{"id":1185,"date":"2024-05-18T16:17:07","date_gmt":"2024-05-18T16:17:07","guid":{"rendered":"https:\/\/ml-gis-service.com\/?p=1185"},"modified":"2024-05-18T16:25:56","modified_gmt":"2024-05-18T16:25:56","slug":"wsknn-session-based-recommendation-engine-and-its-parameters","status":"publish","type":"post","link":"https:\/\/ml-gis-service.com\/index.php\/2024\/05\/18\/wsknn-session-based-recommendation-engine-and-its-parameters\/","title":{"rendered":"Set this right! – How to prepare recommendation system for the real world"},"content":{"rendered":"\n
\n

Session-based recommendation engine in Python<\/p>\nPart 4<\/cite><\/blockquote>\n\n\n\n

Introduction<\/h1>\n\n\n\n

Every model is founded on high-quality data. Assuming a clean and representative dataset, you may start thinking about the model’s hyperparameters. WSKNN<\/code> recommendation model has multiple settings, and we will test them to see how they affect recommendations. Before we start, you should know that if you use WSKNN<\/code> package and its internal algorithm, you can change model parameters on the fly, even during the inference! What’s the catch? The model grows very fast because it duplicates data; thus, it is unreliable for enormous datasets.<\/p>\n\n\n\n

Here is the list of model parameters in the package wsknn==1.2<\/code>:<\/p>\n\n\n\n

    \n
  1. The number of recommendations.<\/li>\n\n\n\n
  2. Sample size.<\/li>\n\n\n\n
  3. Sampling strategy.<\/li>\n\n\n\n
  4. Number of neighbors.<\/li>\n\n\n\n
  5. Weighting function.<\/li>\n\n\n\n
  6. Ranking strategy.<\/li>\n\n\n\n
  7. Known items.<\/li>\n\n\n\n
  8. Required action and its index.<\/li>\n\n\n\n
  9. Custom weights.<\/li>\n\n\n\n
  10. Random recommendations.<\/li>\n<\/ol>\n\n\n\n

    We will describe each parameter and check how single or multiple parameters can change the model scoring and inference time.<\/p>\n\n\n\n

    (1) Number of recommendations<\/h2>\n\n\n\n

    Parameter name:<\/strong> number_of_recommendations<\/code><\/p>\n\n\n\n

    Default value:<\/strong> 5<\/p>\n\n\n\n

    Short description:<\/strong> the number of recommended products<\/p>\n\n\n\n

    Long description:<\/strong>\u00a0This parameter doesn’t switch anything within a trained model but affects the number of outputs. We show five recommendations for online shops typically. Sometimes, we need more results. WSKNN<\/code> returns every recommendation with a weight; thus, we can set this parameter to a large value and cut off low-weighted recommendations. UI\/UX design might impose how many recommendations a model returns. This parameter is related to (7) and (10) – those three parameters profoundly affect the output.<\/p>\n\n\n\n

    (2) Sample size<\/h2>\n\n\n\n

    Parameter name:<\/strong> sample_size<\/code><\/p>\n\n\n\n

    Default value:<\/strong> 1000<\/p>\n\n\n\n

    Short description:<\/strong> how many sessions from the model are sampled to make a recommendation, the possible neighbors space<\/p>\n\n\n\n

    Long description:<\/strong> This is the first parameter that changes the internal workings of the WSKNN architecture. Typically, the model stores billions of sessions. Checking every session’s similarity could be cumbersome. The model makes a recommendation using only a subset of all available sessions. It chooses the subset using the strategy provided by the parameter (3) Sampling Strategy<\/strong>.<\/p>\n\n\n\n

    (3) Sampling strategy<\/h2>\n\n\n\n

    Parameter name:<\/strong> sampling_strategy<\/code><\/p>\n\n\n\n

    Default value:<\/strong> \"common_items\"<\/code><\/p>\n\n\n\n

    Short description:<\/strong> initial session-filtering options<\/p>\n\n\n\n

    Long description:<\/strong> The WSKNN network has thousands of sessions, and it would be unreliable to check the similarity with every session in a set. That’s why we set the sample size<\/strong> (2) and chose a sampling strategy from four options:<\/p>\n\n\n\n