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 ofline 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.
[25] is crucial: “everything is related to everything else, and time-aware recommender systems (LTARS) and ranktions [1] is pre-filtering, i.e. we collect the location of each user and rank geotagged items nearby. However, tors, such as the geodesic distance between the geotag of
Every day, users consume items tagged with locations, e.g. local news articles from a particular city or Twitter tags trending in a specific region. Recommendations are essential to tackle information overload and filter relevant items from a huge set of available articles. In this context, the first law of geography posed by
Tobler this strategy is problematic. Firstly, many users have multiple and dynamic regions of preference, i.e. they might be interested in items near their home, work or recent vacation stay. Secondly, even after filtering on a preferred location, there are some biases resulting from population density, i.e. items geotagged with a big city will likely be more numerous and popular, which is likely not relevant for all users near that city (vice versa, rural locations might lack recent item interactions). Additiontal to inferring regional preferences, i.e. an item might be relevant to many users regardless of the associated Systems and User Modeling, jointly with the 16th ACM Conference on both locations [2].
In this work, we investigate the problem of locationing highly volatile geotagged items by considering both tors correlated with item relevance: (1) geographic facan item and the inferred regional preferences of a user, and (2) user-item preference, i.e. the item-neighbourhood based relevancy, and (3) the recency and popularity of an item. This problem has been studied in the context of time-aware recommendations [8, 3], location-aware recommendations [5, 21, 24, 7, 19], context-aware recommendations [1, 9, 16] and location-based social networks [2, 28, 12]. A key diference with closely related research by Pálovics et al. [21] is that we assume a user is erences are unknown and dynamic. A second key diference we consider is the volatility of items, which is crucial and often less in other domains such as point-of-interest recommendations [2]. A third diference is that we account for naturally occurring biases in the data such as an imbalance in the popularity and location distributions that hinder recommendations based on (context-aware)
We propose an LTARS that is orthogonal to existing ally, we find that the intrinsic popularity bias is detrimen- interested in multiple locations and these locational prefgeotag such as a news item related to an international in certain domains such as news recommendations [15] © 2022 Copyright for this paper by its authors. Use permitted under Creative Commons License collaborative filtering [ 1]. location-aware and context-aware recommender systems that assume known contextual features or static locational preferences [21, 1, 24] and make the following key contributions: or hybrids thereof [15, 22]. Das et al. propose time-aware news recommendations that combine item-based collaborative filtering, pre-filtering news items on recency, and 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 ofline evaltion preferences from users based on their fre- uation protocols [3]. A key diference is that we account quency in the long-term history of geotagged for location preferences and tackle cold-start item recomitems; (2) identify and rank relevant neighbour mendations. We compare with time-aware collaborative locations based on collaborative filtering or geo- ifltering 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 prefplement and publicly available1. It is also highly erences using a combination of collaborative filtering, eficient supporting frequent or online updates social influence, and geographic modelling where locaand 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 ofline 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 cotraining 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 recommender systems on ofline and online experiments for regional new recommendations.
2. Related work The proposed method is a combination of diferent steps and components that take diferent 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 = { 1, 2, … } be the set of users and = erences [21]. Our work is complementary by inferring { 1, 2, … , } the set of items. We consider implicit feedregional 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. = {⟨, , ⟩ | ∈ ∧ ∈ } while we propose an alternative neighbourhood-based and denote the user history using , , i.e. all interacted method. We argue that focusing on neighbouring loca- items up to timestamp . Each item has one or more lotions is essential in applications with a high cardinality cations or geotags where ⊆ = { 1, 2, … , }, e.g. set of locations, such as regional news recommendations. a specific address, city or country. Each item is availFinally, we use experimental validation in a diferent able starting from a certain time , 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- = min({ | ⟨, , ⟩ ∈ ∶ = }) . For time-aware recomaware methods outperform popularity, content-based mendations we imitate the online setting as close as possiand collaborative filtering-based approaches. ble and evaluate ofline based on a sliding window protocol
In time-aware and news recommendations, most algo- where we partition in time given parameters △ train and rithms are content-based, based on collaborative filtering △ test . That is, we train the model at timestamp using train = {⟨ , , ⟩ | − △ train < ≤ } and predict inter
The interaction data is represented using a user-item- 3.2.1. Top frequent locations actions in test = {⟨ , , ⟩ | < < +△ test }. Addition- dependencies for determining nearby user-preferred loat timepoint . location matrix as shown in Figure 2. The notations and definitions used in this paper are summarised in Table 1.
