This work shows the early stages of the development of a collaborative- ltering-inspired adaptive system to streamline the ordering process at restaurants that use electronic menu systems. Among other results, the proposed system achieved a reduction of the session duration, while increasing feedback given by restaurant guests.
Those meaningful suggestions may stem from collaborative ltering. Well-established algorithms for collaborative ltering exist, as well as several hybridization techniques blending this with other approaches. However, since users of electronic menus are mostly anonymous (no log-in necessary), and the o er of dishes and drinks varies greatly from restaurant to restaurant, the choice for a recommendation technique has to be further tweaked to address these issues. The present work is an attempt to embrace this opportunity, building an adaptive recommender system for electronic restaurant menus, to aid guests in their ordering process, based on MenuMate.
MenuMate, developed by the German startup Aberklar1, is an electronic menu that o ers users a picture-centric approach for use at restaurants. Through MenuMate users can place orders, request the bill and provide feedback about 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, and the arrows suggest the repositioning of dishes of the proposed method, that will be described soon.
The dish overview screen shows a title bar for each category,
followed by thumbnails of each dish in that category. From
that screen, the user can go to the dish details screen, with a
full screen picture and detailed description, from where the
dish can be ordered. After orders have been placed and the
tab asked for, the user is invited to provide feedback about
a few di erent variables, such as food, drinks, service and
electronic menu system, using a star-based rating.
The rest of this document is organized as follows: Section 2
will put this work in perspective in terms of the electronic
menu used for its implementation, as well as of the related
work. Section 3 will describe in general terms the proposed
algorithm and how it ts in. Section 4 describes some of the
performed experiments and displays their results. Finally,
section 5 o ers a brief summary as well as points some
possibilities of further work to be explored.
2. BACKGROUND AND RELATED WORK
To achieve the aforementioned goal, once the current state
of MenuMate is known, it is necessary to know what the
current state of the art for this speci c niche is.
Wasinger et al. have proposed an electronic menu with an
embedded recommender system, called Menu Mentor [
There are psychological studies about how to organize a
restaurant menu, with techniques such as placing items with
high prices rst to "smoothen" the e ect of the lower-priced
items following, even if they are not actually cheap [
The goal of this work was to develop an electronic menu system that adapts itself according to a certain target variable, such as session duration, but while avoiding the need both for the setup of a user pro le and extensive trials after each change. This is what will be presented in the next section. 3. ALGORITHM AND ARCHITECTURE The proposed idea is rather simple. It consists in reordering the menu items to optimize a quanti able target variable, that can be either maximized or minimized. For this work, four target variables were studied: tab value, session duration, revenue rate (cents per minute) and feedback rating. Once this de nition is set, as di erent menu sequences are used the value of each target variable is recorded for each session, as well as the position of each dish in the menu. The optimal menu sequence is computed by calculating the correlation coe cient between each dish's position and the performance, and sorting ascending or descending, depending on whether the action is to maximize or minimize, respectively.
Four correlation coe cients were tested: Spearman's [
The basic principle is as follows: after each session, the tablets send the raw data gathered to the system that generates the sequence (there is one such system per restaurant). This system, based on the chosen coe cient and target variable, calculates the value for that variable and the coe cients, generates the menu sequence, and at the beginning of each session the tablets poll this system for the latest sequence to be employed. This system, called Menu Optimizer, is con gurable in respect to the target variable, action (maximize or minimize) and correlation coe cient, together with other parameters relevant for the A/B testing performed, that will be explained in the next section. 4. PRELIMINARY USER STUDIES Since the method proposed does not explicitly recommend a single item or try to predict ratings, some methods usually employed to evaluate recommender systems cannot be employed, such as cross-validation, recall and precision and accuracy.
There are, however, other methods that can be employed
directly: the test was a double-blind A/B test, in which
neither the user nor the waiter knew which the target variable
at the time was, nor the correlation coe cient used. It was
also online (in the sense that real users were using the
system, generally spending real money through it [
Another adaptation made was with regards to serendipity. Instead of measuring it, there was the measurement of catalog coverage: the measurement of how much users tend to order varied dishes, compared to users of plain paper menu. For that, observations were made of tables whose guests did not use MenuMate, and which dishes they ordered, and those were compared to the orders of MenuMate users.
There were two classes of tests performed, the stress tests and the user tests. The stress tests, which will not be displayed in detail here, were used to assess the scalability of the system, which was developed to be run on a Raspberry Pi computer, with an embedded 900MHz ARM processor. It su ces to say that the system could serve between 100 and 160 tablets with unnoticeable performance losses, and that the limits reached were due to the test rig employed, rather than the system itself. Another interesting result on this front is that, on the employed hardware, menu optimization time is increased on average by 3ms for each session stored, which allows prediction of the optimization time based on the size of the history of observations.
