{
  "id": 16974,
  "title": "What are your \"physical sound\" features?",
  "url": "/competitions/flavours-of-physics/discussion/16974",
  "author_name": "phunter",
  "post_date": "2015-10-13T00:27:14.167000",
  "votes": 2,
  "comment_count": 0,
  "views": 233,
  "content": "<p>Let me start this topic as Hors d'Oeuvres. My physical sound features include:</p>\n\n<ol>\n<li>each muon particle's momentum/rapidity as itself has almost no sense, but some combination helps a little bit by using the symmetric among these 3 muons.</li>\n<li>kinematic features using Ds-&gt;tau decay where the mass change is small, so the tau momentum is almost along with the Ds momentum, thus the IP is small, Dira is close to 1 and some combination of these kinematic feature</li>\n<li>tau features: tau flights for some distance while the background decay is almost immediate, so the secondary vertex distance to PV and so on is also interesting.</li>\n<li>vertex features: the background is Ds-&gt;eta mu nu and eta-&gt;mu mu where these 3 muons are not coming from the same vertex as in tau-&gt;3 mu, thus the vertex quality and the track quality for positive and negative samples are not the same.</li>\n<li>selection features: track isolation and cone isolation. I see these features are dropped for some reason in the most famous script from JustFor. Actually, the max min value of these features have good importance and helps my score up. The reason comes from physics about the background events of LHCb, not an easy reason. I learned it from CERN-THESIS-2015-021 for LHCb's tau-&gt;3 muon search, it is good to read.</li>\n</ol>\n\n<p>And a trick for boosting up the score from me can be, using SPDhits model for ensemble. Most models with SPDhits can't pass the KS test if use training.csv only, the reason is that this feature is not perfected simulated, and it needs calibration for simulation vs control. But, giving a small weight of these models to the final ensemble can magically boost the score a little bit with no hurt. </p>",
  "messages": [
    {
      "id": 95926,
      "postDate": "2015-10-13T00:27:14.167Z",
      "content": "<p>Let me start this topic as Hors d'Oeuvres. My physical sound features include:</p>\n\n<ol>\n<li>each muon particle's momentum/rapidity as itself has almost no sense, but some combination helps a little bit by using the symmetric among these 3 muons.</li>\n<li>kinematic features using Ds-&gt;tau decay where the mass change is small, so the tau momentum is almost along with the Ds momentum, thus the IP is small, Dira is close to 1 and some combination of these kinematic feature</li>\n<li>tau features: tau flights for some distance while the background decay is almost immediate, so the secondary vertex distance to PV and so on is also interesting.</li>\n<li>vertex features: the background is Ds-&gt;eta mu nu and eta-&gt;mu mu where these 3 muons are not coming from the same vertex as in tau-&gt;3 mu, thus the vertex quality and the track quality for positive and negative samples are not the same.</li>\n<li>selection features: track isolation and cone isolation. I see these features are dropped for some reason in the most famous script from JustFor. Actually, the max min value of these features have good importance and helps my score up. The reason comes from physics about the background events of LHCb, not an easy reason. I learned it from CERN-THESIS-2015-021 for LHCb's tau-&gt;3 muon search, it is good to read.</li>\n</ol>\n\n<p>And a trick for boosting up the score from me can be, using SPDhits model for ensemble. Most models with SPDhits can't pass the KS test if use training.csv only, the reason is that this feature is not perfected simulated, and it needs calibration for simulation vs control. But, giving a small weight of these models to the final ensemble can magically boost the score a little bit with no hurt. </p>",
      "rawMarkdown": "Let me start this topic as Hors d'Oeuvres. My physical sound features include:\r\n\r\n0. each muon particle's momentum/rapidity as itself has almost no sense, but some combination helps a little bit by using the symmetric among these 3 muons.\r\n1. kinematic features using Ds->tau decay where the mass change is small, so the tau momentum is almost along with the Ds momentum, thus the IP is small, Dira is close to 1 and some combination of these kinematic feature\r\n2. tau features: tau flights for some distance while the background decay is almost immediate, so the secondary vertex distance to PV and so on is also interesting.\r\n3. vertex features: the background is Ds->eta mu nu and eta->mu mu where these 3 muons are not coming from the same vertex as in tau->3 mu, thus the vertex quality and the track quality for positive and negative samples are not the same.\r\n4. selection features: track isolation and cone isolation. I see these features are dropped for some reason in the most famous script from JustFor. Actually, the max min value of these features have good importance and helps my score up. The reason comes from physics about the background events of LHCb, not an easy reason. I learned it from CERN-THESIS-2015-021 for LHCb's tau->3 muon search, it is good to read.\r\n\r\nAnd a trick for boosting up the score from me can be, using SPDhits model for ensemble. Most models with SPDhits can't pass the KS test if use training.csv only, the reason is that this feature is not perfected simulated, and it needs calibration for simulation vs control. But, giving a small weight of these models to the final ensemble can magically boost the score a little bit with no hurt. ",
      "votes": 2
    }
  ],
  "comments": [],
  "raw_markdown_by_id": {
    "95926": "Let me start this topic as Hors d'Oeuvres. My physical sound features include:\r\n\r\n0. each muon particle's momentum/rapidity as itself has almost no sense, but some combination helps a little bit by using the symmetric among these 3 muons.\r\n1. kinematic features using Ds->tau decay where the mass change is small, so the tau momentum is almost along with the Ds momentum, thus the IP is small, Dira is close to 1 and some combination of these kinematic feature\r\n2. tau features: tau flights for some distance while the background decay is almost immediate, so the secondary vertex distance to PV and so on is also interesting.\r\n3. vertex features: the background is Ds->eta mu nu and eta->mu mu where these 3 muons are not coming from the same vertex as in tau->3 mu, thus the vertex quality and the track quality for positive and negative samples are not the same.\r\n4. selection features: track isolation and cone isolation. I see these features are dropped for some reason in the most famous script from JustFor. Actually, the max min value of these features have good importance and helps my score up. The reason comes from physics about the background events of LHCb, not an easy reason. I learned it from CERN-THESIS-2015-021 for LHCb's tau->3 muon search, it is good to read.\r\n\r\nAnd a trick for boosting up the score from me can be, using SPDhits model for ensemble. Most models with SPDhits can't pass the KS test if use training.csv only, the reason is that this feature is not perfected simulated, and it needs calibration for simulation vs control. But, giving a small weight of these models to the final ensemble can magically boost the score a little bit with no hurt. "
  }
}