Unleash the Hidden Patterns: A Guide to Unsupervised Machine Learning for Article Recommender System | by Suhas Maddali | Mar, 2023

There are title, text, and language features that are most important in our data. Other features such as content type and author user agent are not very suitable for this task of building a recommendation system. Steps would be taken to remove such features and only consider the features such as title, text, and language respectively. Note that not all the features are shown in the figure as this would become a large image.

We get the information about the data and determine the number of non-null values present in the dataset. It is seen that there are features such as authorSessionID and authorUserAgent that contain more than 50 percent missing values. We are going to be removing all of these features along with others as discussed above and would be focusing only on the title, text, and language features.

Exploratory Data Analysis (EDA)

Let us now explore the dataset and get an understanding of it by using visuals and graphs. The first step would be to explore the type of text content present in the data.

Observing from the plot, a large portion of the contents are in the form of HTML links to be used for recommendation. These HTML links contain the actual text data and the titles. Our focus here is to extract meaningful insights from text instead of using other formats such as ‘video’ and ‘rich’ formats. Therefore, we can remove these categories from our project.

A large portion of our text and titles are English titles. Since our focus is only on the English language, we can take the steps to remove other categories to improve our recommendation model efficiency.

Wordcloud gives a good understanding of the most frequently occurring words in our corpus of text. The higher the occurrence of words in a corpus, the bigger the text sizes in a word cloud plot. As seen from the plot, word titles having terms such as ‘Google’, ‘Machine Learning’, and ‘Apple’ are often used in our title space.

Now that we explored the title and found the most commonly occurring words, it is time to explore the text itself in our list of articles to find interesting trends and patterns. A large proportion of articles contain words such as ‘data’, ‘user’, and ‘time’.

Converting the given set of titles into a TFIDF format and applying principal component analysis, the above plot showcases the cumulative percentage of variance explained by each of the components. It is to be noted that with the increase in the number of components, there is an increase in the variance explained as more information from components leads to more explanation and information.

This plot shows the variance explained by the tfidf components on the entire text instead of only considering the title. This plot is quite different from the previous plot in that only a few sets of principal components are able to explain a large portion of the variance in the data. Therefore, this can help in dimensionality reduction as fewer features are able to explain a large portion of the variance.

Clustering is a technique where data points that are similar are grouped together to find interesting patterns and commonalities among them. In this way, recommendations are given based on the cluster at which a data point is present.

In order to determine the right number of clusters, it is handy to use the k-means clustering model and follow the elbow method to find the optimum number of clusters. In our case, the best value of k is 11 as this follows the shape of an elbow.

After performing PCA and clustering based on the optimum number of clusters, it is time to visualize the results of the clustering. Based on the plot above, the clustering is working quite well as there is a pattern found in the clusters.

Let’s determine how the plot looks in 3D to find the underlying patterns. As can be seen, there is a lot of room for the separation of clusters, hence guiding our recommender system to give good suggestions to the user based on the previously read text.

There are other dimensionality reduction techniques such as TSNE and Kernal PCA. Going through each of them to determine the best clustering would be handy for a recommender system. Steps must be taken to visualize the data points from the text and generate interesting patterns.

Reducing the dimensions using TSNE and visualizing the representations with 11 clusters, it is seen that the data points are spread out quite well. Therefore, there are lower chances for articles that are similar to be in the same cluster. As a result, PCA was performing well in terms of clustering and determining the optimum number of clusters. We will also use 3D visualizations to guide our thinking and understanding of clustering mechanisms.

After performing the task of clustering, it is seen that a lot of data embeddings of the text are quite close to each other. Therefore, it can be challenging to accurately perform clustering when the points are not spread in various directions. Hence, we might look for alternative clustering approaches and dimensionality reduction techniques.

Kernel PCA is another popular method that is used for dimensionality reduction. As could be seen, the data representations are spread out well for determining and using the number of clusters. Overall, the algorithm reduces the dimensions well and separated data points well. Let us also go over the 3d representation of clusters as a result of performing dimensionality reduction with this technique.

After plotting a 3d generated representation with kernel PCA, clustering took place quite well and the points are quite spread out. Therefore, this approach can help generate recommendations based on the clustering approach.

Say, a user visits a particular website and reads an interesting piece of text. After this step, the items that are present in the cluster based on kernel PCA representation would be recommended to the user. The user can find these articles fascinating, resulting in the growth of the business.

After performing the previous steps, we define a function that generates a list of useful features that could be used by various recommender system models to make recommendations. These features used in the function below are important and give a good representation of the type of text along with its content and readability.

We apply this function to our data frame and generate a new set of features that are used by recommender systems. Finally, cosine similarity is taken into account when determining the distance between the text of interest with the list of all the other possible texts and articles.

After getting the cosine difference between the present text based on the features generated and a kernel PCA representation, it is compared with the list of existing articles to determine the ones with the lowest distance from clusters. As a result, those articles are recommended to users which makes them engaging and fun to read.

We import various useful libraries for measuring the cosine similarities and generating recommendations for articles that are mostly similar in structure and content.

Conclusion

After going through this exhaustive article, you should be getting a good idea about the working details of a particular implementation of recommendation systems on articles. Performing dimensionality reduction can ensure that we would reduce compute resources and also reduce the impact of outliers when making predictions or providing recommendations. Thanks for taking the time to read this article.

