Personalized Ranking Recommendation via Integrating Multiple Feedbacks 3
and they can also have other interactions with businesses, such as “checking in”
and “viewing”. Obviously, our problem setting is a general framework to utilize
feedback information, and existing problems are special cases of our problem set-
ting. In addition, from recommendation perspective, the predicted ranking over
an item is much more meaningful than the predicted rating. Thus in this work,
we focus on developing a personalized ranking model that integrates multiple
feedbacks. Although many methods have been proposed to utilize the feedbacks,
these models are usually designed for special problem settings, and they cannot
be directly applied in multiple-feedback setting.
However, integration of multiple feedbacks faces two challenges. (1) Design a
unified ranking model integrating multiple feedbacks. In order to make the best
use of these feedback information, we need to design an effective mechanism
to handle relations between explicit and implicit feedbacks as well as relations
among implicit feedbacks. (2) Generate training samples. As a ranking method,
we need to generate preference pairs or lists for training. However, there are
multiple types of feedbacks. What kind of feedbacks could we utilize for better
preference pair or sequence?
The major contributions of our paper are summarized as follows: (1) We first
try to solve the personalized ranking recommendation problem by integrating
multiple feedbacks. The problem widely exists in real recommender system, and
it is a general problem setting to encompass existing works. (2) We propose a
Bayesian Personalized Ranking (BPR) based model MFPR to integrate multiple
feedbacks. Moreover, as there are no readily available training data for this
problem, an effective algorithm is designed to generate the training data that
is more consistent with multiple feedbacks for the MFPR model. (3) We crawl
comprehensive Douban Book and Dianping datasets
1
including ratings and
multiple types of implicit feedbacks.
2 Preliminary
2.1 Explicit & Implicit Feedback and Problem Formulation
Formally, when the data is in the form of explicit feedback with single implicit
feedback, each user u is associated with two types of item sets: implicit item set
N(u) and explicit feedback set E(u). Explicit feedback is intentionally provided
by users to directly express user preferences (e.g., likes or dislikes) to items. For
an item i ∈ E(u), the rating given by user u to item i is denoted as R
ui
. Implicit
feedback reflects user opinions indirectly and can imply user probable preferences
[9]. For an item i ∈ N (u), the implicit feedback does not necessarily mean that
user u likes the item i.
When data consists of explicit feedback with multiple types of implicit feed-
backs, each user is associated with single explicit feedback and τ types of implicit
feedbacks (τ ≥ 2). For user u, the explicit item set is still denoted as E(u) which
contains items user u has rated (i.e., rating) on, and the implicit item sets are
denoted as N
1
(u), N
2
(u), · · · , N
τ
(u) where N
t
(u) contains items user u has ex-
pressed the t-type implicit feedback on (t = 1, · · · , τ ).
1
The datasets are available at https://github.com/7thsword/MFPR-Datasets.