S Ou ns tl ai ni ne a Bb el hi ta yv i o r a l A n a l y s i s
Point-of-Interest Recommendations via a Supervised Random Walk Algorithm Guandong Xu and Bin Fu, University of Technology, Sydney Yanhui Gu, Nanjing Normal University
A new point-ofinterest (POI) recommendation framework simultaneously incorporates user check-ins, reviews, and POI side information into a tripartite graph to predict personalized recommendations via a sentimentsupervised random walk algorithm. january/february 2016
W
ith the rise of social networking, wireless technology, and mobile computing, a new kind of social Web application called location-based
social networks (LBSNs) has become increasingly popular. LBSNs such as Foursquare and Whrrl provide a handy means for mobile users to establish personal social networks, provide reviews, or raise complaints about various points of interest (POIs), such as restaurants, hotels, stores, and cinemas, creating valuable social data for hotspots that are of common interest to the public. The emergence of LBSNs has also motivated POI recommendation, a new type of recommender aimed at giving mobile users better exploring experiences, helping them locate the services in which they’re most interested. POI-based recommendations have attracted a lot of research attention from academia and industries, and several different approaches have emerged. Of these, the most popular treat mobile user check-in activities as explicit ratings (or feedback) on items. When a user checks in to a POI—say, a restaurant—we usually assume that the user likes this restaurant and that the visit represents a “1” rating.1,2 In such a scenario, any POI recommendations could be considered as typical collaborative filtering (CF) procedures,
following traditional CF-based approaches. But such approaches, which rely solely on user check-in data, suffer from CF’s inherent limitations—the sparsity and cold-start problems that result in unsatisfactory recommendations. To counter this, researchers have realized that the social and spatial nature of LBSNs could help integrate complementary information about users and POIs via more appropriate models or intelligent learning approaches, such as Random Walk with Restart (RWR) over graph or matrix factorization. For example, some studies integrate user social influence (friend networks),3 while others exploit the geographic influence of POIs to improve recommendations.4 Despite these successes and improvements, some open research questions in POI recommendations remain: • Current approaches treat user check-in and comment activities as a “like” rating,
1541-1672/16/$33.00 © 2016 IEEE Published by the IEEE Computer Society
15
Online behaviOr al analySiS
but they don’t differentiate the sentiment polarity in reviews. For example, sometimes a user checks into the system to make a complaint or raise concerns, indicating a negative impression rather than a positive score on a POI. • The multi-aspect relations among users, POIs, and their associated reviews haven’t been comprehensively and adequately utilized in POI recommendations, resulting in single-purpose recommendations rather than anything personalized or multifaceted. • Current graph-based algorithms typically decide the equal probability of walk transition between two nodes based on the out-degree of edges, so POIs attached with positive reviews can’t be favored for recommendation when the random walk model is applied to POI recommendations. To address these issues, we propose a new approach for POI recommendations that leverages the multisource information embedded in LBSNs, such as users’ relations, POI geographical information, user check-ins, and user reviews, via a supervised random walk algorithm. The key goal behind our work is to help users accurately locate POIs with overall positive reviews.
Other Approaches Before we present our work, let’s fi rst consider the approaches currently available. Cf-based approaches
POI recommendations can be thought of as conventional recommendation problems, 5 so let’s start with the most traditional: CF. User-based CF. In CF, a user’s implicit preference for one POI can be 16
calculated by aggregating the activities exhibited by like-minded users.1 The predicted check-in probability cˆ i , j of ui to pj is calculated as cˆ i , j =
∑ u ∈N (u )si,k ⋅ ck, j , ∑ u ∈N (u )si,k k
i
k
(1)
i
where si,k denotes the similarity between ui and uk, N(ui) is the K-nearest neighbors of ui.