The goal is to generate a set of top- personalised items most relevant to each user given their history
Given a user and history , at timestamp (i.e. all interactions in train) we count the top- most frequent geotags using: (, ) = { | ∃ ∈ , ∶ ∈ ∧rank((, , )) ≤ }
where sup(, , ) = |{ | ∈ , ∶ ∈ )}| | , | .
We assume each user has preferences for multiple locations of interest, i.e. their home address, work address or recent vacation stay. First, we propose a straightforward technique for determining location preferences using the user’s history based on frequency. Next, we The frequency, or support, is a measure of the userlocation preference, i.e. the ratio of location-specific views versus all views for that user. We remark that in our experiments, recommendations based on the top frequent 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 reu1 u2 u3 u4 i 1 ✓ ✓ l i 1 2 ✓ i 3 ✓ ✓ i 4 ✓ ✓ l 2 i 5 ✓ set of users set of items set of locations a user, ∈ an item, ∈ a location, ∈ set of locations ⊆ or a single location association with item a timestamp, often representing current time an interval in time of item publication (or first interaction) timestamp Δtest ] tamp set of interactions, i.e. tuples ⟨, , ⟩ set of interactions for training in [−Δ train, ] set of interactions for evaluation in ], + set of impressionable candidate items at , i.e. published before + Δ test set of items viewed by user before timesfrequency, or support, of location in , user-location preferences, i.e. top- locations with highest support in , top- most frequent locations and nearby locations from either GEO or CF top- geographically nearest locations to conditional probability of visiting location given is visited before top- nearest locations to based on collaborative filtering preferences candidate items matching user-location candidate items matching recency and popularity constraints candidate items matching user-location preference, recency and popularity score based on user-location preference score based on recency and popularity score based on user-location preference, recency and popularity score based on collaborative filtering store co-occurrence counts in a matrix ||×|| . We use the notation ( | )to denote the conditional probability between two locations, e.g. 30% of users who view items from also view items from , defined as: ( | ) =
where , , , = ∑ min ( ∑ 1 ( ∈ ), ∑ 1 ( ∈ )) .
∈ ∈ ,
∈ , the top- nearest locations using: We remark that locations can re-occur in the history for each user and is represented by a multiset, henceforth the number of co-occurrences is the minimum of the total number of occurrences of each pair of locations aggregated over all users. We define for each location
N ( , ) = { | ∈ ∧ rank ( ( | )) ≤ } 3.2.4. Recommending items based on We extend regional preferences (, ) to near locations using: ′(, ) = 1(, ) ∪ { | ∃ ∈ 1(, ) ∶ ∈ 2
(, )} where is either GEO or CF and 1 and 2 are hyperparameters. We remark that if the profile (, ) only contains low-density locations, we might have fewer than top- item recommendations after filtering. Additionally, locations in (, ) are based only on the history of the current users, while we use GEO if users are interested in locations near frequently visited past locations or CF if users are interested in locations frequently co-visited by all users.
Next, we filter impressionable candidate items in impr matching location regional preferences. Given a user and regional preferences ′(, ) we define the set of ifltered candidate items at timestamp :
LOC (, ) = { | ∈ impr ∶ ∩ ′(, ) ≠ ∅}.
We rank items in LOC (, ) based on the user-location preference score using: rankLOC (, , ) = (, , )
if ∈ (, ) ⎧ ⎨ ⎪ ⎪ ⎩
dist( , ) ⎪⎪sup(, , ) ⋅ (1 − max({dist( , ) | ∈N GEO( )})
) sup(, , ) ⋅ (
| ) if ∈ ′(, ) ∧ = GEO if ∈ ′ (, ) ∧ = CF . location of items instead of the items themselves. We , we use the location having the maximal score w.r.t. to In case a candidate item is tagged with multiple locations the user for estimating relevance, i.e. argmax sup(, , ).
