当前位置：搜档网 › Exploiting Geographical Influence for Collaborative Point-of-Interest Recommendation

Exploiting Geographical Influence for Collaborative Point-of-Interest Recommendation

Exploiting Geographical In?uence for Collaborative Point-of-Interest Recommendation

Mao Y e1?,Peifeng Yin1?,Wang-Chien Lee1?,and Dik-Lun Lee2?

?Department of Computer Science and Engineering,The Pennsylvania State University,P A,USA.

?Department of Computer Science and Engineering,HKUST,Hong Kong

1{mxy177,pzy102,wlee}@https://www.sodocs.net/doc/342316387.html,2dlee@https://www.sodocs.net/doc/342316387.html,t.hk

ABSTRACT

In this paper,we aim to provide a point-of-interests(POI) recommendation service for the rapid growing location-based social networks(LBSNs),e.g.,Foursquare,Whrrl,etc.Our idea is to explore user preference,social in?uence and geo-graphical in?uence for POI recommendations.In addition to deriving user preference based on user-based collabora-tive?ltering and exploring social in?uence from friends,we put a special emphasis on geographical in?uence due to the spatial clustering phenomenon exhibited in user check-in ac-tivities of LBSNs.We argue that the geographical in?uence among POIs plays an important role in user check-in behav-iors and model it by power law distribution.Accordingly, we develop a collaborative recommendation algorithm based on geographical in?uence based on naive Bayesian.Further-more,we propose a uni?ed POI recommendation framework, which fuses user preference to a POI with social in?uence and geographical in?uence.Finally,we conduct a compre-hensive performance evaluation over two large-scale datasets collected from Foursquare and Whrrl.Experimental results with these real datasets show that the uni?ed collaborative recommendation approach signi?cantly outperforms a wide spectrum of alternative recommendation approaches. Categories and Subject Descriptors

H.3.3[Information Search and Retrieval]:Information Filtering;J.4[Computer Applications]:Social and Be-havior Sciences

General Terms

Algorithms,Experimentation.

Keywords

Collaborative Recommendation,Location-based Social Net-works,Geographical In?uence.

Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for pro?t or commercial advantage and that copies bear this notice and the full citation on the?rst page.To copy otherwise,to republish,to post on servers or to redistribute to lists,requires prior speci?c permission and/or a fee.

SIGIR’11,July24-28,2011,Beijing,China.

With the rapid development of mobile devices,wireless networks and Web2.0technology,a number of location-based social networking services,e.g.,Loopt1,Brightkite2, Foursquare3and Whrrl4,have emerged in recent years.5 These LBSNs allow users to establish cyber links to their friends or other users,and share tips and experiences of their visits to plentiful point-of-interests(POIs),e.g.,restaurants, stores,cinema theaters,etc.In LBSNs,a POI recommen-dation service,aiming at recommending new POIs to users in order to help them explore new places and know their cities better,is an essential function that has received a lot of research momentum recently[25,

26].

Figure1:Graph representation of user-user friend-ship and user-location check-in activity in a LBSN

Indeed,facilitating POI recommendations in LBSNs is a promising and interesting research problem because valuable information such as the“cyber”connections among users as well as the“physical”interactions between users and lo-cations have been captured in the systems.Nevertheless, these information have not been fully explored in prior re-search studies relevant to POI recommendations.For ex-ample,Zheng et.al.have extracted visited locations from GPS trajectory logs of mobile users for location recommen-dations[25,26].However,their studies consider neither the social links between users nor the interactions between users and locations in the recommendation process.In this paper, we aim to exploit the unique geographical implications https://www.sodocs.net/doc/342316387.html,

https://www.sodocs.net/doc/342316387.html,

5These services are often referred as location based social networks and thus abbreviated as LBSNs in the paper.

bedded in users’interactions with locations,in addition to applying the social in?uence from users’friends,for POI

recommendations in LBSNs.

Users and POIs are two essential types of entities in LB-SNs.As illustrated in Figure1,users in an LBSN,de-

noted as u1,u2,u3,u4,are interconnected via social links to form a user social network.Moreover,POIs,denoted as

l1,l2,...,l6,are connected with users via their“check-in”ac-tivities,which generally re?ects the users’tastes on various POIs.Finally,as also logically illustrated in the?gure,the POIs,geocoded by longitude,latitude ,are constrained ge-ographically.To make recommendations of POIs to users, obviously the records of previous user check-in activities are very useful.With the availability of such information in LBSNs,an intuitive idea for supporting POI recommenda-tions is to employ the conventional collaborative?ltering (CF)techniques by treating POIs as the“items”in many successful CF-based recommender systems.The basic argu-ment for this idea is that users’tastes can be deduced by other users who exhibit similar visiting behaviors to POIs in previous check-in activities.Thus,user-based or item-based collaborative?ltering techniques may be applicable to POI recommendations.Additionally,the social network of users,which is handily available in the LBSN,can be explored to enhance performance of POI recommendations. Recent studies have argued that social friends tend to share common interests and thus can be used in the process of collaborative?ltering for making recommendations[11,12, 2,13,24].

While the ideas above aim to explore the essential infor-

mation available in LBSNs,i.e.,the user-location interactiv-ities and user-user social links,we argue that the geographi-cal in?uence naturally existing in the activities of users and their geographical proximities cannot be ignored.According to Tobler’s First Law of Geography“Everything is related to everything else,but near things are more related than distant things”[19].Thus,a user intuitively tends to visit nearby POIs.There are two major implications that can be derived from this intuition for POI recommendations:(1) people tend to visit POIs close to their homes or o?ces;and (2)people may be interested in exploring nearby POIs of a POI that they are in favor of,even if it is far away from home,e.g.,a user may explore some restaurants and shops around Time Square when she goes there for a broadway show.Due to the geographical nature of the LBSNs,we believe the geographical in?uence between users and POIs as well as that amongst POIs are as important as the so-cial in?uence amongst users,which as indicated earlier may play a positive role for supporting POI recommendations in LBSNs.In short,we are interested in studying the im-pact of geographical in?uence and social in?uence on POI recommendations in LBSNs.

Our approach for supporting POI recommendation ser-

vice in LBSNs is to develop e?ective collaborative recom-mendation techniques that discover POIs of users’interests by incorporating the three complementary factors:i)user preference of POIs;ii)social in?uence;and iii)geographical in?uence.Notice that users’implicit preferences of POIs can be derived from their check-in activities on POIs.By con-sidering two users who have checked into a lot of common POIs as similar users,we may discover the implicit prefer-ence of a user through the previous check-in activities of her similar users.Recall the example in Figure1.Since u1and u2shares many commonly visited POIs,they may be con-sidered as similar users who are assumed to share similar check-in behaviors,i.e.,preference of POIs.As a result,l1 is a good candidate for recommendation to user u2since u1 has visited this POI before.On the other hand,social in?u-ence of friends can be incorporated in the recommendation process.For example,when considering l4as a recommen-dations candidate for u1,the social in?uence of u4on u1may contribute to the decision making.Finally,the geographical in?uence of POIs on nearby POIs can be considered.As shown in the example,since u2has visited l2and l3before, their nearby POIs l1and l5may be considered positively due to the geographical in?uence.

As discussed earlier,the conventional item recommenda-tion techniques based on user preference[9,18,14,15,13] and social in?uence[12,13]seem to be applicable for POI recommendation.Nevertheless,their e?ectiveness on POI recommendations in LBSNs have not been studied.Most importantly,the idea of incorporating the geographical in-?uence between POIs,which is refreshing and promising for POI recommendation,has not been investigated previously. In this paper,we examine the“geographical clustering phe-nomenon”of user check-in activities in LBSNs and propose a power-law probabilistic model to capture geographical in-?uence among POIs.Accordingly,we realize the targeted collaborative POI recommendation service for LBSNs by in-corporating the geographical in?uence of POIs via Bayesian theory.Finally,we propose a uni?ed location recommenda-tion framework to fuse user preference to POIs along with the social in?uence among users and the geographical in?u-ence among POIs.

In summary,the contributions we made in this research work are four-fold.

?We study the problem of supporting POI recommenda-

tion in location-based social networking systems(LB-

SNs),where POIs are uniquely di?erent from other

recommended items in conventional recommender sys-

tems because of the fact that“physical”interactions

are required between users and POIs.Hence,we in-

vestigate the geographical in?uence between POIs and

propose to incorporate geographical in?uence along

with user preference and social in?uence in the col-

laborative recommendation techniques we develop for

POI recommendations in LBSNs.

?We develop a novel idea to capture the geographical

in?uence by investigating the geographical clustering

phenomenon of user check-in activities in LBSNs.We

propose to employ a power-law probabilistic model to

capture the geographical in?uence among POIs,and

realize our collaborative POI recommendations based

on geographical in?uence via naive Bayesian method.

?We propose a uni?ed recommendation framework for

POI recommendations by fusing user preference,social

in?uence and geographical in?uence to devise a check-

in probability prediction model for a given user to visit

a POI.

?Finally,we evaluate the proposed recommendation tech-nique over large-scale datasets we collected from two

well known LBSNs,i.e.,Foursquare and Whrrl.Exper-

imental results show that our proposed collaborative

recommendation technique exhibits superior POI rec-

ommendation performance against other approaches.

Important?ndings of our evaluation are summarized

below.

