Association rule mining is a technique for discovering relationships in large data sets and thus can obtain insights into process-related phenomena described by the data. In pig farming, necrosis (dead tissue) in the rearing period is an important phenomenon because it negatively afects animal welfare and reduces the share of usable young pigs. Pig rearing is a long-lasting process involving multiple stakeholders and management activities. Causes of necrosis are often unknown and their identification requires considerable time and efort. Pig rearing is a long-lasting process involving multiple stakeholders and management activities. Causes of necrosis are often unknown and their identification requires considerable time and efort. The objectives of this research are to (1) develop an association rule mining approach for generating plausible suggestions for cause-efect relationships of necrosis in pig rearing and (2) empirically evaluate the predictive power of the discovered rule set. We propose a procedure for generating and comparing association rules for diferent aggregation intervals. We used data from 672 pigs, collected over a ten-month period (Oct. 2018 - Apr. 2019 & Oct. 2019 - Dec. 2019). Association rules were created on the training set and tested on the test set. Association rules were evaluated using the metrics support, confidence, and lift, and expert knowledge. The association rules focused on temperature-related attributes and achieved confidence values between 0.65 and 0.99. Association rules suggest temperature and underlying processes as (contributing) causes of necrosis. Process experts provided knowledge that supports these suggestions and indicates their plausibility.
Association rule mining is a popular technique for creating relationships between attributes in
data sets. Therefore, association rule analysis can analyze process-related from diferent sources
to obtain insights into relevant process phenomena [
Pig rearing is often a seven-week process for increasing pigs’ weight (e.g., from 5 to 25 kg),
that involves multiple sub-processes, such as providing food or controlling the ventilation
and heating system in a pen (pigs’ environment). These sub-processes afect pigs and their
environment (e.g., air temperature) and can trigger conditions for necrosis development (e.g.,
tail biting due to stress) [
Previous work propose suggestions for causes should be based on association rules for a
training set and a test set (as commonly used for supervised machine learning tasks), diferentially
preprocessed process data, and the incorporation of process-related knowledge [
We propose the plausibility property to assess suggestions. We call a suggestion plausible if it is justified by process knowledge and based on underlying objective interesting association rules. Justification by process knowledge can be integrated into the analysis of the created association rules [16]. Objective interestingness is measured in this work using the metrics of support, confidence and lift. While lift is calculated on support, support and confidence are already necessary metrics for algorithm configuration with respect to minimum support and minimum confidence [ 17, 18]. Minimum support and minimum confidence influence the creation of association rules in terms of inclusion or exclusion of attributes and rules, while the algorithm decides how association rules are created [17, 18]. Consequently, association rules and their analysis depend mainly on the data preprocessing, e.g., the discretization of the continuous values. Against this backdrop, we address the following research question: How to design a procedure to create plausible association rules for suggestions of cause-efect-relationships in pig rearing? The main contributions of this paper are as follows: • We investigate association rule mining in terms plausible suggestions of cause-efect relationships using the example of pig farming. • We propose a procedure for creating association rules based on data from three aggregation intervals (hourly, daily, weekly) to serve as the basis for proposing causes. • We evaluate generated association rules by using an independent test set and justifications based on the process knowledge.
Our paper is structured as follows: In section 2, we provide background literature. In section 3, we describe the data set and procedure. In section 4, we analyze generated association rules by integrating process-knowledge, and in section 5, we provide a brief conclusion.
Association rule mining describes an approach to describing relationships in data by processing binary coded data to create association rules [19]. Given is a set = {0, . . . , } with ∈ N elements (e.g., is a pig). Each element , in turn, is a set with up to ∈ N attributes of the ⇒ Relationship means, is probably true if is given [20]. To assess this relationship, we use the metrics of support (supp), confidence (conf ), and lift. Supp describes how frequent a subset of attributes is in . Conf describes, how often an association rule is true. Lift describes if ∪ occur less or more often than expected. These metrics are calculated as follows [19, 21]: set . contains all possible attributes of . Thereby, has an attribute ∈ , if is true for (e.g., {} ∈ if a pig has necrosis). Given is a set ⊂ , a set ⊂ , and ∩ = ∅ we define an association rule as a relationship between the two sets as follows [19]: (1) (2) (3) (4) supp () = | { ∈ | ⊆ } |
|| conf ( ⇒ ) =
supp () supp ( ∪ ) lift ( ⇒ ) =
supp ( ∪ )
supp () · supp ( )
Lift ( ⇒ ) = 1 means and are independent, while lift ( ⇒ ) > 1 indicates a positive
correlation and lift ( ⇒ ) < 1 indicates a negative correlation. The process of association
rule generation consists of two steps. First: generation of a candidate set consisting of
several sets and or with minimal support (supp). Second, deriving association rules
based on with minimal confidence (conf). Supp afects || and thus the theoretically
number of association rules as well. The selection of supp can be set manually for all
attributes or for each individual attribute [22, 23]. In practice, however, there may be dificulties
in determining supp because of the tradeof between the number of association rules that
can be processed by humans and the loss of rules (and thus of suggestions for causes) [
[
We used process data from pig rearing provided by the Boxberg Teaching and Research Centre in Germany. The Boxberg Teaching and Research Centre is the central educational, experimental, and testing facility of the state of Baden-Wurttemberg in the field of pig farming. This data set describes the rearing process in the winter season between 10/18/2018 and 4/10/2019 and 10/10/2019 and 12/01/2019. A total of 672 pigs were reared in seven groups of 96 pigs each. Our data set consists of eleven pig-related and eleven environmental factors (figure 1). From our point of view, these factors represent events or results of the underlying processes.
