[{"id":272,"title":"When to worry about negative weights in MVA training for HEP analysis","sha":"10ebd8fea8726550c16401ec5622aa44","user_id":null,"state":"published","body":"A discussion at DS@LHC2015 concerned the treatment of negative weights in MVA training. This is relevant for HEP since the advent of MC generators at NLO in QCD, as [the matching of Matrix-Element to Parton-Shower generators](http://dx.doi.org/10.1088/1126-6708/2002/06/029) assigns negative weights to a fraction of the events with non-uniform probability.\r\nMVA experts remarked the technical difficulty of sign-aware training, although preliminary procedures exist for some tools.\r\nWhen should a HEP analyst worry?\r\nTreating all events as positive is a mismodeling; that does not make MVA outcome \"incorrect\", but possibly sub-optimal (similarly to using fast detector simulations for training and [GEANT](http://dx.doi.org/10.1016/S0168-9002(03)01368-8) for testing, as [O(10%)-level mismodelings](http://dx.doi.org/10.1088/1742-6596/513/2/022012) are usually tolerated in the former). Systematic uncertainties (e.g., QCD scale) may lead to worse mismodeling. \r\nBefore investing time on addressing negative weights treatment, an analyst should check if ignoring the sign would lead to large mismodeling in some important kinematic region, with \"large\" defined as \"larger than any systematic uncertainty\"; e.g., by comparing the normalized distributions of positive-only and negative-only events for the most discriminating MVA inputs: if the difference is larger than between typical systematic variations, it is legitimate to worry. Alternatively, after a sign-blind training, this test can be done on the MVA output.","created_at":"2016-01-07T15:28:33.919Z","updated_at":"2016-01-08T06:01:07.812Z","zenodo_id":92471,"doi":"http://dx.doi.org/10.5281/zenodo.44441","subject":null,"tags":["dslhc","hep","lhc","ml","mva","data analysis"],"vote_count":0,"view_count":1,"score":0.1,"deleted":false,"muted":false,"tweeted":false,"attachment_file_name":null,"attachment_content_type":null,"attachment_file_size":null,"attachment_updated_at":null},{"id":260,"title":"Brief Ideas for the Data Science at LHC Workshop 2015","sha":"48be7c9cc3de0c437011525e9d0b0180","user_id":null,"state":"published","body":"The Data Science @ LHC workshop was a resounding success. We do not plan to have traditional proceedings tied to individual talks, but we do want to capture the ideas that were generated during the workshop. With that in mind, we want to try something new: we encourage you to submit contributions to the Journal of Brief Ideas. Each contribution is capped at 200 words, so this provides a lightweight method of recording ideas that came up during the workshop. They do not have to be tied to a particular talk or speaker, and you can even submit several ideas with a different set of authors. The goal is to lay the groundwork for new collaborations and follow up studies, not to document completed research. As this is a new way of doing things, please let us know what you think or if need help with anything. Each publication will receive a DOI so that it can be cited by other publications and collected by INSPIRE as contributions to the conference.\r\nImportant : fill in your brief ideas by 22 January.\r\nVery important: please use this tag: DSLHC when composing the idea so that we can easily aggregate them.","created_at":"2015-12-22T22:54:51.146Z","updated_at":"2016-01-06T06:00:27.519Z","zenodo_id":92248,"doi":"http://dx.doi.org/10.5281/zenodo.44357","subject":null,"tags":["datascience","lhc","dslhc"],"vote_count":0,"view_count":1,"score":0.1,"deleted":false,"muted":false,"tweeted":false,"attachment_file_name":null,"attachment_content_type":null,"attachment_file_size":null,"attachment_updated_at":null},{"id":294,"title":"Kickstarting research into end-to-end trigger systems","sha":"6e2cc286c54020f08d774487315585e8","user_id":null,"state":"published","body":"The data volumes (_**O**_(TB/s)) created at the Large Hadron Collider (LHC) are too large to record. Typical rejection factors are _**O**_(100-1000), and using as little CPU time as possible to reject an event is the goal. More powerful decision features take more CPU time to construct, therefore the discrimination power of decisions is limited by the available CPU time.\r\n\r\nCurrent approaches rely on hand-crafted features, and hand tuned decision cascades. We propose the possibility of using deep learning techniques to construct an end-to-end system (using the raw data electronics data as input) that learns the decision function of the trigger system (the global HLT decision). Deep Neural Networks can be efficiently evaluated on dedicated hardware and can be used to preempt the decision of the full system.\r\n\r\nFor LHCb we estimate that 750GB of data contain about 10000000 candidates of which 100000 will be accepted by the trigger system.