Loading [MathJax]/jax/element/mml/optable/GreekAndCoptic.js

Friday, October 24, 2014

Week 4 of MofMD. Recommender Systems

Two major classes
  • Content based
  • Collaborative filtering
  • Latent factor based
Discussion on "Long Tail" effect. Some products (less popular) may mot exists in physical stores because it's not reasonable to stock them.

Formal model:
C - set of Customers
S - set of Items

Utility function u: C x S -> R (set of Ratings)
R - totally ordered
Key problems:
  • Gathering "known" ratings for matrix
    • Explicit (ask people, don't scale)
    • Implicit (learn from actions, hard to learn low ratings)
  • Extrapolate unknown ratings from the known ones (mostly interested in high unknown ratings)
    • Sparsity
    • Cold start: new items have no ratings, new users have no history
  • Evaluating extrapolation methods

Content-based recommendations

Main idea to recommend customer x items similar to  previous items highly recommended by customer x
Item profiles - set of "important" features
Text features - set of "important" words
Extract using heuristics like TF-IDF (Term frequency * Inverse Doc Frequency)
fij - frequency of ter (feature) i in doc (item) j
TFij=fijmaxkfij
nj - number of docs that mention term i
N- total number of docs
IDFi=log(Nni) TF-IDF score: wij=TFij×IDFi
Doc profile - set of words with highest TF-IDF scores (together with scores)
I.e. word is frequent in (small) set of document - it's important
Steps include building user profile as sort of aggregated features of items and estimate U(x,i)=cos(θ)=xi|x||i| where x is User profile, i Item profile. Aggregation can be mean or normalized mean of features of items.
Pros:
  • No need for data on other users
  • Able to recommend to users with unique tastes
  • Able to recommend new & unpopular items
  • Explanation for recommended items
Cons:
  • Hard to find appropriate features
  • Overspecialization - never recommend items outside user's profile, unable to exploit quality judgments of other users
  • Cold start problem - how to build user profile

Collaborative filtering

Main idea to find N other users whose rating is similar to x's rating and estimate x's rating based on ratings of users in N
Similarity between users
  • Jaccard similarity sim(A,B)=|rArB||rArB| doesn't work well.
  • sim(A,B)=cos(rA,rB)
  • Centered cosine is better sim(A,B)=cos(rA,rB) - center user's rating around average value.( Make 0 average user's rating)
Rating Predictions
Let rx be the vector of user x's ratings
Let N be the set of k users most similar to x who also rated item i
Prediction for user x and item i:
Option 1: rxi=1kyNryi - doesn't take in account "degree of similarity" of users
Option 2: rxi=1kyNsxyryi/yNsxy where sxy=sim(x,y)

Another version is Item-Item filtering (User-User described right above)
  • For item i find other similar items
  • Estimate rating for item i based on ratings for similar items
  • Can use same similarity metrics and prediction functions as in user-user model
rxi=jN(i;x)sijrxjjN(i;x)sij where sij similarity of items i and j rxj rating of user x on item N(i;x) set items rated by x similar to i

In practice item-item outperforms user-user in many use cases. - Items are "simpler" than users. Items similarity is more meaningful than user similarity

Common practice implementing collaborative filtering - some combination of recommendation system plus filtering
rxi=bxi+jN(i;x)sij(rxj-bxj)jN(i;x)sij where bxi=μ+bx+bi, μ - overall mean movie rating (example of movie database), bx - rating deviation of user  x = (avg rating of user x) - μ, bi - rating deviation of movie i

Evaluating Recommender system
Predict ratings on test set and evaluate RMSE
(x,i)T(rxi-rxi)2N

N=|T|, rxI the predicted rating, rxi the actual rating.

Problem with RMSE - Prediction Diversity (all predictions too similar to each other), Prediction Context (user might operate in different context => want different results), Order of Prediction
We care only to predict high ratings - RMSE might penalize method that does well for high ratings and badly with others, Alternative : precision at top k i.e. percentage of prediction in user's top k ratings.

