A Neighbourhood-based Location- and Time-aware Recommender System Len Feremans1 , Robin Verachtert2 and Bart Goethals1 1 University of Antwerp, Belgium 2 Froomle N.V., Belgium Abstract We address the problem of location- and time-aware recommender systems where users with dynamically changing locations are interested in trending and volatile items. Unlike existing work, we do not assume a known static location of each user and derive user-locational preferences from their long-term history of implicit feedback. We propose a recommendation model that accounts for spatial, temporal, popularity and social influences, thereby assuming items tagged with a location, i.e. geotag, city or country. Key ingredients of our online method include: (1) deriving location preferences from the history, (2) learning relevant nearby locations, (3) accounting for recency and popularity jointly, and (4) combining location- and time-aware recommendations with collaborative filtering. Supported by realistic offline and online experiments on a large dataset collected from a popular newspaper, and public datasets, we find that the proposed recommender outperforms content-based and time-aware collaborative filtering approaches. Keywords Context-aware recommender systems, Location-based news recommendations, Collaborative filtering 1. Introduction sports event or a celebrity tweet. Finally, we have to consider the relationship between locations, i.e. in some Every day, users consume items tagged with locations, applications physical distances are more important (i.e. e.g. local news articles from a particular city or Twitter point-of-interest recommendations in Facebook places), tags trending in a specific region. Recommendations while in other applications (i.e. Twitter tags) location sim- are essential to tackle information overload and filter ilarity is higher when many users consecutively prefer relevant items from a huge set of available articles. In both locations [2]. this context, the first law of geography posed by Tobler In this work, we investigate the problem of location- [25] is crucial: “everything is related to everything else, and time-aware recommender systems (LTARS) and rank- but near things are more related than distant things”. ing highly volatile geotagged items by considering both A common strategy for context-aware recommenda- spatio-temporal and user activity trends. We study fac- tions [1] is pre-filtering, i.e. we collect the location of tors correlated with item relevance: (1) geographic fac- each user and rank geotagged items nearby. However, tors, such as the geodesic distance between the geotag of this strategy is problematic. Firstly, many users have an item and the inferred regional preferences of a user, multiple and dynamic regions of preference, i.e. they and (2) user-item preference, i.e. the item-neighbourhood might be interested in items near their home, work or based relevancy, and (3) the recency and popularity of recent vacation stay. Secondly, even after filtering on a an item. This problem has been studied in the context preferred location, there are some biases resulting from of time-aware recommendations [8, 3], location-aware population density, i.e. items geotagged with a big city recommendations [5, 21, 24, 7, 19], context-aware rec- will likely be more numerous and popular, which is likely ommendations [1, 9, 16] and location-based social net- not relevant for all users near that city (vice versa, rural works [2, 28, 12]. A key difference with closely related locations might lack recent item interactions). Addition- research by Pálovics et al. [21] is that we assume a user is ally, we find that the intrinsic popularity bias is detrimen- interested in multiple locations and these locational pref- tal to inferring regional preferences, i.e. an item might erences are unknown and dynamic. A second key differ- be relevant to many users regardless of the associated ence we consider is the volatility of items, which is crucial geotag such as a news item related to an international in certain domains such as news recommendations [15] ORSUM@ACM RecSys 2022: 5th Workshop on Online Recommender and often less in other domains such as point-of-interest Systems and User Modeling, jointly with the 16th ACM Conference on recommendations [2]. A third difference is that we ac- Recommender Systems, September 23rd, 2022, Seattle, WA, USA count for naturally occurring biases in the data such as Envelope-Open len.feremans@uantwerpen.be (L. Feremans); an imbalance in the popularity and location distributions robin.verachtert@froomle.com (R. Verachtert); that hinder recommendations based on (context-aware) bart.goethals@uantwerpen.be (B. Goethals) © 2022 Copyright for this paper by its authors. Use permitted under Creative Commons License collaborative filtering [1]. CEUR Workshop Proceedings Attribution 4.0 International (CC BY 4.0). CEUR Workshop Proceedings (CEUR-WS.org) http://ceur-ws.org ISSN 1613-0073 We propose an LTARS that is orthogonal to existing location-aware and context-aware recommender systems or hybrids thereof [15, 22]. Das et al. propose time-aware that assume known contextual features or static loca- news recommendations that combine item-based collab- tional preferences [21, 1, 24] and make the following key orative filtering, pre-filtering news items on recency, and contributions: age-based discounting to account for the bias towards recent items [8]. Similarly, we consider the short lifetime • We propose novel techniques to (1) extract loca- of items, online scalability issues and realistic offline eval- tion preferences from users based on their fre- uation protocols [3]. A key difference is that we account quency in the long-term history of geotagged for location preferences and tackle cold-start item recom- items; (2) identify and rank relevant neighbour mendations. We compare with time-aware collaborative locations based on collaborative filtering or geo- filtering in Section 4. graphical distance; (3) filter and rank items jointly In location-based social networks point-of-interests on recency and popularity; (4) combine location- items are tagged geographically by users interacting on and