社交网络视频聚类：多模态与集成簇算法的创新应用

15 浏览量更新于2024-08-27 收藏 1.91MB PDF 举报

本文主要探讨了"基于多模式和聚类集成的社交网络视频聚类"这一研究主题，发表在2019年的《神经计算》(Neurocomputing)期刊第366期，234-247页。作者Vinath Mekthanavanh、Tianrui Li、Jie Hu和Yan Yang来自中国西南交通大学信息科学技术学院，他们针对当前社交网络视频资源的爆炸性增长，提出了一个创新的方法来解决视频分类的问题。传统的社交网络视频聚类主要依赖于用户上传的视频文本标签，通过计算文本间的语义关系来进行视频分类。然而，这种方法面临挑战，因为准确衡量视频内容中的语义关联是一项困难任务。为了克服这个问题，研究者们引入了多模式和聚类集成策略，即结合多种数据模态（如视觉、音频、元数据等）以及多种聚类算法的优势，以提高视频分类的精度和效率。首先，文章可能涉及的关键技术包括： 1. **多模式融合**：这是一种综合处理不同类型数据的技术，旨在捕捉视频内容的多元信息，如文本描述、图像特征和音频信号，以更全面地反映视频的主题和内容。 2. **相似度函数**：在融合多个模态后，文章可能会介绍如何设计或选择合适的相似度度量方法，以便于计算不同视频之间的相关性和相似性。 3. **聚类算法集成**：研究可能比较了几种常见的聚类算法（如K-means、DBSCAN、层次聚类等），通过集成它们的优点，如避免单一算法的局限性，提高聚类结果的稳定性和有效性。 4. **语义关系建模**：为了更好地理解视频内容，可能使用自然语言处理技术，如词嵌入或深度学习模型，来捕捉和量化视频中的语义联系。 5. **社交网络特性**：考虑到社交网络的特性，如用户行为、兴趣社区和社交关系，论文可能会探讨这些因素如何影响视频聚类的性能。 6. **算法评估与改进**：文章可能还包含实验部分，展示了集成方法在真实社交网络数据集上的性能，并可能讨论了如何通过调整参数或优化算法进一步提升聚类效果。最后，这项研究的意义在于提升社交网络视频检索的用户体验，通过更精准的视频分类，帮助用户更容易找到符合需求的内容，从而提高用户满意度。同时，也为未来的社交网络内容分析和个性化推荐提供了有价值的研究基础。

236 V. Mekthanavanh, T. Li and J. Hu et al. / Neurocomputing 366 (2019) 234–247

2.2. Semantic relation and semantic similarity

The capture of semantic relation between terms and semantic

similarity has attracted many researchers’ attention in recent

years. A number of semantic relation and semantic similarity

measures have been proposed. The previously studied popular

semantic similarity methods were evaluated using WordNet

an underlying reference ontology. Billhardt et al. [31] proposed the

context vector model based on VSM, which incorporated term de-

pendencies and thus obtained semantically richer representations

of documents. Budanitsky and Hirst [32] showed an evaluation of

resource-based measures of lexical semantic distance, equivalently,

semantic relatedness, for natural language processing applications.

Mikolov and Dean [13] approached the literature of the semantic

relation between words by using the similarity of their context

information. Liu et al. [33] proposed a new short text modeling

method by combining the semantic information obtained from

a hierarchical lexical database and the statistical information ex-

tracted from the corpus involved. Resnik [34] presented a measure

of semantic similarity in an IS-A taxonomy based on the notion of

shared information content. Turney [35] proposed an algorithm to

measure the similarity of pairs of words by using the well-known

measure of semantic similarity: Pointwise Mutual Information

(PMI) and Information Retrieval (IR). Farahat and Kamel [36] pro-

posed new models for document representation that can capture

a semantic similarity between documents based on measures

of correlations between their terms. Gao et al. [37] addressed a

wordnet-based semantic similarity measurement by combining

edge-counting and information content theory.

2.3. Clustering ensemble

Clustering ensemble is an approach which is widely adopted in

clustering research areas. It combines multiple clustering results

to improve the quality of the ﬁnal result. Clustering ensemble

includes two main parts: diversity (creating multiple clusterings)

and consensus function (combining multiple clusterings). Strehl

and Ghosh [38] approached the concept of clustering ensemble by

applying graph theory to achieve the consensus clustering results.

