{
  "id": 371439,
  "title": "[revisit] the problem",
  "url": "/competitions/otto-recommender-system/discussion/371439",
  "author_name": "Kefan Xu",
  "post_date": "2022-12-10T02:23:18.116000",
  "votes": 1,
  "comment_count": 0,
  "views": 0,
  "content": "<p>what this topic trying to accomplish is to map/analyze the host's goal into [project] definition (draft) from an engineer's view. </p>\n<h1><strong>validate the problem</strong></h1>\n<p><strong>why the problem</strong><br>\none/two objective(s) prediction is not good enough, want to explore a <strong>multi-objective</strong> prediction. </p>\n<p><strong>what is the domain of the problem</strong><br>\necommerce. in ecommerce, 3 most important elements are [customer], [product] and [transactions]. <br>\n[transactions] are the gold, they associate the [customer] with the [product]. now translate that into the dataset provided by host, we got, [session] is the customer, [aid] is the product, [events] is the transactions. </p>\n<p><strong>what are we trying to accomplish</strong> <br>\nwe gets input dataset (features), then [a model] to produce useful output dataset (labels).</p>\n<p><strong>can this problem resolved by non-ML method</strong><br>\nyes. recommender system in ecommerce was there almost since the day one when ecommerce was born. [customer]. [product], [sales] are well defined in the relational database, and used for generating list suggested item for users, which means [multi-objective recommender system] in theory, it was already there ( prove me wrong ). but may not be as optimal as ML method. ( may need to do a comparison, to prove the conclusion ). </p>\n<p><strong>is ML the right approach for this problem</strong> <br>\nthe indirect answer is, ML can use the data more effectively than human crafted methods. but the end result ( beyond this competition ) may be a combination of ML and human craft methods ( its hard the define or draw a line between ML and human crafted method in some areas ). if we choose ML, then the dataset is the one we are depended on for the success of resolving this problem. </p>\n<h1><strong>validate the data</strong></h1>\n<p><strong>predictive power</strong><br>\nthe dataset provided by host contains [session], [event],  [aid] and [time](inside of event] which map into the core of the ecommerce ( as described above ), the host provided dataset does has predictive power. in other words, these dataset has more predictive power than other non-core elements dataset in ecommerce domain.</p>\n<p><strong>consistent, reliable, correct and available</strong><br>\nthese are the one of the fun part of the competition and our job to validate.</p>\n<h1><strong>validate the actionability after the problem resolve</strong></h1>\n<p>if the result of prediction of what a session-event will do next matches host expectation, then these predictions ( the model ) can fit into **multi-objective **recommender system, to become part of the system. ths NEW recommender system will lean to a goal of better user-experience which helps users to get the product they really ( its measurable ) NEED, not something pushed to them. and because the predictions are based on click, cart and order, its the full view of user's intension (compare to 1/2 objectives), then one of the direct benefit for the ecommerce shop, is this user will has less chance to return the product recommended by this system ( is this valid conclusion ? ). </p>",
  "messages": [
    {
      "id": 2060487,
      "postDate": "2022-12-10T02:23:18.117Z",
      "content": "<p>what this topic trying to accomplish is to map/analyze the host's goal into [project] definition (draft) from an engineer's view. </p>\n<h1><strong>validate the problem</strong></h1>\n<p><strong>why the problem</strong><br>\none/two objective(s) prediction is not good enough, want to explore a <strong>multi-objective</strong> prediction. </p>\n<p><strong>what is the domain of the problem</strong><br>\necommerce. in ecommerce, 3 most important elements are [customer], [product] and [transactions]. <br>\n[transactions] are the gold, they associate the [customer] with the [product]. now translate that into the dataset provided by host, we got, [session] is the customer, [aid] is the product, [events] is the transactions. </p>\n<p><strong>what are we trying to accomplish</strong> <br>\nwe gets input dataset (features), then [a model] to produce useful output dataset (labels).</p>\n<p><strong>can this problem resolved by non-ML method</strong><br>\nyes. recommender system in ecommerce was there almost since the day one when ecommerce was born. [customer]. [product], [sales] are well defined in the relational database, and used for generating list suggested item for users, which means [multi-objective recommender system] in theory, it was already there ( prove me wrong ). but may not be as optimal as ML method. ( may need to do a comparison, to prove the conclusion ). </p>\n<p><strong>is ML the right approach for this problem</strong> <br>\nthe indirect answer is, ML can use the data more effectively than human crafted methods. but the end result ( beyond this competition ) may be a combination of ML and human craft methods ( its hard the define or draw a line between ML and human crafted method in some areas ). if we choose ML, then the dataset is the one we are depended on for the success of resolving this problem. </p>\n<h1><strong>validate the data</strong></h1>\n<p><strong>predictive power</strong><br>\nthe dataset provided by host contains [session], [event],  [aid] and [time](inside of event] which map into the core of the ecommerce ( as described above ), the host provided dataset does has predictive power. in other words, these dataset has more predictive power than other non-core elements dataset in ecommerce domain.</p>\n<p><strong>consistent, reliable, correct and available</strong><br>\nthese are the one of the fun part of the competition and our job to validate.</p>\n<h1><strong>validate the actionability after the problem resolve</strong></h1>\n<p>if the result of prediction of what a session-event will do next matches host expectation, then these predictions ( the model ) can fit into **multi-objective **recommender system, to become part of the system. ths NEW recommender system will lean to a goal of better user-experience which helps users to get the product they really ( its measurable ) NEED, not something pushed to them. and because the predictions are based on click, cart and order, its the full view of user's intension (compare to 1/2 objectives), then one of the direct benefit for the ecommerce shop, is this user will has less chance to return the product recommended by this system ( is this valid conclusion ? ). </p>",
