I created a Wit.AI entity with about 10k values. When I input a message that is exactly one of the the possible values, Wit.AI doesn't seem to be able to pick that up. I tried this with an entity that has only a few values and it worked. Is this because of the amount of values that the entity has?
If your Wit app has stories, the stories may be taking precedence over any other training data you have provided (e.g. entities and their values), because Wit intentionally overfits on stories.
Make sure you delete any stories if you want your AI to be sensitive to a large training dataset.
Related
I'm kinda new to Neural Networks and just started to learn coding them by trying some examples.
Two weeks ago I was searching for an interesting challenge and I found one. But I'm about to give up because it seems to be too hard for me... But I was curious to know if anyone of you is able to solve this?
The Problem: Assume there are ".htm"-files that contain tables about the same topic. But the table structure isn't the same for every file. For example: We have a lot ".htm"-files containing information about teachers substitutions per day per school. Because the structure of those ".htm"-files isn't the same for every file it would be hard to program a parser that could extract the data from those tables. So my thought was that this is a task for a Neural Network.
First Question: Is it a task a Neural Network can/should handle or am I mistaken by that?
Because for me a Neural Network seemed to fit for this kind of a challenge I tried to thing of an Input. I came up with two options:
First Input Option: Take the HTML Code (only from the body-tag) as string and convert it as Tensor
Second Input Option: Convert the HTML Tables into Images (via Canvas maybe) and feed this input to the DNN through Conv2D-Layers.
Second Question: Are those Options any good? Do you have any better solution to this?
After that I wanted to figure out how I would make a DNN output this heavily dynamic data for me? My thought was to convert my desired JSON-Output into Tensors and feed them to the DNN while training and for every prediction i would expect the DNN to return a Tensor that is convertible into a JSON-Output...
Third Question: Is it even possible to get such a detailed Output from a DNN? And if Yes: Do you think the Output would be suitable for this task?
Last Question: Assuming all my assumptions are correct - Wouldn't training this DNN take for ever? Let's say you have a RTX 2080 ti for it. What would you guess?
I guess that's it. I hope i can learn a lot from you guys!
(I'm sorry about my bad English - it's not my native language)
Addition:
Here is a more in-depth Example. Lets say we have a ".htm"-file that looks like this:
The task would be to get all the relevant informations from this table. For example:
All Students from Class "9c" don't have lessons in their 6th hour due to cancellation.
1) This is not particularly suitable problem for a Neural Network, as you domain is a structured data with clear dependcies inside. Tree based ML algorithms tend to show much better results on such problems.
2) Both you choices of input are very unstructured. To learn from such data would be nearly impossible. The are clear ways to give more knowledge to the model. For example, you have the same data in different format, the difference is only the structure. It means that a model needs to learn a mapping from one structure to another, it doesn't need to know any data. Hence, words can be Tokenized with unique identifiers to remove unnecessary information. Htm data can be parsed to a tree, as well as json. Then, there are different ways to represent graph structures, which can be used in a ML model.
3) It seems that the only adequate option for output is a sequence of identifiers pointing to unique entities from text. The whole problem then is similar to Seq2Seq best solved by RNNs with an decoder-encoder architecture.
I believe that, if there is enough data and htm files don't have huge amount of noise, the task can be completed. Training time hugely depends on selected model and its complexity, as well as diversity of initial data.
I am currently developing a neural network that tries to make a suggestion for a specific user based on his recent activities. I will try to illustrate my problem with an example.
Now, let's say im trying to suggest new music to a user based on the music he recently listened to. Since people often listen to artists they know, one input of such a neural network might be the artists he recently listened to.
The problem is the encoding of this feature. As the id of the artist in the database has no meaning for the neural network, the only other option that comes to my mind would be one-hot encoding every artist, but that doesn't sound to promising either regarding the thousands of different artists out there.
My question is: How can i encode such a feature?
The approach you describe is called content-based filtering. The intuition is to recommend items to customer A similar to previous items liked by A. An advantage to this approach is that you only need data about one user, which tends to result in a "personalized" approach for recommendation. But some disadvantages include the construction of features (the problem you're dealing with now), the difficulty to build an interesting profile for new users, plus it will also never recommend items outside a user's content profile. As for the difficulty of representation, features are usually handcrafted and abstracted afterwards. For music specifically, features would be things like 'artist', 'genre', etc. and abstraction for informative keywords (if necessary) is widely done using tf-idf.
This may go outside the scope of the question, but I think it is also worth mentioning an alternative approach to this: collaborative filtering. Rather than similar items, here we instead try to find users with similar tastes and recommend products that they liked. The only data you need here are some sort of user ratings or values of how much they (dis)liked some data - eliminating the need for feature design. Furthermore, since we analyze similar persons rather than items for recommendation, this approach tends to also work well for new users. The general flow for collaborative filtering looks like:
Measure similarity between user of interest and all other users
(optional) Select a smaller subset consisting of most similar users
Predict ratings as a weighted combination of "nearest neighbors"
Return the highest rated items
A popular approach for the similarity weighting in the algorithm is based on the Pearson correlation coefficient.
