Changes

2018-05-18: Cleaned up demo_day_classifier directory and fleshed out the writeup on the page. 2018-05-16: Wrote a script (classify_all_accelerator.py) to pull all of the unclassified accelerators from the master variable list (if they are not already in the Cohort List page), and then classify them. This works best if the creation years are provided in the Master Variable List. Started the run on the whole dataset. This will definitely pull up a lot of duplicate results, so it might be valuable to run a program at the end to remove duplicates. 2018-05-11/12: Ran on data, predicted html files are saved in positive directory. Also determined that the model extremely overfits, more data is probably the only fix. 2018-05-06: Changed crawl_and_classify so that the html pages are separated based on what they are predicted to be. Added support for individual year searching. Started running on actual data. Tuned hyperparameters too, should save to params.txt. 2018-05-04: Same. Also cleaned up directory, wiki. Model now achieves 0.80 (+/- 0.15) accuracy. 2018-05-03: Played around with different features and increased dataset. 2018-04-23: So auto-generated features actually reduces accuracy, probably because there isn't enough data. I've gone back to my hand picked features and I'm just focusing on making the dataset larger. 2018-04-16: Still working through using auto-generated features. It takes forever. :/ I reduced the number of words looked at to about 3000. This makes it a lot faster, and seems like it should still be accurate, because the most frequent words are words like "demo" and "accelerator". I also switched from using beautiful soup for text extraction to [https://github.com/aaronsw/html2text html2text]. I might consider using [https://nlp.stanford.edu/IR-book/html/htmledition/sublinear-tf-scaling-1.html Sublinear tf scaling] (parameter in the tf model). 2018-04-16: I think I'm going to transition from using hand-picked feature words to automatically generated features. [http://scikit-learn.org/stable/tutorial/text_analytics/working_with_text_data.html This webpage] has a good example. I could also use n-grams, instead of unigrams. I might also consider using a SVM instead of a random forest, or a combination of the two. 2018-04-12: Continued increasing the dataset size as well as going back and correcting some wrong classifications in the dataset. I'm wondering whether the accuracy would be improved most by an increased dataset, a different approach to features, or changes to the model itself. I am considering using something like word2vec with, for example, five words before and after each instance of the words "startup" or "demo day" in the pages. The problem with this is that this would need its own dataset (which would be easier to create). However, semantic understanding of the text might be an improvement. Or, maybe, I could train this on the title of the article, because the title should have enough semantic meaning. But even this dataset might have to be curated, because a lot of the 0's are demoday pages, they just don't list the cohorts. 2018-04-11: Increased the dataset size using the classifier. Ironed out some bugs in the code. 2018-04-09: Wrote the code to put everything together. It runs the google crawler, creates the features matrix from the results, and then runs the classifier on it. This can be used to increase the size of the dataset and improve the accuracy of the classifier.  * Steps to train the model: Put all of the html files to be used in DemoDayHTMLFull. Then run web_demo_features.py to generate the features matrix, training_features.txt. Then, run demo_day_classifier_randforest.py to generate the model, classifier.pkl. * Steps to run: In the file crawl_and_classify.py, set the variables to whatever is wanted. Then, run crawl_and_classify using python3. It will download all of the html files into the directory CrawledHTMLPages, and then it will generate a matrix of features, CrawledHTMLPages\features.txt. It will then run the trained model saved in classifier.pkl to predict whether these pages are demo day pages, and then it will save the results to predicted.txt. 2018-04-05: The classifier doesn't work as well when there is an imbalance of positive and negative training cases, so I made it use the same number of cases from each. Also had a meeting, my next task is to run the google crawler to create a larger dataset, which we can then use to improve the classifier. 2018-04-04: Continued to classify examples, and tried using images as features. It didn't give great results, so I abandoned that. Currently the features are wordcounts in the headers and title. I might consider the number of "simple" links in the page, like "www.abc.com". Complicated links would be used a lot for many things, but simple links are often used to bring someone to a home page, as in a cohort company's page, so this might be a good indicator of a list of cohort companies. 2018-04-02: Mostly worked on increasing the size of the dataset. So far, with just 30ish more examples, there was a ton of increase in accuracy, so I'm just going to spend the rest of the time doing this. As seen in the graph, this is not entirely due to just the size of the data set, but probably the breadth of it. Also used scikit learns built in hyperparameter grid search, will run this overnight once the dataset is large enough. Another thing I'm thinking about is adding images as a feature, for example the number of images over 200 px wide, because some demo day pages use logos as the cohort company list. [[File:Datasizetoaccuracy.png|300px]] This graph the number of training examples given versus the accuracy. 2018-03-28: Changed to using sykitlearn random forest instead of tensorflow, because this would allow me to see which features have a lot of value and might be affecting the model negatively. One observation I made is that certain years affect the model highly... Maybe I should generalize it for the occurrence of any year. Also, I discovered that just using hand-picked features improved accuracy by 10% rather than using all of the word counts. After that, the only other feature I can think of is the number of images in the page or in the center of the page, because often there are images with all of the cohort companies' logos. Tomorrow I am also going to work on hyperparameter tuning and increasing the amount of data we have.