time-aware recommendations with collabo- a social network such as Facebook places or Foursquare rative filtering. [28, 2, 12]. Related to our work, Ye et al. propose a hy- • The proposed method is straightforward to im- brid recommendation system that models user-item pref- plement and publicly available1 . It is also highly erences using a combination of collaborative filtering, efficient supporting frequent or online updates social influence, and geographic modelling where loca- and cold-start item recommendations. tion preferences are proportional to the inverse squared • Motivated by the recent criticism of unrealistic geographic distance [28, 18]. However, their approach offline evaluation protocols [3, 13, 23], we adopt is specific to point-of-interest applications where items an evaluation protocol based on a sliding win- such as restaurants are rated after users physically check dow protocol and create subsets of interactions for in at a specific address with longitude and latitude co- training and testing during consecutive periods ordinates. Location-based social network recommender where we filter candidate items on publication systems typically also model social influence, assume date (or first interaction time). explicit feedback and a long life-time of items. • We find that the proposed method and hybrids thereof outperform popularity, content-based and (time-ware) collaborative filtering-based recom- 3. A neighbourhood-based mender systems on offline and online experi- location- and time-aware ments for regional new recommendations. recommender system 2. Related work The proposed method is a combination of different steps and components that take different signals into account, In most related work on location-aware recommendations, i.e. spatial, temporal, popularity and social influence, as general news [24] or Twitter tags [21] are recommended. shown in Figure 1. In contrast to [24] and extended work [20, 4] we assume user locational preferences are non-stationary. Pálovics 3.1. Task definition et al. propose an online model to recommend volatile items, thereby assuming non-stationary locational pref- Let 𝑈 = {𝑢1 , 𝑢2 , … 𝑢𝑛 } be the set of users and 𝐼 = erences [21]. Our work is complementary by inferring {𝑖1 , 𝑖2 , … , 𝑖𝑚 } the set of items. We consider implicit feed- regional preferences for each user. Pálovics et al. also back where a user interacts with a certain item at a propose a hierarchically organized geolocation structure, certain timestamp, i.e. 𝒟 = {⟨𝑢, 𝑖, 𝑡⟩ | 𝑢 ∈ 𝑈 ∧ 𝑖 ∈ 𝐼 } while we propose an alternative neighbourhood-based and denote the user history using 𝐼𝑢,𝑡 , i.e. all interacted method. We argue that focusing on neighbouring loca- items up to timestamp 𝑡. Each item has one or more lo- tions is essential in applications with a high cardinality cations or geotags 𝐿𝑖 where 𝐿𝑖 ⊆ 𝐿 = {𝑙1 , 𝑙2 , … , 𝑙𝑘 }, e.g. set of locations, such as regional news recommendations. a specific address, city or country. Each item is avail- Finally, we use experimental validation in a different able starting from a certain time 𝑡𝑖 , i.e. the publication domain, i.e. regional news recommendations instead of timestamp. In case this is not available, we compute geotagged tweets. We find that in both domains, location- 𝑡𝑖 = min({𝑡 | ⟨𝑢, 𝑗, 𝑡⟩ ∈ 𝒟 ∶ 𝑖 = 𝑗}). For time-aware recom- aware methods outperform popularity, content-based mendations we imitate the online setting as close as possi- and collaborative filtering-based approaches. ble and evaluate offline based on a sliding window protocol In time-aware and news recommendations, most algo- where we partition 𝒟 in time given parameters △𝑡train and rithms are content-based, based on collaborative filtering △𝑡test . That is, we train the model at timestamp 𝑡 using 𝒟train = {⟨𝑢𝑘 , 𝑖𝑘 , 𝑡𝑘 ⟩ | 𝑡 − △𝑡train < 𝑡𝑘 ≤ 𝑡} and predict inter- 1 https://bitbucket.org/len_feremans/regio-reco/ Identify Filter and rank Pre-processing Create profile near regions candidate articles Filter popular Top frequent Geographic Geographic items (nearby) regions distance Popularity Training window Content-based Near regions using CF Publication User history datetime CF Temporal Location Popularity Social Figure 1: Overview of proposed LTARS recommender system consisting of 4 main steps. At each step there are different factors we account for: spatial influence, temporal influence, popularity influence and social influence actions in 𝒟test = {⟨𝑢𝑘 , 𝑖𝑘 , 𝑡𝑘 ⟩ | 𝑡 < 𝑡𝑘 < 𝑡 +△𝑡test }. Addition- dependencies for determining nearby user-preferred lo- ally, we define the set of impressionable items at times- cations. tamp 𝑡 using 𝐶impr = {𝑖 | 𝑖 ∈ 𝐼 ∶ 𝑡 − △𝑡train < 𝑡𝑖 < 𝑡 + △𝑡test }. The interaction data is represented using a user-item- 3.2.1. Top frequent locations location matrix as shown in Figure 2. The notations and definitions used in this paper are summarised in Table 1. Given a user 𝑢 and history 𝐼𝑢,𝑡 at timestamp 𝑡 (i.e. all The goal is to generate a set of top-𝑛 personalised interactions in 𝒟train ) we count the top-𝑘 most frequent items most relevant to each user 𝑢 given their history 𝐼𝑢 geotags using: at timepoint 𝑡. 𝑃 𝑘 (𝑢, 𝑡) = {𝑙 | ∃𝑖 ∈ 𝐼𝑢,𝑡 ∶ 𝑙 ∈ 𝐿𝑖 ∧rank(𝑠𝑢𝑝(𝑢, 𝑙, 𝑡)) ≤ 𝑘} where 3.2. Identifying regional preferences and |{𝑖 | 𝑖 ∈ 𝐼𝑢,𝑡 ∶ 𝑙 ∈ 𝐿𝑖 )}| sup(𝑢, 𝑙, 𝑡) = . neighbouring locations |𝐼𝑢,𝑡 | We assume each user has preferences for multiple loca- The frequency, or support, is a measure of the user- tions of interest, i.e. their home address, work address location preference, i.e. the ratio of location-specific views or recent vacation stay. First, we propose a straight- versus all views for that user. We remark that in our ex- forward technique for determining location preferences periments, recommendations based on the top frequent using the user’s history based on frequency. Next, we locations improve substantially using a longer history use location-location and user-location (or collaborative) (e.g. a full month instead of the last days) and after re- moving the 5% most popular items as very popular items are detrimental when inferring locational preferences. l1 l2 i1 i2 i3 i4 i5 3.2.2. Geographical neighbour locations u1 ✓ ✓ ✓ We define for each location 𝑙𝑖 the top-𝑘 nearest