hexiangnan / sigir16-eals Goto Github PK

Hello,
I'm having problem implement your code.
The data we use is movieLens100k, and we would like to have the output of predicted rating matrix. Thank you very much.

Hi, Prof. He!
In the ItemKNN.java, the similarity matrix is a diagonal matrix but the method buildModel_items only seems to calculate the values in the lower triangular part and doesn't assign the values to the symmetry elements.
The experimental results show that there are some differences between using itemCount && j != i and itemCount in line 5 (line 71 of the original code).

protected void buildModel_items(int startItem, int endItem) {
		// Build the similarity matrix for selected items.
		for (int i = startItem; i < endItem; i++) {
			HashMap<Integer, Double> map_item_score = new HashMap<Integer, Double>();
			for (int j = 0; j < itemCount && j != i; j++) {
				// Cosine similarity
				double score = trainMatrix.getColRef(i).innerProduct(trainMatrix.getColRef(j));
				if (score != 0) {
					score /= (lengths[i] * lengths[j]);
					map_item_score.put(j, score);
				}
			}
			if (K <= 0) { // All neighbors
				for (int j : map_item_score.keySet()) {
					similarity.setValue(i, j, map_item_score.get(j));
				}
			} else { // Only K nearest neighbors
				for (int j : CommonUtils.TopKeysByValue(map_item_score, K, null)) {
					similarity.setValue(i, j, map_item_score.get(j));
				}
			} // end if
		} // end for

Hi, Professor He. Recently I'm trying to test e-als on ml-1m and some other data sets, but I'm not so familiar with programming with JAVA. I've preprocessed the data sets in the same form as you shown in the yelp data set. However, without any modification, when I run main_MF.java to test the offline performance, I got an extremely bad result, i.e., 0.4xxx for HR and 0.01xx for NDCG. This is unacceptable since a simple MF algorithm can have much better performance. Do you have any suggestion about how to reproduce your results? (ml-1m is tested in your paper "Neural Collaborative Filtering" but I can't reproduce the same results) Or do I need to modify some critical parameters?

Thanks.

Here is the parameters I used:

Data	data/ml-1m.rating
#Users	6040
#Items	3706
#Ratings	 994169 (train), 6040(test)
FastALS: showProgress=false, factors=64, maxIter=500, reg=0.010000, w0=10.00, alpha=0.75