∈ For example, assume 1 has interacted with 100 items 1 was published 15 minutes ago with 0 views and 2 of which 50 are tagged with 1 and 20 with 2 during the was published 5 hours ago with 100 views. Again we training period [ − △ train, [ . Assume that 30% of all rank 1 before 2 since rankPOP (1, ) = 0+10 = 40 and users who interacted with 1 also interacted with 3. The rankPOP (2, ) = 100+10 = 22. 0.25 user-location preferences are then rankLOC ( 1, 1, ) = 5 0.5, rankLOC ( 1, 2, ) = 0.2 and rankLOC ( 1, 3, ) = 0.5 × 0.3.
pop(, , △ pop) = |{⟨ , , ⟩ | ⟨ , , ⟩ ∈ train ∶ where
This section proposes combinations of location-, timeaware and collaborative filtering-based recommendations. 3.4.1. Location- and time-aware recommendations First, we propose to combine scores to recommend items having relevant geotags and are trending. We define the set of candidate items by filtering on both regional preferences, popularity and recency: (, ) = LOC (, ) ∩ POP ().For a user the candidate items ∈ (, ) are ranked using: rankLTARS (u, i, t) = ⋅ rankLOC (, , ) + (1 − ) ⋅ rankPOP+REC (, )
This section considers strategies to filter and rank items on popularity and recency. A popularity filter keeps candidate items above a certain popularity threshold. A recency filter keeps candidate items published before a specific timestamp. Both approaches have their drawbacks, i.e. a popularity filter removes cold-start (or very recent) items, while a recency filter removes slightly aged yet popular items. We define the set of candidate items ifltered on both recency and popularity at timestamp as: POP () = { | ∈ impr ∶ ( − △ 1 < ) ∧ ( − △2 < ∨ pop(, , △ pop) > )
where = ∧ > − △ pop}|.
rankPOP+REC (, ) =
rankPOP+REC (, ) max({rankPOP+REC (, ) | ∈ (, ) }) where we normalise the popularity by the number of hours since the publication and where represents a bias term for cold-start items. For instance, given item 1 that was published 1 hour ago with 100 views and an item 2 published 5 hours ago with 200 views. We rank 1 before 2 since on average more users have viewed 1 items in a single hour. As a second example, assume = 10 and Here, △ 1, △ 2, △ and are hyper-parameters. For instance, by selecting △ 1 = 4 , △ 2 = 2 , △ pop = 4 , and where is a hyper-parameter to control the relative = 1 we exclude items that are more than 4 days old or weight of a user’s regional preferences versus the popubetween 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 efect of selecting can- 3.4.2. Online computation of location- and didate using POP . 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 impr , which in- An advantage of the proposed approach for locationcreases to 0.231 (+4.5%) by filtering articles older than two and time-aware recommendation is that it can be comdays and to 0.239 (+8.1%) using POP w.r.t. the previous puted online. For updating (, ) 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 timepublication timestamp or popularity. However, both ap- independent, we precompute the pairwise distances once proaches have their disadvantages. We propose to rank ofline having a complexity of (|| 2)and store the resultcandidate items on recency and popularity jointly using: ing matrix. For computing neighbouring locations based on collaborative filtering, the complexity is (| | + || 2).
