=Paper=
{{Paper
|id=Vol-1438/paper7
|storemode=property
|title=An Adaptive Electronic Menu System for Restaurants
|pdfUrl=https://ceur-ws.org/Vol-1438/paper7.pdf
|volume=Vol-1438
|dblpUrl=https://dblp.org/rec/conf/recsys/FilhoW15
}}
==An Adaptive Electronic Menu System for Restaurants==
https://ceur-ws.org/Vol-1438/paper7.pdf
An Adaptive Electronic Menu System for Restaurants
Paulo Henrique Azevedo Filho Wolfgang Wörndl
Aberklar GbR Technische Universität München
Helene-Mayer-Ring 7 Boltzmannstraße 3
80809, Munich, Germany 85748, Garching bei München, Germany
paulo.azevedo@aberklar.com woerndl@informatik.tu-muenchen.de
ABSTRACT restaurant to restaurant, the choice for a recommendation
This work shows the early stages of the development of a technique has to be further tweaked to address these issues.
collaborative-filtering-inspired adaptive system to stream-
line the ordering process at restaurants that use electronic The present work is an attempt to embrace this opportu-
menu systems. nity, building an adaptive recommender system for elec-
tronic restaurant menus, to aid guests in their ordering pro-
Among other results, the proposed system achieved a reduc- cess, based on MenuMate.
tion of the session duration, while increasing feedback given
by restaurant guests. MenuMate, developed by the German startup Aberklar1 ,
is an electronic menu that offers users a picture-centric ap-
Categories and Subject Descriptors proach for use at restaurants. Through MenuMate users can
H.4.m [Information Systems Applications]: Miscella- place orders, request the bill and provide feedback about
neous their experience. This is the system that was extended with
the adaptive system described in this paper. Figure 1 gives
an idea of the general structure of the dish overview screen,
General Terms and the arrows suggest the repositioning of dishes of the
Human Factors, Economics proposed method, that will be described soon.
Keywords The dish overview screen shows a title bar for each category,
Recommender Systems, Adaptive, Collaborative-Filtering, followed by thumbnails of each dish in that category. From
Electronic Menu, Target Variable that screen, the user can go to the dish details screen, with a
full screen picture and detailed description, from where the
1. INTRODUCTION dish can be ordered. After orders have been placed and the
There is an increasing trend in the number of tablet-based tab asked for, the user is invited to provide feedback about
systems for ordering food, in order to streamline the ser- a few different variables, such as food, drinks, service and
vice, increasing efficiency and profit. This is a great oppor- electronic menu system, using a star-based rating.
tunity to make a change: instead of mere replacements for
the paper menu and replacement for some of the roles of The rest of this document is organized as follows: Section 2
the waiters, such systems can do much more, such as be- will put this work in perspective in terms of the electronic
coming feedback-gathering devices, or helping guests choose menu used for its implementation, as well as of the related
their dishes better and increasing their satisfaction by offer- work. Section 3 will describe in general terms the proposed
ing meaningful suggestions. algorithm and how it fits in. Section 4 describes some of the
performed experiments and displays their results. Finally,
Those meaningful suggestions may stem from collaborative section 5 offers a brief summary as well as points some pos-
filtering. Well-established algorithms for collaborative filter- sibilities of further work to be explored.
ing exist, as well as several hybridization techniques blend-
ing this with other approaches. However, since users of 2. BACKGROUND AND RELATED WORK
electronic menus are mostly anonymous (no log-in neces- To achieve the aforementioned goal, once the current state
sary), and the offer of dishes and drinks varies greatly from of MenuMate is known, it is necessary to know what the
current state of the art for this specific niche is.
Wasinger et al. have proposed an electronic menu with an
embedded recommender system, called Menu Mentor [9].
The authors come from a perspective of highly personalized
explicit recommendations, processed on the user’s phone,
and that must be scrutable, that is, allow the users to know
why a given recommendation (be it positive or negative) was
RecSys Vienna, Austria, 16th-20th September 2015 1
www.aberklar.com/en
Four correlation coefficients were tested: Spearman’s [4],
Pearson’s [5], Goodman and Kruskal’s [2] and Kendall’s [3].