Matrix factorization
Graph-based approaches
methods as an extension of CF are predominant in recommender system research because of the latent factor model’s unique strength, but they still suffer from sparsity and cold-start problems. Friend-based CF. Slightly differing from user-based CF, friend-based CF6 relies on the social influence of friends (via social networks) along with K-nearest neighbors to predict the potential check-in probability: cˆ i, j =
∑ u ∈F(u )SIi, k ⋅ ck, j , ∑ u ∈F(u )SIi, k k
i
k
(2)
i
where SIi,k denotes the social influence between two users, which is derived from social networks and the co-occurrence of check-in activities, and F(ui) is the friend set of ui. Geolocation-based
recommendation.
Apparently, geolocation information— the “physical” location of a POI—contributes largely to user decisions, with www.computer.org/intelligent
users usually having a greater preference for choices closer to them rather than those further away. Thus, POI recommendations should properly consider geographical distance in addition to dependence on user preferences.4 One study7 devised a new POI recommendation scoring scheme with the supplement of geolocation information to enhance existing user- or friend-based recommendations.
Differing from the user-POI check-in matrix model, in which one user can be modeled as a vector of check-ins over various POIs, the graph model is another mathematical way to represent or imply relations among nodes in a network graph. In graph models, RWR8 is a useful method for capturing the overall associations between graph nodes, and it has been adopted in the context of POI by mapping users and POIs as nodes in a network. Therefore, the check-in probability is determined by a random walker via an iterative process, PR(t + 1) = (1 − a) M × PR(t) + ap. After the convergence of RWR, the derived probability vector, PR, presents us a score list of overall possibility of nodes, that is, POIs to be visited by designated user ui. Matrix factorization methods as an extension of CF are predominant in recommender system research because of the latent factor model’s unique strength, but they still suffer from sparsity and cold-start problems. A graph-based or random walk model could integrate multi-aspect relations between various nodes—not only useritem but also user-user or item-item relations—into an arbitrary graph structure, upon which graph-based approaches are deployed. Graph-based recommendation would perform more accurately due to its consideration of multi-aspect information, but it Ieee InTeLLIGenT SySTeMS
TopN Recommendation Users
Friendship
User’s subject Sentiment extraction Reviews
Subjects Random walk with restart
POI Supervised Spatial distance
POIs
Intrarelationship extraction
Recommendation Sentiment fusion
Figure 1. Framework of our approach. It consists of three components: the multisubject models of users, points of interest (POIs), and reviews; the sentiment analysis module; and the supervised random walk engine.
creates a higher computing overhead, given a large number of nodes and complicated interrelations among nodes. Sentiment Analysis
Sentiment information 9 and rich semantic information10 are two important factors in analyzing user reviews. Through empirical studies conducted on a Foursquare dataset,10 research has shown that user reviews contain rich sentiment information. Leveraging this information for better recommendations motivates our own work.
Our Proposed Approach Now that we’ve reviewed related approaches, we present our own approach. Framework and Models
Figure 1 shows the framework for our approach, which consists of three components: the multisubject models of users, POIs, and reviews; the sentiment analysis module; and the supervised random walk engine. january/february 2016
User multifaceted subjects. Because us-
ers often show interest in particular subjects, it’s preferable to recommend POIs about these subjects rather than making a general recommendation. Therefore, we need to determine the subject spectrum derived from reviews, the subjects each POI belongs to, and the subjects each user is interested in. To this end, we process and analyze review data to capture subject information in the following steps: 1. Preprocess all the review data using standard natural language processing tools. After removing stop words, punctuation marks, and numbers, we extract a set of terms for each review to represent it. 2. We predefine a set of subjects S = {s1, s2, …, sk} by analyzing the reviews—for example, three typical subjects are food, relax, and shopping in our study dataset. We also determine a set of seed terms Si = {t1, t2 , , t|si | } for each subject www.computer.org/intelligent
