Student Research Abstract: Context-Driven Mobile Apps Management and Recommendation Yong Zheng Center for Web Intelligence DePaul University Chicago, IL, USA
[email protected] 1.
PROBLEM AND MOTIVATION
in different contextual situations. Typically, those context information include time, location, occasion, and so forth. For example, a user may choose to play “Angry Birds” if he is on a short queue to buy a bus ticket, while he or she may choose the mobile office to work on some documents if the user is on a long-haul bus. This example shows user’s different choices on two different occasions: short-waiting time in a short queue, and long-waiting time on the bus.
The problem of information overload can be summarized as the difficulty for users to make decisions (e.g., retrieve a document or purchase an item) which is actually caused by the presence of too much information, especially the Web information. Information retrieval and recommender systems have successfully alleviated this problem by either exploring information relevance or adapting to users’ preferences.
Context-aware recommendation has been proved to be effective in many applications, including the mobile apps, since users’ preferences on apps are always changing from contexts to contexts. Meantime, the emergence of CARS brought a new recommendation opportunity: context suggestion [3, 2] which aims to recommend the appropriate contexts (such as time, location, occasion) for users to consume the items. For example, the system may suggest the users to use office for document processing on the bus rather than waiting in the short queue for bus ticketing. Or, it may recommends the users to have more workout at weekend. The underlying goal of context suggestion is to help users maximize the utilities or consumption by doing the right thing on the right time and at the right place or occasion.
With the development of mobile computing and systems, the information overload problem turns out to be more and more serious in the mobile application domain. More specifically, there are so many mobile apps offered in the mobile system (e.g, android market, apple store, etc), and users may install several apps on their mobile devices. How to well-management those apps and recommend the appropriate ones to the end users on the right time at the right place? Apparently, users’ needs on apps are usually dynamic and their preferences may be changed from time to time, where traditional ways cannot meet the requirements of users’ on-time or occasional demands. This paper proposes a context-driven framework for mobile apps management and recommendation, where context is raised to play an important role in users’ interactions with mobile apps and it is able to change the ways of how users use and manage their mobile apps. More specifically, both context-aware mobile app recommendation and context suggestion are embedded into the framework, and we also present our initial results on the context suggestions.
2.
For example, Google Music recently leased a function named as “Listen now”. Everytime when a user clicks it, the page will suggest a list of appropriate contexts for users to listen to the music shown in book Figure 1, so that users may system canasalso provide recommendations which werebe inthe mean rating an terested andoccasion. keep spinning on listening. giftedinbythose other suggestions users in the same significant splits f
inferred from the u with this user.
The splitting ap of dependent ratin context-aware reco [6]) can also be app However, they onl cannot infer how g
BACKGROUND AND RELATED WORK
Recommender system (RS) is a well-known system being able to assist users’ decision makings by recommending appropriate items. Context-aware recommender system (CARS) [1] is a new type of RS which aims to adapt to users’ preferences Permission to make digital or hard copies of part or all of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and the full citation on the first page. Copyrights for third-party components of this work must be honored. For all other uses, contact the Owner/Author. Copyright is held by the owner/author(s).
TABLE II: A Rank User u1 u1 u1 u1
Fig.1:1:Context OccasionsSuggestion suggested by on Google Music Figure Google Music
Split time: week with: girlfr mood: pos mood: nega
ACM 978-1-4503-3739-7/16/04.
Recently, Google released “Browse” functions on Google Music and Youtube channels so that users can view recby browsing specific content categories (e.g., 3. ommendations ARCHITECTURE AND APPROACHES The main challe music genre) and context categories (e.g., mood, activity). The For example, a use In context-aware recommendation, context considered “Listen Now” page on Google Music (shown inisFigure 1) is as = weekend”, “with an additional input, but system may not able to suggest a list of the contexts, where users willobserve get a listusers’ the companion is contexts in real-time. Limited information, such as time and of personalized playlists by clicking one of those suggestions. the movie. In this The suggested will from dynamically changesystem during aand day, GPS location, could becontexts obtained the mobile required for accura and it could be personalized for each user. data, but those information may not be influential contex-
http://dx.doi.org/10.1145/2851613.2852026
tual variables for recommendation purposes. In this section,
SAC 2016, April 4-8, 2016, Pisa, Italy
III.
633
S OLUTIONS AND C HALLENGES
Our previous work [4] has pointed out two categories of potential solutions: one is direct context prediction which views each context condition as a label and try to predict the appropriate labels to be suggested. Multilabel classification is the principal technique as the solution, and we have explored it in [4]. Another solution is named as indirect context recommendation, where users’ preferences on contexts can be inferred by context-aware recommendation algorithms, especially the ones exploring the dependencies between contexts and the other
IV.