There are title, text, and language features that are most important in our data. Other features such as content type and author user agent are not very suitable for this task of building a recommendation system. Steps would be taken to remove such features and only consider the features such as title, text, and language respectively. Note that not all the features are shown in the figure as this would become a large image.

We get the information about the data and determine the number of non-null values present in the dataset. It is seen that there are features such as authorSessionID and authorUserAgent that contain more than 50 percent missing values. We are going to be removing all of these features along with others as discussed above and would be focusing only on the title, text, and language features.

Exploratory Data Analysis (EDA)

Let us now explore the dataset and get an understanding of it by using visuals and graphs. The first step would be to explore the type of text content present in the data.

Observing from the plot, a large portion of the contents are in the form of HTML links to be used for recommendation. These HTML links contain the actual text data and the titles. Our focus here is to extract meaningful insights from text instead of using other formats such as ‘video’ and ‘rich’ formats. Therefore, we can remove these categories from our project.

A large portion of our text and titles are English titles. Since our focus is only on the English language, we can take the steps to remove other categories to improve our recommendation model efficiency.

Wordcloud gives a good understanding of the most frequently occurring words in our corpus of text. The higher the occurrence of words in a corpus, the bigger the text sizes in a word cloud plot. As seen from the plot, word titles having terms such as ‘Google’, ‘Machine Learning’, and ‘Apple’ are often used in our title space.

Now that we explored the title and found the most commonly occurring words, it is time to explore the text itself in our list of articles to find interesting trends and patterns. A large proportion of articles contain words such as ‘data’, ‘user’, and ‘time’.

Converting the given set of titles into a TFIDF format and applying principal component analysis, the above plot showcases the cumulative percentage of variance explained by each of the components. It is to be noted that with the increase in the number of components, there is an increase in the variance explained as more information from components leads to more explanation and information.

This plot shows the variance explained by the tfidf components on the entire text instead of only considering the title. This plot is quite different from the previous plot in that only a few sets of principal components are able to explain a large portion of the variance in the data. Therefore, this can help in dimensionality reduction as fewer features are able to explain a large portion of the variance.

Clustering is a technique where data points that are similar are grouped together to find interesting patterns and commonalities among them. In this way, recommendations are given based on the cluster at which a data point is present.

In order to determine the right number of clusters, it is handy to use the k-means clustering model and follow the elbow method to find the optimum number of clusters. In our case, the best value of k is 11 as this follows the shape of an elbow.

After performing PCA and clustering based on the optimum number of clusters, it is time to visualize the results of the clustering. Based on the plot above, the clustering is working quite well as there is a pattern found in the clusters.

Let’s determine how the plot looks in 3D to find the underlying patterns. As can be seen, there is a lot of room for the separation of clusters, hence guiding our recommender system to give good suggestions to the user based on the previously read text.

There are other dimensionality reduction techniques such as TSNE and Kernal PCA. Going through each of them to determine the best clustering would be handy for a recommender system. Steps must be taken to visualize the data points from the text and generate interesting patterns.

Reducing the dimensions using TSNE and visualizing the representations with 11 clusters, it is seen that the data points are spread out quite well. Therefore, there are lower chances for articles that are similar to be in the same cluster. As a result, PCA was performing well in terms of clustering and determining the optimum number of clusters. We will also use 3D visualizations to guide our thinking and understanding of clustering mechanisms.

After performing the task of clustering, it is seen that a lot of data embeddings of the text are quite close to each other. Therefore, it can be challenging to accurately perform clustering when the points are not spread in various directions. Hence, we might look for alternative clustering approaches and dimensionality reduction techniques.

Kernel PCA is another popular method that is used for dimensionality reduction. As could be seen, the data representations are spread out well for determining and using the number of clusters. Overall, the algorithm reduces the dimensions well and separated data points well. Let us also go over the 3d representation of clusters as a result of performing dimensionality reduction with this technique.

After plotting a 3d generated representation with kernel PCA, clustering took place quite well and the points are quite spread out. Therefore, this approach can help generate recommendations based on the clustering approach.

Say, a user visits a particular website and reads an interesting piece of text. After this step, the items that are present in the cluster based on kernel PCA representation would be recommended to the user. The user can find these articles fascinating, resulting in the growth of the business.

After performing the previous steps, we define a function that generates a list of useful features that could be used by various recommender system models to make recommendations. These features used in the function below are important and give a good representation of the type of text along with its content and readability.

We apply this function to our data frame and generate a new set of features that are used by recommender systems. Finally, cosine similarity is taken into account when determining the distance between the text of interest with the list of all the other possible texts and articles.

After getting the cosine difference between the present text based on the features generated and a kernel PCA representation, it is compared with the list of existing articles to determine the ones with the lowest distance from clusters. As a result, those articles are recommended to users which makes them engaging and fun to read.

We import various useful libraries for measuring the cosine similarities and generating recommendations for articles that are mostly similar in structure and content.

Conclusion

After going through this exhaustive article, you should be getting a good idea about the working details of a particular implementation of recommendation systems on articles. Performing dimensionality reduction can ensure that we would reduce compute resources and also reduce the impact of outliers when making predictions or providing recommendations. Thanks for taking the time to read this article.