Summary:
Previous recommendations using explicit feedback from users, we use implicit feedback;
In this paper, the method optimization process and the data quantity are linear, which can be well fused with the existing system.
Let's talk about an explanation of the method.
1. Introduction
1) e-commerce to recommend a large
2) The traditional recommendation method:
First, content-based, shortcomings: data is not good to collect
Second, collaborative filtering: Advantages: Domain-independent, can capture information that is difficult to obtain based on content, high precision
Cons: Cold start, Content-based no this problem
3) The data type of the recommended system:
I. Dominant data: Scoring, approval/objection (Thumbs-up/down), less data available
Second, recessive data: Purchase, browse, search, etc., data many
4) Characteristics of implicit feedback data:
First, no negative feedback
Second, the data have noise: may not be out of their own needs to act (gift); buy not like; Watch popular videos
Dominant numerical characteristic reaction preference, implicit characteristic reaction confidence
Four, the measurement standard is not OK: explicit feedback with MSE, recessive do not know
2. Basic settings
U,v represents the user; I,j indicates that Item;r (U,i) represents a behavior or rating; no behavior is recorded as 0 points.
3. Previous work
1) Neighborhood Model: First there is user-base, after the item-base;item-base effect is better, because the item responds to user preferences, and similar user estimates are not allowed
ITEM-CF is used more in explicit feedback and can be optimized with user and item bias, but implicit feedback uses data such as frequency, which is not appropriate;
The downside of ITEM-CF is the inability to differentiate user preferences
2) LFM:PLSA,NN,LDA,SVD
SVD for dominant data; better than CF
The core of this paper is to use SVD to recessive data
4. Modeling
1) Model: SVD with weights, weight term representation confidence
2) The data magnitude is too large, SGD is difficult to calculate, using ALS calculation; ALS is good at a large number of miss values, not in dense case
3) Reduce time complexity to linear time based on ALS optimization (deduction not understood)
4) Modeling variants: Changes in P (u,i) and C (u,i)
5. Model explanation
As with the model deduction, I didn't see it.
6.
"Paper reading-ctr" <<collaborative Filtering for implicit Feedback datasets>> reading