Digital Advertising Traffic Operation: Machine Learning for Process Discovery

Massimiliano Dal Mas me @ maxdalmas.com

ABSTRACT In a Web Advertising Traffic Operation it’s necessary to manage the day-to-day trafficking, pacing and optimization of digital and paid social campaigns. The data analyst on Traffic Operation can not only quickly provide answers but also speaks the language of the Process Manager and visually displays the discovered process problems. In order to solve a growing number of complaints in the customer service process, the weaknesses in the process itself must be identified and communicated to the department. With the help of Process Mining for the CRM data it is possible to identify unwanted loops and delays in the process. With this paper we propose a process discovery based on Machine Learning technique to automatically discover variations and detect at first glance what the problem is, and undertake corrective measures.

Keywords: Process Mining, Business process management, Declarative process models, Digital Advertising, Online Advertising, Trafficking, Adv Operations Management, Mixed Integer Linear Programming (MILP), Optimization, Scheduling, Stakeholders, Project Management, End-To-End management, Campaign Workflow, Analytical Skills, People management skills, Relationships, CRM, Operations, Ad Platforms, Creative delivery, Campaign Performance, KPIs

INTRODUCTION to ensure quick and accurate trafficking In a Web Advertising Traffic Operation of campaigns, campaign performance it’s necessary to ensure quick and accurate optimisations, troubleshooting, and trafficking of campaigns, campaign reporting and that campaigns launch on time performance optimizations, troubleshooting, and are fully executed according to the and reporting. Ensure that campaigns launch insertion orders. Different processes are on time and are fully executed according to involving different activities with the insertion order. [1] advertisers, agencies, and publishers to To deliver digital campaigns it’s necessary oversee proper implementation of campaigns (ad tags, beacons, reporting ______discrepancies, click tracking, etc.) with the retail client as the technical digital media If you have read this paper and wish to be included in a mailing list (or other means of expert. communication) that I maintain on the subject, Moreover it is necessary to maintain than send e-mail to: me @ maxdalmas.com proactive communications on account status across multiple account stakeholders actually going on (especially if the products including Account Managers, Sales are services or data). Managers, Senior Management and customer contacts ensuring immediate and Process Mining direct feedback to both internal and external Process mining combines process models stakeholders where necessary monitor and and event data in various novel ways. As a improve customer satisfaction as it relates to result, one can find out what people and successful campaign management. organizations really do. For example, process models can be automatically Usually Traffic Managers create document discovered from event data. Compliance can on operations best practices policies and be checked by confronting models with procedures, including an I/O trafficking event data. Loops can be uncovered by SLA but trafficking is an ongoing process. replaying timed events on discovered or Unlike products, processes are less tangible. normative models. Hence, process mining Processes may only exist in the minds of can be used to identify and understand people and it is difficult to “materialize loops, bottlenecks, inefficiencies, deviations, processes”. and risks. In this paper we define a process mining techniques for the trafficking aim to Machine Learning for Process Mining discover, monitor and improve real As part of Artificial Intelligence we can processes by extracting knowledge from define Machine Learning as the field that event logs. “gives computers the ability to learn without We define a technique that automatically being explicitly programmed” [2]. discover a process model from event data on The field of Machine Learning, which aims a Traffic Office. to develop computer algorithms that improve with experience, holds promise to THE BASIC IDEA enable computers to assist humans in the Processes can be defined from human analysis of large, complex data sets [3 – 12]. behavior partly controlled by procedures While Machine Learning collects past and/or information systems. Normative or information to learn, it is a fact that the past descriptive models may be used to document never predicts the future. processes. But, these models are human- It is mirrored in the “Process Mining” made artifacts that could not necessarily say approach that assumes that collecting more much about the “real” process performed. data on more business processes makes them As a result, there may be a disconnection more predictable. Process Mining is not a between model and reality. Moreover, the reporting tool, but an analysis tool. It people involved in a process are typically enables you to quickly analyze any and very unable to understand the corresponding complex processes. Process mining process model and cannot easily see what is techniques aim to discover, monitor and improve real processes by extracting Given an event E consisting of a collection knowledge from event logs. of activities a, we could construct a Petri net that “adequately” describes the observed Discovery Algorithm behavior. The process discovery algorithm here We define the discovery algorithm as a presented use a Machine Learning as function d (2) that maps any event E (3) Process Mining from example behavior considering a Workflow Net WN (4) and an recorded in an event log E. interest factor f of event E (1). This paper is grounded in Linear Temporal

