推荐基础视频笔记

1推荐问题
information overload 信息过载
everything is personalized 任何东西都是个性化
recommendation as data mining 推荐是数据挖掘
serendipity 新颖性
explaination 可解释性
awareness & diversity 多样性
evolution of the recommender problem 推荐问题的演化
rating -> ranking -> page optimization -> context-aware
2传统方法
collaborative filtering 协同过滤
personalized vs non-personalized CF 个性化 vs 非个性化
based only on similar users vs average of all the users
reasonable -> great 合理向理想演进

RMSE 指标
random 2
average >1.0
cinematch < 1.0  --- user-based
prize < 0.8
perfect 0

user-based CF  - UBCF
item-based CF
similarity cosine-based similarity
memory-based
sparsity  scalability 数据稀疏性和伸缩性差

model-based CF
rules-based approaches (association rules) -- priori algorithm
classification
clustering

两个算法
SVD 奇异值分解   0.8914  gradient descent svd++
RBM restricted boltzman machines 受限玻尔兹曼机    0.8990
linear blend prize RMSE: 0.88

聚类方法
LSH locality-sensitive hashing
k-means
affinity propagation
spectral clusting
non-parametric bayesian clustering

limitations of collaborative filtering 
cold start 
polularity bias

3 beyond traditional methods 现代方法
learing to rank 排序
metrics 排序指标
1normalized discounted cumulative gain
2mean reciprocal rank(MRR)
3fraction of concordant pairs (FCP)

approaches 排序方法
pointwise
pairwise
listwise

similarity 相似性计算
graph-based similarities:simRank

deep learning

multi-armed bandits

链接

http://videolectures.net/kdd2014_amatriain_mobasher_recommender_problem/