Si. Terms such as coffee, pizza, and restaurant are related to the subject food, for example. Hence, for each review rij, which is given to POI pj by user ui, we define a subject’s frequency as the times of rij’s corresponding terms appeared in or semantically related to the subject. 3. The distribution of user ui’s interests across all subjects is represented as a k-length vector siu . u Element sij indicates subject sj’s frequency by aggregating its frequency over all ui’s related reviews. Similarly, for each POI pj, we can also represent its subject distribution as a k-length vecp p tor s j , in which s jk is subject sk’s frequency aggregated over all reviews related to pj. We thus can capture each user and POI’s subject distribution after normalizing corresponding vectors. This information can be used in our tripartite model to represent the user-subject and POI-subject 17
Online behaviOr al analySiS
relations. These subjects are formed via manual selection of seed terms. Of course, automatic subject forming through content analysis would improve user preference fi nding and recommendation, but we leave this to our future work. Sentiment orientation. As discussed earlier, user check-in activities in LBSNs only represent visits to specific POIs, without indicating like or dislike preferences. Sometimes, users enter LBSNs to make complaints, raise concerns, or criticize the service they encountered, rather than to give a positive appraisal. Thus, it’s crucial for recommender systems to differentiate sentiment orientations in reviews and, in turn, use them for favoring POIs with positive polarity. This mainly motivates our proposed approach—incorporating sentiment analysis into our recommendation framework. The common assumption of sentiment orientation calculation is that one phrase has a positive sentiment polarity when it has strong associations with positive terms, and vice versa. To do this, we follow an approach used elsewhere.9 The key steps are to calculate the point-wise mutual information (PMI) between terms carrying sentiment information from reviews and predefined sentiment words, to identify sentiment orientations. Specifically, we adopt PMI-IR, which uses PMI and information retrieval (IR) to compare the similarity of review terms to positive reference words (excellent, cool, and so on) against its similarity to negative reference words (poor, bad, and so on). The numerical sentiment score is assigned by measuring the difference between the mutual information with the positive words and that with the negative words. Negation is a common issue in text processing, and there are different 18
techniques for handling it, especially in our experiments—for example, we adopted a commonly used technique that involves adding NOT to every word between negation and following punctuation. The PMI-IR between two words, w 1 and w 2 , is defi ned as follows: f pmi (w1, w2 ) = log2
p(w1, w2 ) , p(w1) ⋅ p(w2 )
(3)
the common assumption of sentiment orientation calculation is that one phrase has a positive sentiment polarity when it has strong associations with positive terms, and vice versa. where, p(wi) is the probability that wi occurs in corpus, and p(w 1 , w 2) is the probability that w 1 and w 2 cooccur. Because probability is a measure of the expectation that an event will occur or a statement is true, we have p(wi ) = f (wi ) / m, where f(wi) is the frequency that (wi) occurs in the whole corpus, and m is the size of corpus. We calculate the sentiment orientation (SO) of a phrase as follows: SO(phrase) = fpmi (phrase, excellent) − fpmi (phrase, poor). (4) By setting a cut-off value, we can derive the sentiment polarity for review rij by applying Equation 4. Therefore, given a user ui, we can www.computer.org/intelligent