In fact, the experiments tried all of them in different sessions
and made a comparison between them. More coefficients
could be used, the only requirement is that they must yield
values between -1 and 1, which could imply a normalization
step for coefficients that do not yield results in this range.
In order to vary the position of dishes, there is what we
called pre-optimization randomization, which will shuffle the
dishes before sorting them, giving the chance to dishes that
have the same coefficient (within a small delta) to change
places, changing not only absolute but also relative order.
A bias could, otherwise, arise from the fact that stable sort-
ing algorithms were used, thus keeping items with a simi-
Figure 1: Screen of MenuMate with the overview of lar correlation coefficient always in the same relative order,
the dishes, as well as a suggestion of the dynamic preventing them from switching places and moving far away
reordering of the menu items. from each other. Also, dishes for which there is no previous
information always start with coefficient 0.
given, and give them the chance to override it. It does, how- The basic principle is as follows: after each session, the
ever, assume explicit user profiles, acquired through usage of tablets send the raw data gathered to the system that gen-
the system on the user’s smartphone. In order to minimize erates the sequence (there is one such system per restau-
user setup, restaurants should be able to offer the system rant). This system, based on the chosen coefficient and tar-
and the hardware. get variable, calculates the value for that variable and the
coefficients, generates the menu sequence, and at the be-
There are psychological studies about how to organize a ginning of each session the tablets poll this system for the
restaurant menu, with techniques such as placing items with latest sequence to be employed. This system, called Menu
high prices first to ”smoothen” the effect of the lower-priced Optimizer, is configurable in respect to the target variable,
items following, even if they are not actually cheap [6]. There action (maximize or minimize) and correlation coefficient,
are also studies evaluating guests’ gaze and how it corre- together with other parameters relevant for the A/B testing
lates to which dishes get ordered, placing dishes restaura- performed, that will be explained in the next section.
teurs want to have ordered in those positions, such as shown
at [1]. There are as well other possibilities that go in a sim-
ilar direction, such as [8] and [10]. Going for this sort of 4. PRELIMINARY USER STUDIES
psychological study, however, would require extensive trials Since the method proposed does not explicitly recommend
after every change to the menu, as well as the usage of the a single item or try to predict ratings, some methods usu-
same menu in a different place where cultural background ally employed to evaluate recommender systems cannot be
changes. This type of approach is also sensitive to the direc- employed, such as cross-validation, recall and precision and
tion of reading of the mother tongue of the restaurant guest, accuracy.
such as right to left in case of Arabic speakers.
There are, however, other methods that can be employed
The goal of this work was to develop an electronic menu sys- directly: the test was a double-blind A/B test, in which nei-
tem that adapts itself according to a certain target variable, ther the user nor the waiter knew which the target variable
such as session duration, but while avoiding the need both at the time was, nor the correlation coefficient used. It was
for the setup of a user profile and extensive trials after each also online (in the sense that real users were using the sys-
change. This is what will be presented in the next section. tem, generally spending real money through it [7], as was
the case in our tests), and measured a few variables.
3. ALGORITHM AND ARCHITECTURE Since there is no way to directly measure accuracy for this
The proposed idea is rather simple. It consists in reordering system, there is employment of efficiency. In traditional rec-
the menu items to optimize a quantifiable target variable, ommender systems, the goal is to try to predict what the
that can be either maximized or minimized. For this work, user would like to find and show it to them, shortening the
four target variables were studied: tab value, session dura- search. Assuming that users find more easily the informa-
tion, revenue rate (cents per minute) and feedback rating. tion they are looking for, they will more promptly take deci-
Once this definition is set, as different menu sequences are sions based on that information, namely order the dish they
used the value of each target variable is recorded for each intend to. Assuming this to be true, it would derive that
session, as well as the position of each dish in the menu. a reduction of session duration would imply increased ac-
The optimal menu sequence is computed by calculating the curacy in a way. To assess whether this is true, it should
correlation coefficient between each dish’s position and the be coupled with an increase of the feedback ratings, which
performance, and sorting ascending or descending, depend- would show that the items found satisfied the user. This
ing on whether the action is to maximize or minimize, re- metric, that as mentioned before, is called efficiency, and
spectively. was used to assess the system.
