MTH522

Project2-Resubmission

Project1 resubmission

Statistical techniques like the Autocorrelation Function (ACF) are vital for interpreting time series data, as they quantify self-correlations in the series over different time lags. Positive autocorrelation signifies that current and past trends are similar, while negative values indicate an inverse relationship. The ACF uncovers enduring patterns, enabling more accurate forecasts.

It is extensively utilized in domains like economics and environmental science where predicting future behaviors is crucial but relies on grasping historical contingencies. For example, accurately forecasting stock prices requires knowing market autocorrelations, while reliable weather prediction depends on modeling meteorological autocorrelations over time. By revealing meaningful sequential patterns, the ACF allows researchers across fields to anticipate events more dependably. It serves as an essential numerical tool within time series analysis for deciphering structures in temporal data.

The suitability of time series and LSTM models is highly contingent on the data properties and patterns observed. Through experience, I have realized predictive success is rooted in selecting the right model and customizing it for the dataset. As I advance in data analysis, the knowledge gained from applying these models remains integral in shaping my perspective towards sequential data.

Time series forecasting is an essential analytical method to reveal trends and patterns hidden within time-based data, going beyond conventional statistical approaches. It empowers informed decisions by leveraging historical information, decoding temporal relationships, and projecting future outcomes.

In data science and forecasting, time series analysis is pivotal, offering a window into how phenomena progress over time. It enables the dissection of historical data to uncover seasonality, cyclic behavior, and directional tendencies. With this understanding, we can venture into prediction, generating invaluable insights to guide decision-making across domains.

I aimed to rephrase your points concisely while preserving the essential concepts about the significance of time series analysis and forecasting.

My work with Long Short-Term Memory networks (LSTMs) has offered valuable insights. The distinguishing capability of LSTMs is handling extended dependencies in sequential data, a common challenge. Their integrated memory cell and three specialized gates—forget, input and output—allow LSTMs to selectively retain or discard information. This empowers them to capture pertinent information over long sequences, proving extremely useful for my natural language processing and time series projects.

Additionally, investigating time series models has been rewarding. Time series analysis assumes data points collected over time are interrelated and order matters. I have focused largely on two time series model types: univariate and multivariate. While univariate models like ARIMA and Exponential Smoothing spotlight trends and seasonality in individual variables, multivariate models like Vector Autoregression (VAR) and Structural Time Series provide a bigger picture by examining multiple interrelated variables.

I aimed to express the key points in my own words while maintaining the meaning you conveyed about LSTM networks and time series analysis.

Project3

Information Gain stands as a pivotal concept in machine learning, especially in the realm of decision tree algorithms. It serves to quantify how effectively a feature can partition data into target classes, providing a means to prioritize features at each decision point. Essentially, Information Gain measures the difference in entropy before and after splitting a set on a specific attribute.

On the other hand, in the context of forecasting, there’s the method of simple exponential smoothing (SES). Ideal for data lacking strong trends or seasonality, SES assumes that future values will predominantly reflect the most recent observations, giving less weight to older data. This approach is characterized by historical weighting, simplicity in its input requirements, adaptability based on past errors, and a focus on recent data. By emphasizing the most recent information, SES streamlines pattern identification and minimizes the impact of noise and outliers in older data, making it particularly adept at forecasting in dynamic environments where variables exhibit volatility.

Today’s progress involved several steps in working with the code. Initially, I successfully conducted a count of parameters from the custom dataset, revealing that offense code 3115 was the most common. Subsequently, I endeavored to extract the data associated with offense code 3115 into a new dataset for exclusive analysis using latitude and longitude. Although the creation of the new dataset was successful, I encountered a mismatch issue when plotting the data on the map of Boston, specifically with latitude and longitude parameters. I am actively working to resolve this discrepancy.

In addition to spatial analysis, I aimed to plot offense codes along with their frequencies using matplotlib. Unfortunately, an error message pertaining to an invalid built-in function has surfaced. This is perplexing, considering the success of a similar method on another dataset. I am currently investigating the source of this error and will rectify it to proceed with plotting the graph. Furthermore, I plan to generate a Pareto Curve by the weekend, offering a comprehensive analysis of the dataset.

Today, my focus remained on the crime incident report dataset, where I worked on developing a code to extract data for specific crime codes over the span of eight years. The goal is to create a graph plotting latitude and longitude parameters, providing a visual representation of crime distribution in Boston. Additionally, I explored the Pareto curve method as a valuable tool for analyzing the dataset. This method involves plotting individual values in descending order, combining both bar and line charts. The line chart represents the cumulative total of the dataset, offering insights into the percentage contribution of each crime to the overall incident reports. I believe this Pareto curve will provide a nuanced understanding of how police allocate resources and which offenses dominate their workload. In the forthcoming days, my aim is to present the data through well-crafted graphs and visual representations.

Today, I continued to work over analyzing the data from the crime incident report. With initial analysis of the data in the last time, where the maximum number of incidents were reported for investigating a particular person, I started working towards analyzing the insights related to that parameter. I am currently working on a code, through which I will aim to combine the data of that particular crime or offense code with the parameters of longitude and latitude. Combining both the data parameters, we can try to work on the neighborhood which has the most reports for these types of incidents and hence, with that we should be able to say the particular crime prone neighborhood. In the coming days, I will continue to work over the code and also discuss this further with my team