Fred and Jain [39] used evidence accumulation for combining

multiple clusterings and demonstrated that evidence accumulation

outperforms the other combination approaches. Mimaroglu and

Erdil [40,41] approached a combination of multiple clusterings

by using the evidence accumulated from the clusterings. Other

solutions for combining multiple clusterings based on genetic

algorithm were proposed by Mohammadi et al. [42] . Azimi

et al. [43] presented an ensemble method based on the ant colony

algorithm, which can automatically determine the number of

clusters.

Semi-supervised clustering ensemble has become an inter-

esting problem in machine learning. Yang et al. [44] presented

a semi-supervised consensus clustering ensemble based on

multi-ant colonies algorithm. Iqbal et al. [45] proposed a semi-

supervised clustering ensemble by using a voting scheme. Wang

and Pan [46] exploited the spectral clustering to generate con-

sensus clustering with semi-supervised clustering. Mahmood

et al. [21] incorporated Must-Link constraint with graph tree

consensus clustering ensemble. Yang et al. [47] proposed a novel

semi-supervised multi-ant colonies consensus clustering algorithm

and parallelized it on MapReduce. Yu et al. [48] proposed a

new semi-supervised clustering ensemble which referred to an

incremental semi-supervised clustering ensemble approach, where

http://wordnet.princeton.edu/

the contribution is to develop an incremental ensemble member

selection based on local and global objective function.

3. System framework for social WVC

3.1. System overview

Fig. 1 shows the framework of Social WVC model. Firstly, the

textual information from social web videos such as title, tag, and

description is selected and then extracted them based on fea-

ture extraction. Secondly, the possible external sources are em-

ployed, e.g., WordNet, Word2vec and NGD from Google search en-

gine which are expected to capture the semantic information and

the feature relevance of terms list in documents. In this step, an

additional technique for the semantic relation between terms in

documents has been used to capture the relation from the local

views. After that, the similarities from each model are combined by

using the combination function to get single similarity before pass-

ing it to the clustering model. Thirdly, three clustering algorithms,

namely, aﬃnity propagation, spectral, and graph partitioning, have

been selected for the clustering purpose. The related video’s data is

included as pairwise constraint (Must-Link). Finally, we incorporate

the pairwise constraint into clustering ensemble in [38] to obtain

the ultimate results.

3.2. Feature e x traction

Each dataset consists of three subsets, e.g., title, tag and de-

scription, where each subset contains short information with noisy

and incomplete keywords. During the feature extraction, we use a

number of techniques, ( i.e. , word splitter, removing stop word

omit the most common words such as prepositions, articles, and

conjunctions, words stemming or lemmatization, tokenization) for

getting the useful information on videos.

3.3. Vector space model

In order to categorize the social web videos, we use Vector

Space Model (VSM) to compare two videos by using their textual

feature vectors. In this model, each video is represented as vectors

in a common vector space. The similarity between two videos is

measured by the OrS.

Deﬁnition 1 [49] . Term Frequency (TF). Suppose d is a document

and t is a term in d . TF(t,d) represents the frequency of a term in

a document.

T F (t, d) = f

, (1)

where f

is the frequency of term t in document d .

Deﬁnition 2

[49] . Inverse Document Frequency (IDF). Suppose D

is a document space and t is a term in D . IDF(t,D) is deﬁned as

follows:

IDF (t, D ) = log

1 + |{ d ∈ D : t ∈ d}|

, (2)

where N is the total number of documents in the corpus, and

|{ d ∈ D : t ∈ d }| is the number of documents where term t appears.

Deﬁnition 3 [49] . Term Frequency-Inverse Document Frequency

(TF-IDF). Suppose D is a document space, d ∈ D and t is a term in

D . The Term Frequency-Inverse Document Frequency (TF-IDF) of t

to d in D is deﬁned as follows.

T F − IDF (t, d, D ) = T F (t, d) × IDF (t, D ) (3)

www.ranks.nl/stopwords

剩余13页未读，继续阅读

weixin_38500948

粉丝: 3
资源: 915

社交网络视频聚类：多模态与集成簇算法的创新应用

基于知识粒化的层次聚类集成模型

具有高阶联合特征和聚类集成的大规模社交网络的隐写分析

大数据聚类的具有随机投影的模糊c均值和聚类集成

基于聚类和顺序聚类的高校数据挖掘分析.pdf

ClusterEnsembleV20_CSPA_聚类集成_

行业文档-设计装置-一种基于相似度矩阵谱分解的文本聚类集成方法.zip

基于聚类集成的微博话题发现算法优化

利用贝叶斯网络的半监督聚类集成方法

三支决策聚类集成：基于投票的区间集方法

高效谱聚类集成算法发现大规模网络重叠社区

最新资源