a constant inventory rotation with purchase timers. Often offers and steep discounts are paired together, and while things can go on sale multiple times, the underlying catalog is ever-changing itself. There are no guarantees the item will return.
Rotating Stores in Video Games
While some clothing retailers run on a structure similar to a rotating store, the biggest adopter of this monetization strategy has been gaming. While mobile games were the first to explore this space, the rotation feature has landed in some of the most popular video games, such as Fortnite, VALORANT, Apex Legends, and Fall Guys. There are two significant reasons why rotating stores are becoming more commonplace in the industry; the underlying good is digital, and it is exclusive. The digital aspect of items allows for more effortless rotations. While an in-depth algorithm is a possible solution for better selections, a random draw of items from the inventory can act as a reasonable starting point. Also, there are no capacity constraints since this inventory requires no physical space; theoretically, an infinite number of units are available for sale. As for exclusivity, most video games have a monopoly over the content they create; this means less risk for a rotating store model as there are no competitors to swoop in and sell those same items directly to the player.
Recap of Rotating Store Properties
- Rotation — the storefront switches purchasable products frequently, and the change cadence is often known.
- Limited — only a fraction of the entire catalog is purchasable at any given time.
- Discounts — discounts/promotions can occur in rotating stores, and due to the quick feedback loop, they can be optimized more easily.
- Digital — rotating stores are more common in spaces where the product is digital, such as video games.
- Exclusivity — sales success heavily relies on the exclusivity of the content with the store.
The concept of personalization is relatively straightforward and not limited to rotating stores. Instead of serving the same action to everyone, a system assigns an experience that best suits each individual. However, the depth at which that system carries out personalization can vary substantially. At some companies, personalization teams are driven by heuristics and manual curation, whereas at others, these teams consist of data scientists who primarily leverage recommendation systems rooted in machine learning. Overall, there are three standard options for personalization:
A random selection of products for each player may seem like a poor choice, and for the most part, it is. In theory, random meets the requirements for personalization but at the same time feels entirely impersonal. The one redeeming quality of a random system is that it generates an invaluable set of data for optimizing future recommendations that aren’t random. In short, a random solution scales well but is often a poor player experience.
On the opposite end of the spectrum, one could hire multiple product experts who assign individuals items that are deemed best for each player. This situation is not feasible at granular levels, and curation is usually a rough cut across all players using a few heuristics. For example, any player over the age of X will receive item A, while any player below the age of X will receive item B. That said, a curated approach can result in a better player experience but scales very poorly.
Finally, an algorithmic approach can dynamically learn from the player and assign items based on their past actions. The more information this system has from the player, the better their personalization will become. Overall, using an algorithmic method will have the most favorable player experience and scales well, but it requires a healthy data infrastructure for building sophisticated recommendation systems.
Personalization may sound like a no-brainer; we should apply it everywhere! However, in many situations, the cost of creating a personalization service far outweighs the benefit; we want to assess the value of the system before just building it blindly. Luckily, a savvy person can identify most of the critical factors that determine success beforehand; here are a few that are important:
Diversity of Content
A recommendation system tends to perform better when there is a large variance in the underlying content. If all of the content is very similar, the magnitude between users’ preferences will be small. For example, if Netflix only streamed horror films, the underlying recommendation system would be less valuable as variance within horror films is smaller than the variance between other film genres such as comedy or drama.
Diversity of Users
A recommendation system tends to perform better when there is not only a large variance in the underlying content but also the consumers of that content. If all users have very similar preferences or playstyles, a popularity heuristic may be a decent indicator of interest for the entire user base. For example, if Netflix streams all film genres, but those genres’ ranks were similar from one person to the next, personalization would be less valuable.
Size of Content
A recommendation system tends to perform best at scale. If a catalog is small, ranking problems will generally yield unnoticeable results. One reason for this is that awareness is not a problem; if it is easy for the user to iterate through all of the content within a reasonable amount of time, personalization is not paramount in unlocking discovery.
A recommendation system tends to perform better when the interface’s underlying structure can also be optimized. This is hardly the case, and recommendations must fit into a previously designed rigid structure without personalization in mind. The selection between a standard store and a rotating store, such as discussed earlier, can also impact personalization. Here, the lack of access becomes the focus instead of a lack of awareness.
Now that we have a general idea of a rotating store, which factors influence the environments in which recommendations can thrive, and three methods for personalization, let’s go through the steps to determine which of those techniques might bring the most value.
