Website recommendation System Analysis: Recommendation system will never recommend anything to you

Intermediary transaction SEO diagnosis Taobao guest Cloud host technology Hall

Do not think you can Taobao "secretly" look at erotic items, you do on the site all do not escape the system's eyes. To guess your mind, the recommended algorithm needs to gather as much data as you can to determine what you really like. Of course, this is to win your trust and let online retailers make more money.

Source: yoochoose.com

(/joseph A. Konstan & John Riedl) The recommended algorithm is how to "guess what you Like", and now you have a basic idea of how online retailers look at you every time you surf the internet and try to match your preferences to those of others.

The recommendation system has two other features, and it has a significant impact on the results you see at the end of the recommendation: first, before figuring out how similar you are to other shoppers, the referral system must first figure out what you really like; second, the recommendation system runs in accordance with a set of business rules to ensure that the recommended results are useful to you, Also make business profitable.

How does the recommended algorithm win your trust and make the business rich?

Collect your Internet Data

For example, looking at Amazon's Art store, the last time we went to see it, there were 900多万册 prints and posters on sale. Amazon's Art store has several ways to assess your preferences. It will allow you to grade an artwork at 1 to 5 stars, it will also record which pictures you have clicked to enlarge the view, which paintings you have looked over and over many times, what you put into the wish list, and what you finally actually buy. It also tracks which paintings are displayed on every page you browse. Online retailers will use the path you've made on their website (the pages you've browsed and the links you clicked on) to recommend the items you're connected to. In addition, it combines your purchase record with the scoring information to create a file for your long-term buying preferences.

Companies like Amazon collect a lot of such data about customers. When you log in, almost every action you make on its site will be written down and reserved for future use. Thanks to the browser cookie, even anonymous shoppers can maintain online records, which will eventually be linked to customer profiles when anonymous shoppers create accounts or log on. This type of explosive data collection is not unique to online businesses, and Wal-Mart is renowned for its in-depth mining of cash receipts. However, online stores are in a better position to view and record, not just what consumers buy, but also what you have considered, browsed, and decided not to buy. In most parts of the world, all such activities are monitored and recorded by the person, and only in Europe is the data privacy law limiting the operation to some extent.

Of course, no matter what the law is, customers find their data abused by people, and they will have a strong resistance. As early as September 2000, Amazon suffered a painful: some customers found that they received higher prices, because the site will identify them as old customers, rather than anonymous access to or from a comparison site to transfer incoming customers. Amazon claims this is just a random price test, and it's a coincidence that the results are linked to the identity of old customers. That said, it stopped the operation.

Run under business rules

The business rules added to the recommended algorithm are designed to prevent the algorithm from giving stupid recommendations and to help online retailers maximize their turnover without losing your trust. At the very least, recommender systems should avoid what people say about the supermarket paradox (supermarket Paradox). For example, almost everyone who goes to the supermarket likes to eat bananas and often buys some. So shouldn't the recommendation system recommend bananas to every customer? The answer is no--it doesn't help customers, it doesn't increase the sales of bananas. So the smart supermarket recommendation system will always include a rule that explicitly excludes bananas from recommended results.

This example may sound like nothing, but in one of our early projects, our referral system has recommended the Beatles ' white Album album to almost every person visiting our site. Statistically speaking, this is a great recommendation: Customers have not previously purchased this album from this E-commerce site, and most customers have high ratings for the White Album. Still, the recommendation is ineffective-anyone interested in the White Album already has one.

Of course, most of the recommended rules are more subtle. For example, when John was in Netflix's action video in September, The Avengers was not seen in the results, because the blockbuster didn't have a rental version at the time, and the recommendation wouldn't make Netflix rich. So John was led to "Iron Man 2" (Iron Man 2), which is already available in streaming media.

Other rules include banning the recommendation of goods sold at a loss to attract customers (loss leader) and, conversely, encouraging the recommendation of unsalable. During the operation of net perceptions, we worked with a customer who used the recommendation system to identify potential customers of inventory backlog and achieved considerable success.

Win your trust.

However, this kind of thing will soon become tricky. A recommended algorithm for selling highly profitable goods will not win customers ' trust. It's like going to a restaurant where the waiter is trying to recommend a fish to you. Is this fish really what he thinks is the best? Or is the chef urging the bottom man to sell it before the fish goes bad?