split the whole set of POIs into two subsets: PP = (Pp 1 , Pp2 ,…, Ppm), which ui has given positive comments, and NP = (Pn 1 , Pn2 ,…, Pnn) , which ui has given negative comments. Reasonably, a good recommendation function should always favor POIs with positive comments over those with negative comments. Tripartite graph model. To model the triadic relations between users, POIs, and subjects, we use a tripartite graph, G = (V, E), V = U ∪ P ∪ S, E = {ei,j}, i, j ∈ U, P, S, where U, P, and S are sets of users, POIs, and subjects. Their interrelations can be derived by projecting the triadic relationships onto three matrices—that is, AUP, AUS, APS, where AUP denotes the matrix of user check-ins on various POIs alongside a complementary sentiment orientation matrix derived from reviews. AUS and APS are the matrices reflecting the users’ subject distribution and POIs’ subject characteristics, respectively. In addition to user check-in and subject category information, we also take user social connection and POI geographical information into this model to form a unified framework: AUU T A = AUS T AUP
AUS ASS APS
AUP T APS , APP
(5)
where AUU denotes the adjacency matrix of user friend connections, AUU = {f (ui, uj), ui, uj ∈ U}, f(ui, uj) = 1, if ui is a friend of uj, otherwise f(ui, uj) = 0; A SS denotes the semantic similarity matrix between various subject categories, which is derived from corpus-based semantic similarity via Wikipedia as a corpus, A SS = {sem_ sim(sk, sl), sk, sl ∈ S}; and A PP = {g(pm , pn), pm , pn ∈ P}, g(pm , pn) is a function of distance between pm and pn , g(pm , pn) = dist(pm , pn)/max (dist(pi, pj)), pi, pj ∈ P. Ieee InTeLLIGenT SySTeMS
As discussed earlier, user checkins and reviews can both contribute to determining the user’s preference for POIs. Therefore, we use a parameterized function to learn every element of AUP instead of assigning a fixed value as other matrices. Specifically, we use a vector of features y ij to depict user i’s preference for POI j, thus for a given edge ∈ AUP, so AUP (i, j) is defined by AUP(i, j) = fw(y ij).
1 1+ e
− w ∗ψ i , j
(7)
.
Because all entries in A are in different ranges, we perform a normalization process. Thus, matrix A is obtained by normalizing each column of A: A D−1 A D−1 A D−1 UU U US S UP P T −1 T A = AUS DU ASS DS−1 APS DP−1 . (8) T −1 AUP DU APS DS−1 APP DP−1
Here, DU, DS , and DP are diagonal matrices. The ith entry in the diagonal of DU is the sum of column i in matrix A, that is, for any user i, the DU(i, i) is computed as follows: DU−1(i , i) = 1
∑ M∈{U ,T ,P} ∑
The constructed tripartite graph captures multirelations among users, POIs, and subjects. Our goal is then to determine the most preferred POIs given a specific user ui with a preset subject sj and a checked-in POI pk. To fulfill this aim, RWR is performed on the tripartite graph with a restart at user ui, subject sj, and POI pk. After the iterative process converges, the POIs with the top-N scores are considered to be the
(6)
Here, fw(y ij) represents the node transition probability of a walker in the graph, which is parameterized by w, and w is to be learned during the model training stage to minimize the loss function in optimization. Especially in this article, we choose the sigmoid function for f: fw (ψ i , j ) =
Supervised random Walk with Sentiment
M k =1AUM (k, i)
,
(9)
where DS−1 and DP−1 are defi ned similarly. Thus, each column of A will be normalized with a sum of 1. january/february 2016
ideally, the POis recommended by our
This algorithm repeats and converges to a stable state, indicating overall preference scores on various POIs for the given ui, tj, and pl.
approach should have a big gain in positive
Optimization framework
reviews, in addition to their relevance to the user’s specific subject and current position. recommendations satisfying the selection criterion of ui, tj, and pk. More formally, the RWR is executed according to the following equation: PU P S P P
t +1
t
1(u = ui ) PU = (1 − α )A PS + α 1(s = s j ) , P 1(p = p ) P l
(10) where • a is a restart probability, controlling the walker jumping back to its initial state from the current state randomly; • ui denotes the target user, sj represents the chosen user subjects (multifaceted), and pl represents ui’s current www.computer.org/intelligent
location, indicating that ui prefers POIs around pl, and (1 (u = ui), 1 (s = sj), 1 (p = pl))T specifies the initial start state being from ui, sj, and pl; • A is the stochastic state transition probability supervising the walker moving to the next state (it’s normally determined by the graph’s link structure as a constant, but here it’s learned via a supervised algorithm based on the sentiment factor derived from user reviews); and • P = (PUT , PST , PPT )T is a vector of visiting probability of all nodes, and PP presents the recommendation scores for all the POIs we recommend.