Logic (LTL) used to specify constraints on (2) dBu∈→ u ( A ) WN the ordering of activities [13]. We consider an event as composed by different activities (3) E ∈⊆BA( ) ; A uA that can require different times.

Different activities involved in a Traffic (4) A WN=∈ aσσ| ∈ E Operation can involve different stakeholders i ( ) { }

(as: advertisers, agencies, publishers, clients, Where B(A) is the set of all the activities and etc.) having the “working time” as common is u the set of the activities involved in the factor. WN Workflow Net WN. For our purpose we define an interest factor Given an interest factor f containing events f as a ratio between the time spent for an referring to a set of activities A as in (3), we activity (a) and the time spent for an decompose discovery by distributing the antecedent activity (a^-1) and the activities over multiple sets A1, A2, . . . , An. consequent activity (a^+1) as show in (1) The same activity may appear in multiple sets as long as A = A1 U A2 U. . .U An. a For each activity set Ai, we discover a (1) f = −+11 aa× Workflow Net WNi (4) by applying a For an event E a confidence of the constraint discovery algorithm to interest factor f|Ai , is 1 in a stronger dependency between the i.e., the overall interest factor projected onto antecedent and the consequent activities to a subset of activities. By combining the reflect the real correlation between the resulting Workflow Nets we obtain an activities [14]. overall Workflow Net WN = WN1 U WN2 We can use the interest factor f to prune U. . . U WNn describing the behavior in the processes, e.g., consider only those overall event E. constraints f above a minimum confidence threshold. A stronger dependency between Discovery Loops and delay the antecedent and the consequent is It is possible to identify unwanted loops in associated with a value of this measure f that the process defining the set of activities in is further from 1 – cause the activity a is the Workflow Network WN for multiple sets waiting the antecedent to be completed [15]. as long as A = A1 U A2 U. . .U An as defined below in (5) on networks with a few thousands arcs in aa∪−∪−+ aa ( iiii) ( ) less than five seconds of CPU time on a PC. (5) loop = a − We then tested the model's effectiveness i using subset of “specification required” by RELATED WORKS the main worldwide industry leading Discovering a process model from a multiset Advertising Agency. of traces is a very challenging problem and The fast numerical algorithms described various discovery techniques have been above render the relative optimization proposed [16 - 27]. Different approaches are problem treatable. This procedure is based used and it is impossible to provide an on the comparison between measured data exhaustive overview of all techniques. The and the simulated solutions produced by the main recent techniques are: heuristics [20, algorithm, and gives an average percentage 21], inductive logic programming [22], error of about 13%. state-based regions [20, 23], language-based regions [25, 26], and genetic algorithms CONCLUSION [27]. In the future, we intend to test the calibration procedure on a more completed EXPERIMENT workflow network WN, and to construct a We have built a simulation model real-time interface for the automatic reproducing the behavior of event E. representation and visualization of critical For our own interest we did that independent activities on the workflow to detect at first research projects also through specific glance what the problem is, and undertake innovative tasks validate towards average corrective measures. working time