I don’t know what exactly Pandora was a few years ago (perhaps one of the automated music recommendation portals), but now they are positioning themselves as the Internet radio. From my experience of using it for the last week, it works amazingly well — indeed, I like it better than Spotify. To start with, I choose a base station, which can be some artist or genre of music, for example Norah Jones. From then on it works like a radio/music player playing a cassette or CD but the stream of music is never-ending. The user interface could not be characteristically simpler (not a plus point though) but it has the added feature of voting up or down a song. The votes are the feedback that Pandora uses to generate the next songs to play for me. So far they play Norah Jones, Michael Bubble, Louis Armstrong, Adele, Billie Holiday, and many others for me. Though impressive, this shouldn’t be so surprising if we know that Pandora has this huge database of information about millions of music tracks and artists. This database contains the rhythmic styles, acoustic features, genres, and around 400 other features, most of which are manually annotated by professional artists*. Given the data, it is not hard to find similar songs and artists like Norah Jones. Nevertheless, my use case is straightforward as I don’t change my station across different moods or times of the day whereas many users do. From the presentation by one of the Pandora engineers that I attended, it seems the capability to switch the theme of a station depending on the situations (context) is still well beyond the current Pandora system.
*This is a very different approach to the technologies used by other music service providers or its competitor such as the use of machine learning and signal processing to automatically analyze songs and artists. Pandora’s approach is very costly especially in the long-run, but the data they have is invaluable and is perhaps more reliable at least in the current state.
After writing the rebuttal for the reviews of our paper comes a hard-learned lesson in academic writing: show the evidence. In our paper, after some explanation and reasoning, we made a general statement saying ‘A is X’ but we had never shown that A is indeed X theoretically or empirically. As such, we were not able to strongly defend this argument and so it was inexplicable to respond to the reviewer who made the comment. At some point came the thought that it would have been easier to say ‘we are sorry the statement was sloppy’ 🙂 Although the two other reviewers said they enjoyed reading the paper, this does not hide the fact that that sentence lacked scientific validity. Hunches and intuition are acceptable and in fact used outside of the academia all the times; indeed, it is rather tedious if not boring to provide figures following every claim that one makes. However, the lesson remains that to really convince people: show plenty of evidence, emphatically.
It appears to me that, every now and then, us graduate students like to use complaints as a defensive mechanism to get through the days. Why do we love complaining so dear? Well, first and foremost, our wages are minimum compared to our peers of equivalent or even lesser qualifications. But they make sure that our taxes (publications) are collected in due time, often once or more every year. Sounds like peasants in feudalism? Why then does the entire industry of academia seem to function and move forward consistently? Because we’re very low maintenance — we rejoice just by looking at a number (our error measurement) that is smaller, no matter how marginal, than that of our our neighbors (previous authors). That is what I call ‘a game of numbers’ 😀
Recently I’ve been looking at recommender systems , which is one of the most prevalent application of machine learning in the industry. A recommender system looks at your previous consumption behaviors (e.g. movies you watched, articles you read, products you purchased, etc) and suggests items that are (hopefully) relevant to your interests. You actually interact with recommender systems every day when using services like YouTube, Amazon, and Netflix; they often come with statements saying “you may also like…” or “recommended for you…”. Other online news aggregators/publishers also have similar feature under new articles although I’m not sure if their recommendations are automatically generated or hand-picked by their staffs/editors (o_o)). One interesting observation from a dataset I’m experimenting with is that ’emotions’ seems to play an important role in shaping people’s opinion toward a movie. Other factors, including the overhyped social influence, are not as indicative of user reception as the emotions or feelings evoked by the movie. My retrospective thoughts and observations agree with this finding. Unless we can build captivating products that arouse and inspire its users, like some of the Apple tech devices or the wide range of free Google services, they will soon be forgotten..