推荐系统的演化 (Evolution of online recommendation system)

类似于amazon和豆瓣这样的仓库型网站，推荐系统是提高用户黏度、扩大盈利规模的必备手段。

最初的推荐系统其实是类似于黄页的分类系统，比如 bababian和yupoo都属于“网络相册”这一类，那么，介绍bababian的时候就可以推荐yupoo，这个系统最耗费成本，而且效率相当低。而且图片等非文本信息很难被分类。

然后进入了tag横行的时代 ，通过UGC模式产生的大量tag，让仓库里的大量条目建立关联，豆瓣和采麦的推荐系统就是这么做的。tag可以很好的胜任图片信息的推荐，但是对音乐、艺术等抽象信息就无能为力了。

音乐分享鼻祖Pandora只好不辞辛劳的采用了一种特殊的推荐机制：分析每一首曲目并赋予其独特的音乐DNA标记，通过标记的匹配来进行推荐，在pandora还未封闭之前，它一直是我听歌的首选，推荐的准确度相当高，以至于封闭之后让我无论如何也找不到替代品。

现在又有一种新的推荐模式出现：基于相似族群的推荐。最近去了趟toluu，是做博客推荐的，输入你喜欢的博客网址，系统会自动分析与你爱好相似的用户（把你们归类为同一人群），并把他们喜欢的博客推荐给你（同一人群应该有同样的爱好）。我没仔细去体验toluu的服务，也不好说准确度有多高，但是这个模式的确非常诱人，其实这也是最初采麦想采用的推荐模式，但现在还处于用户基数不够的阶段，尚不是合适机会。

相似族群的推荐更像是朋友之间的悄悄话。口味相似的人，互相从对方身上获取自己未接触过的信息，最后把每个人的小圈圈扩大，所有人都融合到一个大圈圈里，实现1+1=4的平方级增长，符合资源优化配置的趋势。

To Mika:

For websites like amazon and douban, recommendation system is the key element for a higher user stickyness and larger sales.

The original recommendation mechanism is an online yellow page, with different catalogs and each item belongs to one or several catalogs. The item under the same catalog will be recommended automatically. The weak point is yellow page can not define non-text information like pictures.

Then came the TAG era from about 5 years ago. Tag is the most representative feature of web 2.0. With enormous UGC (user generated content) emerging in one night, user added tags become the most efficient way for keyword match and recommendation.

But tag also has fatal weakness: you can never add enough accurate tags to virtual information which involves personal taste, such as books, music, movie. I love adventure movies, but Caribbean Pirates is disgusting for me, while everyone tag it with “adventure”. How can we deal with that?

Pandora invented a complicated system to analyze every song’s DNA (inner pattern), and generates a very high accuracy when recommend new songs to users. Frankly, I damn love pandora. It’s the best product I’ve ever found online. Don’t mention Last.fm to me, it sucks. In China there are exact copiers: YOBO and songtaste. Since pandora shut the door to non-US users, I can only turn to YOBO, but saddly, the recommendation system is like shit.

The final solution for recommendation should be based on human TASTE. Similar people have similar taste. If our system can determine that Tom and Jack is 80% similar people (in some field), then we can reach a 80% accuracy when we recommend TOM’s favorites to Jack, and vice versa. I’m not sure if toluu (a blog recommendation system) is based on this mechanism, but I will definitely spend my whole life on it.

This entry was posted on April 03rd, 2008 and is filed under , recommendation, tech, 推荐机制, 推荐系统. You can follow any responses to this entry through the RSS 2.0 feed. You can leave a response or Trackback from your own site.