pop(, , △ pop) + Algorithms exist to update the item similarities incremenrankPOP+REC (, ) = tally [14], which in principle could be adopted for updat − ing location similarities online. However, since location neighbourhoods are typically less dynamic, the need for frequent model updates is less important. Therefore, we choose to re-compute the co-visitation matrix regularly, thereby adopting sparse optimisation techniques that make this computation extremely eficient, even on large datasets [11]. Finally, we update the popularity counts of each item online. We remark that online model training is essential for both computational eficiency is usually more of it where many regions and towns have and accuracy, since in many domains, such as social memultiple articles published every day. dia, news or auction websites recent items are the most relevant. 3.4.3. Hybrid recommendations A limitation of the previous approach is that two users with the same regional preferences ′(, ) receive the same recommendations at timestamp . A natural extension is to adopt existing time-aware collaborative filtering methods [8] and have a hybrid solution where we also account for user-item preferences. Therefore, we adopt item-based collaborative filtering as a strong baseline [10, 6]. We compute conditional probabilities for each item pair based on co-visitations and apply exponential
We load all interactions and article metadata during a 40-day period (from 1st July until 11th August 2021) and exclude all articles containing general news and sport.
Next, we filter items and users having fewer than five interactions and items that are viewed more than once by the same user. By default, we remove candidate items that are more than 4 days old. We remove the overall top 1% most popular items to overcome that the recommendation model is biased towards predicting only popular items. After pre-processing, we have
7.6 million interactions, 458 755 users and 9 493 items. Next, we perform an ofline simulation where we evaluate recommendations using a sliding window of 2 hours (△ test = 2ℎ) in the last week of data since the popularity distribution hourly interval, we train a model based on interactions during the entire test week for users with interactions in both the train and test set. 4.1.2. Public datasets We also use two public location-based social network datasets to promote reproducibility. The datasets contain 227,428 and 573,703 check-ins collected for 10 months from Foursquare in New York City and Tokyo [27]. Each check-in is associated with a venue (or item), timestamp, GPS coordinates and category, which we ignore. Since the dataset is relatively small, we use a single time-based split and use the last month for evaluation.
We preprocess the dataset as before and filter items and users with fewer than five interactions and items viewed more than once. Additionally, we round GPS coordinates to 2 decimals to create location tags, thereby considering all coordinates within 1.11 kilometres identical. The main
In this first experiment, we investigate if users are more interested in geotagged items that are nearby geographically or from similar locations based on collaborative ifltering. We investigate the following methods on the regional news dataset: 1. Using the top- most frequent regions, i.e. (, ) without nearby regions.
ing from NCF . 2. Add geographically near locations from NGEO .
3. Add near locations based on collaborative filter∈ 0 , = ∑ , ⋅ ,
where , = { ⋅ (1 − ) − , otherwise
if ∈ , } .
Here, − , denotes the diference in hours between the current time and the time of interaction , transformed using an exponential time-decay function parameterised by and . For collaborative filtering-based recommendation we compute a score: rankCF (, , ) = ∑ ( | ) = ∈ , ∑ ∈ ,
, , + 1 where (|)
denotes the conditional probability between two items, e.g. 30% of users who (recently) viewed article also (recently) viewed article . and time-aware preferences and collaborative filtering, i.e. given a user and a candidate item ∈ (, )
, we define: rankLTARS+CF (, , ) = teractions [8, 17]. We define a weighted co-visitation age-based discounting to give more weight to recent in- (and results) vary substantially in time [23]. At each twomatrix | |×| | using:
Finally, we recommend items based on both location- properties of each dataset are shown in Table 2. ⋅ rankLTARS (, , ) + (1 − ) ⋅ rankCF (, , ).
filtering
4.1.1. Regional news We collect data from a prominent regional newspaper in Belgium. In the current digital age more and more users look for information on newspaper websites which brings several challenges for the recommendation system. Regional news is diferent from general news in that there Dataset Regional news Foursquare TKY Foursquare NYC
For extending the profile ′(, ) we select hyperparame- 4.4. Comparing popularity, time-aware ters 1 = 3 and 2 = 3, use Δtrain = 30 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 candidate 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 recommenand 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 frequent locations from the history. Therefore, we observe a relative increase of 8.5% using nearby regions determined using geodesic distance. A similar trend holds for recall@10. We observe that adding nearby locations using collaborative filtering does not perform well in this dataset. However, we argue that this variant has poten- with age-based tial in other applications, such as Twitter tag predictions [21], where locations are more distant and international. 1. A popularity baseline ranking the most trending
items. 2. Content-based filtering ranking the most similar items using soft-cosine based on a pre-trained word2vec embedding [26]. 3. Item-based collaborative filtering
discounting [8, 17]. 4. LTARS with geographically near locations and ranking jointly on user-location preference and recency and popularity. 5. A hybrid recommender where we combine
LTARS with item-based collaborative filtering.
