这个slide介绍了不少推荐的算法,但我所收获的确是整个推荐环节的一些通用的过程及数据整理:
Music Recommender Systems
- Data Collection(User rating/collection/listen log/view log)
- Data Cleaning(Missing/Wrong/Noise/Duplicate Data)
- Data Preprocessing
- Data Mining
- Tracking & Optimization
最近有个推荐的实战项目,但在重新看了这个豆瓣的推荐分享后发现我们没考虑到的问题还相当多。对于盛大来说,要做好这块不仅仅是找几个人这么简单,而是需要从业务、用户行为、技术实现和实际运营等多维度来思考。
——————————–
什么样的产品适合推荐(产品特性、用户行为层面):
- 具有媒体性的产品(Media Product)
- 口味(taste)很重要
- 单位成本不重要
- 瀑布效应(Information Cascade),详见附一
什么样的产品适合推荐(业务层面):
- 条目增长相对稳定
- 能够获得快速反馈
- 稀疏性、多样性和时效性的平衡
推荐系统的可扩展性:
推荐系统面临的挑战:
- 推荐是一项技术还是一种产品/功能?
- 推荐能否有独立的产品形态?
附一:information cascade
An
information (or
informational)
cascade occurs when people observe the actions of others and then make the same choice that the others have made, independently of their own private information signals. Because it is usually sensible to do what other people are doing, the phenomenon is assumed to be the result of
rational choice. Nevertheless, information cascades can sometimes lead to arbitrary or even erroneous decisions.