declared on “specification Massimiliano Dal Mas is an engineer working on required” by the main worldwide industry webservices, trafficking and online advertising and is leading Advertising Agency. interested in knowledge engineering. In the last years he http://www.sizmek-sea.com/Spec/ had to play a critical role at Digital Advertising business, cultivating relationships with key publisher partners with http://support.adform.com/...specifications/ experience managing a team. Been responsible for all day general-specifications/ to day operations with partners and consult on the best http://creative-weborama.com/category/ ways to monetize their properties. His interests include: uncategorized/ user interfaces and visualization for information retrieval, automated Web interface evaluation and text analysis, http:// groupm.dk/Banner-sizes-and- empirical computational linguistics, text data mining, specifications knowledge engineering and artificial intelligence. He received BA, MS degrees in Computer Science Engineering The simulator prototype can describe the from the Politecnico di Milano, Italy. He won the evolution of activities over time taking thirteenth edition 2008 of the CEI Award for the best interactions at junctions into consideration. degree thesis with a dissertation on "Semantic We have investigated the numerical validity technologies for industrial purposes" (Supervisor Prof. M. Colombetti). In 2012, he received the best paper award at of the approximation algorithm. It has the IEEE Computer Society Conference on Evolving and proven to be very fast, providing solutions Adaptive Intelligent System (EAIS 2012) at Carlos III University of Madrid, Madrid, Spain. In 2013, he received the best paper award at the ACM Conference on Web Intelligence, Mining and Semantics (WIMS 2013) at [6] M. Dal Mas, “Elastic Adaptive Universidad Autónoma de Madrid, Madrid, Spain. His Ontology Matching on Evolving paper at W3C Workshop on Publishing using CSS3 & Folksonomy Driven Environment” in HTML5 has been recommended as position paper. Proceedings of IEEE Conference on Evolving and Adaptive Intelligent References System (EAIS 2012), Madrid, Spain, 35-40, IEEE, [1] M. Dal Mas, “Digital Advertising (http://ieeexplore.ieee.org/xpl/articleD Traffic Operation: Flow Management etails.jsp?arnumber=6232801) Analysis”. In proceeding for DOI: 10.1109/ EAIS.2012.6232801 pubblication [7] M. Dal Mas, “Intelligent Interface [2] P. Simon, “Too Big to Ignore: The Architectures for Folksonomy Driven Business Case for Big Data”. Wiley. Structure Network” in Proceedings of p. 89. ISBN 978-1-118-63817-0 the 5th International Workshop on (2013) Intelligent Interfaces for Human- [3] M. Dal Mas, “Hyperelastic-Based Computer Interaction (IIHCI-2012), Adaptive Dynamics Methodology in Palermo, Italy, 519– 525, IEEE, 2012 Knowledge Acquisition for (http://ieeexplore.ieee.org/xpl/articleD Computational Intelligence on etails.jsp?arnumber=6245653) DOI: Ontology Engineering of Evolving 10.1109/CISIS.2012.158 Folksonomy Driven Environment”. [8] M. Dal Mas, “Elastic Adaptive Sentiment Analysis and Ontology Dynamics Methodology on Ontology Engineering An Environment of Matching on Evolving Folksonomy Computational Intelligence, 2016 Driven Environment”. Journal 639/2016; 639:141-173 Springer DOI: Evolving Systems 5(1): 33-48 (2014) 10.1007/978-3-319-30319-2_7 (http://link.springer.com/article/10.10 (http://link.springer.com/chapter/10.1 07%2Fs12530-013-9086-5) 007%2F978-3-319-30319-2_7 ) [9] L. Mussel, “Conformazione e struttura [4] M. Dal Mas, “Layered Ontological di bio-polimeri”. Dip. di Scienze Image for