To obtain reliable recommendation scores for POIs using Equation 10, the transition matrix A plays an important role that guides the walker to reach more preferable targets. Thus, the design of A is actually a key issue in applying RWR. Ideally, the POIs recommended by our approach should have a big gain in positive reviews, in addition to their relevance to the user’s specific subject and current position. In the following, we show how we integrate sentiment factors into the random walk process. Given a user ui and POIs for supervised POI recommendations, suppose ui has positively commented on a set of POIs, PP = (Pp 1 , Pp2 ,…, Ppm ), and negatively commented on a set of POIs, NP = (Pn 1 , Pn2 ,…, Pnn ). As mentioned earlier, the RWR following Equation 10 should always rank the POIs with positive labels higher than those with negative labels—that is, PP(u) > PP(v), u ∈ PP, v ∈ NP. Intuitively, we can adopt AUC (area under the ROC curve), a commonly used 19
Online behaviOr al analySiS
metric in machine learning to measure the ranking function quality. More precisely, AUC is defined by AUC =
∑ u∈PP ∑ v∈NP I(PP (u) − PP (v)) , PP ⋅ NP
(11)
i has to be differentiated with respect to all model parameters. More specifically, the derivative of Fi for w are ∂Fi (w) ∂w = c⋅
∑
u∈PPk ∧ v ∈NPk
where I(x) is a unity function, that is, 1 if x > 0; 0 otherwise. Now our goal is to find the best transition matrix A to maximize the AUC. As in Equation 6, matrix AUP is actually determined by parameters w, thus the problem of finding the optimal matrix A is transformed into how to learn the optimal w. For each element (i, j) in AUP, it’s calculated by A(i, j) = fw(yij). Function fw(yij) parameterized by w takes the edge feature vector yij as input and computes the corresponding edge strength A(i, j), which indicates the state transition probability. With edges parameterized by w, the overall optimization task with respect to the ranking AUC and the polarity on POIs is defined by minimizing the averaged (1 − AUC),
∂S(∆ vu ) ∂PP (v) ∂PP (u) (14) , − ∂w ∂∆ vu ∂w
where ∆vu = PP(v) − PP(u), and c is a constant, c = 1/(|PPk| ⋅ |NPk|). According to Equation 14, it’s easy to compute ∂S(∆vu)/ ∂∆vu , thus the remaining
For systems that return a ranked sequence of results, it’s desirable to also consider the order in which the returned results are presented.
min F(w) = w
1 K
K
∑
∑ u∈PP ∑ v∈NP I(PP (v) − PP (u)) , k
k
PPk ⋅ NPk
k =1
(12)
where F(w) = 1 − AUC and K is the user number. Now our goal is to solve the optimization problem of Equation 12. Obviously, minimizing Equation 12 is intractable. Instead, we use the stochastic gradient descent approach to minimize the above loss function. F(w) isn’t a differentiable function because I(x) is a discrete function, so we use an s-shaped logistic function S as a differentiable approximation to replace I(x): S(x) =
1 1 + e− x
.
