=Paper=
{{Paper
|id=Vol-1688/paper-16
|storemode=property
|title=Deep Auto-Encoding for Context-Aware Inference of Preferred Items’ Categories
|pdfUrl=https://ceur-ws.org/Vol-1688/paper-16.pdf
|volume=Vol-1688
|authors=Moshe Unger,Bracha Shapira,Lior Rokach,Ariel Bar
|dblpUrl=https://dblp.org/rec/conf/recsys/UngerSRB16
}}
==Deep Auto-Encoding for Context-Aware Inference of Preferred Items’ Categories==
https://ceur-ws.org/Vol-1688/paper-16.pdf
Deep Auto-Encoding for Context-Aware Inference of
Preferred Items’ Categories
Moshe Unger, Bracha Shapira, Lior Rokach, Ariel Bar
Ben-Gurion University of the Negev
and Telekom Innovation Laboratories at BGU
Beer Sheva, Israel
mosheun@post.bgu.ac.il, {bshapira, liorrk, arielba}@bgu.ac.il
ABSTRACT inferring users' context utilizing high dimensional data. We sense
Context-aware systems enable the sensing and analysis of user the user’s rich contextual feature space (e.g., Wi-Fi networks,
context in order to provide personalized services to users. We accelerometers, light, microphones, etc.) from the user's mobile
observed that it is possible to automatically learn contextual phone and apply an unsupervised deep learning technique that
factors and behavioral patterns when users interact with the extracts the most important features and discovers significant
system. We later utilize the learned patterns to infer contextual correlations between them. We rely on the available users'
user interests within a recommender system. We present a novel feedbacks in order to detect different behavioral patterns from the
context-aware model for detecting users' preferred items’ raw sensor data. Specifically, we split the data according to
categories using an unsupervised deep learning technique applied similar genres (or categories) of items that have been rated by
to mobile sensor data. We train an auto-encoder for each item users in a recommender system (RS). Then we apply auto-
genre, using contextual data that was obtained when users encoding on the divided data for each genre (e.g., auto-encoding
interacted with the system. Given new contextual sensor data from for the 'food' genre, auto-encoding for the 'nightlife spot' genre,
a user, the discovered patterns from each auto-encoder are used to etc.) in order to discover different context patterns. Exploiting the
predict the category of items that should be recommended to the implicit correlations between environmental features that affect
user in the given context. In order to collect rich contextual data, user preferences, can be used to model the dynamic context of a
we conducted an extensive field study over a period of four weeks user. Thus, we aim to obtain unsupervised contexts from the deep
with a group of ninety users. The analysis reveals significant layers in order to determine the type of items relevant to the user's
insights regarding the inference of different granularity levels of current context.
categories that are available within the data. The major contributions of this paper include the following:
first, we show how to infer contextual user preferences and
Keywords availability employing the user's current unsupervised context and
Recommender Systems; Deep Learning; Auto-encoder; Context; context models that were learned from past users' interactions
Mobile with a RS. We suggest to split the rated data according to different
genres/categories of items (e.g., "food," "nightlife spot") and learn
1. INTRODUCTION a different deep model for each category by its contextual data.
Context plays an important role in determining the relevance of The models represent implicit contextual situations related to each
a service provided by an application to the user's needs. A system category. Given new raw sensor data, we predict the category of
is considered context-aware if it can extract, interpret, and use an item that reflects the user's current context. The categories may
context information and adapt its functionality to the users' be defined according to the level of granularity required by the
immediate context [2]. Obtaining explicit contexts is a resource target system's functionality and the available information about
demanding task, since it requires either inputs from users, the the categories. Second, we use an auto-encoders for modeling
knowledge of a domain expert, or collecting labeled contextual users' contexts from the data collected from mobile device
information. However, it is possible to use available users' ratings sensors. We demonstrate our finding with data collected from real
in order to learn implicit behavior patterns from available raw data users during an extensive user study.
[1,4,5]. In this paper, we build on Hull et al. (1997) who defined
context as the user situation. When dealing with context, three 2. METHOD
entities can be distinguished: places (rooms, buildings etc.), Our method infers contextual preferences of users in terms of
people (individuals, groups), and things (physical objects, the users’ availability for receiving recommendations and their
computer components etc.). In order to improve context preferred categories of items. The method consists of two phases,
prediction, it is necessary to discard indiscriminative or highly as presented in Figure 1: The first is the training phase in which
correlated features in order to avoid the curse of dimensionality we apply auto-encoding on contextual data collected from users'
[1]. In addition, when dealing with high dimensional data, it is interactions within a recommender system. We build several deep
important to consider dependencies between characteristics, in neural networks for different splits of the data according to the
order to avoid a large number of model parameters. Baltrunas et items’ categories (Figure 1a); The second phase is the prediction
al. [1] suggested a context based splitting approach in which phase, where we use new sensor data currently recorded from the
ratings of certain items were split according to the value of an users' mobile phone. We utilize the learned deep models to
item-dependent contextual condition. Although this technique reconstruct the input data and select the network that best fits the
reveals the best contexts for each item, it is limited to a single and data with minimal error. We can then predict the current preferred
binary context and thus cannot model relations between several user's category according to the selected contextual model (Figure
contexts. Moreover, while Baltrunas suggest to split the context 1b).
