This paper presents a context-aware mobile shopping recommender system. A critique-based baseline recommender system is enhanced by the integration of context conditions like weather, time, temperature and the user's company. These context conditions are embedded into the recommendation algorithm via pre- and post- ltering. A nearest neighbor algorithm, using the concept of an average selection context, calculates how contextually relevant a recommendation is. Out of 20 clothing items from the hybrid recommendation algorithm, context-aware post- ltering searches for the nine best- tting items. The resulting context-aware recommender system is evaluated in a user study with 100 test participants. The answers of the user study show, that the recommendations were perceived as being better than the recommendations of a non-context aware recommender system.
H.4.2 [Information Systems Applications]: Types of Systems|Decision support
Context-aware recommender systems (CARS) are systems utilizing the user's context such as the user's position, weather or social environment to recommend items. A context-aware recommender system could for example recommend the \Albertina" museum rather than visiting the \Prater" amusement park if the user spends a rainy day in Vienna. This paper evaluates which kind of context information is relevant in a mobile shopping recommender system and how this information could be utilized to improve recommendations of clothing items in a context-aware recommender system. By integrating contextual mobile information into the recommendations it is expected, that the recommended items better t the customer's needs and therefore customers are more satis ed with the recommender system. The paper is organized as follows. We rst start o with some de nitions relevant for context-aware recommender systems and summarize related work. The next section de nes the context factors and describes the system's overall design. The user study evaluating the developed system is discussed in section 4. The paper concludes by summarizing its results and giving an outlook on future research topics. 2.
A widely used de nition in the area of context-aware applications is the de nition by Dey: \Context is any information that can be used to characterize the situation of an entity. An entity is a person, place, or object that is considered relevant to the interaction between a user and an application, including the user and applications themselves" [6, p. 5].
They de ne context as relevant information for an interaction between a user and an application. Therefore, if the context of an entity shall be de ned, it is necessary to ask which information is relevant to the situation.
Context-aware recommender systems (CARS) integrate
context into the recommendation process. This process can
be described by this three dimensional recommendation
function [
R : U ser
(1) The rating function (R) considers the Context (which is dened by all the di erent Context Factors) and recommends items of the item set (Item) to a user by predicting the rating that this user would give to an item. Context complicates the recommendation process as items can be rated in di erent contexts. An umbrella for example can be rated at good weather conditions very highly, due to the fact that it looks nice or is small. However, if it was raining the same umbrella could get a bad rating, due to the fact that it breaks at the slightest wind. So the context in the rating function brings additional complexity as the recommendation algorithm does not only have to match users with items, but also with the context.
Adomavicius and Tuzhilin [
Since the consideration of context can enhance the
usefulness of a recommendation for a user, CARS are recently
receiving a lot of attention [
Baltrunas et al. [
Researches have also been done on automatically
predicting the user's context. For example in [
I'm feeling LoCo [
So far, no research exists that analyzed all the contextual factors that might be useful when recommending clothing items from di erent stores for mobile shoppers and investigated how such a recommender system can be constructed and is perceived. Such an application could help the user detecting new (formerly unknown) brands or stores and nd clothes matching the user's fashion style. Compared to existing mobile recommender systems, clothing items are different in the way, that they frequently change. Such a recommender system has to be frequently trained or being able to provide good recommendations on a sparse dataset. We therefore rst acquire the relevant context factors in a mobile shopping scenario and then come up with a promising approach how to integrate this context into the recommendation process. 3.
We imagine a system that uses the user's mobile context to recommend clothing items available in shops close to the user's position. However, as in our previously developed baseline system (see Section 3.1), the new approach should still allow critiquing of items. As described in Section 2, context can be integrated into the recommender system in three di erent ways: contextual pre- ltering, contextual post- ltering and contextual modeling. In this work, 1https://foursquare.com two approaches (contextual pre- ltering and contextual postltering) are combined to improve the recommendations (see Figure 1). Pre- ltering (Section 3.2) is used to determine which items of the case base are relevant to the user. Relevance for example depends on the distance the user accepts to travel, or the opening hours of a shop. Post- ltering (Section 3.4) is used to lter the items that shall be recommended according to their adequacy to the current context by using a nearest neighbor algorithm. In order to build a database of contextually tagged items, a pre-study was executed asking users to classify items according to contexts (Section 3.3). This data ensures, that some items already are contextually tagged, which is needed for the post- ltering of the recommendations. The user interface and interaction design of our CARS is described in section 3.5. 3.1
The system presented in [
(b) Map view in both systems mendation algorithm. The algorithm uses a two-fold strategy: On a positive critique of an item (touching the thumbs up symbol) it shows items that more closely match the critiqued item. On a negative critique (thumbs down symbol) more diverse items are shown. In both cases, the recommendation algorithm uses a k-nearest neighbor algorithm to nd the k items that best t the current requirements. In this case k is set to nine, meaning that in each cycle, the user is shown nine di erent recommendations. In the following screen, the user selects which of the properties (color, brand, price, type) of the item shall be critiqued. The recommender system then shows more or less items (depending on the critique) of the selected feature(s). By touching an item's picture in the recommendations view, the system displays a result screen, where the user can select the item. The application also shows the immediate surroundings of the user in a map (see Figure 3). The system described in this section without context-awareness is used as a baseline for testing the context-aware recommender system. Furthermore we have made some adjustments to this content-based recommender system due to the changed dataset and performance problems. 3.2
In the contextual pre- ltering step, we make sure that only relevant data is loaded into the recommender system. Therefore, the context factors distance to shop, shop crowdedness, shop opening hours and item in stock are used to restrict the case base and avoid unnecessary search in items the user does not want to see. The user may state preferences for each of these context factors. The user might state a di erent distance to the shops or that she wants to see crowded places as well.
