利用地理位置邻域特性进行位置推荐：CIKM'14论文摘要

需积分: 9 175 浏览量更新于2024-08-26 收藏 731KB PDF 举报

本篇论文《Exploiting Geographical Neighborhood Characteristics for Location Recommendation》发表于2014年的第23届ACM国际信息与知识管理会议（CIKM '14），在新加坡管理大学（Singapore Management University, SMU）的研究集合学校——信息系统学院（School of Information Systems）进行研究。论文的主要作者包括 Yong LIU、Wei WEI、Aixin SUN 和 Chunyan MIAO，分别来自新加坡管理大学和南洋理工大学（Nanyang Technological University）。论文的核心内容探讨了如何利用地理邻域特性来改进位置推荐系统。地理位置信息在推荐系统中扮演着重要角色，特别是当涉及到服务或产品的个性化推荐时。通过分析和理解邻里区域的特征，如人口密度、商业分布、社会经济指标等，论文可能提出了新的算法或模型来预测用户可能感兴趣的位置，例如餐厅、商店或娱乐场所。具体而言，作者可能研究了以下几点： 1. 地理邻域聚类：识别和定义相似的地理位置群体，以便根据这些群组的共同特性来进行推荐。 2. 动态变化：考虑地理邻域随时间和空间的变化，以便提供实时且准确的推荐。 3. 社区感应：分析用户在特定地点的行为模式，以推断他们可能对邻近区域的兴趣。 4. 结合其他数据源：除了地理位置数据，论文可能还融合了其他数据，如用户历史行为、社交媒体信息或实时交通情况，以增强推荐的精度。 5. 实证研究：论文可能展示了在实际场景中的应用案例，以及这种策略如何改善推荐效果和用户体验。此外，论文可能还涉及到了以下方面： - 学术贡献：分享了SMU和NTU在知识机构层面上的合作成果，强调了跨校际研究合作的重要性。 - 方法论：详细解释了研究方法和技术，包括数据收集、处理和模型构建的过程。 - 评估与比较：对比了他们的方法与其他位置推荐技术，展示了优越性或创新之处。 - 可持续性和扩展性：讨论了模型的适用范围和未来可能的改进方向。最后，读者可以通过SMU图书馆的Ink数据库进一步探索作者们的相关研究成果，链接为<https://ink.library.smu.edu.sg/sis_research>，这表明该论文是机构知识库的一部分，反映了SMU在信息系统的学术积累。

mendation (see Section 2.3 for details). This kind of characteristics

are independent from individual users and should have potential

beneﬁts to location recommendation. Motivated by this, in this pa-

per, we propose to study geographical characteristics from location

perspective, with an emphasis on the geographical neighborhood.

Through our empirical analysis, we found that nearest neighbor-

ing locations tend to share more common visitors. This observa-

tion indicates that, in general, a user shares similar preferences

on nearest neighboring locations. In our method, this pattern is

used as the instance-level geographical neighborhood charac-

teristics. Speciﬁcally, by introducing a similarity measure, the re-

lationship underlying a user’s preferences on a few nearest neigh-

boring locations is captured and utilized to characterize her prefer-

ence on a target location.

In addition, our study on check-in data also shows that locations

in a geographical region may share similar user preferences. Each

geographical region is usually associated with a speciﬁc function

(e.g., business, entertainment, and education) [4, 29]. Such geo-

graphical regions contain additional prior knowledge about the in-

teractions between users and locations. Therefore, we use the geo-

graphical region structure as the region-level geographical neigh-

borhood characteristics. Speciﬁcally, a group lasso penalty is

employed to integrate the geographical region structure within the

check-in data into the learning of latent factors of users and loca-

tions. To summarize, the major contributions made in this paper

are as follows:

• We empirically analyze geographical characteristics from loca-

tion perspective using historical check-in data collected from

Gowalla. Two levels (i.e., instance-level and region-level) of

geographical neighborhood characteristics have been studied.

• We propose a novel location recommendation framework, i.e.,

nstance-Region Neighborhood Matrix Factorization (IRenMF),

to incorporate aforementioned geographical neighborhood char-

acteristics for improving location recommendation accuracy. In

particular, to solve the optimization problem in IRenMF, we pro-

pose an alternating optimization strategy, in which an acceler-

ated proximal gradient (APG) method is us ed to learn the latent

factors of locations, considering the geographical region struc-

ture of the check-in data.

• We extensively evaluate IRenMF on the datasets collected from

Gowalla. Experimental results show that: (1) Compared to the

baseline matrix factorization model (i.e., WRMF) without uti-

lizing geographical characteristics, our methods that use either

instance-level or region-level geographical neighborhood char-

acteristics signiﬁcantly improves recommendation accuracy; the

improvement is on average by 30.78% and 17.2% respectively,

in terms of precision metric P@5. (2) Utilizing both two lev-

els of geographical neighborhood characteristics, IRenMF sub-

stantially outperforms two state-of-the-art location recommen-

dation methods, which consider geographical characteristics de-

rived from user perspective.