For now, we’ll assume that content diversity in the upcoming scenarios is ample. If this was not the real-world case, it could probably be addressed by creating more content in spaces where the catalog lacks. On the other hand, it is unlikely a lack of diversity in users and, therefore, user preference across products can be as quickly addressed. Feasibility will need to be assessed beforehand. Still, for most games and products, variety in actions should be ample. For simplicity, let’s assume that our rotating storefront consists of a single item from the catalog that rotates once a week. Note that most of these scenarios are not realistic but are used as building blocks to assimilate personalization value.
Scenario One: Everyone wants everything equally.
Regardless of the item and the player, there is an equal chance that a player will make a purchase. This example is most definitely a stretch and is only used to set a foundation. Note that the average probability of purchase for all twenty items here is 3%, which is arbitrary but used to establish a baseline for comparison with other scenarios.
Here, a personalization strategy is not needed; if you have no way to predict who will purchase what, there is no need to waste time on a curated or algorithmic solution. A random selection for all players, personalized or not, will prove to be just as effective.
Scenario Two: Everyone wants everything unequally.
This scenario is also unlikely but far more realistic than the previous and helps demonstrate that content has varying popularity levels. In many systems, a power law is often assumed, such as the Pareto distribution (see 80–20 rule). Note that the average probability of purchase across all items here is still 3%, just distributed unevenly.
Once again, in-depth personalization carries no value in this case. It should also be clear that a random selection strategy will perform very poorly unless that selection is weighted accordingly. In the previous scenario, the system chooses each of the twenty items at random 5% of the time. To be optimal in this scenario, the system should now select the top item way more often; in fact, some things might not be worth picking at all.
Scenario Three: Only some people want specific items, but the overall demand for an item is equal across the entire population.
An extension of scenario one, but now each user has their preferences. Intuitively, this makes sense; it would be quite a surprise if everyone wanted everything equally. However, this scenario doesn’t expect one item to be more popular than the next when we look at the entire population. Therefore, the users’ aggregate preference will be the same as the image shown in scenario one, but individuals could have various distributions. I’ve naively simulated how some users might distribute their total “60% of purchase probability” (20 items, 3% each) in the chart below:
Under this assumption, the probability of purchasing an individual item can have an extensive range; one user’s likelihood of buying a thing might be double while another user’s likelihood is far less likely. The random method starts to break down here, and detailed personalization is beneficial. In this situation, a random selection would have an expected purchase rate of 3%, regardless of the player. However, if we could perfectly predict each player’s best item and put that in their rotating store for the week, the average chance of purchase across the four users would be 5.64%. Those numbers translate into an 88% increase, but of course, it relies on a toy simulation that is not fully representative of a real-world scenario. If one was to build out real-world predictions for each player, then these simulations could accurately assess the overall value.
Scenario Four: Only some people want specific items, and the overall demand for an item is unequal across the entire population.
We went through all of the previous scenarios to help build-up to this one; here, the underlying distributions, while still assumptions, closely match the outcomes seen in the real-world. This scenario is an extension of scenario two, but now each user has their own distribution, with the items’ overall popularity influencing that distribution.
Notice that while different users have their preferences, the overall popularity of items roughly holds. Recall under scenario two that the selection of items was weighted by their demand. To give a better short-term comparison, assume that we chose to show each user the most popular item; their probability of purchase under that scenario would have been ~20% each. However, what if the underlying demand distribution was best represented by scenario four, but we chose that same tactic anyway? For user one, we would have gotten a lift in the sale probability (20% → 29%), but for the other users, their likelihoods decreased (20% → 16%, 6.6%, 4.2%); altogether, 55.8% when we were expecting ~80%. If we had personalized instead, we would have shown the first user the same item and all other users item “C”. The probabilities in that scenario are 29%, 22.8%, 16.9%, and 25.6% for a total of 94.3%, a 69% increase.
The gains mentioned in the last two scenarios are too good to be true for the long-term, but they hint that improving the quality of recommendations is a worthwhile endeavor. A more in-depth discussion needs to be had around many caveats to determine the actual value of personalization. In part two, we will address these topics:
- Catalog Scale — how does the scale affect personalization value?
- Interface Structure — will a store’s structure be beneficial?
- Long-Term Scenarios — are there diminishing returns through time?
We will also look into detailed simulations across a more extensive set of users to estimate the real advantage that personalization brings. See you in the next rotation!
Part two of this blog has been finished and is located HERE.
This post was originally published by Riley Howsden at Towards Data Science