The case base is ltered in four steps. First, all shops that are not within the speci ed distance, then shops that are not open at the speci ed time and shops that do not match the crowdedness criterion are excluded. Finally, it is veri ed that the item is in stock. After pre- ltering the items based on these conditions, it is veri ed that at least 300 items are available in the case base, as our tests showed that this is the minimum amount of data to adequately react on the user's preferences. However, if there were not enough items available in the case base, these conditions are relaxed and the user is noti ed about this step.
Before being able to recommend items based on context,
the relevant context has to be de ned. A promising
approach to assess the context relevance for a tourism
scenario is presented by Baltrunas et al. [
Based on the pre- ltered item set the critique-based recommender selects 20 items. Out of these 20 items only nine are actually displayed. Therefore, the contextual postltering algorithm (illustrated in algorithm 1) has to eliminate eleven items in each cycle. The context factors time of the day, day of the week, company, temperature and weather are used to post- lter the recommendations. For this purpose, we use a k-nearest neighbor method because this technique has proven to be adequate in di erent CARS. The most important component in nearest neighbor algorithms is the used distance metric. In our approach, the user is not able to rate an item within a given context, but only to select it (and therefore implicitly rating it as good). Based on this consideration, we came up with a distance metric that de nes an average context in which an item is selected. The average context speci es in which context an item is selected. If an item was not selected in any context, it can be assumed, that this item neither is liked by a lot of users nor in a speci c context and can therefore receive a higher distance to the current context. Popular items, which are selected in many di erent contexts will receive a distance which is close to 0.5. However, as they are very popular, they should not receive a high distance and therefore their distance is reduced by a de ned percentage of their distance. avgContextDist(c; i) = b2ia
N (ia)
N (f ) P wj j2ij (2) Equation 2 de nes the distance metric. It calculates the distance between an item's (i) average context (in which the item is selected) and the current context (c). The rst quotient calculates the average distance to the current context whereas the second quotient normalizes the distances.
The set of all context conditions in which an item has
been chosen is de ned by ia. An individual context
condition in which an item has been chosen is de ned by b. For
each clothing type, the context factors are of di erent
importance. Hence, di erent weights (wi;b) can be assigned to
context conditions. We assigned the weights for each
clothing type based on a previously conducted experiment [
In order to make di erent items (with di erent overall weights) comparable, we normalize the distance between zero and one by multiplying with the second quotient of the function. Here N (f ) de nes the number of context factors ( ve) we use for post- ltering. The number of context factors is divided by the sum of weights of all context factors (wj ) for the speci c item (j 2 ij ).
dist(x; y) =
(graphDistance(x; y) if y is nominal
rjaxngyejy otherwise
(3)
If the context factor is ordinal, interval or ratio-scaled, the
distances are calculated based on the euclidean distance.
Otherwise the graphDistance, a pre-de ned distance for
nominal attributes, is used. This graphDistance is similar
to the distance used by Lee and Lee [
Algorithm 1 Post- ltering by current and item context 1: procedure ContextPostFilter(items; context; k) 2: for all item in items do 3: avgContextDistance(context; item) 4: if itemDistance == null then
6: end if 7: end for
10: end procedure
The algorithm's disadvantage is that it weights each factor independently without taking into consideration possible connections between the individual context factors. For example the connection of rain and being with a friend might be more di erent from rain and being with the family, than the individual distances between being with the family and being with a friend. This detection of dependencies could be done by decision trees or other machine learning techniques. Nevertheless, we expect that the algorithm provides reasonable recommendations for the user's current context without these dependencies. The algorithm calculates the context distances in less than 100 ms on a Samsung Galaxy S3 mini for 20 items with the items being set in (overall) 200 di erent contexts. It allows weighting of context factors for each clothing type separately and distances for nominal attributes. The method kN earestN eighbors(items; k) sorts the items by their distance to the current context. In case of any ties it uses the similarity measure that has already been applied in our baseline system (Section 3.1). 3.5
When starting the application, the user is asked to set the following context conditions manually: preferred distance to the shop, opening hours, temperature, weather and company. Moreover the user can specify if she wants to exclude items that are not in stock and shops that are too crowded. The conditions for time of the day and day of the week, are not captured, as it is expected that the users are aware of these conditions subconsciously. The context determination interfaces can be seen in Figure 5.
