Turning Google Search Data into Predictions

Every day, millions of people turn to Google Search to find information about various topics, from new cars to flu symptoms. The words they type into the search bar contain important information about what they are looking for and what they are doing. However, making sense of these search terms has not been easy. Typically, users have used categories to filter search data, but this method often misses a lot of details.

In this study, we introduce a new way to condense search data into a smaller size while keeping the essential information from the individual terms, without relying on user-defined categories. Our approach includes two main ideas: first, we propose a method called SLaM Compression, which uses pre-trained language models to create a summary of search data. Second, we present a model called CoSMo, which estimates real-world events using only search data. We show that our methods can accurately predict U.S. car sales and flu rates using only Google Search data.

#The Importance of Google Search Data

Google Search is the leading search engine globally, providing a wealth of information about the terms that users search for and their connection to real-world events, such as purchasing behavior, economic activity, or health trends. Research has already shown that Google search data can improve predictions and models. The current methods mainly use two types of data: Google Trends and search logs.

Google Trends organizes search terms into categories and gives an index value for search volume based on the category for specific days and regions. While useful, this method treats diverse queries as if they belong to the same group, limiting the depth of analysis. For example, it groups all car-related searches without distinguishing between the types of cars. Researchers have used this data to predict economic activities and other trends, but they usually rely on additional information, such as historical sales data.

On the other hand, search logs contain pairs of search terms and how often they were searched over a certain time. Although search logs offer more detailed data, they also present challenges due to the sheer number of unique terms, making it hard to convert this data into manageable features for models. Some researchers have filtered terms or used one-hot encoding for specific searches to make it more digestible.

In our work, we aim to summarize search logs more effectively, allowing us to use them for prediction tasks without the need for extensive filtering.

#Our Approach

We divide our modeling strategy using search data into two main parts: 1) condensing search data into useful features and 2) selecting a model that fits these features.

We leverage language models to reduce the complexity of search data while retaining meaningful information. Instead of mapping search terms to binary vectors, we use language models to represent terms as points in a high-dimensional space. We then combine these search terms into a single representative vector, which we call a search embedding.

With this framework, we can automatically create search embeddings without needing user-defined filters, allowing for flexibility in the timeframe used for analysis. Our method provides a memory-efficient representation of search data that is still very effective for prediction.

#SLaM Compression

SLaM Compression works by taking all the searches within a specific time frame and condensing them into a fixed-length vector that summarizes all search terms. Each search term is transformed into a fixed-length vector by a language model, allowing us to group similar terms together based on their meaning.

This process helps us capture the nuances of search terms without generating an overwhelming amount of data. Our compression method does not require filtering search terms in advance, enabling us to work with larger datasets without losing important information.

We break our representation into two parts: the total search volume and the normalized search embedding. By leveraging search volume data along with our search embeddings, we can establish connections between individual search terms and broader trends.

#CoSMo Model

The CoSMo model is designed to predict real-world events using the search embeddings we generate. Instead of relying on complex filtering or categorization, CoSMo uses a more straightforward approach that allows for flexibility in the data being analyzed.

Using the search embeddings, CoSMo outputs a score indicating the likelihood of a given event occurring based on user search terms. The flexibility of our model allows it to adapt to different regions and timeframes, leading to more accurate predictions.

#Real-World Applications

We test our methods using two real-world examples: predicting flu rates and U.S. auto sales. Through these case studies, we demonstrate how our approach can significantly enhance the accuracy of predictions based solely on search data.

#Predicting U.S. Auto Sales

When predicting auto sales, we compare our results with existing methods. By using our search embeddings, we improve accuracy from approximately 58% to 75%. This means our model can better capture the connection between search queries and actual sales figures.

Our model can account for regional differences in search behavior and adoption, making it more adaptable and accurate in various contexts. With our method, we successfully predicted sales trends without relying on historical data or external variables, which suggests that our approach can hold promise for broader economic predictions.

#Predicting Flu Rates

For flu prediction, we model rates of Influenza-Like Illness (ILI) at the national level. We use Google Search data related to flu symptoms to forecast flu rates over several years.

Our model similarly performs well, estimating actual flu rates closely and demonstrating the potential of search data to provide insights into public health trends. Unlike traditional methods that often rely on historical data and external factors, our model uses only search patterns, highlighting the efficacy of our approach in public health monitoring.

#Model Performance and Testing

We evaluate our methods extensively using various experimental setups. For both automotive sales and flu predictions, we compare our performance against previous models and methods to show the improvements our approach brings to the table.

#Automotive Sales Experiments

We benchmark our model against existing models in forecasting vehicle sales. We observe a considerable boost in predictive accuracy when using our search embeddings compared to traditional classification methods. Even with a simple model structure, our method manages to capture complex relationships between search behavior and sales outcomes.

#Flu Rate Experiments

For flu rate predictions, we conduct similar experiments. Our method performs better than other models that only utilize search data. We also explore different variations of our model to identify optimal configurations, optimizing performance for different flu seasons.

#Insights from the Model

One valuable aspect of our approach is the interpretability of the model. We can analyze how individual search terms contribute to the overall predictions, allowing us to understand the factors driving search behavior and their implications for real-world events.

By examining the search terms associated with high scores, we reveal how users interact with search engines in relation to flu symptoms. This insight not only helps refine our model but also informs public health strategies and marketing approaches.

#Handling Misspellings and Variability

Our method proves capable of managing tasks such as misspellings and synonyms effectively. The language models we utilize can understand variations of search terms, enhancing our model’s robustness and reliability.

#Future Directions

Although we have demonstrated the potential of our methods, there are still opportunities for further exploration and refinement. We look forward to applying our approach to other domains and refining our models to achieve even greater accuracy and flexibility.

#Expanding to Other Areas

We believe that the methods we've developed can be beneficial in many other areas beyond flu predictions and automotive sales. Our approach could be extended to various industries, including retail, sports, and more, tapping into the rich insights that Google Search data provides.

#Enhancements and Adaptations

As technology evolves, we will continue to adapt our methods to leverage advancements in language modeling and machine learning. By integrating new tools and techniques, we can refine our models, enhance their predictive capabilities, and provide more accurate insights into consumer behavior and trends.

#Conclusion

Our study illustrates the significant value of Google Search data in creating predictive models. By developing SLaM Compression and CoSMo, we have found ways to summarize search data effectively while retaining essential information. These methods not only improve predictive power across various contexts but also provide interpretable insights that can inform decision-making.

As we move forward, we aim to expand our approach's applicability, demonstrating the versatility and strength of using language models in understanding and predicting real-world events through search data. With billions of searches happening every day, there are countless opportunities to harness this information for better predictions and insights across multiple fields.