C ON
In this paper, we d ommendation tasks formally introduce challenges. Specific by deviation-based as user splitting. algorithms [7] coul correlation between Also, we’d like to
We split the data to training (80%) and testing (20%). We use a multi-label classification algorithm (we choose binary relevance and use decision trees as the classifier) to build the model for suggesting contexts to based on the training set. For testing set, we assume the mobile apps in the testing set are the ones users are interested in. And we use the trained model to suggest contexts for each entry. In other words, we obtain multiple ranked list of contexts for each user, since this user may use multiple items in the testing set. Then, for each user, we aggregate the ranking of all the suggested and ranked contexts in the testing set and provide a single ranked contexts for this user.
we introduce our context-driven mobile framework and discuss the initial results on the context suggestion.
3.1
Context-Driven Mobile Framework
By combining the context suggestion and context-aware recommendation, it is able to solve the context collection and contextual recommendation problems simultaneously. The context-driven mobile framework can be described as: Context-aware' Recommenda,ons'
Context' Sugges,on'
c $
' Apps$list$1$ ' ' Apps$list$2$ ' ' ' Apps$List$3$ '
Mobile' Usage'
In terms of the baseline, we converted the data shown in Table 1 to a 2-dimensional one where users are rows, contexts are columns, and the value for each entry is the average usage frequency for each user on each context by aggregating the usage data across of items. And then we apply matrix factorization to this user-context data, and generate a ranked context for each user.
Figure 2: Context-Driven Mobile Framework First of all, a list of contexts will be suggested to each user. Those contexts could be the time or location collected from mobile system and GPS, and also could be any potential and influential contexts, such as holiday, workout, driving, trekking, mountain climbing, etc, which could be the domain-specific contexts and created manually. By users’ clicks on a specific context (e.g., c1, c2 or c3), the system will provide a list of recommended apps corresponding to selected context. The system will record users’ usage on those apps with selected contextual information, where those data will be finally used for context suggestions. Apparently, it is a self-cycling framework which enables the users interact with the mobile apps in suggest contexts. It is helpful to collect users’ contexts information, as well as utilize contextaware recommendations to adapt users’ dynamic demands.
3.2
The ground truth (i.e., the real ranking of contexts for each user) is obtained from the testing data in the converted usercontext matrix. And we use normalized discounted cumulative gain (NDCG) as the evaluation metric for top-5 recommendation. The results show that we obtain a 34% NDCG by using the matrix factorization on the converted data, and a 56% NDCG by using the MLC model, which reveals that by using MLC modelwe can obtain better ranked list of contexts for suggestions to each user.
4.
CONCLUSIONS AND FUTURE WORK
In this paper, we introduce a context-driven framework to fuse context suggestion and context-aware recommendation, where this way enables users to interact with the mobile apps in specific contexts directly, which further helps context collection and contextual recommendations.
Approaches
In terms of context-aware recommendation, there are several effective recommendation algorithms developed and the state-of-the-art algorithms have been implemented in the toolkit CARSKit [4]. The challenge is on the context suggestion, where previous work suggests to use the multi-label classification (MLC) to recommend contexts to a pair of . In this paper, we introduce a way to recommend contexts to each user by reusing the MLC technique.
In addition, we propose a solution using multi-label classification technique to recommend contexts to each user which is demonstrated to work better than simply using a recommendation model based on the evaluation using a real-world mobile apps data. In future, we’d like to explore more context suggestion algorithms and evaluate the context suggestion and context-aware recommendation on more data sets.
We use Frappe data1 which is a mobile app data set with context information and users’ usage (in terms of usage frequencies within 2 months). We sampled a subset of the data for evaluation, where there are 100 users, 237 apps and 6851 usage records. There are 6 contexts involved: we used 4 contexts provided by the data: at home, at work, weekend and weekday, and also randomly created values for additional two contexts: relaxed and busy. A sample of the data can be show in Table 1.
5.
Table 1: Example of Data user u1 u1 u2 1
app a1 a1 a2
usage 21 119 3
wkend 1 0 1
wkday 0 1 0
home 1 0 1
work 0 1 0
relaxed 1 0 0
busy 0 1 1
http://baltrunas.info/research-menu/frappe
634
REFERENCES
[1] G. Adomavicius, B. Mobasher, F. Ricci, and A. Tuzhilin. Context-aware recommender systems. AI Magazine, 32(3):67–80, 2011. [2] Y. Zheng. Context suggestion: Solutions and challenges. In Proceedings of the 15th IEEE International Conference on Data Mining Workshops. IEEE, 2015. [3] Y. Zheng, B. Mobasher, and R. Burke. Context recommendation using multi-label classification. In Proceedings of the 13th IEEE/WIC/ACM International Conference on Web Intelligence, pages 288–295, 2014. [4] Y. Zheng, B. Mobasher, and R. Burke. CARSKit: A java-based context-aware recommendation engine. In Proceedings of the 15th IEEE International Conference on Data Mining Workshops. IEEE, 2015.
