diff --git a/NEWS.md b/NEWS.md
index 79f8f92..9b9229d 100644
--- a/NEWS.md
+++ b/NEWS.md
@@ -3,9 +3,10 @@
 ### general
 
 - strongly speed-up contingency table computation for heterogeneous=true and max_k=0/1
-- use the more compiler-friendly `stack` (introduced in Julia v1.9) instead of `hcat` for large numbers of columns (if available)
+- use the more compiler-friendly `stack` (introduced in Julia v1.9) instead of `hcat` for large numbers of columns (if available), introduce fast method for sparse columns
 - improve univariate pvalue filtering
 - remove performance bottleneck in three-way `adjust_df`
+- catch incorrect usage of 'meta_data_path' as keyword argument
 
 # v0.19.2 (latest)
 
diff --git a/src/io.jl b/src/io.jl
index f2d139d..cd00842 100644
--- a/src/io.jl
+++ b/src/io.jl
@@ -52,7 +52,7 @@ function load_data(data_path::AbstractString, meta_path::StrOrNoth=nothing; tran
         end
         ld_results = load_jld(data_path, otu_data_key, otu_header_key, meta_data_key, meta_header_key, transposed=transposed)
     else
-        error("$(file_ext) not a valid output format. Choose one of $(valid_data_formats)")
+        error("$(file_ext) not a valid input format. Choose one of $(valid_data_formats)")
     end
 
     ld_results
diff --git a/src/learning.jl b/src/learning.jl
index a1d0b51..82d3d7e 100644
--- a/src/learning.jl
+++ b/src/learning.jl
@@ -298,7 +298,7 @@ end
 function make_table(data_path::AbstractString, meta_data_path::StrOrNoth=nothing;
     otu_data_key::StrOrNoth="otu_data", otu_header_key::StrOrNoth="otu_header", 
     meta_data_key::StrOrNoth="meta_data", meta_header_key::StrOrNoth="meta_header", 
-    verbose::Bool=true, transposed::Bool=false)
+    transposed::Bool=false)
     
     data, header, meta_data, meta_header = load_data(data_path, meta_data_path, otu_data_key=otu_data_key,
                                                      otu_header_key=otu_header_key, meta_data_key=meta_data_key,
@@ -316,6 +316,24 @@ function make_table(data_path::AbstractString, meta_data_path::StrOrNoth=nothing
     data, header, meta_mask
 end
 
+#learn_local_neighborhood(target_var::AbstractString, header, args...; kwargs...) =
+#    learn_local_neighborhood(findfirst(==(target_var), header), args...; kwargs...)
+#
+#function learn_local_neighborhood(target_var::Int, data_path::AbstractString, meta_data_path::StrOrNoth=nothing;
+#    otu_data_key::StrOrNoth="otu_data",
+#    otu_header_key::StrOrNoth="otu_header", meta_data_key::StrOrNoth="meta_data",
+#    meta_header_key::StrOrNoth="meta_header", verbose::Bool=true,
+#    transposed::Bool=false, kwargs...)
+#
+#    verbose && println("\n### Loading data ###\n")
+#    data, header, meta_mask = make_table(data_path, meta_data_path, otu_data_key=otu_data_key,
+#                                         otu_header_key=otu_header_key, meta_data_key=meta_data_key,
+#                                         meta_header_key=meta_header_key, transposed=transposed)
+#    
+#   
+#
+#end
+
 """
     learn_network(data_path::AbstractString, meta_data_path::AbstractString) -> FWResult{<:Integer}
 
@@ -339,6 +357,11 @@ function learn_network(data_path::AbstractString, meta_data_path::StrOrNoth=noth
     meta_header_key::StrOrNoth="meta_header", verbose::Bool=true,
     transposed::Bool=false, kwargs...)
 
