Resulting Rules:
After applying Apriori algorithm for association rule mining for our dataset, approximately 6300 rules were generated with support as 0.01 and confidence as 0.5. A snippet of these rules can be seen below. The top 15 rules for support, the top 15 rules for confidence, and the top 15 rules for lift can also be found in the snippets below in the respective order. Entire results of each of these can be found in the google colab notebook here.
Visualizations:
Network graphs play a crucial role in visualizing the relationships between items or features in association rule mining. By representing items as nodes and the associations between them as edges, network graphs provide a clear and intuitive depiction of the underlying patterns discovered by the mining process. This visualization helps analysts to identify frequent itemsets, assess the strength of associations, and comprehend the structure of rules more effectively. Moreover, network graphs enable the identification of clusters or communities of related items, facilitating the interpretation and extraction of actionable insights from the mined rules. Overall, network graphs enhance the interpretability and usability of association rule mining results, aiding decision-making and knowledge discovery processes.
Analyzing the bar graph:
The rule with the most support is "{true} => {worldnews}". This rule states that if something is classified as true, it is very likely to be classified as world news. The rule with the least support is "{Government News} => {fake}". This rule states that if something is classified as government news, it is most likely to also be classified as fake. There are several other rules that have very little support, including "{left-news} => {fake}", "{Middle-east} => {fake}", "{News} => {fake}", and "{politics} => {fake}". These rules all suggest that people who support the other rules are skeptical of information from these sources. The graph does not show the total number of people who expressed support for each rule, so it is difficult to say definitively how widespread these views are just from this bar graph.
The wide range suggests the model is trained on varying amounts of data for different types of fake news information. This could potentially impact the overall performance and solve the problem of biases. The centralized distribution indicates the model is generally confident in its predictions, but there might be individual cases where the confidence score is lower where it suggests less certainty. The tight clustering implies the model consistently provides a significant improvement in detecting fake news compared to the baseline across different types.
The correlation graph below shows us that Support and confidence have a weak positive correlation, Support and lift have a weak negative correlation, and Confidence and lift have a moderate negative correlation. We know that, correlation does not imply causation. Just because two variables are correlated does not mean that one variable causes the other to change.
Results and Conclusions:
High Confidence Rules: Rules with a high confidence value (close to 1) indicate strong associations between the antecedent (lhs) and consequent (rhs) items. For example, the rules {subject=politics} => {target=fake} and {subject=News} => {target=fake} have a confidence of 1, suggesting that if an article belongs to the "politics" subject or the "News" subject, it is highly likely to be classified as fake news.
Support and Lift: Support indicates the proportion of transactions that contain both the antecedent and consequent items. Lift measures how much more likely the consequent is given the antecedent, compared to its overall occurrence. Higher lift values indicate stronger associations. For example, the rule {subject=politics} => {target=fake} has a lift of 1.92, indicating that articles about politics are almost twice as likely to be fake news compared to the baseline probability.
Coverage: Coverage represents the proportion of transactions that contain the antecedent item. Higher coverage values indicate that the antecedent item is more frequent in the dataset. For example, the rule {subject=worldnews} => {target=true} has a coverage of 0.24, meaning that 24% of the transactions contain articles about world news.
Count: Count shows the absolute frequency of occurrences of the rule in the dataset. For example, the rule {subject=News} => {target=fake} has a count of 222, indicating that this rule applies to 222 transactions in the dataset.
Overall, these association rules provide insights into the relationships between different attributes (e.g., subject, cluster) and the target variable (fake or true news). You can use this information to understand patterns in your data and potentially identify characteristics associated with fake news articles.