Finally, something to consider here is the need for performance/scalability: calculating pairwise similarities for millions of users is not really light-weight on a normal computer.
I'm doing a research on the author name disambiguation problem. I want to make some experiments. I want to perform clustering on citation records. My dataset consist of 2000 xml records. I need testing data. The dataset that I'm using is not popular and I need to make testing data manually. I don't know how to do so. I need instruction of how to make testing data manually. Note: I want to compare the performance of a set of techniques in solving the author name disambiguation problem, So I must perform testing.
Even though it is not really clear what kind of testing you want to perform, but general answer to the issue at hand - trying to artificially create more data from the data you have at hand - is a bootstrap. In general it is technique when you perform sampling with replacement from your dataset as many times as you want. It randomly picks up some element from your data repetitively untill you get a sample of the size you want. The sample you get could be larger than your original dataset but should have similar (from statistical point of view) as your original dataset. Bootstrap sampling is available in sklearn.
P.S. You need to keep in mind that this solution is not optimal - best solution to this problem is to actually get more real data somehow.
Classification vs. Clustering
For author name disambiguation, I don't think you want clustering. What you want is classification.
You have a features for each author / publication. Now you give the classifier two of those feature vectors. It classifies "it is the same author" or "those are different authors".
Training / testing data
Having a binary classification problem, the testing suddenly becomes simple: Just use one of the measures used in literature so often (accuracy, precision, recall, confuscation matrix).
Getting the data might be a bit more complicated. You wrote that you have an XML file of 2000 records. I guess you can derive features from those records automatically and authors have an identifier? Then you can simply generate negative examples by having different authors and positive examples by checking if the identifier is the same.
Otherwise you can have a look at http://dblp.uni-trier.de/. Although there are likely many publications under the same author which should be different, they do distinguish authors not only by name but give them identifiers.
Alternatively, you can train a classifier to classify each of the known authors with e.g. > 30 publications. Then remove the softmax layer and use those features to distinguish the authors.
We're starting to build an SSAS tabular model and wondering if most people have one model or multiple. If multiple, do you duplicate tables that are needed by each, or is there a way to share tables between models? I think I know the answer, but I'm hoping those with more experience can confirm what we've found...
From what I've researched I think...
- you can't share tables across models - any "common" tables would have to be duplicated in and deployed with each model and would take up memory
- we should create one model, use perspectives to organize the tables and make it easier to work with
- multiple models could be acceptable if there is little or no common data across models
thanks
You are correct, there is no way to share tables between models.
Perspectives can help.
The question of whether to have one model or more depends on the user audience. Who are the users? How analytically savvy are they? Will they have a reasonable understanding of the model structure?
One issue that affects my rather unsophisticated users is when a dimension does not relate to all fact tables. In this scenario, as is expected, measures on the fact table calculate identical values for every member of the unrelated dimension. For less knowledgeable users, this situation is confusing.
I agree with Ari's answer, and am posting this answer to explain my own experience.
We use a few large models for more sophisticated users that are in memory and processed once a week. We have agreed with the business that these models will not be available during processing, so we are able to processes without holding a transaction open which allows us to keep many more smaller models because we do not need to keep 2x the size of our largest model available to the instance. We use perspectives to simplify the presentation and reduce the confusion for the multiple fact tables. Even with perspectives, the models are rather complex, and it takes some training to get users used to working with the different facts.
We also use smaller models, usually more targeted to a specific audience/need. Many are processed daily, and use transactional processing to ensure the are available to users as much as possible. There are several dimensions that are used in several of our small models, but we are able to filter them so that user's do not see the full list of members which reduces size, and has been a huge benefit for my users, because they only see members that have a fact that they are analyzing instead of a list of every member associated with any fact.
We use views to ensure conformity between models when a dimension is used in multiple models. In my opinion this is very important, as it is very confusing when I have the same dimension with slightly different attribute names.
To sum up (pun intended)...
I like developing and working with large models. I think they answer more questions with less work.
Most users I have worked with prefer smaller, more concise models. Your server hardware/processing requirements may direct you to smaller models as well, even though some of the dimensions will be duplicated.
I am developing an application that uses a neural network. Currently I am looking at either trying to put it into a relational database based on SQL (probably SQL server) or a graph database.
From a performance viewpoint, the neural net will be very large.
My questions:
Do relational databases suffer a performance hit when dealing with a neural net in comparison to graph databases?
What graph-database technology would be best suited to dealing with a large neural net?