locations u2 ✓ ✓ using u3 ✓ ✓ 𝑘 (𝑙 ) = {𝑙 | 𝑙 ∈ 𝐿 ∧ rank(dist(𝑙 , 𝑙 )) ≤ 𝑘} NGEO 𝑖 𝑗 𝑗 𝑖 𝑗 u4 ✓ where we use dist to denote the geodesic distance be- Figure 2: Illustration of the user-item-location matrix. Based tween two geotags in kilometres. In practice, we use the on this example, we infer that user 𝑢1 is primarily interested set 𝐿train ⊆ 𝐿 containing any location associated with in items from location 𝑙1 , i.e. sup(𝑢1 , 𝑙1 , 𝑡) = 1 and user 𝑢3 in any item in the training data since for many (smaller) both location 𝑙1 and 𝑙2 , where sup(𝑢3 , 𝑙1 , 𝑡) = sup(𝑢3 , 𝑙2 , 𝑡) = 0.5. locations we have no items published in the past period, For simplicity, we assume a single geotag per item in the i.e. cold-start locations and a lower likelihood for future illustration. recommendations. Symbol Description the notation 𝑃(𝑙𝑗 | 𝑙𝑖 ) to denote the conditional probability 𝑈 set of users between two locations, e.g. 30% of users who view items 𝐼 set of items from 𝑙𝑖 also view items from 𝑙𝑗 , defined as: 𝐿 set of locations 𝑢 a user, 𝑢 ∈ 𝑈 𝑀𝑖,𝑗 𝑖 an item, 𝑖 ∈ 𝐼 𝑃(𝑙𝑗 | 𝑙𝑖 ) = where 𝑙 a location, 𝑙 ∈ 𝐿 𝑀𝑖,𝑖 𝐿𝑖 or 𝑙𝑖 set of locations 𝐿𝑖 ⊆ 𝐿 or a single location 𝑙𝑖 association with item 𝑖 𝑀𝑖,𝑗 = ∑ min ( ∑ 1 (𝑙𝑖 ∈ 𝐿𝑎 ) , ∑ 1 (𝑙𝑗 ∈ 𝐿𝑏 )) . 𝑡 a timestamp, often representing current 𝑢∈𝑈 𝑎∈𝐼𝑢,𝑡 𝑏∈𝐼𝑢,𝑡 time Δ𝑡 an interval in time We remark that locations can re-occur in the history for 𝑡𝑖 publication (or first interaction) timestamp each user and is represented by a multiset, henceforth of item 𝑖 the number of co-occurrences is the minimum of the 𝒟 set of interactions, i.e. tuples ⟨𝑢, 𝑖, 𝑡⟩ total number of occurrences of each pair of locations 𝒟train set of interactions for training in [𝑡−Δtrain , 𝑡] aggregated over all users. We define for each location 𝑙𝑖 𝒟test set of interactions for evaluation in ]𝑡, 𝑡 + Δtest ] the top-𝑘 nearest locations using: 𝐶impr set of impressionable candidate items at 𝑡, 𝑘 (𝑙 , 𝑡) = i.e. published before 𝑡 + Δtest N𝐶𝐹 𝑖 {𝑙𝑗 | 𝑙𝑗 ∈ 𝐿 ∧ rank (𝑃(𝑙𝑗 |𝑙𝑖 )) ≤ 𝑘} 𝐼𝑢,𝑡 set of items viewed by user 𝑢 before times- tamp 𝑡 3.2.4. Recommending items based on sup(𝑢, 𝑙, 𝑡) frequency, or support, of location 𝑙 in 𝐼𝑢,𝑡 neighbouring locations 𝑃 𝑘 (𝑢, 𝑡) user-location preferences, i.e. top-𝑘 loca- We extend regional preferences 𝑃 𝑘 (𝑢, 𝑡) to near locations tions with highest support in 𝐼𝑢,𝑡 using: 𝑃 ′ (𝑢, 𝑡) top-𝑘 most frequent locations and nearby locations from either 𝑁GEO or 𝑁CF 𝑘 𝑘 𝑃 ′ (𝑢, 𝑡) = 𝑃 𝑘1 (𝑢, 𝑡) ∪ {𝑙𝑖 | ∃ 𝑙 ∈ 𝑃 𝑘1 (𝑢, 𝑡) ∶ 𝑙𝑖 ∈ 𝑁𝛼 2 (𝑙, 𝑡)} 𝑁GEO (𝑙) top-𝑘 geographically nearest locations to 𝑙 𝑃(𝑙𝑗 | 𝑙𝑖 ) conditional probability of visiting location where 𝑁𝛼 is either 𝑁GEO or 𝑁CF and 𝑘1 and 𝑘2 are hyper- 𝑙𝑗 given 𝑙𝑖 is visited before 𝑘 𝑁CF (𝑙, 𝑡) top-𝑘 nearest locations to 𝑙 based on collab- parameters. We remark that if the profile 𝑃 𝑘 (𝑢, 𝑡) only orative filtering contains low-density locations, we might have fewer than 𝐶LOC (𝑢, 𝑡) candidate items matching user-location top-𝑛 item recommendations after filtering. Additionally, preferences locations in 𝑃 𝑘 (𝑢, 𝑡) are based only on the history of the 𝐶POP (𝑡) candidate items matching recency and pop- current users, while we use 𝑁GEO if users are interested ularity constraints in locations near frequently visited past locations or 𝑁CF 𝐶(𝑢, 𝑡) candidate items matching user-location if users are interested in locations frequently co-visited preference, recency and popularity by all users. rank LOC score 𝑖 based on user-location preference Next, we filter impressionable candidate items in 𝐶impr rank POP+REC score 𝑖 based on recency and popularity matching location regional preferences. Given a user 𝑢 rank LTARS score 𝑖 based on user-location preference, recency and popularity and regional preferences 𝑃 ′ (𝑢, 𝑡) we define the set of rank CF score 𝑖 based on collaborative filtering filtered candidate items at timestamp 𝑡: Table 1 𝐶LOC (𝑢, 𝑡) = {𝑖 | 𝑖 ∈ 𝐶impr ∶ 𝐿𝑖 ∩ 𝑃 ′ (𝑢, 𝑡) ≠ ∅}. Overview of notation and definitions We rank items in 𝐶LOC (𝑢, 𝑡) based on the user-location preference score using: 3.2.3. Collaborative filtering-based neighbour locations 𝑠𝑢𝑝(𝑖, 𝑙𝑖 , 𝑡) if 𝑙𝑖 ∈ 𝑃 𝑘 (𝑢, 𝑡) ⎧ dist(𝑙𝑖 ,𝑙𝑗 ) ⎪sup(𝑢, 𝑙 , 𝑡) ⋅ (1 − ) An alternative technique is to learn nearby locations ⎪ 𝑗 max({dist(𝑙 ,𝑙 ) | 𝑙 ∈N 𝑘 (𝑙 )}) 𝑗 𝑘 𝑘 GEO 𝑗 rankLOC (𝑢, 𝑖, 𝑡) = based on the geotagged item histories of all users, result- ⎨ if 𝑙𝑖 ∈ 𝑃 ′ (𝑢, 𝑡) ∧ 𝑁𝛼 = 𝑁GEO ing in because you interacted with location 𝑙𝑖 we recom- ⎪ sup(𝑢, 𝑙𝑗 , 𝑡) ⋅ 𝑃(𝑙𝑖 |𝑙𝑗 ) ⎪ mend location 𝑙𝑗 type of inference. Hence, we compute ⎩ if 𝑙𝑖 ∈ 𝑃 ′ (𝑢, 𝑡) ∧ 𝑁𝛼 = 𝑁CF . the co-visitations for each location similar to item-based collaborative filtering but counting co-occurrences of the In case a candidate item 𝑖 is tagged with multiple locations location of items instead of the items themselves. We 𝐿𝑖 , we use the location having the maximal score w.r.t. to store co-occurrence counts in a matrix 𝑀 |𝐿|×|𝐿| . We use the user for estimating relevance, i.e. argmax sup(𝑢, 𝑙𝑗 , 𝑡). 𝑙𝑗 ∈𝐿𝑖 For example, assume 𝑢1 has interacted with 100 items 𝑖1 was published 15 minutes ago with 0 views and 𝑖2 of which 50 are tagged with 𝑙1 and 20 with 𝑙2 during the was published 5 hours ago with 100 views. Again we training period [𝑡 − △𝑡train , 𝑡[. Assume that 30% of all rank 𝑖1 before 𝑖2 since rankPOP (𝑖1 , 𝑡) = 0+10 0.25 = 40 and users who interacted with 𝑙1 also interacted with 𝑙3 . The rankPOP (𝑖2 , 𝑡) = 100+10 = 22. user-location preferences are then rankLOC (𝑢1 , 𝑙1 , 𝑡) = 5 0.5, rankLOC (𝑢1 , 𝑙2 , 𝑡) = 0.2 and rankLOC (𝑢1 , 𝑙3 , 𝑡) = 0.5 × 0.3. 