Intelligent Interaction to Chimiche – Università di Padova Extend User Capa-bilities on Padova Digital University Archive, Multimedia Systems in a 1964. Folksonomy” Driven Environment. [10] I. Vragovic, E. Louis, A. Diaz- Smart Innovation, Systems and Guilera, “Efficiency of information Technologies 01/2015; 36:103-122 transfer in regular and complex Springer DOI: 10.1007/978-3-319- networks.”, Phys Rev E 71:026122, 17744-1_7 2005 (http://link.springer.com/chapter/10.1 [11] U. Brandes, T. Erlebach, “Network 007%2F978-3-319-17744-1_7 ) analysis: methodological [5] M. Dal Mas, “Folksodriven Structure foundations”. Lecture notes in Network”. Ontology Matching computer science, vol 3418. Springer, Workshop (OM-2011) collocated with New York, 2005 the 10th International Semantic Web [12] G. Bretti, R. Natalini, B. Piccoli, “A Conference (ISWC-2011), CEUR WS Fluid-Dynamic Traffic Model on vol. 814 (http://ceur-ws.org/Vol-814), Road Networks”, Archives of 2011 Computational Methods in Serebrenik, “Process Discovery using Engineering 14 (2007), 139-172; Integer Linear Programming”. [13] E. Damaggio, A. Deutsch, R. Hull, Fundamenta Informaticae, 94:387– V.Vianu, “Automatic verification of 412, 2010. datacentricbusiness processes”. In: [21] J.E. Cook and A.L. Wolf, BPM 2011. pp. 3–16 “Discovering Models of Software [14] S. Brin, R. Motwani, C. Silverstein, Processes from EventBased Data”. “Beyond Market Baskets: ACM Transactions on Software Generalizing Association Rules to Engineering and Methodology, Correlations”. In: ACM SIGMOD 7(3):215–249, 1998. 1997. pp. 265–276 [22] S. Goedertier, D. Martens, J. [15] R. Haberman, “Mathematical Vanthienen, and B. Baesens, “Robust models.” Prentice-Hall, Inc. New Process Discovery with Artificial Jersey, 1977, 255-394. Negative Events”. Journal of Machine Learning Research, 10:1305–1340, [16] W.M.P. van der Aalst, V. Rubin, H.M.W. Verbeek, B.F. van Dongen, 2009. E. Kindler, and C.W. G¨unther, [23] J. Carmona, J. Cortadella, and M. “Process Mining: A Two-Step Kishinevsky, “A Region-Based Approach to Balance Between Algorithm for Discovering Petri Nets Underfitting and Overfitting”. from Event Logs”. In Business Software and Systems Modeling, Process Management (BPM2008), 9(1):87–111, 2010. pages 358–373, 2008 [17] W.M.P. van der Aalst, A.J.M.M. [24] R. Bergenthum, J. Desel, R. Lorenz, Weijters, and L. Maruster, “Workflow and S. Mauser, “Process Mining Mining: Discovering Process Models Based on Regions of Languages”. In from Event Logs”. IEEE Transactions G. Alonso, P. Dadam, and M. on Knowledge and Data Engineering, Rosemann, editors, International 16(9):1128–1142, 2004. Conference on Business Process Management (BPM 2007), volume [18] R. Agrawal, D. Gunopulos, and F. Leymann, “Mining Process Models 4714 of Lecture Notes in Computer from Work-flow Logs”. In Sixth Science, pages 375–383. Springer- International Conference on Verlag, Berlin, 2007. Extending Database Technology, [25] J.M.E.M. van der Werf, B.F. van volume 1377 of Lecture Notes in Dongen, C.A.J. Hurkens, and A. Computer Science, pages 469–483. Serebrenik, “Process Discovery using SpringerVerlag, Berlin, 1998. Integer Linear Programming”. Fundamenta Informaticae, 94:387– [19] A.J.M.M. Weijters and W.M.P. van der Aalst, “Rediscovering Workflow 412, 2010 Models from Event-Based Data using [26] A.K. Alves de Medeiros, W.M.P. van Little Thumb”. Integrated Computer- der Aalst, ”Genetic Process Mining: Aided Engineering, 10(2):151–162, An Experimental Evaluation”. Data 2003. Mining and Knowledge Discovery, 14(2):245–304, 2007. [20] J.M.E.M. van der Werf, B.F. van Dongen, C.A.J. Hurkens, and A.