Can a geospatial database such as PostGIS be used to represent a neural net efficiently?
That depends on the intent of progress on the model.
Do you have a fixated idea on an immutable structure of the network? Like a Kohonnen map. Or an off-the-shelf model.
Do you have several relationship structures you need to test out, so that you wish to be able flip a switch to alternate between various structures.
Does your model treat the nodes as fluid automatons, free to seek their own neighbours? Where each automaton develops unique characteristic values of a common set of parameters, and you need to analyse how those values affect their "choice" of neighbours.
Do you have a fixed set of parameters for a fixed number of types/classes of nodes? Or is a node expected to develop a unique range of attributes and relationships?
Do you have frequent need to access each node, especially those embedded deep in the network layers, to analyse and correlate them?
Is your network perceivable as, or quantizable into, set of state-machines?
Disclaimer
First of all, I need to disclaim that I am familiar only with Kohonnen maps. (So, I admit having been derided for Kohonnen as being only entry-level of anything barely neural-network.) The above questions are the consequence of personal mental exploits I've had over the years fantasizing after random and lowly-educated reading of various neural shemes.
Category vs Parameter vs Attribute
Can we class vehicles by the number of wheels or tonnage? Should wheel-quantity or tonnage be attributes, parameters or category-characteristics.
Understanding this debate is a crucial step in structuring your repository. This debate is especially relevant to disease and patient vectors. I have seen patient information relational schemata, designed by medical experts but obviously without much training in information science, that presume a common set of parameters for every patient. With thousands of columns, mostly unused, for each patient record. And when they exceed column limits for a table, they create a new table with yet thousands more of sparsely used columns.
Type 1: All nodes have a common set of parameters and hence a node can be modeled into a table with a known number of columns.
Type 2: There are various classes of nodes. There is a fixed number of classes of nodes. Each class has a fixed set of parameters. Therefore, there is a characteristic table for each class of node.
Type 3: There is no intent to pigeon-hole the nodes. Each node is free to develop and acquire its own unique set of attributes.
Type 4: There are fixed number of classes of nodes. Each node within a class is free to develop and acquire its own unique set of attributes. Each class has a restricted set of attributes a node is allowed to acquire.
Read on EAV model to understand the issue of parameters vs attributes. In an EAV table, a node needs only three characterising columns:
node id
attribute name
attribute value
However, under constraints of technology, an attribute could be number, string, enumerable or category. Therefore, there would be four more attribute tables, one for each value type, plus the node table:
node id
attriute type
attribute name
attribute value
Sequential/linked access versus hashed/direct-address access
Do you have to access individual nodes directly rather than traversing the structural tree to get to a node quickly?
Do you need to find a list of nodes that have acquired a particular trait (set of attributes) regardless of where they sit topologically on the network? Do you need to perform classification (aka principal component analysis) on the nodes of your network?
State-machine
Do you wish to perceive the regions of your network as a collection of state-machines?
State machines are very useful quantization entities. State-machine quatization helps you to form empirical entities over a range of nodes based on neighbourhood similarities and relationships.
Instead of trying to understand and track individual behaviour of millions of nodes, why not clump them into regions of similarity. And track the state-machine flow of those regions.
Conclusion
This is my recommendation. You should start initially using a totally relational database. The reason is that relational database and the associated SQL provides information with a very liberal view of relationship. With SQL on a relational model, you could inquire or correlate relationships that you did not know exist.
As your experiments progress and you might find certain relationship modeling more suitable to a network-graph repository, you should then move those parts of the schema to such suitable repository.
In the final state of affairs. I would maintain a dual mode information repo. You maintain a relational repo to keep track of nodes and their attributes. So you store the dynamically mutating structure in a network-graph repository but each node refers to a node id in a relational database. Where the relational database allows you to query nodes based on attributes and their values. For example,
SELECT id FROM Nodes a, NumericAttributes b
WHERE a.attributeName = $name
AND b.value WItHIN $range
AND a.id = b.id
I am thinking, perhaps, hadoop could be used instead of a traditional network-graph database. But, I don't know how well hadoop adapts to dynamically changing relationships. My understanding is that hadoop is good for write-once read-by-many. However, a dynamic neural network may not perform well in frequent relationship changes. Whereas, a relational table modeling network relationships is not efficient.
Still, I believe I have only exposed questions you need to consider rather than providing you with a definite answer, especially with a rusty knowledge on many concepts.
Trees can be stored in a table by using self-referencing foreign keys. I'm assuming the only two things that need to be stored are topology and the weights; both of these can be stored in a flattened tree structure. Of course, this can require a lot of recursive selects, which depending on your RDBMS may be a pain to implement natively (thus requiring many SQL queries to achieve). I cannot comment on the comparison, but hopefully that helps with the relational point of view :)