–Geographical in?uence shows a more signi?cant impact on the e?ectiveness of POI recommen-dations in LBSNs than social in?uence,improv-ing the recommendation performance by at least 13.8%against social in?uence.–Random Walk and Restart [12]may not be suit-able for POI recommendation in LBSNs.Based on our analysis over the real data and the ex-perimental results,we ?nd that friends still re-?ect signi?cantly di?erent preferences and social tie cannot re?ect the similarity of check-in behav-ior among users.–Item-based collaborative ?ltering do not provide results comparable to user-based collaborative ?l-tering,because many POIs,in the current state of LBSNs,only show a few user check-ins.Thus item similarity is not as accurate as user similar-ity.

The rest of this paper is organized as follows.In Section 2,we provide some background on conventional recommenda-tion techniques according to user’s own preference and so-cial in?uence and review related works in the literature.In Section 3,we describe the location recommendation process according to geographical in?uence.In Section 4,we pro-pose a location recommendation framework,which uni?es all three factors together.In Section 5,we perform an empirical study on the di?erent location recommendation algorithms upon two large scale datasets crawled from Foursquare and Whrrl,respectively.Finally,in Section 6,we conclude the paper.

2.PRELIMINARIES

In this section,we ?rst provide background on user-based collaborative ?ltering and friend-based recommendation,which serve as the building blocks in our fusion approach to exploit user preference and social in?uence.Next we review some relevant studies in recommender systems.

2.1User-based Collaborative Filtering

Based on collaborative ?ltering,users’implicit preference can be discovered by aggregating the behaviors of similar users.Let U and L denote the user set and the POI set in an LBSN,which keeps track of check-in activities in the system.The check-in activity a user u i ∈U has at a POI l j ∈L is denoted as c i,j where c i,j =1represents u i has a check-in at l j before and c i,j =0means there is no record of u i visiting l j .These recorded user check-in activities at POIs are thus used to discover a user’s implicit preference of a POI,which can be represented as a probability to predict how likely the user would like to have a check-in at an unvisited POI.We denote this prediction by c i,j and obtain this predicted check-in probability of u i to l j as follows.

c i,j =

u k w i,k ·c k,j u k w i,k (1)where w i,k is the similarity weight between users u i and u k .To compute the similarity weights w i,k between users u i and u k ,several similarity measures can be adopted,e.g.,cosine similarity and Pearson correlation.In our study,we choose cosine similarity due to its simplicity.The cosine

similarity weight between users u i and u k ,denoted as w U

i,k ,

is de?ned as follows.

w i,k =

l j ∈L c i,j c k,j

l j ∈L c 2

i,j

l j ∈L c 2

k,j

(2)

2.2Friend-based Collaborative Filtering

Friends tend to have similar behavior because they are friends and might share a lot of common interests,thus lead-ing to correlated check-in behaviors [15,13].For example,two friends may hang out to see a movie together sometimes,or a user may go to a restaurant highly recommended by her friends.All those possible reasons suggest that friends might provide good recommendation for a given user due to their potential correlated check-in behavior.In other words,we can turn to user’s friends for recommendation,and we call it recommendation based social in?uence from friends.

POI recommendations based on social in?uence can be realized by the friend-based collaborative ?ltering approach as described in [13]. c i,j =

u k ∈F i SI k,i ·c k,j u k ∈F i SI k,i

(3)where c i,j is the predicted check-in probability of u i at l j ,F i is the friends set of u i ,and SI k,i is directional social in?uence weight u k has on u i [14,15,13].

On the one hand,we think friends who have closer social tie may have better trust in terms of their recommendation;on the other hand,friends who show more similar check-in behavior should have more similar tastes with the active user,thus suggestions from those friends are more worthy.Thus,in the following,we introduce how to derive the social in?uence weight by combining the above two aspects.

One way to derive the social in?uence weight between two friends is based on both of their social connections and sim-ilarity of their check-in activities [12].

SI k,i =η·|F k ∩F i |

|F k ∪F i |+(1?η)·|L k ∩L i ||L k ∪L i |

(4)

where ηis a tuning parameter ranging within [0,1],and F k and L k denote the friend set and POI set of user u k ,respectively.6

Another way of measurement is via the Random Walk with Restart (RWR)technique [20]over the graph that cap-tures both the social connections among users as well as the check-in activities between users and POIs [24].Starting from a node k ,an RWR is performed by randomly following a link to another node at each step.Notice that there is a probability a in every step to restart at node k .By it-erating RWR repeatedly until the whole process converges,a stationary (or steady-state)probability for each node can be obtained.The stationary probabilities of nodes give us a long-term visit rate for each user node (e.g.,user u i )given a bias towards a particular starting node (https://www.sodocs.net/doc/342316387.html,er u k ).This can be interpreted as the social in?uence weight user u k have on u i ,i.e.,SI k,i .

2.3Related Work

Content-based and collaborative ?ltering techniques are

two widely adopted approaches for recommender systems [1].A content-based system selects items for recommendation based on the similarity between item content (e.g.,key-words/tags describing the items)and user pro?le [3,8,17].6

The friend set of a user refers to the socially connected friends of the user in the LBSN,while her POI set refers to the set of POIs she has check-in activities.

Since it mainly relies on dictionary-bound relations between the terms used in user pro?les and item content,implicit associations between users are not considered.

The collaborative?ltering systems are divided into two categories,i.e.,memory-based and model-based.Memory-based systems can be further classi?ed into user-based and item-based systems.For user-based systems[9],the simi-larity between all pairs of users is computed based on their ratings on associated items using some selected similarity measurement such as cosine similarity or Pearson correla-tion.Based on the user similarity,missing rating corre-sponding to a given user-item pair can be derived by com-puting a weighted combination of the ratings upon the same item from similar users.For item-based systems[18],instead of using similarity between users to predict missing rating, predications are made by?nding similarly rated items?rst in order to compute a weighted combination of user ratings upon similar items.On the other hand,the model-based collaborative?ltering systems assume that users may form clusters based on their similar behavior in rating items.A model can be learned based on patterns recognized in the rating behaviors of users using machine learning techniques such as clustering algorithms or Bayesian networks[5,23]. Under the context of social networking systems,social friendship is shown to be bene?cial for collaborative?ltering based recommendation systems,e.g.,memory-based[11,12] and random walk based[2,11,12].These works argue that social friends tend to share common interests and thus their relationships should be considered in the process of collab-orative?ltering.Random walk captures a social network as a graph with probabilistic weighted links to represent social relations and thus is able to accurately predict user prefer-ences to items[12]and social in?uence to other users[24]. On the other hand,social friendship has also been explored in the model-based systems[14,13].These work mostly focus on conventional recommendation systems for recom-mending items such as movies.

Recently,location recommendation and mining has at-tracted a lot of attentions from the research community[27, 26,7,25,22].Among them,[27,26,7,25]are mainly fo-cused on GPS datasets which do not consider social rela-tionships among users.In these works,unfortunately,the geographical in?uence among POIs are not explored[27,26, 25].Recently,the correlation of locations in GPS trajecto-ries are explored[7].In this work,however,locations are still treated as conventional items.As such,the correlations between locations are established through users’activities instead of their geographical in?uence.[22]is the?rst re-search to provide location recommendations services in LB-SNs,but with the goal of improving e?ciency of location recommendation.

Our study di?erentiates itself from all these prior works in four aspects:i)the application domain of location-based social networking systems,embracing both social and geo-graphical features in the captured data,is new and unique; ii)the study of social in?uence and geographical in?uence in recommender systems for LBSNs is unexplored previously; iii)the proposal of uni?ed collaborative recommendation approach,which incorporates geographical in?uence along with user preference and social in?uence,is new and inno-vative;iv)two large-scale real dataset collected from well known LBSNs,namely,Foursquare and Whrrl datasets,are adopted for performance evaluation.

Figure2:Geographical in?uence probability distri-bution

3.RECOMMENDATION VIA GEOGRAPH-

ICAL INFLUENCE

As mentioned earlier,the check-in activities of users in LBSNs record their physical interactions(i.e.,visits)at POIs, Thus,we argue that the geographical proximities of POIs have a signi?cant in?uence on users’check-in behavior.To better understand this geographical in?uence on users,we perform a spatial analysis on real datasets of user check-in activities collected from two well known LBSNs,i.e.,Foursquare and Whrrl.Speci?cally,we aim to study the implication of distance on user check-in behavior by measuring how likely two of a user’s check-in POIs are within a given distance. To obtain this measurement,we calculate the distances be-tween all pairs of POIs that a user has checked in and plot

a histogram(actually probability density function)over the distance of POIs checked in by the same user.As shown in Figure2,a signi?cant percentage of POIs pairs checked in

by the same user appears to be within short distance,indi-cating a geographical clustering phenomenon in user check-in activities.7This phenomenon may be attributed to the ge-ographical in?uence which may be intuitively explained by the following tendencies:(1)people tend to visit POIs close

to their homes or o?ces;and(2)people may be interested

in exploring nearby POIs of a POI that they are in favor of,even if it is far away from home.As a result,the POIs visited by the same user tend to be clustered geographically. We believe that this geographical clustering phenomenon in user check-in activities can be exploited for POI recommen-dations in LBSNs.Thus,in the following,we study and model this geographical in?uence on user check-in behavior

at POIs,aiming to utilize it in POI recommendations.

