optimal for transitioning from watch video with topic τi to
a related video with topic τi, according to the loss criterion
defined in Equation 3.
5. SYSTEM OVERVIEW
In this section, we provide a holistic overview of the general
architecture of the experimental video suggestion system,
which was designed for evaluating the topic retrieval
methods described in the previous sections. This system
was designed specifically for the purpose of the experiments
in this paper, and does not correspond to any system used
in production.
We built this experimental system in order to integrate
both the co-view based and the topic-based retrieval methods
into a single video suggestion system, shown in Figure
4. The purpose of this integration is to demonstrate the
benefits of our topic-based video suggestion approach, as
compared to the standard co-view based video suggestion
approaches [7, 13].
For a given watch video VW , two retrieval processes are
run in parallel. First, we retrieve a set of k related candidates
based on the link structure of the co-view graph. See
Baluja et al. [7] (surveyed in Section 2.2) for the detailed
description of an implementation of such a process. Most
generally, this process retrieves all the videos that were most
oftenly co-watched with the watch video VW , regardless of
their topicality.
Second, we retrieve top-k ranking candidate videos using a
topic-based retrieval process. This process is based either on
the weighted topic retrieval or the diagonal topic transitions,
described in Section 3 and Section 4, respectively.
This set of k +k highest ranked results from both co-view
and topic retrieval processes is sent to the reranking model.
For fairness of our experimental setup, we ensure that the
number of candidate videos from both co-view and topic
retrieval processes are equal.
The final reranking model is out of the scope of this paper.
However, for our purposes, it is important to make two
general comments about it.
First, the reranking model takes into account a very large
number of features based on the watch-related video pair.
This makes the reranking model prohibitively expensive to
run on the entire video corpus. Instead, the retrieval processes
(based on either co-views or topics) provide a small
set of tens of potentially relevant video candidates to the
reranking model. By augmenting the co-view retrieval with
the top results from the topic retrieval, we seek to diversify
and improve this candidate set.
Second, to avoid bias in our evaluation, the reranking
model excludes any topic-related features, which are used
in the topic retrieval process. Therefore, we strive to ensure
that the candidates from the topic retrieval process are not
given an unfair advantage by the reranking model.
In the experiments in Section 7 we describe how the related
video suggestion system design outlined in Figure 4
is used to evaluate the benefits of the topic-based video retrieval
using both user simulation and a live traffic experiment.
6. RELATED WORK
In this paper, we propose two retrieval methods for related
video suggestion. The first retrieval method, described