ایجاد مدل ماشین لرنینگ با استفاده از الگوریتم‌های رتبه‌بندی

درود به فریلنسرهای عزیز، اگر متن زیر را به صورت به هم ریخته میبینید، فایل PDF ضمیمه شده را مطالعه بفرمایید. ما یک دیتاست شامل 200 کیوورد داریم که 50 وبسایت اول گوگل برای هر یک از آنها جمع آوری شده است (پس در مجموع ده هزار وبسایت در این دیتاست وجود دارد). همچنین 37 پارامتر یکسان برای هر یک از این وبسایت‌ها تهیه و به دیتاست اضافه شده است. میخواهیم به کمک یک فریلنسر آشنا با ماشین لرنینگ و دیتاساینس موارد زیر رو انجام بدیم: تکمیل کردن دیتاست با اضافه شدن semantic text relatedness به عنوان پارامتر 38ام: باید 50 وبسایت مربوط به هر کیوورد به کمک مدل "Learning-To-Rank with Bert in TF-Ranking" رتبه بندی شده و نتیجه این رتبه بندی به دیتاست اضافه شود. این مدل توسط محققین گوگل توسعه داده شده و کد مربوط به آن در گیت‌هاب موجود است که لینک و مقاله مربوط به آن را به پروژه ضمیمه کردیم. توضیح این بخش به زبان انگلیسی: Adding text semantic relatedness factor using Learning-To-Rank with Bert in TF-Ranking: We don’t want to train a model in this step as we use a pretrained model that is developed by Google researchers. We use this model to rank webpages for each of the 200 queries or keywords that we are analyzing in this research. Then, this rank will be added as an additional feature to the dataset that will be used for training. By doing this, we can find how much the new feature, called text relatedness rank of a webpage affects the final rank of that webpage in search results. . حال که دیتاست ما با ده هزار وبسایت و 38 پارامتر آماده شده است، به سراغ بخش دیتاساینس پروژه میرویم که به منظور حفظ شفافیت به صورت انگلیسی ارائه شده است: 1. Training and fine-tuning models on training and validation data: As this is a ranking problem, we need to use Learning to Rank (LTR) algorithms, notably XGBoost, LightGBM and RankNet that use supervised machine learning to solve ranking algorithms [21]. We will divide our data into two segments: training and test data. We will keep the unseen test data for final evaluation, and we use k-fold cross validation to train our model [22]. In general, with cross validation, we divide our data into k folds, but for LTR algorithms, cross validation is a bit different since we need to divide our keywords into train and validation parts. Also, for the test data it is the same, and we need to keep part of our keywords and related webpages to those keywords as our test set. To evaluate our models, we use a metric called Normalized Discounted Cumulative Gain (NDCG) that takes ranking into account [26]. We can use NDCG@K with K as the number of results which in our case is 50 because we consider the first 5 pages of Google search results. For example, we know the true Google ranking for the first 50 webpages related to a specific keyword. We put these webpages in a table and change the ranking. Then, we ask the model to rank this list, and we compare this ranking that is generated by our model to the true Google ranking for that keyword. Our models involve parameters that can be tuned to improve the model results. We will use randomized search cross validation (CV) that is a famous hyper parameter tuning to fine tune our model and choose a combination of parameters that results in a better NDCG score. 2. Testing the models on test data: We will check the performance of the trained models against unseen test data based on NDCG@50 evaluation metric [22]. 3. Selecting the best option: We will compare the models and find the one that has the best result and can better predict on our test data. 4. Framework Construction After we selected the model that can help us know which factors are more important in SEO, we should construct a framework. In this framework, we explain the prediction of ranks with the selected model by creating a feature importance table to show and rank the importance of each factor. In more details, more important features have a higher impact on the webpage’s rank, and online marketers can use this table to prioritize their plans. Also, we visualize our explanation using some techniques like SHAP (SHapley Additive exPlanations) values. SHAP values help us better understand complex machine learning models [27]. The feature importance shows us which features have a higher impact on the rank of a webpage in search engine results, and SHAP values define this impact is positive or negative. Also, we try to provide actionable insights for online marketers and content creators and explain the results in a clear and simple way and help them how to use the model for different industries and keywords. Based on these results, online marketers know where to focus and which features are more important and should be improved as a higher priority when they want their webpages to appear on the first Google search results. دو منبع اصلی که پروژه از آنها الهام گرفته شده است به پروژه ضمیمه شده اند که در صورت نیاز میتوانید به آن مراجعه فرمایید. سقف بودجه ما مبلغ 8 میلیون تومان میباشد.