3.4. Combining location- and time-aware recommendations with collaborative 3.3. Ranking items on popularity and filtering recency This section proposes combinations of location-, time- This section considers strategies to filter and rank items aware and collaborative filtering-based recommenda- on popularity and recency. A popularity filter keeps can- tions. didate items above a certain popularity threshold. A recency filter keeps candidate items published before a 3.4.1. Location- and time-aware recommendations specific timestamp. Both approaches have their draw- First, we propose to combine scores to recommend items backs, i.e. a popularity filter removes cold-start (or very having relevant geotags and are trending. We define the recent) items, while a recency filter removes slightly aged set of candidate items by filtering on both regional pref- yet popular items. We define the set of candidate items erences, popularity and recency: 𝐶(𝑢, 𝑡) = 𝐶 LOC (𝑢, 𝑡) ∩ filtered on both recency and popularity at timestamp 𝑡 𝐶 (𝑡). For a user 𝑢 the candidate items 𝑖 ∈ 𝐶(𝑢, 𝑡) are POP as: ranked using: 𝐶POP (𝑡) = {𝑖 | 𝑖 ∈ 𝐶impr ∶ (𝑡 − △𝑡1 < 𝑡𝑖 ) ∧ (𝑡 − △𝑡2 < 𝑡𝑖 ∨ rankLTARS (u, i, t) = pop(𝑖, 𝑡, △𝑡pop ) > 𝜖) where 𝛼 ⋅ rankLOC (𝑢, 𝑖, 𝑡) + (1 − 𝛼) ⋅ rank POP+REC (𝑖, 𝑡) pop(𝑖, 𝑡, △𝑡pop ) = |{⟨𝑢𝑘 , 𝑖𝑘 , 𝑡𝑘 ⟩ | ⟨𝑢𝑘 , 𝑖𝑘 , 𝑡𝑘 ⟩ ∈ 𝒟train ∶ where 𝑖𝑘 = 𝑖 ∧ 𝑡𝑘 > 𝑡 − △𝑡pop }|. rankPOP+REC (𝑖, 𝑡) rank POP+REC (𝑖, 𝑡) = max({rankPOP+REC (𝑗, 𝑡) | 𝑗 ∈ 𝐶(𝑢, 𝑡)}) Here, △𝑡1 , △𝑡2 , △𝑡𝑝𝑜𝑝 and 𝜖 are hyper-parameters. For in- stance, by selecting △𝑡1 = 4𝑑, △𝑡2 = 2𝑑, △𝑡pop = 4𝑑, and where 𝛼 is a hyper-parameter to control the relative 𝜖 = 1 we exclude items that are more than 4 days old or weight of a user’s regional preferences versus the popu- between 2 to 4 days old and have fewer than 1 interac- larity of an item normalised over its age. tions during the last 4 days. For brevity of this paper we omit detailed experiments on the effect of selecting can- 3.4.2. Online computation of location- and didate using 𝐶POP . But we remark that on the regional time-aware recommendations news dataset, we have a recall@10 of 0.221 when using the default set of impressionable items 𝐶impr , which in- An advantage of the proposed approach for location- creases to 0.231 (+4.5%) by filtering articles older than two and time-aware recommendation is that it can be com- days and to 0.239 (+8.1%) using 𝐶POP w.r.t. the previous puted online. For updating 𝑃 𝑘 (𝑢, 𝑡) for each user we store parameters. the frequency of each location in memory and update Traditionally, we rank candidate items on recency, i.e. counts when new interactions arrive. Since, NGEO is time- publication timestamp or popularity. However, both ap- independent, we precompute the pairwise distances once proaches have their disadvantages. We propose to rank offline having a complexity of 𝑂(|𝐿|2 ) and store the result- candidate items 𝑖 on recency and popularity jointly using: ing matrix. For computing neighbouring locations based on collaborative filtering, the complexity is 𝑂(|𝒟 | + |𝐿|2 ). pop(𝑖, 𝑡, △𝑡pop ) + 𝑐 Algorithms exist to update the item similarities incremen- rank POP+REC (𝑖, 𝑡) = tally [14], which in principle could be adopted for updat- 𝑡 − 𝑡𝑖 ing location similarities online. However, since location where we normalise the popularity by the number of neighbourhoods are typically less dynamic, the need for hours since the publication and where 𝑐 represents a bias frequent model updates is less important. Therefore, term for cold-start items. For instance, given item 𝑖1 that we choose to re-compute the co-visitation matrix reg- was published 1 hour ago with 100 views and an item 𝑖2 ularly, thereby adopting sparse optimisation techniques published 5 hours ago with 200 views. We rank 𝑖1 before that make this computation extremely efficient, even on 𝑖2 since on average more users have viewed 𝑖1 items in large datasets [11]. Finally, we update the popularity a single hour. As a second example, assume 𝑏 = 10 and counts of each item online. We remark that online model training is essential for both computational efficiency is usually more of it where many regions and towns have and accuracy, since in many domains, such as social me- multiple articles published every day. dia, news or auction websites recent items are the most We load all interactions and article metadata during a relevant. 40-day period (from 1st July until 11th August 2021) and exclude all articles containing general news and sport. 3.4.3. Hybrid recommendations Next, we filter items and users having fewer than five interactions and items that are viewed more than once A limitation of the previous approach is that two users by the same user. By default, we remove candidate items with the same regional preferences 𝑃 ′ (𝑢, 𝑡) receive the that are more than 4 days old. We remove the overall top same recommendations at timestamp 𝑡. A natural exten- 1% most popular items to overcome that the recommen- sion is to adopt existing time-aware collaborative filtering dation model is biased towards predicting only popular methods [8] and have a hybrid solution where we also items. After pre-processing, we have 7.6 million interac- account for user-item preferences. Therefore, we adopt tions, 458 755 users and 9 493 items. Next, we perform item-based collaborative filtering as a strong baseline an offline simulation where we evaluate recommenda- [10, 6]. We compute conditional probabilities for each tions using a sliding window of 2 hours (△𝑡 test = 2ℎ) in item pair based on co-visitations and apply exponential the last week of data since the popularity distribution age-based discounting to give more weight to recent in- (and results) vary substantially in time [23]. At each two- teractions [8, 17]. We define a weighted co-visitation hourly interval, we train a model based on interactions matrix 𝐹 |𝐼 |×|𝐼 | using: in the past △𝑡train days and report recall@10 and ndcg@10 𝑡 𝑡 during the entire test week for users with interactions in 𝐹𝑖,𝑗 = ∑ 𝑤𝑢,𝑖 ⋅ 𝑤𝑢,𝑗 where both the train and test set. 𝑢∈𝑈 𝑡−𝑡𝑢,𝑘 𝑡 = {𝑎 ⋅ (1 − 𝑏) 𝑤𝑢,𝑘 if 𝑖𝑘 ∈ 𝐼𝑢,𝑡 }. 