To achieve our goal,we would like to compute the like-lihood that a user u i would check in both POI l j and l k. Based on Figure2,we intuitively think the check-in proba-bility may follow the power-law distribution.Nevertheless, we observe that the check-in probability of POI pairs vis-ited by the same person over distance is not a standard power-law distribution.Even though the left part of the

?gure decreases linearly(i.e.,decreases exponentially in reg-ular scale)and thus?ts power-law distribution very well,the right part may sometimes deviate irregularly(i.e.,the prob-ability is high at some points).A reasonable explanation is that users may travel to di?erent places and thus create mul-tiple check-in spatial clusters.Generally speaking,the fact that a user’s check-in POIs tend to be in a short distance is con?rmed in our data analysis.As mentioned earlier,nearby POIs are more related to each other,which exhibits strong geographical in?uence.Moreover,the linear portion of the

7Note that the?gure has been shown in log-log scale.

plot in Figure2covers the majority(90%)of the POI pairs. Thus,we propose to use power law distribution to model the check-in probability to the distance between two POIs visited by the same user as follows.

y=a×x b(5) where a and b are parameters of a power-law distribution, and x and y refer to the distance between two POIs visited by the same user and its check-in probability,respectively. Equation(5)can transformed into Equation(6)in“log-log”scale to?t a linear model.

log y=w0+w1log x(6) Thus,the original power-law distribution can be recovered via the following equation.

a=2w0b=w1(7) Hence,we can simply apply a linear curve?tting method to realize regression as follows.More speci?cally,let y = log y and x =log x.We shall?t data as follows

y (x ,w)=w0+w1·x (8) where w0and w1are the linear coe?cients,collectively de-noted by w.In other words,the model can be learned in form of w.In order to avoid over-?tting,we approach the weight coe?cients by least square error method and add a penalty term(i.e.,regularization term)to discourage the coe?cients from reaching large values as below[6].

E(w)=1

n=1

{y (x n,w)?t n}2+λ

||w||2(9)

where E(w)denotes the loss function,N presents the cardi-nality of input dataset,t n is the ground truth corresponding to x n,andλis the regularization term.

Accordingly,the optimal values of a and b form the setting which minimizes the loss function E(w)as below.

opt{a,b}=arg min

a,b

E(w)(10) In the following,we introduce a collaborative recommen-dation method based on the naive Bayesian method to real-ize POI recommendation in LBSNs.For a given user u i and its visited POI set L i,we de?ne the probability that u i has check-in activities at all locations in L i by considering the pair-wise distances of POIs in L i as follows.

P r[L i]=

l m,l n∈L i∧m=n

P r[d(l m,l n)](11)

where d(l m,l n)denotes the distance between POIs l m and l n,and P r[d(l m,l n)]=a×d(l m,l n)b which follows the pow-law distribution model we obtained above.Note that here we assume the distances of POI pairs are independent. Thus,for a given POI l j(i.e.,the recommendation can-didate),user u i,and her visited POI set L i,we have the likelihood probability for u i to check in l j as follows.

P r[l j|L i]=P r[l j∪L i]

P r[L i]

P r[L i]×

l y∈L i

P r[d(l j,l y)]

P r[L i]

l y∈L i

P r[d(l j,l y)]

(12)

To make a POI recommendation,we sort all the POIs in L?L i in accordance with their P r[l j|L i](l j∈L?L i)to return the POI with the highest P r[l j|L i]to the user.4.UNIFIED COLLABORATIVE POI REC-

OMMENDATION

In this section,we propose a uni?ed framework to perform

collaborative recommendation,which fuses ideas factors of

user preference,social in?uence and geographical in?uence in POI recommendation.Notice that,di?erent from pre-

dicting a POI’s rating,we aim to return a ranked list of

candidate POIs,which is very similar to conventional infor-mation retrieval[4].

4.1Fusion Framework

As discussed,each factor,i.e.,user preference,social in-

?uence or geographical in?uence,can be utilized to realize

POI recommendation.Thus,we intuitively can implement three di?erent recommender systems.We propose to use a

linear fusion framework to integrate ranked lists provided

by the three above-mentioned recommenders into the?nal ranked list[4,21].By integrating multiple recommenders,

top-ranked POIs from each of the recommendation algo-

rithms could increase both recall(due to the di?erent highly

ranked POIs)and precision(giving that the recommender systems have a high density of user-preferred POIs on top

of the results lists.

Let S i,j denote the check-in probability score of user u i at POI l j,i.e.,the more likely u i has a check-in activity at l j,

the larger S i,j is.Let S u i,j,S s i,j and S g i,j denote the check-

in probability scores of user u i at POI l j,corresponding to recommender systems based on user preference,social

in?uence and geographical in?uence,respectively.We have

S i,j as follows.

S i,j=(1?α?β)S u i,j+αS s i,j+βS g i,j(13) where the two weighting parametersαandβ(0≤α+β≤1)

denote the relative importance of social in?uence and ge-

ographical in?uence comparing to user preference.Here α=1states that S i,j depends completely on the prediction based on social in?uence;β=1states that S i,j depends

completely on the predication based on geographical in?u-ence;whileα=β=0states that S i,j counts only on user preference.

4.2Check-in Probability Score Estimation According to the above fusion framework,in order to esti-mate the check-in probability score S i,j,we need to predict the check-in probability score of S u i,j,S s i,j and S g i,j corre-sponding to user preference,social in?uence and geograph-ical in?uence,respectively.Accordingly,we estimate the check-in probability p u i,j,p s i,j and p g i,j for a user u i to visit a POI l j in order to obtain S u i,j,S s i,j and S g i,j,respectively. First,the predication of p u i,j can be estimated based on the idea of user-based collaborative?ltering as discussed before. More speci?cally,we utilize the behavior of similar users to realize the predication as Equation(1).Thus we have

p u i,j=

u k

w i,k·c k,j

u k

w i,k

(14) where w i,k is the similarity weight between users u i and u k. Similarly,the prediction of p s i,j can be estimated based on based on the idea of friend-based collaborative?ltering. Thus,according to Equation(3),we have

p s i,j=

u k∈F i

SI k,i·c k,j

u k∈F i

SI k,i

(15) where F i is the friends set of u i,SI k,i is the weight measuring social in?uence from u k to u i.

Finally,p g i,j can be directly obtained from Equation(12) p g i,j=P r[l j|L i]=

l y∈L i

P r[d(l j,l y)](16)

where L i is the visited POI set of u i,and d(l j,l y)denotes the distance between POIs l j and l y.

After we get the check-in probability estimation,we ob-tain the corresponding scores as follows.

S u i,j=p u i,j

Z u i

,where Z u i=max l j∈L?L i{p u i,j}

S s i,j=p s i,j

Z s i

,where Z s i=max l j∈L?L i{p s i,j}

S g i,j=p g i,j

Z g i

,where Z g i=max l j∈L?L i{p g i,j}

(17)

where Z u i,Z s i and Z g i are normalization terms.

5.EMPIRICAL EV ALUATION

In this section,we design and conduct several experiments to compare the recommendation qualities of the proposed collaborative recommendation algorithms with some state-of-the-art recommendation techniques,including collabora-tive?ltering and random walk with restart,and to investi-gate several interesting questions.Speci?cally,the design of the experiments aims to achieve the following goals.(1)As our proposed method factors in user preference,social in-?uence from friends and geographical in?uence from nearby location,we intent to study parametersαandβto under-stand the roles/weights of the above-mentioned factors in obtaining optimal recommendations.(2)We intend to val-idate our ideas by comparing the e?ectiveness of the pro-posed approach with other state-of-the-art techniques.(3) Due to the growing research interests in social in?uence from friends,we intend to further study the similarity of check-in behaviors in terms of the strength of“social ties”between two friends.(4)In our proposal,user-based collaborative ?ltering approach has been employed to discover user pref-erence.We intend to explore the feasibility and necessity of integrating item-based collaborative?ltering approach to further enhance the recommendation quality.(5)We would like to understand how data sparsity may a?ect POI rec-ommendations in LBSNs.(6)How well our techniques deal with cold start users,who do not have many check-in records for discovery of their interests[10].

5.1Dataset Description

We crawled the websites of Foursquare and Whrrl,two of the most representative LBSNs,for a month to collect two datasets consisting of153,577users and96,229POIs in Foursquare,and5,892users and53,432POIs in Whrrl, respectively.Our performance evaluation is conducted on these two large-scale real datasets.After summarizing the check-in records,we get the user-POI check-in matrix densi-ties as4.24×10?5for Foursquare dataset and2.72×10?4for Whrrl datasets,respectively.Note that,the e?ectiveness of recommendation service for sparse dataset(i.e.,low density user-POI check-in matrix)is usually not high due to the lim-ited information provided by the dataset.For example,the reported precision in[12]is0.17over a pre-prossed dataset with7.8×10?4density.Thus,in our experiments,we focus on observing the relative performance of algorithms instead of their absolute e?ectiveness measures,which we expect to improve as the number of LBSN users continues to grow and more check-in activities are logged.To facilitate our evalu-ation,for each individual user in the datasets,we randomly mark o?x%(x=10,30,50(with30as the default value)of

all POIs visited by the user.In the experiments,the eval-uated POI recommendation algorithms are used to recover the missing user-POI pairs that have been marked o?.

5.2Performance Metrics

A POI recommendation algorithm under evaluation com-putes a ranking score for each candidate POI(i.e.,POI that user has not visited)and returns the top-N highest ranked POIs as recommendations to a targeted user.To evalu-ate the prediction accuracy,we are interested in?nding out how many POIs previously marked o?in the preprocessing step recovered in the returned POI recommendations.More speci?cally,we examine two metrics:(1)the ratio of recov-ered POIs to the N recommended POIs,and(2)the ratio of recovered POIs to the set of POIs deleted in preprocessing. The former is precision@N while the latter is recall@N,and collectively referred as performance@N.In our experiment, we test the performance when N=5,10,20with5as the default value.