In order to build trust, more complex recommendation algorithms will try to maintain a certain degree of transparency, giving customers a general idea of why the system recommends the product to themselves, and can change their profile when they do not like the recommended results. For example, you can delete the shopping record you bought on Amazon; after all, those things don't reflect your personal preferences. You can also see why the system recommends certain products to you. When Amazon picked Jonathan Franzen's novel Freedom, John clicked on the link on the tag, "Why do you recommend it to me?" A brief explanation was immediately displayed, and it turned out that the book he had placed in his wish list triggered the recommendation. However, since he had not read the books on the list of wishes, John would not have been in charge of the recommendation of liberty. An explanation like this will let the user know if the recommended results are useful.

However, perfecting the personal data and interpreting the recommendation results are often insufficient to ensure that the system is not wrong. Recently, Amazon bombed Joe with a promotional email from a high-definition large-screen television (HDTV)--3 a week and one months in a row. In addition to sending too many emails to Joe, the retailer did not realise that Joe had bought a television set with his wife's account. In addition, these emails do not provide an obvious way for Joe to say "Thank you, but I am not interested". In the end, Joe canceled some of his email subscriptions at Amazon; he didn't care about receiving all sorts of information, and he had more time to really watch his TV.

How much does the recommendation algorithm work?

alumni.berkeley.edu

Well, how much does the referral algorithm really do? They have certainly been increasing online sales, according to Jack Aaronson, an analyst at Aaronson Group, who estimates that investment in recommended algorithms can be 10% because of the growth in sales generated by recommended algorithms. -30% of the proceeds. And they're just getting started. Now, for those of us who study recommender systems, the biggest challenge is figuring out how to judge new methods and algorithms best. This is not as simple as benchmarking the microprocessor because different recommender systems have very different goals.

The simplest way to evaluate an algorithm is to see how much difference there is between its predictions and the actual evaluation of the user. For example, if John gave the romantic novel Twilight (Twilight) A star, Amazon might have noticed that the algorithm predicted by other similar users that John would give two stars, and that there would be a star deviation. However, the seller is more concerned about the algorithm in the user evaluation of high goods on the wrong, because the more praise of the goods are more likely to buy customers; John will not buy Twilight city anyway. So, taking this evaluation into consideration is not helpful in understanding how much the recommendation algorithm works.

Another common method is to see how high the number of matches between the proposed results and what the customer actually buys. However, this method may also be misleading, because this analysis will be the user himself to find the product incorrectly counted on the head of the recommended algorithm, and the user himself to find things is the most should not be recommended! In view of the shortcomings of these methods, researchers have been studying new evaluation indicators, not just precision, It also focuses on other attributes like discovering unexpected surprises and diversity.

The discovery of unexpected surprises (serendipity) will weighted unusual recommendations, especially those that are valuable to a particular user but are useless to other users of the same category. Adjusting the algorithm to find unexpected surprises notes that the White Album appears to be a good recommendation for almost everyone, and therefore instead looks for a less common option-perhaps Joan Armatrading's love and affection. This less-popular recommendation is unlikely to hit the target, but once it encounters it will bring a much bigger surprise to the user.

Looking at the diversity of recommended results is also a good illustration of the problem. For example, a user who loves to watch Dick Francis's mystery novels may still be disappointed when he sees all of Dick Francis's work in the recommendation form. A truly diverse list of recommended forms will include different authors and different types of books, as well as movies, games and other products.

Recommendation system research needs to break through a variety of obstacles, much more than fine-tuning on existing systems. The researchers are now considering how the recommended algorithm should help users discover what they don't know about a site's content collection. For example, send people who buy books to the apparel department in the Amazon, rather than giving some safe, customer-accepted recommendations. Outside the retail world, recommended algorithms can help people get in touch with new ideas; even if we disagree with some of them, the overall effect will probably be positive because it will help reduce the social Balkan (Balkanization, i.e., fragmentation). It remains to be seen whether the recommended algorithm can do this, or whether it is annoying or distrustful.

But one thing is clear: the referral system will only get better, gather more and more data about you, and show it in other, unexpected places. If you like this article, Amazon will be happy to recommend all the other books you might like about the recommendation system.

Joseph A. Konstan and John Riedl are both computer science professors at the University of Minnesota. Konstan, an IEEE senior member, and Riedl of IEEE have been involved in the creation of the Movielens recommendation system. "Guess what you Like" is how you guessed your mind? Is the first half of this article.

Compiled from: IEEE Technology overview Deconstructing recommender BAE

Nevin Pictures: yoochoose.com;alumni.berkeley.edu

Article title: netregistry.com.au