for each item, we learn all context patterns that related to the same
category of items. We suggest a novel approach for modeling and
splitting method, on the test data. Table 1 presents results for
prediction of the two settings. As observed, our model
significantly outperforms all of the tested classification algorithms
by at least 45% in terms of accuracy and by 23% in terms of
AUC. This phenomenon can be explained by the fact that the
auto-encoder can better represent high dimensional data and
correlations between features using all available features. We can
also notice that the tested classifiers obtained much lower results
in the prediction accuracy of seven categories. This can be
explained by the fact that the detailed items' categories are
conceptually more similar to each other, compared to categories in
(a) Training Phase previous setting.
Table 1. Prediction of High Level Categories
Prediction of Prediction of
Three Seven
Prediction Model Categories Categories
Accuracy AUC Accuracy AUC
Auto-Encoding 0.971 0.927 0.884 0.918
Random Forest 0.657 0.739 0.484 0.685
SVM 0.655 0.671 0.5 0.617
(b) Prediction Phase
Figure 1: Method Overview C4.5 0.668 0.751 0.51 0.289
3. EVALUATION 4. CONCLUSION
In this paper we presented a novel approach for inferring
3.1 Field Experiment and Data Collection contextual user preferences by applying auto-encoding on sensor
We aim to infer availability for receiving recommendations data. Our solution relies on the identification and usage of positive
("busy") and preferred categories of items ("food," "nightlife feedbacks acquired in contextual situations. In order to evaluate
spot") collected from RS in order to provide meaning and further our suggested model, we conducted an extensive user study over a
explanation regarding the unsupervised contexts. We evaluated period of four weeks with a developed application which displays
our method on data that was collected from mobile device sensors POI (point of interest) recommendations to users. The
in relation to a recommender system that provided experimental results show that we were successfully able to
recommendations of points of interest (POIs) obtained from predict preferred items’ categories at different granularity levels.
Foursquare1 API and received users’ feedback about the provided In all settings the auto-encoding approach was superior to the
recommendations. We developed an Android application which traditional state-of-the-art classification methods in terms of
monitors the user's sensors and recommends popular POIs nearby. accuracy and AUC. The results indicate that auto-encoder is the
90 students between the ages of 20-45 (53 male and 37 female) most effective method tested for modeling contextual patterns
participated in the experiment. The overall experiment was when dealing with high dimensional data.
conducted for a period of a month. Overall, the system collected
21,397 instances of positive user feedback (11,051 regrading food 5. REFERENCES
POIs, 812 regarding nightlife POIs and 9,534 reported "busy"). [1] Baltrunas, Linas, and Francesco Ricci. "Context-based
splitting of item ratings in collaborative
In order to acquire data from a variety of sensors, we used
filtering." Proceedings of the third ACM conference on
Android APIs and the Funf [3] application and collected nine
sensors: accelerometer, Wi-Fi, battery, light, orientation, magnetic Recommender systems. ACM, 2009.
field, gravity, audio level, and location. Additional information [2] Byun and Chverst B. Sarwar, G. Karypis, J. Konstan, and J.
was derived as well, such as day of the week, time of day, weather Riedl. "Analysis of Recommendation Algorithms for
conditions, activity recognition, screen-log, call-log, and traffic ECommerce". Electronic Commerce Proceedings of the 2nd
statistics of certain applications installed on the user's mobile ACM conference on Electronic commerce, 158 – 167, 2000.
device (i.e., Facebook, WhatsApp, etc.). Raw data collected by the [3] Nadav Aharony, et al. "Social fMRI: Investigating and
sensors and the additional information listed above were shaping social mechanisms in the real world." Pervasive and
aggregated, analyzed, and engineered to generate 247 features.
Mobile Computing 7.6 (2011): 643-659.
We aimed to train a model that would be able to classify a test
[4] Guo, Lei, et al. "Learning to recommend with social
set of categories at different granularity levels. Specifically, we
contextual information from implicit feedback." Soft
trained three auto-encoders that distinguished between three
different contexts: food, nightlife spot, and busy. Second, we Computing 19.5 (2015): 1351-1362.
distinguished between seven contextual categories including: [5] Wang, Hao, Naiyan Wang, and Dit-Yan Yeung.
restaurant, vegetarian, fast food, desserts, café, nightlife spot, and "Collaborative deep learning for recommender
busy. We applied several prediction models, using the time-based systems." Proceedings of the 21th ACM SIGKDD
International Conference on Knowledge Discovery and Data
1 https://developer.foursquare.com/ Mining. ACM, 2015.