5.3Evaluated Recommendation Approaches Three factors,namely user preference(U),social in?uence from friends(S)and geographical in?uence from POIs(G), are incorporated in our uni?ed collaborative recommenda-tion algorithm,denoted by USG in our evaluation.A num-ber of state-of-the-art and new collaborative?ltering ap-proaches,some of which can be con?gured by controlling the weight parameters,0<α,β<1,in USG,are also evaluated

for comparison.In addition of USG,the recommendation approaches under evaluation are listed below.

?user-based CF(denoted by U)-this is a special case

of USG by setting bothαandβas zeros.In other words,

only user preference is considered for recommendation.

?friend-based CF(denoted by S)-this is also a spe-

cial case of USG,whereα=1.Here,only friends of the

targeted user are used in making a speci?c recommen-

dation.As introduced before,there are two alterna-

tive methods to derive the social in?uence weight be-

tween friends.One is to compute the social in?uence

weight based on friends based on Equation(4)[12]

and the other is to derive social in?uence weight be-

tween friends using Random Walk and Restart tech-

nique[24].To di?erentiate these two approaches,we

denote them as S and S rwr,respectively.

?GI-based recommendation(denoted by G)-this ap-

proach,considering only the factor of geographical in-

?uence,is a special case of USG whereβ=1.

?Random Walk with Restart(denoted by RWR)-this

is a state-of-the-art algorithm recently developed for

collaborative item recommendation based on social net-

works[12].Users’preferences to items are predicted

by Random Walk and Restart over a graph capturing

social graph and user-item matrix.

?User preference/social influence based recommen-

dation(denoted by US)-this method,considering

both user preference and social in?uence from friends,

is a special case of USG,where0<α<1andβ=0.

?User preference/geographical influence based rec-ommendation(denoted by UG)-this approach,con-

sidering both user preference and geographical in?u-

(a)Precision@5-Foursquare (b)Recall@5-Foursquare

(c)Precision@5-Whrrl (d)Recall@5-Whrrl

Figure 3:Tuning parameters

(a)Precision@N -Foursquare (b)Recall@N -Foursquare (c)Precision@N -Whrrl (d)Recall@N -Whrrl

Figure 4:Performance comparison

ence,is a special case of USG ,where 0<β<1and α=0.

5.4Tuning Parameters

As mentioned,two parameters α(for social in?uence fac-tor)and β(for geographical in?uence factor)can be con-trolled to tune the performance of USG and to con?gure it into other recommendation approaches for evaluation.Here we vary them in USG to understand the roles of user pref-erence,social in?uence from friends and geographical in?u-ence from POIs played in achieving the optimal USG per-formance.Similarly,we tune αin US and βin UG to ?ndout their optimal settings as well.Figure 3shows the perfor-mance@5results of USG under di?erent αand βsettings,where the best parameter settings are indicated in the ?g-ures.The optimal settings for US and UG can also be observed in the ?gures,i.e.,dashed line for US and solid line for UG .Those optimal parameter settings are also summarized in Table 1.

Precision@5Recall@5

αβαβFoursquare

US 0.1–0.1–UG –0.2–0.2USG 0.10.10.20.1Whrrl

US 0.1–0.1–UG –0.1–0.1USG

0.10.2

0.10.1

Table 1:Optimal parameter settings

Through this study,we can easily observe that user pref-erence plays a dominate role in contributing to the optimal recommendation,while both social in?uence and geograph-ical in?uence are innegligible.More speci?cally,as shown in Table 1,the factor of user preference contributes at least 70%in making the best recommendation,while both social in?uence and geographical in?uence contribute at least 10%.

5.5Performance Comparison

Next,we compare the e?ectiveness of the recommendation

approaches under evaluation.Figure 4shows the performance@N (N =5,10,20)of all approaches in terms of their best per-formance (i.e.,the performance under the optimal param-eter settings).The experiments used both Foursquare and Whrrl datasets.The precision and recall for them are plot-ted in Figure 4(a)and Figure 4(b),and Figure 4(c)and Figure 4(d),respectively.In these ?gures,USG always ex-hibits the best performance in terms of precision and recall under all values of N s,showing the strength of combines all three factors of user preference,social in?uence and ge-ographical in?uence.Notice that both of our real datasets (i.e.,Foursquare and Whrrl)have low density.According to the empirical study in [12],the reported precision is about 0.17over a pre-processed dataset with 7.8×10?4density of user-item matrix.Thus,the measured low precision over our datasets (which are not preprocessed)is reasonable.Most importantly,USG outperforms the baseline approach U (i.e.,user-based CF)by about 50%percentage of performance improvement in both datasets.

Between the two alternative social in?uence measurement methods (i.e.,S and S rwr )for friend-based CF,we ?nd S to have much better performance than S rwr .Moreover,RWR shows poor performance for POI recommendation in these experiments.This raises a very interesting question of whether Random Walk and Restart technique is suitable for POI recommendations.In a later section,we shall answer this question by analyzing the correlation between (i)the similarity of check-in behavior among friends and (ii)social ties among friends.For the rest of the experiments,we use S as the component of social in?uence from friends in US and USG .

Figure 4also indicates that both social in?uence and ge-ographical in?uence can be utilized to perform POI recom-mendation.As shown,both S and G provide comparable results against U .Notice that,in LBSNs,since the check-

(a)Performance@5-Foursquare(b)Performance@5-Whrrl

Figure5:Fusion of U and

L (a)Performance@5-Foursquare(b)Performance@5-Whrrl

Figure6:Fusion of L and G

(a)Similarity computed via RWR-(b)Number of common friend-(c)Common friend ratio-

(d)Similarity computed via RWR-

Whrrl

(e)Number of common friend-

Whrrl

(f)Common friend ratio-Whrrl

Figure7:Social tie and its in?uence implication

in activities involve physical interaction between users and

POIs,geographical in?uence matters a lot,which is con-

?rmed in the study.As shown,G usually outperforms S and

sometimes even performs better than U,e.g.,when N=20.

Also,UG always show better performance than US.This is

due to the spatial clustering phenomenon appearing in user

check-in activities.Thus,when N is relatively large,there

is very good chance to discover most of user’s check-in ac-

tivities based on social in?uence.

In both Foursquare and Whrrl datasets,we?nd that when

more factors are considered the performance turns out to be

better.For example,US is better than U and S,UG is better

than U and G,and USG shows the best performance.

5.6Study on Item-based CF

In addition to user-based CF,item-based CF can also es-

timate a user’s preference to an item,by exploring the sim-

ilarity between items instead of users[18].In[21],a CF

technique has been proposed to fuse both user-based and

item-based similarity to overcome the data sparsity prob-

lem[10].Thus,a potential idea for POI recommendations is

to employ the item-based CF(denoted by L).Additionally,

geographical in?uence,which models the in?uence among

POIs,may be seemingly similar to“item similarity”in item-

based CF.However,we would like to point out that they

are conceptually di?erent and thus should not be mistaken.

In this section,we explore the idea of further incorporating

L into our framework by examining whether fusing L with U

and G respectively into new approaches denoted by UL and

GL would outperform U and G alone.

Similar to[21],we introduce a weighting parameterλin

UL.Whenλ=1,UL is reduced to U;and whenλ=0,L is

obtained.Similarly,we introduce a weighting parameterγ

in GL.Figure5and Figure6show the performance of UL and

GL on Foursquare and Whrrl datasets under various settings

ofλandγ.Surprisingly,these?gures show that L brings no

advantage at all in enhancing U or L in POI recommenda-

tions,indicating item-based CF is not an e?ective approach

in our application.Our explanation is that,at the current

stage,POIs in LBSNs may not have been visited by su?-

cient many users to make item-based CF work well.In other

words,the computed similarity between two POIs may not

provide a good clue to decide whether a user likes a POI or

not.Since U or G alone show much better performance than

L,we don’t integrate L in our recommendation framework.

5.7Study on Social In?uence

As shown earlier,Random Walk with Restart[12]does not

perform well for POI recommendations.To obtain a com-

prehensive understanding of the reasons behind,we analyze

the correlation between the similarity of user check-in be-

haviors and the user similarity calculated based on Random

(a)Precision@5-Foursquare(b)Recall@5-Foursquare(c)Precision@N-Whrrl(d)Recall@N-Whrrl

(a)Precision@5-Foursquare(b)Recall@5-Foursquare(c)Precision@5-Whrrl(d)Recall@5-Whrrl

Figure9:Performance for cold start users

Walk and Restart.Note that,based on[12],user similarity can be derived from the social graph matrix and user-POI check-in matrix.Figure7(a)and Figure7(d)show the plots on Foursquare and Whrrl datasets under the best RWR set-tings.Both?gures show that similar users do not necessarily have high similarity in their check-in behaviors.For exam-ple,user pairs with similarity larger than0.1usually share nothing in their check-in behavior in both Foursquare and Whrrl datasets.The results indicate that the tastes of a user’s friends may actually vary signi?cantly,which has also be discussed in[16]recently.To further verify this?nding, we also examine the correlation between the similarity of check-in behaviors between two friends and the strength of their social ties.In our tests,the social tie is de?ned in two forms:1)number of common friends(see Figure7(b) and Figure7(e)for experimental results)and common friend ratio(see Figure7(c)and Figure7(f)for experimental re-sults),where common friend ratio is measured by Jaccard coe?cient.For friends who have very strong social tie(i.e., larger number of common friends or larger common friend ratio),we again?nd their check-in behaviors are not neces-sarily similar as shown in the?gures.