4.1.2. Public datasets 0 otherwise We also use two public location-based social network Here, 𝑡 − 𝑡𝑢,𝑘 denotes the difference in hours between the datasets to promote reproducibility. The datasets contain current time 𝑡 and the time of interaction 𝑡𝑢,𝑘 transformed227,428 and 573,703 check-ins collected for 10 months using an exponential time-decay function parameterised from Foursquare in New York City and Tokyo [27]. Each by 𝑎 and 𝑏. For collaborative filtering-based recommen- check-in is associated with a venue (or item), timestamp, dation we compute a score: GPS coordinates and category, which we ignore. Since the dataset is relatively small, we use a single time-based 𝐹𝑖,𝑗 rankCF (𝑢, 𝑖, 𝑡) = ∑ 𝑃(𝑖 | 𝑗) = ∑ split and use the last month for evaluation. We pre- 𝑗∈𝐼𝑢,𝑡 𝑗∈𝐼𝑢,𝑡 𝐹𝑗,𝑗 + 1 process the dataset as before and filter items and users with fewer than five interactions and items viewed more where 𝑃(𝑖|𝑗) denotes the conditional probability between than once. Additionally, we round GPS coordinates to 2 two items, e.g. 30% of users who (recently) viewed article decimals to create location tags, thereby considering all 𝑗 also (recently) viewed article 𝑖. coordinates within 1.11 kilometres identical. The main Finally, we recommend items based on both location- properties of each dataset are shown in Table 2. and time-aware preferences and collaborative filtering, i.e. given a user 𝑢 and a candidate item 𝑖 ∈ 𝐶(𝑢, 𝑖), we 4.2. Comparing regional preferences and define: neighbouring locations determined rankLTARS+CF (𝑢, 𝑖, 𝑡) = using geodistance and collaborative 𝛽 ⋅ rankLTARS (𝑢, 𝑖, 𝑡) + (1 − 𝛽) ⋅ rankCF (𝑢, 𝑖, 𝑡). filtering In this first experiment, we investigate if users are more 4. Experimental setup and results interested in geotagged items that are nearby geograph- ically or from similar locations based on collaborative 4.1. Dataset and offline evaluation filtering. We investigate the following methods on the 4.1.1. Regional news regional news dataset: We collect data from a prominent regional newspaper 1. Using the top-𝑘 most frequent regions, i.e. 𝑃 𝑘 (𝑢, 𝑡) in Belgium. In the current digital age more and more without nearby regions. users look for information on newspaper websites which 2. Add geographically near locations from NGEO . brings several challenges for the recommendation system. 3. Add near locations based on collaborative filter- Regional news is different from general news in that there ing from NCF . Dataset #users #items #locations #interactions Regional news 458 755 9 493 329 7 649 178 Foursquare TKY 2 292 7 057 732 128 555 Foursquare NYC 1 083 3 908 527 40 935 Table 2 Properties of datasets after preprocessing. ′ For extending the profile 𝑃 (𝑢, 𝑡) we select hyperparame- 4.4. Comparing popularity, time-aware ters 𝑘1 = 3 and 𝑘2 = 3, use Δtrain = 30𝑑 and filter the top collaborative filtering, content-based 5% most popular items before learning regional prefer- filtering and location- and ences and the location similarity matrix. We rank candi- date items matching regional preferences on recency. time-aware recommendation systems In Figure 3, we show each method result’s for ndcg@10 In this experiment, we compare the following recommen- and recall@10. The mean ndcg@10 is 0.184 when using dation systems on the proprietary regional news dataset collaborative filtering, 0.222 when using geographically and two public datasets: nearby locations and 0.203 when only using the top-k fre- quent locations from the history. Therefore, we observe 1. A popularity baseline ranking the most trending a relative increase of 8.5% using nearby regions deter- items. mined using geodesic distance. A similar trend holds 2. Content-based filtering ranking the most similar for recall@10. We observe that adding nearby locations items using soft-cosine based on a pre-trained using collaborative filtering does not perform well in this word2vec embedding [26]. dataset. However, we argue that this variant has poten- 3. Item-based collaborative filtering with age-based tial in other applications, such as Twitter tag predictions discounting [8, 17]. [21], where locations are more distant and international. 4. LTARS with geographically near locations and ranking jointly on user-location preference and recency and popularity. 4.3. Comparing ranking methods 5. A hybrid recommender where we combine Intuitively purely ranking on recency as we do in the LTARS with item-based collaborative filtering. previous experiment does not result in the best top-𝑛 recommendations. In this experiment, we compare the 4.4.1. Regional news dataset following ranking methods on the regional news dataset: We select hyperparameters △𝑡pop = 12ℎ for popularity, 1. On recency △𝑡train = 3𝑑 for collaborative filtering and △𝑡train = 30𝑑 2. On popularity for LTARS. For collaborative filtering we set the weight- 3. On rankPOP+REC decay parameters to 𝑎 = 1 and 𝑏 = 0.1. For LTARS we use 4. On rankLTARS the extended profile 𝑃 ′ (𝑢, 𝑡) where we use the top-3 most frequent regions (𝑘1 = 3) and top-3 geographically near- We filter candidate items using the top-3 most frequent est neighbours (𝑘2 = 3) and for ranking we set 𝛼 to 0.25 locations for each user and set hyperparameters △𝑡pop = thereby giving more relatively more weight to the user- 12ℎ for the popularity window, 𝑐 = 0 for rankPOP+REC location preference score. For the hybrid recommender and 𝛼 = 0.5 for rankLTARS giving equal weight to the user- we set 𝛽 to 0.5. location preference score rankLOC and rankPOP+REC . The resulting ndcg@10 and recall@10 values over a In Figure 4, we show the results for ndcg@10 and re- week are shown in Figure 5 and the average values in call@10. The mean ndcg@10 is 0.205 by ranking on pop- Table 3. Concerning ndcg@10 the hybrid method works ularity, 0.204 using recency and 0.218 using rankPOP+REC best, i.e. with an ndcg@10 of 0.270 we observe a 6.2% (+5.9%). If we rank using rankLTARS the ndcg@10 in- increase over item-based collaborative filtering, a 13.7% creases to 0.226 (+9.3%). We remark that by ignoring increase over LTARS, and a 50% increase over popularity. recency and ranking on user-location preference score We omit the results from the plot for the content-based only the ndcg@10 decreases to 0.122. The recall@10 recommender: with an ndcg@10 of only 0.019 it performs is respectively 0.285, 0.296, 0.300 and 0.309 by ranking poorly. With a recall@10 of 0.301 the proposed LTARS on popularity, recency, rankPOP+REC and rankLTARS . We method has the highest recall@10 and we observe a 4.3% conclude that ranking items on user-location preference increase compared to item-based collaborative filtering, a and popularity/age outperforms baseline ranking func- 24.9% increase compared to the popularity baseline, and a tions by a wide margin. small 1.3% increase compared to the hybrid recommender. 