We have presented a new learning problem tailored to learning
from clustering feedback called collaborative clustering, which can
be viewed as a clustering analogue to collaborative filtering. We
proposed a latent factor model to learn the space of clustering variability
within a user population. We conducted empirical evaluations
based on usage data collected from a clustering interface developed
for a sensemaking task of exploring and organizing attractions in
Paris. Our results show that our approach significantly outperforms
conventional feature-based approaches on several realistic use cases.
In a sense, our approach for collaborative clustering is the simplest
one that inherits the benefits of both tensor factorization as
well as metric learning. It may be interesting to incorporate other
advancements in collaborative filtering, such as localized latent
embeddings, implicit feedback, and temporal dynamics [16, 15].
Another limitation of our approach is the inability to boost performance
by using a joint model of both latent and observed features.
Having a feature-based model is important for tackling the so-called
cold-start problem for brand-new items with no feedback. This limitation
may be due to the relatively simple linear content-based model
we employed. For instance, recent work in collaborative filtering
has demonstrated the ability to achieve the “best of both worlds” by
effectively combining a latent factor model with a content-based
(non-linear) topic model [30], and a similar approach may be fruitful
for collaborative clustering. It may also be that user features are
more useful than item features since training data per user is low.
Although we developed and validated our LCC approach for
learning personalized clustering models, we did not formally model
the full interactive setting where the system adaptively adjusts its
recommendations based on user feedback. This full interactive setting
can be considered as a type of interactive clustering problem.
The most relevant related work on interactive clustering are [5, 9],
although neither addressed how to model the end-to-end interaction
sequence.12 Other work such as [3] does provide guarantees for the
interactive setting, but assumes a very different (and less realistic)
interaction model where users must provide feedback on full clusterings
of all items. Other adaptive clustering work include learning
a single similarity function via crowdsourcing [28].
The broader goal is to design personalization frameworks that
learn from multiple types of rich interactions (e.g., dynamic rankings
[8] and zoomable metro maps [25]), as well as reason about (and
optimize for) long-term user utility over entire interactive sessions.
Progress towards this goal will require a confluence of progress in in-
We have presented a new learning problem tailored to learningfrom clustering feedback called collaborative clustering, which canbe viewed as a clustering analogue to collaborative filtering. Weproposed a latent factor model to learn the space of clustering variabilitywithin a user population. We conducted empirical evaluationsbased on usage data collected from a clustering interface developedfor a sensemaking task of exploring and organizing attractions inParis. Our results show that our approach significantly outperformsconventional feature-based approaches on several realistic use cases.In a sense, our approach for collaborative clustering is the simplestone that inherits the benefits of both tensor factorization aswell as metric learning. It may be interesting to incorporate otheradvancements in collaborative filtering, such as localized latentembeddings, implicit feedback, and temporal dynamics [16, 15].Another limitation of our approach is the inability to boost performanceby using a joint model of both latent and observed features.Having a feature-based model is important for tackling the so-calledcold-start problem for brand-new items with no feedback. This limitationmay be due to the relatively simple linear content-based modelwe employed. For instance, recent work in collaborative filteringhas demonstrated the ability to achieve the “best of both worlds” byeffectively combining a latent factor model with a content-based(non-linear) topic model [30], and a similar approach may be fruitfulfor collaborative clustering. It may also be that user features aremore useful than item features since training data per user is low.Although we developed and validated our LCC approach forlearning personalized clustering models, we did not formally modelthe full interactive setting where the system adaptively adjusts itsrecommendations based on user feedback. This full interactive settingcan be considered as a type of interactive clustering problem.The most relevant related work on interactive clustering are [5, 9],although neither addressed how to model the end-to-end interactionsequence.12 Other work such as [3] does provide guarantees for theinteractive setting, but assumes a very different (and less realistic)interaction model where users must provide feedback on full clusteringsof all items. Other adaptive clustering work include learninga single similarity function via crowdsourcing [28].The broader goal is to design personalization frameworks thatlearn from multiple types of rich interactions (e.g., dynamic rankings[8] and zoomable metro maps [25]), as well as reason about (andoptimize for) long-term user utility over entire interactive sessions.Progress towards this goal will require a confluence of progress in in-
การแปล กรุณารอสักครู่..
