{
  "id": 548201,
  "title": "Outlier customers",
  "url": "/competitions/h-and-m-personalized-fashion-recommendations/discussion/548201",
  "author_name": "",
  "post_date": "2024-11-25T16:39:52.468977900Z",
  "votes": null,
  "comment_count": 1,
  "views": 0,
  "content": "<p>I searched but didn't find any discussion of outlier customers.</p>\n<p>I was surprised to see some (a few thousand) customers with close to 2000 purchases and many more with 200 purchases.</p>\n<p>In retail this can sometimes happen if:</p>\n<ul>\n<li>There are some otherwise anonymous transactions associated with some internal account</li>\n<li>There are resellers … could be a company selling into foreign countries with trade barriers or could be an ebay seller or some other shop getting large quantities at discount</li>\n<li>Could be a very enthusiatic   customer … but 2000 items seems excessive </li>\n</ul>\n<p>Does anyone have thoughts - or is there some other discussion directly with the client H&amp;M where they have outlined some of the \"quirks\" that might be noticed?</p>\n<p>What I plan … running recos with and without \"outliers\" … at this point I'm not positive what approach I'll take, ignore some customers, or limit the influence.</p>\n<p><code>SELECT customer_id, count(1)\nFROM transactions\nGROUP BY customer_id\nORDER BY 2 desc\nLIMIT 10000</code></p>\n<p>I have not yet, but I plan to, look at a few of these top customers to see if they have bought the same article numerous times (which to me indicates a reseller).</p>",
  "messages": [
    {
      "id": "3055233",
      "postDate": "11/25/2024 16:39:52",
      "content": "<p>I searched but didn't find any discussion of outlier customers.</p>\n<p>I was surprised to see some (a few thousand) customers with close to 2000 purchases and many more with 200 purchases.</p>\n<p>In retail this can sometimes happen if:</p>\n<ul>\n<li>There are some otherwise anonymous transactions associated with some internal account</li>\n<li>There are resellers … could be a company selling into foreign countries with trade barriers or could be an ebay seller or some other shop getting large quantities at discount</li>\n<li>Could be a very enthusiatic   customer … but 2000 items seems excessive </li>\n</ul>\n<p>Does anyone have thoughts - or is there some other discussion directly with the client H&amp;M where they have outlined some of the \"quirks\" that might be noticed?</p>\n<p>What I plan … running recos with and without \"outliers\" … at this point I'm not positive what approach I'll take, ignore some customers, or limit the influence.</p>\n<p><code>SELECT customer_id, count(1)\nFROM transactions\nGROUP BY customer_id\nORDER BY 2 desc\nLIMIT 10000</code></p>\n<p>I have not yet, but I plan to, look at a few of these top customers to see if they have bought the same article numerous times (which to me indicates a reseller).</p>",
      "rawMarkdown": "I searched but didn't find any discussion of outlier customers.\n\nI was surprised to see some (a few thousand) customers with close to 2000 purchases and many more with 200 purchases.\n\nIn retail this can sometimes happen if:\n- There are some otherwise anonymous transactions associated with some internal account\n- There are resellers ... could be a company selling into foreign countries with trade barriers or could be an ebay seller or some other shop getting large quantities at discount\n- Could be a very enthusiatic   customer ... but 2000 items seems excessive \n\nDoes anyone have thoughts - or is there some other discussion directly with the client H&M where they have outlined some of the \"quirks\" that might be noticed?\n\nWhat I plan ... running recos with and without \"outliers\" ... at this point I'm not positive what approach I'll take, ignore some customers, or limit the influence.\n\n`SELECT customer_id, count(1)\nFROM transactions\nGROUP BY customer_id\nORDER BY 2 desc\nLIMIT 10000`\n\nI have not yet, but I plan to, look at a few of these top customers to see if they have bought the same article numerous times (which to me indicates a reseller).",
      "votes": null
    },
    {
      "id": "3055310",
      "postDate": "11/25/2024 17:21:35",
      "content": "<p>Related to this, I can imagine three separate recommendation \"engines\" and backtesting the three separately (holding out the last 7 days of transactions):</p>\n<ol>\n<li>Normal buyers</li>\n<li>Non-buyers</li>\n<li>Outliers</li>\n</ol>\n<p>Product recommendations for resellers is, in my experience, not the point in sending out an email with \"recommended for you\" .. its the normal buyers that are important.   </p>\n<p>Non-buyers - where you know so much demographic information, but none of their web browse history, seems unusual.</p>",
      "rawMarkdown": "Related to this, I can imagine three separate recommendation \"engines\" and backtesting the three separately (holding out the last 7 days of transactions):\n1. Normal buyers\n2. Non-buyers\n3. Outliers\n\nProduct recommendations for resellers is, in my experience, not the point in sending out an email with \"recommended for you\" .. its the normal buyers that are important.   \n\nNon-buyers - where you know so much demographic information, but none of their web browse history, seems unusual.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 3055310,
      "author_name": "dodger",
      "author_url": "",
      "post_date": "11/25/2024 17:21:35",
      "content": "<p>Related to this, I can imagine three separate recommendation \"engines\" and backtesting the three separately (holding out the last 7 days of transactions):</p>\n<ol>\n<li>Normal buyers</li>\n<li>Non-buyers</li>\n<li>Outliers</li>\n</ol>\n<p>Product recommendations for resellers is, in my experience, not the point in sending out an email with \"recommended for you\" .. its the normal buyers that are important.   </p>\n<p>Non-buyers - where you know so much demographic information, but none of their web browse history, seems unusual.</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "3055233": "I searched but didn't find any discussion of outlier customers.\n\nI was surprised to see some (a few thousand) customers with close to 2000 purchases and many more with 200 purchases.\n\nIn retail this can sometimes happen if:\n- There are some otherwise anonymous transactions associated with some internal account\n- There are resellers ... could be a company selling into foreign countries with trade barriers or could be an ebay seller or some other shop getting large quantities at discount\n- Could be a very enthusiatic   customer ... but 2000 items seems excessive \n\nDoes anyone have thoughts - or is there some other discussion directly with the client H&M where they have outlined some of the \"quirks\" that might be noticed?\n\nWhat I plan ... running recos with and without \"outliers\" ... at this point I'm not positive what approach I'll take, ignore some customers, or limit the influence.\n\n`SELECT customer_id, count(1)\nFROM transactions\nGROUP BY customer_id\nORDER BY 2 desc\nLIMIT 10000`\n\nI have not yet, but I plan to, look at a few of these top customers to see if they have bought the same article numerous times (which to me indicates a reseller).",
    "3055310": "Related to this, I can imagine three separate recommendation \"engines\" and backtesting the three separately (holding out the last 7 days of transactions):\n1. Normal buyers\n2. Non-buyers\n3. Outliers\n\nProduct recommendations for resellers is, in my experience, not the point in sending out an email with \"recommended for you\" .. its the normal buyers that are important.   \n\nNon-buyers - where you know so much demographic information, but none of their web browse history, seems unusual."
  },
  "source": "meta"
}