0.30 Pk P k + NGEO 0.25 P k + NCF ncdg 0.20 0.15 0.10 2021-08-05 2021-08-06 2021-08-07 2021-08-08 2021-08-09 2021-08-10 2021-08-11 time 0.40 Pk P k + NGEO 0.35 P k + NCF 0.30 recall 0.25 0.20 0.15 2021-08-05 2021-08-06 2021-08-07 2021-08-08 2021-08-09 2021-08-10 2021-08-11 time Figure 3: Ndcg@10 and recall@10 over one week for different methods for determining regional preferences and neighbouring locations on the regional news dataset. Using the top-3 frequent location the mean ndcg@10 is 0.203, by adding near location from NGEO the ndcg@10 increases to 0.222 (+8.5%), and by adding nearby locations from NCF the ndcg@10 decreases to 0.184 (-9.3%). rankLTARS 0.30 rankPOP + REC recency 0.25 popularity ncdg 0.20 0.15 2021-08-05 2021-08-06 2021-08-07 2021-08-08 2021-08-09 2021-08-10 2021-08-11 time rankLTARS 0.40 rankPOP + REC recency 0.35 popularity recall 0.30 0.25 0.20 0.15 2021-08-05 2021-08-06 2021-08-07 2021-08-08 2021-08-09 2021-08-10 2021-08-11 time Figure 4: Ndcg@10 and recall@10 over one week for different methods for ranking candidate items on the regional news dataset. By ranking on popularity results the a mean ndcg@10 is 0.205 and on recency the mean ndcg@10 is 0.204. By ranking on rankPOP+REC the ndgc@10 increases to 0.218 (+5.9%). By ranking on rankLTARS the ndcg@10 increases to 0.226 (+9.3%). Interestingly, there is no clear winner over the entire LTARS we use the extended profile 𝑃 𝑘 (𝑢, 𝑡) where we use week, i.e. on the last day we observe a severe drift in the top-20 most frequent regions and set 𝛼 to 0.75. For popularity bias better captured by collaborative filtering the hybrid recommender we set 𝛽 to 0.5. since LTARS filters out (popular) items not matching We show the results in Table 3. We find that, concern- regional preferences. However, the accuracy of LTARS ing ndg@10 the hybrid method performs best, followed validates the premise that local items are often more by LTARS with a large margin. Concerning recall@10 relevant. LTARS performs worse than the popularity baseline. We remark that by ignoring ranking on item recency would 4.4.2. Public datasets further improve results. We conclude that our method outperforms the popularity and item-based collaborative We repeat the previous experiment using two public filtering methods by a large margin concerning ndcg@10. location-based social network datasets. We remark that in both datasets, there is no significant preference to- 4.4.3. Execution time wards more recent venues. We set the popularity and training window to use all available data, i.e. we select In Table 4 we show the total execution time in seconds for hyperparameters △𝑡pop = △𝑡train = 9𝑚 for collaborative training the model and computing predictions. Runtimes filtering and LTARS and do not use weight-decay. For are measured on a laptop with an 2,3 GHz 8-Core Intel 0.35 0.30 0.25 ncdg 0.20 0.15 rankLTARS + CF rankLTARS 0.10 rankCF 0.05 popularity 2021-08-05 2021-08-06 2021-08-07 2021-08-08 2021-08-09 2021-08-10 2021-08-11 0.5 time 0.4 0.3 recall 0.2 rankLTARS + CF rankLTARS rankCF 0.1 popularity 2021-08-05 2021-08-06 2021-08-07 2021-08-08 2021-08-09 2021-08-10 2021-08-11 time Figure 5: Ndcg@10 and recall@10 for different recommendations techniques over a week on the regional news dataset. The mean ndcg@10 is 0.135 for the popularity baseline, 0.253 for item-based collaborative filtering, 0.233 for LTARS and 0.270 for the hybrid method. The mean recall@10 is 0.226 for the popularity baseline, 0.288 for item-based collaborative filtering, 0.301 for LTARS and 0.297 for the hybrid method. We omit the results for the content-based recommender having only a mean ndcg@10 of 0.019. Dataset Popularity ItemKNN LTARS Hybrid ndcg@10: Regional news 0.135 0.253 0.233 0.270 Foursquare TKY 0.034 0.077 0.095 0.129 Foursquare NYC 0.031 0.052 0.084 0.106 recall@10: Regional news 0.226 0.288 0.301 0.297 Foursquare TKY 0.042 0.042 0.036 0.040 Foursquare NYC 0.043 0.041 0.028 0.033 Table 3 Comparing top-𝑛 recommender systems on private and public datasets. The best result on each dataset are in bold (best) or underlined (second best). Core i9 and 16 GB of RAM. The publicly available im- 4.5. Online evaluation plementation is in Python. We remark, that concerning To validate the findings of the proposed approach we per- complexity, model training is 𝑂(|𝐼 |2 ) for item-based col- formed an online A/B trial on two regional news websites laborative filtering and 𝑂(|𝐿|2 ) for LTARS with geodesic in Belgium. The goal of the recommendations is to sur- nearby regions. At test time methods have a compa- face relevant regional articles for each user. Users have rable cost, i.e. for LTARS we filter items on regional the option to explicitly specify one or more locations preference and compute the user-location preference and they are interested in, however only 1 in 4 users provide popularity-based scores for each item, while for item- this preference. Finding the right regions to recommend based collaborative filtering we compute the dot product articles from, therefore is an important problem for these between the history and the (time-weighted) item-item websites. similarity matrix. In the Regional news datasets we have During the trials users were randomly assigned to ei- 6 624 156 interactions, 456 578 users and 8 462 items in ther the control or treatment group during the test period the first window, yet total training time is only 27.3𝑠. For of 9 days. Both groups received an equal amount of users. making predictions we require 14.1s for 22 921 test users, The control algorithm recommends the most recent items which is less then 1 millisecond per user on average. We from the explicit interest locations if available and oth- conclude that LTARS is highly efficient and scalable to erwise recommends articles from each user’s most read large datasets. region. In the experimental group a user profile 𝑃 𝑘 (𝑢, 𝑡) is constructed with 𝑘 = 3 following the LTARS method described in this work, ignoring the 1% most popular Dataset Popularity ItemKNN LTARS Hybrid runtime (s) Regional news 3.6s 141.7s 34.7s 182.4s Foursquare TKY 0.4s 4.8s 2.4s 8.4s Foursquare NYC 0.8s 14.0s 7.5s 25.3s Table 4 Comparing the execution time of top-𝑛 recommender systems on private and public datasets. items. If a user has given an explicit regional preference, Parame- Description these regions are always included in their regional profile. ter So a user that indicated explicit interest in two regions, Δ𝑡train training window will have a third region deduced from their historical Δ𝑡test test window Δ𝑡pop training window for computing popularity interactions. For this group items are filtered on regional 𝑘1 top-𝑘1 locations with highest frequency preferences and ranked according to recency. 