The rest of this paper is organized as follows. Section 2 reviews

the most relevant work about this study. Section 3 introduces our

empirical analysis of check-in data and provides some background

about matrix factorization. Next, Section 4 presents the details of

IRenMF and describes the optimization solution to IRenMF. Then,

in Section 5, we report the experimental results conducted on real-

world datasets, which demonstrate that the geographical character-

istics derived from location perspective can signiﬁcantly improve

location recommendation accuracy. Finally, Section 6 draws the

conclusion of this study.

2. RELATED WORK

This study relates with three research areas: collaborative ﬁlter-

ing, structured sparsity learning, and personalized location recom-

mendation. Next, we will present an overview of the most related

work in each area.

2.1 Collaborative Filtering

The collaborative ﬁltering (CF) technique has been widely used

for personalized recommendation tasks [1]. In the literature, there

are two main categories of CF methods, namely memory-based CF

and model-based CF.

The most popular memory-based CF approaches are user-based

CF and item-based CF. The user-based CF predicts a user’s prefer-

ence on a given item based on the preferences of other similar users

on the target item. In contrast, the item-based CF ﬁrst ﬁnds similar

items to the target item and then predicts a user’s preference on the

target item, according to her preferences on other similar items. In

previous work [26,27], the user-based CF has been successfully ap-

plied for the location recommendation tasks, and it has been found

to signiﬁcantly outperform item-based CF. Similar results have also

been found in our experiments (see Section 5.2.1).

The model-based CF approaches aim to learn an algorithmic

model to explain the observed user-item interactions. Most exist-

ing model-based CF models deal with rating prediction problems,

in which users’ preferences are explicitly indicated by the rating

values given to items. However, in most practical scenarios, users’

rating data are unavailable. The personalized recommendations are

based on users’ implicit feedback, e.g., purchase history [20], click

history [5], and check-in history [3]. In [8], Hu et al. introduced the

ﬁrst attempts applying model-based CF approaches for large scale

implicit feedback datasets. Instead of ignoring all missing entries

in the user-item interaction matrix, they treated all missing entities

as negative examples and assigned vastly varying conﬁdence lev-

els to the positive and negative examples. Then, a least square loss

function was used to learn the latent features of users and items.

To simplify the computation, Pan et al. [19] proposed a sampling

framework to collect a fraction of missing entries in the user-item

matrix as negative examples. Recently, Rendle et al. [21] presented

a Bayesian Personalized Ranking (BPR) approach for recommen-

dation with users’ implicit feedback. This method only assumed

that the unobserved items (or negative items) were less interesting

than the observed items (or positive items) to an individual user. Its

objective was to rank the positive items above the negative items in

the positive-negative item pairs of each user.

2.2 Structured Sparsity Learning

For feature learning problems in high-dimensional space, spar-

sity is a desirable property of feature vectors. The L

-norm penalty

has been widely used to enforce sparsity to the learned feature

vectors. Numerous methods have been proposed to solve the L

regularization regression problems in s tatistics, machine learning

and computer vision. The readers can refer to [6] (Chapter 18) f or

more dis cussions.

However, the standard L

-norm penalty does not consider any

dependency (e.g., the group structure) among inputs. This limits

its applicability in many complex application scenarios. In recent

years, a lot of structured sparsity learning approaches have been

proposed to induce different joint sparsity patterns to related vari-

ables, by employing more s tructured constraints on inputs. Yuan

and Lin [30] proposed the group lasso model that assumes disjoint

group structures of inputs and enforces sparsity on the pre-deﬁned

groups of features. It employed a L

mixed norm penalty, with

p > 1, to achieve group level variable selection. In practice, L

剩余10页未读，继续阅读

Dronme

粉丝: 0
资源: 1

利用地理位置邻域特性进行位置推荐：CIKM'14论文摘要

BlueHat IL 2020PPT汇总（10份）.zip - 零信任

藏经阁-Exploiting-Network-Printers.pdf

HDNET- Exploiting HD Maps for 3D Object Detection.pdf

memetic q-learning

exploiting bert for end-to-end aspect-based sentiment analysis

exploiting cloze questions for few shot text classification and natural language inference

Explain 'int func(int x, int (*p)(int))'. exploiting this fucntion to rewrite this program in C programming languag

Defect-free Squares

adversarial tactics，techniques，and common knowledge是什么

融合麻雀算法中的麻雀飞行机制

最新资源