Pandora Internet Radio is a music streaming service that is doing a great job in terms of music recommendation. Their recommendation system is based on the Music Genome Project, which began in 2000 and is one of the most thorough analysis of popular music ever undertaken. It took 30 experts in music theory five years to complete. The Music Genome Project is based on an intricate analysis by actual humans (about 20 to 30 minutes per four-minute song) of the music of 10,000 artists from the past 100 years. This means that the analysis of new music continues every day. The difference with recommendation systems of other players in the market is that Pandora does not make use of the popular method of “people who like this also like this” or other users ratings.

It starts with entering a song or artist that you want to listen to and Pandora will generate a continuous playlist. While a song is playing, it is possible to provide positive or negative feedback for songs chosen by the service. Based on your feedback, Pandora is analyzing the musical structures present in the songs you like and it will add other songs that possess similar musical traits to your playlist.

But, after years of customizing playlists to individual listeners, the company has started data-mining users’ musical tastes for clues about the kinds of ads most likely to engage them. Eric Bieschke, Pandora’s chief scientist says: “It’s becoming quite apparent to us that the world of playing the perfect music to people and the world of playing perfect advertising to them are strikingly similar”.

A lot of companies make use of behavioral targeting. Behavioral targeting comprises a range of technologies and techniques used by online website publishers and advertisers aimed at increasing the effectiveness of advertising by using users’ web-browsing behavior and personal information. Amazon is one of the companies that is really differentiating themselves by using their huge amount of data to make a deeper understanding about individuals and try to influence their behavior.

Pandora adds another layer on behavioral targeting: music. They seek correlations between the listening habits of their users and the kinds of ads they might be most receptive to. The company beliefs that people’s music, movie or book choices may reveal much more than commercial likes and dislikes. Certain product or cultural preferences can give glimpses into consumers’ political beliefs, religious faith, sexual orientation or other intimate issues.

In time of elections, Pandora uses their political ad-targeting system. They are able to deconstruct users’ song preferences to predict their political party of choice. This system has already been used in former presidential and congressional campaigns. During these campaigns, Pandora users tuning into country music or Christian bands might see ads for republican candidates and others listening to hip-hop tunes or classical music might see ads for democrats.

I think providing ads based on music preferences is a new interesting way to target the right audience. Of course, the system will not be able to predict everyone’s political affiliations right, but I can imagine that the algorithm has improved a lot since the last presidential elections. So in 2016, Hillary Clinton might be winning the elections thanks to the ads on Pandora.

Deschene, L. (2008). What Is Behavioral Targeting?. Available: http://i.bnet.com/pdf/199800-What_Is_Behavioral_Targeting_.pdf. Last accessed 03-05-2015.

Layton, J. (2015). How Pandora Radio Works. Available: http://computer.howstuffworks.com/internet/basics/pandora1.htm. Last accessed 03-05-2015.

Singer, N. (2014). Listen to Pandora, and It Listens Back. Available: http://www.nytimes.com/2014/01/05/technology/pandora-mines-users-data-to-better-target-ads.html?_r=0. Last accessed 03-05-2015.

The availability of real-time data on customer characteristics has encouraged companies to personalize operational decisions for each arriving customer (Golrezaei, et al., 2014). For instance, Orbitz.com has found that Mac users spend on average $20 to $30 more per night on a hotel than Windows users . Therefore the online travel agency can show different and more expensive hotels to Mac users (Mattioli, 2012).

The key question in this paper is: Given the complexity of coordinating real-time, front-end, customer-facing decisions with the back-end supply chain constraints, what policies should companies use to take advantage of real-time data?

Once a customer arrives on the website, his or her customer type will be revealed. This can be based on computer type , zip code, gender or on any available information that is relevant for the company. Based on the customer’s type and the remaining inventory, the firm offers an assortment. An example of a website that shows different assortments based on customer type is Zalando. When I visited the website last week, I was automatically directed to the ladies department without logging in. Associated with each customer type is the probability of purchasing each product under each assortment. This means that the company calculates, for example, the probability that a 40 year old woman who lives in a high-income neighbourhood purchases a certain dress which is showed between other dresses under Assortment A but also the probability that she purchases the same dress that is showed between different dresses under Assortment B. The authors of this paper want to design an revenue-maximizing algorithm that determines the assortment to offer to each arriving customer, taking into account the customer type and current inventories.

In the paper, the authors propose a couple inventory-balancing algorithms. An inventory-balancing algorithm makes use of a discount factor that depends on the fraction of the product’s remaining inventory. This means that when the inventory of a product drops, the discount becomes higher which results in a lower discounted revenue. Upon the arrival of each customer, based on the customer’s type, the algorithm offers the assortment that maximizes the expected discounted revenue. By adjusting the revenue of each product according to its remaining inventory, the algorithms hedge against the uncertainty in the types of future customers by reducing the rate at which products with low inventory are offered. For example, a pair of red leather boots (low inventory) that normally would be showed at page 1 to a 40 year old woman who lives in a high-income neighbourhood (in case of enough inventory) might now be showed on page 4, because it is likely that she is willing to buy another, more expensive, pair of shoes. When a 35 year old woman who lives in a low-income neighbourhood arrives at the website 5 minutes later, these red leather boots will actually be showed at page 1 because this maximizes the revenue.

These inventory-balancing algorithms work very well in cases with significant uncertainty in the market size, yielding 5%-11% more revenues than reoptimization methods. Reoptimization methods work extremely well and yield nearly optimal revenue because they can effectively ration the inventory to all customers. This means that inventory-balancing algorithms are more suitable for situations with a lot of uncertainty. Also, the inventory- balancing algorithms have a strong performance under both nonstationary and stationary demand processes. This implies that, even when there are sudden shocks in the customers’ arrival patterns, for example in case of seasonality, the algorithm maintains a strong performance guarantee. Another advantage is that this inventory-balancing algorithms are simple and flexible which makes it possible to combine them with reoptimization methods.

Golrezaei, N., Nazerzadeh, H. & Rusmevichientong, P. (2014) Real-Time Optimization of Personalized Assortments. Management Science. 60 (6), 1532-1551.

Mattioli, D. (2012) On Orbitz, Mac Users Steered to Pricier Hotels. Available: http://www.wsj.com/articles/SB10001424052702304458604577488822667325882. Last accessed 23-4-2015.