𝑘2 top-𝑘2 nearby locations in 𝑁GEO or 𝑁CF A total of 1.7 million boxes were requested for 235k 𝑐 bias term in rank POP+REC (𝑖, 𝑡) users on the first newspaper and 2.5 million boxes for 𝜖 threshold when filtering candidate items 375k users on the second. We find that the experimental on popularity group has a 5.1% (relative) increase in click-through-rate 𝑎, 𝑏 weights for exponential decay in time- on the first newspaper and an 11.4% increase on the sec- aware item-based collaborative filtering ond. Both results were statistically significant at the 99% 𝛼 relative weight for ranking on regional pref- confidence level. In addition to the CTR results, the re- erences versus popularity and recency gional profiles also cover more of the user’s regions of 𝛽 relative weight for ranking on location and interest, recommending them articles from more diverse time versus collaborative filtering regions. Table 5 Overview of hyperparameters 4.6. Sensitivity of hyperparameters In Table 5 we summarise the hyperparameters introduced to evaluate the weighted sum w.r.t. 𝛼 and 𝛽 using the by the proposed LTARS. We investigate the sensitivity of cached partial scores for location, recency and collabora- our model with respect to the most important parameters, tive filtering-based recommendations for each user, item 𝛼 and 𝛽 for giving more weight to location, recency or pair. collaborative filtering. To clearly show the influence of these parameters, we report recall@10 with different parameter settings on two datasets. 5. Conclusion In Figure 6 we plot the recall@10 for varying values of 𝛼 and 𝛽 between 0.0 and 1.0 while keeping other pa- In this paper, we tackle the important problem of opti- rameters fixed as discussed before (i.e. Δ𝑡train = 30𝑑, mising location- and time-aware recommendations. We Δ𝑡pop = 12ℎ, 𝑘1 = 3 and 𝑐 = 0 for Regional news). On propose techniques for determining regional preferences the Regional news dataset we find that the hand-picked and neighbouring locations of interest. Additionally, we values for 𝛼 and 𝛽 in the last experiment are sub-optimal. consider ranking functions that consider spatial, tempo- With a value of 𝛼 set to 0.75 the recall@10 is 0.361 which ral and behavioural factors. We performed an extensive increases to 0.378 (+ 4.6%) using the optimal value of comparison offline using a realistically time-aware pro- 𝛼 = 0.3 and with a value of 𝛽 = 0.5 the recall@10 is 0.348 tocol based on sliding windows. Experiments show that which increases to 0.360 (+ 3.4%) using 𝛽 = 0.9. A similar the neighbourhood-based location- and time-aware rec- trend is visible in Foursquare TKY where also a local ommendation system and hybrids thereof outperform maximum is found with values of 𝛼 and 𝛽 in between. popularity, content-based and time-aware collaborative This suggests that we should optimise hyperparameters filtering-based methods on a large regional news dataset to further increase accuracy. We remark that in a dy- and two public location-based social network datasets. namic environment where there is a potential drift in Additionally, we performed an online A/B trial showing popularity bias, it make sense to adjust hyperparame- a clear increase in click-through-rate. ters periodically, i.e. using the last batch (or window) A limitation of our work is that the proposed model is of interactions for tuning. Note that optimising 𝛼 and straightforward and many of the proposed components 𝛽 is computationally inexpensive since we only have consist of heuristics. We motivate this by the fact that for rankLTARS 0.36 rankLTARS + CF 0.375 0.35 0.370 0.34 recall@10 recall@10 Regional news 0.365 0.33 0.360 0.32 0.355 0.31 0.350 0.30 0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0 0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0 α β 0.035 rankLTARS + CF 0.042 0.030 0.040 recall@10 recall@10 Foursquare TKY 0.025 0.020 0.038 0.015 0.036 rankLTARS 0.010 0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0 0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0 α β Figure 6: Recall@10 on the Regional news (up) and Foursquare Tokyo (bottom) datasets for varying hyperparameter. In the first experiment (left) we vary 𝛼 between 0.0 (ranking only on rankPOP+REC ) and 1.0 (ranking only on rankLOC ). In both datasets we observe a local maximum for 𝛼 in between, advocating that ranking on both popularity/recency and user-location preferences is important. In the second experiment (right) we varying 𝛽 between 0.0 (ranking only on collaborative filtering) and 1.0 (only on rankLTARS ). Again, we observe a local maximum for 𝛽 in between, advocating that the hybrid method outperforms LTARS and CF-based recommendations. location- and time-aware recommendation, the implicit and robust baseline when comparing future research in interaction data is biased and extremely sparse where we online location- and time-aware recommender systems. have relatively few interactions specific to one period and For future research, it would be of interest to update pa- location. Moreover, Dacrema et al. have recently shown rameters dynamically, i.e. by selecting the best value of that well-tuned simple baselines, such as ItemKNN, are 𝛽 and 𝛼 based on the evaluation of those parameters on difficult to beat when using more realistic evaluation the previous period and adapt to the current temporal strategies [6]. A second limitation is that the method is context, i.e. drift in the popularity distribution. specific to applications where items are tagged with one or more locations and the volatility of items is crucial. Acknowledgements We find that the intrinsic simplicity and heuristic na- ture make our model efficient to compute online and the The authors would like to thank the VLAIO project on capacity to predict fresh and cold-start items. We con- qualitative evaluation for online recommender systems clude that the proposed algorithm is useful as an efficient for funding this research. References ing the cold-start problem in location recommen- dation using geo-social correlations. Data Mining [1] Gediminas Adomavicius and Alexander Tuzhilin. and Knowledge Discovery, 29(2):299–323, 2015. Context-aware recommender systems. In Recom- [13] Florent Garcin, Boi Faltings, Olivier Donatsch, Ayar mender systems handbook, pages 217–253. Springer, Alazzawi, Christophe Bruttin, and Amr Huber. Of- 2011. fline and online evaluation of news recommender [2] Jie Bao, Yu Zheng, David Wilkie, and Mohamed systems at swissinfo. ch. In Proceedings of the 8th Mokbel. Recommendations in location-based social ACM Conference on Recommender systems, pages networks: a survey. GeoInformatica, 19(3):525–565, 169–176, 2014. 