From the above observations,we conclude that friends have di?erent tastes.The similarity in friends’check-in behaviors may not necessarily be re?ected in terms of the strength of their social ties.As a matter of fact,in measur-ing the social in?uence between friends,we?nd the factor of check-in behavior to be more important than the fac-tor of social tie.Through our experiments on the S algo-rithm,we?nd the optimal setting forηin Equation(4)to be smaller than0.05in both Foursquare and Whrrl datasets, which indicates the factor of check-in behavior weighs more than the factor of social tie.While RWR treats both factors equally,thus degrading the performance in our Foursquare and Whrrl datasets.

5.8Impact of Data Sparsity

Here,we study how USG deals with the data sparsity prob-lem.In order to produce user-POI check-in matrix with di?erent sparsity,we mark o?x%=10%,30%and50%of user’s check-in activity records from the original check-in datasets for three groups of tests as shown in Figure8. The larger the mark-o?ratio x is,the sparser the user-POI check-in matrix is.As shown,USG always exhibits the best performance@5under all mark-o?ratios.Particularly,when the data is very sparse,e.g.,x%=50%,geographical in?u-ence plays an extremely important role in recommending interesting POIs to users.The reason is that both users and their social friends have relatively small check-in logs.Thus, the similarity weight or social in?uence score derived from such sparse data may be misleading.On the other hand,ge-ographical in?uence,re?ecting a global behavior a?ected by geography,?ts the behaviors of most users in LBSNs.Thus, the approaches incorporating geographical in?uence factor, i.e.,G,UG and USG,show great strengths under various data sparsity scenarios.

5.9Test for Cold Start Users

Finally,we test the performance of POI recommendations for cold start users.Here,we consider those users who have less than5check-in activities in the user-POI check-in ma-trix after removing30%check-ins as cold start users.As shown in Figure9,in all cases we tested,USG always shows the best performance.Note that in POI recommendations for cold start users,user preference is hard to capture as POIs visited by this user are few.Consequently,U shows the worst performance as it only considers user preference. G,which explores the spatial clusters of user check-in activi-ties,is also a?ected.On the other hand,S overcomes the lack of user’s check-ins as social friends may supply many use-ful check-ins,potentially useful for POI recommendations. Thus,in this experiment,we?nd that the recommendation performance of S usually works better than U and G do.No-tice that,we?nd the performance of G to be better than S in extremely sparse scenario in Figure8because in that scenario,social friends’check-in records are very limited as well.Thus,geographical in?uence prevails due to its appli-cability to most of the people.However,it is noteworthy that all three factors are very important for the POI recom-mendations to cold start users,as USG is always the best.

6.CONCLUSIONS AND FUTURE WORK This research attempts to facilitate a POI recommenda-tion service in location-based social networks.Our idea is to incorporate user preference,social in?uence and geograph-ical in?uence in the recommendation.In addition to de-riving user preference by user-based collaborative?ltering and capturing social in?uence from friends,we model the geographical in?uence among POIs by employing power law distribution to uncover the spatial clustering phenomenon in user check-in activities.Furthermore,we propose a uni?ed POI recommendation framework,which fuses user prefer-ence to a POI with social in?uence and geographical in?u-ence.We conduct a comprehensive performance evaluation over two large-scale real datasets collected from Foursquare and Whrrl.Experimental results show that the uni?ed col-laborative recommendation technique is superior to all other recommendation approaches evaluated.Additional?ndings have been uncovered through analysis of the experimental results,including1)geographical in?uence shows a more signi?cant impact on the e?ectiveness of POI recommenda-tions than social in?uence;2)Random Walk and Restart may not be suitable for POI recommendation in LBSNs,be-cause friends exhibit signi?cantly di?erent preferences(i.e., the strength of social ties do not re?ect the similarity of check-in behavior among users in LBSNs;3)Item-base CF is not an e?ective approach in our application due to insuf-?cient number of visitors to many locations at the current state of LBSNs.

The semantic tags of POIs contain very rich information brought in by LBSN users.As for the next step,we plan to incorporate the semantic tags of POIs to further improve the uni?ed POI recommendation framework we proposed in this paper.

7.REFERENCES

[1]G.Adomavicius and A.Tuzhilin.Toward the Next

Generation of Recommender Systems:A Survey of the State-of-the-Art and Possible Extensions.IEEE

TKDE,17(6):734–749,2005.

[2]R.Andersen,C.Borgs,J.T.Chayes,U.Feige,A.D.

Flaxman,A.Kalai,V.S.Mirrokni,and

M.Tennenholtz.Trust-based recommendation

systems:an axiomatic approach.In WWW,pages

199–208,2008.

[3]M.Balabanovic and Y.Shoham.Content-based

collaborative recommendation.CACM,40(3):66–72,

1997.

[4]B.T.Bartell,G.W.Cottrell,and R.K.Belew.

Automatic combination of multiple ranked retrieval

systems.In SIGIR,pages173–181,1994.

[5]D.Billsus and M.J.Pazzani.Learning collaborative

information?lters.In ICML,pages46–54,1998.

[6]C.M.Bishop.Pattern Recognition and Machine

Learning.2006.

[7]X.Cao,G.Cong,and C.S.Jensen.Mining signi?cant

semantic locations from gps data.PVLDB,

3(1):1009–1020,2010.

[8]A.M.Ferman,J.H.Errico,P.van Beek,and M.I.

Sezan.Content-based?ltering and personalization

using structured metadata.In JCDL,page393,2002.

[9]J.L.Herlocker,J.A.Konstan,A.Borchers,and

J.Riedl.An algorithmic framework for performing

collaborative?ltering.In SIGIR,pages230–237,1999.[10]Z.Huang,H.Chen,and D.D.Zeng.Applying

associative retrieval techniques to alleviate the

sparsity problem in collaborative?ltering.ACM

TOIS,22(1):116–142,2004.

[11]M.Jamali and M.Ester.TrustWalker:a random walk

model for combining trust-based and item-based

recommendation.In KDD,pages397–406,2009. [12]I.Konstas,V.Stathopoulos,and J.M.Jose.On social

networks and collaborative recommendation.In

SIGIR,pages195–202,2009.

[13]H.Ma,I.King,and M.R.Lyu.Learning to

recommend with social trust ensemble.In SIGIR,

pages203–210,2009.

[14]H.Ma,M.R.Lyu,and I.King.Learning to

recommend with trust and distrust relationships.In

RecSys,pages189–196,2009.

[15]H.Ma,H.Yang,M.R.Lyu,and I.King.SoRec:

social recommendation using probabilistic matrix

factorization.In CIKM,pages931–940,2008.

[16]H.Ma,D.Zhou,C.Liu,M.R.Lyu,and I.King.

Recommender Systems with Social Regularization.In WSDM,2011.

[17]P.Melville,R.J.Mooney,and R.Nagarajan.

Content-Boosted Collaborative Filtering for Improved Recommendations.In AAAI/IAAI,pages187–192,

2002.

[18]B.M.Sarwar,G.Karypis,J.A.Konstan,and

J.Riedl.Item-based collaborative?ltering

recommendation algorithms.In WWW,pages

285–295,2001.

[19]W.R.Tobler.A Computer Movie Simulating Urban

Growth in the Detroit Region.Economic Geography,

46:234–240,1970.

[20]H.Tong,C.Faloutsos,and J.-Y.Pan.Fast Random

Walk with Restart and Its Applications.In ICDM,

pages613–622,2006.

[21]J.Wang,A.P.de Vries,and M.J.T.Reinders.

Unifying user-based and item-based collaborative

?ltering approaches by similarity fusion.In SIGIR,

pages501–508,2006.

[22]M.Ye,P.Yin,and W.-C.Lee.Location

Recommendation in Location-based Social Networks.

In GIS,pages458–461,2010.

[23]H.Yildirim and M.S.Krishnamoorthy.A random

walk method for alleviating the sparsity problem in

collaborative?ltering.In RecSys,pages131–138,2008.

[24]Q.Yuan,S.Zhao,L.Chen,S.Ding,X.Zhang,and

W.Zheng.Augmenting collaborative recommender by fusing explicit social relationships.In ACM

RecSys-Workshop,pages49–56,2009.

[25]V.W.Zheng,B.Cao,Y.Zheng,X.Xie,and Q.Yang.

Collaborative?ltering meets mobile recommendation:

A user-centered approach.In AAAI,2010.

[26]V.W.Zheng,Y.Zheng,X.Xie,and Q.Yang.

Collaborative location and activity recommendations

with gps history data.In WWW,pages1029–1038,

2010.

[27]Y.Zheng,L.Zhang,X.Xie,and W.-Y.Ma.Mining

interesting locations and travel sequences from gps

trajectories.In WWW,pages791–800,2009.