Latent Factor Models

Global effects => like Baseline Estimation
Local neighborhood (CF/NN) => Final estimate
Collaborative filtering - derive unknown rating from k-nearest neighbors like
ˆrxi=jN(i;x)SijrxjjN(i;x)Sij
sij similarity of items i and j
ruj rating of user x on item j
N(i;x) set of items similar to item i that were rated by x


Combine with global effects
ˆrxi=bxi+ jN(i;x)Sij(rxj-bxj)jN(i;x)Sij
bxi=μ+bx+bi
μ - overall mean rating
bx - rating deviation of user  x
bi - rating deviation of item i

Problems:
1) Similarity are arbitrary
2) Pairwise similarities neglect interdependencies among users
3) Taking weighted average can be restricting


SVD
A - input data matrix
U - left singular vec
V - right singular vec
Σ - singular values

A~~U Sigma V^T => "SVD" R~~Q*P^T, A=R, Q=U, P^T=Sigma V^T
R - utility matrix approximated by Q and P matrices of lower dimensions

SVS minimizes reconstruction error
SSE = {::}_{U,V,Sigma}^"min" sum_{i,j in A}(A_{ij}-[U Sigma V^T]_{ij})^2
SSE - Sum of Squared Errors
SSE and MSE are monotonically related RMSE = 1/{const} sqrt(SSE) => also minimizing RMSE
But R has missing entries!

So modify SVD to find P,Q such as
{::}_{P,Q}^"min" sum_{(i,x in R)}(r_{"xi"}-q_i * p_x^T)^2 (use elements with defined scores)
also
  • we don't require P,Q to be orthogonal/unit length
  • P,Q map users/movies to latent space
We want a large k, but if k>2 SSE on training data will increase, because model will adapt noise. So we may use regularization like
{::}_{P,Q}^"min" sum_{(i,x in R)}(r_{"xi"}-q_i * p_x^T)^2+ lambda[ sum_x||p_x||^2+sum_i||q_i||^2] 
and use gradient descent

Initialize P and Q (using SVD - pretend missed ratings are 0)
Do gradient descent
  • P larr P - eta * gradP
  • Q larr Q - eta * gradQ
where gradQ = [gradq_{ik}] and gradq_{ik}=sum_"xi"-2(r_"xi"-q_ip_x^T)p_"xk"+2lambdaq_"ik"and similary for gradP

Modeling biases and interactions
hat r_"xi" = mu + b_x + b_i + q_i*p_x^T
where
'mu' - overall mean rating
b_x - bias for user x
b_i - bias for movie i
q_i*p_x^T - user movie interaction

and we can fit a new model
{::}_{P,Q}^"min" sum_{(i,x in R)}(r_{"xi"}- (mu + b_x + b_i + q_i*p_x^T))^2 + lambda[ sum_x||p_x||^2+sum_i||q_i||^2+sum_x||b_x||^2+sum_i||b_i||^2] 
where b_u and b_i become additional parameters for the estimation

Dimensionality Reduction

Matrix rank - number of lineary independent columns/rows. The amount of coordinates can be reduced to rank.

Why reduce?
  • Discover hidden correlation/topics
  • Remove redundant and noisy features
  • Interpretation and visualization
  • Easier storage and processing of data

SVD definition

A_[m x n] = U_[m x r] Sigma_[r x r] (V_[n x r])^T
A: input matrix (say, m documents, n terms)
U: Left singular martix m x n matrix (say, m documents, r concepts)
Sigma Singular values, r x r diagonal matrix, values on diagonal sorted in decreasing order ("strength of each concept")
V: Right singular vector n x r matrix ( n terms, r concepts)

A~~USigmaV^T = sum_isigma_iu_inu_i^T

Decomposition is always possible, matrices are unique, U and  V columns orthonormal ( U^TU=I, V^TV=I, columns are orthogonal unit vectors), Sigma diagonal, singular values are positive and sorted in decreasing order

We can remove minor (tail nu).
Frobenius norm is minimal ||M||_F=sqrt(Sigma_"ij"M_"ij^2")

SVD gives best low rank approximation
Rule-of-a thumb - Keep 80-90% of 'energy' (=sumsigma_i^2)
SVD complexity O(nm^2) or O(n^2m)whichever is less but less work if we just nee a singular values, or k first singular vectors, or the matrix is sparse


Query SVD
q_"concept"= q V - transform query q to "concept" space and it will give a matrix of "strength" of each concept

Relation to Eigen-decomposition
A=XLambdaX^T
  • A is symmetric
  • U,V,X are orthonormal (U^TU=I)
  • Lambda, Sigma are diagonal
or diagonal matrix lambda_i=sigma_i^2

So
+ Optimal low rank approximation
- Interpretability problem
- Lack of sparsity (singular vectors are dense)

CUR decomposition
Goal - express A as product of C,U,R matrices and keep ||A - C*U*R||_F small

"Constraints" on U - contains set of columns from A, similarly on R - set of rows from A

In practice we can choose about 4K rows/cols and do almost as good as SVD does.

See book for details.

Pros:
Easy interpretation
Sparcity
Cons:
Rows and columns with large norm may be choosen few times (probabilistic method)




P.S. ASCIIMathML

No comments:

Post a Comment