Figure 2a shows an example of a calculated set of recommendations. With the thumbs up or thumbs down button the user is able to critique the item's attributes such as price, brand, clothing type and color (see Figure 2b). Besides this critiquing possibility, the user is able to see some explanations such as why the particular item is recommended. By clicking on an item's picture, the user gets to another screen with more detailed information about the item and the store. Here, the user can nally select the item (see Figure 3a). This information should enhance the trust the user has in the recommendations as she can check whether the initial preferences (about distances to shop, crowdedness, etc.) were incorporated. Moreover, we implemented a map showing all available shops. On click of a shop we show the shop's opening hours, the crowdedness, the name, the distance to the current position and how many items (out of the current recommendations) are available at this shop (see Figure 3b). 4.
The user study was designed in order to test the di erences in user perceptions between the baseline application and the context-aware recommender system. We want to nd out whether the users perceive a di erence in the accuracy of recommendations. A second goal of the study is to nd out whether users are more satis ed with a recommender system that takes the mobile context into account. Therefore, the goal of the user study is to evaluate if the following hypotheses are true: Hypothesis 1: The integration of context-awareness leads to better perceived accuracy compared to non-contextaware recommendations.
Hypothesis 2: The integration of context-awareness improves the overall user satisfaction.
Hypothesis 1 is tested by comparing the ratings of recommendations in a context-aware system and a baseline system. The users should rate how they perceived the recommendations on a seven-point Likert-scale. Hypothesis 2 shall test whether the users are more likely to use, reuse or recommend the application. This is an indication on how well the system adapts to the users and how satis ed they are. 4.1
The user study is designed as a supervised within-subjects user survey to minimize the number of survey participants and improve the comparability between the applications. Each user tests both applications (the baseline system and the CARS) and answers a questionnaire afterwards. Which system is tested rst is ipped in between subjects so that a bias because of learning e ects could be reduced. The participants are asked to imagine being in the scenario, the tool generated for them, whereby the location is always Munich. The participant's task is to nd one item only, which they would like to try on. As soon as the users have found a suitable item, they are asked to select it, such that they can nish the test and answer the corresponding questionnaire. The target population of this application are young smartphone users that like to go shopping. In the user survey qualitative and quantitative data are collected. Qualitative data is measured via a questionnaire. It mainly consists of statements, the user should assess on a seven-point Likert-scale (from 1 - strongly agree, to 7 - strongly disagree), e.g. how satis ed the user is with the recommendations and the application in general. The quantitative data is directly measured within the application and includes the number of critiquing cycles, the time between viewing the rst recommendations and selecting an item, and the item diversity. Before the user starts using the application, a scenario describing the user's location, weather and company is generated for her (see Figure 6).
The participants are asked to actively select their context in the application and imagine it. This scenario is visually displayed to the users throughout the whole survey on a computer screen directly in front of them. The context conditions not mentioned in the scenario description, such as the crowdedness, can be selected by the user based on her own preferences.
The dataset used to test the application includes 5157 randomly selected fashion items, that were extracted from the Zalando API2 of their UK-store in February and March 2015. Since our dataset is arti cial, we distributed the items equally across all 129 shops and made realistic assumptions for our shops. The shop's opening hours were set to realistic values with moderate modi cations to have some di erences 2https://www.zalando.co.uk in the dataset. The crowdedness was set randomly with probability of 20 % and an item is in stock with a probability of 90 %. 4.2
All in all 100 participants (48 females, 52 males), between the ages of 17 and 30, took part in the user study. The answers to the Likert-statements (from 1 - strongly agree, to 7 strongly disagree) in this work either followed a positively or negatively skewed distribution and are ordinal scaled instead of interval scaled. Therefore, a two-tailed paired Wilcoxon signed rank test is executed, rather than a paired t-test, to detect whether there are any signi cant di erences between the distributions. The results of the two-sided tests are reported by stating a V and a p value. The V is the sum of ranks assigned to di erences with a positive sign. Therefore, a higher V stands for higher di erences in the user's decisions. The p value de nes how signi cant the results are. In general we evaluate whether the null hypothesis is likely to be true. The means, as well as the V and p values of the most important metrics of the two systems are shown in Table 1.