Pig-related factors were weights, sex, daily gain between birth and start of rearing, mother,
father, age at the start of rearing, and the presence of necrosis and bloody tails. Weights, sex, daily
gain, age at start of rearing, and identification of mother and father were recorded manually. The
presence of pig necrosis and bloody tails was recorded during a weekly assessment according
to a uniform grading scale (see [24]). Environmental factors were temperatures, humidity,
brightness, and water consumption, recorded at five-second intervals, and daily dispensed food
and manipulable material. One part of the pen was covered by a false ceiling about 0.75 cm high,
while the larger part had a ceiling height of 2.9 meters. Thus, air and ground temperatures are
divided into temperatures under and outside the false ceiling. Our approach consists of analyzing
association rules created for three aggregation intervals of environmental factors: hourly, daily,
and weekly (figure 1) [
During the preprocessing, we performed the steps: (1) unification of labels and data formats, (2)
refilling environmental data, (3) removal of outliers, (4) creation of samples, (5) one hot encoding
of categorical features, and (6) creation of training and test sets. The source data of temperatures,
water, humidity, and brightness consisted of entries when a change was measured. Missing
values were filled in a five-second interval by forward fill. We detected outliers using lower
and upper limits set by experts. Outliers (e.g., temperature = 3, 000∘ ) are errors and have
been removed [
For the aggregation interval hourly, we calculated average values for temperatures, humidity, brightness and water for each hour in a week. We also calculated the diference between the flow and return flow temperatures, and the designated amount for food and manipulable materials on a daily basis. For the daily aggregation interval, we used the hourly data to calculate the n o it c e ll o c a t a D g n i s s e c o r p e r p a t a D g n iir n a T t s e T
aggregation interval hourly
aggregation interval daily
with Y={necrosis}, suppmin=0.3, confmin=0.4
minimum, maximum, and maximum diference (between two hours) for temperatures, brightness (without minimum), humidity, and the diference between flow and return flow temperatures for each day in a week. We also added the sum of water, food and manipulable material for each day. For the weekly aggregation interval, we calculated the minimum, maximum, and maximum diference between two hours for temperatures, brightness (without minimum), humidity, and the diference between flow and return flow temperatures in a week. We also added the weekly sum and maximum daily diference of water, food, and manipulable material.
Then we assigned continuous values to the corresponding classes with the value 1 (true) if
the continuous value is in the corresponding range, and 0 (false) otherwise (discretization into
binary values, table A) [
In step (5), one-hot coding was performed for sex and mother and father identifiers. In step
(6), we created our training and independent testing set based on the chronological start of
rearing of the groups used (see figure 3). Due to discretization, our training set consists of more
attributes than samples, which can lead to overfitting and potentially exponential growth of
possible association rules [
with random forest as estimator to select 150 features, reducing 50 per round. We also performed random undersampling to balance necrosis in the training set to specify a higher supp. A brief description of the training and test set is provided in table 1.
We used the FP-growth algorithm to create association rules because FP-growth does not
require candidate generation [
For testing, the created association rules were re-identified on the test set using the FP-Growth algorithm and supp = 0.1 and conf = 0.0000001. We re-identified these associations rules to discard rare associations rules in terms of support so that an association rule that only applies to one or two pigs is ignored. Association rules with the highest confidence on the training set tested on the test set are shown in tables 2 (aggregation level hourly), 3 (level daily), and 4 (level weekly). − corresponds to the number of days before an assessment.