(13)
To update the gradient descent, the AUC loss Fi with respect to any user 20
question is how to compute ∂Pp(v)/∂w and ∂Pp(u)/∂w; due to space limitations, we omit the derivation part here. Then, F(w) is optimized via a stochastic gradient descent approach. Multifaceted POI recommendations
With the transition matrix A, we can use Equation 10 to calculate the recommendation scores for POIs with respect to specific user ui and subject sj. The steady state of P = PUT , PST , PPT after the convergence of random walk indicates the probability that the nodes will be visited. Specifically, sorting the derived PP in a descending order and taking the top-N POIs will result in the recommendations we requested. www.computer.org/intelligent
Experimental Evaluation To evaluate our approach performance, we conducted experiments. Here, we present the comparison results and a case study. Dataset and Preprocessing
Many LBSN-based datasets are available on the Internet from sites such as Foursquare, Gowalla, and Whrrle. For our work, we chose Foursquare10 because it includes textual reviews. This data was crawled from September 2010 to late January 2011, resulting in a total collection of 22,387,930 checkins. We removed duplicated check-ins to obtain 22,272,103 unique check-ins and 225,061 users. Because the check-in data is extremely sparse (around 1.36 × 10–4), we conducted data preprocessing to ensure dataset quality in terms of integrity, sparsity, and scalability. We empirically adjusted the dataset size and sparsity by setting various frequency thresholds and removing the corresponding POIs with checkin frequency less than the threshold. Then, we formed the experiment dataset, which contains 225,061 users and 801,795 check-ins. To find the topic being talked about, we counted the term frequency of all the textual data in the check-ins. We also evaluated the semantic relatedness of each term pair—for example, the terms “mocha” and “latte” are considered to be the same and are categorized to “coffee.” To determine the subject distribution of all users and POIs, we analyzed the review data using the process introduced earlier. After this processing, we extracted three subjects and their occurrence frequencies from the dataset: food (110,245), relax (98,560), and tour (73,851), respectively. Based on the extracted subjects, we further marked each review of check-in data with appropriate subject information, for multifaceted recomIeee InTeLLIGenT SySTeMS
0.9
SSR
0.7
0.8
0.6
0.7
0.5
0.6 Hit ratio
Precision
0.8
0.4 0.3 0.2
0.5 0.4
0.1
0.3
0.0
0.2
–1.0 –0.8 –0.6 –0.4 –0.2
0.0
0.2
0.4
0.6
0.8
1.0
SSR
–1.0 –0.8 –0.6 –0.4 –0.2
Sentiment coefficient
0.0
0.2
0.4
0.6
0.8
1.0
Sentiment coefficient
Figure 2. The effect of sentiment coefficient selection. The optimal result is achieved when the cut-off value is set to 0.2, possibly implying that a slightly higher cut-off value helps better differentiate the sentiment orientations.
0.9
0.9 0.8
0.8 0.7
0.6
Hit ratio
Precision
0.7
CF USG RWR SSR
0.5
0.6 0.5
0.4
0.4
0.3
0.3
0.2
@5
@10
@15
@20
CF USG RWR SSR
0.2
@5
@10
@15
@20
TopN
TopN
Figure 3. Precision and hit ratio of each strategy. Our proposed sentiment supervised random walk (SSR) performs the best, followed by user social and POI geographical influence-based recommendation (USG),7 random walk with restart (RWR), 8 and collaborative filtering (CF).
mendations. For social information, we utilized retweeting behaviors among users to derive their connections. Evaluation Metrics and Baselines
For the recommendation evaluation, we randomly divided the whole dataset into two parts by 70 percent (training set) and 30 percent (test set), adopting precision and hit ratio january/february 2016
as evaluation metrics. In precision evaluation, for each given user from a test set, we determined the topN POIs based on the calculated PP values. Then, we calculated the ratio of POIs in ground truth (that is, those who really visited POIs) to the recommendation size (N) as the precision, so precision = |Recom ∩ GroundTruth|/|N|. In hit-ratio www.computer.org/intelligent
evaluation, for each user check-in, we used the leave-one-out strategy, with the user and subjects serving as input for multifaceted recommendations and leaving the visited POI as the ground truth. We determined Hiti = 1 given the existence of real POIs in the recommendation set, otherwise Hiti = 0. By averaging the sum of hits in the test set, we could calculate the hit ratio 21
Online Behavior al Analysis Table 1. Toy example of multifaceted recommendation.* Categories ID
Distance (m)
Textual comments
1
17.2
Nice view, fresh air, nice place for hiking.
2
27.2
Indian pancake tastes sweet and nice.
√
3
57.2
Great Chinese restaurant, but it’s expensive.
