{
  "id": 79136,
  "title": "Hypothesis on Reference Date and Loyalty Score",
  "url": "/competitions/elo-merchant-category-recommendation/discussion/79136",
  "author_name": "",
  "post_date": "2019-01-31T09:34:39.224177500Z",
  "votes": 21,
  "comment_count": 2,
  "views": 0,
  "content": "<p>We are not given much information about reference date and loyalty score. Loyalty score matter a lot since it is our target, and like many public kernels show, oldness of the reference date influences loyalty score a lot. So apparently understanding these two concepts is important.</p>\n\n<p>This is my <strong>hypothesis</strong> of the situation.\nElo is setting <strong>two periods and two points in time order</strong>.</p>\n\n<h2>Period 1) Historical period</h2>\n\n<p>In this period, Elo watches the card transactions in order to provide appropriate personalized recommendations later.</p>\n\n<h2>Point 1) Reference point</h2>\n\n<p>It is the point of time when Elo starts giving personalized recommendations of <strong>new</strong> merchant categories, based on card holder's historical transactions. </p>\n\n<h2>Period 2) Evaluation period</h2>\n\n<p>In this period, Elo gathers transactions information to <strong>evaluate</strong> card holders if they are likely to be loyal to Elo's recommendations in the future. \nSpecifically, Elo wants to see if card holders made transactions with the merchants recommended. I think it would have been more reasonable if Elo only included the transactions with the merchant they've recommended, but due to whatever reason (maybe they wanted to secure their recommendation algorithm), they've decided not to do so. </p>\n\n<h2>Point 2) Evaluation point</h2>\n\n<p>With the information gathered in evaluation period, Elo seeks to predict card holder's loyalty in the <strong>future</strong> period. If the predicted loyalty score is low, Elo will stop giving recommendations. If predicted loyalty score is high, Elo will continue giving recommendations.</p>\n\n<p>The given true target loyalty scores in the competition dataset are calculated in the period after evaluation date. For example, if the evaluation date is 18-05-30, Elo could have counted how many transactions a card holder made with recommended merchants in 18-06-01~18-10-31, and used that count as a true loyalty score.</p>\n\n<hr>\n\n<p>To conclude, our goal is to predict which card holder will respond actively to Elo’s recommendations in the future based on their transactions before given recommendations, and 2 months of transactions while receiving recommendations.  Elo will use the best algorithm Kagglers provide, maybe combine it with their secret recommendation information, and use that algorithm at an evaluation point to sort out which card holder to continue giving promotions to.</p>\n\n<p>If my hypothesis is correct, I guess how the features of ‘historical transactions’ changed in ‘new merchant transactions’ will also play an important role in predicting future loyalty scores.</p>\n\n<p>But still, I don't understand why oldness of <code>reference date</code> is so important when predicting loyalty.</p>",
  "messages": [
    {
      "id": "464166",
      "postDate": "01/31/2019 09:34:39",
      "content": "<p>We are not given much information about reference date and loyalty score. Loyalty score matter a lot since it is our target, and like many public kernels show, oldness of the reference date influences loyalty score a lot. So apparently understanding these two concepts is important.</p>\n\n<p>This is my <strong>hypothesis</strong> of the situation.\nElo is setting <strong>two periods and two points in time order</strong>.</p>\n\n<h2>Period 1) Historical period</h2>\n\n<p>In this period, Elo watches the card transactions in order to provide appropriate personalized recommendations later.</p>\n\n<h2>Point 1) Reference point</h2>\n\n<p>It is the point of time when Elo starts giving personalized recommendations of <strong>new</strong> merchant categories, based on card holder's historical transactions. </p>\n\n<h2>Period 2) Evaluation period</h2>\n\n<p>In this period, Elo gathers transactions information to <strong>evaluate</strong> card holders if they are likely to be loyal to Elo's recommendations in the future. \nSpecifically, Elo wants to see if card holders made transactions with the merchants recommended. I think it would have been more reasonable if Elo only included the transactions with the merchant they've recommended, but due to whatever reason (maybe they wanted to secure their recommendation algorithm), they've decided not to do so. </p>\n\n<h2>Point 2) Evaluation point</h2>\n\n<p>With the information gathered in evaluation period, Elo seeks to predict card holder's loyalty in the <strong>future</strong> period. If the predicted loyalty score is low, Elo will stop giving recommendations. If predicted loyalty score is high, Elo will continue giving recommendations.</p>\n\n<p>The given true target loyalty scores in the competition dataset are calculated in the period after evaluation date. For example, if the evaluation date is 18-05-30, Elo could have counted how many transactions a card holder made with recommended merchants in 18-06-01~18-10-31, and used that count as a true loyalty score.</p>\n\n<hr>\n\n<p>To conclude, our goal is to predict which card holder will respond actively to Elo’s recommendations in the future based on their transactions before given recommendations, and 2 months of transactions while receiving recommendations.  Elo will use the best algorithm Kagglers provide, maybe combine it with their secret recommendation information, and use that algorithm at an evaluation point to sort out which card holder to continue giving promotions to.</p>\n\n<p>If my hypothesis is correct, I guess how the features of ‘historical transactions’ changed in ‘new merchant transactions’ will also play an important role in predicting future loyalty scores.</p>\n\n<p>But still, I don't understand why oldness of <code>reference date</code> is so important when predicting loyalty.</p>",