情态动词表推测用法总结及专项练习

情态动词表推测用法总结及专项练习 1.can / could用于表推测的用法 (1) 从使用句型上看，can 通常只用于否定句或疑问句，一般不用于肯定句，而could 可用于肯定句、否定句和疑问句。两者没有时间上的差别，只是could 比can 更委婉，更不确定。如：It can’t [couldn’t] be true. 那不可能是真的。 What can [could] they be doing? 他们会在干什么呢? We could go there this summer. 今年夏天我们可能要去那儿。注：can 有时也用于肯定句中表示推测，主要用于表示理论上的可能性(即从理论上看是可能的，但实际未必会发生)，或表示“有时”之意。如： Even experienced teachers can make mistakes. 即使是有经验的教师也可能出错。 She can be very unpleasant. 她有时很令人讨厌。 (2) 从时间关系看，对现在或将来情况作推测，后接动词原形；对正在进行的情况作推测，后接be doing 结构；对过去情况作推测，后接动词完成式。如： He could have gone home. 他可能已经回家了。 He can’t [couldn’t] have understood. 他不可能理解了。

Why does he know this? Can [Could] someone have told him about it? 他怎么知道? 会是哪个人告诉他了吗? (3) “could+完成式”除表示对过去的推测外，还有以下重要用法： ①表示过去没有实现的可能性，常译为“本来可以”。如： I could have lent you the money．Why didn’t you ask me? 我本来可以借这笔钱给你的。你为什么不向我提出? ②用来委婉地责备某人过去应该做某事而没有去做，常译为“本来应该”。如： You could have helped him. 你本来应该帮助他的。 ③表示“差点儿就要”。如： I could have died laughing. 我差点儿笑死了。 2. may / might用于表推测的用法表示推测，两者都可用，只是might 比may 语气更不确定，表示的可能性更小。 (1) 在句型使用方面：两者均可用于肯定句和否定句，但用于疑问句时，may通常不用于句首，但可用于疑问句的句中（如特殊疑问句等），而might尽管可以用于疑问句的句首，但不算普通，通常会改用其他句式（如用could等）。如： He may [might] know the answer. 他可能知道答案。

不及物动词归纳

1.只是不及物的： faint,hesitate,lie,occur,pause,rain,remain,sleep,sneeze. 2.常见的及物，不及物的： answer,ask,begin,borrow,choose,climb,dance,eat,enter,fail,f ill,grow,help,hurry,jump,know,leave,marry,meet,obey,pull,re ad,see,sell,touch,wash,watch,win,write 3.及物不及物意义变化的lift.升高beat vi.跳动vt. 敲、打; grow vi.生长vt. 种植play vi.玩耍vt. 打（牌、球）,演奏smell vi.发出（气味）vt. 嗅ring vi.（电话、铃）响vt.打电话speak vi.讲话vt. 说（语言）hang vi. 悬挂vt. 绞死operate vi.动手术vt. 操作 4.意义不变的 start, answer, sing, close, consider, insist, read, learn, prepare, pay, hurt, improve....

live, go, work, listen, look, come, die, belong, fa ll, exist, rise, arrive, sit, sail, hurry, fail, su cceed. agree... 不及物动词 agree, go, work, listen, look, come, die, belong, f all, exist, rise, arrive, sit, sail, hurry, fail, s ucceed、beat、buy, catch, invent, found, like, obs erve, offer, prevent, promise, raise, find, forget, receive, regard, see, say, seat, supply, select, s uppose, show, make, take, tell 6.不及物动词短语 down (stop functioning 坏了，不好使了) That old Jeep had a tendency to break down just w hen I needed it the most. on (become popular 出名) Popular songs seem to catch on in California first and then spread eastward. 3. come back ( return to a place 返回)

lie-lie-lay三个动词的区别

动词lie，lie，lay的区别 1)lie用作规则动词(lie，lied，lied)时意为“说谎”。如： ①He wasn't telling the truth. He lied again/He was lying. 他没讲实话，他又撒谎了。/他在撒谎。 ②She lied to us about her job. 她就她的工作对我们撒了谎。 ③Your watch must be lying. 你的表肯定不准。【注意】表示这一意思时，lie可用作名词。我们一般说tell a lie或tell lies，而不说say/talk/speak a lie(或lies)。再如：Lies cannot cover up facts.(谎言掩盖不了事实。)如要说“我痛恨说谎话”，可以说I hate lying或I hate telling lies，一般不说I hate lie或I hate to lie。hate to lie尽管句法正确，但似乎含有“虽然不愿，但仍不得不说谎”之意。

2)lie用作不规则动词(lie，lay，lain)时，表示“躺”，“(东西)平放”，“位于”等意。如： ①He's still lying in bed. 他还躺在床上。 ②He felt tired, so he went and lay down for a rest. 他感到疲劳，所以去躺下休息了。 ③His books lay open on the desk when I went in. 我进去时，发现他的书平摊在书桌上。 ④Snow lay thick on the fields. 田野里覆盖着厚厚的一层雪。 ⑤Don't leave your things lying about. 别把东西四处乱丢。

英语情态动词的用法大全附解析

英语情态动词的用法大全附解析一、初中英语情态动词 1．—We've got everything ready for the picnic. —Do you mean I __________ bring anything with me? A. can't B. mustn't C. couldn't D. needn't 【答案】 D 【解析】【分析】句意：—我们已经为野餐准备好了一切。—你的意思是我不必带任何东西吗？A. can't 不能；B. mustn't 表示禁止，一定不要；C. couldn't不能，表示过去时态；D. needn't不必。结合句意，故选D。【点评】本题考查情态动词的用法。 2．—How amazing this robot is! —Wow, it has video cameras in its eyes, so it “see” and interact with pe ople. A. may B. can C. must D. should 【答案】 B 【解析】【分析】句意：—这个机器人多么惊人啊！—哇，在它的眼睛里有摄像机，因此它能看见和人打交道。A. may 可以，可能；表示许可，B. can 能；表示能力，C. must 必须；D. should应该；根据it has video cameras in its eyes,可知是有能力看见，故选B。【点评】考查情态动词辨析。熟记情态动词的含义和用法。 3．—Where is George? —He _______ be here just now. His coffee is still warm. A. need B. can't C. must D. shouldn't 【答案】 C 【解析】【分析】句意：——乔治在哪里？——他刚才一定在这里，他的咖啡还是热的。A. need需要； B. can't 不能，不可能（表示推测）； C. must 必须，一定（表示推测）； D. shouldn't不应该；根据His coffee is still warm.可知表示肯定推测一定在这儿；故答案为C。【点评】考查情态动词。掌握情态动词表推测时的意义和用法。 4．— Sorry, I forgot to take money with me. Maybe I can't buy the book you like. — Mum, you ______ worry about it. We can pay by Alipay （支付宝）. A. can't B. needn't C. mustn't D. shouldn't 【答案】 B 【解析】【分析】句意：——对不起，我忘记带钱了。也许我不能买你喜欢的书。——妈妈，你不用担心。我们可以用支付宝支付。A. can't不能，指不允许或否定推测；B. needn't 不需，指没必要；C. mustn't不能，表禁止；D. shouldn't不应该，表建议。根据句意语境，本句是说妈妈不需要担心，故答案为B。【点评】考查情态动词。理解句意并掌握情态动词的意义和用法区别。

enjoy的详细用法

enjoy的详细用法今天给大家带来enjoy的用法，我们一起来学习吧，下面就和大家分享，来欣赏一下吧。 enjoy的用法 1.enjoy表示“享受…的乐趣”，如： He was a guy who enjoyed life to the full. 他是个尽情享受生活乐趣的人。 2.enjoy表示“享有”，如： The average German will enjoy 40 days paid holiday this year. 德国人今年平均将享有40天带薪假期。 3.enjoy doing sth(enjoy后要接动名词，而不接不定式。)如： I enjoy swimming.我喜欢游泳。 4.enjoy onself (自) 得其乐，过得快活，玩得开心，如： I am really enjoying myself at the moment.

我此刻很是自得其乐。 Did you enjoy yourself at the English evening? 这个英语晚会你玩得开心吗? 新万能动词ENJOY 想怎么用就怎么用在写作的过程中，动词往往是最能体现我们的语言表达水平的。当我们遇到名词不会翻译，我们会想尽一切办法来描述，从而达到表达的目的，比如考试时你突然不记得affinity(亲密关系)怎么翻，我们可以简化，尝试说“a close relationship”。但是，当遇到不会翻的动词的时候，我们就会陷入一种使用万能动词的怪圈，那么整篇*就会充斥着六大万能动词do, get, give, make, have 和take，在写作考试中，即便是搭配基本准确，也会因为表达的单一性对我们的分数会有影响。峰回路转，还好有enjoy，这个动词小词的搭配极强，可以起到丰富表达的作用。凡是表达积极意义的，基本上都可以跟enjoy搭配，除原意“享受”之外，还能译成“受到、得到、拥有、具备等”等含义。 1. enjoy + 支持类先看例句：China will enjoy more and more support internationally. 如何理解此处的enjoy support?其实就是表达“受到支持”。所以我们可以把enjoy搭配支持类的名词，形成的常用

英语情态动词用法总结(完整)