√
4
62.1
Had a rest at this small gas station for next trip.
5
68.5
Cheap price, nice service, delicious dishes, Chinese town.
6
71.2
The one thing that makes me happy right now is iced coffee.
7
129.2
Spicy! Which is more spicy, Chinese or Indian food?
8
200.1
Fantastic, great actor.
9
210.2
Why is Chinese food salty and oily? Cheap but not healthy!
10
272.4
The music. The wifi. The life.
Food
Relax
Strategies Tour
CF
USG
RWR
SSR
√ 1
1
√ √
1 √
√ √ √
3 3
2
2
2
1
3
3
2 √
*√ in categories denotes the comments in which the category belongs; the numbers in strategies represent the recommendation order by that strategy. SSR = sentiment supervised random walk, USG = user social and POI geographical influence-based recommendation, RWR = random walk with restart, CF = collaborative filtering.
and investigated the impact of various aspects of recommendations.
The Authors Guandong Xu, corresponding author, is a senior lecturer and program leader of social
and text analytics in the Advanced Analytics Institute at the University of Technology, Sydney. His research interests cover Web mining and Web search, data mining and machine learning, recommender systems, social network analysis, and social media mining. Xu received a PhD in computer science from Victoria University, Australia. He’s the AEIC of World Wide Web Journal and founder of the International Conference on Behavioral, Economic, and Socio-Cultural Computing. Xu is a member of IEEE and ACM. Contact him at
[email protected].
Bin Fu is a PhD student in computer science in the Advanced Analytics Institute at the University of Technology, Sydney, Australia. His research interests include machine learning and data mining, multilabel learning, and recommendation systems. Fu received an MSc in computer science from Beijing Jiaotong University. Contact him at Bin.Fu@ student.uts.edu.au. Yanhui Gu, corresponding author, is a lecturer at Nanjing Normal University. His research
interests include natural language processing, data mining, and information retrieval. Gu received a PhD in computer science and technology from the University of Tokyo. Contact him at
[email protected].
= (1 / N ) ∑ iN−1 Hiti . From the definitions,
we can see that the bigger the precision and hit ratio, the better the algorithm performs. Three baselines are used for comparison (see Figure 3). For systems that return a ranked sequence of results, it’s desirable to also consider the order in which the returned results are presented. Here, we judiciously propose hitratio metrics that can evaluate the 22
recommender system’s sequence performance, but other metrics such as MAP (mean average procession) and NDCG (normalized discounted cumulative gain) can also do this. Due to space limitations, we omit MAP and NDCG from this discussion.
Results and Discussions We compared our approach’s evaluation results to state-of-the-art techniques www.computer.org/intelligent
Effect of Sentiment Polarity
The sentiment of reviews plays an important role in guiding the random walker and determining the recommending scores of POIs. Although the sentiment value is determined through Equation 4, the polarity of sentiment orientation is actually derived by setting various cut-off values to separate positive and negative polarity. In the experimental study, we also wanted to examine the impact of these values on recommendations. Figure 2 shows our results in terms of precision and hit ratio. From the figure, we can see that the optimal result is achieved when the cut-off value is set to 0.2, possibly implying that a slightly higher cut-off value helps better differentiate the sentiment orientations. Recommendation Performance Comparisons
Figure 3 shows effectiveness comparisons with respect to each strategy. From the figure, we can see that our IEEE INTELLIGENT SYSTEMS