      "rawMarkdown": "We are not given much information about reference date and loyalty score. Loyalty score matter a lot since it is our target, and like many public kernels show, oldness of the reference date influences loyalty score a lot. So apparently understanding these two concepts is important.\n\nThis is my **hypothesis** of the situation.\nElo is setting **two periods and two points in time order**.\n\n## Period 1) Historical period\nIn this period, Elo watches the card transactions in order to provide appropriate personalized recommendations later.\n\n## Point 1) Reference point\nIt is the point of time when Elo starts giving personalized recommendations of **new** merchant categories, based on card holder's historical transactions. \n\n## Period 2) Evaluation period\nIn this period, Elo gathers transactions information to **evaluate** card holders if they are likely to be loyal to Elo's recommendations in the future. \nSpecifically, Elo wants to see if card holders made transactions with the merchants recommended. I think it would have been more reasonable if Elo only included the transactions with the merchant they've recommended, but due to whatever reason (maybe they wanted to secure their recommendation algorithm), they've decided not to do so. \n\n## Point 2) Evaluation point\nWith the information gathered in evaluation period, Elo seeks to predict card holder's loyalty in the **future** period. If the predicted loyalty score is low, Elo will stop giving recommendations. If predicted loyalty score is high, Elo will continue giving recommendations.\n\nThe given true target loyalty scores in the competition dataset are calculated in the period after evaluation date. For example, if the evaluation date is 18-05-30, Elo could have counted how many transactions a card holder made with recommended merchants in 18-06-01~18-10-31, and used that count as a true loyalty score.\n\n---\n\nTo conclude, our goal is to predict which card holder will respond actively to Elo’s recommendations in the future based on their transactions before given recommendations, and 2 months of transactions while receiving recommendations.  Elo will use the best algorithm Kagglers provide, maybe combine it with their secret recommendation information, and use that algorithm at an evaluation point to sort out which card holder to continue giving promotions to.\n\nIf my hypothesis is correct, I guess how the features of ‘historical transactions’ changed in ‘new merchant transactions’ will also play an important role in predicting future loyalty scores.\n\nBut still, I don't understand why oldness of `reference date` is so important when predicting loyalty.",
      "votes": null
    },
    {
      "id": "464491",
      "postDate": "02/01/2019 01:13:08",
      "content": "<p>This is nice point. Thanks for sharing.</p>",
      "rawMarkdown": "This is nice point. Thanks for sharing.",
      "votes": null
    },
    {
      "id": "464541",
      "postDate": "02/01/2019 04:00:37",
      "content": "<p>Seems make sense</p>",
      "rawMarkdown": "Seems make sense",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 464491,
      "author_name": "longyin2",
      "author_url": "",
      "post_date": "02/01/2019 01:13:08",
      "content": "<p>This is nice point. Thanks for sharing.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 464541,
      "author_name": "codlife",
      "author_url": "",
      "post_date": "02/01/2019 04:00:37",
      "content": "<p>Seems make sense</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "464166": "We are not given much information about reference date and loyalty score. Loyalty score matter a lot since it is our target, and like many public kernels show, oldness of the reference date influences loyalty score a lot. So apparently understanding these two concepts is important.\n\nThis is my **hypothesis** of the situation.\nElo is setting **two periods and two points in time order**.\n\n## Period 1) Historical period\nIn this period, Elo watches the card transactions in order to provide appropriate personalized recommendations later.\n\n## Point 1) Reference point\nIt is the point of time when Elo starts giving personalized recommendations of **new** merchant categories, based on card holder's historical transactions. \n\n## Period 2) Evaluation period\nIn this period, Elo gathers transactions information to **evaluate** card holders if they are likely to be loyal to Elo's recommendations in the future. \nSpecifically, Elo wants to see if card holders made transactions with the merchants recommended. I think it would have been more reasonable if Elo only included the transactions with the merchant they've recommended, but due to whatever reason (maybe they wanted to secure their recommendation algorithm), they've decided not to do so. \n\n## Point 2) Evaluation point\nWith the information gathered in evaluation period, Elo seeks to predict card holder's loyalty in the **future** period. If the predicted loyalty score is low, Elo will stop giving recommendations. If predicted loyalty score is high, Elo will continue giving recommendations.\n\nThe given true target loyalty scores in the competition dataset are calculated in the period after evaluation date. For example, if the evaluation date is 18-05-30, Elo could have counted how many transactions a card holder made with recommended merchants in 18-06-01~18-10-31, and used that count as a true loyalty score.\n\n---\n\nTo conclude, our goal is to predict which card holder will respond actively to Elo’s recommendations in the future based on their transactions before given recommendations, and 2 months of transactions while receiving recommendations.  Elo will use the best algorithm Kagglers provide, maybe combine it with their secret recommendation information, and use that algorithm at an evaluation point to sort out which card holder to continue giving promotions to.\n\nIf my hypothesis is correct, I guess how the features of ‘historical transactions’ changed in ‘new merchant transactions’ will also play an important role in predicting future loyalty scores.\n\nBut still, I don't understand why oldness of `reference date` is so important when predicting loyalty.",
    "464491": "This is nice point. Thanks for sharing.",
    "464541": "Seems make sense"
  },
  "source": "meta"
}