Figure 2: The evolution of feedback ratings and ses- Figure 3: How the values of the correlation coeffi-
sion duration at one of the restaurants, that suggests cients evolved over time at the two restaurants.
good efficiency.
back at El Patio was too low to assume it was meaningful).
Another adaptation made was with regards to serendipity. It suggests that indeed over time, independent of those be-
Instead of measuring it, there was the measurement of cat- ing the target variables, feedback tends to rise and session
alog coverage: the measurement of how much users tend duration to reduce, indicating good efficiency.
to order varied dishes, compared to users of plain paper
menu. For that, observations were made of tables whose Table 1 shows how the choice of a target variable influences
guests did not use MenuMate, and which dishes they or- the performance of that variable, as well as of other vari-
dered, and those were compared to the orders of MenuMate ables. In this table only sessions optimized with Spearman’s
users. correlation coefficient are used, because it was the only co-
efficient with a high enough number of observations from
There were two classes of tests performed, the stress tests which to derive conclusions (21 sessions in total). The ”none”
and the user tests. The stress tests, which will not be dis- line, represents results for a non-optimized menu. The sys-
played in detail here, were used to assess the scalability of tem did well in optimizing for the target variable, with the
the system, which was developed to be run on a Raspberry best results for both feedback and session duration, a close
Pi computer, with an embedded 900MHz ARM processor. It second for tab value, while not delivering good results for
suffices to say that the system could serve between 100 and revenue rate, because the higher tab value was not enough
160 tablets with unnoticeable performance losses, and that to compensate for the roughly halved session duration time
the limits reached were due to the test rig employed, rather achieved when the target variable was the duration. It is
than the system itself. Another interesting result on this also worth noticing that indeed, revenue rate is a direct con-
front is that, on the employed hardware, menu optimization sequence of both tab value and session duration, but the av-
time is increased on average by 3ms for each session stored, erage revenue rate is not necessarily the same as the revenue
which allows prediction of the optimization time based on rate of the averaged tab values and durations. For clarity,
the size of the history of observations. the best results for each variable are displayed in boldface.
The preliminary user studies were performed at two differ- Figure 3 shows another facet of the inner workings of the
ent restaurants in Munich, for about a week in each, time system: although the number of observed sessions was low,
during which the system would gather session data and au- it is very fast to converge the internal coefficients, suggesting
tomatically adapt itself, switching the target variable at reg- some stability after approximately 10 sessions. Another in-
ular intervals. The number of observations was rather small teresting effect is that the absolute values of the coefficients
due to the reduced number of days allowed for observations. tend to be higher for dishes that ended up being ordered,
There were 13 sessions observed in one restaurant, called El which means the system can in fact predict which dishes
Patio, and 35 in the other, called Wendlinger. The restau- will be ordered, by checking the dishes with highest abso-
rants had, respectively, 201 and 290 menu items, split in 24 lute value of the coefficients.
and 34 categories. The original menu sequence was devised
by their respective owners. Due to operational constraints, catalog coverage was only
measured at El Patio. There, 12 of the 13 sessions were
Figure 2 shows the evolution over time of the efficiency at with menu optimization enabled, that will be considered for
Wendlinger (the number of sessions to which users gave feed- this measurement. 50 sessions with paper menus were also
Target Variable Feedback Value (e) Duration Rev. Rate
None 2,0 25,17 12.865,6 20,9
Feedback 5,0 13,25 15.779,9 33,8
Tab Value 4,4 38,00 7.499,5 59,1
Session Duration 4,3 36,85 2.406,9 154,9
Revenue Rate 5,0 40,10 4.400,6 56,6
Table 1: Cross-references between target variables and the results for all interest variables.