\r\n\r\nTo kickstart research into this new area of research each of the four LHC experiments should release a few seconds of HLT output as well as the corresponding input. This will allow collaboration with experts from the field of machine-learning and cooperation between the experiments to solve this challenging problem.","created_at":"2016-04-01T11:31:52.210Z","updated_at":"2016-04-16T13:09:00.853Z","zenodo_id":105137,"doi":"http://dx.doi.org/10.5281/zenodo.48784","subject":null,"tags":["lhc","deep learning","machinelearning","data_exchange"],"vote_count":1,"view_count":1,"score":1.1,"deleted":false,"muted":false,"tweeted":true,"attachment_file_name":null,"attachment_content_type":null,"attachment_file_size":null,"attachment_updated_at":null},{"id":162,"title":"Learning computationally expensive functions","sha":"1670e0f0f5125e7292dec2f10f09ed59","user_id":349,"state":"published","body":"Analysing data from an experiment at the Large Hadron Collider at CERN requires large amounts of computing power.\r\n\r\nIn addition to large amounts of experimental data, each individual analysis requires large amounts of simulated data. This production of simulated data is the single largest consumer of CPU resources of the LHC computing grid [[Atlas Simulation](https://indico.cern.ch/event/279530/session/5/contribution/11), [LHCb Computing Resource usage in 2014](http://cds.cern.ch/record/1984010)]. Producing enough simulated data is already challenging and will become a limiting factor in the future with the size of experimental datasets poised to increase by O(10)--O(100).\r\n\r\nIn individual analyses sophisticated features are computed from basic quantities measured in the detectors. These sophisticated features improve the sensitivity of an analysis, but require a large amount of CPU time [[Mtop matrix element](http://arxiv.org/abs/hep-ex/0406031), [Measurement of the Single Top Quark Production Cross Section at CDF](http://arxiv.org/abs/0809.2581), [Matrix element for Higgs](http://arxiv.org/abs/1502.02485), [Missing mass calculator](http://arxiv.org/abs/1012.4686)]. The large CPU needs limit the number of cases in which these techniques are used, resulting in sub-optimal analyses.\r\n\r\nI propose replacing the calculations by a regression model [[Greedy Function Approximation: A Gradient Boosting Machine](http://www.jstor.org/stable/2699986), [Scikit-learn: Machine Learning in Python](http://jmlr.csail.mit.edu/papers/v12/pedregosa11a.html)] which is cheap to evaluate. The full simulation or computation of sophisticated features is performed on a subset of events, which are then used to train a regression model. This model is used to perform the computation for a large number of events. A proof of concept for the [Missing mass calculator](http://arxiv.org/abs/1012.4686) using [the dataset from the ATLAS Higgs Boson Machine Learning Challenge](http://doi.org/10.7483/OPENDATA.ATLAS.ZBP2.M5T8) is available: [Learning expensive functions](http://betatim.github.io/posts/learning-expensive-functions/).","created_at":"2015-05-18T11:39:18.249Z","updated_at":"2015-12-18T13:22:56.608Z","zenodo_id":29587,"doi":"http://dx.doi.org/10.5281/zenodo.17720","subject":null,"tags":["hep","cern","lhc","ml","machinelearning","regression","physics"],"vote_count":1,"view_count":1,"score":1.1,"deleted":false,"muted":false,"tweeted":true,"attachment_file_name":null,"attachment_content_type":null,"attachment_file_size":null,"attachment_updated_at":null},{"id":412,"title":"Recasting through reweighting","sha":"8106c030eba22dd3a8d268940d5e42d8","user_id":null,"state":"published","body":"*Recasting* refers to reinterpreting the results of searches for new particles or standard model measurements in the context of different theoretical models\u0026nbsp;\\[[1](http://inspirehep.net/record/872781)\\]. The fundamental task is to replace the original hypothesis $p_0(x)$ with a new hypothesis $p_1(x)$, where $x$ is some observed quantity. The effect of the detector response and analysis cuts can be encoded in a *folding* operator $\\int\\,W(x|z)dz$ acting on the truth-level distribution $p(z)$. By keeping the  analysis fixed, $W(x|z)$ does not change, thus recasting amounts to:\r\n\\begin{equation}p_0(x)=\\int\\,p_0(z)W(x|z)dz\\hspace{.5cm}\\Longrightarrow\\hspace{.5cm}p_1(x)=\\int\\,p_1(z)W(x|z)dz\\end{equation}\r\n\r\nThere are two primary approaches:\r\n  * **folding**: Samples from $p_1(z)$ are run through a detector simulation and analysis chain to estimate $p_1(x)$\u0026nbsp;\\[[2](https://indico.cern.ch/event/567550/contributions/2638695/)\\]. This is common when $z$ is high-dimensional, $p_0(z)$ and $p_1(z)$ are very different, or $W(x|z)$ is sensitive to experimental details.\r\n  * **unfolding**: An alternate theory $p_1(z)$ is compared directly to an unfolded distribution $\\hat{p}(z)$ obtained from applying an approximate inverse operation to the observed data. Typically, unfolding is restricted to low-dimensional $x, z$ and Gaussian uncertainties.\r\n\r\nWe point out a third option\r\n   * **reweighting**: Reweight pre-folded events $(x_i,z_i)~\\sim\\,p_0(x,z)$ by the factor $r(z_i)=p_1(z_i)/p_0(z_i)$, as in\r\n\\begin{equation}p_1(x)=\\int p_1(z)W(x|z)dz=\\int\\,p_0(z)\\underbrace{\\frac{p_1(z)}{p_0(z)}}_{\\textrm{reweighting}}W(x|z)dz\\end{equation}\r\nThis approach does not require simulating new events or the approximations used in unfolding. Note, sample variance becomes a problem if $r(z_i)\\gg\\,1$.","created_at":"2017-10-14T14:16:18.901Z","updated_at":"2022-06-05T20:51:32.689Z","zenodo_id":1013926,"doi":"https://doi.org/10.5281/zenodo.1013926","subject":null,"tags":["physics","reinterpretation","lhc"],"vote_count":3,"view_count":1,"score":3.1,"deleted":false,"muted":false,"tweeted":true,"attachment_file_name":null,"attachment_content_type":null,"attachment_file_size":null,"attachment_updated_at":null}]