Climbing is a popular sport and recreational activity. Unfortunately, there is a lack of technologies for supporting climbers in choosing what climbing route to climb next. We introduce a project aimed at developing a Climbing Recommender System for suggesting routes that are suited for training and practicing sport climbing. We model a climber by relying on both explicit and implicit feedback. Implicit feedback is acquired by an automatic activity recognition component (in climbing gyms), while explicit feedback is acquired by means of a mobile application. We also present a recommendation approach based on the prediction of the subjective evaluation of climbing routes' dificulty. In fact, often climbers perceive the dificulty of a route diferently from the oficial grade. The prediction method is based on the analysis of how climbers deviate their assessment of routes' dificulty from the oficial dificulty grade, and it generates explanations for the predictions.
important information about available climbing routes.
and this large set of alternative routes makes it dificult for any climber, either novice or expert, to make the right choice. There is a clear need for Climbing Recommender Systems (CRS) that could support climbers in choosing the most suitable next routes. Ideally, such a system should manage all the aspects considered by climbers in their choices. The recommended routes should be: suitable for training, challenging but within their capabilities, enjoyable, compatible with the contextual situation (weather conditions), and good for a group of climbers. Building such a system requires knowledge about climbers’ preferences, their physical and technical level, the group’s characteristics, and, importantly, the climbing routes’ characteristics, such as, dificulty grade, livered by means of a well-designed human-computer (mobile) interaction where climbers can leave explicit feedback about the routes they climbed. Moreover, sensor data (body and device sensors) should be used to implicitly detect climbers’ activities and performance. ing are at an early stage of development, and furthermore, nEvelop-O 3rd Edition of Knowledge-aware and Conversational Recommender Systems (KaRS) & 5th Edition of Recommendation in Complex profiling of climbers, and then we introduce a CRS for
Unfortunately, current sensing technologies for climb- yet been exploited in sport climbing, as there is a lack of location, and safety level. This can be acquired and de- a knowledge-rich user and item profile and then recomoutdoor climbing routes suggestion. 2. Routes Recommendation assigned by the climber to the route. This prediction can be used to recommend to the climber routes that are in the right dificulty range for her. In order to address this task, we have applied data mining techniques to explicit feedback data collected by means of the Vertical-Life mobile app where climbers can rate, grade, and comment on routes that they have climbed. Our work is motivated by Draper [24], who reports that climbers often have diferent opinions about the dificulty grade of the same route. The app ofers climbers the possibility to express their disagreement with the ‘oficial’ climbing grade of a route, as it is given in the guidebook. We have adopted a knowledge-based approach to understanding the reasons of such disagreements. We have compared the predictions, for the subjective evaluations of routes’ dificulty, generated by a regression model, based on features aimed at capturing why climbers disagree with the oficial grade, to a standard collaborative filtering algorithm.
We describe the results we have obtained in developing a knowledge-based CRS. Recommendations are generated by understanding users’ characteristics (profile), which are obtained by semi-automatically analyzing their behavior [19]. We consider two types of sources of climberrelated knowledge, which are acquired by means of either implicit or explicit feedback [20]. Implicit feedback is collected through sensor data analysis, in order to automatically detect the climber’s performed activities, i.e., which route was climbed and how the climber performed (e.g., the duration, or the efort made). Explicit feedback relates instead to climbers’ manual input, typically describing the experience of climbing a route, such as, the safety of the route, or the perceived dificulty. In order to collect this type of data, we are collaborating with Vertical-Life (www.vertical-life.info), which ofers a mo- 2.1. Route Grade Prediction bile application for climbers. We plan to augment their We focus our investigation on ascents performed on application with recommendation technologies. routes located in mountains and crags in the lead style;
Related to implicit feedback, we have developed some however, the approach described below can be extended initial solutions and technologies for activity detection, to ascents performed in climbing gyms and other climbwhich are currently designed only for indoor lead climb- ing disciplines, such as bouldering. ing. We have employed a ‘smart quickdraw sensor’ [21]. The data set of climbers’ entries into the mobile app A standard quickdraw is a piece of climbing equipment about their ascents is summarized in Table 1. The data set used by climbers to allow the climbing rope to run freely is restricted to the most frequently climbed route grades. through protection such as a bolt anchors, hence it is Grades are represented with integers ranging from 6 (5a) used for securing the climber at a specific point on the to 22 (7c). Observed deviations of climbers’ grades with climbing route. A smart quickdraw is a regular one aug- respect to routes’ oficial grades range from -3 to 3. It is mented with an accelerometer. The movements of the important to note that, on average, the climber’s grade rope, used for securing the climber, are propagated to the coincides with route’s oficial grade in 92% of ascents. quickdraw. Then, they are detected by the sensor and sent wirelessly to a computer for data analysis. By using this ‘smart quickdraw sensor’, we were able to detect with Table 1 an accuracy of 93% the activity of ‘rope pulling’, which Data set of climbers’ ascents description. ’% climber grades’ happens shortly after the climber finishes an ascent and raenfeerxsptliocitthgerapdeerceevnatlaugaetioofnatshcaetndtsifewrshferroemththeecloimficibaelrosngea.ve when she removes the rope from the wall. Recognizing rope pulling is instrumental in measuring the number of Outdoor climbs ascents made on a line equipped with this sensor, and it # climbs 157,576 can also be employed to distinguish beginners from ex- # climbers 2,624 pert climbers [22]. More sophisticated data analysis has # routes 10,738 also shown that the energy of the quickdraw movement % climber grades 8% can be used for climber’s performance measurement. In addition to this, we employed video cameras to detect The first personalized grade prediction method, which when the climber is lowered to the ground after her as- we call ‘knowledge-based’ uses a linear regression model, cent is finished [ 23]. We plan to adopt a similar solution for which we generated features representing climbersalso for outdoor climbing. routes interactions. The assumption is that the oficial
Explicit feedback data, includes instead the subjective dificulty of the routes, as defined by the route setters may evaluations of ascended routes’ dificulty grades, which also depend on their skills, while a climber’s perceived are often registered by climbers after they have climbed dificulty of the routes may depend on the climber’s physa route. By using this data we aim at predicting the sub- ical level and on contextual conditions (e.g., the season). jectively perceived dificulty of a route, which would be Therefore, we included the time factor, as the outdoor routes can change with the time (rocks might deteriorate, the climber-assigned grade of a route coincides with the equipment might break): we generate similar features oficial grade of the route. Hence, predicting correctly which take into consideration the specified time inter- the grades assigned by climbers when they deviate from val limit. As a result, we predict the climber’s perceived the oficial grade is challenging. grade as a linear dependent variable from the identified It is also worth noting that a major disadvantage of features. the SVD rating prediction model is that users and items
Assume that we have a target climber , and a route . are here represented in a joint latent factor space, which We are interested in the prediction ⋀(, , ) of how is hardly useful for explaining the predictions [26, 27]. In climber would grade the route at a particular point in fact, in this domain it is pivotal to properly convince the time . The predictive features that we have introduced climber about the reliability of the provided information: are as follows: there are important safety issues to consider.
We have measured the root-mean-square error (RMSE) • ( ) : the ‘oficial’ route dificulty grade. of the proposed models (linear regression knowledge• ( , ) : given the grade assessments of climbers based and SVD based collaborative filtering) and comwho previously climbed the target route, before pared them to the baseline. The results show that both time , we compute the average deviation of these models give a lower error than the baseline, meaning grades from the route’s oficial grade ( ) . that they are capable of correctly using climbers’ feed• ( , ) : this feature is a variant of feature back about dificulty level of routes (see Table 2). In the ( , ) such that only the final year and a half of table, ‘per user RMSE’ indicates the RMSE computed for data collection is included for feature computa- each user and then averaged, while ‘RMSE’ is the global tion. The feature is meant to capture change in average of all prediction errors. the dificulty of an ascent that may result from deterioration of the rock over time due to frequent Table 2 climbing activity (the so-called polished routes) Performance of perceived dificulty grade prediction on the or recent maintenance work. data set of climbers’ ascents on outdoor routes. • (, ( ), ) : given the grade assess- Model Outdoor routes ments of the target climber , before time , for RMSE per user RMSE routes of the same oficial grade ( ) , this feature expresses the mean deviation of these baseline 0.339 0.191 (± 0.306) grades from ( ) within a three-month pe- Lin. Regression KB 0.317 0.176 (± 0.284) riod in the final year of data collection. This fea- SVD CF 0.322 0.174 (± 0.284) ture is supposed to capture the influence of environmental conditions on a climber’s perception of route dificulty. For example, climbing outdoors 2.2. Explanations and GUI in the summer months is typically considered as harder than climbing in the spring.
a novel GUI of the Vertical-Life climbing application,
As we have noted above, climbers do not often devi- which in addition to the oficial grade shows the preate with their subjective evaluations of route dificulty dicted climber’s perceived grade. Additionally, we have from the oficial evaluations. Hence, a strong baseline employed the coeficients of the linear regression model method for predicting (, , ) is actually using di- to generate explanations to the climbers for the predicted rectly ( ) . grades of the selected routes. The linear regression model
In addition to this baseline prediction method, we com- along with its coeficients is shown in the following equapare the predictions generated by the above-mentioned tion: linear regression knowledge-based model with those ⋀ computed by a standard collaborative filtering (CF) al- (, , ) = 0.027 + 0.998 ⋅ ( ) + 0.410 ⋅ ( , ) gorithm. In the application of collaborative filtering, we + 1.051 ⋅ ( , ) + 0.279 ⋅ (, ( ), ) model the grade prediction task as a special rating predic- Clearly the oficial grade ( ) has the largest imtion problem, where the rating of a route is the deviation portance, but also the other features, related to the gradof the climber’s grade from the oficial grade, namely: ing behavior of the climbers, do have important weights (, , ) − ( ) . We use matrix factorization, in the model. The explanation sentence, which is comnamely singular value decomposition (SVD), to generate menting why the predicted grade is diferent from the such predictions [25]. We note that the rating matrix oficial grade, is generated by considering, case by case, is in this case very sparse (0.994 sparsity) and with a the features that, for a given route and climber combinaprevalence of 0 ratings: these are the evaluations where tion, have the largest positive (or negative) impact on the predicted grade deviation. Such an explanation may be useful if a climber would like to better understand why the predicted dificulty of the route is diferent from the oficial grade. Potentially, this can lead to fewer accidents, as the climber may choose routes that better match her capabilities.
Figure 1 shows an explanation example sentence given in the prototype app to the climber when she points to the perceived grade column by touching the screen on route 04.
that we plan to address in the future.
Firstly, we have developed sensor data analysis technologies for automatic activity recognition in climbing gyms, but we have not yet developed a similar solution for outdoor climbing. Moreover, knowledge extracted from low-level sensor data, e.g., the average speed of a climber during an ascent, has not yet been integrated with the explicit feedback collected from the mobile application. In fact, one related research question is how the climber’s skill level acquired by the implicit feedback can be used to improve the subjective route dificulty grade predictions that we have computed by relying only on explicit feedback.
Secondly, as we have mentioned in the introduction, climbing route recommendations may support diverse users’ needs. One important aspect is to identify routes that fit a specified training plan or to give explicit feedback to climbers about their ‘mistakes’ in trying the wrong routes. In fact, the training aspect is very important, as many climbers find that being able to track and achieve gradual progress is a crucial motivation for them. One possible approach to this problem is trying to intelligently revise or complete an existing climber’s training program [28, 29, 30, 31], which the climber typically stores in the app.
Thirdly, the implemented explanation component can be improved in order to give a more convincing explanation, not only of the predicted dificulty, but also of the recommended route [32]. Moreover, for motivational and training purposes, climbers sometimes repeat the routes which they tried, and specific explanations should be generated in these cases [33]. For instance, the system may argue: ‘This lead climbing route was climbed by you with 3 stops, try it again with fewer stops this time’. In fact, the specific rationale of a recommendation should be made clear to the climber. As a matter of fact, some routes are more enjoyable and should be recommended for climbers’ satisfaction; other routes are more important for training and motivation; other routes are relevant because they may better satisfy the needs of the group of climbers the target user belongs to.
Finally, we must properly evaluate the proposed system prototype, and understand whether such a CRS would be suitable and interesting for climbers. For this purpose, we have created an online survey [34] to collect climbers’ opinions on the proposed CRS.
In conclusion, in this paper, we have presented raw components and preliminary results that will be integrated into a novel CRS. We want to create a rich knowledge-based climber’s profile taking into consideration climber’s preferences, current physical level, behavior and skills. Such knowledge should be extracted from log data of the interaction of climbers with the routes that they have tried and evaluated. By better exploiting the bulk of knowledge contained in electronic guidebooks and climbers’ diaries, we aim at increasing climbers’ satisfaction but also their safety, as climbers will be supported to choose routes that are more aligned with their skills and expectations.
This work has been partly supported by the project “Sensors and data for the analysis of sports activities (SALSA)”, funded by the EFRE-FESR programme 20142020 (CUP: I56C19000110009). The authors thank Andrea Janes, Ben Lepesant and the Vertical-Life (https: //www.vertical-life.info/) company for the data provided for this research. [16] J. Pansiot, R. C. King, D. G. McIlwraith, B. P. L. ifeld, 2008.
Lo, Guang-Zhong Yang, Climbsn: Climber perfor- [30] M. L. Michailov, Workload characteristic, performance monitoring with bsn, in: 2008 5th Interna- mance limiting factors and methods for strength tional Summer School and Symposium on Medical and endurance training in rock climbing, Medicina Devices and Biosensors, 2008, pp. 33–36. Sportiva 18 (2014) 97–106. [17] A. Tonoli, R. Galluzzi, E. C. Zenerino, D. Boero, [31] T. O. Bompa, C. Buzzichelli, Periodization-: theory Fall identification in rock climbing using wearable and methodology of training, Human kinetics, 2019. device, Proceedings of the Institution of Mechanical [32] F. Ricci, L. Rokach, B. Shapira, Recommender Engineers, Part P: Journal of Sports Engineering systems: Introduction and challenges, in: Recand Technology 230 (2016) 171–179. ommender systems handbook, Springer, 2015, pp. [18] A. Bonfitto, A. Tonoli, S. Feraco, E. C. Zenerino, 1–34.
R. Galluzzi, Pattern recognition neural classifier for [33] K. Tsukuda, M. Goto, Explainable recommendafall detection in rock climbing, Proceedings of the tion for repeat consumption, in: Fourteenth ACM Institution of Mechanical Engineers, Part P: Journal Conference on Recommender Systems, 2020, pp. of Sports Engineering and Technology 233 (2019) 462–467.
478–488. [34] S. team, Poll for climbers ‘predicted perceived climb[19] M. De Gemmis, P. Lops, C. Musto, F. Narducci, G. Se- ing dificulty grades’, 2021. URL: https://forms.gle/ meraro, Semantics-aware content-based recom- YhKaCM3aqN8hAY557. mender systems, in: Recommender systems handbook, Springer, 2015, pp. 119–159. [20] I. Ivanova, Climber behavior modeling and recommendation, in: Proceedings of the 29th ACM Conference on User Modeling, Adaptation and Personalization, 2021, pp. 298–303. [21] I. Award, Winner | OUTDOOR Digital Services, https://www.ispo.com/en/ awards/ispo-award/2019/outdoor/winners/ vertical-life-smart-quickdraw, 2019. [Online; accessed 22-July-2021]. [22] I. Ivanova, M. Andric, A. Janes, F. Ricci, F. Zini,
Climbing activity recognition and measurement with sensor data analysis, in: Companion Publication of the 2020 International Conference on
Multimodal Interaction, 2020, pp. 245–249. [23] I. Ivanova, M. Andrić, S. Moaveninejad, A. Janes,
F. Ricci, Video and sensor-based rope pulling detection in sport climbing, in: Proceedings of the 3rd International Workshop on Multimedia Content
Analysis in Sports, 2020, pp. 53–60. [24] N. Draper, Climbing grades: Systems and subjectivity, The Science of Climbing and Mountaineering (2016) 227. [25] X. Ning, C. Desrosiers, G. Karypis, A comprehensive survey of neighborhood-based recommendation methods, Recommender systems handbook (2015) 37–76. [26] N. Tintarev, J. Masthof, Explaining recommendations: Design and evaluation, in: Recommender systems handbook, Springer, 2015, pp. 353–382. [27] Y. Koren, R. Bell, Advances in collaborative filtering
(2015) 77–118. [28] E. Horst, How to climb 5.12, Rowman & Littlefield,
2011. [29] E. Horst, Training for climbing: The definitive guide to improving your performance, Rowman & Little