In order to test the user's perceived prediction accuracy, we asked if the recommended products tted the individual preferences. The baseline application's mean is 2.71 whereas the CARS mean is 2.34 (M edian = 2 for both systems). The Wilcoxon signed rank test reveals, that the recommendations of the CARS tted signi cantly better to the user's preferences than the baseline's recommendations (V = 1807, p < :01).
The context-awareness of the applications is evaluated by asking whether the products were in line with the provided scenario. The baseline application's mean is 2.82 whereas the CARS mean is 2.66 (M edian = 2 for both applications). The Wilcoxon signed rank test shows V = 1346, p = :54. This means that the users did not perceive any of the systems as being more context-aware than the other.
When asking the users whether they are likely to use the application again, the users stated that they are signi cantly more likely to use the CARS (M edian = 2, M ean = 2:64) again, than the baseline (M edian = 3, M ean = 3:06) application (V = 1563, p < :01).
The maximum time needed to nd an item in the baseline application was 867 seconds (M edian = 142s, M ean = 179s) and in the CARS application 697 seconds (M edian = 149s, M ean = 182s). The time needed to select an item is not signi cantly di erent between the applications (V = 2302:5, p = :45).
Another measure for the e ectiveness of the recommendation algorithm is the number of critiquing cycles until an item was selected. Participants completed their task in average 1.24 cycles less using CARS (M edian = 5, M ean = 6:1 with CARS, M edian = 5, M ean = 7:34 with the baseline system). Again a Wilcoxon signed rank test was executed (V = 2393:5, p = :11). However, the result is not signi cant, meaning that the null hypothesis cannot be rejected.
One of the goals of the CARS was to reduce the number of times an individual item is shown (item frequency) and thus increase the number of di erent items (item coverage). All in all the baseline application showed 7506 (1690 different; 22.5 % unique) and the CARS 6390 (1754 di erent; 27.4 % unique) items. We measured every time that an item was displayed to any user. The maximum number of times an item was shown was 115 (M edian = 3, M ean = 4:392) for the baseline application and 53 (M edian = 2, M ean = 3:622) for the CARS. A Wilcoxon signed rank test reveals that there is a signi cant di erence between the samples (V = 285253:5, p < :01), meaning that the CARS showed items signi cantly less frequent than the baseline. Although the CARS showed less items overall, more di erent items have been shown. This indicates that the recommended items have been more diverse.
Overall, 59 participants reported that they prefer the context-aware application (CARS). This are signi cantly more compared to a random distribution of answers as a chisquared test reveals (X 2 = 30:38, with 2 df [degrees of freedom], p < .001).
The test participants found that the CARS recommendations tted signi cantly better to their preferences. Therefore, hypothesis 1 that the recommendations by a contextaware system are perceived as better is retained. Hypothesis 2 that the overall user satisfaction is improved can also be retained to a certain degree as users were more satis ed with the CARS. The results might be less signi cant than expected as only six users rated items in context as an initial dataset. However, we wanted the dataset to be sparse as there are frequent changes to fashion collections.
In this work, a context-aware recommender system was developed and evaluated in a mobile shopping scenario. Our CARS is based on an active learning algorithm and uses a nearest neighbor algorithm. Compared to a system without context-awareness, the recommendations were perceived as signi cantly better in the CARS. Interestingly, the users did not attribute the better recommendation quality to the more context-aware recommendations but to better adaptability to their preferences and their clothing style, although the only di erence from an algorithmic perspective is the context-awareness. It should be investigated in more detail, whether context-awareness is only perceived subconsciously. The next step for this application would be to test it in an online-experiment where real context-aware information is elicited. In a rst approach the clothing data of some selected retailers would be enough to test this application online. In the future, we plan to conduct a user study where real context-aware information is elicited. Still a major challenge for context-aware applications is to acquire contextaware data to train or tweak a context-aware algorithm. For this user study, selected users classi ed the contexts in which they would try the clothes on. As the users in the user survey also imagined these contexts, we expect no signi cant di erences between the classi cation of the items and the imagined scenario in the user study. This approach might help in narrowing down the problem of acquiring relevant context data as a quick start for a context-aware application. However, it has to be evaluated how close real contextual ratings can be estimated with this method. In order to adapt the existing approaches of estimating a rating to a yes or no decision we had to develop the concept of an average context, in which an item is selected. We believe that every context-aware recommender system relying on yes or no decisions might have bene ts from adapting its context incorporation by using our approach. We also aim to nd out whether the results of this work can be transferred to other application scenarios, such as for grocery shopping or leisure activity recommendation systems.