+    # Catch incorrect usage of data_path/meta_data_path as keyword argument
+    for key in (:data_path, :meta_data_path)
+        @assert !(key in keys(kwargs)) "'$key' is a positional argument, please use 'learn_network(<data_path>, <meta_data_path>; <kwargs...>)'."
+    end
+
     verbose && println("\n### Loading data ###\n")
     data, header, meta_mask = make_table(data_path, meta_data_path, otu_data_key=otu_data_key,
                                          otu_header_key=otu_header_key, meta_data_key=meta_data_key,
@@ -353,10 +376,12 @@ end
 Works like learn_network(data_path::AbstractString, meta_data_path::AbstractString), but takes paths to multiple data sets (independent sequencing experiments (e.g. 16S + ITS) for the same biological samples) which are normalized independently.
 """
 function learn_network(all_data_paths::AbstractVector{<:AbstractString}, meta_data_path::StrOrNoth=nothing;
-    otu_data_key::StrOrNoth="otu_data",
-    otu_header_key::StrOrNoth="otu_header", meta_data_key::StrOrNoth="meta_data",
-    meta_header_key::StrOrNoth="meta_header", verbose::Bool=true,
-    transposed::Bool=false, kwargs...)
+    otu_data_key::StrOrNoth="otu_data", otu_header_key::StrOrNoth="otu_header", transposed::Bool=false, kwargs...)
+
+    # Catch incorrect usage of data_path/meta_data_path as keyword argument
+    for key in (:all_data_paths, :meta_data_path)
+        @assert !(key in keys(kwargs)) "'$key' is a positional argument, please use 'learn_network(<all_data_paths>, <meta_data_path>; <kwargs...>)'."
+    end
 
     data_path = all_data_paths[1]
     if length(all_data_paths) > 1
@@ -447,6 +472,9 @@ function learn_network(data::AbstractMatrix; sensitive::Bool=true,
     cache_pcor::Bool=false, time_limit::AbstractFloat=-1.0, update_interval::AbstractFloat=30.0, parallel_mode="auto",
     extra_data::Union{AbstractVector,Nothing}=nothing, share_data::Bool=true, experimental_kwargs...)
 
+    @assert !(:meta_data_path in keys(experimental_kwargs)) "You provided a OTU matrix together with a meta data path, this is currently not supported. 
+    Use either 'learn_network(<otu_table_path>, <meta_data_path>; <kwargs...>)' or 'learn_network(<otu_matrix>; <kwargs...>)'."
+
     start_time = time()
 
     cont_mode = sensitive ? "fz" : "mi"
diff --git a/src/preprocessing.jl b/src/preprocessing.jl
index c3169a2..9eb8d12 100644
--- a/src/preprocessing.jl
+++ b/src/preprocessing.jl
@@ -1,10 +1,38 @@
+function _fast_stack_sparse(vecs::Vector{SparseVector{T1, T2}}) where {T1 <: Real, T2 <: Integer}
+    """Fast method for stacking sparse columns"""
+    n_rows = length(vecs[1])
+    @assert all(length(x) == n_rows for x in vecs)
+    
+    rids, cids, nzvals = Int[], Int[], T1[]
+
+    for (col_i, v) in enumerate(vecs)
+        n_val = nnz(v)
+
+        if n_val > 0
+            append!(rids, rowvals(v))
+            append!(cids, repeat([col_i], n_val))
+            append!(nzvals, nonzeros(v))
+        end
+    end
+
+    n_cols = length(vecs)
+    return sparse(rids, cids, nzvals, n_rows, n_cols)
+end
+
 function stack_or_hcat(vecs::AbstractVector{<:AbstractArray})
     # use more efficient stack (introduced in Julia v1.9) if available
-    if isdefined(Base, :stack)
-        return stack(vecs)
+    stacked_matrix = if isdefined(Base, :stack)
+        # use even faster custom implementation for sparse vectors
+        if isa(vecs, AbstractVector{<:SparseVector})
+            _fast_stack_sparse(vecs)
+        else
+            stack(vecs)
+        end
     else
-        return hcat(vecs...)
+        hcat(vecs...)
     end
+
+    return stacked_matrix
 end
 
 
@@ -187,7 +215,7 @@ function discretize(X::AbstractMatrix{ElType}; n_bins::Integer=3, nz::Bool=true,
         rank_method::String="tied", disc_method::String="median", nz_mask::BitMatrix=BitMatrix(undef, (0,0))) where ElType <: AbstractFloat
     if nz
         if issparse(X)
-            disc_vecs = SparseVector{Int}[]
+            disc_vecs = SparseVector{Int,Int}[]
             for j in 1:size(X, 2)
                 push!(disc_vecs, discretize_nz(X[:, j], n_bins, rank_method=rank_method, disc_method=disc_method))
             end
diff --git a/src/types.jl b/src/types.jl
index b2c9f59..ed5feeb 100644
--- a/src/types.jl
+++ b/src/types.jl
@@ -149,7 +149,7 @@ end
 ## RESULT TYPES ##
 ##################
 
-const RejDict{T} = Dict{T,Tuple{Tuple{Int64,Vararg{Int64,N} where N},TestResult,Tuple{Int,Float64}}}
+const RejDict{T} = Dict{T,Tuple{Tuple{Int64,Vararg{Int64,N}} where N,TestResult,Tuple{Int,Float64}}}
 
 struct HitonState{T}
     phase::Char