Intuitively purely ranking on recency as we do in the previous experiment does not result in the best top- recommendations. In this experiment, we compare the following ranking methods on the regional news dataset: 1. On recency 2. On popularity 3. On rankPOP+REC 4. On rankLTARS We filter candidate items using the top- 3 most frequent locations for each user and set hyperparameters △ pop = 12ℎ for the popularity window, = 0 for rankPOP+REC and = 0.5 for rankLTARS giving equal weight to the userlocation preference score rankLOC and rankPOP+REC .
In Figure 4, we show the results for ndcg@10 and recall@10. The mean ndcg@10 is 0.205 by ranking on popularity, 0.204 using recency and 0.218 using rankPOP+REC (+5.9%). If we rank using rankLTARS the ndcg@10 increases to 0.226 (+9.3%). We remark that by ignoring recency and ranking on user-location preference score only the ndcg@10 decreases to 0.122. The recall@10 is respectively 0.285, 0.296, 0.300 and 0.309 by ranking on popularity, recency, rankPOP+REC and rankLTARS . We conclude that ranking items on user-location preference and popularity/age outperforms baseline ranking functions by a wide margin. 4.4.1. Regional news dataset We select hyperparameters △ pop = 12ℎ for popularity, △ train = 3 for collaborative filtering and △ train = 30 for LTARS. For collaborative filtering we set the weightdecay parameters to = 1 and = 0.1 . For LTARS we use the extended profile ′(, ) where we use the top-3 most frequent regions ( 1 = 3) and top-3 geographically nearest neighbours ( 2 = 3)and for ranking we set to 0.25 thereby giving more relatively more weight to the userlocation preference score. For the hybrid recommender we set to 0.5.
The resulting ndcg@10 and recall@10 values over a week are shown in Figure 5 and the average values in Table 3. Concerning ndcg@10 the hybrid method works best, i.e. with an ndcg@10 of 0.270 we observe a 6.2% increase over item-based collaborative filtering, a 13.7% increase over LTARS, and a 50% increase over popularity.
We omit the results from the plot for the content-based recommender: with an ndcg@10 of only 0.019 it performs poorly. With a recall@10 of 0.301 the proposed LTARS method has the highest recall@10 and we observe a 4.3% increase compared to item-based collaborative filtering, a 24.9% increase compared to the popularity baseline, and a small 1.3% increase compared to the hybrid recommender. 2021-08-05 2021-08-06 2021-08-07 2021-08-09 2021-08-10
2021-08-11 2021-08-08
time 2021-08-08
time 2021-08-05 2021-08-06 2021-08-07 2021-0t8i-m0e8 2021-08-09 2021-08-10 2021-08-11 Figure 4: Ndcg@10 and recall@10 over one week for diferent 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 (, ) where we use week, i.e. on the last day we observe a severe drift in the top-20 most frequent regions and set to 0.75. For popularity bias better captured by collaborative filtering the hybrid recommender we set to 0.5. since LTARS filters out (popular) items not matching We show the results in Table 3. We find that, concernregional 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 ifltering 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 △ pop = △ train = 9 for collaborative training the model and computing predictions. Runtimes ifltering and LTARS and do not use weight-decay. For are measured on a laptop with an 2,3 GHz 8-Core Intel 2021-08-05 2021-08-06 2021-08-07 2021-08-09 2021-08-10
2021-08-11 2021-08-08 time rankLTARS+CF rankLTARS rankCF popularity rankLTARS+CF rankLTARS rankCF popularity 0.35 0.30 g0.25 cd0.20 n 0.15 0.10 0.05
Core i9 and 16 GB of RAM. The publicly available im- 4.5. Online evaluation plementation is in Python. We remark, that concerning complexity, model training is (| | 2)for item-based col- To validate the findings of the proposed approach we perlaborative filtering and (|| 2)for LTARS with geodesic formed an online A/B trial on two regional news websites nearby regions. At test time methods have a compa- in Belgium. The goal of the recommendations is to surrable cost, i.e. for LTARS we filter items on regional face relevant regional articles for each user. Users have preference and compute the user-location preference and the option to explicitly specify one or more locations popularity-based scores for each item, while for item- they are interested