proposed sentiment supervised random walk (SSR) performs the best, followed by user social and POI g eographical influence-based recommendation (USG), RWR, and CF in terms of precision and hit ratio. Compared to its three counterparts, SSR can achieve up to 6.4, 16.0, and 24.8 percent improvement on precision, and 8.4, 11.9, and 21.1 percent improvement on the hit ratio, respectively. Therefore, we can conclude that by taking sentiment information into consideration, our approach can recommend more preferred POIs. For multifaceted recommendation, we randomly selected a user with 1,012 reviews in the dataset. We deliberately filtered out this user’s reviews about the subject of “food,” such as “cake, hot, latte,” as the user’s subject for recommendation and obtained a hit ratio of around 85.3 percent in the test set. If we make recommendations based on CF, USG, and RWR approaches without the chosen subject, the obtained hit-ratio results lower to 34.3, 61.5, and 50.2 percent, respectively. Example of Multifaceted Recommendation
We found more than 80 percent of textual information in reviews is related to food. Therefore, we conducted evaluations with the intent category of “food” to evaluate the effectiveness of multifaceted recommendation. A toy example illustrates how the different strategies impact the order of top-3 recommendation results generated. We had human experts evaluate the output to see whether the results reflect real user intention. When a user wanted to find “food,” we listed the recommended POIs with reviews and geographic distance to the user’s current location. january/february 2016
From Table 1, you can see that different algorithms generate recommendation results differently, given the input “food” and the user’s current geolocation. CF strategy only takes the pairwise information into account, so the recommended result “8” seems unrelated to “food.” USG incorporates social information into the f ramework, but pairwise social information sometimes doesn’t reflect a real user’s intent. RWR can retrieve more precise results, but it can’t reflect user sentiment (the recommended “3” is about a good, but expensive, Chinese restaurant). In contrast, our proposed strategy achieves the most precise result, which we attribute to the joint consideration of user intent and sentiment information conveyed by each POI. For example, POIs “3” and “5” are both related to “Chinese food,” but “5” reflects the positive information that’s more suitable for recommendation.
B
y incorporating sentiment analysis, we can differentiate the polarity of user reviews on POIs and supervise the random walk over a multirelational graph of users, POIs, and reviews. In the future, we’ll study the automatic formation of subjects via content analysis, evaluate our approach on other datasets, and investigate other aspects of our framework as well as develop new recommendation algorithms.
3. M. Jiang et al., “Social Recommendation with Cross-Domain Transferable Knowledge,” IEEE Trans. Knowledge and Data Eng., vol. 27, no. 11, 2015, pp. 3084–3097. 4. C. Cheng et al., “Fused Matrix Factorization with Geographical and Social Influence in Location-Based Social Networks,” Proc. AAAI, 2012, pp. 17–23. 5. N. Zheng, X. Jin, and L. Li, “CrossRegion Collaborative Filtering for New Point-of-Interest Recommendation,” Proc. World Wide Web (WWW), 2013, pp. 45–46. 6. X. Yang et al., “On Top-k Recommendation Using Social Networks,” Proc. ACM Conf. Recommender Systems, 2012, pp. 67–74. 7. M. Ye et al., “Exploiting Geographical Influence for Collaborative Point-ofInterest Recommendation,” Proc. Special Interest Group Information Retrieval (SIGIR), 2011, pp. 325–334. 8. M. Ye et al., “On the Semantic Annotation of Places in LocationBased Social Networks,” Proc. Knowledge and Data Discovery (KDD), 2011, pp. 520–528. 9. P.D. Turney, “Thumbs Up or Thumbs Down: Semantic Orientation Applied to Unsupervised Classification of Reviews,” Proc. Assoc. Computational Linguists (ACL), 2002, pp. 417–424. 10. Z. Cheng et al., “Exploring Millions of Footprints in Location Sharing Services,” Proc. Int’l Conf. Weblogs and Social Media (ICWSM), 2011, pp. 81–88.
References 1. D. Lian and X. Xie, “Learning Location Naming from User Checkin Histories,” Proc. GIS, 2011, pp. 112–121. 2. D. Xu et al., “Graph-Based Residence Location Inference for Social Media Users,” IEEE MultiMedia, vol. 21, no. 4, 2014, pp. 76–83. www.computer.org/intelligent
Selected CS articles and columns are also available for free at http://ComputingNow.computer.org. 23