英语情态动词用法总结(完整) 一、单项选择情态动词 1．--- Difficulties always go with me! --- Cheer up! If God closes door in front of you, there be a window opened for you. A．must B．would C．could D．can 【答案】A 【解析】【详解】考查情态动词辨析。句意：——困难总是伴随着我！——高兴点! 如果上帝在你面前关上了门，一定有一扇窗户为你打开。A. must必须；B. would将要；C. could能，会；D. can能，会。must表示对现在的状态推测时，意为“一定”，表示可能性很大的推测。符合语境。故选A。【点睛】 1) must用在肯定句中表示较有把握的推测，意为"一定"。 2) must表对现在的状态或现在正发生的事情的推测时， must 后面通常接系动词be 的原形或行为动词的进行式。 3) must 表示对已发生的事情的推测时，must 要接完成式。 4) must表示对过去某时正发生的事情的推测，must 后面要接完成进行式。 5) 否定推测用can't。本句中的。must表示对现在的状态推测时，意为一定，表示可能性很大的推测。符合第2点用法。 2．Paul did a great job in the speech contest. He many times last week. A．need have practised B．might practise C．must have practised D．could practise 【答案】C 【解析】【详解】考查情态动词。句意：保罗在演讲比赛中表现得很好。他上星期一定练习了很多次。must have done是对过去发生的动作最有把握的猜测，意思是“一定”。故C选项正确。 3．He is a bad-tempered fellow, but he ________ be quite charming when he wishes. A．shall B．should C．can D．must 【答案】C 【解析】【详解】考查情态动词辨析。句意：他是个脾气不好的家伙，但当他希望自己有魅力的时候，他可

lie的用法和短语例句(20201027140422)

lie的用法和短语例句 lie有躺下;位于;在于;平放;存在等意思，那么你知道lie的用法吗？下面跟着一起来学习一下，希望对大家的学习有所帮助！ lie的用法： lie的用法1 : lie的基本意思是指人“平躺〔卧〕，也可表示“（使）某物平放”。 lie的用法2 : lie是不及物动词，常与around, down, in, on 等词连用,不能用于被动结构。 lie的用法3 : lie有时可用作系动词，后接形容词（而不接副词）、现在分词或过去分词作表语。 lie的用法4: lie可用于表示“存在”的There...句型中位于there 之后，主谓倒装。 lie的用法5 : lie作“展现”解时，多以物作其主语。 lie的用法6 : lie的基本意思是“说谎”，指为了欺骗别人，通过已有预谋地或故意地歪曲了事实真相而直接生硬地所说的假话，是可数名词。 lie的用法7 :在非正式文体中，可用lying代替lie。 lie的常用短语： lie about 〔around〕（v.+adv.）

lie back (v.+adv.) lie before (v.+prep.) lie beh indl (v.+adv.) lie beh ind2 (v.+prep.) lie bey ondl (v.+adv.) lie bey ond2 (v.+prep.) lie by1 (v.+adv.) lie by2 (v.+prep.) lie dow n (v.+adv.) lie heavy on (v.+adj.+prep.) lie in1 (v.+adv.) lie in2 (v.+prep.) lie off (v.+adv.) lie on 〔upon〕(v.+prep.) lie over (v.+adv.) lie to1 (v.+adv.) lie to2 (v.+prep.) lie的用法例句： 1. The blame for the Charlest on fiasco did not lie with him. 查尔斯顿惨败错不在他。 2. Lie face upwards with a cushi on un der your head. 头枕垫子仰卧。

英语副词用法详解

英语副词用法详解一、单项选择副词 1．This car is environmentally friendly because it uses electricity instead of gas. , it is less expensive. A．Besides B．However C．Instead D．Meanwhile 【答案】A 【解析】【详解】考查副词。A. Besides另外；B. However然而；C. Instead 相反；D. Meanwhile与此同时。句意：这辆汽车是环保的，因为它用电而不用汽油。此外，它更便宜。结合句意可知答案为A。 2．(安徽新余市模拟)Recently stock prices have been falling________because of the influence of the global financial crisis. A．sharply B．absolutely C．narrowly D．universally 【答案】A 【解析】考查副词辨析。句意为：受全球经济危机的影响，最近股票价格急剧下跌。A.锋利地，敏锐地，急剧地；B.绝对地；C狭窄地，勉强地；D.全体地，无一例外地。根据题干中“global financial crisis”可知，此处指股票下跌。答案：A 3．Linda's health is improving ________ day by day, and the doctor has advised her to stay home for a couple of weeks. A．frequently B．gradually C．fluently D．actually 【答案】B 【解析】根据improving和day by day可知，琳达的身体在一天天“逐渐地(gradually)”康复。frequently频繁地；fluently流利地；actually实际上，均不符合语境。答案：B 4．(浙江卷)I’ve been writing this report________for the last two weeks，but it has to be handed in tomorrow. A．finally B．immediately C．occasionally D．certainly 【答案】C

高中英语情态动词的用法详解

高中英语情态动词用法归纳与练习 A: can / could = be able to 1 表示能力两种时态can (could)，其他时态要用be able to的形式 He can speak French. Was/were able to 可以表示成功做成某事的含义* managed to do / succeeded i n doing With our help, he was able to build the house. 2 表示请求或许可Could 比Can 更加客气* may Can I go now? Yes, you can. 3 表示推测或可能性，常用于否定句和疑问句 Can he be at school? Today is Sunday. He can't be at school. Can’t couldn’t/ Can ? Could? have done There is no light in the room. Can she have gone to bed? She can’t have gone to school----- it’s Sunda y. Could have done 可用在肯定句中表示过去存在的可能性* may have done She could have gone out with some friends yesterday. Could have done 可以表示虚拟语气某事可能发生（并没有发生） It was silly to throw the TV out of the window. It could have hit somebody. 批评某人没有完成某事(本可以做某事) You are late again. You could have got up earlier. 4 can 表示客观上的可能性* Anybody can make mistakes. B: may and might 1 表示请求和允许类似用法can could(更加口语化) May I watch TV now? Yes, you may. 2 表示可能性; 猜测might比may更加不确定may/might have done 对过去的揣测* He may be busy these days. May not 可能不；不可能；一定不可 3 表示祝愿Wish you success! May you succeed!* C: must and have to 1 表示必须；一定要可以用have to 代替 must 现在；将来的必须而且是主观看法* have to 有多种时态而且强调客观情况 I really must stop smoking now. I had to leave early because I wasn’t feeling well. Must I come here?

【英语】英语名词用法总结(完整)

【英语】英语名词用法总结(完整) 一、单项选择名词 1．Actively involve yourself in voluntary work and you will see what a(an) ________ it will make to your future life. A．influence B．contribution C．difference D．variety 【答案】C 【解析】【详解】本题考查固定短语。解题步骤：1.确定各项意思：A. influence影响；B. contribution贡献； C. difference不同； D. variety多样化；2.确定答案：固定短语make a difference有影响。句意：积极地参与志愿者的工作，你就会看见它对你未来的生活有多么大的影响。故选C。 2．An advantage of living on the top floor of a high rise building is that you can get a good _________. A．scene B．scenery C．sight D．view 【答案】D 【解析】【详解】考查名词辨析。句意：住在高楼大厦顶层的优点是你能看很好地欣赏风景。A. scene 情景B. scenery景色 C. sight 景象 D. view风景。view“风景”普遍用语,指从某个角度所看到的风景。根据句意，故D选项正确。【点睛】 Scene:1. 某地的景色、风景，指展现在眼前的情景，也可以指scenery的一部分 3．Only when women are enabled to become strong will our families, our economies, and our societies reach their fullest . A．distinction B．innovation C．potential D．appreciation 【答案】C 【解析】【详解】考查名词。A. distinction区别； B. innovation创新； C. potential潜能； D. appreciation 欣赏。句意：只有当妇女能够变得强大时，我们的家庭、经济和社会才能充分发挥其潜力。结合句意可知答案为C。 4．With the help of the English teacher, he has made in his English writing. A．a great progress B．much progress C．little progress D．many progresses

lie作平躺、撒谎时的过去式和过去分词

lie作平躺、撒谎时的过去式和过去分词 lie有平躺，撒谎等意思。你知道lie作平躺、撒谎时的过去式和过去分词是什么吗?下面是为你整理的lie 平躺、撒谎时的过去式和过去分词，希望大家喜欢! lie作平躺、撒谎时的过去式和过去分词平躺时的过去式和过去分词：lie;lied lied 撒谎时的过去式和过去分词：lie; lay lain lie的用法v. 躺下;位于;在于;平放;存在;延伸;说谎 n. 谎言用作动词用作不及物动词 S+～(+A) He lay there, flooded with pain, fear and hatred. 他躺在那儿,心中充满了痛苦、恐惧和憎恨。 As you make your bed, so you must lie. 自作自受。 Lasa lies too far away from Beijing. 拉萨离北京很远。 There lies〔lay〕... There lay a boy.

那里躺着一个男孩。 There lies a vast expense of golden crops before us. 一望无际的金灿灿的作物展现在我们的眼前。 There lies a bright future before us. 我们前途似锦。 S+～+ v -ed The national hero lies buried here. 那位民族英雄就安葬在这里。 His motives lay concealed. 他的动机不明。 S+～+ v -ing She lies sleeping soundly. 她躺在床上,睡得香甜极了。 The goods lie wasting in the warehouse. 货物闲放在仓库里。 lie的词语用法v. lie的基本意思是指人“平躺〔卧〕”,也可表示“(使)某物平放”。 lie是不及物动词,常与around, down, in, on等词连用,不能用于被动结构。 lie有时可用作系动词,后接形容词(而不接副词)、现在分词或过去分词作表语。 lie可用于表示“存在”的There...句型中,位于there之后,主谓

高中情态动词讲解

情态动词专项讲解 1. China is developing its high-speed train technology fast, so it _____be good enough to catch up with the best. A. may B. might C. must D. should 2. Then some other interesting reasons appear, such as “It's so smoggy that I ______find my way to office." A. mustn’t B. won’t C. can’t D. shouldn’t 3. "The world is big, and I want to see it," wrote a teacher in her resignation letter. _____ you quit your job to travel the world A. Must B. Might C. Should D. Would 【答案揭晓】CCD 一、情态动词的定义情态动词是一种本身有一定的词义，但要与动词原形或其被动语态一起使用，给谓语动词增添情态色彩，表示说话人对有关行为或事物的态度和看法，认为其可能.应该或必要等。情态动词后面加动词原形。 1. No driving electric motorbikes in some areas is a rule that we shall obey in Fuzhou.

