In many approaches to recommendation diversification, a recommender scores items for relevance and then re-ranks them to balance relevance with diversity. In intent-aware diversification, diversity is formulated in terms of coverage of aspects, where aspects are either explicit such as movie genres or implicit such as the latent factors found during matrix factorization. Typically, the same set of aspects is used across all users. In this paper, we propose a form of personalized intent-aware diversification, which we call SPAD (SubProfile-Aware Diversification). The aspects we use in SPAD are subprofiles of the user's profile. They are not defined in terms of explicit or implicit features. We compare SPAD to other forms of intent-aware diversification. We present empirical results in support of SPAD.
It has long been recognized that it is not enough for recommendations to be accurate or relevant. In many domains, recommendations must be novel to the user or serendipitous, and a set of recommendations must be diverse. Diversity is one response to uncertainty. A recommender cannot be certain of a user’s short-term or longer-term interests, both because some user profiles are small and some, while they may not be so small, will contain preferences over diferent kinds of items. In the face of uncertainty, a diverse set of recommendations is more likely to contain one or more items that will satisfy the user.
In many approaches to recommendation diversification, a recommender scores items for relevance and then re-ranks them to balance relevance with diversity. In intent-aware diversification [3], the idea is that the re-ranked recommendations should cover the diferent tastes or interests revealed by the user’s profile. The most common way to characterize a user’s tastes is as a probability distribution over so-called aspects of the items in the user’s profile. Aspects are usually either explicit features such as movie genres or implicit features such as the latent factors found during matrix factorization. Hence, typically, the same set of aspects is used across all users — only the probablilities vary across users. 2
The dominant approach to diversification is greedy re-ranking,
in which sets of recommendations RS for a user u are re-ranked
by considering the marginal contribution that would be made by
adding an item i to the result set RL. The marginal contribution
is measured by an objective function fobj (i, RL) which is typically
a linear combination of the item’s relevance score s (u, i ) and the
marginal contribution item i makes to the diversity of RL, div(i, RL),
the trade-of between the two being controlled by a parameter λ
(0 ≤ λ ≤ 1):
fobj (i, RL) = (1 − λ)s (u, i ) + λ div(i, RL)
(
Intent-aware diversification methods assume a set of aspects A which describe the items and for which user interests can be estimated. The aspects might be explicit: like genres such as comedy in a movie recommender. Alternatively, aspects might be implicit, e.g. corresponding to the latent factors found by a matrix factorization recommender system.
User u’s interests can be formulated as a probability distribution
p (a|u ) for aspects a ∈ A. The probability of choosing an item i
from the set of recommendations RS given an aspect a of user u is
denoted by p (i |u, a). In the Query Aspect Diversification framework
(xQuAD) [2, 4] , diversification can be achieved by re-ranking a
conventional recommender’s recommendation set as Equation (
X p (a|u )p (i |u, a) a ∈A
Y (1 − p (j |u, a))
j ∈RL
(
What characterizes the work on intent-aware diversification in recommender systems that we have described so far is the use of a global set of aspects. In our work, we infer the aspects from the user’s profile, making them personalized: the aspects for one user need not be the same for another. 3
In this section, we explain our new approach to diversification in recommender systems, which we call SubProfile Aware Diversification (SPAD). It is a greedy re-ranking approach; it is intent-aware; but it is also personalized, based on identifiable subprofiles within the user’s profile.
Let I be the set of all items. Subprofile detection works on positively-rated items in the user’s profile. In the case of positiveonly feedback, user u’s profile, Iu ⊆ I , is the set of items she has interacted with (liked, clicked on, purchased, etc.). In the case of explicit ratings rui (e.g. 1-5 stars), then Iu must be defined in terms of items the user liked, which will usually involve thresholding the ratings, e.g. in our experiments, we use Iu = {i |rui ≥ 4}. A user’s subprofiles are subsets of Iu .
Our approach to detecting user subprofiles is based on a method for recommending to shared accounts, called DAMIB-COVER [5]. DAMIB-COVER identifies diferent tastes within the profile of a shared account (which it assumes come from the diferent users who share that account) and recommends items to satisfy each taste. We adapt DAMIB-COVER to take in the profile for a single-user account u and to extract the diferent subprofiles Su that correspond to the diferent tastes of that user.
In the work on intent-aware diversification that we described earlier, the same set of aspects A was used for all users. In SPAD, aspects are user-specific: user u has set of aspects Au . And, in the earlier work, aspects were often based on explicit features F , i.e. A = F . In SPAD, aspects are user subprofiles, i.e. Au = Su . Each subprofile S ∈ Su contains a set of items from Iu . Diferent subprofiles can be of diferent lengths; the number of subprofiles can difer across users.
In SPAD, the set RS is greedily re-ranked using the objective
function given as Equation (
We compare SPAD to other re-ranking approaches on the MovieLens1M dataset with 5-fold cross validation. We show the results of taking recommendations made by matrix factorization (MF) and probabilistic latent semantic analysis (pLSA) algorithms and then re-ranking them using SPAD and other re-ranking approaches: 0.16 0.10 0.30 0.27 G DC0.24 n α0.21 0.25 0.25
MF cplsa xQuAD cplsa xQuAD
In Figure 1, we plot precision (an accuracy metric) and α -nDCG (a diversity metric) for diferent values of λ, which controls the amount of diversification. Notice that α -nDCG measures diversity with respect to the explicit features F (the meta-data). It therefore may favour recommenders that re-rank using those features. Our new method, SPAD, makes no use of the features and so it is at a disadvantage in these experiments.
For both baseline algorithms (MF and pLSA), SPAD has the highest precision. For the pLSA baseline, SPAD also has the highest diversity; for the MF baseline, SPAD’s diversity is competitive with the other re-ranking algorithms despite being at a disadvantage as mentioned earlier.
We plan to further explore the efectiveness of SPAD on other datasets, and with more baseline algorithms. We also plan to develop other subprofile detection methods instead of using DAMIBCOVER. We will also explore the interpretability of SPAD’s recommendations in terms of subprofiles.
This publication has emanated from research supported in part by a research grant from Science Foundation Ireland (SFI) under Grant Number SFI/12/RC/2289.