The preliminary user studies were performed at two di erent restaurants in Munich, for about a week in each, time during which the system would gather session data and automatically adapt itself, switching the target variable at regular intervals. The number of observations was rather small due to the reduced number of days allowed for observations. There were 13 sessions observed in one restaurant, called El Patio, and 35 in the other, called Wendlinger. The restaurants had, respectively, 201 and 290 menu items, split in 24 and 34 categories. The original menu sequence was devised by their respective owners.
Figure 2 shows the evolution over time of the e ciency at Wendlinger (the number of sessions to which users gave feedback at El Patio was too low to assume it was meaningful). It suggests that indeed over time, independent of those being the target variables, feedback tends to rise and session duration to reduce, indicating good e ciency.
Table 1 shows how the choice of a target variable in uences the performance of that variable, as well as of other variables. In this table only sessions optimized with Spearman's correlation coe cient are used, because it was the only coe cient with a high enough number of observations from which to derive conclusions (21 sessions in total). The "none" line, represents results for a non-optimized menu. The system did well in optimizing for the target variable, with the best results for both feedback and session duration, a close second for tab value, while not delivering good results for revenue rate, because the higher tab value was not enough to compensate for the roughly halved session duration time achieved when the target variable was the duration. It is also worth noticing that indeed, revenue rate is a direct consequence of both tab value and session duration, but the average revenue rate is not necessarily the same as the revenue rate of the averaged tab values and durations. For clarity, the best results for each variable are displayed in boldface. Figure 3 shows another facet of the inner workings of the system: although the number of observed sessions was low, it is very fast to converge the internal coe cients, suggesting some stability after approximately 10 sessions. Another interesting e ect is that the absolute values of the coe cients tend to be higher for dishes that ended up being ordered, which means the system can in fact predict which dishes will be ordered, by checking the dishes with highest absolute value of the coe cients.
Due to operational constraints, catalog coverage was only measured at El Patio. There, 12 of the 13 sessions were with menu optimization enabled, that will be considered for this measurement. 50 sessions with paper menus were also
None Feedback
Tab Value Session Duration
Revenue Rate 2,0 5,0 4,4 4,3 5,0 25,17 13,25 38,00 36,85 40,10 12.865,6 15.779,9 7.499,5 2.406,9 4.400,6 20,9 33,8 59,1 154,9 56,6 observed. In total, 97 di erent dishes and drinks were ordered, in di erent quantities. From those 97, 19 were ordered both with and without MenuMate, 63 only by users of the paper menu and 15 only by users of MenuMate with menu optimization.
It is not easy to extrapolate how those proportions would be in case an equal number of observations was available, and discarding paper menu-based sessions could arbitrarily lead to any results. Assuming, however, that for both systems at each session there is an equal probability of adding not previously ordered items, and that this probability is inherent to either MenuMate or the paper menu, a proportion rule may be followed.
In the case of the paper menu, 19 + 63 = 82 di erent items were ordered in 50 sessions, which yields an average of 1,64 new items per session. With MenuMate in use, there were 19 + 15 = 34 di erent items ordered in 12 sessions, resulting in 2,83 new items per session. That may indicate that thus, the menu optimizations led to guests ordering a bigger variety of items. Alternatively, it could be that as more sessions would be measured, these sessions would progressively get less "innovative", in terms of the ordered dishes, which could revert this balance. An extended evaluation, with a similar number of sessions in both conditions would help settle this matter. 5. CONCLUSIONS AND FUTURE WORK The proposed menu optimizer harnesses principles of recommender systems to improve restaurant menus according to an arbitrary interest variable. In fact, it can be used to facilitate user interaction for any system to which access is anonymous and the number of items not overwhelming. Another use that comes to mind is the choice of which presentations to attend at a conference, or main sights to visit in a city.
One of the main contributions is the strong suggestion that this purely statistical method may improve the menu even if the underlying mechanism that drives the change is not understood by the system (i.e. the psychological implications). The developed system comprises the algorithm for menu optimization, coupled with a robust and scalable implementation of it. It was followed by qualitative and quantitative tests, with real paying users, of which only very few results were presented this time due to page number constraints, but that nevertheless suggest e cacy of the proposed method.
There are, however, some promising improvements to the method. Among which, the highlights are: Feature extraction, to allow dishes of di erent restaurants to be matched and correlation information exchange, possibly improving results; Extra variable isolation, that would allow control over in uential external variables such as weather, time of the day or season; If individual user pro ling is done, further personalization, such as ltering out dishes based on allergies or taste preferences, could be done; Stochastic exploration of dishes, which could allow for recommendation of the next course based on previously ordered dishes in a session.