2015. [14] Olivier Jeunen, Koen Verstrepen, and Bart Goethals. [3] Pedro G Campos, Fernando Díez, and Iván Canta- Efficient similarity computation for collaborative dor. Time-aware recommender systems: a compre- filtering in dynamic environments. In Proceedings of hensive survey and analysis of existing evaluation the 13th ACM Conference on Recommender Systems, protocols. User Modeling and User-Adapted Interac- pages 251–259, 2019. tion, 24(1):67–119, 2014. [15] Mozhgan Karimi, Dietmar Jannach, and Michael [4] Cheng Chen, Xiangwu Meng, Zhenghua Xu, and Jugovac. News recommender systems–survey and Thomas Lukasiewicz. Location-aware personalized roads ahead. Information Processing & Management, news recommendation with deep semantic analysis. 54(6):1203–1227, 2018. IEEE Access, 5:1624–1638, 2017. [16] Sonal Linda and KK Bharadwaj. A genetic algo- [5] Lisi Chen, Shuo Shang, Zhiwei Zhang, Xin Cao, rithm approach to context-aware recommendations Christian S Jensen, and Panos Kalnis. Location- based on spatio-temporal aspects. In Integrated In- aware top-k term publish/subscribe. In 2018 IEEE telligent Computing, Communication and Security, 34th international conference on data engineering pages 59–70. Springer, 2019. (ICDE), pages 749–760. IEEE, 2018. [17] Nathan N Liu, Min Zhao, Evan Xiang, and Qiang [6] Maurizio Ferrari Dacrema, Paolo Cremonesi, and Yang. Online evolutionary collaborative filtering. Dietmar Jannach. Are we really making much In Proceedings of the fourth ACM conference on Rec- progress? a worrying analysis of recent neural ommender systems, pages 95–102, 2010. recommendation approaches. In Proceedings of the [18] Yiding Liu, Tuan-Anh Nguyen Pham, Gao Cong, 13th ACM conference on recommender systems, pages and Quan Yuan. An experimental evaluation 101–109, 2019. of point-of-interest recommendation in location- [7] Shaojie Dai, Yanwei Yu, Hao Fan, and Junyu Dong. based social networks. Proceedings of the VLDB Spatio-temporal representation learning with social Endowment, 10(10):1010–1021, 2017. tie for personalized poi recommendation. Data [19] Lorenzo Massai, Paolo Nesi, and Gianni Pantaleo. Science and Engineering, 7(1):44–56, 2022. Paval: A location-aware virtual personal assistant [8] Abhinandan S Das, Mayur Datar, Ashutosh Garg, for retrieving geolocated points of interest and and Shyam Rajaram. Google news personalization: location-based services. Engineering Applications scalable online collaborative filtering. In Proceed- of Artificial Intelligence, 77:70–85, 2019. ings of the 16th international conference on World [20] Yunseok Noh, Yong-Hwan Oh, and Seong-Bae Park. Wide Web, pages 271–280, 2007. A location-based personalized news recommenda- [9] María del Carmen Rodríguez-Hernández and Sergio tion. In 2014 International Conference on Big Data Ilarri. Ai-based mobile context-aware recommender and Smart Computing (BIGCOMP), pages 99–104. systems from an information management perspec- IEEE, 2014. tive: Progress and directions. Knowledge-Based [21] Róbert Pálovics, Péter Szalai, Júlia Pap, Erzsé- Systems, 215:106740, 2021. bet Frigó, Levente Kocsis, and András A Benczúr. [10] Mukund Deshpande and George Karypis. Item- Location-aware online learning for top-k recom- based top-n recommendation algorithms. ACM mendation. Pervasive and mobile computing, 38: Transactions on Information Systems (TOIS), 22(1): 490–504, 2017. 143–177, 2004. [22] Shaina Raza and Chen Ding. News recommender [11] Len Feremans, Boris Cule, Celine Vens, and Bart system: a review of recent progress, challenges, and Goethals. Combining instance and feature neigh- opportunities. Artificial Intelligence Review, pages bours for extreme multi-label classification. Inter- 1–52, 2021. national Journal of Data Science and Analytics, 10 [23] Teresa Scheidt and Joeran Beel. Time-dependent (3):215–231, 2020. evaluation of recommender systems. In Perspec- [12] Huiji Gao, Jiliang Tang, and Huan Liu. Address- tives@ RecSys, 2021. [24] Jeong-Woo Son, A-Yeong Kim, and Seong-Bae Park. Tenth International Conference on Language Re- A location-based news article recommendation sources and Evaluation (LREC 2016), Paris, France, with explicit localized semantic analysis. In Pro- may 2016. European Language Resources Associa- ceedings of the 36th international ACM SIGIR con- tion (ELRA). ISBN 978-2-9517408-9-1. ference on Research and development in information [27] Dingqi Yang, Daqing Zhang, Vincent W Zheng, and retrieval, pages 293–302, 2013. Zhiyong Yu. Modeling user activity preference [25] Waldo R Tobler. A computer movie simulating by leveraging user spatial temporal characteristics urban growth in the detroit region. Economic geog- in lbsns. IEEE Transactions on Systems, Man, and raphy, 46(sup1):234–240, 1970. Cybernetics: Systems, 45(1):129–142, 2014. [26] Stephan Tulkens, Chris Emmery, and Walter Daele- [28] Mao Ye, Peifeng Yin, Wang-Chien Lee, and Dik- mans. Evaluating unsupervised dutch word Lun Lee. Exploiting geographical influence for col- embeddings as a linguistic resource. In Nico- laborative point-of-interest recommendation. In letta Calzolari (Conference Chair), Khalid Choukri, Proceedings of the 34th international ACM SIGIR con- Thierry Declerck, Marko Grobelnik, Bente Mae- ference on Research and development in Information gaard, Joseph Mariani, Asuncion Moreno, Jan Odijk, Retrieval, pages 325–334, 2011. and Stelios Piperidis, editors, Proceedings of the