in, however only 1 in 4 users provide based collaborative filtering we compute the dot product this preference. Finding the right regions to recommend between the history and the (time-weighted) item-item articles from, therefore is an important problem for these similarity matrix. In the Regional news datasets we have websites. t6h6e2fir4s1t5w6i nindtoewr a,cyteitontost,a4l5tr6a5in78inugsteimrseainsdon8l4y622i7t.e3m. sFoinr theDruthriencgotnhterotlroiarlstrueastemrsewntegreroruapnddoumrinlyg athsseigtensetdpetorioedimaking predictions we require 14.1s for 22 921 test users, of 9 days. Both groups received an equal amount of users. which is less then 1 millisecond per user on average. We The control algorithm recommends the most recent items conclude that LTARS is highly eficient and scalable to from the explicit interest locations if available and othlarge datasets. erwise recommends articles from each user’s most read region. In the experimental group a user profile (, ) is constructed with = 3 following the LTARS method described in this work, ignoring the 1% most popular Dataset runtime (s) 375k users on the second. We find that the experimental group has a 5.1% (relative) increase in click-through-rate on the first newspaper and an 11.4% increase on the second. Both results were statistically significant at the 99% confidence level. In addition to the CTR results, the regional profiles also cover more of the user’s regions of interest, recommending them articles from more diverse regions.
In Table 5 we summarise the hyperparameters introduced by the proposed LTARS. We investigate the sensitivity of our model with respect to the most important parameters, and for giving more weight to location, recency or collaborative filtering. To clearly show the influence parameter settings on two datasets. Parame5. Conclusion of and between 0.0 and 1.0 while keeping other pa- In this paper, we tackle the important problem of optirameters fixed as discussed before (i.e. Δ pop = 12ℎ, 1 = 3 and = 0 for Regional news). On the Regional news dataset we find that the hand-picked Δ train = 30 , mising location- and time-aware recommendations. We propose techniques for determining regional preferences and neighbouring locations of interest. Additionally, we values for and in the last experiment are sub-optimal. consider ranking functions that consider spatial, tempoincreases to 0.378 (+ 4.6%) using the optimal value of which increases to 0.360 (+ 3.4%) using = 0.9 . A similar trend is visible in Foursquare TKY where also a local maximum is found with values of and in between.
This suggests that we should optimise hyperparameters ral and behavioural factors. We performed an extensive comparison ofline using a realistically time-aware protocol based on sliding windows. Experiments show that the neighbourhood-based location- and time-aware recommendation system and hybrids thereof outperform popularity, content-based and time-aware collaborative ifltering-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) of interactions for tuning. Note that optimising and is computationally inexpensive since we only have
A limitation of our work is that the proposed model is straightforward and many of the proposed components consist of heuristics. We motivate this by the fact that for
rankLTARS+CF
rankLTARS+CF 0.35 010.34 0.31 0.30 0.042 0.036 0.015 0.010
rankLTARS 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 palocation. 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 and based on the evaluation of those parameters on dificult 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 nature make our model eficient 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 eficient for funding this research. [24] Jeong-Woo Son, A-Yeong Kim, and Seong-Bae Park. Tenth International Conference on Language ReA location-based news article recommendation sources and Evaluation (LREC 2016), Paris, France, with explicit localized semantic analysis. In Pro- may 2016. European Language Resources Associaceedings 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 Dikmans. Evaluating unsupervised dutch word Lun Lee. Exploiting geographical influence for colembeddings 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 conThierry 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