Association rules in the hourly aggregation interval on the training set (table 2) focused on the flow temperature and air temperature based on the confidence level and frequency when the time period is not considered. Each association rule consisted of only one combination of the attributes air temperature 18∘ ≤ < 21∘ , flow temperature 20∘ ≤ < 25∘ , or dif. between flow and return flow temperature 0∘ ≤ < 3∘ . The highest confidence and lift on the test set is achieved by rule no. 1, while rule no. 3., which consists of one more temperature attribute in , achieves the highest confidence on the training set. According to rule no. 3, the temperature between flow and return flow temperature was < 3∘ , indicating the heating system emitted no or less heat. As a first conclusion, temperature (especially air temperature and flow temperature) and the corresponding heating system are candidates for suggestions on the causes of necrosis. These association rules are assigned to a specific hourly period. To support the first conclusion, we compare this preliminary conclusion with association rules of the daily and weekly aggregation interval.
Association rules at the aggregation interval daily on the training set focused on temperature attributes (see table 3). In particular, maximum air temperature 18∘ ≤ < 21∘ , and maximum air temperature under the false ceiling 24∘ ≤ < 27∘ , and return flow temperature 25∘ ≤ < 30∘ were frequent attributes. Rules no. 4 to no. 6 with the highest confidence level in the training set were not applicable to the test set. The attributes maximum air temperature 18∘ ≤ < 21∘ , minimum and maximum air temperature under the false ceiling 24∘ ≤ < 27∘ were also included in other rules during training in other periods, but reached a lower conf. No other process environment attributes such as brightness were present in the association rules of the training set. Although the rules in table 3 were not applicable to the test set and the confidence of the association rules in the test set was lower than in the training set, we conclude that the association rules on the aggregation interval daily support the preliminary conclusion that temperature is a possible cause.
At the aggregation interval weekly, the association rules (table 4) on the training set focused on the air temperature under the false ceiling, expressed by two association rules describing the minimum 21∘ ≤ < 24∘ ) and maximum 24∘ ≤ < 27∘ air temperature under the false ceiling. However, the rule no was not applicable on the test set. The air temperature itself occurs only in one association rule, which refers to the maximum diference between two consecutive hours (1∘ ≤ < 2∘ ), indicating that the air temperature was almost constant during one week. While another association rule consisted of the maximum diference in manipulable material output between two days, there are no other association rules that do not consist of temperatures. Also based on the association rules for this aggregation interval, we ifnd that temperature is a possible suggestion for the cause for necrosis.
Analysis of association rules per aggregation interval in both the training and test propose temperatures as a cause of necrosis in the pig’s environment. According to process experts, an air temperature of 18∘ < 21∘ indicates a low air temperature. During the study period, tests were conducted in the pens with diferent heating zones. Air temperatures in the areas under the false ceiling were warmer than in the areas without the false ceiling. Therefore, process-related information is available to support temperatures as a suggestion for the causes of necrosis. Process experts said in discussions that the temperatures in the association rules are quite plausible. Plausibility refers to subjective interestingness by including of processrelated knowledge [16, 14]. Thus, we conclude that association rule mining is able to create plausible association rules for use cases where there are few or no possibilities to identify causes of undesirable process outputs. Lack of specific sensors (e.g., necrosis-specific sensors), non-speaking process stakeholders (e.g., pigs), or an environment that limits reproducibility (e.g., diferent numbers of pigs afected by necrosis living in a similar environment) are examples of such reasons that make the root cause analysis dificult. Other possible applications include diseases, quality variations and yield fluctuations in agriculture or animal husbandry. To use our procedure in these use cases, the application-specific class adaptation in table A is necessary.
Temperature is associated with underlying processes, like heating or ventilation control. Thus, process-related causes could lie in the underlying processes. Although association rules establish a relationship between temperatures (or underlying processes) and necrosis, they describe no causality. The suggestion that temperature is a cause needs to be investigated in further experiments, e.g., similar to [12].
As in [
This work is subject to the following limitations. First, pig farming consists of sub-processes
that are carried out in an orderly structure. Association rules use sets of attributes and therefore
do not consider the order of characteristics of the underlying processes. To maintain the ordered
structure, sequential rule discovery is one approach [
This research explored association rule mining for plausible suggestions of process-related cause-efect relationships in pig farming. We used real data over a ten-month period and created association rules based on diferent intervals of data aggregation. Association rules pointing to temperatures and thus the underlying process regarding heat or air control as a suggestion for causes of necrosis. Moreover, there is process-related knowledge (e.g., diferent heating zones) that supports this suggestion, so we assume that association rule mining is able to provide plausible (justified by process knowledge) suggestions.