      "rawMarkdown": "what this topic trying to accomplish is to map/analyze the host's goal into [project] definition (draft) from an engineer's view. \n\n\n#**validate the problem**\n\n**why the problem**\none/two objective(s) prediction is not good enough, want to explore a **multi-objective** prediction. \n\n**what is the domain of the problem**\necommerce. in ecommerce, 3 most important elements are [customer], [product] and [transactions]. \n[transactions] are the gold, they associate the [customer] with the [product]. now translate that into the dataset provided by host, we got, [session] is the customer, [aid] is the product, [events] is the transactions. \n\n**what are we trying to accomplish** \nwe gets input dataset (features), then [a model] to produce useful output dataset (labels).\n\n**can this problem resolved by non-ML method**\nyes. recommender system in ecommerce was there almost since the day one when ecommerce was born. [customer]. [product], [sales] are well defined in the relational database, and used for generating list suggested item for users, which means [multi-objective recommender system] in theory, it was already there ( prove me wrong ). but may not be as optimal as ML method. ( may need to do a comparison, to prove the conclusion ). \n\n**is ML the right approach for this problem** \nthe indirect answer is, ML can use the data more effectively than human crafted methods. but the end result ( beyond this competition ) may be a combination of ML and human craft methods ( its hard the define or draw a line between ML and human crafted method in some areas ). if we choose ML, then the dataset is the one we are depended on for the success of resolving this problem. \n\n\n#**validate the data**\n**predictive power**\nthe dataset provided by host contains [session], [event],  [aid] and [time](inside of event] which map into the core of the ecommerce ( as described above ), the host provided dataset does has predictive power. in other words, these dataset has more predictive power than other non-core elements dataset in ecommerce domain.\n\n**consistent, reliable, correct and available**\nthese are the one of the fun part of the competition and our job to validate.\n\n\n#**validate the actionability after the problem resolve**\nif the result of prediction of what a session-event will do next matches host expectation, then these predictions ( the model ) can fit into **multi-objective **recommender system, to become part of the system. ths NEW recommender system will lean to a goal of better user-experience which helps users to get the product they really ( its measurable ) NEED, not something pushed to them. and because the predictions are based on click, cart and order, its the full view of user's intension (compare to 1/2 objectives), then one of the direct benefit for the ecommerce shop, is this user will has less chance to return the product recommended by this system ( is this valid conclusion ? ). \n",
      "votes": 1
    }
  ],
  "comments": [],
  "raw_markdown_by_id": {
    "2060487": "what this topic trying to accomplish is to map/analyze the host's goal into [project] definition (draft) from an engineer's view. \n\n\n#**validate the problem**\n\n**why the problem**\none/two objective(s) prediction is not good enough, want to explore a **multi-objective** prediction. \n\n**what is the domain of the problem**\necommerce. in ecommerce, 3 most important elements are [customer], [product] and [transactions]. \n[transactions] are the gold, they associate the [customer] with the [product]. now translate that into the dataset provided by host, we got, [session] is the customer, [aid] is the product, [events] is the transactions. \n\n**what are we trying to accomplish** \nwe gets input dataset (features), then [a model] to produce useful output dataset (labels).\n\n**can this problem resolved by non-ML method**\nyes. recommender system in ecommerce was there almost since the day one when ecommerce was born. [customer]. [product], [sales] are well defined in the relational database, and used for generating list suggested item for users, which means [multi-objective recommender system] in theory, it was already there ( prove me wrong ). but may not be as optimal as ML method. ( may need to do a comparison, to prove the conclusion ). \n\n**is ML the right approach for this problem** \nthe indirect answer is, ML can use the data more effectively than human crafted methods. but the end result ( beyond this competition ) may be a combination of ML and human craft methods ( its hard the define or draw a line between ML and human crafted method in some areas ). if we choose ML, then the dataset is the one we are depended on for the success of resolving this problem. \n\n\n#**validate the data**\n**predictive power**\nthe dataset provided by host contains [session], [event],  [aid] and [time](inside of event] which map into the core of the ecommerce ( as described above ), the host provided dataset does has predictive power. in other words, these dataset has more predictive power than other non-core elements dataset in ecommerce domain.\n\n**consistent, reliable, correct and available**\nthese are the one of the fun part of the competition and our job to validate.\n\n\n#**validate the actionability after the problem resolve**\nif the result of prediction of what a session-event will do next matches host expectation, then these predictions ( the model ) can fit into **multi-objective **recommender system, to become part of the system. ths NEW recommender system will lean to a goal of better user-experience which helps users to get the product they really ( its measurable ) NEED, not something pushed to them. and because the predictions are based on click, cart and order, its the full view of user's intension (compare to 1/2 objectives), then one of the direct benefit for the ecommerce shop, is this user will has less chance to return the product recommended by this system ( is this valid conclusion ? ). \n"
  }
}