必须.一定（法律.法规等）动词原形 2. Women who are exposed to second-hand smoke during their pregnancy can be at risk of abortion.能，会，可以(表示有能力或机会) 动词原形二、常考情态动词的关键用法 (一) can 1. －Jim，it is time you went to bed．You need to get up early tomorrow. －It’s not f air，Mary can stay up till ten hut I have to go to bed at eight. 可以(表示允许) it were not for the fact that she can't sing，I would invite her to the party. 能，会，可以(表示有能力) 3. Luckily, iron can be reworked and mistakes don't have to be thrown away 能够，可以(表示某事物的特点) can be really difficult to get along with at times even though he’s a

初中情态动词用法总结

专项：情态动词一考点：情态动词的用法和辨析，情态动词表示推测和可能，由情态动词引导的一般疑问句的回答。二类型：1 只是情态动词：can, could, may, might, must 2 可做情态动词，可做实义动词：need, dare 3 可做情态动词，可做助动词：will, would, shall, should 4 特殊：have to, ought to, used to 三特征：1 有一定的词义，但不能单独作谓语，必须与行为动词和系动词连用构成谓语。 2 无人称和数的变化。（have to 除外） Eg: He has to stay here. 3 后接动词原形。 4 具有助动词作用，可构成否定，疑问或简短回答。四用法： 1. can ①表示能力，“能，会”。Eg : Can you play basketball? ②表示怀疑，猜测，常用于否定句或疑问句。 Eg :Li hua can’t be in the classroom. ③表示请求，允许，多用于口语，译“可以”= may. Eg: you can go now. ④can 开头的疑问句，肯定句，否定句用can或can’t. 2.could①can 的过去式，表示过去的能力。 Eg :I could swim when I was seven years old. ②could 开头的疑问句，肯定和否定回答用could, couldn’t如果could 表示现在的委婉，用can 回答。 Eg: Could I have a drink? Yes, you can. 3.may①表示推测，“可能，也许”，用于肯定句。 Eg: He may come tomorrow. ②表示请求，“许可,可以”。Eg: May I borrow your book? 注：表示请求，许可时，主语为第一人称的一般疑问句，否定回

lie 用法总结

lie—lied—lied—lying （vt．＆．vi．）（撒谎）lie—lay—lain—lying （vi．）（躺下，位于）lay—laid—laid—laying （vt．＆vi．）（平放、产卵）【口诀记忆】

撒谎lie,lied,lied,don't be a liar; 一“赖”到底是说谎（发音都是【lai】）躺lie,lay,lain, lie in bed again; 三个不一样是平躺（原型，过去式，过去分词都不一样）

下蛋 lay,laid,laid,a hen laid an egg; 一“累”到底是下蛋（发音都是【lei】）放置lay,laid,laid laid it in the bag. 下蛋不就是把蛋放置好嘛，所以变法跟下蛋完全一样。

tell a lie (opp) tell the truth lie to sb; lie on one's back How do they lie to each other? The book lay open on the desk.

A bright future lies ahead. He lay on his back. The trouble lies here. Japan lies to the east of China. An oil pipes is being laid between the two cities.

The hunters laid a trap for the tiger. Rainstorms have laid crops. Laying eggs is its full time job. She always lays her books on the shelf.

高中英语语法总结大全之情态动词

高中英语语法总结大全之情态动词情态动词的语法特征 1) 情态动词不能表示正在发生或已经发生的事情,只表示期待或估计某事的发生。２）情态动词除oｕgｈt和have外,后面只能接不带to 的不定式。 3）情态动词没有人称，数的变化，即情态动词第三人称单数不加-s。 4）情态动词没有非谓语形式，即没有不定式，分词，等形式。比较can 和be aｂlｅto １)cancould 表示能力；可能(过去时用ｃould)，只用于现在式和过去式（ｃould)。be ａble ｔｏ可以用于各种时态。 Tｈey will ｂe aｂlｅｔo ｔeｌｌyou tｈe ｎews ｓｏon．他很快就能告诉你消息了。２）只用bｅablｅto a. 位于助动词后。 b.情态动词后。 c．表示过去某时刻动作时。 d. 用于句首表示条件。ｅ. 表示成功地做了某事时,只能用wａs/were able to, 不能用cｏuld。Ｈeｗas ablｅto ｆlee Eurｏpｅbefｏre tｈeｗar broｋe ｏut. =He ｍaｎａgeｄtｏfｌee Euroｐeｂefore thｅwar broke out. 注意:could不表示时态 1)提出委婉的请求，(注意在回答中不可用cｏｕｌd)。 ---CoulｄI haｖe tｈｅtｅlｅｖiｓioｎon? --－Yes,ｙou ｃan. / Nｏ，you can'ｔ. 2）在否定,疑问句中表示推测或怀疑。 Hｅcouldn't be a bａd ｍaｎ. 他不大可能是坏人。

比较ｍay和might 1）表示允许或请求;表示没有把握的推测;mａｙ放在句首,表示祝愿。 May Gｏd bｌｅss yｏu！Ｈｅmiｇht be aｔhome．注意：might 表示推测时,不表示时态。只是可能性比may 小。 2) 成语: may/might as well,后面接不带to 的不定式，意为"不妨"。 Ifｔhａt is thｅcase, we may ａs ｗell try. 典型例题Ｐetｅｒ_＿＿come wｉｔh us tonighｔ，but ｈe iｓn't ｖery sｕrｅyet. A. mｕstB．ｍayC. canD. wｉll 答案Ｂ. 表可能性只能用may.此句意可从后半句推出。比较hａｖe ｔo和mｕst 1)两词都是＇必须＇的意思，ｈave tｏ表示客观的需要，ｍｕｓｔ表示说话人主观上的看法，既主观上的必要。 My ｂroｔher ｗas ｖery ilｌ, ｓo Ｉhａd toｃaｌl tｈe docｔｏr iｎtｈｅmiddlｅof tｈｅniｇhｔ.我弟弟病得很厉害，我只得半夜里把医生请来。(客观上需要做这件事） He said that theyｍust wｏrｋｈａrd．他说他们必须努力工作。(主观上要做这件事) 2)hａve to有人称、数、时态的变化,而must只有一种形式。但must 可用于间接引语中表示过去的必要或义务。 Hｅｈaｄtｏlｏok after ｈｉｓｓisterｙｅsｔerday． 3) 在否定结构中：dｏn＇t hａvｅto表示"不必" mustｎ＇t表示"禁止"， You ｄｏn'ｔhａvｅto tell hiｍabout iｔ．你不一定要把此事告诉他。Ｙｏu mustn＇ｔｔｅlｌhim aboｕt ｉt.你一定不要把这件事告诉他。 must表示推测 1)muｓｔ用在肯定句中表示较有把握的推测，意为"一定＂。

英语副词用法总结(完整)

英语副词用法总结(完整) 一、单项选择副词 1．________ when your ticket has been checked will you be allowed to take the plane. A．Just B．Only C．Still D．Yet 【答案】B 【解析】考查倒装句。根据后面的主语用的是倒装形式will you be ……可知，此句采用的是部分倒装句式。四个选项中only+状语放在句首时使用部分倒装，其余三项没有这种用法，所以选B。 2．Each working day，The Daily is updated on the website every hour，unless indicated. A．otherwise B．therefore C．nevertheless D．moreover 【答案】A 【解析】句意：每个工作日，《日报》每小时都会在网站上更新一次，除非另有注明。otherwise意为“否则，另外”，符合语境。therefore因此；nevertheless然而，不过；moreover而且。 3．（2018·江苏97校联考）More and more people tend to go to E-libraries in that they have access to books when they are unable to get to a library. A．accidentally B．physically C．voluntarily D．appropriately 【答案】B 【解析】考查副词。A. accidentally偶然地；B. physically身体上地；C. voluntarily自动地；D. appropriately适当地。句意：越来越多的人倾向于去电子图书馆，因为他们在无法去图书馆的时候就能接触到书籍。此处指亲身去图书馆，故选B. 4．--Our company had a hard time last year. --But we managed ______. Now everything is progressing steadily. A．moreover B．meanwhile C．somehow D．therefore 【答案】C 【解析】【详解】考查副词词义辨析。句意：--我们公司去年经历了一段艰难的时期。 ---但我们还是设法做到了。现在一切都在稳步发展。A. moreover 而且； B. meanwhile 同时； C. somehow 以某种方式； D. therefore 因此。根据句意和四个选项的内容可知，不管怎样，我们度过了难关，故选C。

Exploiting Geographical Influence for Collaborative Point-of-Interest Recommendation

情态动词表推测用法总结及专项练习

不及物动词归纳

最新英语语法情态动词归纳总结

lie-lie-lay三个动词的区别

英语情态动词的用法大全附解析

enjoy的详细用法

英语情态动词用法总结(完整)

lie的用法和短语例句(20201027140422)

英语副词用法详解

高中英语情态动词的用法详解

【英语】英语名词用法总结(完整)

lie作平躺、撒谎时的过去式和过去分词

高中情态动词讲解

初中情态动词用法总结

lie 用法总结

高中英语语法总结大全之情态动词

英语副词用法总结(完整)

相关文档

最新文档