We thank the LABEL-FIT project team, especially Eva Gallmann, William Gordillo, Barbara Keßler, and Svenja Opderbeck for data collection and data provision. This work was supported by the project “Landwirtschaft 4.0: Info-System (Phase 2)”, funded by the Ministry for Food, Rural Areas and Consumer Protection of Baden-Wurttemberg, Germany. This work was also supported by the project LABEL-FIT by funds of the Federal Ministry of Food and Agriculture (BMEL) based on a decision of the Parliament of the Federal Republic of Germany via the Federal Ofice for Agriculture and Food (BLE) under the innovation support programme (2819200415). Joint Conference on Computer Science and Software Engineering (JCSSE), IEEE, 2013, pp. 206–210. doi:10.1109/JCSSE.2013.6567346. [11] B. U. Ayhan, N. B. Doğan, O. B. Tokdemir, An association rule mining model for the assessment of the correlations between the attributes of severe accidents, Journal of Civil Engineering and Management 26 (2020) 315–330. doi:10.3846/jcem.2020.12316. [12] X. Zhang, Y. Tang, Q. Liu, G. Liu, X. Ning, J. Chen, A fault analysis method based on association rule mining for distribution terminal unit, Applied Sciences 11 (2021) 5221. doi:10.3390/app11115221. [13] R. Agrawal, R. Srikant, Fast algorithms for mining association rules in large databases, in: Proceedings of the 20th International Conference on Very Large Data Bases, VLDB ’94, Morgan Kaufmann Publishers Inc, San Francisco, CA, USA, 1994, pp. 487–499. [14] A. Silberschatz, A. Tuzhilin, On subjective measures of interestingness in knowledge discovery, in: Proceedings of the First International Conference on Knowledge Discovery and Data Mining, KDD’95, AAAI Press, 1995, pp. 275–281. [15] G. Dong, J. Li, Interestingness of discovered association rules in terms of neighborhoodbased unexpectedness, in: X. Wu, R. Kotagiri, K. B. Korb (Eds.), Research and Development in Knowledge Discovery and Data Mining, volume 1394 of Lecture Notes in Computer Science, Springer Berlin Heidelberg, Berlin, Heidelberg, 1998, pp. 72–86. doi:10.1007/ 3-540-64383-4_7. [16] G. Piatetsky-Shapiro, C. J. Matheus, The interestingness of deviations, in: Proceedings of the 3rd International Conference on Knowledge Discovery and Data Mining, AAAIWS’94, AAAI Press, 1994, pp. 25–36. [17] Z. Zheng, R. Kohavi, L. Mason, Real world performance of association rule algorithms, in: F. Provost, R. Srikant, M. Schkolnick, D. Lee (Eds.), Proceedings of the seventh ACM SIGKDD international conference on Knowledge discovery and data mining - KDD ’01, ACM Press, New York, New York, USA, 2001, pp. 401–406. doi:10.1145/502512.502572. [18] J. Hipp, U. Güntzer, G. Nakhaeizadeh, Algorithms for association rule mining — a general survey and comparison, ACM SIGKDD Explorations Newsletter 2 (2000) 58–64. doi:10. 1145/360402.360421. [19] R. Agrawal, T. Imieliński, A. Swami, Mining association rules between sets of items in large databases, ACM SIGMOD Record 22 (1993) 207–216. doi:10.1145/170036.170072. [20] P. Hájek, I. Havel, M. Chytil, The guha method of automatic hypotheses determination,
Computing 1 (1966) 293–308. doi:10.1007/bf02345483. [21] S. Brin, R. Motwani, J. D. Ullman, S. Tsur, Dynamic itemset counting and implication rules for market basket data, ACM SIGMOD Record 26 (1997) 255–264. doi:10.1145/253262. 253325. [22] B. Liu, W. Hsu, Y. Ma, Mining association rules with multiple minimum supports, in: U. Fayyad, S. Chaudhuri, D. Madigan (Eds.), Proceedings of the fifth ACM SIGKDD international conference on Knowledge discovery and data mining - KDD ’99, ACM Press, New York, New York, USA, 1999, pp. 337–341. doi:10.1145/312129.312274. [23] F. Z. El Mazouri, M. C. Abounaima, K. Zenkouar, Data mining combined to the multicriteria decision analysis for the improvement of road safety: case of france, Journal of Big Data 6 (2019). doi:10.1186/s40537-018-0165-0. [24] K. Bönisch, A. vom Brocke, S. Dippel, A. Grümpel, L. Hagemann, C. Jais, D. Lösel,
Data Age (start of rearing) Birth weight
Brightness Control weight Daily increase
Food
Humidity Manipulable material Temperatures (air, ground)
Temperatures (flow, return flow)
Water Weight (start of rearing) Unit kg kg lx kg g kg kg kg % kg kg kg ∘ ∘ ∘ ∘ L L
L