observed. In total, 97 different dishes and drinks were or- • Feature extraction, to allow dishes of different restau-
dered, in different quantities. From those 97, 19 were or- rants to be matched and correlation information ex-
dered both with and without MenuMate, 63 only by users change, possibly improving results;
of the paper menu and 15 only by users of MenuMate with
menu optimization. • Extra variable isolation, that would allow control over
influential external variables such as weather, time of
It is not easy to extrapolate how those proportions would be the day or season;
in case an equal number of observations was available, and • If individual user profiling is done, further personaliza-
discarding paper menu-based sessions could arbitrarily lead tion, such as filtering out dishes based on allergies or
to any results. Assuming, however, that for both systems at taste preferences, could be done;
each session there is an equal probability of adding not pre-
viously ordered items, and that this probability is inherent • Stochastic exploration of dishes, which could allow for
to either MenuMate or the paper menu, a proportion rule recommendation of the next course based on previ-
may be followed. ously ordered dishes in a session.
In the case of the paper menu, 19 + 63 = 82 different items 6. REFERENCES
were ordered in 50 sessions, which yields an average of 1,64 [1] J.-G. Choi, B.-W. Lee, and J.-w. Mok. An experiment
new items per session. With MenuMate in use, there were on psychological gaze motion: a re-examination of
19 + 15 = 34 different items ordered in 12 sessions, resulting item selection behavior of restaurant customers.
in 2,83 new items per session. That may indicate that thus, Journal of Global Business and Technology, 6(1):68,
the menu optimizations led to guests ordering a bigger vari- 2010.
ety of items. Alternatively, it could be that as more sessions
[2] L. A. Goodman and W. H. Kruskal. Measures of
would be measured, these sessions would progressively get
association for cross classifications iii: Approximate
less ”innovative”, in terms of the ordered dishes, which could
sampling theory. Journal of the American Statistical
revert this balance. An extended evaluation, with a similar
Association, 58(302):310–364, 1963.
number of sessions in both conditions would help settle this
[3] M. G. Kendall. A new measure of rank correlation.
matter.
Biometrika, pages 81–93, 1938.
[4] J. L. Myers, A. Well, and R. F. Lorch. Research design
5. CONCLUSIONS AND FUTURE WORK and statistical analysis. Routledge, 2010.
The proposed menu optimizer harnesses principles of rec- [5] K. Pearson. Notes on regression and inheritance in the
ommender systems to improve restaurant menus according case of two parents. In Proceedings of the Royal
to an arbitrary interest variable. In fact, it can be used Society of London, volume 58, pages 240–242, 1895.
to facilitate user interaction for any system to which access
[6] W. Poundstone. Priceless: The myth of fair value
is anonymous and the number of items not overwhelming.
(and how to take advantage of it). Macmillan, 2010.
Another use that comes to mind is the choice of which pre-
[7] G. Shani and A. Gunawardana. Evaluating
sentations to attend at a conference, or main sights to visit
recommendation systems. In Recommender systems
in a city.
handbook, pages 257–297. Springer, 2011.
One of the main contributions is the strong suggestion that [8] B. Wansink, J. Painter, and K. Van Ittersum.
this purely statistical method may improve the menu even if Descriptive menu labels’ effect on sales. The Cornell
the underlying mechanism that drives the change is not un- Hotel and Restaurant Administration Quarterly,
derstood by the system (i.e. the psychological implications). 42(6):68–72, 2001.
[9] R. Wasinger, J. Wallbank, L. Pizzato, J. Kay,
The developed system comprises the algorithm for menu op- B. Kummerfeld, M. Böhmer, and A. Krüger. Scrutable
timization, coupled with a robust and scalable implemen- user models and personalised item recommendation in
tation of it. It was followed by qualitative and quantita- mobile lifestyle applications. In User Modeling,
tive tests, with real paying users, of which only very few Adaptation, and Personalization, pages 77–88.
results were presented this time due to page number con- Springer, 2013.
straints, but that nevertheless suggest efficacy of the pro- [10] S. S. Yang, M. M. Sessarego, et al. $ or dollars: Effects
posed method. of menu-price formats on restaurant checks. Cornell
Hospitality Reports, pages 6–11, 2009.
There are, however, some promising improvements to the
method. Among which, the highlights are: