Recommender systems are becoming essential tools in many scenarios, as they help users extract hidden knowledge and useful insights from datasets. In many real domains, the temporal order between events, combined with their contextualization, improves the accuracy of provided suggestions. In this paper, we introduce a framework designed to mine personalized, in time, contextual, and explainable sequential rules useful to provide recommendations for a predefined target parameter. Specifically, this framework is composed of the 3 Associative Classifier and LookBackApriori, a modification of Apriori algorithm. Our proposal takes historical data and contextual information as input and generates two sets of rules: the first set comprises rules that allow enhancement of the target parameter, and the second makes it worse. The proposed technique is applied to a real-world scenario involving data collected by Fitbit wearable devices to improve the user's sleep score after performing fitness activities in diferent contexts. The idea has been evaluated on two real datasets, and the results confirm the positive efects of the combination of 3 with LookBackApriori.
conditions in the user location and holidays.
More in detail, our use case is based on data gathered
The widespread popularity of sensors and wearable de- with Fitbit and focuses on suggesting the intensity of
vices, like smartwatches and fitness trackers, has in- physical activities and rest periods to carry out during
creased the amount of available data that can be lever- the current day to sleep better. This is done by mining
aged to monitor and enhance various aspects of their the historical contextualized physical activities during a
users’ well-being. Such devices are often equipped with specified temporal window, which represents the number
intuitive apps for activity tracking that mainly provide of consecutive observation days taken into account.
aggregate parameters and trend analysis, thus leaving We use two datasets consisting of activity logs from
room for more personalized and insightful suggestions to Fitbit wearable devices: PMDataset [
To achieve advanced insights, historical data, possi- specific information about the user context.
bly integrated with external information describing the The main aim of this paper is the construction of a
user context, needs to be analyzed for each user to ofer novel recommender system that combines the strengths
tailored and context-aware suggestions to improve their of two algorithms: the 3 associative classifier [
In this work, we propose a framework that aims to to predict a specific target parameter based on associative give personalized, and in time, contextual suggestions to classification. It takes as input all the historical physical a specific user to improve a target parameter (e.g., sleep activity and the related contextual information (i.e., the quality) along with an explanation about the provided context at the time the physical activity was performed) suggestions. To achieve this goal we integrate monitored and outputs the predicted sleep score. We leverage the data with contextual information, e.g., current weather second algorithm to provide an explainable recommendation about what activity to do and what to avoid to Published in the Proceedings of the Workshops of the EDBT/ICDT 2024 increase the predicted sleep quality and not decrease it. Joint Conference (March 25-28, 2024), Paestum, Italy With this framework, we overcome the limitations of * Corresponding author. the two algorithms and, in particular, the state explo†$Thaensneaa.duatlhlaovrsecccohnitar@ibuunteivdre.iqtu(Aal.lyD. alla Vecchia); sion problems of LBA when managing wide temporal niccolo.marastoni@univr.it (N. Marastoni); windows are less severe in L3. In addition, LBA allows elisa.quintarelli@univr.it (E. Quintarelli) the production of explainable recommendations; indeed, 0000-0001-7026-5205 (A. Dalla Vecchia); 0000-0001-6988-1203 since LBA is based on Apriori, it mines sequential rules (N. Marastoni); 0000-0001-6092-6831 (E. Quintarelli)
© 2024 Copyright for this paper by its authors. Use permitted under Creative Commons License that contain in their antecedent the explanation of the CPWrEooUrckReshdoinpgs IhStpN:/c1e6u1r3-w-0s.o7r3g ACttEribUutRion W4.0oInrtekrnsahtioonpal (PCCroBYce4.0e).dings (CEUR-WS.org) provided sleep score present in the consequent.
OutC
R+ RThe framework proposed in this paper combines an associative classifier, 3, to predict the value of a target parameter (e.g., the sleep score for the current day, the stress level) and an algorithm based on Apriori, called LookBackApriori (LBA), to generate timely and explainable contextual recommendations helpful to suggest what to do to improve the predicted value (i.e. the fitness activities to undertake in the current day to increase the sleep score, whenever it is possible).
The rule consists of a sequence of itemsets, each
representing either fitness activities ( ), sleep quality (),
or both (), for a specific day, where 0 represents the
current day. The parameter defines the temporal
window, that is, the number of consecutive days that the
Associative classifier The first part of our framework algorithm can consider and thus may be present in a rule.
comprises the 3 associative classifier [
Then, the classification of unlabeled data is executed activity data is present for the current day. Indeed, this in two steps: first, by considering a subset of high-quality itemset, 0 , represents the physical activity suggested to rules for the classification process, and second, by adding the user by our framework to improve their sleep score a larger set of rules when it fails to find rules for certain for the same night, represented in the consequent as 0. data points. The associative classifier 3 is used for historical data
In the green section of Fig. 1, we show the classifier processing and prediction to address memory-related iscomponent, which is employed for predicting the value sues faced by LBA that stem from the size of the input. of a target parameter, represented in our scenario by the After the sleep score prediction step, LBA is used to prosleep score for the current day. It takes historical data vide explainable positive and negative recommendations (i.e., the user’s log of their fitness activities and sleep thanks to the form of the mined rules. To this end, the scores) as input and integrates it with past contextual antecedent contains the sequence of events that will lead information, including the context of the current day at to the sleeping score in the consequent; thus, it provides the time of the prediction. an explanation for the recommendation.
Three possible rules mined by the LBA Algorithm are the following: : − ( − 1) ∧ · · · ∧ − 2 ∧ − 1 ∧ 0 → 0 [, ] 1 : { : 3, : 2}− 1∧{ : 3, : 2}0 → { : 1}0 2 : { : 2, : 3}− 2 ∧ { : 1}− 1 → { : 3}0 3 : { : 2, : 3}− 2 ∧ { : 1}0 → { : 3}0 1 states that if yesterday the user performed a high level of heavy physical activity ( : 3) and a medium level of light activity ( : 2), and today they perform the same activities, the resulting sleep score will have a low value ( : 1). 2 is an incomplete rule since it does not contain information about the physical activity the user should perform during the current day. Although the rule is valid, it is not helpful for making a recommendation to improve sleep quality, as it does not have any itemset labeled 0 in the antecedent. 3 is another incomplete rule, but it gives us information about the physical activity the user should do during the current day to sleep well; thus, it can be used to provide a recommendation.
In the orange part of Fig. 1, we show the contribution of LBA to our framework. Firstly, it mines a set of rules using the Rule Generator, which takes as input the current context, also used by 3 for the prediction step, and data at the time of the prediction. Secondly, taking advantage of the label produced by the classifier, the rules generated are split into two sets: those that improve the target parameter value w.r.t. the predicted one is labeled +, and those that do not are labeled − . Then, the rules in both sets are ordered according to the completeness of the rule antecedent, confidence, and support. The rules recommended to the user are the most complete ones with the highest confidence and support.
Our experiments focus on wearable device data: their logs contain information about daily physical activity levels and sleep scores. Whenever possible (i.e., when we have enough data about the user), we integrate such logs with additional information, e.g., holidays, day of the week, and weather conditions related to the user’s location, to better contextualize the gathered fitness and sleep quality data.
We consider two datasets for this domain: PMdata [
The Custom dataset was collected from 4 users specifically for this study, the earliest of which started recording in August 2021 and ended in September 2022. The participants are evenly split between males and females; their ages vary from 16 to 55.
From both datasets, we make use of the logs about
"light", "medium", and "heavy" activity, along with rest
periods and the sleep score for each day. Fitbit records
these features as minutes spent in each activity type; thus,
we discretize them to obtain categorical data as described
in [
Regarding the context, for both datasets, we also have information on whether a day falls on a weekend ( ) or not ( ).
In addition, for the Custom dataset, we integrate the information about the user’s vacations ( / ) and the weather conditions. For this last aspect, we have simplified its representation as follows: if there has been rain, snow, fog, or other bad weather conditions, the label is . In all the other cases, it is . Regarding the temperature, we use the average daily temperature as a feature, and the result is labeled into or , using the yearly average temperature as a threshold. 3.1. Practical example t-3 t-2 t-1
t0 R t0 t-3 t-2 t-1
t0
Rule Generator t0
Ordering and Splitting
R Training
Associative Classifier
t0 R which improve R which deteriorate t0 t0 t0 t0 classifier then returns a sleep score for the current day.
The rules generated during the training phase state correlations that are only related to the current day, and are in the form shown in the purple rectangle, i.e., contextual information together with physical activity in the antecedent of the rule and the sleep score in the consequent.
Some examples of mined rules are the following: 1 : {, , , : 1, : 2}0 → { : 1}0 2 : {, , : 2, : 3}0 → { : 3}0 • testing the eficacy of the proposed framework on 3 w.r.t. to its ability to predict the value of the predefined target parameter.
For all the experiments, we have reserved the first 80% of the Fitbit logs of each user as a training set and the remaining 20% of the data for testing. Due to the sequential nature of the problem, we cannot randomize the sampling of the two sets. Thus, we maintain the sequential order based on the timestamp of the logs and select the last 20% of the dataset for the tests.
Rule 1 tells us that on a weekday with clear weather, if the user performs low levels of heavy activity ( : 1) 4.1. Relevance of historical data and medium levels of light activity ( : 2), their sleep To conduct this part of the experiments, we use the 3 score for the same night will be low. Whereas rule 2 associative classifier to predict the sleep label related to states that, in the case of a stormy weekday, the user will the current day _0. sleep well after performing medium levels of light activity First, we perform the prediction by selecting only the ( : 2) and high levels of medium activity ( : 3). physical activity and context of the current day as input
Thanks to the sleep score obtained by the associative to the classifier without considering historical data. An classifier, it is possible to split the rules into those that example of input data for a Custom user is: increase or decrease the predicted sleep score. At this stage, the user can explain why their sleep quality may (_0, _0, _0, _0, _3_0, improve or not by looking at the antecedent of the rules. _1_0, _1_0, _2_0)
To test the validity of our framework, we have performed several experiments on the following aspects: • verifying the relevance of historical data and their context in the ability to predict the value of a target parameter, i.e., the sleep score. • testing the performance of the two algorithms used by the framework w.r.t. execution time and memory consumption.
This can be interpreted as: on clear weather and hot weekdays during a holiday, the user performs a high level of light physical activity, low levels of both medium and heavy activity, and a medium level of rest.
Then, we add the historical data (i.e., physical activity, context, and sleep score of the past days) to the input used for the first set of experiments.
Fig. 3 depicts the recorded accuracies of these two experiments for each user in both PMdata and Custom datasets. These results show that for 77% of users, using historical data instead of only using data from the current Fig. 6 and Fig. 7 depict the elapsed time and the memday improves the accuracy of the classifier. ory consumption of the framework when using data from
Fig. 4 shows that, regardless of the presence of con- one of the users of the Custom dataset, showing the imtextual information, having historical data improves the portant contribution of 3. During the experiments, the accuracy of most users. The performance of the algo- input is the physical activity log and the available conrithm decays quickly as the temporal window increases, textual information of the chosen user. We maintain the especially in the absence of contextual data. Thus, it same support and confidence and gradually increase the seems clear that sleep quality does not depend on data temporal windows, i.e., increase the historical data given that is temporally distant from the current day. in input. LBA shows an exponential trend for memory consumption and time elapsed, until its memory allocation fails when the temporal window reaches value 4.
On the other hand, 3 can manage at most a temporal window of 6, maintaining a relatively constant trend.
Additionally, the accuracy value obtained by some users increases gradually in the presence of contextual information as the length of the temporal window increases. One example is shown in Fig. 5, where we can also confirm that sleep does not depend on the activities performed six days earlier. We can also observe that contextual information improves the final result.
Accuracy of the classifier for user p03 with different temporal windows
Without Context With Context 300 ) (se200 m i T 100
0 3500 3000 4.2. Time and memory performance Despite LBA’s intrinsic capability to provide explainable recommendations for achieving a better sleep score, we still employ the associative classifier 3 to process historical data. Thanks to the associative classifier, we can process significantly larger volumes of data, e.g., longer sequences of data with their contextual information, without the risk of memory errors. Additionally, 3 is significantly faster in obtaining the results. 4.3. Evaluation of the framework The last set of experiments performed is on the complete framework, as explained in Sec. 3.1. The idea is to validate, on real user logs, the prediction of 3 enriched with the recommendation produced by LBA. In order to validate the results of the whole framework, the best approach would be to ask for inputs directly from the users.
Due to time constraints, the evaluation of the framework is strictly empirical. 0.4
As before, we use the first 80% of the dataset to train both 3 and LBA. The first step is setting the length of the temporal window, which is done empirically by With the spread of smart devices and the availability of analyzing the experiments in the previous case study. their large datasets, we have the possibility to extract The input of the classifier is composed of the physical both explicit and implicit knowledge about monitored activities, contextual information, and sleep score of the parameters. For this reason, there are many intelligent past days and the contextual information of the current techniques proposed in the literature to improve the cusday. The classifier then predicts the sleep score on the tomization of data exploitation. Recommender Systems current day. (RS) ofer suggestions on items, services, or news that
Separately, the rule generator produces a set of rules may interest users and afect their decisions based on
correlating the physical activities and contextual infor- their profile, history, and preferences [
At this point, a trained classifier and a set of rules exist tions starting from data collected by a Fitbit wearable
for each user. To test the complete framework, we take device. The physical activity information collected by
from the remaining 20% of the dataset a temporal window Fitbit is also used in [
In Fig. 8, we report both the accuracy of the classifier, able sleep model that exploits the correlation between without any knowledge about physical activity for the daily activities and sleep quality, providing recommendacurrent dayIt can be noted that the accuracy of 3 for tions to improve sleep quality. While the outcome of this most users is less than 0.45. framework aligns closely with our proposal, the approach
The output of 3 produced is used as a threshold to does not account for sequences of events that occurred separate the rules mined by the Rule Generator in the in the days leading up to the prediction intended for the tdwatoiosnetss, re+spaencdtive− ly.oDfpuoestitoivteheanndatnuergeatoifvethreecpormobmleemn- puoserra.teFuexrttherenrmalocroen,ttehxetumaol dinefloprrmesaetinotnedbefayiolsn dtoseinncsoerdat hand, it would not be accurate to use historical data to humidity and temperature. As highlighted in Subsection check whether the recommendation given by the frame- 4.1, historical data are important to improve the quality work will actually result in a change in sleep score. of provided predictions.
In the state of the art, contextual information is often integrated into RS to improve the precision of recommendations. In general, user preferences may vary depending on the environment and the situation in which they are
In this paper, we have combined the 3 associative classifier and the LBA algorithm to provide explainable recommendations. Such an approach helps give insights to users who desire to know what afects their sleep score
M. Smith, P. Steggles, Towards a better under- [24] M. Alshammari, O. Nasraoui, S. Sanders, Mining standing of context and context-awareness, in: semantic knowledge graphs to add explainability H. Gellersen (Ed.), Handheld and Ubiquitous Com- to black box recommender systems, IEEE Access 7 puting, First International Symposium, HUC’99, (2019) 110563–110579.
Karlsruhe, Germany, September 27-29, 1999, Pro- [25] R. Guidotti, A. Monreale, S. Ruggieri, F. Turini, F. Giceedings, volume 1707 of Lecture Notes in Computer annotti, D. Pedreschi, A Survey of Methods for Science, Springer, 1999, pp. 304–307. Explaining Black Box Models, ACM Comput. Surv. [12] G. Adomavicius, R. Sankaranarayanan, S. Sen, 51 (2018). doi:10.1145/3236009.
A. Tuzhilin, Incorporating contextual information [26] R. Meo, R. Nai, E. Sulis, Explainable, Interpretable, in recommender systems using a multidimensional Trustworthy, Responsible, Ethical, Fair, Verifiable approach, ACM Trans. Inf. Syst. 23 (2005) 103–145. AI... What’s Next?, in: In Proceedings of 26th [13] G. Adomavicius, B. Mobasher, F. Ricci, A. Tuzhilin, European Conference on Advances in Databases Context-aware recommender systems, AI Mag. 32 and Information Systems, ADBIS, 2022, pp. 25–34. (2011) 67–80. doi:10.1007/978-3-031-15740-0_3. [14] N. M. Villegas, C. Sánchez, J. Díaz-Cely, G. Tamura,
Characterizing context-aware recommender systems: A systematic literature review, Knowledge
Based Systems 140 (2018) 173–200. [15] P. Fournier-Viger, J. C.-W. Lin, R. U. Kiran, Y. S. Koh,
A survey of sequential pattern mining, Data Science and Pattern Recognition 1 (2017) 54–77. [16] J. Pei, J. Han, B. Mortazavi-Asl, J. Wang, H. Pinto,
Q. Chen, U. Dayal, M. Hsu, Mining sequential patterns by pattern-growth: The prefixspan approach,
IEEE Trans. Knowl. Data Eng. 16 (2004) 1424–1440. [17] R. Srikant, R. Agrawal, Mining sequential patterns:
Generalizations and performance improvements, in: P. M. G. Apers, M. Bouzeghoub, G. Gardarin (Eds.), Advances in Database Technology - EDBT’96, volume 1057 of Lecture Notes in Computer Science,
Springer, 1996, pp. 3–17. [18] B. Ma, B. Liu, Y. Hsu, Integrating classification and association rule mining, in: Proceedings of the fourth international conference on knowledge discovery and data mining, volume 50, 1998, pp.
55–00. [19] J. Li, G. Dong, K. Ramamohanarao, L. Wong, Deeps:
A new instance-based lazy discovery and classification system, Machine learning 54 (2004) 99–124. [20] W. Li, J. Han, J. Pei, Cmar: Accurate and eficient classification based on multiple class-association rules, in: Proceedings 2001 IEEE international conference on data mining, IEEE, 2001, pp. 369–376. [21] X. Yin, J. Han, Cpar: Classification based on predictive association rules, in: Proceedings of the 2003 SIAM international conference on data mining, SIAM, 2003, pp. 331–335. [22] X. Zhang, G. Dong, K. Ramamohanarao,
Information-based classification by aggregating emerging patterns, in: International Conference on Intelligent Data Engineering and Automated
Learning, Springer, 2000, pp. 48–53. [23] Y. Zhang, X. Chen, Explainable recommendation:
A survey and new perspectives, Found. Trends Inf.
Retr. 14 (2020) 1–101. doi:10.1561/1500000066.